Silent Risk
Silent Risk
2015 (DRAFT)
ABSTRACT
"Empirical evidence that the boat is safe", or how we tend to be fooled by silent risks. Factum stultus
cognoscit (The fool only understand risks after the harm is done). Risk is both precautionary (fragility based)
and evidentiary (statistical based); it is too serious a business to be left to mechanistic users of probability
theory.
This figure encapsulates the scientific "nonsucker" approach to risk and probability. Courtesy George Nasr.
***
This book provides a mathematical framework for decision making and the analysis of (consequential) hidden risks, those tail events undetected or improperly detected by statistical machinery; and substitutes fragility as a more reliable measure of exposure. Model error is mapped as
risk, even tail risk.1
Risks are seen in tail events rather than in the variations; this necessarily links them mathematically to an asymmetric response to intensity of shocks, convex or concave.
The difference between "models" and "the real world" ecologies lies largely in an additional
layer of uncertainty that typically (because of the same asymmetric response by small probabilities to additional uncertainty) thickens the tails and invalidates all probabilistic tail risk measurements models, by their very nature of reduction, are vulnerable to a chronic underestimation
of the tails.
1 This is a polite way to say No-BS approach to probability.
iii
So tail events are not measurable; but the good news is that exposure to tail events is. In "Fat
Tail Domains" (Extremistan), tail events are rarely present in past data: their statistical presence
appears too late, and time series analysis is similar to sending troops after the battle. Hence the
concept of fragility is introduced: is one vulnerable (i.e., asymmetric) to model error or model
perturbation (seen as an additional layer of uncertainty)?
Part I looks at the consequences of fat tails, mostly in the form of slowness of convergence
of measurements under the law of large number: some claims require 400 times more data
than thought. Shows that much of the statistical techniques used in social sciences are either
inconsistent or incompatible with probability theory. It also explores some errors in the social
science literature about moments (confusion between probability and first moment, etc.)
Part II proposes a more realistic approach to risk measurement: fragility as nonlinear (concave) response, and explores nonlinearities and their statistical consequences. Risk management
would consist in building structures that are not negatively asymmetric, that is both "robust" to
both model error and tail events. Antifragility is a convex response to perturbations of a certain
class of variables.
iv
CONTENTS
Acknowledgments
xvi
Notes for Reviewers xviii
i figuring out probability and what it means
3
1 what is probability? what is exposure?
5
1.1 The Conflation of Events and Exposures
6
1.1.1 Contract Theory
9
1.2 Payoff Classes P1 through P4
12
1.2.1 Atomic Payoff P1
14
1.2.2 Binary Payoff Class P2
15
1.2.3 Vanilla Payoff Class P3 , building blocks for regular exposures.
16
1.2.4 Composite/Sigmoidal Payoff Class P4
20
1.3 Achieving Nonlinearity through P4
20
1.4 Main Errors in the Literature
20
1.5 The Applicability of Some Psychological Biases
21
1.6 Misfitness of Prediction Markets
21
1.6.1 The Black Swan is Not About Probability But Payoff
21
1.6.2 Chernoff Bound
22
1.6.3 Fatter tails lower the probability of remote events (the binary) and raise the
value of the vanilla.
23
1.6.4
The law of large numbers works better with the binary than the variable
23
1.7 Finding Inconsistencies in Scholarly Treatments of Events
25
1.8 Metaprobability and the Payoff Kernel
27
1.9 Classification and Codification of Exposures
31
1.10 Numeraire definition
32
1.11 What is an Insurable Risk?
32
1.12 Ruin problems
32
1.13 Skepticism, Uncertainty, and Scale of a Distributon
32
1.14 Why Pascal Wager Has Nothing to Do With the Left Tail
32
2 the "real world" rigor project
33
2.1 A Course With an Absurd Title
35
2.2 Problems and Inverse Problems
37
2.3 Fragility, not Just Statistics, For Hidden Risks
46
2.4 Solution: The Convex Heuristic
48
2.4.1 Convex Heuristics, Rationality, and Revelation of Preferences
48
2.5 Fragility and Model Error
50
2.5.1 Why Engineering?
50
2.6 General Classification of Problems Related To Fat Tails
52
General Classification of Problems Related To Fat Tails
52
2.7 Closing the Introduction
54
a whats a charlatan in risk and probability?
55
a.1 Charlatan
55
Contents
ii fat tails: the lln under real world ecologies
59
3 fat tails and the problem of induction
61
3.1 The Problem of (Enumerative) Induction
61
3.2 Empirical Risk Estimators
61
3.3 Fat Tails, the Finite Moment Case
64
3.4 A Simple Heuristic to Create Mildly Fat Tails
69
3.5 The Body, The Shoulders, and The Tails
69
3.6 Fattening of Tails With Skewed Variance
72
3.7 Fat Tails in Higher Dimension
74
3.8 Scalable and Nonscalable, A Deeper View of Fat Tails
75
3.9 Subexponential as a class of fat tailed distributions
76
3.10 Joint Fat-Tailedness and Elliptical Distributions
83
3.11 Different Approaches For Statistical Estimators
86
3.12 Econometrics imagines functions in L2 Space
89
3.13 Typical Manifestations of The Turkey Surprise
90
2
3.14 Metrics for Functions Outside L Space
95
3.15 Using the Hilbert Transform
96
3.16 A Comment on Bayesian Methods in Risk Management
96
b special cases of fat tails
99
b.1 Multimodality and Fat Tails, or the War and Peace Model
99
b.2 Transition probabilites: what can break will break
102
c quick and robust measure of fat tails
105
c.1 Introduction
105
c.2 First Metric, the Simple Estimator
105
c.3 Second Metric, the 2 estimator
107
4 hierarchy of distributions for asymmetries
109
4.1 Permissible Empirical Statements
109
4.2 Masquerade Example
110
4.3 The Probabilistic Version of Absense of Evidence
111
4.4 Via Negativa and One-Sided Arbitrage of Statistical Methods
111
4.5 Hierarchy of Distributions in Term of Tails
112
4.6 How To Arbitrage Kolmogorov-Smirnov
115
4.7 Mistaking Evidence for Anecdotes & The Reverse
119
5 effects of higher orders of uncertainty
121
5.1 Meta-Probability Distribution
121
5.2 Metadistribution and the Calibration of Power Laws
122
5.3 The Effect of Metaprobability on Fat Tails
124
5.4 Fukushima, Or How Errors Compound
124
5.5 The Markowitz inconsistency
124
5.6 Psychological pseudo-biases under second layer of uncertainty.
126
5.6.1 The pathologization fallacy 126
6 large numbers and clt in the real world
131
6.0.2 The "Pinker Problem" 131
6.1 The problem of Matching Errors
132
6.2 Generalizing Mean Deviation as Partial Expectation
134
6.3 Class of Stable Distributions
134
6.3.1 Results 135
vi
Contents
6.3.2 Stochastic Alpha or Mixed Samples
136
Symmetric NonStable Distributions in the Subexponential Class
138
6.4.1 Symmetric Mixed Gaussians, Stochastic Mean
138
6.4.2 Half cubic Student T (Lvy Stable Basin)
138
6.4.3 Cubic Student T (Gaussian Basin)
139
6.5 Asymmetric NonStable Distributions in the Subexponetial Class
141
6.5.1 One-tailed Pareto Distributions
141
6.5.2 The Lognormal and Borderline Subexponential Class
141
6.6 Asymmetric Distributions in the Superexponential Class
141
6.6.1 Mixing Gaussian Distributions and Poisson Case
141
6.6.2 Skew Normal Distribution
141
6.6.3 Super-thin tailed distributions: Subgaussians
141
6.7 Acknowledgement
142
6.7.1 Cumulants
142
6.7.2 Derivations using explicit E(| X |)
143
6.7.3 Derivations using the Hilbert Transform and = 0 143
in progress derivations for lln across fat tails
145
d.1 Comments on LLN and Bounds 145
d.1.1 Speed of Convergence for Simple Cases
145
d.1.2 Comparing N = 1 to N = 2 for a symmetric power law with 1< 2. 147
d.2 Digression Into Inversion of Characteristic Function of NonStable Power Law
149
d.2.1 Integrable Characteristic Functions
150
d.3 Pulling the PDF of an n-Summed Student T
150
preasymptotics and central limit in the real world
153
7.1 Using Log Cumulants to Observe Preasymptotics
159
7.2 Convergence of the Maximum of a Finite Variance Power Law
163
7.3 Sources and Further Readings 163
7.4 Convergence for Non-Lvy StablePower Laws Temporarily here)
165
where standard diversification fails
167
fat tails and random matrices
171
some misuses of statistics in social science
175
8.1 Mechanistic Statistical Statements
175
8.2 Attribute Substitution
175
8.3 The Tails Sampling Property 177
8.4 A discussion of the Paretan 80/20 Rule
179
8.5 Survivorship Bias (Casanova) Property
182
8.6 Left (Right) Tail Sample Insufficiency Under Negative (Positive) Skewness
184
8.7 Why N=1 Can Be Very, Very Significant Statistically
186
8.8 The Instability of Squared Variations in Regressions
186
8.9 Statistical Testing of Differences Between Variables
189
8.10 Studying the Statistical Properties of Binaries and Extending to Vanillas
190
8.11 Why Economics Time Series Dont Replicate
190
8.12 A General Summary of The Problem of Reliance on Past Time Series
196
8.13 Conclusion
197
on the instability of econometric data
199
fat tails from recursive uncertainty
201
9.1 Layering uncertainty
201
6.4
e
f
8
g
9
vii
Contents
10
11
12
13
236
viii
Contents
15
16
17
18
19
20
21
22
23
ix
Contents
24 the convexity of wealth to inequality
325
24.1 The One Percent of the One Percent are Divorced from the Rest
25 nonlinearities and risk in medicine
329
25.1 Antifragility from Uneven Distribution
329
325
Contents
Index
370
xi
CHAPTER SUMMARIES
1 Probability defined first things first. Why and how we cannot dissociate probability
from decision. The notion of contract theory. Fallacies coming from verbalistic descriptions of probability. The difference between classes of payoffs with probabilistic
consequences. Formal definition of metaprobability. . . . . . . . . . . . . . . . . . . . . .
2 Outline of the book and project of the codification of Risk and decision theory as related
to the real world (that is "no BS") in nonmathematical language (other chapters are
mathematical). Introduces the main fallacies treated in the project. What can and
should be mathematized. Presents the central principles of risk bearing. Introduces
the idea of fragility as a response to volatility, the associated notion of convex heuristic,
the problem of invisibility of the probability distribution and the spirit of the book.
Explains why risk is in the tails not in the variations. Explains that the layering of
random variables makes more ecological a view that is corresponds tot the "real world"
and how layering of model errors generates fat tails. . . . . . . . . . . . . . . . . . . . .
33
3 Introducing mathematical formulations of fat tails. Shows how the problem of induction gets worse. Empirical risk estimator. Introduces different heuristics to "fatten" tails.
Where do the tails start? Sampling error and convex payoffs. . . . . . . . . . . . . . . .
61
4 Using the asymptotic Radon-Nikodym derivatives of probability measures, we construct a formal methodology to avoid the "masquerade problem" namely that standard
"empirical" tests are not empirical at all and can be fooled by fat tails, though not by
thin tails, as a fat tailed distribution (which requires a lot more data) can masquerade
as a low-risk one, but not the reverse. Remarkably this point is the statistical version of the logical asymmetry between evidence of absence and absence of evidence. We
put some refinement around the notion of "failure to reject", as it may misapply in
some situations. We show how such tests as Kolmogorov Smirnoff, Anderson-Darling,
Jarque-Bera, Mardia Kurtosis, and others can be gamed and how our ranking rectifies
the problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5 The Spectrum Between Uncertainty and Risk. There has been a bit of discussions about
the distinction between "uncertainty" and "risk". We believe in gradation of uncertainty
at the level of the probability distribution itself (a "meta" or higher order of uncertainty.)
One end of the spectrum, "Knightian risk", is not available for us mortals in the real
world. We show how the effect on fat tails and on the calibration of tail exponents and
reveal inconsistencies in models such as Markowitz or those used for intertemporal
discounting (as many violations of "rationality" arent violations . . . . . . . . . . . . . . 121
6 The Law of Large Numbers is the foundation of statistical knowledge or, even (inductive) knowledge tout court. The behavior of the sum of random variables allows us to
get to the asymptote and use handy asymptotic properties. However real life is more
complicated. We cannot
talk about LLN without figuring out the speed of convergence,
which, when it is at n, is only so asymptotically. Further, in some cases the LLN
doesnt work at all. For very fat tailed, under the slightest parametric error, it will be
more than 400 times slower than thought. . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
xiii
Chapter Summaries
7 The behavior of the sum of random variables allows us to get to the asymptote and use
handy asymptotic properties, that is, Platonic distributions. But the problem is that in
the real world we never get to the asymptote, we just get "close" Some distributions get
close quickly, others very slowly (even if they have finite variance). We examine how
fat tailedness worsens the process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8 We apply the results of the previous chapter on the slowness of the LLN and list misapplication of statistics in social science, almost all of them linked to misinterpretation
of the effects of fat-tailedness (and often from lack of awareness of fat tails), and how
by attribute substitution researchers can substitute one measure for another. Why for
example, because of chronic small-sample effects, the 80/20 is milder in-sample (less
fat-tailed) than in reality and why regression rarely works. . . . . . . . . . . . . . . . . .
9 Error about Errors. Probabilistic representations require the inclusion of model (or
representation) error (a probabilistic statement has to have an error rate), and, in the
event of such treatment, one also needs to include second, third and higher order errors
(about the methods used to compute the errors) and by a regress argument, to take the
idea to its logical limit, one should be continuously reapplying the thinking all the way
to its limit unless when one has a reason to stop, as a declared a priori that escapes
quantitative and statistical method. We show how power laws emerge from nested
errors on errors of the standard deviation for a Gaussian distribution. We also show
under which regime regressed errors lead to non-power law fat-tailed distributions. .
10 We present case studies around the point that, simply, some models depend quite a
bit on small variations in parameters. The effect on the Gaussian is easy to gauge, and
expected. But many believe in power laws as panacea. Even if one believed the r.v. was
power law distributed, one still would not be able to make a precise statement on tail
risks. Shows weaknesses of calibration of Extreme Value Theory. . . . . . . . . . . . . .
11 A less technical demarcation between Black Swan Domains and others. . . . . . . . . .
12 We present the difference between ruin problems, particularly when systemic, and
other risk management of a more computable nature. . . . . . . . . . . . . . . . . . . . .
13 Standard economic theory makes an allowance for the agency problem, but not the
compounding of moral hazard in the presence of informational opacity, particularly
in what concerns high-impact events in fat tailed domains (under slow convergence
for the law of large numbers). Nor did it look at exposure as a filter that removes
nefarious risk takers from the system so they stop harming others. (In the language
of probability, skin in the game creates an absorbing state for the agent, not just the
principal). But the ancients did; so did many aspects of moral philosophy. We propose
a global and morally mandatory heuristic that anyone involved in an action which
can possibly generate harm for others, even probabilistically, should be required to
be exposed to some damage, regardless of context. While perhaps not sufficient, the
heuristic is certainly necessary hence mandatory. It is supposed to counter voluntary
and involuntary risk hiding and risk transfer in the tails. . . . . . . . . . . . . . .
14 Deeper into the conflation between a random variable and exposure to it. . . . . . . . .
15 We provide a mathematical approach to fragility as negative sensitivity to a semimeasure of dispersion and volatility (a variant of negative or positive "vega") and examine the link to nonlinear effects. We link to the litterature on model "robustness" and
show how we add nonlinearity to the conventional approaches. . . . . . . . . . . . . . .
16 Presents the fragility theorems and the transfer function between nonlinear response
and the benefits and harm from increased uncertainty. . . . . . . . . . . . . . . . . . . .
17 Presents the IMF fragility heuristics, particularly in the improvement of stress testing. .
xiv
153
175
201
215
227
229
239
251
263
271
287
Chapter Summaries
18 The literature of heavy tails starts with a random walk and finds mechanisms that lead
to fat tails under aggregation. We follow the inverse route and show how starting
with fat tails we get to thin-tails from the probability distribution of the response to a
random variable. We introduce a general dose-response curve show how the left and
right-boundedness of the reponse in natural things leads to thin-tails, even when the
underlying variable of the exposure is fat-tailed. . . . . . . . . . . . . . . . . . . . . . .
19 We extract the effect of size on the degradation of the expectation of a random variable,
from nonlinear response. The method is general and allows to show the "small is beautiful" or "decentralized is effective" or "a diverse ecology is safer" effect from a response
to a stochastic stressor and prove stochastic diseconomies of scale and concentration
(with as example the Irish potato famine and GMOs). We apply the methodology to
environmental harm using standard sigmoid dose-response to show the need to split
sources of pollution across independent . . . . . . . . . . . . . . . . . . . . . . . . . . .
20 Explains why the fragilefragile is necessarily in the nonlinear. . . . . . . . . . . . . . . .
21 Information is convex to noise. The paradox is that increase in sample size magnifies the
role of noise (or luck); it makes tail values even more extreme. There are some problems
associated with big data and the increase of variables available for epidemiological and
other "empirical" research. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22 The one
R percent of the one percent has tail properties such that the tail wealth (expectation K x p(x) dx) depends far more on inequality than wealth. . . . . . . . . . . . . .
23 Examines nonlinearities in medicine /iatrogenics as a risk management problem. . . .
24 A list of common problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25 American Options have hidden optionalities. Using a European option as a baseline we
heuristically add the difference. We also show how these hidden options are extremely
significant compared to the focus of the research literature on insignificant marginal
improvements of the pricing equations but in the Black Scholes World. . . . . . . . . . .
26 Much of the work concerning martingales and Brownian motion has been idealized; we
look for holes and pockets of mismatch to reality, with consequences. Infinite (or undefined) higher moments are not compatible with Ito calculus outside the asymptote.
Path dependence as a measure of fragility. . . . . . . . . . . . . . . . . . . . . . . . . . .
27 Proof that under simple assumptions, such as constraints of Put-Call Parity, the probability measure for the valuation of a European option has the mean derived from
the forward price which can, but does not have to be the risk-neutral one, under any
general probability distribution, bypassing the Black-Scholes-Merton dynamic hedging
argument, and without the requirement of complete markets and other strong assumptions. We confirm that the heuristics used by traders for centuries are both more robust,
more consistent, and more rigorous than held in the economics literature. We also show
that options can be priced using infinite variance (finite mean) distributions. . . . . . .
297
303
315
319
325
329
335
337
343
355
xv
acknowledgments
The text is not entirely that of the author. Four chapters contain recycled text written with collaborators in standalone articles: the late Benoit Mandelbrot (section of slowness of LLN under
power laws, even with finite variance), Elie Canetti and the stress-testing staff at the International
Monetary Fund (for the heuristic to detect tail events), Phil Tetlock (binary vs variable for forecasting), Constantine Sandis (skin in the game) and Raphael Douady (mathematical mapping
of fragility). But it is the latter paper that represents the biggest debt: as the central point of
this book is convex response (or, more generally, nonlinear effects which subsume tail events),
xvi
Chapter Summaries
the latter paper is the result of 18 years of mulling that single idea, as an extention of Dynamic
Hedging [108] applied outside the options domain, with 18 years of collaborative conversation
with Raphael before the actual composition!
This book is in debt to three persons who left us. In addition to Benoit Mandelbrot, this author
feels deep gratitude to the late David Freedman, for his encouragements to develop a rigorous
model-error based, real-world approach to statistics, grounded in classical skeptical empiricism,
and one that could circumvent the problem of induction: and the method was clear, of the type
"dont use statistics where you can be a sucker" or "figure out where you can be the sucker".
There was this "moment" in the air, when a group composed of the then unknown John Ioannidis, Stan Young, Philip Stark, and others got together I was then an almost unpublished and
argumentative "volatility" trader, something people couldnt quite understand unless expressed
as "nonlinear payoffs", even then (Dynamic Hedging was unreadable to nonspecialists) and felt
that getting David Freedmans attention was more of a burden than a blessing, as it meant a
severe obligation.2
Indeed this exact book project was born from a 2012 Berkeley statistics department commencement lecture, given in his memory, with the message: "statistics is the most powerful weapon
today, it comes with responsibility" (since numerical assessments increase risk taking) and the
corrolary, directly taken from his legacy:
"Understand the models errors before you understand the model".
leading to the theme of this book, that all one needs to do is figure out the answer to the
following question:
Are you convex or concave to model errors?
xvii
Chapter Summaries
"outsider" and strange "volatility" trader or "quant" wild animal. But they soon got disappointed
that the animal was not much of a wild animal but an orthodox statistician, actually overzealous
about a nobullshit approach. I thank Wolfgang Hrtle for, before this book was started in the
following form, a full-day seminar at Humboldt University and Pantula Sastry for providing the
inaugurating lecture of the International Year of Statistics at the National Science Foundation.
Carl Tony Fakhry has taken the thankless task of diligently rederiving every equation (at the
time of writing he has just reached Chapter 3). I also thank Wenzhao Wu and Mian Wang for
list of typos.
to the reader
The text can be read by (motivated) non-quants: everything mathematical in the text is accompanied with a "literary" commentary, so in many sections the math can be safely skipped. Its
mission, to repeat, is to show a risk-taker perspective on risk management, integrated into the
mathematical language, not to lecture on statistical concepts.
On the other hand, when it comes to math, it assumes a basic "quant level" advanced or
heuristic knowledge of mathematical statistics, and is written as a monograph; it is closer to a
longer research paper or old fashioned treatise. As I made sure there is little overlap with other
books on the subject, I calibrated this text to the wonderful textbook by the late A. Papoulis
Probability, Random Variables, and Stochastic Processes [84]: there is nothing basic discussed in this
text that is not defined in Papoulis.
For more advanced, more mathematical, or deeper matters such as convergence theorems, the
text provides definitions, but the reader is recommended to use Loeves two volumes Probability
Theory [70] and [71] for a measure theoretic approach, or Fellers two volumes, [40] and [39] and,
for probability bounds, Petrov[86]. For extreme value theory, Embrecht et al [31] is irreplaceable.
status/completion of book
This is a first draft for general discussion, not for equation-wise verification. There are still typos,
errors and problems progressively discovered by readers thanks to the dissemination on the web.
The bibliographical references are not uniform, they are in the process of being integrated into
bibtex.
Note that there are redundancies that will be removed at the end of the composition.
August 2014 status: After completing most of the math sections, I started putting words and
structure around the concepts, so I am progressively introducing "definitions", "remarks",
and comments in plain English, inspired by both Bourbaki and legal codifications. The idea
is to codify and explain all terms to clear up the usual confusions. As of August 2014, I am
only 15% done.
Below is the list of the incomplete sections.
Incomplete Sections in Part I (mostly concerned with limitations of measurements of tail probabilities)
i A list of symbols.
xviii
Chapter Summaries
ii Chapter 3 proposes a measure of fattailedness based on ratio of Norms for all( superexponential, subexponential, and powerlaws with tail exponent >2); it is more powerful than
Kurtosis since we show it to be unstable in many domains. It lead us to a robust heuristic
derivation of fat tails. We will add an Appendix comparing it to the Hill estimator.
iii An Appendix on the misfunctioning of maximum likelihood estimators (extension of the
problem of Chapter 3).
iv In the chapter on pathologies of stochastic processes, a longer explanation of why a stochastic
integral "in the real world" requires 3 periods not 2 with examples (event information for
computation of exposureXt order Xt+ t execution Xt+2t ).
v The "Weron" effect of recovered from estimates higher than true values.
vi A lengthier (and clearer) exposition of the variety of bounds: MarkovChebychevLusin
BerhshteinLyapunov Berry-Esseen Chernoff bounds with tables.
vii A discussion of the Von Mises condition. A discussion of the Cramr condition. Connected:
Why the research on large deviations remains outside fat-tailed domains.
viii A discussion of convergence (and nonconvergence) of random matrices to the Wigner semicirle, along with its importance with respect to Big Data
ix A section of pitfalls when deriving slopes for power laws, with situations where we tend to
overestimate the exponent.
incomplete sections
(mostly concerned with building exposures and convexity of payoffs: What is and What is Not
"Long Volatility")
i A discussion of gamblers ruin. The interest is the connection to tail events and fragility.
"Ruin" is a better name because the idea of survival for an aggregate, such as probability of
ecocide for the planet.
ii An exposition of the precautionary principle as a result of the fragility criterion.
iii A discussion of the "real option" literature showing connecting fragility to the negative of
"real option".
iv A link between concavity and iatrogenic risks (modeled as short volatility).
v A concluding chapter.
Best Regards,
Nassim Nicholas Taleb
November 2014
xix
Figure 0.1: Connecting disciplines. Most economists lack too much in erudition (mathematical and philosophical) to connect the point discussed here
to its historical context so they mistake the idea of The Black Swan for something they may have been exposed to in their past. Further, shows the "Fat
Tonyism", i.e. real world rigor.
Chapter Summaries
Chapter Summaries
Tail Risk
98% tunnel
Domain of
(Mechanistic)
Statistical
Techniques
-15
-10
-5
Thin-tailed
Distribution
(probability
Fat-tailed
swamps payoff)
Distribution
(payoff swamps
probability)
-22
-20
-18
-16
-14
-12
-10
Part I
F I G U R I N G O U T P R O B A B I L I T Y A N D W H AT I T M E A N S
W H AT I S P R O B A B I L I T Y ? W H AT I S
EXPOSURE?
Chapter Summary 1: Probability defined first things first. Why and how we cannot
dissociate probability from decision. The notion of contract theory. Fallacies coming
from verbalistic descriptions of probability. The difference between classes of payoffs
with probabilistic consequences. Formal definition of metaprobability.
The larger mission is as follows:
The project both real-world and anti-anecdotal is inspired of the many historical efforts
and projects aimed to instil rigor in domains that grew organically in a confused way, by
starting from the basics and expanding, Bourbaki-style, in a self-contained manner but
aiming at maximal possible rigor. This would be a Bourbaki approach but completely
aiming at putting the practical before the theoretical, a real-world rigor (as opposed to
Bourbakis scorn of the practical and useful, justified in a theoretical field like mathematics).
The "first thing" is not quite defining probability but rather formally mapping the pair probability and "event" under consideration, subsumed by the notion of probability the two are
inseparable. There has been a long tradition of attempts to define probability, with tons of discussions on what probability is, should be, can be, and cannot be. But, alas these discussions are
at best minute Byzantine nuances, the importance of which has been magnified by the citation
ring mechanism described in Chapter2; these discussions are "academic" in the worst sense of
the word, dwarfed by the larger problem of:
What is the random "event" under concern? Is it an "event" or something more complicated,
like a distribution of outcomes with divergent desirability?
How should we use "probability": probability is not an end product but an input in a larger
integral transform, a payoff kernel.
We have done statistics for a century without a clear definition of probability (whether it is
subjective, objective, or shmobjective plays little role in the equations of probability: all these
probabilities end up adding up to 1 and following the same rules of calculus). But what matters
significantly is the event of concern, which is not captured by the verbalistic approaches to probability.1 Trying to define "what is fire" with academic precision is not something a firefighter
should do, as interesting as it seems, given his higher priorities of figuring out the primary
(nonacademic) variable, what is (and what is not) burning. Almost all these definitions of fire will
end up burning the building in the same manner. People whithout skin in the game (nonfirefighters) who spend time worrying about the composition of fire, but not its effect, would remain
in the gene pool and divert scientific pursuit into interesting but inconsequential directions.2
1 In my basement I have shelves and shelves of treatises trying to define probability, from De Finetti, Keynes, von Mises,
... See Gillies for the latest approach. Compared to the major problems with metaprobability are mere footnote, as I am
showing here by confining such discussion to a footnote.
2 For an example of Byzantine concerns about probability so detailed and diverted from planet earth that they miss
everything of relevance to risk, see the works of David Aldous on the central difference between "finite additivity" or
"countable additivity", which can be classified as the hijacking of the most important discipline in the world, probability,
by scholastic distinctions without (or with relatively minor) real-world difference.
1.1
Probability Distribution of x
Figure 1.1: The conflation of x and f (x): mistaking the statistical properties of the exposure to a variable
for the variable itself. It is easier to modify exposure to get tractable properties than try to understand x.
This is more general confusion of truth space and consequence space.
Figure 1.2: When you use the services of a lawyer for a contract, you are working on limiting or constructing
f (x) your exposure, where your risk and liability start and end. This 13th C. treatise by the legal and
theological scholastic philosopher Pierre de Jean Olivi provide vastly more rigorous codification and deeper
treatment of risk and probability than the subsequent mathematical ones grounded in the narrower ludic
dimension (i.e., confined to games) by Fermat, Pascal, Huyguens, even De Finetti. Why? Because one can
control exposure via contracts and structures rather than just narrowly defined knowledge of probability.
Further, a ludic setup doesnt allow or perturbation of contractual agreements, as the terms are typically
fixed.
I grew up under the rule that it is more reliable to modify f (x) to the point where one can be
satisfied with the reliability of the risk properties than try to understand the statistical properties
of x, particularly under fat tails.4
4 The reason decision making and risk management are inseparable is that there are some exposure people should never
take if the risk assessment is not reliable, which, as we will see with the best map fallacy, is something people understand
in real life but not when modeling.
This is not a joke: the "give us something better" has been a recurring problem this author has
had to deal with for a long time.
1.1.1
Contract Theory
The rigor of the 13th Century legal philosopher Pierre de Jean Olivi is as close to our ambition as
that of Kolmogorov and Paul Lvy. It is a fact that stochastic concepts such as probability, contingency, risk, hazard, and harm found an extreme sophistication in philosophy and legal texts,
from Cicero onwards, way before probability entered our vocabulary and of course probability
was made poorer by the mental gymnastics approach and the ludic version by Fermat-PascalHuygens-De Moivre ...
Remark 1.1 (Science v/s Contract Theory).
Science is typically in binary space (that is, True/False) as defined below, not about exposure, while risk
and decisions are necessarily in standard real-world full payoff space. Contract theory is in exposure space.
Risk management is closer to the latter.
Remark 1.2 (Derivatives Theory).
Option theory is mathematical contract theory.5
5 I thank Eric Briys for insights along these lines.
10
Fourier Transform
of (+ ): b1 (t)
1
E()+
Atomic
Function
notation
1
P2
Binary
+2 , 2
(1 K ) + eit K
P3
Vanilla
+3 , 3
(1R K )
+ K eit dPx
K K +
R
K x dPx
P4a
P4b
Comp.
Gen. Sigm.
bi (wi t)
R i E(i )
E(i ) d
Class
Pi
P1
Name
p(x)
Beyond the trivial, this example illustrates the common confusion between a bet and an
exposure. A bet is a binary outcome, an exposure has more nuanced results and depends on
full distribution.
When we go deeper into the subject, many less obvious, or less known paradox-style problems
occur. Simply, it is of the opinion of the author, that it is not rigorous to talk about probability
as a final product, or even as a foundation of decisions.
The vanillas add a layer of complication: profits for companies or deaths due to terrorism or
war can take many, many potential values. You can predict the company will be profitable,
but the profit could be $1 or $10 billion.
The conflation binary-vanilla is a mis-specification often made in probability, seen in as fundamental texts as in J.M. Keynes approach to probability [64]. Such a conflation is almost always
present in discussions of "prediction markets" and similar aberrations; it affects some results
in research. It is even made in places by De Finetti in the assessment of what makes a good
"probability appraiser"[22].6
The central point here is that decision-making is not about being a good probability appraiser
life is not about probability as a standalone concept but something more complex in which
probability only enters as a kernel, or integral transform.
The designation "vanilla" originates from definitions of financial contracts.7
Example 1.2 (Too much snow).
The owner of a ski resort in the Lebanon, deploring lack of snow, deposited at a shrine of the Virgin Mary
a $100 wishing for snow. Snow came, with such abundance, and avalanches, with people stuck in the cars,
so the resort was forced to close, prompting the owner to quip "I should have only given $25". What the
owner did is discover the notion of nonlinear exposure under tail events.
Example 1.3 (Predicting the "Crisis" yet Blowing Up).
The financial firm Morgan Stanley correctly predicted the onset of a subprime crisis, but they misdefined
the event they called "crisis"; they had a binary hedge (for small drop) and ended up losing billions as the
crisis ended up much deeper than predicted.
6 The misuse comes from using the scoring rule of the following type:if a person gives a probability p for an event A, he
is scored (p 1)2 or p2 , according to whether A is subsequently found to be true or false. Consequences of A or the fact
that there can be various versions of such event are, at best, an afterthought.
7 The vanilla designation comes from option exposures that are open-ended as opposed to the binary ones that are called
exotic; it is fitting outside option trading because the exposures they designate are naturally occurring continuous
variables, as opposed to the binary that which tend to involve abrupt institution-mandated discontinuities.
11
10
!10
!20
xi
50
0
!50
!100
!150
!200
!250
Figure 1.3: Comparing payoff in classes P2 to those in P3 (top), or binaries to the vanilla. The vertical
payoff shows xi , (x1 , x2 , ...) and the horizontal shows the index i= (1,2,...), as i can be time, or any other
form of classification. We assume in the first case payoffs of {-1,1}, and open-ended (or with a very remote
and unknown bounds) in the second.
As we will see, under fat tails, there is no such thing as a typical event, and nonlinearity
widens the difference between verbalistic and precisely contractual definitions.
1.2
12
Payoff
20
20
40
60
80
100
120
-20
-40
-60
Serious
Harm from
Decline
Benefits from
Decline
Starting
Point
Payoff
10
80
100
120
140
Figure 1.6: Even more confusing: exposure to events in class P4 that escape straightforward verbalistic descriptions. Option traders call this a "butterfly
x exposure" in the jargon.
-5
13
-2
-1
The "primitive" state variable xt is continuously observed between discrete periods T t and
T. The payoff or exposure function is 1t> where = {inf(t) : xt
/ A , t T }, a stopping-time
conditional discretization of the continuously sampled time.8
The "payoff kernel" at time T is a member of the exhaustive and mutually exclusive followb (t) (the
ing 4 classes. We write its probability distribution () and characteristic function
distributions of the payoff under the law of state variable x between T t and T, itself
taken as a random variable) at T, and p( x ) the probability law for x at T.
Note that the various layers are obtained by integration over the state variable x over segments
of the domain D :
Z
i =
i1 ( x ) d x
1.2.1
Atomic Payoff P 1
K D
(x K) dx = 1 and
K D
/
(x K) dx = 0 otherwise.
14
1.5
- (K + 2) +
1
2
+ (K - 2)
1.0
-6
-4
-2
i
(t, K) = i
t 1
t
Z
=
D
(i t (x K))
(1.1)
(x K)p(x)dx
Hence
1.2.2
Z K
D
1 (x)dx
which gives us, writing (for clarity) x for the state variable in the integrand and X for the observed one:
+2 (X, K) =
and
2 (K)
1
0
Z D+
K
if X K;
if X < K.
1 (x)dx
giving us:
2 (X, K)
=
0
1
if X > K;
if X K.
15
Remark 1.6.
The class P 2 is closed under affine transformation a H + a L , for all combinations { a H , a L : a H x +
a L D } . This is true for affine transformations of all payoff functions in 2 , the unit of payoff
becoming a H + a L and the lower (upper) bound a L (a H ).
Proposition 1.1 (Binaries are Thin-Tailed).
The probability distribution ( 2 ), a "binary" payoff is a Bernouilli regardless of the underlying probability distribution over the state variable x.
Proof. First consider that +2 can be written as +2 ( x ) =
b 2+ (t , K ):
function
b 2+ (t , K ) =
Z
<K
1
2 (1
e 2 i t ( 1+sgn( x K ) ) p( x ) d x
Z
p( x ) d x +
(1.2)
e i t p( x ) d x
So, with K P( X K ),
b 2+ (t , K ) = (1 K ) + e i t K
Note that we proved that 2 is subgaussian as defined in [61] regardless of p( x ) the probability distribution of the state variable, even if p( x ) has no moments.
1.2.3
Z X
D
2 (x K)dx
XK
0
and
3 (X, K) =
Z D+
X
if X K;
if X < K.
2 (x)dx
giving us:
3 (X, K)
=
KX
0
if X K;
if X > K.
Assume the support spans the real line. The characteristic function (t, K) can be expressed
as:
Z
1
(t, K) =
p(X)e 2 it(X K)(sgn(X k)+1) dX
16
3, I
3 , II
7
6
5
4
3
2
1
10
8
6
4
2
5
-5
xT
3 , III
-5
-2
-4
-6
-8
-10
-5
xT
3 , IV
5
xT
-5
-1
-2
-3
-4
-5
-6
-7
xT
which becomes
(t, K) = (1 K ) + eitK
Z
K
eitx p(x)dx
(1.3)
E (+3 ) =
Z
K
x p(x) dx K K
(1.4)
which maps to common derivatives pricing such as the Bachelier approach[5] or it Lognormal
generalizations popularized with [11].
As we can see Eq. 1.4 doesnt depend on the portion of the tail of the distribution below K. Of
interest is the "stub" part of the pricing, which represents the difference between the vanilla and
the binary of same strike K:
+ (K) E(+3 K+2 ) =
Z
K
x p(x) dx
(1.5)
17
Figure 1.10: Stable Distributions: remarkably the three have exactly the same mean
and mean deviation, but different symmetry parameter.
0.15
0.10
0.05
-10
-5
CVar
PVar
0.75
0.16
0.70
0.15
0.65
0.14
0.60
0.55
0.13
0.50
0.12
0.45
-1.0
-0.5
10
0.11
0.5
1.0
-1.0
-0.5
0.0
0.5
1.0
The contract has the convenience of sensitivity to fat tails (or other measures of uncertainty
such as the scale of the distribution), as it extracts the "tail", segment of the distribution above
(below) K.
R
R
The idea is to compare K x p( x ) d x and K p( x ) d x and see how they react in opposite
directions to certain parameters that control the fatness of tails.
Remark 1.7 (Symmetry/Skewness Problem).
There exists a nondegenerate distribution p*(x) with E p (X) = E p (X) and E p (| X |s ) = E p (| X |s ) for
s 2 such that:
Z
Z
sgn
x p (x) dx
x p(x) dx
K
K
Z
(1.6)
Z
= sgn
p (x) dx
p(x) dx
K
Proof. The sketch of a proof is as follows. Just consider two "mirror" asymmetric distributions,
p1 and p2 , with equal left and right side expectations.
R
R0
+
18
3
p1 (x) = 2
( x)2
3
3( x)Ai
2
3
( x)
3 22/3 32
+ 3 3 2Ai0
2/3 3
2
e
32
( x)3
273
and
3 32 / 3
3
p2 (x) = 2
( x)3
273
3
2
3( x)Ai
( x)2
3 22/3 3 32
+ 3 2Ai0
(x )2
3 22/3
32
3 32 / 3
3
E+p1
E+p2
2
, E
p1 =
2
2
2
3
3
3
3
2
2
=
=
,
E
p
2
2
3
23
2
1
, P+p1 =
3
3
2
1
P+p2 = , P+p1 =
3
3
Moving the beta parameter which controls symmetry (and, only symmetry) to change the distribution
have the effect of moving probabilities without altering expectations.
P+p1 =
Stochastic Volatility Divergence Let s be the scale of the distribution with density ps (x). Consider the ratio of densities;
(x )2
22
x
erfc
9 For instance, the Skew-Normal N(, , ; x), where R controls the skewness, with PDF
, has mean
2
q
r
2
2
2 + and standard deviation 1 22 , meaning the manipulation of leads to change in expectation and
( +1)
+1
scale. The same applies to the mirrored Lognormal (where skewness and expectation depends on variance) and the
Pareto Distribution (where the tail exponent controls the variance and the mean deviation if these exist.
19
R
We will ferret out situations in which K x p(x) dx (the "Cvar" or conditional value at risk)
R
and K p(x) dx (the Probability associated with "VaR" or value-at-risk) react to tail fattening
situations in opposite manner.
1.2.4
+j +i > 1 (K j ) + j i> 1 (K j )
j=1
1.3
1.4
20
1.5
1.6
Alleged
Bias
Derived
P2
Misspecified
domain
Justified
main
Comparing
risks of driving
vs flying
Overestimation Open-ended
of
small payoffs in fatprobabilities tailed domains
Bounded bets
in
laboratory
setting
Long
bias
Lotteries
in
Dread Risk
1.6.1
do-
shot
In short, the vanilla has another dimension, the payoff, in addition to the probability, while the
binary is limited to the probability. Ignoring this additional dimension is equivalent to living in
a 3-D world but discussing it as if it were 2-D, promoting the illusion to all who will listen that
such an analysis captures all worth capturing.
Now the Black Swan problem has been misunderstood. We are saying neither that there must
be more volatility in our complexified world nor that there must be more outliers. Indeed, we
may well have fewer such events but it has been shown that, under the mechanisms of fat tails,
their impact gets larger and larger and more and more unpredictable.
Two points.
Binary predictions are more tractable than standard ones First, binary predictions tend to
work; we can learn to be pretty good at making them (at least on short timescales and with
rapid accuracy feedback that teaches us how to distinguish signals from noise all possible in
forecasting tournaments as well as in electoral forecasting see Silver, 2012). Further, these
are mathematically tractable: your worst mistake is bounded, since probability is defined on the
interval between 0 and 1. But the applications of these binaries tend to be restricted to manmade
things, such as the world of games (the ludic domain).
21
Application
Questionable
domain
Justified
main
do-
Prediction
markets
Revolutions
Elections
Prediction
markets
"Crashes"
in
Natural
Markets (Finance)
Sports
Forecasting
Judging
by
frequency
in
venture
capital and other
winner take all
domains;
It is important to note that, ironically, not only do Black Swan effects not impact the binaries,
but they even make them more mathematically tractable, as will see further down.
Binary predictions are often taken as a substitute for standard ones Second, most non-decision
makers tend to confuse the binary and the vanilla. And well-intentioned efforts to improve
performance in binary prediction tasks can have the unintended consequence of rendering us
oblivious to catastrophic vanilla exposure.
Remark:More technically, for a heavy tailed distribution (defined as part of the subexponential family),
with at least one unbounded side to the random variable (one-tailedness), the variable prediction record
over a long series will be of the same order as the best or worst prediction, whichever in largest in absolute
value, while no single outcome can change the record of the binary.
1.6.2
Chernoff Bound
The binary is subjected to very tight bounds. Let ( Xi )1<in bea sequence independent Bernouilli
trials taking values in the set {0, 1}, with P(X = 1]) = p and P(X = 0) = 1 p, Take the sum
Sn = 1<in Xi . with expectation E(Sn )= np = . Taking as a distance from the mean, the
Chernoff bounds gives:
For any > 0
P(S (1 + ))
22
e
(1 + )1+
2
3
Let us compute the probability of coin flips n of having 50% higher than the true mean, with
2
p= 12 and = n2 : P S 23 n2 2e 3 = en/24
which for n = 1000 happens every 1 in 1.24 1018 .
1.6.3
Fatter tails lower the probability of remote events (the binary) and raise the value of the
vanilla.
The following intuitive exercise will illustrate what happens when one conserves the variance
of a distribution, but fattens the tails by increasing the kurtosis. The probability of a certain
type of intermediate and large deviation drops, but their impact increases. Counterintuitively,
the possibility of staying within a band increases.
Let x be a standard Gaussian random variable with mean 0 (with no loss of generality) and
standard deviation
. Let P>1 be the probability of exceeding one standard deviation. P>1 =
1 12 erfc 1 , where erfc is the complementary error function, so P>1 = P<1 '15.86% and
2
the probability of staying within the stability tunnel between 1 is 1 P>1 P<1 '
68.3%.
Let us fatten the tail in a variance-preserving manner, using the barbell standard method
of linear
combination of two Gaussians with two standard deviations separated by 1 + a
and 1 a , a (0,1), where a is the vvol (which is variance preserving, technically of no
big effect here, as a standard deviation-preserving spreading gives the same qualitative result).
E( x 4 )
Such a method leads to the immediate raising of the standard Kurtosis by 1 + a2 since
2 =
E( x 2 )
3 a2 + 1 , where E is the expectation operator.
P >1 = P<1
1
1
1
1
= 1 erfc
erfc
2
2
2 1a
2 a+1
(1.7)
So then, for different values of a in Eq. 1 as we can see in Figure 2, the probability of staying
inside 1 sigma rises, rare events become less frequent.
Note that this example was simplified for ease of argument.
the tunnel
r In fact
r inside of
1
1
which fat tailedness increases probabilities is between 2 5 17 and
2 5 17
(even narrower than 1 in the example,
r as it numerically corresponds to the area between -.66
and .66), and the outer one is 12 5 + 17 , that is the area beyond 2.13 .
1.6.4
The law of large numbers works better with the binary than the variable
Getting a bit more technical, the law of large numbers works much faster for the binary than the
variable (for which it may never work, see Taleb, 2013). The more convex the payoff, the more
observations one needs to make a reliable inference. The idea is as follows, as can be illustrated
by an extreme example of very tractable binary and intractable variable.
23
Binary
Bet
Level
x
Let xt be the realization of the random variable X (-, ) at period t, which follows a
Cauchy distribution with p.d.f. f ( xt ) (( x 11)2 +1) . Let us set x0 = 0 to simplify and make
0
the exposure symmetric
R around 0. The variable exposure maps to the variable xt and has an
expectation E ( xt ) = xt f (x)dx, which is undefined (i.e., will never converge to a fixed value).
A bet at x0 has a payoff mapped by as a Heaviside Theta Function R> x0 (xt ) paying 1 if xRt > x0 and
0 otherwise. The expectation of the payoff is simply E((x)) = > x0 (x) f (x)dx= x f (x)dx,
0
which is simply P(x > 0). So long as a distribution exists, the binary exists and is Bernouilli
distributed with probability of success and failure p and 1p respectively .
The irony is that the payoff of a bet on a Cauchy, admittedly the worst possible distribution
to work with since it lacks both mean and variance, can be mapped by a Bernouilli distribution,
about the most tractable of the distributions. In this case the variable is the hardest thing to
estimate, and the binary is the easiest thing to estimate.
n
xti the average payoff of a variety of variable bets xti across periods ti , and S n =
Set Sn = n1 i=1
n
i=1 > x0 ( xti ). No matter how large n, limn S n has the same properties the exact same
probability distribution as S1 . On the other hand limn S n= p; further the presaymptotics
of S n are tractable since it converges to 21 rather quickly, and the standard deviations declines
q
q
q
V (S 1 )
(1 p)p
at speed n , since V S n =
=
(given that the moment generating function
n
n
n
for the average is M(z) = pez/n p + 1 ).
1
n
24
n
P ( X > i=1
x ti )
=1
(1.8)
(1.9)
The sum of binaries converges at a speed faster or equal to that of the variable.
The sum of binaries is never dominated by a single event, while that of the variable can be.
How is the binary more robust to model error?
R
In the more general case, the expected payoff ofR the variable is expressed as A xdF(x) (the
unconditional shortfall) while that of the binary= `A dF(x), where A is the part of the support of
interest for the exposure, typically A[K,), or (,K]. Consider model error as perturbations
in the parameters that determine the calculations of the probabilities. In the case of the variable,
the perturbations effect on the probability is multiplied by a larger value of x.
As an example, define
R a slighly more complicated
R variable than before, with option-like characteristics, V(, K) K x p (x)dx and B(, K) K p (x) dx, where V is the expected payoff
of variable, B is that of the binary, K is the strike equivalent for the bet level, and with x[1,
) let p (x) be the density of the Pareto distribution with minimum value 1 and tail exponent ,
so p (x) x 1 .
Set the binary at .02, that is, a 2% probability of exceeding a certain number K, corresponds
to an =1.2275 and a K=24.2, so the binary is expressed as B(1.2, 24.2). Let us perturbate ,
the tail exponent, to double the probability from .02 to .04. The result is B(1.01,24.2)
B(1.2,24.2) = 2. The
corresponding effect on the variable is
more sensitive than the binary.
-
1.7
V(1.01,24.2)
V(1.2,24.2)
Effectively the probability of war dropped, but the risk got bigger, yet historians insisted that
their business is not probability. Their answer was of the sort "we deal with events defined as
25
Let us play the game. With the event space, define as a binary relation such that event
A is more "important" than B if A, B , A BorA % B at least equal to B, then it looks like we
can elicit from the historian that, "in general" (being careful about what the "general" means):
A % B if c(A) c(B)
N+
where c :
is quantitative measure of casualties, measured in number of death or similar
metrics. Our questions Q1 and Q2 can establish monotonicity of the ordering relation.
We can assert that the historian is in fact not in binary space, even if he lives somewhat in the
illusion that he is, otherwise it would lead to inconsistencies in the simplified ordering relation.10
We can continue the game for the satisfaction of certain axioms which would allow us to
assert that in fact our judgment of historical events lines up to their risk, which, unavoidably, is
quantitative. We can even adjust for the "severity" of events, where the binary relation is violated
"except" when casualties are k greater, such that
k 1 : c(A) k c(B) A % B
and still find out that, for large events, history while not being quantitative still depends on a
quantitative ranking of severity. Given that we are in the tails business (thats what risk is
about), history is in fact convincingly vanilla not binary.
"Exposure" (Hence Risk) Needs to be Far More Rigorous Than "Science"
People claiming a "scientific" approach to risk management needs to be very careful about what
"science" means and how applicable it is for probabilistic decision making. Science consists in
a body of rigorously verifiable (or, equivalently, falsifiable), replicable, and generalizable claims
and statements and those statements only, nothing that doesnt satisfy these constraints. Science
scorns the particular. It never aimed at covering all manner of exposure management, and never
about opaque matters. It is just a subset of our field of decision making. We need to survive
by making decisions that do not satisfy scientific methodologies, and cannot wait a hundred
years or so for these to be establishedsimply, extinction is an absorbing barrier. So phronetic
approaches such as [41] or a broader class of matters we can call "wisdom" and precautionary
actions are necessary. But not abiding by naive "evidentiary science", we embrace a larger set
of human endeavors; it becomes necessary to build former protocols of decision akin to legal
codes: rigorous, methodological, precise, adaptable, but certainly not standard "science" per se.
We will discuss the scientism later; for now consider a critical point. Textbook knowledge is
largely about "True" and "False", which doesnt exactly map to payoff and exposure.
10 We could go deeper and express "fuzziness" about the importance of an event or a set of events as second-order effect
similar to metaprobability modeling.
26
1.8
One must never accept a probability without probabilizing the source of the statement. In other
words, if someone who is your sole source of information tells you "I am 100% certain", but you
think that there is a 1% probability that the person is a liar, the probability must no longer be
treated as 100% but 99% or so, perhaps even lower.11 If you look at trust as an "epistemological
notion" (Origgi, [83]), then the degree of trust maps directly to the metaprobability.
Risk, Uncertainty, and Layering: Historical Perspective
Principle 1.3 (The Necessity of Layering).
No probability without metaprobability. One cannot make a probabilistic statement without considering
the probability of a statement being from an unreliable source, or subjected to measurement errors.
We can generalize to research giving certain conclusions in a dubious setup, like many "behavioral economics" results about, say, hyperbolic discounting (aside from the usual problem of
misdefining contracts).
Definition 1.8 (Knightian Uncertainty).
It corresponds to a use of distribution with a degenerate metadistribution, i.e., fixed parameters devoid of
stochasticity from estimation and model error.
Remark 1.9 (A heuristic to spot incompetence).
There is no such thing as "Knightian risk" in the real world, but gradations of computable risk. A good
heuristic is to disqualify any adult who uses the idea of "Knightian risk" as incompetent.
We said that no probability without a payoff, and no probability without a metaprobability (at
least), which produces a triplet 1) exposure, 2) probability, 3) metaprobability.
11 I owe this to a long discussion with Paul Boghossian; it is remarkable how nonphilosophers have a rough time thinking
of the meta-issue.
27
p= i p i
p1
p2
p3
p4
f (x) p (x)
1
1
f (x) p (x)
2
2
f (x) p (x)
3
3
f (x) p (x)
4
4
Figure 1.13: The idea of metaprobability Consider that uncertainty about probability can still get us a
unique measure P equals the weighted average of the states i , with i = 1; however the nonlinearity
of the response of the probability to requires every possible value of to be taken into account. Thus
we can understand why under metaprobabilistic analysis small uncertainty about the probability in the
extreme left tail can cause matters to blow up.
28
i D,
Z
D
p i (x) dx = 1
p, f , (x) [p f ](x)
i D
f (x, i )i p i (x).
(1.10)
i D
i p i (x).
(1.11)
Z
D
(1.12)
Z
D
() p (x) d.
(1.13)
In the rest of the book we can consider the simplified case outside of more serious cases of
cross-dependence and derive a metaprobality adjusted distribution, as a measure or a density
on its own. Thus:
Z
p(x)
() p (x) d
D
29
Figure 1.14: Metaprobability: we add another dimension to the probability distributions, as we consider
the effect of a layer of uncertainty over the probabilities. It results in large effects in the tails, but, visually,
these are identified through changes in the "peak" at the center of the distribution.
Figure 1.15: Fragility to Error (and Stressors): Can be seen in the slope of the sensitivity of payoff across
metadistributions
(x) = lim
There are many other smooth, analytic approximations to the step function. Among the
possibilities are:
H(x) = lim
30
1 1
+ arctan(kx)
2
State
20
40
60
80
100
Time
Absorbing
Barrier
Figure 1.16: A variety of temporal states for a process subjected to an absorbing barrier. Once the absorbing
barrier is hit, the process terminates, regardless of its future potential.
As to sign of x
1
sgn(x) =
Z
sin(ux)
du,
which is useful for derivatives under the integral. Further, for more complex payoffs, in which
the decision (payoff kernel) is the random variable, we can use the convolution theorem to prove
that the Fourier transform of the product of two functions f (x) and g(x) is given as:
F [ f (x)g(x)] =
Z +
F( 0 )G( 0 ) d 0 ,
the convolution of F( 0 )G( 0 ), where F( 0 ) and G( 0 ) are the Fourier transforms of f (x) and
g(x) respectively.
These tools will be useful for working with payoffs analytically.
1.9
31
1.10
numeraire definition
A critical problem with numeraire, in which the payoff is expressed, which is effectively problematic in many situations where the "barbell" (to be defined in section x) is implemented and
something truly "risk-free" needs to be found. Well. only an invariant and unchanging metric is
really risk-free.
Definition 1.20 (Numeraire related payoffs).
1.11
1.12
ruin problems
1.13
1.14
32
Chapter Summary 2: Outline of the book and project of the codification of Risk and
decision theory as related to the real world (that is "no BS") in nonmathematical language
(other chapters are mathematical). Introduces the main fallacies treated in the project.
What can and should be mathematized. Presents the central principles of risk bearing.
Introduces the idea of fragility as a response to volatility, the associated notion of convex
heuristic, the problem of invisibility of the probability distribution and the spirit of the
book. Explains why risk is in the tails not in the variations. Explains that the layering of
random variables makes more ecological a view that is corresponds tot the "real world"
and how layering of model errors generates fat tails.
This chapter outlines the main ideas of the book; it can be read on its own as a summary of
the project.
We start with via negativa, the definition of a negative, to show why fallacies matter (particularly that risk analysis and management are negative activities):
Definition 2.1 (Via Negativa).
Consists in focusing on decision making by substraction, via the identification of errors. In theology and
philosophy, it is the focus on what something is not, an indirect definition. In action, it is a recipe for what
to avoid, what not to do subtraction, not addition, say, in medicine.
Clearly, risk management is a via negativa endeavor, avoiding a certain class of adverse events.
Fallacy
Description
Section(s)
Triffat Fallacy
Chapters 3, 6
33
Fallacy
Description
Counter of Triffat
Fallacy
Section(s)
Belief that loss function and required sample size in estimator for x is the same for
f (x) when f is convex.
Section 3.11
LLN Fallacy
Chapter 6
Binary/Vanilla
Conflation
Crossing the Street
Fallacy
CLT Error
Fallacy of Silent Evidence
Froot
Insurance
fallacy/Pisano
biotech
fallacy
(Harvard
professors)
34
2.1
Fallacy
Description
Pinker Fallacy, 1
(another Harvard
professor1 )
Pinker Fallacy, 2
Ignoring the effect of maximum divergence (Lvy, Kolmogorov) in disconfirmatory empiricism. (Counterfallacy is "n
large" for confirmatory empiricism)
Section(s)
This author is currently teaching a course with the absurd title "risk management and decisionmaking in the real world", a title he has selected himself; this is a total absurdity since risk
management and decision making should never have to justify being about the real world, and
what s worse, one should never be apologetic about it.
In "real" disciplines, titles like "Safety in the Real World", "Biology and Medicine in the Real
World" would be lunacies. But in social science all is possible as there is no exit from the gene
pool for blunders, nothing to check the system, no skin in the game for researchers. You cannot
blame the pilot of the plane or the brain surgeon for being "too practical", not philosophical
enough; those who have done so have exited the gene pool. The same applies to decision
making under uncertainty and incomplete information. The other absurdity in is the common
separation of risk and decision making, since risk taking requires reliability, hence our guiding
principle in the next section.
Indeed something is completely broken in risk management.
And the real world is about incompleteness : incompleteness of understanding, representation,
information, etc., what one does when one does not know what s going on, or when there is
a non - zero chance of not knowing what s going on. It is based on focus on the unknown,
not the production of mathematical certainties based on weak assumptions; rather measure
the robustness of the exposure to the unknown, which can be done mathematically through
metamodel (a model that examines the effectiveness and reliability of the model by examining
1 Harvard University, because of the pressure to maintain a high status for a researcher in the academic community, which
conflicts with genuine research, provides a gold mine for those of us searching for example of fooled by randomness effects.
35
"Evidence" without
rigor
Science/Evidentiary (A)
Practice
Real Life
(B)
Math/Logic (C)
Fooled by
Randomness
Math without
subtance
Science (A)
Practice (B)
(B
36
Formalize/
Expand
Intersection.
B (B C)
not reverse
Math (C)
C)
(A B
C)
Figure 2.1: Wrong! The Symmetric Difference of the Three Sets The unhappy
merger of theory and practice. Most academics and practitioners of risk and probability do not understand what "intersection" means. This explains why Wall Street
"quants" blow up. It is hard trying to explain that yes, it is very mathematical but
bringing what we call a math genius or acrobat wont do. It is jointly mathematical
and practical.
"Math/Logic" includes probability theory,
logic, philosophy.
"Practice" includes ancestral heuristics, inherited tricks and is largely convex, precautionary and via negativa .
2.2
37
Typical
Methods
Times
series
statistics, etc.
Expression
Variance, Value at
Risk
Use of estimated
probability distribution Forecasting
models
Variance, Value at
Risk, Tail exposure,
(Shortfall)
Dependence on
Characteristic both past sample
and parameters
Dependence on parameters
Erratic,
Unreliable for tails
Erratic, Unreliable
for tails
Performance
38
39
Science/Evidentiary (A)
Practice
Real Life
(B)
40
Math/Logic (C)
PrHxL
More data shows
nondegeneracy
Apparently
degenerate case
Additional Variation
x
1
x
10
20
30
40
Figure 2.5: The Masquerade Problem (or Central Asymmetry in Inference). To the left, a degenerate
random variable taking seemingly constant values, with a histogram producing a Dirac stick. One cannot
rule out nondegeneracy. But the right plot exhibits more than one realization. Here one can rule out
degeneracy. This central asymmetry can be generalized and put some rigor into statements like "failure to
reject" as the notion of what is rejected needs to be refined. We produce rules in Chapter 4.
41
dist 1
dist 2
dist 3
dist 4
dist 5
dist 6
dist 7
dist 8
dist 9
dist 10
dist 11
dist 12
dist 13
dist 14
Distributions
that cannot be
ruled out
Distributions
ruled out
Generating
Distributions
Observed
Distribution
Observable
"True"
distribution
THE VEIL
Nonobservable
Figure 2.6: "The probabilistic veil". Taleb and Pilpel (2000,2004) cover the point from an epistemological standpoint with the "veil" thought experiment by which an observer is supplied with data
(generated by someone with "perfect statistical information", that is, producing it from a generator of
time series). The observer, not knowing the generating process, and basing his information on data
and data only, would have to come up with an estimate of the statistical properties (probabilities, mean,
variance, value-at-risk, etc.). Clearly, the observer having incomplete information about the generator,
and no reliable theory about what the data corresponds to, will always make mistakes, but these
mistakes have a certain pattern. This is the central problem of risk management.
The example in Figure 2.6 provides an idea of the methodolody, and Chapter 4 produces
a formal "hierarchy" of statements that can be made by such an observer without violating a
certain inferential rigor. For instance he can "reject" that the data is Gaussian, but not accept it
as easily. And he can produce inequalities or "lower bound estimates" on, say, variance, never
"estimates" in the standard sense since he has no idea about the generator and standard estimates
require some associated statement about the generator.
Definition 2.5.
(Arbitrage of Probability Measure). A probability measure A can be arbitraged if one can produce data
fitting another probability measure B and systematically fool the observer that it is A based on his
metrics in assessing the validity of the measure.
42
0.8
0.6
Shortfall
Figure 2.7: The "true" distribution as expected from the Monte Carlo generator
0.4
0.2
x
-6
-5
-4
-3
-2
-1
250
200
150
100
50
-5
-4
-3
-2
-1
p(x) =
1/ 1
1
k1/ ( x) 1 kx
+1
+x 0
(2.1)
otherwise
By generating a Monte Carlo sample of size N with parameters = 3/2, = 1, k = 2, and = 3/4
and sending it to a friendly researcher to ask him to derive the properties, we can easily gauge
j N M
what can "fool" him. We generate M runs of N-sequence random variates ((xi )i=1
) j=1
The expected "true" mean is:
( k (+1)()
E(x) =
()
Indeterminate
>
otherwise
43
Figure 2.9: The Recovered Standard Deviation, which we insist, is infinite. This
means that every run j would deliver a different average
0.4
0.3
0.2
0.1
STD
5
10
15
V(x) =
k2 (()(2+1)(2)(+1)2 ()2 )
> 2
Indeterminate
otherwise
()2
(2.2)
which in our case is "infinite". Now a friendly researcher is likely to mistake the mean, since
about 60%
of the measurements will produce a higher value than the true mean, and, most
certainly likely to mistake the variance (it is infinite and any finite number is a mistake).
Further, about 73% of observations fall above the true mean. The CDF= 1
R
where is the Euler Gamma function (z) = 0 et tz1 dt.
RK
RK
x p(x) dx
p(x) dx
(+1)()
()
1
+1
the "tail risk" below 1% and 99% for more severe risks. This exercise was a standard one but
there are many more complicated distributions than the ones we played with.
44
Rules
Dutch Book
R1
R1
Inequalities
R2
Asymmetry
R3
R4
Nonlinear
Response
Conditional
Precautionary Principle
R5
Decisions
Description
Probabilities need to add up to 1* but cannot exceed 1
It is more rigorous to work with probability inequalities and bounds than probabilistic estimates.
Some errors have consequences that are largely, and
clearly one sided.**
Fragility is more measurable than probability***
Domain specific precautionary, based on fat tailedness of errors and asymmetry of payoff.
Exposures ( f (x))can be more reliably modified, instead of relying on computing probabilities of x.
* The Dutch book can be expressed, using the spirit of quantitative finance, as a no arbitrage situation, that
is, no linear combination of payoffs can deliver a negative probability or one that exceeds 1. This and the
corrollary that there is a non-zero probability of visible and known states spanned by the probability distribution adding up to <1 confers to probability theory, when used properly, a certain analytical robustness.
** Consider a plane ride. Disturbances are more likely to delay (or worsen) the flight than accelerate it or
improve it. This is the concave case. The opposite is innovation and tinkering, the convex case.
*** The errors in measuring nonlinearity of responses are more robust and smaller than those in measuring
responses. (Transfer theorems).
45
The Supreme Scientific Rigor of The Russian School of Probability One can believe
in the rigor of mathematical statements about probability without falling into the trap of
providing naive computations subjected to model error. If this author were to belong to a
school of thought designated by a nationality, the
{Nationality} school of {discipline},
it would be the Russian school of probability.
Members across three generations: P.L. Chebyshev, A.A. Markov, A.M. Lyapunov, S.N. Bernshtein (ie. Bernstein), E.E. Slutskii, N.V. Smirnov, L.N. Bolshev, V.I. Romanovskii, A.N. Kolmogorov, Yu.V. Linnik, and the new generation: V. Petrov, A.N. Nagaev, A. Shyrayev, and a
few more.
They had something rather potent in the history of scientific thought: they thought in
inequalities, not equalities (most famous: Markov, Chebyshev, Bernstein, Lyapunov). They
used bounds, not estimates. Even their central limit version was a matter of bounds, which
we exploit later by seeing what takes place outside the bounds. They were world apart from
the new generation of users who think in terms of precise probability or worse, mechanistic
social scientists. Their method accommodates skepticism, one-sided thinking: "A is > x,
AO(x) [Big-O: "of order" x], rather than A = x.
For those working on integrating the mathematical rigor in risk bearing they provide a great
source. We always know one-side, not the other. We know the lowest value we are willing
to pay for insurance, not necessarily the upper bound (or vice versa).a
a The way this connects to robustness, which we will formalize next section, is as follows. Is robust what does not
change across perturbation of parameters of the probability distribution; this is the core of the idea in Part II with
our focus on fragility and antifragility. The point is refined with concave or convex to such perturbations.
2.3
Let us start with a sketch of the general solution to the problem of risk and probability, just to
show that there is a solution (it will take an entire book to get there). The following section will
outline both the problem and the methodology.
46
47
2.4
Next we give the reader a hint of the methodology and proposed approach with a semi-informal
technical definition for now.
In his own discussion of the Borel-Cantelli lemma (the version popularly known as "monkeys
on a typewriter")[13], Emile Borel explained that some events can be considered mathematically
possible, but practically impossible. There exists a class of statements that are mathematically
rigorous but practically nonsense, and vice versa.
If, in addition, one shifts from "truth space" to consequence space", in other words focus on
(a function of) the payoff of events in addition to probability, rather than just their probability,
then the ranking becomes even more acute and stark, shifting, as we will see, the discussion
from probability to the richer one of fragility. In this book we will include costs of events as part
of fragility, expressed as fragility under parameter perturbation. Chapter 5 discusses robustness
under perturbation or metamodels (or metaprobability). But here is the preview of the idea of
convex heuristic, which in plain English, is at least robust to model uncertainty.
Definition 2.6 (Convex Heuristic).
In short what exposure is required to not produce concave responses under parameter perturbation.
Summary of a Convex Heuristic (from Chapter 16) Let { f i } be the family of possible
functions, as "exposures" to x a random variable with probability measure (x), where
is a parameter determining the scale (say, mean absolute deviation) on the left side of
the distribution (below the mean). A decision rule is said "nonconcave" for payoff below K
with respect to up to perturbation if, taking the partial expected payoff
EK ( f i ) =
Z K
f i (x) d (x),
One brilliant contribution by economists is the concept of "cheap talk", or the difference between
"stated preferences" (what you say) and "revealed preferences" (those that can be inferred from
actions). Actions are louder than words: what people say (in opinion polls or elsewhere) isnt as
relevant, as individuals reveal their preferences with hard cash or, more generally costly action,
or even more generally risky action (which, invariably, brings us to skin in the game). This is
why opinion polls are considered largely entertainment. Further, the notion of "belief" is largely
misunderstood.
48
Those who engage in actions that threaten their survival will eventually disappear, if their
skin is in their game. Same with populations with the wrong heuristics.
Belief is deeply connected to probability (belief in the epistemic sense). Which means that
violations of probability axioms and bounds can imply irrationality.
We showed here that the notion of "probability" raw is largely verbalistic and empty (prob
ability maps to "degrees of belief" mathematically, is belief),
largely incomplete, more "binary"
while revealed preferences via decisions is what matters (more technically probability is something deeply mathematical, useless on its own, an integral transform into something larger, and
cannot be "summarized" in words). And decisions and decisions only can be a metric for "rationality"
Psychologists and journalistic types who make a living attacking "overreactions" and bubbles
based on superficial assessments typically discuss "rationality" without getting what rationality
means in its the decision-theoretic sense (the only definition that can be consistent, in terms of
absence of violations of the standard axioms [CITE AXIOMS] and only from actions). But as
we saw with convex heuristics the cause behind an action leading to survival is not necessarily
apparent. Many seemingly irrational actions have led populations to survive. Dread risk and
overreactions arent just rational, but may be the only way to survive in some domains. [Cite
Taleb and Read]
As an interesting application, one can even show that it is rational to "believe" in the supernatural if it leads to an increase in survival as a side effect. 2
2 Many authors claim to be arbiters of "rationality" and, as we can see in Chapter on meta-distribution and the debunking
of "pseudo-biases", accuse others of irrationality, but cannot come up with a coherent definition of rationality (unless
model dependent, which means that a breakdown of a model or misspecification can justify actions otherwise deemed
"irrational"); we can however certainly define irrationality in terms of violations of a certain set of axioms, so our
definition is via negativa.
49
2.5
Crucially, we can gauge the nonlinear response to a parameter of a model using the same method
and map "fragility to model error". For instance a small perturbation in the parameters entering the probability provides a one-sided increase of the likelihood of event (a convex response),
then we can declare the model as unsafe (as with the assessments of Fukushima or the conventional Value-at-Risk models where small parameters variance more probabilities by 3 orders of
magnitude). This method is fundamentally option-theoretic.
2.5.1
Why Engineering?
[Discussion of the problem- A personal record of the difference between measurement and working on reliability. The various debates.]
Risk is not Variations
On the common confustion between risk and variations. Risk is tail events, necessarily.
What Do Fat Tails Have to Do With This?
The focus is squarely on "fat tails", since risks and harm lie principally in the high-impact events,
The Black Swan and some statistical methods fail us there. But they do so predictably. We end
Part I with an identification of classes of exposures to these risks, the Fourth Quadrant idea, the
class of decisions that do not lend themselves to modelization and need to be avoided in other
words where x is so reliable that one needs an f (x) that clips the left tail, hence allows for a
computation of the potential shortfall. Again, to repat, it is more, much more rigorous to modify
your decisions.
Fat Tails and Model Expansion
Next wee see how model uncertainty (or, within models, parameter uncertainty), or more generally, adding layers of randomness, cause fat tails.
Part I of this volume presents a mathematical approach for dealing with errors in conventional
probability models For instance, if a "rigorously" derived model (say Markowitz mean variance,
or Extreme Value Theory) gives a precise risk measure, but ignores the central fact that the
parameters of the model don t fall from the sky, but have some error rate in their estimation,
then the model is not rigorous for risk management, decision making in the real world, or, for
that matter, for anything.
So we may need to add another layer of uncertainty, which invalidates some models but not
others. The mathematical rigor is therefore shifted from focus on asymptotic (but rather irrelevant because inapplicable) properties to making do with a certain set of incompleteness and
50
Figure 2.11: A Version of Savages Small World/Large World Problem. In statistical domains assume Small
World= coin tosses and Large World = Real World. Note that measure theory is not the small world, but
large world, thanks to the degrees of freedom it confers.
51
2.6
The Black Swan Problem Incomputability of Small Probalility: It is is not merely that events
in the tails of the distributions matter, happen, play a large role, etc. The point is that these
events play the major role for some classes of random variables and their probabilities are not
computable, not reliable for any effective use. And the smaller the probability, the larger the
error, affecting events of high impact. The idea is to work with measures that are less sensitive
to the issue (a statistical approch), or conceive exposures less affected by it (a decision theoric
approach). Mathematically, the problem arises from the use of degenerate metaprobability.
52
Description
Chapters
Preasymptotics,
Incomplete
Convergence
Inverse Problems
Degenerate
Metaprobability
53
2.7
We close the introduction with De Finettis introduction to his course "On Probability":
The course, with a deliberately generic title will deal with the conceptual and controversial
questions on the subject of probability: questions which it is necessary to resolve, one way or
another, so that the development of reasoning is not reduced to a mere formalistic game of mathematical expressions or to vacuous and simplistic pseudophilosophical statements or allegedly
practical claims. (emph. mine.)
The next appendix deplores academic treatment of probability so we get it out of the way.
54
W H AT S A C H A R L ATA N I N R I S K A N D
PROBABILITY?
We start with a clean definition of charlatan. Many of us spend time fighting with charlatans;
we need a cursory and useable definition that is both compatible with our probability business
and the historical understanding of the snake-oil salesman.
a.1
charlatan
55
Subdiscipline of Bullshittology I
am being polite here. I truly believe
that a scary share of current discussions of risk management and probability by nonrisktakers fall into the
category called obscurantist, partaking of the "bullshitology" discussed
in Elster: "There is a less polite word
for obscurantism: bullshit. Within
Anglo-American philosophy there is
in fact a minor sub-discipline that
one might call bullshittology." [29].
The problem is that, because of
nonlinearities with risk, minor bullshit can lead to catastrophic consequences, just imagine a bullshitter piloting a plane. My angle is that the
bullshit-cleaner in the risk domain is
skin-in-the-game, which eliminates
those with poor understanding of
risk.
56
a.1 charlatan
Science (A)
Math (C)
C A' B'
57
2 It has been fashionable to invoke the vague idea of mathematical "charlatanry" in the history of economics, first with
Alfred Marshall famous procedure "(1) Use mathematics as shorthand language, rather than as an engine of inquiry. (2)
Keep to them till you have done. (3) Translate into English. (4) Then illustrate by examples that are important in real
life (5) Burn the mathematics. (6) If you cant succeed in 4, burn 3. This I do often.". Similarly, J.M. Keynes: "(...)we
will endeavour to discredit the mathematical charlatanry by which, for a hundred years past, the basis of theoretical
statistics have been greatly undermined", in A Treatise On Probability [64]. As one can see, these types of warnings proved
ineffectual owing to citation rings. So our tack is different, largely constrained by the idea of skin in the game that would
bring things to the missing link of reality.
58
Part II
FAT TA I L S : T H E L L N U N D E R R E A L W O R L D E C O LO G I E S
FAT TA I L S A N D T H E P R O B L E M O F
INDUCTION
3.1
Turkey and Inverse Turkey (from the Glossary in Antifragile): The turkey is fed by the butcher
for a thousand days, and every day the turkey pronounces with increased statistical confidence
that the butcher "will never hurt it"until Thanksgiving, which brings a Black Swan revision of
belief for the turkey. Indeed not a good day to be a turkey. The inverse turkey error is the mirror
confusion, not seeing opportunities pronouncing that one has evidence that someone digging
for gold or searching for cures will "never find" anything because he didnt find anything in the
past.
What we have just formulated is the philosophical problem of induction (more precisely of
enumerative induction.) To this version of Bertrand Russels chicken we add: mathematical
difficulties, fat tails, and sucker problems.
3.2
Let us define an empirical risk estimator that we will work with throughout the book. We start
with a partial first moment.
Definition 3.1.
(Estimator) Let X be, as of time T, a standard sequence of n+1 observations, X = xt0 +it 0in (with xt
R, i N), as the discretely monitored history of a stochastic process Xt over the closed interval [t0 , T ]
(with realizations at fixed interval t thus T = t0 + nt). 1
The empirical estimator MTX (A, f ) is defined as
MTX (A,
n 1 A f xt0 +it
f ) i=0 n
i=0 1D 0
(3.1)
1 It is not necessary that t follows strictly calendar time for high frequency observations, as calendar time does not necessarily correspond to transaction time or economic time, so by a procedure used in option trading called "transactional
time" or "economic time", the observation frequency might need to be rescaled in a certain fashion to increase sampling
at some windows over others a procedure not dissimilar to seasonal adjustment, though more rigorous mathematically. What matters is that, if there is scaling of t, the scaling function needs to be fixed and deterministic. But this
problem is mostly present in high frequency. The author thanks Robert Frey for the discussion.
61
150
out of
sample
In sample
100
50
T
t
Figure 3.1: A rolling window: to estimate the errors of an estimator,it is not rigorous to compute in-sample
properties of estimators, but compare properties obtained at T with prediction in a window outside of it.
Maximum likelihood estimators should have their variance (or other more real-world metric of dispersion)
estimated outside the window.
Let us stay in dimension 1 for now not to muddle things. Standard Estimators tend to be
variations about MtX (A, f ) where f(x) =x and A is defined as the domain of the distribution of
X, standard measures from x, such as moments of order z, etc., are calculated "as of period" T.
Such measures might be useful for the knowledge of some properties, but remain insufficient for
decision making as the decision-maker may be concerned for risk management purposes with
the left tail (for distributions that are not entirely skewed, such as purely loss functions such as
damage from earthquakes, terrorism, etc.), or any arbitrarily defined part of the distribution.
62
n
x 1A
i=0
n
1D 0
i=0
(3.2)
n
x1A
i=0
n
1A
i=0
(3.3)
One of the uses of the indicator function 1 A , for observations falling into a subsection* A of
the distribution, is that we can actually derive the past actuarial value of an option with X as an
underlying struck as K as MTX (A, x), with A = (, K] for a put and A = [K, ) for a call, with
f (x) = x K or K x.
Criterion 3.1.
The measure M is considered to be an estimator over interval [ t- N t, T] if and only if it holds in
expectation over a specific period XT+it for a given i>0, that is across counterfactuals of the process,
with a threshold e (a tolerated relative absolute divergence; removing the absolute sign reveals the bias) so
X ( A , f )
E MTX ( Az , f ) M>
z
T
X
<e
(3.4)
(MT ( Az , f )) =
M X ( Az , f )
T
when MTX ( Az , f ) is computed; but while working with the opposite problem, that is, trying to
guess the spread in the realizations of a stochastic process, when the process is known, but not
X ( A , 1) as a divisor.
the realizations, we will use M>
z
T
In other words, the estimator as of some future time, should have some stability around the
"true" value of the variable and stay below an upper bound on the tolerated bias.
We use the loss function (.) = |.| measuring mean absolute deviations to accommodate
functions and exposures and that do not have finite second moment, even if the process
has such moments. Another reason is that in the real world gains and losses are in straight
numerical deviations.a
a Using absolute deviations would sound more robust than squared deviations, particularly for fat-tailed domains;
it seems that the resistance comes, among other things, from the absence of derivability at 0.
So we skip the notion of "variance" for an estimator and rely on absolute mean deviation so
can be the absolute value for the tolerated bias. And note that we use mean deviation as
the equivalent of a "loss function"; except that with matters related to risk, the loss function is
embedded in the subset A of the estimator.
This criterion makes our risk estimator compatible with standard sampling theory. Actually,
it is at the core of statistics. Let us rephrase:
Standard statistical theory doesnt allow claims on estimators made in a given set unless these
are made on the basis that they can "generalize", that is, reproduce out of sample, into the part
of the series that has not taken place (or not seen), i.e., for time series, for >t.
63
3.3
Some harmless formalism: L p space. Lets look at payoff in functional space, to work with
the space of functions having a certain integrability. Let Y be a measurable space with
Lebesgue measure . The space L p of f measurable functions on Y is defined as:
p
L () =
f :
Z
| f | d
1/ p
<
with p 1. The application of concern for our analysis in this section is where the measure
is a counting measure (on a countable set).
Fat tails are not about the incidence of low probability events, but the contributions of events
away from the "center" of the distribution to the total properties.2 As a useful heuristic, consider
the ratio h
p
E (X2 )
h=
E(| X |)
where E is the expectation operator (under the probability measure of concern and x is a centered
variable such E(x) = 0); the ratio increases with the fat tailedness of the distribution; (The general
1
( MX ( A,xn )) n
case corresponds to MTX (A,| x|) , n > 1, under the condition that the distribution has finite
T
64
f(x)
x2
x
x
Figure 3.2: The difference between the two weighting functions increases for large values of x.
The effect is due to the convexity differential between both functions, |x| is piecewise linear
and loses the convexity effect except for a zone around the origin.3
Proof: By Jensens inequality under the counting measure.
As a convention here, we write L p for space, L p for the norm in that space.
n
, The L p Norm is defined (for our purpose) as, with p N , p 1):
Let X ( xi )i=1
kXk p
n
| xi | p
i=1
n
1/ p
The idea of dividing by n is to transform the norms into expectations,i.e., moments. For the
Euclidian norm, p = 2.
The norm rises with higher values of p, as, with a > 0.4 ,
1
n
|xi | p+a
i=1
!1/(p+a)
>
1
n
| xi | p
!1/ p
i=1
What is critical for our exercise and the study of the effects of fat tails is that, for a given
norm, dispersion of results increases values. For example, take a flat distribution, X= {1, 1}.
3 TK Adding an appendix "Quick and Robust Estimates of Fatness of Tails When Higher Moments Dont Exist" showing
how the ratios STD/MAD (finite second moment) and MAD(MAD)/STD (finite first moment) provide robust estimates
and outperform the Hill estimator for symmetric power laws.
4 An application of Hlders inequality,
1/ p
1 1
1
n
n
.
| xi | p+a a+p n a+p p i=1
| xi | p
i=1
65
1 3
2, 2
produces rising
Trying again, with a wider spread, we get even higher values of the norms, X =
q
4
5
5 3 43
1201
2101
2
,
,
.
{|| X ||n }5n=1 = 1, ,
4 2
4
2 23/5
(3.5)
n
1 7
4, 4
(3.6)
k X k = sup
1
|x |
n i
n
(3.7)
i=1
Gaussian Case For a Gaussian, where x N(0, ), as we assume the mean is 0 without loss of
generality,
1/ N
A, X N
MTX
MTX (A, | X |)
N 1
2N
1
N
N
2 2 1 (1) N + 1 N+1
2
or, alternatively
1 12 N2 N + 1
1 (N 3)
N
2
1
+
(
1)
=
2
(3.8)
2
2
MTX (A, | X |)
R
where (z) is the Euler gamma function; (z) = 0 tz1 et dt. For odd moments, the ratio is 0.
For even moments:
r
MTX A, X 2
X
2
MT ( A, | X |)
MTX A, X N
hence
q
MTX ( A, X 2 )
Standard Deviation
=
=
Mean Absolute Deviation
MTX ( A, | X |)
For a Gaussian the ratio 1.25, and it rises from there with fat tails.
Example: Take an extremely fat tailed distribution with n=106 , observations are all -1 except
for a single one of 106 ,
n
o
X = 1, 1, ..., 1, 106 .
The mean absolute deviation, MAD (X) = 2. The standard deviation STD (X)=1000. The ratio
standard deviation over mean deviation
qis 500.
As to the fourth moment, it equals 3
3
2
For a power law distribution with tail exponent =3, say a Student T
66
Figure 3.3: The Ratio Standard Deviation/Mean Deviation for the daily returns
of the SP500 over the past 47 years, with a
monthly window.
1.5
1.4
1.3
1.2
1.1
Time
MTX ( A, X 2 )
MTX
( A, | X |)
Standard Deviation
=
Mean Absolute Deviation
2
We will return to other metrics and definitions of fat tails with power law distributions when
the moments are said to be "infinite", that is, do not exist. Our heuristic of using the ratio of
moments to mean deviation works only in sample, not outside.
"Infinite" moments Infinite moments, say infinite variance, always manifest themselves as computable numbers in observed sample, yielding an estimator M, simply because the sample is
finite. A distribution, say, Cauchy, with infinite means will always deliver a measurable mean
in finite samples; but different samples will deliver completely different means. Figures 3.4 and
3.5 illustrate the "drifting" effect of M a with increasing information.
67
MTX HA, xL
4
3
2
1
T
2000
4000
6000
8000
10 000
-1
-2
MTX IA, x2 M
4.0
Figure 3.5: The standard deviation of a series with infinite variance (St(2)).
3.5
3.0
T
2000
68
4000
6000
8000
10 000
3.4
Since higher moments increase under fat tails, as compared to lower ones, it should be possible
so simply increase fat tails without increasing lower moments.
Note that the literature sometimes separates "Fat tails" from "Heavy tails", the first term
being reserved for power laws, the second to subexponential distribution (on which, later).
Fughtetaboutdit. We simply call "Fat Tails" something with a higher kurtosis than the Gaussian, even when kurtosis is not defined. The definition is functional as used by practioners of
fat tails, that is, option traders and lends itself to the operation of "fattening the tails", as we will
see in this section.
A Variance-preserving heuristic. Keep E X 2 constant and increase E X 4 , by "stochasticizing" the variance of the distribution, since < X 4 > is itself analog to the varianceof < X 2 >
2
measured across samples ( E X 4 is the noncentral equivalent of E X 2 E X 2
). Chapter x will do the "stochasticizing" in a more involved way.
An effective heuristic to get some intuition about the effect of the fattening of tails consists
in simulating a random variable set to be at mean 0, but with the following
variance-preserving
1
a
with probability p
tail fattening
trick:
the
random
variable
follows
a
distribution
N
0,
1
1
= 2 and N 0, 1 + a with the remaining probability 2 , with 0 6 a < 1.
The characteristic function is
(t, a) =
2 2
1 1 (1+a)t2 2
e 2
1 + e at
2
3.5
We assume tails start at the level of convexity of the segment of the probability distribution to
the scale of the distribution.
69
Peak
(a2 , a3 L
0.5
Shoulders
Ha1 , a2 L,
Ha3 , a4 L
0.4
0.3
a
Right tail
Left tail
0.2
0.1
a1
-4
-2
a2
a3
a4
2
Figure 3.6: Fatter and Fatter Tails through perturbation of . The mixed distribution with values for the
stochastic volatility coefficient a: {0, 14 , 21 , 34 }. We can see crossovers a1 through a4 . The "tails" proper start
at a4 on the right and a1 on the left.
Let X be a random variable, the distribution of which p(x) is from a general class of all
unimodal one-parameter continous pdfs p with support D R and scale parameter . Let
p(.) be quasi-concave on the domain, but neither convex nor concave. The density function p(x)
satisfies: p(x) p(x + e) for all e > 0, and x > x and p(x) p(x e) for all x < x with
{ x : p(x ) = maxx p(x)}. The class of quasiconcave functions is defined as follows: for all x and
y in the domain and [0, 1],
p ( x + (1 ) y) min ( p(x), p(y))
A- If the variable is "two-tailed", that is, D = (-,), where p (x)
70
p(x,+)+p(x, )
2
e 22 24 52 (x )2 + (x )4
= 0,
27
which produces the following crossovers:
(
r
r
1
1
5 + 17 ,
5 17 ,
{ a1 , a2 , a3 , a4 } =
2
2
)
r
r
1
1
5 17 , +
5 + 17
+
2
2
(3.9)
In figure 3.6, the crossovers for the intervals are numerically {2.13, .66, .66, 2.13 }.
As to a symmetric power law(as we will see further down), the Student T Distribution with
scale s and tail exponent :
! +1
2
2
+ x2
s
p(x)
sB
{ a1 , a2 , a3 , a4 } =
5 (+1)(17+1)+1
s
1
q
,
1
2, 2
5 (+1)(17+1)+1
s
1
5+ (+1)(17+1)+1
s
1
q
,
5+ (+1)(17+1)+1 o
s
1
o
{ a1 , a2 , a3 , a4 } = 4 13s, 4 + 13s, 4 13s, 4 + 13s
We can verify that when , the crossovers become those of a Gaussian. For instance, for
a1 :
q
r
5 (+1)(17+1)+1
s
1
1
lim
=
5 17 s
2
2
71
5+ (+1)(17+1)+1
s
1
, with infinite in the stochastic volatility Gaussian case and s the stan2
dard deviation. The "tail" is located between around 2 and 3 standard deviations. This flows
from our definition: which part of the distribution is convex to errors in the estimation of
the scale.
But in practice, because historical measurements of STD will be biased lower because of
small sample effects (as we repeat fat tails accentuate small sample effects), the deviations
will be > 2-3 STDs.
B- For some one-tailed distribution that have a "bell shape" of convex-concave-convex shape,
under some conditions, the same 4 crossover points hold. The Lognormal is a special case.
{ a1 , a2 , a3 , a4 } = e
1
2
2 2 52 172
3.6
1
2
2 2
172 +52
,e
1
2
2+ 2 52 172
,e
1
2
o
172 +52
2+ 2
We can improve on the fat-tail heuristic in 3.4, (which limited the kurtosis to twice the Gaussian)
as follows. We Switch between Gaussians with variance:
(
2 (1 + a),
with probability p
2 (1 + b),
with probability 1 p
p
(t, a) = p e 2 (a+1)
with Kurtosis
3((1 a2 ) p1)
p 1
2 t2
(p 1) e
2 t2 (ap+p1)
2(p1)
thus allowing polarized states and high kurtosis, all variance preserv-
1 p
p .
Thus with p = 1/1000, and the maximum possible a = 999, kurtosis can reach as high a level
as 3000.
This heuristic approximates quite well the effect on probabilities of a lognormal weighting
for the characteristic function
(t, V) =
72
Z
e
0
2
t 2v
log(v)v0+ Vv
2
2Vv2
2vVv
2
dv
Pr
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
Figure 3.7: Stochastic Variance: Gamma distribution and Lognormal of same mean and variance.
where v is the variance and Vv is the second order variance, often called volatility of volatility.
Thanks to integration by parts we can use the Fourier transform to obtain all varieties of payoffs
(see Gatheral, 2006). But the absence of a closed-form distribution can be remedied as follows.
Gamma Variance A shortcut for a full lognormal distribution without the narrow scope of
heuristic is to use Gamma Variance. Assume that the variance of the Gaussian follows a gamma
distribution.
v
v1 V
e V
(v) =
()
2
with mean V and standard deviation V . Figure 3.7 shows the matching to a lognormal with
same first two moments
get the lognormal with mean and standard deviation, respectively,
r as we
n
3
1
V
V
log
and
log
2
V+1
V+1 . The final distribution becomes (once again, assuming, without loss, a mean of 0):
f ,V (x) =
Z
0
x2
e 2v
(v)dv
2 v
allora:
2
f ,V (x) =
3
42
V
1
1
x2
1
4
()
K 1
2
2
q V
1
x2
!
(3.10)
73
-4
-2
Why power laws? There are a lot of theories on why things should be power laws, as sort of
exceptions to the way things work probabilistically. But it seems that the opposite idea is never
presented: power should can be the norm, and the Gaussian a special case as we will see in
Chapt x, of concave-convex responses (sort of dampening of fragility and antifragility, bringing
robustness, hence thinning tails).
3.7
x
2
(3.11)
*
1
1 * * T 1 * *
m /2
1/2
fa x
= (2)
|1 |
exp
x 1
x
2
2
1
1 * * T 1 * *
m /2
1/2
+ (2)
|2 |
exp
x 2
x
2
2
74
(3.12)
2
0
-4
-2
-2
4
2
2
0
-2
-2
-4
-4
-2
-2
0
2
Figure 3.9: Multidimensional Fat Tails: For a 3 dimentional vector, thin tails (left) and fat tails (right) of
the same variance. Instead of a bell curve with higher peak (the "tunnel") we see an increased density of
points towards the center.
3.8
So far for the discussion on fat tails we stayed in the finite moments case. For a certain class of
distributions, those with finite moments, PPXX>>nK
depends on n and K. For a scale-free distribution,
K
L(t x)
=1
L(x)
5 We can simplify by assuming as we did in the single dimension case, without any loss of generality, that = (0, . . . , 0).
75
log P>x
Student (3)
0.1
10-4
10-7
Gaussian
-10
10
LogNormal-2
10-13
log x
2
10
20
Figure 3.10: Three Types of Distributions. As we hit the tails, the Student remains scalable while the
Standard Lognormal shows an intermediate position before eventually ending up getting an infinite slope
on a log-log plot.
(with
STD and mean ) , by taking the PDF this time instead of the exceedance probability
(x )2
log f (x) = 22 log( 2) 21 2 x2 which goes to faster than log(x) for x .
So far this gives us the intuition of the difference between classes of distributions. Only
scalable have "true" fat tails, as others turn into a Gaussian under summation. And the tail
exponent is asymptotic; we may never get there and what we may see is an intermediate version
of it. The figure above drew from Platonic off-the-shelf distributions; in reality processes are
vastly more messy, with switches between exponents.
Estimation issues Note that there are many methods to estimate the tail exponent from data,
what is called a "calibration. However, we will see, the tail exponent is rather hard to guess, and
its calibration marred with errors, owing to the insufficiency of data in the tails. In general, the
data will show thinner tail than it should.
We will return to the issue in Chapter 10.
3.9
We introduced the category "true fat tails" as scalable power laws to differenciate it from the
weaker one of fat tails as having higher kurtosis than a Gaussian.
Some use as a cut point infinite variance, but Chapter 3 will show it to be not useful, even
misleading. Many finance researchers (Officer, 1972) and many private communications with
76
P(X >k)
P(X >2 k)
P(X >k)
P(X >2 k)
P(X >k)
P(X >2 k)
(Gaussian)
(Gaussian)
Student(3)
Student (3)
Pareto(2)
Pareto (2)
44
720
14.4
4.9
31600.
5.1 1010
71.4
6.8
64
1.01 109
5.5 1023
216
7.4
216
1.61 1015
9 1041
491
7.6
512
10
1.31 1023
9 1065
940
7.7
1000
12
5.63 1032
fuhgetaboudit
1610
7.8
1730
14
1.28 1044
fuhgetaboudit
2530
7.8
2740
16
1.57 1057
fuhgetaboudit
3770
7.9
4100
18
1.03 1072
fuhgetaboudit
5350
7.9
5830
20
3.63 1088
fuhgetaboudit
7320
7.9
8000
finance artists reveal some kind of mental block in seeing the world polarized into finite/infinite
variance.
Another useful distinction: Let X = ( xi )1in be realizations of i.i.d. random variables in R+ ,
with cumulative distribution function F; then by the Teugels (1975)[113] and Pitman [89] (1980)
definition:
1 F2 (x)
=2
x 1 F(x)
lim
77
(R
RK
K
f 2 (x, ) dx
)
=
f (x, ) dx
=1,2
Rz
0
t a1 (1 t)b1 dt
6)
2(K + log(K 1)) 2 K(K(K+3)
+ 6 log(K 1)
K 1
,
K
K2
and, for = 5,
(R
RK
K
f 2 (x, ) dx
f (x, ) dx
)
=
=5
1
K(K(K(K(K(K(K(K(4K + 9) + 24) + 84) + 504) 5250) + 10920) 8820) + 2520) + 2520(K 1)4 log(K 1)
2(K 1)4 K5
We know that the limit is 2 for all three cases, but it is important to observe the preasymptotics.
As we can see in Figure 3.10, finite or nonfinite variance is of small importance for the effect in
the tails.
Example 2 Case of the Gaussian. Since the Gaussian belongs to the family of the stable distribution (Chapter x), the convolution will produce a Gaussian of twice the variance. So taking a
Gaussian, N (0, 1) for short
(0 mean
and unitary standard deviation), the densities of the convo
lution will be Gaussian 0, 2 , the ratio of the exceedances
R 2
erfc K2
f
(x)
dx
RK
=
erfc K
K f (x) dx
2
78
1 - F2
1-F
3.0
2.8
=1
2.6
=2
=5
2.4
2.2
2.0
20
40
60
80
100
Figure 3.11: The ratio of the exceedance probabilities of a sum of two variables over a single one: power
law
1 - F2
1-F
HGaussianL
10 000
8000
6000
4000
2000
1
K HSTDev.L
Figure 3.12: The ratio of the exceedance probabilities of a sum of two variables over a single one: Gaussian
79
3.0
2.5
2.0
K
200
400
600
800
1000
Application: Two Real World Situations We are randomly selecting two people, and the sum
of their heights is 4.1 meters. What is the most likely combination? We are randomly selecting
two people, and the sum of their assets, the total wealth is $30 million. What is the most likely
breakdown?
Assume two variables X1 and X2 following an identical distribution, where f is the density
function,
P [X1 + X2 = s] = f 2 (s)
=
f (y) f (s y) dy.
s
e 2 ns 2
+ f
=
2
2
2
s
where x
For a Power law, say a Pareto distribution with tail exponent, f (x)= x 1 xmin
min
s
s
is minimum value , 2 xmin , and 2 xmin
s
s
s
s 1
2
f +
f
= 2 xmin
+
2
2
2
2
80
More generally, distributions are called subexponential when the exceedance probability declines more slowly in the tails than the exponential.
For a one-tailed random variable7 ,
a) limx
PX >x
PX > x
b) limx
PX >x
P(X >max(x))
The sum is of the same order as the maximum (positive) value, another way of saying that the
tails play a large role.
Clearly F has to have no exponential moment:
Z
0
eex dF(x) =
6 Technical comment: we illustrate some of the problems with continuous probability as follows. The sets 4.1 and 30 106
have Lebesgue measures 0, so we work with densities and comparing densities implies Borel subsets of the space, that
is, intervals (open or closed) a point. When we say "net worth is approximately 30 million", the lack of precision in
the statement is offset by an equivalent one for the combinations of summands.
7 for two-tailed variables, the result should be the same by splitting the observations in two groups around a center. BUT
I NEED TO CHECK IF TRUE
81
x2
0.00020
2
x2
x2
0.00015
2
x2
2 x2
0.00010
2
x2
3 x-
0.00005
2
10
15
20
Figure 3.14: Multiplying the standard Gaussian density by emx , for m = {0, 1, 2, 3}.
35
1
- log2 HxL
2
30
2 x
log2 HxL
25
x2
2 x
20
log2 HxL
2 x2
15
2 x
log2 HxL
10
3 x-
2 x
1.2
1.4
1.6
1.8
2.0
Figure 3.15: Multiplying the Lognormal (0,1) density by emx , for m = {0, 1, 2, 3}.
The standard Lognormal belongs to the subexponential category, but just barely so (we used
in the graph above Log Normal-2 as a designator for a distribution with the tail exceedance
82
Max
10 000
8000
6000
4000
2000
0
Figure 3.16: A time series of an extremely fat-tailed distribution (one-tailed). Given a long enough series,
the contribution from the largest observation should represent the entire sum, dwarfing the rest.
3.10
There is another aspect, beyond our earlier definition(s) of fat-tailedness, once we increase the
dimensionality into random vectors:
Stochastic Parameters The problem of elliptical distributions is that they do not map the return
of securities, owing to the absence of a single variance at any point in time, see Bouchaud and
Chicheportiche (2010) [18]. When the scales of the distributions of the individuals move but not
in tandem, the distribution ceases to be elliptical.
Figure 3.17 shows the effect of applying the equivalent of stochastic volatility methods: the
more annoying stochastic correlation. Instead of perturbating the correlation matrix as a unit
as in section 3.7, we perturbate the correlations with surprising effect.
83
84
Figure 3.19: Elliptically Contoured Joint Returns for for a multivariate distribution (x, y, z) solving to the
same density.
Figure 3.20: NonElliptical Joint Returns, from stochastic correlations, for a multivariate distribution (x, y, z)
solving to the same density.
85
3.11
There are broadly two separate ways to go about estimators: nonparametric and parametric.
The nonparametric approach It is based on observed raw frequencies derived from sample-size
n. Roughly, it sets a subset of events A and MTX (A, 1) (i.e., f(x) =1), so we are dealing with the
n
1 A . Thus these estimates dont allow discussions on frequencies < n1 ,
frequencies (A) = n1 i=0
at least not directly. Further the volatility of the estimator increases with lower frequencies. The
error is a function of the frequency itself (or rather, the smaller of the frequency and 1-). So
n
if i=0
1 A =30 and n = 1000, only 3 out of 100 observations are expected to fall into the subset A,
restricting the claims to too narrow a set of observations for us to be able to make a claim, even
if the total sample n = 1000 is deemed satisfactory for other purposes. Some people introduce
smoothing kernels between the various buckets corresponding to the various frequencies, but
in essence the technique remains frequency-based. So if we nest subsets, A1 A2 A, the
expected "volatility" (as we will see later in the chapter, we mean MAD, mean absolute deviation,
not STD) of MTX ( Az , f ) will produce the following inequality:
X
X
E MTX A<z , f M>
A<z , f
E MTX Az , f M>
Az , f
T
T
M X Az , f
M X A<z , f
T
T
for all functions f (Proof via twinking of law of large numbers for sum of random variables).
The parametric approach it allows extrapolation but emprisons the representation into a specific off-the-shelf probability distribution (which can itself be composed of more sub-probability
distributions); so MTX is an estimated parameter for use input into a distribution or model and
the freedom left resides in differents values of the parameters.
Both methods make is difficult to deal with small frequencies. The nonparametric for obvious
reasons of sample insufficiency in the tails, the parametric because small probabilities are very
sensitive to parameter errors.
The Sampling Error for Convex Payoffs
This is the central problem of model error seen in consequences not in probability. The literature
is used to discussing errors on probability which should not matter much for small probabilities.
But it matters for payoffs, as f can depend on x. Let us see how the problem becomes very bad
when we consider f and in the presence of fat tails. Simply, you are multiplying the error in
probability by a large number, since fat tails imply that the probabilities p(x) do not decline fast
enough for large values of x. Now the literature seem to have examined errors in probability,
not errors in payoff.
Let MTX ( Az , f ) be the estimator of a function of x in the subset Az = (1 , 2 ) of the support of
the variable. Let (MTX ( Az , f )) be the mean absolute error in the estimation of the probability in
the small subset Az = (1 , 2 ), i.e.,
X ( A , 1)
E MTX ( Az , 1) M>
z
T
X
MT ( A z , f )
MTX ( Az , 1)
Assume f(x) is either linear or convex
in the form C+ x , with both >
X(but not concave)
0 and 1. Assume E[X], that is, E M>T ( AD , f ) < , for Az AD , a requirement that is not
necessary for finite intervals.
86
|1 2 | min (|2 | , |1 |) 1
MTX Az , f
X
X
E MT (Az , 1) M> T (Az , 1)
+ min (|2 | , |1 |)
MTX (Az , 1)
X (A , f )
E [ M>
]
z
T
R 2
f (x)p(x) dx
,
R 2
p(x) dx
E[
X ( A ,1)
M>
z
T
Rule 3.1.Under our conditions above, if for all (0,1) and f {i,j} (x) Az ,
(1) f i (x )+ f i (x+)
f i (x)
(1) f j (x )+ f j (x+)
,
f j (x)
X
Rule 3.2.Let ni be the number of observations required for M>
, f i the estimator under f i
T A zi
X
j under f j with observation
to get an equivalent expected mean absolute deviation as M>
T Az j , f
X
X (A , f j ) , then
size n j , that is, for MT,n
Azi , f i ) = MT,n
zj
i
j
ni n j
This inequality becomes strict in the case of nonfinite first moment for the underlying distribution.
The proofs are obvious for distributions with finite second moment, using the speed of convergence of the sum of random variables expressed in mean deviations. We will not get to them
until Chapter x on convergence and limit theorems but an example will follow in a few lines.
We will discuss the point further in Chapter x, in the presentation of the conflation problem.
For a sketch of the proof, just consider that the convex transformation of a probability distribu1/ x
( x )
1/
over
tion p(x) produces a new distribution f (x) x with density p f (x) =
its own adjusted domain, for which we find an increase in volatility, which requires a larger n
to compensate, in order to maintain the same quality for the estimator.
Example For a Gaussian distribution, the variance of the transformation becomes:
!
22 2 2
2 + 1 2
1
2
2 (1) + 1 +
(1) + 1
V x =
2
2
and to adjust the scale to be homogeneous degree 1, the variance of
V x
22 2
=
!
2 + 1 2
1
2
2 (1) + 1 +
(1) + 1
2
2
For =1, we get an idea of the increase in variance from convex transformations:
87
Variance V()
Kurtosis
15
15 6
231
5
96 8
207
10
945
10170 12
46189
63
38787711
12769
Since the standard deviation drops at the rate n for non powerlaws, the number of n(),
that is, the number of observations
needed to incur the same error on the sample in standard
V()
V(1)
, hence n1 = 2 n 2 . But to equalize the errors in mean
deviation space will be n =
n
1
deviation space, since Kurtosis is higher than that of a Gaussian, we need to translate back into
L1 space, which is elementary in most cases.
V x =
2
2 xmin
.
( 2)( 1)2
Log characteristic functions allows us to deal with the difference in sums and extract the speed
of convergence.
Example illustrating the Convex Payoff Inequality Let us compare the "true" theoreticalvalue
X
(, 3], n=200, across m simulations > 10 by estimating E MT A, x M>T A, x / MT A, x
using
j
n 1 A xij
1 m n 1 A xi
X
.
=
M>
/
T A, x
m j=1 i=1
1A
1
A
i=1
It produces the following table showing an explosive
relative error . We compare the effect to
a Gausian with matching standard deviation, namely 3. The relative error becomes infinite as
approaches the tail exponent. We can see the difference between the Gaussian and the power
law of finite second moment: both "sort of" resemble each others in many applications but...
not really.
88
St(3)
G(0,3)
0.17
0.05
3
2
0.32
0.08
0.62
0.11
5
2
1.62
0.13
f uhgetaboudit
0.18
Utility Theory Note that under a concave utility of negative states, decisions require a larger
sample. By CPSEI the magnification of errors require larger number of observation. This is
typically missed in the decision-science literature. But there is worse, as we see next.
Tail payoffs The author is disputing, in Taleb (2013), the results of a paper, Ilmanen (2013), on
why tail probabilities are overvalued by the market: naively Ilmanen (2013) took the observed
probabilities of large deviations, f (x) = 1 then made an inference for f (x) an option payoff based
on x, which can be extremely explosive (a error that can cause losses of several orders of magnitude the initial gain). Chapter x revisits the problem in the context of nonlinear transformations
of random variables. The error on the estimator can be in the form of parameter mistake that
inputs into the assumed probability distribution, say the standard deviation (Chapter x and
discussion of metaprobability), or in the frequency estimation. Note now that if 1 -, we
may have an infinite error on MTX ( Az , f ), the left-tail shortfall while, by definition, the error on
probability is necessarily bounded.
If you assume in addition that the distribution p(x) is expected to have fat tails (of any of the
kinds seen in 3.83.9, then the problem becomes more acute.
Now the mistake of estimating the properties of x, then making a decisions for a nonlinear
function of it, f (x), not realizing that the errors for f (x) are different from those of x is extremely
common. Naively, one needs a lot larger sample for f (x) when f (x) is convex than when f (x) = x.
We will re-examine it along with the "conflation problem" in Chapter x.
3.12
There is something Wrong With Econometrics, as Almost All Papers Don t Replicate. Two
reliability tests in Chapter x, one about parametric methods the other about robust statistics,
show that there is something rotten in econometric methods, fundamentally wrong, and that the
methods are not dependable enough to be of use in anything remotely related to risky decisions.
Practitioners keep spinning inconsistent ad hoc statements to explain failures.
We will show how, with economic variables one single observation in 10,000, that is, one single
day in 40 years, can explain the bulk of the "kurtosis", a measure of "fat tails", that is, both a
measure how much the distribution under consideration departs from the standard Gaussian, or
the role of remote events in determining the total properties. For the U.S. stock market, a single
day, the crash of 1987, determined 80% of the kurtosis for the period between 1952 and 2008.
The same problem is found with interest and exchange rates, commodities, and other variables.
Redoing the study at different periods with different variables shows a total instability to the
kurtosis. The problem is not just that the data had "fat tails", something people knew but sort
of wanted to forget; it was that we would never be able to determine "how fat" the tails were
within standard methods. Never.
89
The implication is that those tools used in economics that are based on squaring variables (more
technically, the L2 norm), such as standard deviation, variance, correlation, regression, the kind
of stuff you find in textbooks, are not valid scientifically(except in some rare cases where the variable is bounded). The so-called "p values" you find in studies have no meaning with economic
and financial variables. Even the more sophisticated techniques of stochastic calculus used in
mathematical finance do not work in economics except in selected pockets.
3.13
Two critical (and lethal) mistakes, entailing mistaking inclusion in a class Di for D<i because of
induced slowness in the convergence under the law of large numbers. We will see that in the
hierarchy, scale (or variance) is swamped by tail deviations.
Great Moderation (Bernanke, 2006) consists in mistaking a two-tailed process with fat tails
for a process with thin tails and low volatility.
Long Peace (Pinker, 2011) consists in mistaking a one-tailed process with fat tails for a process
with thin tails and low volatility and low mean.
Some background on Bernankes severe mistake. When I finished writing The Black Swan, in
2006, I was confronted with ideas of "great moderation" stemming from the drop in volatility
in financial markets. People involved in promulgating such theories did not realize that the
process was getting fatter and fatter tails (from operational and financial, leverage, complexity,
interdependence, etc.), meaning fewer but deeper departures from the mean. The fact that nuclear
bombs explode less often that regular shells does not make them safer. Needless to say that with
the arrival of the events of 2008, I did not have to explain myself too much. Nevertheless
90
Figure 3.21: The Turkey Problem: This is the shortest explanation of the link between evidentiary and
nonprecautionary risk management and the problem of induction. Looking for the name of the author for
credit/premission.
people in economics are still using the methods that led to the "great moderation" narrative, and
Bernanke, the protagonist of the theory, had his mandate renewed.
When I contacted social scientists I discovered that the familiarity with fat tails was pitifully
small, highly inconsistent, and confused.
The Long Peace Mistake . Later, to my horror, I saw an identical theory of great moderation
produced by Steven Pinker with the same naive statistically derived discussions (>700 pages of
them!). Except that it applied to security. The problem is that, unlike Bernanke, Pinker realized
the process had fat tails, but did not realize the resulting errors in inference.
Chapter x will get into the details and what we can learn from it.
91
10
200
400
600
800
1000
-10
-20
-30
-40
-50
92
93
Figure 3.25: High Water Mark in Palais de la Cit in Paris. The Latin poet Lucretius, who did not attend
business school, wrote that we consider the biggest objeect of any kind that we have seen in our lives as the
largest possible item: et omnia de genere omni / Maxima quae vivit quisque, haec ingentia fingit. The high water
mark has been fooling humans for millennia: ancient Egyptians recorded the past maxima of the Nile, not
thinking that the worst could be exceeded. The problem has recently affected the UK. floods with the "it
never happened before" argument. Credit Tony Veitch
94
3.14
We can see from the data in Chapter 3 that the predictability of the Gaussian-style cumulants
is low, the mean deviation of mean deviation is 70% of the mean deviation of the standard
deviation (in sample, but the effect is much worse in practice); working with squares is not a
good estimator. Many have the illusion that we need variance: we dont, even in finance and
economics (especially in finance and economics).
We propose different cumulants, that should exist whenever the mean exists. So we are not
in the dark when we refuse standard deviation. It is just that these cumulants require more
computer involvement and do not lend themselves easily to existing Platonic distributions. And,
unlike in the conventional Brownian Motion universe, they dont scale neatly.
Note finally that these measures are central since, to assess the quality of the estimation
MTX ,
X
X ( A , f )
we are concerned with the expected mean error of the empirical expectation, here E MT ( Az , f ) M>
z
T
, where z corresponds to the support of the distribution.
C0
T
xi
i=1
T
1
T
1A
i=1
T
xi 1 A or C0
i=1
1
T
D
i=1
T
xi 1 A ,
i=1
depending on whether the function of concern for the fragility metric requires conditioning or
not). The first cumulant,
C1
T
1
| x C0 |
T 1 i=1 i
produces the Mean Deviation (but centered by the mean, the first moment). The second cumulant,
C2
T
1
|| x Co |C1 |
T 2 i=1 i
1
TN
p(x) =
+(
x 2
s
)
.
s B 2 , 21
95
3.15
In the cases where | X | is hard to compute by integration, particularly with Lvy Stable distributions that do not allow no explicit densities, we can make use of the Hilbert Transform to extract
the expected mean deviations.
H ( f ) = F 1 ( i sgn( ) F ( f )),
where
1
def
[ H ( f )]( x ) = p.v.
f (t)
dx.
xt
3.16
= lim lim
a b0
Z b
a
Z a
b
[This section will be developed further; how the statemennt "but this is my prior" can be nonsense with risk management if such a prior is not solid. ]
Brad Efron (2013)[28]:
Sorry. My own practice is to use Bayesian analysis in the presence of genuine prior
information; to use empirical Bayes methods in the parallel cases situation; and otherwise to be cautious when invoking uninformative priors. In the last case, Bayesian
calculations cannot be uncritically accepted and should be checked by other methods,
which usually means frequentistically.
Diaconis and Friedman [24] show the difficulty for an agent to formulate a prior.
Further Reading
Pitman [89], Embrechts and Goldie (1982)[32]Embrechts (1979 Doctoral thesis?)[33], Chistyakov
(1964) [19], Goldie (1978)[52], Pitman[89], Teugels [113], and, more general, [34].
96
Figure 3.26: Terra Incognita: Brad Efrons positioning of the unknown that is certainly out of reach for any
type of knowledge, which includes Bayesian inference.(Efron, via Susan Holmes)
97
S P E C I A L C A S E S O F FAT TA I L S
condition
0
time
-20
-40
Low Probability
Region
-60
-80
-100
Figure B.1: The coffee cup is less likely to incur "small" than large harm; it is exposed to (almost) everything
or nothing. The same type of payoff is prevalent in markets with, say, (reval)devaluations, where small
movements beyond a barrier are less likely than larger ones.
For monomodal distributions, fat tails are the norm: one can look at tens of thousands
of time series of the socio-economic variables without encountering a single episode of
"platykurtic" distributions. But for multimodal distributions, some surprises can occur.
b.1
We noted in 1.x that stochasticizing, ever so mildly, variances, the distribution gains in fat tailedness (as expressed by kurtosis). But we maintained the same mean.
But should we stochasticize the mean as well, and separate the potential outcomes wide
enough, so that we get many modes, the "kurtosis" (as measured by the fourth moment) would
drop. And if we associate different variances with different means, we get a variety of "regimes",
each with its set of probabilities.
99
Either the very meaning of "fat tails" loses its significance under multimodality, or takes on
a new one where the "middle", around the expectation ceases to matter.[6, 72].
Now, there are plenty of situations in real life in which we are confronted to many possible
regimes, or states. Assuming finite moments for all states, s 1 a calm regime, with expected
mean m 1 and standard deviation 1 , s 2 a violent regime, with expected mean m 2 and standard
deviation 2 , and more. Each state has its probability p i .
Assume, to simplify a one-period model, as if one was standing in front of a discrete slice
of history, looking forward at outcomes. (Adding complications (transition matrices between
different regimes) doesnt change the main result.)
The Characteristic Function (t) for the mixed distribution becomes:
N
(t) =
1 2 2 +i t m
i
i
pi e 2 t
i=1
M1 = p1 m1 + ( 1 p1 ) m2
M2 = p1
m 21 + 12
+ ( 1 p 1 ) m 22 + 22
M 3 = p 1 m 31 + ( 1 p 1 ) m 2
m 22 + 322
+ 3m 1 p 1 12
M 4 = p 1 6m 21 12 + m 41 + 314 + 1 p 1
6m 22 22 + m 42 + 324
Let us consider the different varieties, all characterized by the condition p 1 < (1 p 1 ), m 1 <
m 2 , preferably m 1 < 0 and m 2 > 0, and, at the core, the central property: 1 > 2 .
Variety 1: War and Peace. Calm period with positive mean and very low volatility, turmoil with
negative mean and extremely low volatility.
100
b.1 multimodality and fat tails, or the war and peace model
Pr
S1
S2
Variety 2: Conditional deterministic state Take a bond B, paying interest r at the end of a
single period. At termination, there is a high probability of getting B(1 + r ), a possibility of
defaut. Getting exactly Bis very unlikely. Think that there are no intermediary steps between
war and peace: these are separable and discrete states. Bonds dont just default "a little bit".
Note the divergence, the probability of the realization being at or close to the mean is about nil.
Typically, p(E( x )) the probabilitity densities of the expectation are smaller than at the different
means of regimes, so P( x = E( x )) < P ( x = m 1 ) and < P ( x = m 2 ) , but in the extreme case
(bonds), P( x = E( x )) becomes increasingly small. The tail event is the realization around the
mean.
The same idea applies to currency pegs, as devaluations cannot be "mild", with all-or- nothing
type of volatility.
With option payoffs, this bimodality has the effect of raising the value of at-the-money options
and lowering that of the out-of-the-money ones, causing the exact opposite of the so-called
"volatility smile".
Note the coffee cup has no state between broken and healthy. And the state of being broken
can be considered to be an absorbing state (using Markov chains for transition probabilities),
since broken cups do not end up fixing themselves.
Nor are coffee cups likely to be "slightly broken", as we see in figure B.1.
A brief list of other situations where bimodality is encountered:
1. Currency pegs
2. Mergers
101
b.2
So far we looked at a single period model, which is the realistic way since new information may
change the bimodality going into the future: we have clarity over one-step but not more. But let
us go through an exercise that will give us an idea about fragility. Assuming the structure of the
model stays the same, we can look at the longer term behavior under transition of states. Let P
be the matrix of transition probabilitites, where pi,j is the transition from state
i to state
j over
t, (that is, where S(t) is the regime prevailing over period t, P S(t + t) = s j S(t) = s j
!
p1,1 p2,1
P=
p1,2 p2,2
After n periods, that is, n steps,
n
P =
an
bn
cn
dn
Where
an =
bn =
cn =
dn =
The extreme case to consider is the one with the absorbing state, where p1,1 = 1, hence (replacing pi,6=i|i=1,2 = 1 pi,i ).
P =
N
1 p2,2
N
p2,2
lim P =
102
1 p1,1
1 p2,2
103
C
c.1
Q U I C K A N D R O B U S T M E A S U R E O F FAT
TA I L S
introduction
We propose a new measure of fatness of tails. We also propose a quick heuristic to extract the
tail exponent and get distributions for a symmetric power law distributed variable. It is based
on using whatever moments are believed to be reasonably finite, and replaces kurtosis which in
financial data has proved to be unbearingly unstable ([109], [? ]). The technique also remedies
some of the instability of the Hill estimator, along with its natural tradoff between how much
data one must discard in otder to retain in the tails that is relevant to draw the slope. Our
estimators use the entire data available. This paper covers two situations:
1. Mild fat tails: a symmetric distribution with finite second moment, > 2 , preferably in
the neighborhood of 3. (Above 4 the measure of kurtosis becomes applicable again).
2. Extremely fat tails: a symmetric distribution with finite first moment, 1 < < 3.
Let x be a r.v. on the real line. Let x be distributed according to a Student T distribution.
! +1
2
p(x) =
(x )2
2
B
1
2, 2
(C.1)
We assume that = 0 for data in high enough frequency as the mean will not have an effect
on the estimation tail exponent.
c.2
E(x2 )
.
E(| x |)
v
u
u
u
u
tR
() =
+1
x2
2
+ x 2
1
B 2 , 2
dx
+1
)
2
| x | 2
+ x 2
1
B 2,2
q
2 2
=
1
2
(C.2)
dx
The tail from the observations: Consider a random sample of size n, (Xi )1in . Get a sample
metric
105
Hill10
Hill20
Hill100
Hill10
Cumulant Ratio
Hill20
Hill100
106
STD
MAD
for the sample (these measures do not necessarily need to be central). The estimation of m using
maximum likelihood methods [FILL]
The recovered tail .
= 1 (m) = { : () = m}
which is computed numerically.
The Hm corresponds to the measure of the m largest deviation in the right tails= (a negative
value for m means it is the left tail). We rank X(1) X(2) ... X(m) ... X(n) . The Hill
estimator
1
m
log XXi
i=1
m+1
Hm =
m
c.3
Method
Estimate
STD Error
H10
3.09681
1.06873
H20
2.82439
0.639901
H50
2.4879
0.334652
H100
2.14297
0.196846
3.26668
0.422277
2 () =
( 1)B
1
,
2 2
1
E(| x E| x ||)
E(| x |)
! 1
2
1 2
,
+4
2 2
2
+2
1
2B , 1
2 ( 1) 2 F1 2 , +1
;
;
1)
2
2
4
2 2
( 1)2 B
2 2 F1
1 +1 3
2, 2 ; 2;
+
( 1)B
0
m =
4
2
(1)2 B( 2 , 12 )
1
2, 2
2
!
+
1
2
(C.3)
n
1 i=1
| Xi MAD |
n
MAD
107
108
Method
Estimate
STD Error
H10
1.92504
0.677026
H20
1.80589
0.423783
H50
1.68919
0.237579
H100
1.56134
0.149595
1.8231
0.243436
4.1
One can make statements of the type "This is not Gaussian", or "this is not Poisson"(many people dont realize that Poisson distributions are generally thin tailed owing to finite moments);
but one cannot rule out a Cauchy tail or other similar power laws. So this chapter puts some
mathematical structure around the idea of which "empirical" statements are permissible in acceptance and rejection and which ones are not. (One can violate these statements but not from
data analysis, only basing oneself on a priori statement of what belongins to some probability
distributions.)12
Let us get deeper into the masquerade problem, as it concerns the problem of induction and
fat-tailed environments, and get to the next step. Simply, if a mechanism is fat tailed it can
deliver large values; therefore the incidence of large deviations is possible, but how possible,
how often these occur should occur, will be hard to know with any precision beforehand. This
is similar to the standard water puddle problem: plenty of ice cubes could have generated it.
As someone who goes from reality to possible explanatory models, I face a completely different
spate of problems from those who do the opposite.
We said that fat tailed series can, in short episodes, masquerade as thin-tailed. At the worst,
we dont know how long it would take to know for sure what is going on. But we can have
a pretty clear idea whether organically, because of the nature of the payoff, the "Black Swan"
can hit on the left (losses) or on the right (profits). This point can be used in climatic analysis.
Things that have worked for a long time are preferable they are more likely to have reached
their ergodic states.
1 Classical statistical theory is based on rejection and failure to reject, which is inadequade as one can reject fat tails, for
instance, which is not admissible here. Likewise this framework allows us to formally "accept" some statements.
2 This chapter was motivated by the findings in an article by Clauset, Aaron, Cosma Rohilla Shalizi, and Mark EJ Newman. "Power-law distributions in empirical data." SIAM review 51.4 (2009): 661-703, deeming that wealth data "cannot
plausibly be considered to follow a power law". The methodology they used is based on a class of "naive" power law
fitting methods than ignore the properties of out-of-sample parts.
109
4.2
masquerade example
Figure 4.1: N=1000. Sample simulation. Both series have the exact same means and variances at the level
of the generating process. Naive use of common metrics leads to the acceptance that the process A has thin
tails.
Figure 4.2: N=1000. Rejection: Another realization. there is 1/2 chance of seeing the real properties of A.
We can now reject the hypothesis that the smoother process has thin tails.
with probability (1 p)
p
with p [0,1); a, b (-1,1) and (to conserve the variance) b= a 1 p , which produces a Kurtosis
3( (1 a2 ) p1)
=
thus allowing polarized states and high kurtosis, with a condition that for a
p1
110
4.3
Our concern is exposing some errors in probabilistic statements and statistical inference, in
making inferences symmetric, when they are more likely to be false on one side than the other,
or more harmful one side than another. Believe it or not, this pervades the entire literature.
Many have the illusion that "because Kolmogorov-Smirnoff is nonparametric, it is therefore
immune to the nature specific distribution under the test (perhaps from an accurate sentence
in Feller (1971), vol 2 as we will see further down). The belief in Kolmogorov-Smirnoff is also
built in the illusion that our concern is probability rather than expected payoff, or the associated
problem of "confusing a binary for a vanilla, where by attribute substitution, one tests a certain
variable in place of another, simpler one.
In other words, it is a severe mistake to treat epistemological inequalities as equalities. No
matter what we do, we end up going back to the problem of induction, except that the world
still exists and people unburdened with too many theories are still around. By making one-sided
statements, or decisions, we have been immune to the muddle in statistical inference.
Remark 4.1 (Via negativa and the problem of induction).
Test statistics are effective (and robust) at rejecting, but not at accepting, as a single large deviation
allowed the rejection with extremely satisfactory margins (a near-infinitesimal P-Value). This illustrates
the central epistemological difference between absence of evidence and evidence of absence.3
4.4
Via negativa In theology and philosophy, corresponds to the focus on what something is not,
an indirect definition. In action, it is a recipe for what to avoid, what not to do subtraction, not
addition, say, in medicine. In epistemology: what to not accept, or accept as false. So a certain
body of knowledge actually grows by rejection. ( Antifragile[111], Glossary).
The proof and the derivations are based on climbing to a higher level of abstraction by focusing
the discussion on a hierarchy of distributions based on fat-tailedness.
Remark Test statistics can be arbitraged, or "fooled"in one direction, not the other.
Let us build a hierarchy of distributions based on tail events. But, first, a discussion of the
link to the problem of induction.
From The Black Swan (Chapter 16):
This author has learned a few tricks from experience dealing with power laws: whichever
exponent one try to measure will be likely to be overestimated (recall that a lower exponent
implies a smaller role for large deviations)what you see is likely to be less Black Swannish
than what you do not see. Lets say I generate a process that has an exponent of 1.7. You do
3 ab esse ad posse valet consequentia.
111
4.5
Let D i be a class of probability measures, D i D > i means in our terminology that a random
event "in" D i would necessarily "be in" D j , with j > i, and we can express it as follows. Let A K
+
be a one-tailed interval in R, unbounded on one side K, s.a. A
K = ( , K or A K = [K , ) ,
and ( A) the probability measure on the interval, which corresponds to i ( A K ) the cumulative
distribution function for K on the left, and i ( A +K ) = 1 the CDF (that is, the exceedance
probability) on the right.
For continuous distributions, we can treat of the Radon-Nikodym derivatives for two measures
i
over as the ratio of two probability with respect to a variable in A K .
j
such that j ( A
K 0 > i ( A K 0 ) and j ( A K ) > i ( A K ) for all K < K 0 .
The derivations are as follows. Simply, the effect of the scale of the distribution (say, the
variance in the finite second moment case) wanes in the tails. For the classes of distributions up
to the Gaussian, the point is a no brainer because of compact support with 0 measure beyond
a certain K. As as far as the Gaussian, there are two brands, one reached as a limit of, say, a
sum of n Bernouilli variables, so the distribution will have compact support up to a multiple
of n at infinity, that is, in finite processes (what we call the "real world"where things are finite).
The second Gaussian category results from an approximation; it does not have compact support
but because of the exponential decline in the tails, it will be dominated by power laws. To quote
Adrien Douady, it has compact support for all practical purposes.4
Let us focus on the right tail.
Case of Two Powerlaws
For powerlaws, let us consider the competing effects of scale, say (even in case of nonfinite
variance), and tail exponent, with > 1 . Let the density be
P, (x) = L(x)x 1
where L(x) is a slowly varying function,
4 Van Zwet,[cite]: Given two cumulative distribution functions F(x) and G(x), F has lighter tails than G (and G has heavier
tails than F) if the function G 1 (F(x)) is convex for x 0.
112
r,k (x)
P,k (x)
P, (x)
By only perturbating the scale, we increase the tail by a certain factor, since limx r1,k (x) = k ,
which can be significant. But by perturbating both and looking at the limit we get limx r,k (x)
(1+)
= k Lx
, where L is now a constant, thus making the changes to the tail exponent
leading for large values of x.
Obviously, by symmetry, the same effect obtains in the left tail.
Rule 4.1.When comparing two power laws, regardless of parametrization of the scale parameters for
either distributions, the one with the lowest tail exponent will have higher density in the tails.
x2
22
log2 (x)
2k2 s2
log2 (x)
2k2 s2
ksx
d=
log2 (x)
2k2 s2
ksx
2
x2
2
and
2
x2
2
log(ksx) + log(), which is dominated by the first term x2 as it is convex when the
other terms are concave, so it will be > 0 for large values of x independently of parameters.
Rule 4.2.Regardless of parametrization of the scale parameter (standard deviation) for either distribution, a lognormal will produce asymptotically higher tail densities in the positive domain than the
Gaussian.
Case of Mixture of Gaussians
Let us return to the example of the mixture distribution N(0, ) with probability 1 p and
N(0, k ) with the remaining probability p. The density of the second regime weighted by p
x2
2 2
p e 2k
k 2
x2
113
Class
Description
D1
D2
Thin tails
D3a
D3b
Conventional Thin
tails
Starter Fat Tails
D5
D6
Subexponential
Supercubic
D7
Infinite Variance
D8
Undefined
Moment
Gaussian approximation of a
natural phenomenon
Higher kurtosis than the Gaussian but rapid convergence to
Gaussian under summation
(e.g. lognormal)
Cramer conditions
do not hold
R
for t > 3, etx d(Fx) =
Levy
<
2 ,
R tx Stable
e dF(x) =
Fuhgetaboutdit
First
A comment on 4.3
Gaussian From Convergence is Not Gaussian : We establish a demarcation between two levels
of Gaussians. Adding Bernouilli variables or Binomials, according to the random walk idea (or
similar mechanism that generate Gaussians) always leads to thinner tails to the true Gaussian.
Subgaussian domain for a review,[17], Kahanes "gaussian shift"5 :
Mixtures distributions entailing Di and D j are classified with the highest level of fat tails
Dmax(i,j) regardless of the mixing. A mixture of Gaussians remains Gaussian for large deviations,
even if the local properties can be confusing in small samples, except for the situation of infinite
nesting of stochastic volatilities discussed in Chapter 6. Now a few rapidly stated rules.
Rule 4.4.(General Decision Making Heuristic). For any information entailing nonbinary decision (see definition in Chapter x), rejection or acceptance of fitness to pre-specified probability distributions, based on suprema of distance between supposed probability distributions (say Kolmogorov
Smirnoff and similar style) should only be able to "accept" the fatter tail one and "reject"the lower
tail, i.e., based on the criterion i > j based on the classification above.
Warning 1 : Always remember that one does not observe probability distributions, only realizations. Every probabilistic statement needs to be discounted by the probability of the parameter being away from the true one.
5 J.P. Kahane, "Local properties of functions interms of random Fourier series," Stud. Math., 19, No. i, 1-25 (1960)
114
Fuhgetaboudit
Lvy-Stable <2
L1
Supercubic 3
Subexponential
CRAMER
CONDITION
Gaussian from Lattice Approximation
COMPACT
SUPPORT
Degenerate
LAW OF LARGE NUMBERS (WEAK)
CONVERGENCE ISSUES
Figure 4.3: The tableau of Fat tails, along the various classifications for convergence purposes (i.e., convergence to the law of large numbers, etc.)A variation around Embrechts et al [31], but applied to the
Radon-Nikodym derivatives.
Warning 2 : Recall that we do not live in probability space, but payoff space.
Rule 4.5.(Decision Mistakes). Fatter tailed distributions are more likely to produce a lower insample variance (using empirical estimators) than a distribution of thinner tail of the same variance
(in the finite variance case).
For the derivation, recall that (from 3.5), there in increase in observations in the "tunnel"( a2 , a3 )
in response to increase in fat-tailedness.
4.6
Counterintuitively, when one raises the kurtosis, as in Figure 4.1.4.1 the time series looks "quieter. Simply, the storms are rare but deep. This leads to mistaken illusion of low volatility
115
Table of the "fake"Gaussian when not busted Let us run a more involved battery of statistical
tests (but consider that it is a single run, one historical simulation).
Statistic
P-Value
Anderson-Darling
0.406988
0.354835
Cramr-von Mises
0.0624829 0.357839
Jarque-Bera ALM
1.46412
0.472029
Kolmogorov-Smirnov 0.0242912 0.167368
0.0424013 0.110324
Kuiper
Fake Distr
Mardia Combined
1.46412
0.472029
Mardia Kurtosis
0.876786 0.380603
Mardia Skewness
0.7466
0.387555
2
43.4276
0.041549
Pearson
Shapiro-Wilk
0.998193
0.372054
Watson U 2
0.0607437 0.326458
Anderson-Darling
Cramr-von Mises
Jarque-Bera ALM
Kolmogorov-Smirnov
Kuiper
Genuine
Mardia Combined
Mardia Kurtosis
Mardia Skewness
Pearson 2
Shapiro-Wilk
Watson U 2
Statistic
0.656362
0.0931212
3.90387
0.023499
0.0410144
3.90387
1.83609
0.620678
33.7093
0.997386
0.0914161
P-Value
0.0854403
0.138087
0.136656
0.204809
0.144466
0.136656
0.066344
0.430795
0.250061
0.107481
0.116241
when in fact it is just high kurtosis, something that fooled people big-time with the story of
the "great moderation"as risks were accumulating and nobody was realizing that fragility was
increasing, like dynamite accumulating under the structure.
Kolmogorov - Smirnov, Shkmolgorov-Smirnoff Remarkably, the fat tailed series passes general test of normality with better marks than the thin-tailed one, since it displays a lower variance. The problem discussed with with Avital Pilpel (Taleb and Pilpel, 2001, 2004, 2007) is that
Kolmogorov-Smirnov and similar tests of normality are inherently self-referential.
These probability distributions are not directly observable, which makes any risk calculation suspicious
since it hinges on knowledge about these distributions. Do we have enough data? If the distribution is,
say, the traditional bell-shaped Gaussian, then yes, we may say that we have sufficient data. But if the
116
Anderson-Darling
Cramr-von Mises
Jarque-Bera ALM
Kolmogorov-Smirnov
Kuiper
Busted Fake
Mardia Combined
Mardia Kurtosis
Mardia Skewness
Pearson 2
Shapiro-Wilk
Watson U 2
Statistic
376.05
80.734
4.21 107
0.494547
0.967
4.21 107
6430.
166432.
30585.7
0.014
80.58
P-Value
0.
0.
0.
0.
0.
0.
1.5 108979680
1.07 1036143
3.28 106596
1.91 1057
0.
distribution is not from such well-bred family, then we do not have enough data. But how do we know
which distribution we have on our hands? Well, from the data itself .
If one needs a probability distribution to gauge knowledge about the future behavior of the distribution
from its past results, and if, at the same time, one needs the past to derive a probability distribution in the
first place, then we are facing a severe regress loopa problem of self reference akin to that of Epimenides
the Cretan saying whether the Cretans are liars or not liars. And this self-reference problem is only the
beginning.
(Taleb and Pilpel, 2001, 2004)
Also,
From the Glossary in The Black Swan . Statistical regress argument (or the problem of the
circularity of statistics): We need data to discover a probability distribution. How do we know if we have
enough? From the probability distribution. If it is a Gaussian, then a few points of data will suffice. How
do we know it is a Gaussian? From the data. So we need the data to tell us what probability distribution
to assume, and we need a probability distribution to tell us how much data we need. This causes a severe
regress argument, which is somewhat shamelessly circumvented by resorting to the Gaussian and its kin.
A comment on the Kolmogorov Statistic It is key that the Kolmogorov-Smirnov test doesnt
affect payoffs and higher moments, as it only focuses on probabilities. It is a severe problem
because the approximation will not take large deviations into account, and doesnt make it
useable for our purpose. But thats not the only problem. It is, as we mentioned, conditioned on
sample size while claiming to be nonparametric.
Let us see how it works. Take the historical series and find the maximum point of divergence
with F(.) the cumulative of the proposed distribution to test against:
!n
1 J
where n =
T t0
t
117
0.8
0.6
0.4
0.2
D
x
1
We will get more technical in the discussion of convergence, take for now that the Kolmogorov
statistic, that is, the distribution of D, is expressive of convergence, and should collapse with n.
The idea is that, by a Brownian Bridge argument (that is a process
pinned on both sides,
with
J
i=1
Xti+t0
intermediate steps subjected to double conditioning), D j =
F Xtj+t0 which is
j
Uniformly distributed.
The probability of exceeding D, P> D = H
nD , where H is the cumulative distribution
function of the Kolmogorov-Smirnov distribution,
2 t2
i=1
We can see that the main idea reposes on a decay of nD with large values of n. So we can
easily fool the testing by proposing distributions with a small probability of very large jump,
where the probability of switch . n1 .
The mistake in misinterpreting Feller: the distribution of Dwill be uniform independently of
the distribution under scrutiny, or the two distributions to be compared. But it does not mean
that the test is immune to sample sizen, that is, the possibility of jump with a probability an
inverse function of n.
Use of the supremum of divergence
Note another manifestation of the error of ignoring the effect of the largest deviation. As we saw
with Kolmogorov-Smirnoff and other rigorous methods in judging a probability distribution,
one focuses on the maximum divergence, the supremum, as information. Another unused today
but very potent technique, initially by Paul Levy (1924), called the concentration function, also
reposes on the use of a maximal distance:
From Petrov (1995):
Q (X) sup P(x X x + )
x
for every 0.
118
4.7
119
120
Chapter Summary 5: The Spectrum Between Uncertainty and Risk. There has been a
bit of discussions about the distinction between "uncertainty" and "risk". We believe in
gradation of uncertainty at the level of the probability distribution itself (a "meta" or
higher order of uncertainty.) One end of the spectrum, "Knightian risk", is not available
for us mortals in the real world. We show how the effect on fat tails and on the calibration
of tail exponents and reveal inconsistencies in models such as Markowitz or those used
for intertemporal discounting (as many violations of "rationality" arent violations .
5.1
meta-probability distribution
When one assumes knowledge of a probability distribution, but has uncertainty attending the
parameters, or when one has no knowledge of which probability distribution to consider, the
situation is called "uncertainty in the Knightian sense" by decision theorisrs(Knight, 1923). "Risk"
is when the probabilities are computable without an error rate. Such an animal does not exist
in the real world. The entire distinction is a lunacy, since no parameter should be rationally
computed witout an error rate. We find it preferable to talk about degrees of uncertainty about
risk/uncertainty, using metadistribution, or metaprobability.
() d ,
(5.1)
() d
(5.2)
In other words, if one is not certain about a parameter , there is an inescapable layer of stochasticity; such stochasticity raises the expected (metaprobability-adjusted) probability if it is < 12
and lowers it otherwise. The uncertainty is fundamentally epistemic, includes incertitude, in the
sense of lack of certainty about the parameter.
121
0.1
ni=1 p HX i L i
pHX * L
0.001
10-5
10
-7
pIX M
X
5
10
50
100
500
1000
Figure 5.1: Log-log plot illustration of the asymptotic tail exponent with two states.
The model bias becomes an equivalent of the Jensen gap (the difference between the two
sides of Jensens inequality), typically positive since probability is convex away from the center
of the distribution. We get the bias A from the differences in the steps in integration
Z
Z
A =
() p( x | ) d p x | () d
With f ( x ) a function , f ( x ) = x for the mean, etc., we get the higher order bias A 0
Z
Z Z
Z
A0 =
() f ( x ) p( x | ) d d x
f ( x ) p x | () d d x
(5.3)
n
Now assume the distribution of as discrete n states, with = ( i ) i=1
each with associated
n
probability = i _i=1^n, i=1 i = 1. Then 5.2 becomes
!
n
p( x ) = i
p ( x |i )
(5.4)
i=1
5.2
Remark 5.1.
In the presence of a layer of metadistributions (from uncertainty about the parameters), the asymptotic tail
exponent for a powerlaw corresponds to the lowest possible tail exponent regardless of its probability.
This explains "Black Swan" effects, i.e., why measurements tend to chronically underestimate
tail contributions, rather than merely deliver imprecise but unbiased estimates.
122
p( x ) =
i
i
i x i 1 x min
i=1
+1
i
i = K
i x i 1 x min
i=1
n
i
where K is a strictly positive constant and = mi n i . In other words i=1
i x i 1 x min
i
1in
is asymptotically equivalent to a constant times x +1 . The lowest parameter in the space of all
possibilities becomes the dominant parameter for the tail exponent.
P>x
0.0004
0.0003
0.0002
Bias A
0.0001
STD
1.3
1.4
1.5
1.6
1.7
1.8
Figure 5.2: Illustration of the convexity bias for a Gaussian from raising small probabilities: The plot
shows the STD effect on P>x, and compares P>6 with a STD of 1.5 compared to P> 6 assuming a linear
combination of 1.2 and 1.8 (here a(1)=1/5).
n
Figure 5.1 shows the different situations: a) p(x | ), b) i=1
p ( x |i ) i and c) p ( x | ). We can
see how the last two converge. The asymptotic Jensen Gap A becomes p(x | ) p(x | ).
Implications
Whenever we estimate the tail exponent from samples, we are likely to underestimate the thickness of the tails, an observation made about Monte Carlo generated -stable variates and the
estimated results (the Weron effect)[121].
The higher the estimation variance, the lower the true exponent.
123
5.3
Recall that the tail fattening methods in 3.4 and 3.6.These are based on randomizing the variance.
Small probabilities rise precisely because they are convex to perturbations of the parameters (the
scale) of the probability distribution.
5.4
Risk management failed on several levels at Fukushima Daiichi. Both TEPCO and its captured
regulator bear responsibility. First, highly tailored geophysical models predicted an infinitesimal
chance of the region suffering an earthquake as powerful as the Tohoku quake. This model uses
historical seismic data to estimate the local frequency of earthquakes of various magnitudes;
none of the quakes in the data was bigger than magnitude 8.0. Second, the plants risk analysis
did not consider the type of cascading, systemic failures that precipitated the meltdown. TEPCO
never conceived of a situation in which the reactors shut down in response to an earthquake, and
a tsunami topped the seawall, and the cooling pools inside the reactor buildings were overstuffed
with spent fuel rods, and the main control room became too radioactive for workers to survive,
and damage to local infrastructure delayed reinforcement, and hydrogen explosions breached
the reactors outer containment structures. Instead, TEPCO and its regulators addressed each of
these risks independently and judged the plant safe to operate as is.Nick Werle, n+1, published
by the n+1 Foundation, Brooklyn NY
5.5
Assume that someone tells you that the probability of an event is exactly zero. You ask him
where he got this from. "Baal told me" is the answer. In such case, the person is coherent, but
would be deemed unrealistic by non-Baalists. But if on the other hand, the person tells you "I
estimated it to be zero," we have a problem. The person is both unrealistic and inconsistent.
Something estimated needs to have an estimation error. So probability cannot be zero if it is
estimated, its lower bound is linked to the estimation error; the higher the estimation error, the
higher the probability, up to a point. As with Laplaces argument of total ignorance, an infinite
estimation error pushes the probability toward 12 . We will return to the implication of the
mistake; take for now that anything estimating a parameter and then putting it into an equation
is different from estimating the equation across parameters. And Markowitz was inconsistent
by starting his "seminal" paper with "Assume you know E and V" (that is, the expectation and
the variance). At the end of the paper he accepts that they need to be estimated, and what is
124
x
(5.5)
2
*
N
weights matrix = (1 , . . . , m ),normalized, with i=1
i = 1 (allowing exposures to be both
positive and negative): The scalar of concern is; r = T .X, which happens to be normally
distributed, with variance
~ T ..
~
v=
125
5.6
Often psychologists and behavioral economists find "irrational behavior" (or call it under something more polite like "biased") as agents do not appear to follow a normative model and violate
their models of rationality. But almost all these correspond to missing a second layer of uncertainty by a dinky-toy first-order model that doesnt get nonlinearities it is the researcher who
is making a mistake, not the real-world agent. Recall that the expansion from "small world"
to "larger world" can be simulated by perturbation of parameters, or "stochasticization", that
is making something that appears deterministic a random variable itself. Benartzi and Thaler
[8], for instance, find an explanation that agents are victims of a disease labelled "myopic loss
aversion" in not investing enough in equities, not realizing that these agents may have a more
complex, fat-tailed model. Under fat tails, no such puzzle exists, and if it does, it is certainly not
from such myopia.
This approach invites "paternalism" in "nudging" the preferences of agents in a manner to fit
professors-without-skin-in-the-game-using-wrong-models.
5.6.1
Today many mathematical or conceptual models that are claimed to be rigorous are based upon
fixed parameters that miss a layer of uncertainty. Such models are deemed rational in the sense
that they are logically derived from their assumptions, and, further, can be used to assess rationality by examining deviations from such models, as indicators of irrationality. Except that it
is often the modeler who is using an incomplete representation of the reality, hence using an
erroneous benchmark for rationality. Often the modelers are not familiar with the dynamics
of complex systems or use antiquated statistical methods that do not take into account fat-tails
and make inferences that would not be acceptable under different classes of probability distributions. Many biases, such as the ones used by Cass Sunstein, about the overestimation of the
probabilities of rare events in fact correspond to the testers using a bad probability model that
is thin-tailed.
It has became popular to claim irrationality for GMO and other skepticism on the part of the
general publicnot realizing that there is in fact an "expert problem" and such skepticism is
healthy and even necessary for survival. For instance, in The Rational Animal [? ], the authors
pathologize people for not accepting GMOs although "the World Health Organization has never
found evidence of ill effects," a standard confusion of evidence of absence and absence of evidence. Such pathologizing is similar to behavioral researchers labeling hyperbolic discounting
as "irrational" when in fact it is largely the researcher who has a very narrow model and richer
models make the "irrationality" go away, as we will see further down.
These researchers fail to understand that humans may have precautionary principles against
systemic risks, and can be skeptical of the untested consequences of policies for deeply rational
reasons, even if they do not express such fears in academic format.
Hidden convexities Let us use our approach in detecting convexity to three specific problems:
1) the myopic loss aversion that we just discussed, 2) time preferences, 3) probability matching.
126
-0.04
-0.05
-0.06
-0.07
-0.08
Higher values
of a
-0.09
0.10
0.15
0.20
0.25
H 1
a,
H1
1.6
1.5
Ha, 1 (t)
Figure 5.4: The ratio H20 or the degradation of "utility" under second order effects.
1.4
1.3
1.2
1.1
0.2
0.4
0.6
0.8
w, (x)t,t (x) dx
(5.6)
127
/2
q 1
t2
1
1
t /2
+1
t
1 F1 2 ; 2 ; 22
2
2 t
2 t
+
1
2
+1
+1 t 2 +1
1
2 t
+1
2
+1
t 2
1
2 t
/2
+ 2t
1
2 t
1 F1
t2
1 3
2 ; 2 ; 22
(5.7)
We can see from 5.7 that the more frequent sampling of the performance translates into worse
utility. So what Benartzi and Thaler did was try to find the sampling period "myopia" that
translates into the sampling frequency that causes the "premium" the error being that they
missed second order effects.
Now under variations of with stochatic effects, heuristically captured, the story changes:
what if there is a very small probability that the variance gets multiplied by a large number,
with the total variance remaining the same? The key here is that we are not even changing
the variance at all: we are only shifting the distribution to the tails. We are here generously
assuming that by the law of large numbers it was established that the "equity premium puzzle"
was true and that stocks really outperformed bonds.
So we switch between two states, (1 + a) 2 w.p. p and (1 a) w.p. (1 p).
Rewriting 5.6
Ha,p (t) =
w, (x) p t,1+a t (x) + (1 p) t,1 a t (x) dx
(5.8)
Result Conclusively, as can be seen in figures 5.3 and 5.4, second order effects cancel the statements made from "myopic" loss aversion. This doesnt mean that myopia doesnt have effects,
rather that it cannot explain the "equity premium", not from the outside (i.e. the distribution
might have different returns", but from the inside, owing to the structure of the KahnemanTversky value function v(x).
Comment We used the (1+a) heuristic largely for illustrative reasons; we could use a full
distribution for 2 with similar results. For instance the gamma distribution with density
v
v1 e V ( V )
with expectation V matching the variance used in the "equity premium"
f (v) =
()
theory.
Rewriting 5.8 under that form,
Z Z
0
Which has a closed form solution (though a bit lengthy for here).
Time preference under model error
This author once watched with a great deal of horror one Laibson [67] at a conference in
Columbia University present the idea that having one massage today to two tomorrow, but
reversing in a year from now is irrational and we need to remedy it with some policy. (For
128
e t () d
It is easy to prove the general case that under symmetric stochasticization of intensity (that
is, with probabilities 12 around the center of the distribution) using the same technique we did
in 3.4:
H 0 (t, ) =
1 ()t (+)t
e
+e
2
Example: Gamma Distribution Under the gamma distribution with support in R+ , with parameters and , () =
we get:
1 e
()
H(t, , ) =
Z
0
so
lim
1 e
()
d =
1
+t
H(t, , )
=1
H(t , , )
Meaning that preferences become flat in the future no matter how steep they are in the present,
which explains the drop in discount rate in the economics literature.
1 I discovered that [38] Farmer and Geanakoplos have applied a similar approach to Hyperbolic discounting
129
()=
we get the normatively obtained (not empirical pathology) so-called hyperbolic discounting:
H(t) =
130
1
1+kt
L A R G E N U M B E R S A N D C LT I N T H E R E A L
WORLD
Chapter Summary 6: The Law of Large Numbers is the foundation of statistical knowledge or, even (inductive) knowledge tout court. The behavior of the sum of random
variables allows us to get to the asymptote and use handy asymptotic properties. However real life is more complicated. We cannot
talk about LLN without figuring out the
speed of convergence, which, when it is at n, is only so asymptotically. Further, in some
cases the LLN doesnt work at all. For very fat tailed, under the slightest parametric error,
it will be more than 400 times slower than thought.
You observe data and get some confidence that the average is represented by the sample
thanks to a standard metrified "n". Now what if the data were fat tailed? How much more do
you need? What if the model were uncertain we had uncertainty about the parameters or the
probability distribution itself?
Figure 6.1: How thin tails (Gaussian) and fat tails (1< 2) converge to the mean.
Main Results In addition to explicit extractions of partial expectations for alpha stable
distributions, one main result in this paper is the expression of how uncertainty about
parameters (in terms of parameter volatility) translates into a larger (or smaller) required n.
Model Uncertainty The practical import is that model uncertainty worsens inference, in a
quantifiable way.
6.0.2
It is also necessary to debunk a fallacy: we simply do not have enough data with commonly
discussed fat-tailed processes to naively estimate a sum and make series of claims about stability
of systems, pathology of people reacting to risks, etc. A surprising result: for the case with
131
6.1
By the weak law of large numbers, consider a sum of random variables X1 , X2 ,..., Xn independent
and identically distributed with finite mean m, that is E[Xi ] < , then n1 1in Xi converges
to m in probability, as n . And the idea is that we live with finite n.
We get most of the intuitions from closed-form and semi-closed form expressions working
with:
stable distributions (which allow for a broad span of fat tails by varying the exponent,
along with the asymmetry via the coefficient
stable distributions with mixed exponent.
other symmetric distributions with fat-tails (such as mixed Gaussians, Gamma-Variance
Gaussians, or simple stochastic volatility)
More complicated situations entailing more numerical tinkering are also covered: Pareto classes,
lognormal, etc.
Instability of Mean Deviation
Indexing with p the property of the variable X p and g
n
!
(
p Xp m
p
i
n p : E
=E
np
(6.1)
1 The Pinker Problem A class of naive empiricism. It has been named so in reference to sloppy use of statistical techniques
in social science and policy making, based on a theory promoted by the science writer S. Pinker [88] about the drop of
violence that is based on such statistical fallacies since wars unlike domestic violence are fat tailed. But this is a very
general problem with the (irresponsible) mechanistic use of statistical methods in social science and biology.
132
()
1.7
1.6
1.5
1.4
1.0
1.5
2.0
2.5
3.0
And since we know that convergence for the Gaussian happens at speed n 2 , we can compare
to convergence of other classes.
We are expressing in Equation 6.1 the expected error (that is, a risk function) in L1 as mean absolute deviation from the observed average, to accommodate absence of variance but assuming
of course existence of first moment without which there is no point discussing averages.
Typically, in statistical inference, one uses standard deviations of the observations to establish
the sufficiency of n. But in fat tailed data standard deviations do not exist, or, worse, when they
exist, as in powerlaw with tail exponent > 3, they are extremely unstable, particularly in cases
where kurtosis is infinite.
Using mean deviations of the samples (when these exist) doesnt accommodate the fact that
fat tailed data hide properties. The "volatility of volatility", or the dispersion around the mean
deviation increases nonlinearly as the tails get fatter. For instance, a stable distribution with
tail exponent at 23 matched to exactly the same mean deviation as the Gaussian will deliver
measurements of mean deviation 1.4 times as unstable as the Gaussian.
Using mean absolute deviation for "volatility", and its mean deviation "volatility of volatility"
expressed in the L1 norm, or C1 and C2 cumulant:
C1 = E(| X m|)
C2 = E (| X E(| X m|)|)
We can compare that matching mean deviations does not go very far matching cumulants.(see
Appendix 1)
Further, a sum of Gaussian variables will have its extreme values distributed as a Gumbel
while a sum of fat tailed will follow a Frchet distribution regardless of the the number of summands.
The difference is not trivial, as shown in figures , as in 106 realizations for an average with 100
summands, we can be expected observe maxima > 4000 the average while for a Gaussian we
can hardly encounter more than > 5 .
133
6.2
It is unfortunate that even if one matches mean deviations, the dispersion of the distributions
of the mean deviations (and their skewness) would be such that a "tail" would remain markedly
different in spite of a number of summands that allows the matching of the first order cumulant.
So we can match the special part of the distribution, the expectation > K or < K, where K can
be any arbitrary level.
Let (t) be the characteristic function of the random variable. Let be the Heaviside theta
function. Since sgn(x) = 2(x) 1
(t) =
eitx (2(x K) 1) dx =
2ieiKt
t
R
And define the partial expectation as E+K K x dF(x) = E(X | X >K )P(X > K). The special
expectation becomes, by convoluting the Fourier transforms; where F is the distribution function
for x:
Z
(t u) (u)du|t=0
(6.2)
E+K = i
t
R
R
Mean deviation becomes a special case of equation 6.2, E(| X |) = x dF(x) x dF(x).
6.3
Assume alpha-stable the class S of probability distribution that is closed under convolution:
S(, , , ) represents the stable distribution with tail index (0, 2], symmetry parameter
[0, 1], location parameter R, and scale parameter R+ . The Generalized Central
Limit Theorem gives sequences an and bn such that the distribution of the shifted and rescaled
sum Zn = (in Xi an ) /bn of n i.i.d. random variates Xi the distribution function of which FX (x)
has asymptotes 1 cx as x + and d( x) as x weakly converges to the stable
distribution
cd
S(,2 , 10<<2
, 0, 1).
c+d
We note that the characteristic functions are real for all symmetric distributions. [We also note
that the convergence is not clear across papers[119] but this doesnt apply to symmetric distributions.]
Note that the tail exponent used in non stable cases is somewhat, but not fully, different
for = 2, the Gaussian case where it ceases to be a powerlaw the main difference is in the
asymptotic interpretation. But for convention we retain the same symbol as it corresponds to
tail exponent but use it differently in more general non-stable power law contexts.
The characteristic function (t) of a variable X with scale will be, using the expression for
> 1, See Zolotarev[125], Samorodnitsky and Taqqu[98]:
= exp it |t | 1 i tan
sgn(t)
2
which, for an n-summed variable (the equivalent of mixing with equal weights), becomes:
1
(t) = exp int n t 1 i tan
sgn(t)
2
134
Results
1
2
|u|2 1
+ i tan
sgn(u) exp |u | 1 i tan
sgn(u) + iKu du
2
2
(6.3)
1/
1/
1
1
+ 1 i tan
.
1 + i tan
2
2
2
E+K (, , , , K) =
2
1 k
k+1
k k
2 tan2
(1)k 1 + i tan
i K
2 +1
+
2k1 k!
k=1
(6.4)
1/
k1
+ 1 i tan
k1
(6.5)
Our formulation in Equation 6.5 generalizes and simplifies the commonly used one from Wolfe
[122] from which Hardin [55] got the explicit form, promoted in Samorodnitsky and Taqqu [98]
and Zolotarev[125]:
!!
21
1 tan
tan
1
1 2
2
E(| X |) = 2 1
tan2
+1
cos
1
n = 2 1 22
ng
1 i tan
+ 1 + i tan
(6.6)
2
2
with alternative expression:
n = 22
2
sec
1
1
tan (tan(
2 ))
sec
1
ng
1
2
2
(6.7)
1
n g
n =
2(1)
(6.8)
135
(6.9)
n
Fughedaboudit
9
8
= 12
=1
6.09 1012
2.8 1013
1.86 1014
5
4
574,634
895,952
1.88 106
11
8
5,027
6,002
8,632
3
2
567
613
737
13
8
165
171
186
7
4
75
77
79
15
8
44
44
44
30.
30
30
Remark 6.1.
1
The ratio mean deviation of distributions in S is homogeneous of degree k . 1 . This is not the case for
other classes "nonstable".
Proof. (Sketch) From the characteristic function of the stable distribution. Other distributions
need to converge to the basin S.
6.3.2
(X )
i (Xi )
i=1
where =
136
m
i=1
i i and
m
i=1
i = 1.
(|X|)
= 5/4
3.5
3.0
2.5
-1.0
2.0
= 3/2
1.5
= 7/4
0.5
-0.5
1.0
2
2
350
300
250
200
150
100
50
1.4
1.6
1.8
2.0
137
=
n
+ H 1 + log(n) + 2 H 1 + log(n) +
2 4
which is positive for values in the specified range, keeping < 2 as it would no longer converge
to the Stable basin.
Which is also negative with respect to al pha as can be seen in Figure 6.4. The implication is
that ones sample underestimates the required "n". (Commentary).
6.4
6.4.1
While mixing Gaussians the kurtosis rises, which makes it convenient to simulate fattailedness.
But mixing means has the opposite effect, as if it were more "stabilizing". We can observe a
similar effect of "thin-tailedness" as far as the n required to match the standard benchmark. The
situation is the result of multimodality, noting that stable distributions are unimodal (Ibragimov
and Chernin) [60] and infinitely divisible Wolfe [123]. For Xi Gaussian with mean , E =
q
2
+ 2 e 22 , and keeping the average with probability 1/2 each. With the
erf
2
perfectly symmetric case = 0 and sampling with equal probability:
2
2
2
2
erf
2
1
e 2
1
erf e +
2
(E+ + E ) =
+ erf
2
2
2
2
2
q
2
2
+
exp
6.4.2
2
2
2
e
+ erf
22
Relative convergence:
Theorem 6.1.
For all so and so, (details), etc.
X m
E kn i n
c2
c1
X m
E n i n
where:
138
c 1 = k 1
(6.10)
0.12
5
2
0.08
0.10
0.06
0.04
0.02
20 000
40 000
60 000
80 000
100 000
20 000
40 000
60 000
80 000
100 000
Figure 6.5: Different Speed: the fatter tailed processes are not just more uncertain; they also converge more
slowly.
2
1
c2 = 27/2 1/2
4
Note that because the instability of distribution outside the basin, they end up converging to
S Min(,2) , so at k = 2, n = 1, equation 6.10 becomes an equality and k we satisfy the equalities
in ?? and 6.9.
Proof. (Sketch)
The characteristic function for = 23 :
33/8 |t|3/4 K 3
4
(t) =
8
2
q
3
4
3
2
|t|
2
3 34 74
6.4.3
Student T with 3 degrees of freedom (higher exponent resembles Gaussian). We can get a semiexplicit density for the Cubic Student T.
p(x) =
6 3
( x 2 + 3)
we have:
(t) = E[eitX ] = (1 +
3 | t |) e
3 |t|
139
1
n
xi |
0.7
0.6
0.5
Figure 6.6: Student T with exponent =3. This applies to the general class of symmetric power
law distributions.
0.4
0.3
0.2
0.1
10
20
30
40
50
Z +
using
Z
0
k t
t e
(1 +
3 t)n en
3t
cos(tx) dt
where Ta (x) is the T-Chebyshev polynomial,2 the pdf p(x) can be writen:
p(x) =
2 n 1
n2 + x3
3
n! n2 +
n
k=0
1k +n
2
x
3
Tk+1 r
1
x2 +1
3n2
(n k)!
which allows explicit solutions for specific values of n, not not for the general form:
{ En } 1 n <
(
)
2 3 3 3 34 71 3 3138 3
899
710162 3 425331 3
33082034
5719087 3
,
,
=
,
, ,
,
,
,
,
,...
2 With thanks to Abe Nassen and Jack DAurizio on Math Stack Exchange.
140
0.5
p=
1
2
0.4
p=
1
8
Gaussian
0.3
0.2
4
10
6.5
6.5.1
6.5.2
6.6
6.6.1
6.6.2
This is the most untractable case mathematically, apparently though the most present when we
discuss fat tails [124].
6.6.3
Consider a sum of Bernoulli variables X. The average n in xi follows a Binomial Distribution. Assuming np N+ to simplify:
n
E (|n |) = 2 (x np) p x
(1 p)n x
x
i 0np
E |n | = 2(1 p)n( p)+n2 pnp+1 (np + 2)
n
n
(p 1)
1 p(np + 2)
2
np + 1
np + 2
where:
1 =2 F1 1, n(p 1) + 1; np + 2;
p
p1
141
p=
0.5
0.4
1
2
p=
1
100
Gaussian
0.3
0.2
0.1
20
40
and
6.7
60
80
100
2 =2 F1 2, n(p 1) + 2; np + 3;
p
p1
acknowledgement
Colman Humphrey,...
Cumulants
+e
C2 = erf(
C1
which is 1.30 C1 .
For a powerlaw distribution, cumulants are more unwieldy:
=3/2
C1
6
5
4
34
Move to appendix
p q
1
9/2
9/2
9/4
9/4
5/4 3 5/2
=
384
+24
2
2 4 21 + 23 1 H1
2
2 1
1
2 6 3/2 31 21 + 23 5/4
q
q
4
5 4
3 4
3
3/4
2
2
2
2
+ 15362 1 + 3 H2 +
1 + 3 3 21 + 33 (H2 + 2) 2 2 H1
=3/2
C2
142
6.7 acknowledgement
where 1 =
6.7.2
3
4
, 2 =
5
4
, 3 =
1
4
, H1 = 2 F1
1
3 5 7
4 , 4 ; 4 ; 2
, and H2 = 2 F1
23
1 5 3
2 , 4 ; 2 ; 2
See Wolfe [122] from which Hardin got the explicit form[55].
6.7.3
Section obsolete since I found forms for asymmetric stable distributions. Some commentary on
Hilbert transforms for symmetric stable distributions, given that for Z = | X |, dFz (z) = dFX (x)(1
sgn(x)), that type of thing.
Hilbert Transform for a function f (see Hlusel, [57], Pinelis [87]):
H( f ) =
1
p.v.
Z
f (x)
tx
dx
Here p.v. means principal value in the Cauchy sense, in other words
p.v.
E(| X |) =
= lim lim
a b 0 a
H((0)) =
p.v.
t
t
E(| X |) =
In our case:
E(| X |) =
Z b
1
p.v.
1
p.v.
Z a
b
Z
(z)
Z
(z)
z2
tz
dz|t=0
dz
2
e|t|
dt =
t2
143
D
d.1
I N P R O G R E S S D E R I VAT I O N S F O R L L N
A C R O S S FAT TA I L S
Recall from Chapter 3 that the quality of an estimator is tied to its replicability outside the
set in which it was derived: this is the basis of the law of large numbers which deals with the
limiting behavior of relative frequencies.
(Hint: we will need to look at the limit without the common route of Chebychevs inequality
which requires E[Xi2 ] < . Chebychevs inequality and similar ones eliminate the probabilities
of some tail events).
So long as there is a mean, observations should at some point reveal it.
The law of iterated logarithms For the thin-tailed conditions, we can see in Figure x how by
the law of iterated logarithm, for xi i.i.d. distributed with mean 0 and unitary variance, lim sup
n
xi
i=1
2n log log(n)
n
xi
i=1
2n log log(n)
Let us examine the speed of convergence of the average N1 1i N Xi . For a Gaussian distribution (m, ), the characteristic function for the convolution is:
(t/ N) N =
e
imt s2 t2
N
2N 2
N
,
c
c
which, derived twice at 0 yields (i)2 t
2 i t /. t 0 which produces the standard deviation
(1)
(n) =
so one can say that sum converges at a speed N.
N
imt s2 t2
N
2N 2
N
= eimt
Now eimt is the characteristic function of the degenerate distribution at m, with density
p(x) = (m x) where is the Dirac delta with values zero except at the point m x. (Note that
the strong law of large numbers implies that convergence takes place almost everywhere except
for a set of probability 0; for that the same result should be obtained for all values of t).
145
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
10
15
20
25
Maximum
observation
50
30
100
150
200
250
300
Figure D.1: The distribution (histogram) of the standard deviation of the sum of N=100 =13/6. The second
graph shows the entire span of realizations. If it appears to shows very little information in the middle, it
is because the plot is stretched to accommodate the extreme observation on the far right.
But things are far more complicated with power laws. Let us repeat the exercise for a Pareto
distribution with density L x 1 , x> L,
iLt N
N
N
(t/ N) = E+1
,
N
R
where E is the exponential integral E; En (z) = 1 ezt /tn dt.
At the limit:
lim
t
N
N
= e 1 iLt ,
1 L,
1 N
E+1 (0) N 2 E1 (0)E+1 (0) + E (0)2 N N E+1 (0) N + N 1
.
N
The trap After some tinkering, we get (N) = (1) , the same as with the Gaussian, which is
N
a trap. For we should be careful in interpreting (N), which will be very volatile since (1) is
already very volatile and does not reveal itself easily in realizations of the process. In fact, let p(.)
be the PDF of a Pareto distribution with mean m, variance v, minimum value L and exponent .
Infinite variance of variance The distribution of the variance, v can be obtained analytically:
intuitively its asymptotic tail is v 2 1 . Where g(.) is the probability density of the variance:
L
g(v) =
with support:[(L
2 L
2 L
1 +
2 v
1
v
)2 , ).
Cleaner: the expected mean deviation of the variance for a given will be =
146
1
v
R
(x m)2 v p(x)dx.
L
d.1.2
Let (t) be the characteristic function of the symmetric Student T with degrees of freedom.
After two-fold convolution of the average we get:
41 /2 |t| K
2
(t/2)2 =
2
2
|t| 2
2
1
2
(t/2)2 i
t x
dt,
+1 +2
x2
2 ; 2 ;
2 F1 (a, b; c; z)
k!
k=0
N
xi =x))
We can compare the twice-summed density to the initial one (with notation: p N (x)= P(i=1
p1, (x) =
+x2
B
+1
2
1
2, 2
147
0.9
0.8
0.7
1.5
2.0
2.5
3.0
The double effect of summing fat tailed random variables: The summation of random
variables performs two simultaneous actions, one, the thinning of the tails by the CLT
for a finite variance distribution (or convergence to some basin of attraction for infinite
variance classes); and the other, the lowering of the dispersion by the LLN. Both effects are
fast under thinner tails, and slow under fat tails. But there is a third effect: the dispersion of
observations for n=1 is itself much higher under fat tails. Fatter tails for power laws come
with higher expected mean deviation.
From there, we see that in the Cauchy case (=1) the sum conserves the density, so
p1,1 (x) = p2,1 (x) =
1
(1 + x 2 )
() =
and
21 12
2
2
1
lim () =
2
148
d.2
1
2 2 /4 | t | /2 K ( | t | )
2
The Characteristic function of the Student T with degrees of freedom, C(t) =
( 2 )
f n (x) =
!n Z
33/8
2
3
4
i tx
|t|
3n
4
r
K3
4
3
|t|
2
!n
dt
I tried all manner of expansions and reexpressions of the Bessel into other functions (Hypergeometric, Gamma) to no avail. One good news is that n = 2 works on Mathematica because the
Wolfram library has the square of a Bessel function. It would be great to get the solution for at
least n = 3.
C(t)
t n
n/2 e
2
t
5t n
K5
8 2
1
5
4
3
2
K3
4
7
4
K7
3
2t
7t
2
K1
2t n
n
K 9 3t
2
8
q
5
n
K5
2t
4
11t n
K 11
8
23t
n
(3t+ 3)
n /2 e
8
2
9
4
5
2
11
4
/2
t
N
21 2 /4 Nt K
n
2
35n/4
t3/2
149
Student T
Pareto
etsgn(t)
5| t |
55/16 |t|5/8 K 5
2
8
4
5
2(
8)
q
3 t
33 / 8 | t | 3 / 4 K 3
2| |
4
8
2( 34)
7| t |
77/16 |t|7/8 K 7
2
E2 (iLt)
5
4
3
2
7
4
2
9
4
5
2
11
4
d.3
5
4 E 94 (iLt)
3
2 E 52 (iLt)
7
(iLt)
4 E 11
4
23/4
( 78 )
2 | t | K1
6 23/4
2 |t|
3| t |9/8 K 9
( 18 )
55 / 8 | t | 5 / 4 K 5
27/8 ( 54 )
1111/16 |t|11/8 K 11
8
3| t |
2
5 t
2| |
2E3 (iLt)
9
(iLt)
4 E 13
4
5
2 E 72 (iLt)
11|t|
2
11
(iLt)
4 E 15
4
3/4 11
2 2
( )
8
3
t
|
|
e
3 |t| + 1
3E4 (iLt)
6 3
( x 2 + 3)
we have:
(t) = E[eitX ] = (1 +
3 | t |) e
3 |t|
1
2
Z +
(1 +
3 | t |)n e n
3 |t| itx
dt =
Z +
0
(1 +
3 t)n en
3t)n =
n
k ( 3t)k
k=0
n
150
k t
t e
3t
cos(tx) dt
n 1
6 3 n2 + s2
Sn =
18
n! 2 2 12 k +n
1
Tk+1 q
(n k)! n + s
s2
k=0
+1
n
n2
1 with thanks to Abe Nassen and Jack DAurizio on Math Stack Exchange.
151
Chapter Summary 7: The behavior of the sum of random variables allows us to get to the
asymptote and use handy asymptotic properties, that is, Platonic distributions. But the
problem is that in the real world we never get to the asymptote, we just get "close" Some
distributions get close quickly, others very slowly (even if they have finite variance). We
examine how fat tailedness worsens the process.
An intuition: how we converge mostly in the center of the distribution
We start with the Uniform Distribution, patently the easiest of all.
f (x) = {
1
HL
Lx H
0
elsewhere
where L = 0 and H =1
0.2
0.4
0.6
0.8
500
400
300
200
100
0.5
1.5
153
800
600
400
200
0.5
1.5
2.5
As we can see, we get more observations where the peak is higher. Now some math. By convoluting 2, 3, 4 times we can see the
progress and the decrease of mass in the tails:
f 2 (z 2 ) =
(
( f (z x ))( f x ) d x =
2 z2
1 < z2 < 2
z2
0 < z2 1
(7.1)
23
Z 3
(z 3 3)z 3 23
f 3 (z 3 ) =
( f 2 (z 3 2)) f ( x 2 ) d x 2 =
0
12 (z 3 3)(z 3 1)
1
2
2 (z 3 3)
0 < z3 1
1 < z3 < 2
z3 = 2
(7.2)
2 < z3 < 3
With N = 3 we square terms, and the familiar "bell" shape starts to emerge thanks to such
squaring.
f4 x =
Z 4
0
154
22
z4
4
41 z 24 + 4z 4
1 (z 4) 2
4
4
z4 = 3
z4 = 2
0 < z4 1
(7.3)
0.5
1.5
We can see how quickly, after one single addition, the net probabilistic weight is going to
be skewed to the center of the distribution, and the vector will weight future densities..
2.5
1.5
0.5
-0.5
0.5
1.5
2.5
0.7
0.6
0.5
0.4
0.3
0.2
0.1
1
155
0.4
0.3
0.2
0.1
N (0, 1)as n
n
D
Where is converges in distribution and N(0,1) is the Gaussian with mean 0 and unit
standard deviation.
Granted convergence in distribution is about the weakest form of convergence. Effectively
we are dealing with a double problem.
The first, as uncovered by Jaynes, corresponds to the abuses of measure theory: Some properties
that hold at infinity might not hold in all limiting processes .
There is a large difference between convergence a.s. (almost surely) and the weaker forms.
Jaynes 2003 (p.44):The danger is that the present measure theory notation presupposes the
infinite limit already accomplished, but contains no symbol indicating which limiting process
was used (...) Any attempt to go directly to the limit can result in nonsense.
We accord with him on this point along with his definition of probability as information
incompleteness, about which later.
The second problem is that we do not have a clean limiting process the process is itself
idealized.
Now how should we look at the Central Limit Theorem? Let us see how we arrive to it assuming independence.
The Kolmogorov-Lyapunov Approach and Convergence in the Body 2 The CLT works does
not fill-in uniformily, but in a Gaussian way indeed, disturbingly so. Simply, whatever your
1 Feller 1971, Vol. II
2 See Loeve for a presentation of the method of truncation used by Kolmogorov in the early days before Lyapunov started
using characteristic functions.
156
n
Xi
i=0
u Z =
u
n
Z2
2
Ru
=
Z2
e 2 dZ
, for u z u
dZ
inside the tunnel [-u,u] the odds of falling inside the tunnel itself,
and
Z u
Z ( N )d z +
Z
u
Z 0 ( N )d z
u
N
u
and u. So the speed of convergence to the Gaussian will depend on N u as can be seen in the
2a3
.
x4
will be 2Na
. Now the center, by the Kolmogorov version of the central limit theorem, will have
x4
a variance of Na2 in the center as well, hence
x2
e 2a2 N
P( x) =
2a N
Setting the point u where the crossover takes place,
157
Figure 7.1: Q-Q Plot of N Sums of variables distributed according to the Student T with 3 degrees of
freedom, N=50, compared to the Gaussian, rescaled
into standard deviations. We see on both sides a
higher incidence of tail events. 106 simulations
x2
2Na3
e 2aN
'
,
x4
2a N
u2
3 aNN
22a
,
u = 2a N
W
1
,
2N 1/4 (2)1/4
158
N
2000
4000
6000
8000
10 000
> Xi
P> X
for [NOT] 0 x
m
2
1 N log(N)
Generalizing for all exponents > 2 More generally, using the reasoning for a broader set and
getting the crossover for powelaws of all exponents:
( 2)e
2
x
2aN
2 aN
'
1
x2
Beta
1+
h
/2
1
2, 2
,]
2+
s
x
a ( + 1) N W()
( 2)
Where
(2)
=
7.1
1
+1
2 14 12
B( 2 , 21 )
2
+1
a ( + 1) N
The normalized cumulant of order n, n is the derivative of the log of the characteristic function
which we convolute N times divided by the second cumulant (i,e., second moment).
This exercise show us how fast an aggregate of N-summed variables become Gaussian, looking
at how quickly the 4th cumulant approaches 0. For instance the Poisson get there at a speed that
159
Distr.
PDF
Normal(, )
(x )2
e 22
e^-x
2
2
2
iz
z
N log e(1+e ) N log iz
N log eiz 2
th
th
th
10
Distr.
PDF
6 th
2
x 2
e 21
1
N
1
N 2 2
1
N 4 4
1
N 6 6
1
N 8 8
2
N
3!2
N2
5!4
N4
7!6
N6
9!8
N8
2
a b N
3!
a2 b2 N 2
5!
a4 b4 N 4
7!
a6 b6 N 6
9!
a8 b8 N 8
2
x 2
e 22
p
+ (1 p)
22
212 2
z 1
z2 2 2
N log pe 2 + (1 p)e 2
1
0
2
3(1 p)p(12 22 )
b a e b x a 1
(a)
Nconvoluted
log
Characteristic
2nd Cum
3 rd
4 th
Exponentl() (a, b)
e x
x!
Nconvoluted
Log
Characteristic
2 nd Cu- 1
mulant
3 rd
0
th
Poisson( )
N 2 ( p12 (1+p)22 )
(15(1+p)p(1+2p)(12 22 )3 )
( N 4 ( p12 (1+p)22 )5 )
StudentT(3)
StudentT(4)
5/2
6 3
1
12
2
x2 +4
( x2 +3)
N log
3 |z| + 1 N log 2 |z|2 K2 (2 |z|)
3 | z |)
1
"fuhgetaboudit"
1
TK
"fuhgetaboudit"
"fuhgetaboudit"
"fuhgetaboudit"
"fuhgetaboudit"
160
p 2 s
z0,
o
n
p
k
s (zH2r+s ) Ys,r (k1)k
2k2
k=3
+ 1
s
2 s! 2r+ 2
where k is the cumulant of order k. Yn,k ( x1 , . . . , xk+n+1 ) is the partial Bell polynomial given
by
Yn,k ( x1 , . . . , xk+n+1 )
n
m1 =0
n
n!
xs ms
1[nmn +m1 +2m2 +=nmn +m1 +m2 +=k]
m1 ! m n !
s!
mn =0
s =1
Xi = c n X + d n
i=1
Yi + an X
i=1
161
0.14
0.15
0.12
0.10
0.10
0.08
0.06
0.05
0.04
0.02
-20
10
-10
20 -30
-25
-20
-15
-10
-5
Figure 7.4: Disturbing the scale of the alpha stable and that of a more natural distribution, the gamma
distribution. The alpha stable does not increase in risks! (risks for us in Chapter x is defined in thickening
of the tails of the distribution). We will see later with convexification how it is rare to have an isolated
perturbation of distribution without an increase in risks.
Warning: perturbating the skewness of the Levy stable distribution by changing without
affecting the tail exponent is mean preserving, which we will see is unnatural: the transformation of random variables leads to effects on more than one characteristic of the distribution.
S(, , , )represents the stable distribution Stype with index of stability , skewness parameter
, location parameter , and scale parameter .
The Generalized Central Limit Theorem gives sequences an and bn such that the distribution
of the shifted and rescaled sum Zn = (in Xi an ) /bn of n i.i.d. random variates Xi whose
distribution function FX (x) has asymptotes 1 cx as x-> + and d( x) as x-> weakly
converges to the stable distribution S1 (, (c d)/(c + d), 0, 1):
Note: Chebyshevs Inequality and upper bound on deviations under finite variance. [To
ADD MARKOV BOUNDS CHEBYCHEV CHERNOV BOUNDS.]
Even when the variance is finite, the bound is rather far. Consider Chebyshevs inequality:
P(X > )
2
2
P(X > n)
1
,
n2
which effectively accommodate power laws but puts a bound on the probability distribution
of large deviations but still significant.
162
Gaussian
7. 102
ChebyshevUpperBound
9
4
3. 104
16
3. 106
25
1. 109
36
8. 1011
49
2. 1015
64
9. 1018
81
10
1. 1023
100
7.2
An illustration of the following point. The behavior of the maximum value as a percentage of a
sum is much slower than we think, and doesnt make much difference on whether it is a finite
variance, that is >2 or not. (See comments in Mandelbrot & Taleb, 2011)
(N) E ()
MaxSum
0.05
0.04
=1.8
0.03
0.02
0.01
=2.4
N
2000
7.3
4000
6000
8000
10 000
Limits of Sums
Paul Lvy [68], Gnedenko and Kolmogorov [51], Prokhorov [92], [91], Hoeffding[58], Petrov[86],
Blum[12].
163
164
7.4
1
2 2 /4 | t | /2 K ( | t | )
2
The Characteristic function of the Student T with degrees of freedom, C(t) =
( 2 )
33/8
3
f n (x) =
ei tx |t|3n/4 K 3
dt
|t|
8
4
2
2 43
I tried all manner of expansions and reexpressions of the Bessel into other functions (Hypergeometric, Gamma) to no avail. One good news is that n = 2 works on Mathematica because the
Wolfram library has the square of a Bessel function. It would be great to get the solution for at
least n = 3.
StudentT
Pareto(1, )
1
5
4
3
2
7
4
2
9
4
5
2
11
4
etsgn(t)
E2 (iLt)
5| t |
55/16 |t|5/8 K 5
2
8
4
5
2(
8)
q
3/4
3 t
3
/
8
3 |t| K 3
|
|
2
4
8
2( 34)
7| t |
77/16 |t|7/8 K 7
2
5
4 E 94 (iLt)
3
2 E 52 (iLt)
7
(iLt)
4 E 11
4
23/4
( 78 )
2 | t | K1
23/4 8 3|t|9/8 K
(
2 |t|
9
8
1
8 q
3| t |
2
2E3 (iLt)
9
(iLt)
4 E 13
4
55 / 8 | t | 5 / 4 K 5
27/8 ( 54 )
1111/16 |t|11/8 K 11
8
11
8
5 t
2| |
11|t|
2
5
2 E 72 (iLt)
3/4
2 2
( )
e 3| t |
3 |t| + 1
11
(iLt)
4 E 15
4
3E4 (iLt)
165
0.3
0.2
0.1
-30
-20
10
-10
20
30
-30
-20
0.02
20
30
-30
-10
0.05
0.04
10
20
30
-30
-20
-10
35
0.04
0.03
0.03
0.02
0.02
0.02
0.01
0.01
0.01
10
20
30
-30
-20
-10
20
30
10
20
30
40
0.04
-10
10
0.01
0.01
30
30
0.02
0.02
20
20
0.03
0.03
10
10
25
0.05
0.04
0.03
10
20
30
166
-20
20
0.04
-20
0.04
0.06
30
-30
0.06
10
-10
15
0.07
0.06
0.05
0.04
0.03
0.02
0.01
-30 -20 -10
10
0.08
0.12
0.10
0.08
0.06
0.04
0.02
W H E R E S TA N DA R D D I V E R S I F I C AT I O N
FA I L S
Risk
Markowitz
Real World
Underestimation
of risk
20
Overerestimation
of diversification
40
60
80
100
Number of Assets
Figure E.1: The "diversification effect": difference between promised and delivered. Markowitz Mean
Variance based portfolio construction will stand probably as one of the most empirically invalid theory
ever used in modern times.
This is an analog of the problem with slowness of the law of large number: how a portfolio
can track a general index (speed of convergence) and how high can true volatility be compared
to the observed one (the base line).
Model Structure under Markowitz Mean Variance Historically, Modern Portfolio Theory (MPT),
as a normative theory of risk bearing, has made the central assumption that, for a set return, the
investor rationally minimizes the variance of his exposure, defined as the mean square variability (or, which was proven by MPT to be equivalent to maximizing the mean conditional on a set
variance). The standard models, Markowitz (1952, 1959) [75] [76], and the extensions such as
Treynor(1965) [116] and Sharpe (1966)[102], all base their criterion on minimum variance. [See
Constantinides and Malliaris (1995)[20], Elton and Gruber (1997)[30] for a survey, Huang and
Litzenberger (1988)[59] for an analytical discussion. Distributional Condition: The agent is
supposed to have full grasp of the probability distribution with all the joint returns known and
deemed to be Gaussian. Further, no error rate is allowed. Utility Condition: The agent is supposed to have a utility function that allows optimization via minimum variance, as the agent
cares only about the first two moments of the distribution. So, conveniently, under quadratic
utility of wealth, U(W) aW bW 2 , where W is wealth, a random variable, the expected wealth
167
168
e
v(x) =
(+log(1 x))2
22
2(x 1)
With a fatter tailed distribution, such as the a standard powerlaw used in finance (Gabaix,
2008,[47]), where is the tail exponent,
! +1
2
x
v(x) =
(log(1 x)1)2
+
a2
(a ax)B
1
2, 2
vHxL
0.6
0.5
0.4
0.3
0.2
0.1
x
-10
-8
-6
-4
-2
169
FAT TA I L S A N D R A N D O M M AT R I C E S
[The equivalent of fat tails for matrices. This will be completed, but consider for now that the
4th moment reaching Gaussian levels (i.e. 3) in the chapter is equivalent to eigenvalues reaching
Wigners semicircle. ]
171
0.0030
0.0025
0.0020
0.0010
0.0005
-100
100
200
p=10 -4 a=9998
0.006
0.005
0.004
0.003
0.002
0.001
-400
-200
200
400
0.0004
0.0003
Figure F.3: Student T
0.0002
0.0001
-20 000
172
20 000
40 000
3
2
5. 10 -6
4. 10 -6
3. 10 -6
Figure F.4: Cauchy
2. 10 -6
1. 10 -6
-4 10 7 -2 10 7
2 10 7 4 10 7 6 10 7
173
S O M E M I S U S E S O F S TAT I S T I C S I N
SOCIAL SCIENCE
Chapter Summary 8: We apply the results of the previous chapter on the slowness of
the LLN and list misapplication of statistics in social science, almost all of them linked
to misinterpretation of the effects of fat-tailedness (and often from lack of awareness of
fat tails), and how by attribute substitution researchers can substitute one measure for
another. Why for example, because of chronic small-sample effects, the 80/20 is milder
in-sample (less fat-tailed) than in reality and why regression rarely works.
8.1
Recall from the Introduction that the best way to figure out if someone is using an erroneous
statistical technique is to use such technique on a dataset for which you have the answer. The best
way to know the exact properties is to generate it by Monte Carlo. So the technique throughout
the chapter is to generate fat-tailed data, the properties of which we know with precision, and
check how such standard and mechanistic methods detect the true properties, then show the
wedge between observed and true properties.
Also recall from Chapter 6 (D.1) that fat tails make it harder for someone to detect the true
properties; for this we need a much, much larger dataset, more rigorous ranking techniques
allowing inference in one direction not another ( Chapter 4), etc. Hence this chapter is a direct
application of the results and rules of Chapter 4.
8.2
attribute substitution
Attribute substitution occurs when an individual has to make a judgment (of a target attribute)
that is complicated complex, and instead substitutes a more easily calculated one. There have
been many papers (Kahneman and Tversky [118] , Hoggarth and Soyer, [104] and comment
[107]) showing how statistical researchers overinterpret their own findings, as simplication leads
to the fooled by randomness effect.
Dan Goldstein and this author (Goldstein and Taleb [53]) showed how professional researchers and practitioners substitute norms in
q the evaluation of higher order properties of time
series, mistaking k x k1 for k x k2 (or n1 | x | for nx ). The common result is underestimating the
randomness of the estimator M, in other words read too much
into it (and, what is worse, un2
derestimation of the tails, since, as we saw in 3.4, the ratio | x| increases with "fat-tailedness"
to become infinite under tail exponents 2). Standard deviation is ususally explained and
interpreted as mean deviation. Simply, people find it easier to imagine that a variation of, say,
(-5,+10,-4,-3, 5, 8) in temperature over successive day needs to be mentally estimated by squaring
the numbers, averaging them, then taking square roots. Instead they just average the absolutes.
x2
175
But, what is key, they tend to do so while convincing themselves that they are using standard
deviations.
There is worse. Mindless application of statistical techniques, without knowledge of the
conditional nature of the claims are widespread. But mistakes are often elementary, like lectures
by parrots repeating "N of 1" or "p", or "do you have evidence of?", etc. Many social scientists
need to have a clear idea of the difference between science and journalism, or the one between
rigorous empiricism and anecdotal statements. Science is not about making claims about a
sample, but using a sample to make general claims and discuss properties that apply outside
the sample.
Take M (short for MTX (A, f )) the estimator we saw above from the realizations (a sample path)
for some process, and M* the "true" mean that would emanate from knowledge of the generating
process for such variable. When someone announces: "The crime rate in NYC dropped between
2000 and 2010", the claim is limited M the observed mean, not M the true mean, hence the
claim can be deemed merely journalistic, not scientific, and journalists are there to report "facts"
not theories. No scientific and causal statement should be made from M on "why violence has
dropped" unless one establishes a link to M* the true mean. M cannot be deemed "evidence"
by itself. Working with M alone cannot be called "empiricism".
What we just saw is at the foundation of statistics (and, it looks like, science). Bayesians
disagree on how M converges to M*, etc., never on this point. From his statements in a dispute
with this author concerning his claims about the stability of modern times based on the mean
casualy in the past (Pinker [88]), Pinker seems to be aware that M may have dropped over time
(which is a straight equality) and sort of perhaps we might not be able to make claims on M*
which might not have really been dropping.
176
600
500
400
300
200
100
100
200
300
400
500
600
700
In some areas not involving time series, the differnce between M and M* is negligible.
So I rapidly jot down a few rules before showing proofs and derivations (limiting M to the
arithmetic mean, that is, M= MTX ((, ), x)).
Note again that E is the expectation operator under "real-world" probability measure P.
8.3
From the derivations in D.1, E[| M- M*|] increases in with fat-tailedness (the mean deviation
of M* seen from the realizations in different samples of the same process). In other words,
fat tails tend to mask the distributional properties. This is the immediate result of the
problem of convergence by the law of large numbers.
On the difference between the initial (generator) and the "recovered" distribution
(Explanation of the method of generating data from a known distribution and comparing realized outcomes to expected ones)
Case Study: "Long Peace" Type Claims On The Stability of the Future Based on Past Data
When the generating process is power law with low exponent, plenty of confusion can take
place.
For instance, Pinker [88] claims that the generating process has a tail exponent 1.16 but made
the mistake of drawing quantitative conclusions from it about the mean from M and built theories
about drop in the risk of violence that is contradicted by the data he was showing, since fat tails
plus negative skewness/asymmetry= hidden and underestimated risks of blowup. His study
is also missing the Casanova problem (next point) but let us focus on the error of being fooled
by the mean of fat-tailed data.
Figures 8.2 and 8.3 show the realizations of two subsamples, one before, and the other after
the turkey problem, illustrating the inability of a set to naively deliver true probabilities through
calm periods.
177
Casualties H000L
4000
3000
2000
1000
TimeHYearsL
CasualtiesH000L
1.4 106
Figure 8.3: The Turkey Surprise: Now 200 years, the second 100 years dwarf the first;
these are realizations of the exact same process, seen with a
longer window and at a different scale.
1.2 106
1.0 106
800 000
600 000
400 000
200 000
TimeHYearsL
178
800
600
400
200
M1
200
400
600
800
M2
12 000
10 000
8000
6000
4000
2000
M1
5000
10 000
15 000
20 000
25 000
30 000
The next simulations shows M1, the mean of casualties over the first 100 years across 104 sample
paths, and M2 the mean of casualties over the next 100 years.
So clearly it is a lunacy to try to read much into the mean of a power law with 1.15 exponent
(and this is the mild case, where we know the exponent is 1.15. Typically we have an error rate,
and the metaprobability discussion in Chapter x will show the exponent to be likely to be lower
because of the possibility of error).
Claims Made From Power Laws
The Cederman graph, Figure 8.7 shows exactly how not to make claims upon observing power
laws.
8.4
Next we will see how when one hears about the Paretan 80/20 "rule" (or, worse, "principle"), it
is likely to underestimate the fat tails effect outside some narrow domains. It can be more like
95/20 or even 99.9999/.0001, or eventually 100/e. Almost all economic reports applying power
179
1.4
1.2
1.0
0.8
M1
1.0
1.5
2.0
Figure 8.7: Cederman 2003, used by Pinker [88] . I wonder if I am dreaming or if the exponent is really
= .41. Chapters x and x show why such inference is centrally flawed, since low exponents do not allow claims
on mean of the variableexcept to say that it is very, very high and not observable in finite samples. Also, in
addition to wrong conclusions from the data, take for now that the regression fits the small deviations, not
the large ones, and that the author overestimates our ability to figure out the asymptotic slope.
180
Where does this 80/20 business come from? Assume the power law tail exponent, and an
exceedant probability PX > x = xmin x , x (xmin , ). Simply, the top p of the population gets
1
S = p of the share of the total pie.
=
log(p)
log(p) log(S)
which means that the exponent will be 1.161 for the 80/20 distribution.
Note that as gets close to 1 the contribution explodes as it becomes close to infinite mean.
x 1 , x x
Derivation: Start with the standard density f (x) = xmin
min .
1) The Share attributed above K, K xmin , becomes
R
K x f (x) dx
R
= K 1
x
f
(x)
dx
x
min
Z
K
f (x)dx = K
of the result
Why the 80/20 Will Be Generally an Error: The Problem of In-Sample Calibration
Vilfredo Pareto figured out that 20% of the land in Italy was owned by 80% of the people, and
the reverse. He later observed that 20 percent of the peapods in his garden yielded 80 percent
of the peas that were harvested. He might have been right about the peas; but most certainly
wrong about the land.
For fitting in-sample frequencies for a power law does not yield the proper "true" ratio since
the sample is likely to be insufficient. One should fit a powerlaw using extrapolative, not interpolative techniques, such as methods based on Log-Log plotting or regressions. These latter
methods are more informational, though with a few caveats as they can also suffer from sample
insufficiency.
Data with infinite mean, 1, will masquerade as finite variance in sample and show about
80% contribution to the top 20% quantile. In fact you are expected to witness in finite samples a
lower contribution of the top 20%/
n , ordered x x
Let us see: Figure 8.8. Generate m samples of =1 data X j =(xi,j )i=1
i,j
i 1,j , and
181
0.03
0.02
0.01
0.5
0.6
0.7
0.8
0.9
1.0
Z 15
Figure 8.8: The difference betwen the generated (ex ante) and recovered (ex post) processes; = 20/100,
N = 107 . Even when it should be 100/.0001, we tend to watch an average of 75/20
8.5
E(M0 M ) increases under the presence of an absorbing barrier for the process. This is the
Casanova effect, or fallacy of silent evidence see The Black Swan, Chapter 8. ( Fallacy of silent
evidence: Looking at history, we do not see the full story, only the rosier parts of the process, in
the Glossary)
History is a single sample path we can model as a Brownian motion, or something similar
with fat tails (say Levy flights). What we observe is one path among many "counterfactuals", or
alternative histories. Let us call each one a "sample path", a succession of discretely observed
states of the system between the initial state S0 and ST the present state.
Arithmetic process: We can model it as S(t) = S(t t) + Zt where Zt is noise drawn from
any distribution.
Wt
Geometric process: We can model it as S(t) = S(t t)e
S(t t)et+s tZt but Wt
typically
can be noise drawn from any distribution. Typically, log S(tS(t)
it) is treated as Gaussian, but we
can use fatter tails. The convenience of the
Gaussian is stochastic calculus and the ability to skip
steps in the process, as S(t)=S(t-t)et+s tWt , with Wt N(0,1), works for all t, even allowing
for a single period to summarize the total.
The Black Swan made the statement that history is more rosy than the "true" history, that is, the
mean of the ensemble of all sample path.
Take an absorbing barrier H as a level that, when reached, leads to extinction, defined as
becoming unobservable or unobserved at period T.
When you observe history of a family of processes subjected to an absorbing barrier, i.e., you
see the winners not the losers, there are biases. If the survival of the entity depends upon
not hitting the barrier, then one cannot compute the probabilities along a certain sample path,
without adjusting.
182
250
200
150
100
Barrier H
50
Time
200
400
600
800
1000
Begin The "true" distribution is the one for all sample paths, the "observed" distribution is the
T .
one of the succession of points (Sit )i=1
Bias in the measurement of the mean In the presence of an absorbing barrier H "below", that
is, lower than S0 , the "observed mean" > "true mean"
Bias in the measurement of the volatility The "observed" variance (or mean deviation) 6 "true"
variance
The first two results are well known (see Brown, Goetzman and Ross (1995)). What I will set
to prove here is that fat-tailedness increases the bias.
First, let us pull out the "true" distribution using the reflection principle.
Thus if the barrier is H and we start at S0 then we have two distributions, one f(S), the other
f(S-2( S0 -H))
By the reflection principle, the "observed" distribution p(S) becomes:
(
p(S) =
f (S) f (S 2 (S0 H ))
if S > H
if S < H
183
8.6
E[ M- M*] increases (decreases) with negative (positive) skeweness of the true underying variable.
Some classes of payoff (those affected by Turkey problems) show better performance than
"true" mean. Others (entrepreneurship) are plagued with in-sample underestimation of the
mean. A naive measure of a sample mean, even without absorbing barrier, yields a higher
oberved mean than "true" mean when the distribution is skewed to the left, and lower when the
skewness is to the right.
This can be shown analytically, but a simulation works well.
To see how a distribution masks its mean because of sample insufficiency, take a skewed
distribution with fat tails, say the standard Pareto Distribution we saw earlier.
both N and T. Next we calculate the median value of Tj=1 Mj T , such that P> = 21 where, to
repeat, M* is the theoretical mean we expect from the generating distribution.
Entrepreneurship is penalized by right tail insufficiency making performance look worse than
it is. Figures 0.1 and 0.2 can be seen in a symmetrical way, producing the exact opposite effect
of negative skewness.
184
8.6 left (right) tail sample insufficiency under negative (positive) skewness
Probability
Outcomes
-140
-120
-100
-80
-60
-40
-20
0.95
0.90
j
MT
in simula-
(106
tions
Monte Carlo runs). We can observe the underestimation of the mean of a
skewed power law distribution as exponent gets lower. Note that lower values of
imply fatter tails.
0.85
0.80
0.75
1.5
2.0
2.5
185
8.7
The Power of Extreme Deviations: Under fat tails, large deviations from the mean are vastly
more informational than small ones. They are not "anecdotal". (The last two properties corresponds to the black swan problem, inherently asymmetric).
We saw the point earlier (with the masquerade problem) in ??.??. The gist is as follows, worth
repeating and applying to this context.
A thin-tailed distribution is less likely to deliver a single large deviation than a fat tailed
distribution a series of long calm periods. Now add negative skewness to the issue, which
makes large deviations negative and small deviations positive, and a large negative deviation,
under skewness, becomes extremely informational.
Mixing the arguments of ??.?? and ??.?? we get:
Asymmetry in Inference: Under both negative [positive] skewness and fat tails, negative
[positive] deviations from the mean are more informational than positive [negative] deviations.
8.8
Probing the limits of a standardized method by arbitrage. We can easily arbitrage a mechanistic
method of analysis by generating data, the properties of which are known by us, which we call
"true" properties, and comparing these "true" properties to the properties revealed by analyses,
as well as the confidence of the analysis about its own results in the form of "p-values" or other
masquerades.
This is no different from generating random noise and asking the "specialist" for an analysis
of the charts, in order to test his knowledge, and, even more importantly, asking him to give
us a probability of his analysis being wrong. Likewise, this is equivalent to providing a literary
commentator with randomly generated giberish and asking him to provide comments. In this
section we apply the technique to regression analyses, a great subject of abuse by the social
scientists, particularly when ignoring the effects of fat tails.
In short, we saw the effect of fat tails on higher moments. We will start with 1) an extreme
case of infinite mean (in which we know that the conventional regression analyses break down),
then generalize to 2) situations with finite mean (but finite variance), then 3) finite variance but
infinite higher moments. Note that except for case 3, these results are "sort of" standard in the
econometrics literature, except that they are ignored away through tweaking of the assumptions.
Fooled by =1 Assume the simplest possible regression model, as follows. Let yi = 0 + 1 xi +
s zi , with Y=(yi )1<in the set of n dependent variables and X= (xi )1<in , the independent one;
Y, X e R, i e N. The errors zi are independent but drawn from a standard Cauchy (symmetric,
with tail exponent =1), multiplied by the amplitude or scale s; we will vary s across the
thought experiment (recall that in the absence and variance and mean deviation we rely on s as
a measure of dispersion). Since all moments are infinite, E[zin ] = for all n1, we know ex ante
that the noise is such that the "errors" or residuals" have infinite means and variances but the
problem is that in finite samples the property doesnt show. The sum of squares will be finite.
The next figure shows the effect of a very expected large deviation, as can be expected from a
Cauchy jump.
186
40
60
80
100
-1000
-2000
The big deviation
-3000
-4000
= 1; s = 5
Pr
0.4
0.3
0.2
0.1
0.0
0.2
0.4
0.6
0.8
1.0
R2
Next we generate T simulations (indexed by j) of n pairs (yi , xi )1<in for increasing values of
and multiplied z by the scaling constant
x, thanks to Cauchy distributed variables variable zi,j
i,j
s, leaving us with a sequence
n T
0 + 1 xi + szi,j
.
i=1 j=1
Using standard regression techniques of estimation we "regress" and obtain the standard equaest
est is the estimated Y, and E a vector of unexplained residuals
tion Y est = est
0 + X 1 , where Y
n T
T
est est x
E ei,j yest
. We thus obtain T simulated values of j j=1 ,
i
j
0
1
i,j
i=1 j=1
where j 1-
n
ei,j 2
i=1
,
n
i=1 (yi,j y j )2
n
the R-square for a sample run j, where y j = n1 i=1
yi,j , in other words 1-
( squared residuals) / (squared variations). We examine the distribution of the different realizations of .
Arbitraging metrics For a sample run which, typically, will not have a large deviation,
R-squared: 0.994813 (When the "true" R-squared would be 0)
The P-values are monstrously misleading.
187
=1; s=.5
Pr
0.15
0.10
0.05
0.2
0.4
0.6
0.8
1.0
R2
yHxL
15
10
Figure 8.17: We can fit different regressions to the same story (which is no story).
A regression that tries to accommodate the
large deviation.
x
20
40
60
80
100
-5
-10
yHxL
15
10
x
20
-5
188
40
60
80
100
0.08
0.06
0.04
0.02
0.00
0.2
0.4
0.6
0.8
1.0
R2
Estimate
Std Error
T-Statistic
P-Value
4.99
0.417
11.976
7.8 1033
0.10
0.00007224
1384.68
9.3 1011426
8.9
A pervasive attribute substitution: Where X and Y are two random variables, the properties of
X-Y, say the variance, probabilities, and higher order attributes are markedly different from the
difference in properties. So E ( X Y ) = E(X) E(Y) but of course, Var(X Y) 6= Var(X)
Var(Y), etc. for higher norms. It means that P-values are different, and of course the coefficient
of variation ("Sharpe"). Where is the Standard deviation of the variable (or sample):
E(X Y) E(X) E(Y))
6=
(X Y)
(X)
(Y)
In Fooled by Randomness (2001):
A far more acute problem relates to the outperformance, or the comparison, between
two or more persons or entities. While we are certainly fooled by randomness when
it comes to a single times series, the foolishness is compounded when it comes to the
comparison between, say, two people, or a person and a benchmark. Why? Because
both are random. Let us do the following simple thought experiment. Take two individuals, say, a person and his brother-in-law, launched through life. Assume equal
odds for each of good and bad luck. Outcomes: lucky-lucky (no difference between
189
8.10
See discussion in Chapter ??. A lot of nonsense in discussions of rationality facing "dread
risk" (such as terrorism or nuclear events) based on wrong probabilistic structures, such as
comparisons of fatalities from falls from ladders to death from terrorism. The probability of falls
from ladder doubling is 1 1020 . Terrorism is fat-tailed: similar claims cannot be made.
A lot of unrigorous claims like "long shot bias" is also discussed there.
8.11
190
The implication is that those tools used in economics that are based on squaring variables
(more technically, the Euclidian, or L2 norm), such as standard deviation, variance, correlation,
regression, the kind of stuff you find in textbooks, are not valid scientifically(except in some
rare cases where the variable is bounded). The so-called "p values" you find in studies have
no meaning with economic and financial variables. Even the more sophisticated techniques of
stochastic calculus used in mathematical finance do not work in economics except in selected
pockets.
The results of most papers in economics based on these standard statistical methods are thus
not expected to replicate, and they effectively dont. Further, these tools invite foolish risk taking.
Neither do alternative techniques yield reliable measures of rare events, except that we can tell
if a remote event is underpriced, without assigning an exact value.
P(t)
P(tit)
, take the measure MtX (, ), X 4 of
1
MtX (, ), X 4
n
Xt4it
i=0
Q(n)
n
Max Xt4ti )i=0
n
Xt4ti
i=0
For a Gaussian (i.e., the distribution of the square of a Chi-square distributed variable) show
Q 104 the maximum contribution should be around .008 .0028. Visibly we can see that the
distribution 4th moment has the property
!
n
4
4
P X > max(xi )i2n P X > xi
i=1
Recall that, naively, the fourth moment expresses the stability of the second moment. And the
second moment expresses the stability of the measure across samples.
191
0.8
0.6
0.2
0.0
Security
Max Q
Years.
Silver
0.94
46.
SP500
0.79
56.
CrudeOil
0.79
26.
Short Sterling
0.75
17.
Heating Oil
0.74
31.
Nikkei
0.72
23.
FTSE
0.54
25.
JGB
0.48
24.
Eurodollar Depo 1M
0.31
19.
Sugar #11
0.3
48.
Yen
0.27
38.
Bovespa
0.27
16.
Eurodollar Depo 3M
0.25
28.
CT
0.25
48.
DAX
0.2
18.
Note that taking the snapshot at a different period would show extremes coming from other
variables while these variables showing high maximma for the kurtosis, would drop, a mere
result of the instability of the measure across series and time. Description of the dataset:
All tradable macro markets data available as of August 2008, with "tradable" meaning actual
closing prices corresponding to transactions (stemming from markets not bureaucratic evaluations, includes interest rates, currencies, equity indices).
Performance of Standard NonParametric Risk Estimators, f(x)= x or |x| (Norm L1), A =(-, K]
Does the past resemble the future in the tails? The following tests are nonparametric, that is
entirely based on empirical probability distributions.
So far we stayed in dimension 1. When we look at higher dimensional properties, such as
covariance matrices, things get worse. We will return to the point with the treatment of model
error in mean-variance optimization.
192
30
20
10
Monthly Vol
0.8
0.6
0.4
0.2
Vol of Vol
0.20
0.15
0.10
0.05
0.00
193
M@t+1D
0.004
0.003
0.002
0.001
M@tD
0.001
0.002
0.003
0.004
M@t+1D
0.030
0.025
0.020
0.015
0.010
0.005
M@tD
0.005
0.010
0.015
0.020
0.025
0.030
M@t+1D
0.0004
0.0003
0.0002
0.0001
M@tD
0.0001
194
0.0002
0.0003
0.0004
0.0005
Figure 8.27: Correlations are also problematic, which flows from the instability of single variances and the
effect of multiplication of the values of random variables.
195
8.12
The four aspects of what we will call the nonreplicability issue, particularly for mesures that are
in the tails. These are briefly presented here and developed more technically throughout the
book:
a- Definition of statistical rigor (or Pinker Problem). The idea that an estimator is not about
fitness to past data, but related to how it can capture future realizations of a process seems
absent from the discourse. Much of econometrics/risk management methods do not meet this
simple point and the rigor required by orthodox, basic statistical theory.
b- Statistical argument on the limit of knowledge of tail events. Problems of replicability
are acute for tail events. Tail events are impossible to price owing to the limitations from the size
of the sample. Naively rare events have little data hence what estimator we may have is noisier.
c- Mathematical argument about statistical decidability. No probability without metaprobability. Metadistributions matter more with tail events, and with fat-tailed distributions.
1. The soft problem: we accept the probability distribution, but the imprecision in the calibration (or parameter errors) percolates in the tails.
2. The hard problem (Taleb and Pilpel, 2001, Taleb and Douady, 2009): We need to specify
an a priori probability distribution from which we depend, or alternatively, propose a
metadistribution with compact support.
3. Both problems are bridged in that a nested stochastization of standard deviation (or the
scale of the parameters) for a Gaussian turn a thin-tailed distribution into a power law (and
stochastization that includes the mean turns it into a jump-diffusion or mixed-Poisson).
d- Economic arguments: The Friedman-Phelps and Lucas critiques, Goodharts law. Acting on statistical information (a metric, a response) changes the statistical properties of
some processes.
196
8.13 conclusion
8.13
conclusion
This chapter introduced the problem of "surprises" from the past of time series, and the invalidity
of a certain class of estimators that seem to only work in-sample. Before examining more deeply
the mathematical properties of fat-tails, let us look at some practical aspects.
197
O N T H E I N S TA B I L I T Y O F E C O N O M E T R I C
DATA
Table 15: Fourth noncentral moment at daily, 10-day, and 66-day windows for the random variables
K(1)
K(10)
K(66)
Max
Quartic
Years
6.3
3.8
2.9
0.12
22.
Australia TB 10y
7.5
6.2
3.5
0.08
25.
Australia TB 3y
7.5
5.4
4.2
0.06
21.
BeanOil
5.5
7.0
4.9
0.11
47.
Bonds 30Y
5.6
4.7
3.9
0.02
32.
Bovespa
24.9
5.0
2.3
0.27
16.
British Pound/USD
6.9
7.4
5.3
0.05
38.
CAC40
6.5
4.7
3.6
0.05
20.
Canadian Dollar
7.4
4.1
3.9
0.06
38.
Cocoa NY
4.9
4.0
5.2
0.04
47.
Coffee NY
10.7
5.2
5.3
0.13
37.
Copper
6.4
5.5
4.5
0.05
48.
Corn
9.4
8.0
5.0
0.18
49.
Crude Oil
29.0
4.7
5.1
0.79
26.
CT
7.8
4.8
3.7
0.25
48.
DAX
8.0
6.5
3.7
0.20
18.
Euro Bund
4.9
3.2
3.3
0.06
18.
Euro Currency/DEM
previously
5.5
3.8
2.8
0.06
38.
Eurodollar Depo 1M
41.5
28.0
6.0
0.31
19.
Eurodollar Depo 3M
21.1
8.1
7.0
0.25
28.
FTSE
15.2
27.4
6.5
0.54
25.
Gold
11.9
14.5
16.6
0.04
35.
Heating Oil
20.0
4.1
4.4
0.74
31.
Hogs
4.5
4.6
4.8
0.05
43.
40.5
6.2
4.2
0.19
16.
17.2
16.9
4.3
0.48
24.
Live Cattle
4.2
4.9
5.6
0.04
44.
Nasdaq Index
11.4
9.3
5.0
0.13
21.
Australian
lar/USD
Dol-
199
200
K(1)
K(10)
K(66)
Max
Quartic
Years
Natural Gas
6.0
3.9
3.8
0.06
19.
Nikkei
52.6
4.0
2.9
0.72
23.
Notes 5Y
5.1
3.2
2.5
0.06
21.
Russia RTSI
13.3
6.0
7.3
0.13
17.
Short Sterling
851.8
93.0
3.0
0.75
17.
Silver
160.3
22.6
10.2
0.94
46.
Smallcap
6.1
5.7
6.8
0.06
17.
SoyBeans
7.1
8.8
6.7
0.17
47.
SoyMeal
8.9
9.8
8.5
0.09
48.
Sp500
38.2
7.7
5.1
0.79
56.
Sugar #11
9.4
6.4
3.8
0.30
48.
SwissFranc
5.1
3.8
2.6
0.05
38.
TY10Y Notes
5.9
5.5
4.9
0.10
27.
Wheat
5.6
6.0
6.9
0.02
49.
Yen/USD
9.7
6.1
2.5
0.27
38.
FAT TA I L S F R O M R E C U R S I V E
U N C E R TA I N T Y
Second Version. An earlier version was presented at Benoit Mandelbrots Scientific Memorial,
New Haven, April 11, 2011, under the title: The Future Will Be More Fat Tailed Than The Past
Chapter Summary 9: Error about Errors. Probabilistic representations require the inclusion of model (or representation) error (a probabilistic statement has to have an error
rate), and, in the event of such treatment, one also needs to include second, third and
higher order errors (about the methods used to compute the errors) and by a regress argument, to take the idea to its logical limit, one should be continuously reapplying the
thinking all the way to its limit unless when one has a reason to stop, as a declared a
priori that escapes quantitative and statistical method. We show how power laws emerge
from nested errors on errors of the standard deviation for a Gaussian distribution. We
also show under which regime regressed errors lead to non-power law fat-tailed distributions.
9.1
layering uncertainty
With the Central Limit Theorem: we start with a distribution and, under some conditions, end
with a Gaussian. The opposite is more likely to be true. We start with a Gaussian and under
error rates we end with a fat-tailed distribution.
Unlike with the Bayesian compounding the:
1. Numbers of recursions
and
2. Structure of the error of the error (declining, flat, multiplicative or additive)
determine the final moments and the type of distribution.
Note that historically, derivations of power laws have been statistical (cumulative advantage,
preferential attachment, winner-take-all effects, criticality), and the properties derived by Yule,
Mandelbrot, Zipf, Simon, Bak, and others result from structural conditions or breaking the
independence assumptions in the sums of random variables allowing for the application of
the central limit theorem. This work is entirely epistemic, based on the projection of standard
philosophical doubts into the future, in addition to regress arguments.
Missing the point
Savage, in his Foundation of Statistics [99]:
Estimate of the accuracy of estimates:
The doctrine is often expressed that a point estimate is of little, or no, value unless accompanied
201
H1 - a1 L Ha2 + 1L Ha3 + 1L
H1 - a1 L Ha2 + 1L
H1 - a1 L Ha2 + 1L H1 - a3 L
H1 - a1 L
H1 - a1 L H1 - a2 L Ha3 + 1L
H1 - a1 L H1 - a2 L
H1 - a1 L H1 - a2 L H1 - a3 L
Figure 9.1: Three levels of multiplicative relative error rates for the standard deviation , with (1 an ) the
relative error on an1
by an estimate of its own accuracy. This doctrine, which for the moment I will call the doctrine
of accuraty estimation, may be a little old-fashioned, but 1 think some critical discussion of it here
is in order for two reasons. In the first place, the doctrine is still widely considered to contain
more than a grain of truth. For example, many readers will think it strange, and even remiss,
that I have written a long chapter (Chapter 15) on estimation without even sugesting that an
estimate should be accompanied by an estimate of its accuracy. In the second place, it seems to
me that the concept of interval estimation, which is the subject of the next section, has largely
evolved from the doctrine of accuracy estimation and that discussion of the doctrine will, for
some, pave the way for discuasion of interval estimation. The doctrine of accuracy estimation
is vague, even by the standards of the verbalistic tradition, for it does not say what should be
taken as a measure of accuracy, that is, what an estimate of accuracy ahould estimate.
So we got diverted into the wrong direction for all these years as it did not hit Savage that we
should perhaps see what effect would the estimation error have via structured perturbation.1
1 I thank Dane Rook for the discussion.
202
So as we will see in this chapter, we can actually see the effect with a great deal of clarity.
Layering Uncertainties
Take a standard probability distribution, say the Gaussian. The measure of dispersion, here ,
is estimated, and we need to attach some measure of dispersion around it. The uncertainty
about the rate of uncertainty, so to speak, or higher order parameter, similar to what called
the volatility of volatility in the lingo of option operators here it would be uncertainty
rate about the uncertainty rate. And there is no reason to stop there: we can keep nesting
these uncertainties into higher orders, with the uncertainty rate of the uncertainty rate of the
uncertainty rate, and so forth. There is no reason to have certainty anywhere in the process.
Main Results
Note that unless one stops the branching at an early stage, all the results raise small probabilities
(in relation to their remoteness; the more remote the event, the worse the relative effect).
1. Under the first regime of proportional constant (or increasing) recursive layers of uncertainty about rates of uncertainty expressed as standard deviation, the distribution converges to a power law with infinite variance, even when one starts with a standard Gaussian.
2. Under the same first regime, expressing uncertainty about uncertainty in terms of variance,
the distribution converges to a power law with finite variance but infinite (or undefined)
higher moments.
3. Under the other regime, where the errors are decreasing (proportionally) for higher order
errors, the ending distribution becomes fat-tailed but in a benign way as it retains its finite
variance attribute (as well as all higher moments), allowing convergence to Gaussian under
Central Limit.
We manage to set a boundary between these two regimes.
In both regimes the use of a thin-tailed distribution is not warranted unless higher order errors
can be completely eliminated a priori.
Higher order integrals in the Standard Gaussian Case
We start with the case of a Gaussian and focus the uncertainty on the assumed standard deviation. Define (,,x) as the Gaussian PDF for value x with mean and standard deviation
.
A 2nd order stochastic standard deviation is the integral of across values of R+ , under
the distribution f ( , 1 , ) , with 1 its scale parameter (our approach to trach the error of the
error), not necessarily its standard deviation; the expected value of 1 is 1 .
f (x)1 =
Z
0
(, , x) f ( , 1 , ) d
203
f (x) N =
Z
0
...
Z
0
(, , x) f ( , 1 , )
f (1 , 2 , 1 ) ... f (N 1 , N , N 1 ) d d1 d2 ... dN
(9.1)
The problem is that this approach is parameter-heavy and requires the specifications of the
subordinated distributions (in finance, the lognormal has been traditionally used for 2 (or Gaus2
sian for the ratio Log[ t2 ] since the direct use of a Gaussian allows for negative values). We would
need to specify a measure f for each layer of error rate. Instead this can be approximated by
using the mean deviation for , as we will see next2 .
1
(, (1 + a1 ), x ) + (, (1 a1 ), x)
f (x)1 =
2
(9.2)
Now assume uncertainty about the error rate a1 , expressed by a2 , in the same manner as before.
Thus, as a first method, the multiplicative effect, in place of 1 a1 we have (1 a1 )(1 a2 ).
Later we will use the non-multiplicative (or, rather, weakly multiplicative) error expansion (1
(a1 (1 (a2 (1 a3 ( ...))).
The second order stochastic standard deviation:
f (x)2 =
1
, (1 + a1 )(1 + a2 ), x +
4
!
, (1 a1 )(1 + a2 ), x) + (, (1 + a1 )(1 a2 ), x + , (1 a1 )(1 a2 ), x
(9.3)
204
f (x) N =
where
MiN
is the
ith
1
2N
2N
(, MiN , x)
i=1
(a j Ti,j + 1)
!2 N
j=1
ith line
i=1
jth column
1
T=
1
1
1
1
1
1
1
1
1
1
1
1
1
1
and
3
M =
(1 a1 ) (1 a2 ) (1 a3 )
(1 a1 ) (1 a2 ) (a3 + 1)
(1 a1 ) (a2 + 1) (1 a3 )
(1 a1 ) (a2 + 1) (a3 + 1)
(a1 + 1) (1 a2 ) (1 a3 )
(a1 + 1) (1 a2 ) (a3 + 1)
(a1 + 1) (a2 + 1) (1 a3 )
205
0.6
0.5
0.4
0.3
0.2
0.1
-6
-4
-2
9.2
,
(a
+
1)
(1
a)
,
x
(9.4)
j
j=0
Because of the linearity of the sums, when a is constant, we can use the binomial distribution
as weights for the moments (note again the artificial effect of constraining the first moment in
the analysis to a set, certain, and known a priori).
M1 (N) =
N
M2 (N) = 2 a2 + 1 + 2
N
M3 (N) = 3 2 a2 + 1 + 3
N
N
M4 (N) = 6 2 2 a2 + 1 + 4 + 3 a4 + 6a2 + 1 4
For clarity, we simplify the table of moments, with =0
M1 (N) = 0
N
M2 (N) = a2 + 1 2
M3 (N) = 0
N
M4 (N) = 3 a4 + 6a2 + 1 4
M5 (N) = 0
N
M6 (N) = 15 a6 + 15a4 + 15a2 + 1 6
206
affect the first absolute moment = 2 (that is, the initial assumed ). The situation would be
different under addition of x.
Every recursion multiplies the variance of the process by (1 + a2 ). The process is similar to a
stochastic volatility model, with the standard deviation (not the variance) following a lognormal
distribution, the volatility of which grows with M, hence will reach infinite variance at the limit.
Consequences
For a constant a > 0, and in the more general case with variable a where an an1 , the moments
explode.
N
Even the smallest value of a >0, since 1 + a2 is unbounded, leads to the second moment
going to infinity (though not the first) when N . So something as small as a .001% error
rate will still lead to explosion of moments and invalidation of the use of the class of L2
distributions.
In these conditions, we need to use power laws for epistemic reasons, or, at least, distributions outside the L2 norm, regardless of observations of past data.
Note that we need an a priori reason (in the philosophical sense) to cutoff the N somewhere,
hence bound the expansion of the second moment.
9.3
Convergence to power law would require the following from the limit distribution. Where P> x
is short for P(X > x), P> x = L(x) x and L(x) is a slowly varying function.
= lim lim (x, N)
x N
We know from the behavior of moments that, if convergence is satisfied, (1, 2).
We can have a visual idea with the Log-Log plot (Figure 9.3) how, at higher orders of
stochastic volatility, with equally proportional stochastic coefficient, (where a1 = a2 = ... = an =
1
10 ) the density approaches that of a power law, as shown in flatter density on the LogLog plot.
The probabilities keep rising in the tails as we add layers of uncertainty until they seem to reach
the boundary of the power law, while ironically the first moment remains invariant.
The same effect takes place as a increases towards 1, as at the limit the tail exponent P>x
approaches 1 but remains >1.
(x, N) = 1
log f (x |,,N)
x
log(x)
x1
207
a=
10
Log PrHxL
0.1
10-4
10-7
10-10
10-13
Log x
1.5
2.0
3.0
5.0
7.0
10.0
15.0
20.0
30.0
x 1 (N)
2 (N)
(9.5)
where
K
1 (N) = x(a + 1)3j (1 a)3j3K
j=0
K
1 2
2j
2j2K
exp x (a + 1) (1 a)
j
2
2 (N) = (a + 1) j (1 a) jK
j=0
K
1
exp x2 (a + 1)2j (1 a)2j2K
j
2
Making the variable continuous (binomial as ratio of gamma functions) makes it equivalent,
at large N, to:
(x, N) = 1
x(1 a) N 1 (N)
2 2 (N)
(9.6)
where
1 (N)
Z N
0
1
exp x2 (a + 1)2y (1 a)2y2N
2
208
dy
2 (N)
Z N
0
2
a+1
y
(N + 1)
2 (y + 1)(N y + 1)
1
exp x2 (a + 1)2y (1 a)2y2N dy
2
P > K| N = 2
N 1
N
j
j=0
erfc
(9.7)
2(a + 1) j (1 a) N j
Rz
0
et dt
Convexity effect The next two tables shows the ratio of exceedant probability under different
values of N divided by the probability in the case of a standard Gaussian.
P>3,N
P>3,N=0
P>5,N
P>5,N=0
1
10
P>10,N
P>10,N=0
1.01724
1.155
10
1.0345
1.326
45
15
1.05178
1.514
221
20
1.06908
1.720
922
25
1.0864
1.943
3347
1
100
P>3,N
P>3,N=0
P>5,N
P>5,N=0
P>10,N
P>10,N=0
2.74
146
1.09 1012
10
4.43
805
8.99 1015
15
5.98
1980
2.21 1017
20
7.38
3529
1.20 1018
25
8.64
5321
3.62 1018
209
H1 - a1 L Ha2 + 1L Ha3 + 1L
H1 - a1 L Ha2 + 1L
H1 - a1 L Ha2 + 1L H1 - a3 L
1 - a1
H1 - a1 L H1 - a2 L Ha3 + 1L
H1 - a1 L H1 - a2 L
H1 - a1 L H1 - a2 L H1 - a3 L
9.4
M1 (N) =
M2 (N) =
2 + 2
M3 (N) =
3 + 32
M4 (N) = 34 a2 + 1
Hence (3, 4)
9.5
N
+ 4 + 62 2
As we said, we may have (actually we need to have) a priori reasons to decrease the parameter a
or stop N somewhere. When the higher order of ai decline, then the moments tend to be capped
(the inherited tails will come from the lognormality of ).
210
(9.8)
i=1
N 1
1 aqi
We can reexpress ( 9.8) using the Q-Pochhammer symbol (a; q) N = i=1
M2 (N) = 2 a21 ; 2 N
Which allows us to get to the limit
lim M2 (N) =
2 ; 2
a21 ; 2
( 2 1) ( 2 + 1)
N 1
6a21 2i + a41 4i + 1
i=0
M4 (N) = 34
2 2 3 a21 ; 2
lim M4 (N) = 34
2 2 3 a21 ; 2
3 + 2 2 a21 ; 2 N
3 + 2 2 a21 ; 2
(9.9)
(9.10)
So the limiting second moment for =.9 and a_1=.2 is just 1.28 2 , a significant but relatively
benign convexity bias. The limiting fourth moment is just 9.884 , more than 3 times the Gaussians (3 4 ), but still finite fourth moment. For small values of a and values of close to 1, the
fourth moment collapses to that of a Gaussian.
211
1
L
x, , 1 + TN .AN i
i=1
a2 ,
a3
a3 + a2 + a
a3 + a2 +
a3 a2 +
a3 a2 +
A3 T3 =
a3 + a2
a3 + a2
a3 a2
a3 a2
a 2i
i=0
At the limit:
lim M 4 ( N ) =
12 2
+ 4 + 12 2
1 a2
9.6
212
3 See the discussion of the Gosset and Fisher approach in Chapter 3 of Mosteller and Tukey [78].
4 I thank Andrew Gelman and Aaron Brown for the discussion.
213
10
PA R A M E T R I Z A T I O N A N D T A I L S
Chapter Summary 10: We present case studies around the point that, simply, some models
depend quite a bit on small variations in parameters. The effect on the Gaussian is easy
to gauge, and expected. But many believe in power laws as panacea. Even if one believed
the r.v. was power law distributed, one still would not be able to make a precise statement
on tail risks. Shows weaknesses of calibration of Extreme Value Theory.
This chapter is illustrative; it will initially focus on nonmathematical limits to producing estimates of M TX ( A , f ) when A is limited to the tail. We will see how things get worse when one is
sampling and forecasting the maximum of a random variable.
10.1
We saw the shortcomings of parametric and nonparametric methods so far. What are left are
power laws; they are a nice way to look at the world, but we can never really get to know the
exponent , for a spate of reasons we will see later (the concavity of the exponent to parameter
uncertainty). Suffice for now to say that the same analysis on exponents yields a huge in-sample
variance and that tail events are very sensitive to small changes in the exponent.
For instance, for a broad set of stocks over subsamples, using a standard estimation method
(the Hill estimator), we get subsamples of securities. Simply, the variations are too large for a
reliable computation of probabilities, which can vary by > 2 orders of magnitude. And the effect
on the mean of these probabilities is large since they are way out in the tails.
The way to see the response to small changes in tail exponent with probability: considering
P>K K , the sensitivity to the tail exponent P>K = K log(K).
Now the point that probabilities are sensitive to assumptions brings us back to the Black Swan
problem. One might wonder, the change in probability might be large in percentage, but who
cares, they may remain small. Perhaps, but in fat tailed domains, the event multiplying the
probabilities is large. In life, it is not the probability that matters, but what one does with it,
such as the expectation or other moments, and the contribution of the small probability to the
total moments is large in power law domains.
R
For all powerlaws, when K is large, with > 1, the unconditional "shortfall" S+ = K x(x)dx
R K
( 1)2
There is a deeper problem related to the effect of model error on the estimation of , which
compounds the problem, as tends to be underestimated by Hill estimators and other methods,
but let us leave it for now.
215
1.5
2.0
2.5
3.0
Figure 10.1: The effect of small changes in tail exponent on a probability of exceeding a certain point. To the
left, a histogram of possible tail exponents across >4 103 variables. To the right the probability, probability
of exceeding 7 times the scale of a power law ranges from 1 in 10 to 1 in 350. For further in the tails the
effect is more severe.
10.2
We saw earlier how difficult it is to compute risks using power laws, owing to excessive model
sensitivity. Let us apply this to the Extreme Value Theory, EVT. (The idea is that is useable by
the back door as test for nonlinearities exposures not to get precise probabilities).
On its own it can mislead. The problem is the calibration and parameter uncertainty in the
real world we dont know the parameters. The ranges in the probabilities generated we get are
monstrous.
We start with a short presentation of the idea, followed by an exposition of the difficulty.
M j = max Xi,j
n
i=1
Xi,j
n p
i=1 j=1
Figure 10.2 and 10.2 plot a histogram of the result of both the simulation and the fitting of a
distribution.
Let us now fit to the sample from the simulation to g, the density of an Extreme Value Distribution for x (or the Gumbel for the negative variable x), with location and scale parameters
and , respectively: g(x; , ) =
216
x
x e
1.5
1.0
0.5
3.5
4.0
4.5
5.0
5.5
1.5
1.0
Figure 10.3: Fitting an extreme value distribution (Gumbel for the maxima) =
3.97904, = 0.235239
0.5
3.5
4.0
4.5
5.0
5.5
217
n
\
P(Xi ) = F(x)n ,
i=1
where F is the cumulative d.f of the Gaussian. Taking the first derivative of the cumulative
distribution to get the density of the distribution of the maximum,
1
pn (x) x ( F(x)n ) =
2 2 n ne
x2
2
n 1
erf x + 1
2
x
x e
, where
After some derivations[see below], g(x) =
1
1
2 n2 , where erfc is the
= 2erfc
inverse error function, and
1
1
2
2
= 2 erfc
2 n erfc
2 en
For n = 30K, {, } = {3.98788, 0.231245}
p
1
1
log(log(n))+log(4)
2 log(n)
x
1
x
for x>0
218
0.04
0.03
0.02
0.01
100
200
300
400
500
600
0.04
0.03
0.02
0.01
50
100
150
200
219
1
P>3
1
P>10
1
P>20
1
P>40
1
P>80
1.
1.25
1.5
1.75
2
2.25
2.5
2.75
3.
3.25
3.5
3.75
4.
4.25
4.5
4.75
5.
4.
4.
6.
7.
10.
12.
16.
21.
28.
36.
47.
62.
82.
107.
141.
185.
244.
11.
18.
32.
57.
101.
178.
317.
563.
1001.
1779.
3163.
5624.
10001.
17783.
31623.
56235.
100001.
21.
43.
90.
190.
401.
846.
1789.
3783.
8001.
16918.
35778.
75660.
160001.
338359.
715542.
1.5106
3.2106
41.
101.
253.
637.
1601.
4024.
10120.
25449.
64001.
160952.
404772.
1.01106
2.56106
6.43106
1.61107
4.07107
1.02108
81.
240.
716.
2140.
6400
19141.
57244.
171198.
512001.
1.5 106
4.5106
1.3107
4.0107
1.2108
3.6108
1.1109
3.27109
Table 18: EVT for different tail parameters . We can see how a perturbation of moves the probability of
a tail event from 6, 000 to 1.5 106 . [ADDING A TABLE FOR HIGHER DIMENSION WHERE THINGS
ARE A LOT WORSE]
220
10.3
We can use power laws in the "near tails" for information, not risk management. That is, not
pushing outside the tails, staying within a part of the distribution for which errors are not
compounded.
I was privileged to get access to a database with cumulative sales for editions in print that had
at least one unit sold that particular week (that is, conditional of the specific edition being still
in print). I fit a powerlaw with tail exponent ' 1.3 for the upper 10% of sales (graph), with
N=30K. Using the Zipf variation for ranks of powerlaws, with r x and ry the ranks of book x and
y, respectively, Sx and Sy the corresponding sales
Sx
=
Sy
rx
ry
100
1000
1
1.3
Note this is only robust in deriving the sales of the lower ranking edition (ry > r x ) because of
inferential problems in the presence of fat-tails.
P>X
1
0.1
Near tail
0.01
=1.3
0.001
10-4
X
100
104
106
221
222
H
h.1
P O I S S O N V S . P O W E R L A W TA I L S
By the masquerade problem, any power law can be seen backward as a Gaussian plus a series
of simple (that is, noncompound) Poisson jumps, the so-called jump-diffusion process. So the
use of Poisson is often just a backfitting problem, where the researcher fits a Poisson, happy with
the "evidence".
The next exercise aims to supply convincing evidence of scalability and NonPoisson-ness of
the data (the Poisson here is assuming a standard Poisson). Thanks to the need for the probabililities add up to 1, scalability in the tails is the sole possible model for such data. We may not
be able to write the model for the full distribution but we know how it looks like in the tails,
where it matters.
The Behavior of Conditional Averages With a scalable (or "scale-free") distribution, when K is
"in the tails" (say you reach the point when 1 F(X > x) = Cx where C is a constant and
the power law exponent), the relative conditional expectation of X (knowing that X >K) divided
E[X | X >K]
by K, that is,
is a constant, and does not depend on K. More precisely, the constant is
K
.
1
R
x f (x, ) dx
K
RK
=
1
f
(x,
)
dx
K
This provides for a handy way to ascertain scalability by raising K and looking at the averages
in the data.
Note further that, for a standard Poisson, (too obvious for a Gaussian): not only the conditional expectation depends on K, but it "wanes", i.e.
R
lim
mx
.
K (x) dx
R mx
K
K x! dx
!
=1
Calibrating Tail Exponents In addition, we can calibrate power laws. Using K as the cross-over
point, we get the exponent above it the same as if we used the Hill estimator or ran a regression above some point.
We heuristically defined fat tails as the contribution of the low frequency events to the total
properties. But fat tails can come from different classes of distributions. This chapter will present
the difference between two broad classes of distributions.
This brief test using 12 million pieces of exhaustive returns shows how equity prices (as well
as short term interest rates) do not have a characteristic scale. No other possible method than a
Paretan tail, albeit of unprecise calibration, can charaterize them.
223
h.2
This exercise was done using about every piece of data in sight: single stocks, macro data,
futures, etc.
Equity Dataset We collected the most recent 10 years (as of 2008) of daily prices for U.S. stocks
(no survivorship bias effect as we included companies that have been delisted up to the last
trading day), n= 11,674,825 , deviations expressed in logarithmic returns.
We scaled the data using various methods.
The expression in "numbers of sigma" or standard deviations is there to conform to industry
language (it does depend somewhat on the stability of sigma). In the "MAD" space test we
used the mean deviation.
log Sti
Sti 1
MAD(i) =
1
N
log Sti j
t n Si
j+t1
We focused on negative deviations. We kept moving K up until to 100 MAD (indeed) and
we still had observations.
Implied|K =
MAD
1.
2.
5.
10.
15.
20.
25.
50.
70.
75.
100.
E [ X | X <K ]
1.75
3.02
7.96
15.32
22.32
30.24
40.87
101.75
156.70
175.42
203.99
n(forX < K)
1.32 106
300806.
19285.
3198.
1042.
418.
181.
24.
11.
9.
7.
E[ X | X <K ]
K
1.75
1.51
1.59
1.53
1.48
1.51
1.63
2.03
2.23
2.33
2.03
E [ X | X <K ]
E [ X | X <K ] K
Implied
2.32
2.95
2.68
2.87
3.04
2.95
2.57
1.96
1.80
1.74
1.96
Sigma-Space In the "sigma space" test we used a rolling 22 day window scaled by the noncentral
standard deviations. We did not add a mean for reasons explained elsewhere.
Short term Interest Rates Literally, you do not even have a large number K for which scalability
drops from a small sample effect.
224
E [ X | X <K ]
3.01
8.02
15.60
30.41
113.324
170.105
180.84
192.543
251.691
E[ X |
X <K
n(forX < K)
K
343952.
1.50
21156.
1.60
3528.
1.56
503.
1.52
20.
2.26
10.
2.43
9.
2.26
8.
2.13
5.
2.51
Implied
2.97
2.65
2.78
2.91
1.78
1.69
1.79
1.87
1.65
Implied
1.38361
1.7076
2.68491
2.84496
3.16464
3.32782
UK Rates 1990-2007
n=4143
MAD E [ X | X <K ]
0.5
1.68802
1.
2.23822
3.
4.97319
5.
8.43269
6.
9.56132
7.
11.4763
E[ X |
X <K
n(forX < K)
K
1270
3.37605
806
2.23822
140
1.65773
36
1.68654
26
1.59355
16
1.63947
Implied
1.42087
1.80761
2.52038
2.45658
2.68477
2.56381
225
226
11
Chapter Summary 11: A less technical demarcation between Black Swan Domains and
others.
Let us return to M[A, f (x)] of Chapter 3. A quite significant result is that M[A,x n ] may not
converge, in the case of, say power laws Rwith exponent < n, but M [ A, x m ] where m < n,
would converge. Well, where the integral f (x)p(x) dx does not exist, by clipping tails, we
can make the payoff integrable. There are two routes;
R 1) Limiting f (turning an open payoff to a binary): when f (x) is a constant as in a binary
Kp(x)dx will necessarily converge if p is a probability distribution.
2) Clipping tails: (and this is the business we will deal with in Antifragile, Part II), where the
RH
payoff is bounded, A = [L, H], or the integral L f (x)p(x)dx will necessarily converge.
11.1
M0 depends on the 0th moment, that is, Binary, or simple, i.e., as we saw, you just care if
something is true or false. Very true or very false does not matter. Someone is either pregnant
or not pregnant. A statement is true or false with some confidence interval. (I call these M0
as, more technically, they depend on the zeroth moment, namely just on probability of events,
and not their magnitude you just care about raw probability). A biological experiment in
the laboratory or a bet with a friend about the outcome of a soccer game belong to this category.
M1+ Complex, depend on the 1st or higher moments. You do not just care of the frequency
but of the impact as well, or, even more complex, some function of the impact. So there is
another layer of uncertainty of impact. (I call these M1+, as they depend on higher moments of
Simple payoffs
Complex
payoffs
Distribution 1
(thin tailed)
First Quadrant
Extremely
Safe
Second
Quadrant:
Safe
Distribution 2
(no or unknown
characteristic
scale)
Fourth
Quadrant:
Dangers
227
Mo
True/False
f(x)=0
Medicine (health
not epidemics)
Psychology experiments
M1
Expectations
LINEAR PAYOFF
f(x)=1
Finance : nonleveraged Investment
Insurance, measures of expected
shortfall
Bets (prediction
markets)
Binary/Digital
derivatives
Climate
Life/Death
Economics (Policy)
Security:
Terrorism, Natural catastrophes
Epidemics
Casinos
M2+
NONLINEAR PAYOFF
f(x) nonlinear(= x 2 ,
x 3 , etc.)
Derivative payoffs
Dynamically
hedged portfolios
Leveraged portfolios (around the
loss point)
Cubic payoffs (strips
of out of the money
options)
Errors in analyses of
volatility
Calibration of nonlinear models
Expectation
weighted by nonlinear utility
Kurtosis-based positioning (volatility
trading)
Conclusion The 4th Quadrant is mitigated by changes in exposures. And exposures in the 4th
quadrant can be to the negative or the positive, depending on if the domain subset A exposed
either on the left or on the right.
228
12
Chapter Summary 12: We present the difference between ruin problems, particularly
when systemic, and other risk management of a more computable nature.
A more technical exposition of the fourth quadrant, in which we replace the 4th Q with the
precautionary principle. Simply one need to "show" that a given policy does not belong to IV in
Table 21, or, alternatively exert more vigilance there.
Table 21: The Four Quadrants and Exposure Classes H A and H B
Z = i i f i (Xi ) H A
Z = i i f i (Xi ) H B
XA
II
XB
III
IV: Domain of PP
(x)
B = { X B : limx+ 11FF(x)
= 2}, where F 2 = F 0 F is the cumulative distribution of X1 + X2 , the
sum of two copies of X.
P(Sn > x)
n > x)
Or, equivalently, For a given n 2, a) limx P(S
P(X > x) = n, b) limx P( Mn > x ) = 1.
(12.1)
229
12.1
The precautionary principle (PP) states that if an action or policy has a suspected risk of causing
severe harm to the public domain (affecting general health or the environment globally), the
action should not be taken in the absence of scientific near-certainty about its safety. Under
these conditions, the burden of proof about absence of harm falls on those proposing an action,
not those opposing it. PP is intended to deal with uncertainty and risk in cases where the absence of evidence and the incompleteness of scientific knowledge carries profound implications
and in the presence of risks of "black swans", unforeseen and unforeseable events of extreme
consequence.
This non-naive version of the PP allows us to avoid paranoia and paralysis by confining precaution to specific domains and problems. Here we formalize PP, placing it within the statistical
and probabilistic structure of ruin problems, in which a system is at risk of total failure, and
in place of risk we use a formal"fragility" based approach. In these problems, what appear to
be small and reasonable risks accumulate inevitably to certain irreversible harm. Traditional
cost-benefit analyses, which seek to quantitatively weigh outcomes to determine the best policy
option, do not apply, as outcomes may have infinite costs. Even high-benefit, high-probability
outcomes do not outweigh the existence of low probability, infinite cost optionsi.e. ruin. Uncertainties result in sensitivity analyses that are not mathematically well behaved. The PP is
increasingly relevant due to man-made dependencies that propagate impacts of policies across
the globe. In contrast, absent humanity the biosphere engages in natural experiments due to
random variations with only local impacts.
The aim of the precautionary principle (PP) is to prevent decision makers from putting society
as a wholeor a significant segment of itat risk from the unexpected side effects of a certain
type of decision. The PP states that if an action or policy has a suspected risk of causing severe
harm to the public domain (such as general health or the environment), and in the absence of
scientific near-certainty about the safety of the action, the burden of proof about absence of harm
230
0.6
0.4
0.2
2000
4000
6000
8000
10 000
Exposure
falls on those proposing the action. It is meant to deal with effects of absence of evidence and
the incompleteness of scientific knowledge in some risky domains.1
We believe that the PP should be evoked only in extreme situations: when the potential harm
is systemic (rather than localized) and the consequences can involve total irreversible ruin, such
as the extinction of human beings or all life on the planet.
Precautionary Approach
localized harm
nuanced cost-benefit
statistical
statistical
variations
convergent probabibilities
recoverable
independent factors
evidence based
thin tails
bottom-up, tinkering
evolved
systemic ruin
avoid at all costs
fragility based
probabilistic non-statistical
ruin
divergent probabilities
irreversible
interconnected factors
precautionary
fat tails
top-down engineered
human-made
Table 22: Two different types of risk and their respective characteristics compared
231
Low model
uncertainty
Ruin
High model
uncertainty
Ruin
probability
-5
10
232
12.2
12.3
We show in Figures 12.2 and 12.3 that an increase in uncertainty leads to an increase in the
probability of ruin, hence "skepticism" is that its impact on decisions should lead to increased,
not decreased conservatism in the presence of ruin. More skepticism about models implies
more uncertainty about the tails, which necessitates more precaution about newly implemented
techniques, or larger size of exposures. As we said, Nature might not be smart, but its longer
track record means smaller uncertainty in following its logic.
Mathematically, more uncertainty about the future or about a model increases the scale
of the distribution, hence thickens the "left tail" (as well as the "right one") which raises the
potential ruin. The survival probability is reduced no matter what takes place in the right tail.
Hence skepticim about climate models should lead to more precautionary policies.
In addition, such increase uncertainty matters far more in Extremistan and has benign effects
in Mediocristan. Figure 12.3 shows th asymmetries between costs and benefits as far as ruin
probabilities, and why these matter more for fat-tailed domains than thin-tailed ones. In thintailed domains, an increase in uncertainty changes the probability of ruin by several orders of
magnitude, but the effect remains small: from say 1040 to 1030 is not quite worrisome. In
fat-tailed domains, the effect is sizeable as we start with a substantially higher probability of
ruin (which is typically underestimated, see [? ]).
12.4
233
234
235
236
Today many mathematical or conceptual models that are claimed to be rigorous are based upon
unvalidated and incorrect assumptions and are not robust to changes in these assumptions. Such
models are deemed rational in the sense that they are logically derived from their assumptions,
and, further, can be used to assess rationality by examining deviations from such models, as
indicators of irrationality. Except that it is often the modeler who is using an incomplete representation of the reality, hence using an erroneous benchmark for rationality. Often the modelers
are not familiar with the dynamics of complex systems or use antiquated statistical methods
that do not take into account fat-tails and make inferences that would not be acceptable under different classes of probability distributions. Many biases, such as the ones used by Cass
Sunstein (mentioned above), about the overestimation of the probabilities of rare events in fact
correspond to the testers using a bad probability model that is thin-tailed. See Ref. [? ] for a
deeper discussion.
It has became popular to claim irrationality for GMO and other skepticism on the part of the
general publicnot realizing that there is in fact an "expert problem" and such skepticism is
healthy and even necessary for survival. For instance, in The Rational Animal [? ], the authors
pathologize people for not accepting GMOs although "the World Health Organization has never
found evidence of ill effects," a standard confusion of evidence of absence and absence of evidence. Such pathologizing is similar to behavioral researchers labeling hyperbolic discounting
as "irrational" when in fact it is largely the researcher who has a very narrow model and richer
models make the "irrationality" go away.
These researchers fail to understand that humans may have precautionary principles against
systemic risks, and can be skeptical of the untested consequences of policies for deeply rational
reasons, even if they do not express such fears in academic format.
237
13
S K I N I N T H E G A M E A N D R I S K TA K I N G
Chapter Summary 13: Standard economic theory makes an allowance for the agency
problem, but not the compounding of moral hazard in the presence of informational
opacity, particularly in what concerns high-impact events in fat tailed domains (under
slow convergence for the law of large numbers). Nor did it look at exposure as a filter
that removes nefarious risk takers from the system so they stop harming others. (In
the language of probability, skin in the game creates an absorbing state for the agent,
not just the principal). But the ancients did; so did many aspects of moral philosophy.
We propose a global and morally mandatory heuristic that anyone involved in an action
which can possibly generate harm for others, even probabilistically, should be required
to be exposed to some damage, regardless of context. While perhaps not sufficient, the
heuristic is certainly necessary hence mandatory. It is supposed to counter voluntary and
involuntary risk hiding and risk transfer in the tails.
The literature in risk, insurance, and contracts has amply dealt with the notion of information
asymmetry (see Ross, 1973, Grossman and Hart, 1983, 1984, Tirole 1988, Stiglitz 1988), but not
with the consequences of deeper information opacity (in spite of getting close, as in Hulmstrom,
1979), by which tail events are impossible to figure out from watching time series and external
signs: in short, in the "real world" (Taleb, 2013), the law of large numbers works very slowly, or
does not work at all in the time horizon for operators, hence statistical properties involving tail
events are completely opaque to the observer. And the central problem that is missing behind the
abundant research on moral hazard and information asymmetry is that these rare, unobservable
events represent the bulk of the properties in some domains. We define a fat tailed domain as
follows: a large share of the statistical properties come from the extremum; for a time series
involving n observations, as n becomes large, the maximum or minimum observation will be of
the same order as the sum. Excursions from the center of the distributions happen brutally and
violently; the rare event dominates. And economic variables are extremely fat tailed (Mandelbrot,
1997). Further, standard economic theory makes an allowance for the agency problem, but not
for the combination of agency problem, informational opacity, and fat-tailedness. It has not
yet caught up that tails events are not predictable, not measurable statistically unless one is
causing them, or involved in increasing their probability by engaging in a certain class of actions
with small upside and large downside. (Both parties may not be able to gauge probabilities in
the tails of the distribution, but the agent knows which tail events do not affect him.) Sadly,
the economics literatures treatment of tail risks , or "peso problems" has been to see them as
outliers to mention en passant but hide under the rug, or remove from analysis, rather than a core
center of the modeling and decision-making, or to think in terms of robustness and sensitivity to
unpredictable events. Indeed, this pushing under the rug the determining statistical properties
explains the failures of economics in mapping the real world, as witnessed by the inability of
the economics establishment to see the accumulation of tail risks leading up to the financial
crisis of 2008 (Taleb, 2009). The parts of the risk and insurance literature that have focused on
tail events and extreme value theory, such as Embrechts (1997), accepts the large role of the
239
240
241
242
Figure 13.1: The most effective way to maximize the expected payoff to the agent at the expense of the
principal.
13.1
This section will analyze the probabilistic mismatch or tail risks and returns in the presence of a
principal-agent problem.
Transfer of Harm: If an agent has the upside of the payoff of the random variable, with no downside,
and is judged solely on the basis of past performance, then the incentive is to hide risks in the left tail
using a negatively skewed (or more generally, asymmetric) distribution for the performance. This can be
generalized to any payoff for which one does not bear the full risks and negative consequences of ones
actions.
Let P(K, M) be the payoff for the operator over M incentive periods
M
j
P(K, M) qt+(i1)t xt+it K + 1t(i1)+t<
(13.1)
i=1
243
R
x f (x) dx
RK j
K f j (x) dx
and E
j
RK
RK
x f j (x) dx
f j (x)
dx
Fj
Fj+ ,
with values
>1 for positive asymmetry, and <1 for negative ones. Intuitively, skewness has probabilities
and expectations moving in opposite directions: the larger the negative payoff, the smaller the
probability to compensate.
We do not assume a fair game, that is, with unbounded returns m (-,), Fj+ E+j +
Fj E
j = m, which we can write as
m+ + m = m.
E+j
1t+it<
!
(13.2)
i=1
Since assuming independent and identically distributed agents payoffs, the expectation at
stopping time corresponds to the expectation
of stopping
com time
multiplied by the
expected
M
1t(i1)+t< = E
pensation to the agent E j + . And E i=1
M
1t(i1)+t< M .
i=1
The expectation of stopping time can be written as the probability of success under the condition of no previous loss:
E
i=1
1t+it<
We can express the stopping time condition in terms of uninterrupted success runs. Let be
the ordered set of consecutive success runs {{ F }, {SF}, {SSF}, ..., {(M 1) consecutive S, F }},
where S is success and F is failure over period t, with associated corresponding probabilities:
244
Fj+ (i1)
1 Fj+ = 1 Fj+ M ' 1
(13.3)
i=1
For M large, since Fj+ (0,1) we can treat the previous as almost an equality, hence:
1t+(i1)t<
!
=
i=1
+ (i 1)
+
(i
1)
F
1
F
'
j
j
i=1
Fj+
1 Fj+
Fj+
1 Fj+
which increases by i) increasing E+j , ii) minimizing the probability of the loss Fj , but, and thats
the core point, even if i) and ii) take place at the expense of m the total expectation from the
package.
Alarmingly, since E+j =
mm
Fj+ ,
return m if it comes from the left side of the distribution, m . Seen in skewness space, the
expected agent payoff maximizes under the distribution j with the lowest value of j (maximal
negative asymmetry). The total expectation of the positive-incentive without-skin-in-the-game
depends on negative skewness, not on m.
Figure 13.2: Indy Mac, a failed firm during the subprime crisis (from Taleb 2009). It is a representative of
risks that keep increasing in the absence of losses, until the explosive blowup.
245
P(K, M)
q (i)
i=1
xt+i t K
1t+(i1)t< ,
(13.4)
(i) 1t(i1)+t<
(13.5)
i=1
Assuming the rate of conditional growth is a constant r [0,) , and making the replacement
(i) eri , we can call the last term in equation 13.5 the multiplier of the expected return to the
agent:
!
E
eir 1t(i1)+t<
i=1
(13.6)
i=1
(F+ 1) (F+ ) M ( Me(M+1)r F+ (M1)e(M+2)r ) F+ e2r
(F+ er 1)2
(13.7)
We can get the table of sensitivities for the "multiplier" of the payoff:
F=.6
0.7
0.8
0.9
r=0
1.5
2.32
3.72
5.47
0.1
2.57
4.8
10.07
19.59
0.2
4.93
12.05
34.55
86.53
0.3
11.09
38.15
147.57
445.59
246
2 2
Forecasters We can see how forecasters who do not have skin in the game have the incentive
of betting on the low-impact high probability event, and ignoring the lowerR probability ones,
even if these are high impact. There isR a confusion between digital payoffs f j (x) dx and full
distribution, called vanilla payoffs, x f j (x) dx, see Taleb and Tetlock (2013)11 .
Opacity and Risk Hiding: NonMathematical Summary We will next proceed to summarize the mathematical argument in verbal form.
A) If an agent has the upside of the payoff of the random variable, with no downside
[OR A DISPROPORTIONATE SHARE OF UPSIDE WITH RESPECT TO THE DOWNSIDE],
and is judged solely on the basis of past performance, then the incentive is to hide risks in
the left tail using a negatively skewed (or more generally, asymmetric) distribution for the
performance. This can be generalized to any payoff for which one does not bear the full
Zs
actions.
risks and negative consequences of oneA
B) Further, even if it is not intentional, i.e., the agent does not aim at probabilistic rent at
the expense of the principal (at variance with the way agents are treated in the economics
literature); by a survival argument, those agents without skin in the game who tend to
engage in strategies that hide risk in the tail tend to fare better and longer and populate the
agent population. So the argument is not one of incentive driving the agents, but one of
survival.
We can sketch a demonstration of these statements with the following reasoning. Assume
that an agent has a payoff as a proportional cut of his performance or the benefits to the
principal, and can get a percentage at year end, his compensation being tied to the visible
income. The timing of the compensation is periodic, with no total claw back (subsequent
obligation to completely return past compensation). The expected value to the agent is that
of a stream, a sum of payoffs over time, extending indefinitely (or bounded by the life of
the agent). Assume that a loss will reduce his future risk-taking, or even terminate it, in
terms of shrinking of such contracts, owing to change in reputation. A loss would hurt the
track record, revealing it so to speak, making such a stream of payoffs stop. In addition, the
10 This discussion of a warped probabilistic incentive corresponds to what John Kay has called the "Taleb distribution",
John Kay "A strategy for hedge funds and dangerous drivers", Financial Times, 16 January 2003.
11 Money managers do not have enough skin in the game unless they are so heavily invested in their funds that they can
end up in a net negative form the event. The problem is that they are judged on frequency, not payoff, and tend to cluster
together in packs to mitigate losses by making them look like "industry event". Many fund managers beat the odds by
selling tails, say covered writes, by which one can increase the probability of gains but possibly lower the expectation.
They also have the optionality of multi-time series; they can manage to hide losing funds in the event of failure. Many
fund companies bury hundreds of losing funds away, in the "cemetery of history" (Taleb, 2007) .
247
248
Part III
(ANTI)FRAGILITY AND NONLINEAR RESPONSES TO
RANDOM VARIABLES
14
EXPOSURES AS TRANSFORMED
RANDOM VARIABLES
Chapter Summary 14: Deeper into the conflation between a random variable and exposure to it.
14.1
A convex and linear function of a variable x. Confusing f(x) (on the vertical) and x (the horizontal)
is more and more significant when f(x) is nonlinear. The more convex f(x), the more the statistical
and other properties of f(x) will be divorced from those of x. For instance, the mean of f(x) will be
different from f(Mean of x), by Jensens ineqality. But beyond Jensens inequality, the difference
in risks between the two will be more and more considerable. When it comes to probability,
the more nonlinear f, the less the probabilities of x matter compared to the nonlinearity of f.
Moral of the story: focus on f, which we can alter, rather than the measurement of the elusive
properties of x.
There are infinite numbers of functions F depending on a unique variable x.
All utilities need to be embedded in F.
14.1.1 Limitations of knowledge
. What is crucial, our limitations of knowledge apply to x not necessarily to f(x). We have no
control over x, some control over F(x). In some cases a very, very large control over f(x).
This seems naive, but people do, as something is lost in the translation.
Probability Distribution of x
251
14.2
Say x follows a distribution p(x) and z = f (x) follows a distribution g(z). Assume g(z) continuous,
increasing, and differentiable for now.
The density p at point r is defined by use of the integral
D(r)
Z r
p(x)dx
hence
Z r
In differential form
252
p(x) dx =
Z f (r)
g(z) dz
g(z)dz = p(x)dx
[ASSUMING f is Borel measurable, i.e. has an inverse that is a Borel Set...]
since x = f (1) (z), one obtains
g(z)dz = p f (1) (z) d f (1) (z)
Now, the derivative of an inverse function
f (1) (z) =
1
,
f 0 ( f 1 (z))
(14.1)
| f 0 (u)|u = f (1) (z)
f (x + x)) f (x). Let us simplify with sole condition, assuming f(.) twice differentiable, x2 0
for all values of x in the convex case and <0 in the concave one. [WILL DISCUSS OTHER
CASES WHERE WE NEED TO SPLIT THE R.V. IN TWO DOMAINS BECAUSE INVERSE NOT
UNIQUE]
Some Examples.
Squaring x: p(x) is a Gaussian(with mean 0, standard deviation 1) , f(x)= x2
x
e 2
g(x) = , x > 0
2 2 x
which corresponds to the Chi-square distribution with 1 degrees of freedom.
Exponentiating x :p(x) is a Gaussian(with mean , standard deviation )
g(x) =
(log(x))2
22
2x
253
0.30
0.25
Figure 14.2: Simulation, first. The distribution of the utility of changes of wealth,
when the changes in wealth follow a power
law with tail exponent =2 (5 million Monte
Carlo simulations).
0.20
0.15
0.10
0.05
-20
-15
-10
-5
10
14.3
There is a conflation of fat-tailedness of Wealth and Utility: Happiness ( f (x))does not have the
same statistical properties as wealth (x)
Case 1: The Kahneman Tversky Prospect theory, which is convex-concave
v(x) =
xa
x0
( x a ) x < 0
+x2
B
+1
2
1
2, 2
R1
Where B is the Euler Beta function, B(a, b) = (a)(b)/(a + b) = 0 t a1 (1 t)b1 dt; we get
(skipping the details of z= v(u) and f(u) du = z(x) dx), the distribution z(x) of the utility of
happiness v(x)
+1
1 a
2
x a
2
/
a
+x
x0
a B( 2 , 2 )
+1
z(x |, a, ) =
2
1 a
x ) a
(
2/ a
+( x )
x<0
1
a B( 2 , 2 )
Fragility: as defined in the Taleb-Douady (2012) sense, on which later, i.e. tail sensitivity
below K, v(x) is less fragile than x.
v(x) has thinner tails than x more robust.
254
0.35
Distribution of x
0.30
0.25
0.20
Figure 14.3: The same result derived analytically, after the Monte Carlo runs.
0.15
Distribution of V(x)
0.10
0.05
-20
10
-10
20
0.020
Tail of x
0.015
0.010
Tail of v(x)
0.005
-18
-16
-14
-12
-10
-8
-6
255
-1
-1
-2
-3
-4
-5
-6
ASYMPTOTIC TAIL More technically the asymptotic tail for V(x) becomes
z( x ) K x a 1
We can see that V(x) can easily have finite variance when x has an infinite one. The dampening
of the tail has an increasingly consequential effect for lower values of .
Case 2: Compare to the Monotone concave of Classical Utility
Unlike the convex-concave shape in Kahneman Tversky, classical utility is monotone concave.
This leads to plenty of absurdities, but the worst is the effect on the distribution of utility.
Granted one (K-T) deals with changes in wealth, the second is a function of wealth.
The Kahneman-Tversky Prospect function is part of the class of the generalized "S curves",
bounded on the left or the right, or with waning derivatives that effectively gives the same
result of boundedness (or more exactly, soft-boundedness). Utility theory is unbounded on
the left, with preserving derivatives.
Researchers tend to consider that K-T is "empirical" while general utility is "normative".
This distinction is absurd: K-T corresponds to normal left-boundedness, as with all doseresponses present in nature. For an organism, death is the bounded worst outcome. You
cannot keep having detriment unless you have infinite resources or capabilities.
The next section illustrates the more general statement that anything with concave-left
exposure has a mean that degrades enormously under increases in uncertainty. Utility
theory, under transformation, gives absurd results mathematically.
Take the standard concave utility function g(x)= 1- e a x . With a=1
The distribution of v(x) will be
e
v(x) =
256
(+log(1 x))2
22
2(x 1)
0.6
0.5
0.4
0.3
0.2
0.1
x
-10
-8
-6
-4
-2
Which can be tolerated owing to the rapid drop in probabilities in the Gaussian tail. But with
a fatter tailed distribution, such as the standard powerlaw (a Student T Distribution) (Gabaix,
2008,[47]), where is the tail exponent,
! +1
2
x
v(x) =
(log(1 x)1)2
+
a2
(a ax)B
1
2, 2
14.4
High Uncertainty
Outcome
2
2
257
14.5
A simple way to see the point that convex payoffs have larger estimation errors: the Ilmanen
study assumes that one can derive strong conclusions from a single historical path not taking
into account sensitivity to counterfactuals and completeness of sampling. It assumes that what
one sees from a time series is the entire story. 1
Probability
Outcomes
Figure 1: The Small Sample Effect and Naive Empiricism: When one looks at historical
returns that are skewed to the left, most missing observations are in the left tails, causing an
overestimation of the mean. The more skewed the payoff, and the thicker the left tail, the worst
the gap between observed and true mean.
Now of concern for us is assessing the stub, or tail bias, that is, the difference between M
and M*, or the potential contribution of tail events not seen in the window used for the analysis.
When the payoff in the tails is powerful from convex responses, the stub becomes extremely
large. So the rest of this note will go beyond the Ilmanen (2012) to explain the convexities of the
payoffs in the tails and generalize to classical mistakes of testing strategies with explosive tail
exposures on a finite simple historical sample. It will be based on the idea of metaprobability
(or metamodel): by looking at effects of errors in models and representations. All one needs is
an argument for a very small probability of a large payoff in the tail (devastating for the option
seller) to reverse long shot arguments and make it uneconomic to sell a tail option. All it takes
is a small model error to reverse the argument.
The Nonlineatities of Option Packages There is a compounding effect of rarety of tail events
and highly convex payoff when they happen, a convexity that is generally missed in the literature.
To illustrate the point, we construct a return on theta (or return on time-decay) metric for a
delta-neutral package of an option, seen at t0 o given a deviation of magnitude NK .
1 The same flaw, namely missing convexity, is present in Bodarenko ??.
258
1
S0 ,t0 ,
where 0 (S0 , K, T t0 , K )is the European option price valued at time t0 off an initial asset
value S0 , with a strike price K, a final expiration at time T, and priced using an implied
standard deviation K . The payoff of is the same whether O is a put or a call, owing to the
delta-neutrality by hegding using a hedge ratio S0 ,t0 (thanks to put-call parity, S0 ,t0 is negative
if O is a call and positive otherwise). S0 ,t0 is the discrete change in value of the option over a
time increment (changes of value for an option in the absence of changes in any other variable).
With the increment = 1/252, this would be a single business day. We assumed interest rate
are 0, with no loss of generality (it would be equivalent of expressing the problem under a
risk-neutral measure). What 14.2 did is re-express the Fokker-Plank-Kolmogorov differential
equation (Black Scholes), in discrete terms, away from the limit of 0. In the standard BlackScholes World, the expectation of (N,K) should be zero, as N follows a Gaussian distribution
with mean 2 . But we are not about the Black Scholes world and we need to examine payoffs
to potential distributions. The use of K neutralizes the effect of expensive for the option as
we will be using a multiple of K as N standard deviations; if the option is priced at 15.87%
volatility, then one standard deviation would correspond to a move of about 1%, e
14.5.1 Convexity and Explosive Payoffs
Of concern to us is the explosive
nonlinearity in the tails. Let us examine the payoff of across
K
many values of K = S0 e
, in other words how many sigmas away from the money the
strike is positioned. A package about 20 out of the money , that is, =20, the crash of 1987
would have returned 229,000 days of decay, compensating for > 900 years of wasting premium
waiting for the result. An equivalent reasoning could be made for subprime loans. From this we
can assert that we need a minimum of 900 years of data to start pronouncing these options 20
standard deviations out-of-the money expensive, in order to match the frequency that would
deliver a payoff, and, more than 2000 years of data to make conservative claims. Clearly as we
can see with =0, the payoff is so linear that there is no hidden tail effect.
8000
PHNL
L = 20
6000
4000
L = 10
2000
L=0
5
10
15
N
20
Figure 2: Returns for package (N,K= S0 Exp[ K ] ) at values of = 0,10,20 and N, the
conditional sigma deviations.
259
0.8
0.6
0.4
0.2
-800
-600
-400
-200
400 000
300 000
200 000
100 000
10
15
20
Figure 3: The extreme convexity of an extremely out of the money option, with =20
Visibly the convexity is compounded by the fat-tailedness of the process: intuitively a convex
transformation of a fat-tailed process, say a powerlaw, produces a powerlaw of considerably
fatter tails. The Variance swap for instance results in 12 the tail exponent of the distribution of
the underlying security, so it would have infinite variance with tail 32 off the cubic exonent
discussed in the literature (Gabaix et al,2003; Stanley et al, 2000) -and some out-of-the money
options are more convex than variance swaps, producing tail equivalent of up to 51 over a broad
range of fluctuations.
For specific options there may not be an exact convex transformation. But we can get a Monte
Carlo simulation illustrating the shape of the distribution and visually showing how skewed it
is.
2
Fragility Heuristic and Nonlinear Exposure to Implied Volatility Most of the losses from
option portfolios tend to take place from the explosion of implied volatility, therefore acting as
if the market had already experienced a tail event (say in 2008). The same result as Figure 3
can be seen for changes in implied volatility: an explosion of volatility by 5 results in a 10
2 This convexity effect can be mitigated by some dynamic hedges, assuming no gaps but, because of local time for
stochastic processes; in fact, some smaller deviations can carry the cost of larger ones: for a move of -10 sigmas followed
by an upmove of 5 sigmas revision can end up costing a lot more than a mere -5 sigmas.Tail events can come from a
volatile sample path snapping back and forth.
260
ATM
=5
= 10
= 20
2
2
5
27
7686
3
3
10
79
72741
4
4
16
143
208429
option gaining 270 (the VIx went up > 10 during 2008). (In a well publicized debacle,
the speculator Niederhoffer went bust because of explosive changes in implied volatility in his
option portfolio, not from market movement; further, the options that bankrupted his fund
ended up expiring worthless weeks later).
The Taleb and Douady (2012)[105] , Taleb Canetti et al (2012)[100] fragility heuristic identifies
convexity to significant parameters as a metric to assess fragility to model error or representation:
by theorem, model error maps directly to nonlinearity of parameters. The heuristic corresponds
to the perturbation of a parameter, say the scale of a probability distribution and looks at the
effect of the expected shortfall; the same theorem asserts that the asymmetry between gain and
losses (convexity) maps directly to the exposure to model error and to fragility. The exercise
allows us to re-express the idea of convexity of payoff by ranking effects.
261
15
A N U N C E R TA I N T Y A P P R OA C H TO
FRAGILITY
15.1
This section is incomplete; it will present a general review the literature and the variety of
definitions of "robustness" in:
Dynamical systems Economic modelling (see Hansen and Sargent) We explain why what we do is add a nonlinear dimension to the loss models and solve some
of the issues with loss models since our funtion (nonlinearity of exposure) solves the problem
of mini-max with unbounded losses, and how our payoff function ("exposure") maps to the loss
function.
15.2
introduction
In short, fragility as we define it is related to how a system suffers from the variability of its
environment beyond a certain preset threshold (when threshold is K, it is called K-fragility),
while antifragility refers to when it benefits from this variability in a similar way to vega of
an option or a nonlinear payoff, that is, its sensitivity to volatility or some similar measure of
scale of a distribution.
We are not saying that are no other definitions and representations of fragility (although we
could not find any that clashes with, or does not fit within our variability approach). Our point
is that such a definition allows us to perform analyses based on nonlinearity. Our method is, in a
263
Breaking Point
Dose
Response
Deterministic and
invariant
Broken Glass
stressor
Stochastic
Stressor
Higher dispersion of
stressor increases probab.
of breaking.
Dose
Figure 15.1: Illustrates why the coffee cup is fragile because it doesnt like variability. Imagine a stressor
of intensity k. A deterministic stressor at k will always be more favorable than a stochastic stressor with
The cup breaks at k + , even if there are compensating effects at k that lower the average. The
average k.
given the same average, the higher the probability of breakage. This illustrates
more dispersion around k,
the dependence of fragility on dispersion ("vega") and the theorems showing how fragility lies in the second
order effect. [INCLUDE STOCHASTIC RESONANCE AS AN OPPOSITE EFFECT]
way, inverse "real option" theory ([117],[1]), by with studies of contingent claims are generalized
to all decision-making under uncertainty that entail asymmetry of payoff.
Simply, a coffee cup on a table suffers more from large deviations than from the cumulative
effect of some shocksconditional on being unbroken, it has to suffer more from tail events
264
15.2 introduction
Prob Density
HK, s- + Ds- L =
HK, s- L =
Hx - WL f
Hx - WL f
Hs_ L
Hs_ +Ds- L
HxL x
HxL x
K
Figure 15.2: A definition of fragility as left tail-vega sensitivity; the figure shows the effect of the perturbation of the lower semi-deviation s on the tail integral of (x ) below K, being a centering constant.
Our detection of fragility does not require the specification of f the probability distribution.
than regular ones around the center of the distribution, the at the money category. This is
the case of elements of nature that have survived: conditional on being in existence, then the
class of events around the mean should matter considerably less than tail events, particularly
when the probabilities decline faster than the inverse of the harm, which is the case of all used
monomodal probability distributions. Further, what has exposure to tail events suffers from
uncertainty; typically, when systems a building, a bridge, a nuclear plant, an airplane, or a
bank balance sheet are made robust to a certain level of variability and stress but may fail
or collapse if this level is exceeded, then they are particularly fragile to uncertainty about the
distribution of the stressor, hence to model error, as this uncertainty increases the probability of
dipping below the robustness level, bringing a higher probability of collapse. In the opposite
case, the natural selection of an evolutionary process is particularly antifragile, indeed, a more
volatile environment increases the survival rate of robust species and eliminates those whose
superiority over other species is highly dependent on environmental parameters.
Figure 15.2 show the tail vega sensitivity of an object calculated discretely at two different
lower absolute mean deviations. We use for the purpose of fragility and antifragility, in place of
measures in L2 such as standard deviations, which restrict the choice of probability distributions,
the broader measure of absolute deviation, cut into two parts: lower and upper semi-deviation
above the distribution center .
This article aims at providing a proper mathematical definition of fragility, robustness, and
antifragility and examining how these apply to different cases where this notion is applicable. 1
1 Hansen and Sargent, in [54]:"A long tradition dating back to Friedman (...) advocates framing macroeconomic policy
rules and interpreting econometric findings in light of doubts about model specification, though how th0se doubts have
been formalized in practice has varied". In fact what we are adding to the story as far as economics here is local and
global convexity of the variation and the asymmetry in one side or the other.
265
266
15.2 introduction
application of the transfer function to the Kahneman-Tversky value function, convex in the
negative domain and concave in the positive one, shows that its decreases fragility in the left tail
(hence more robustness) and reduces the effect of the right tail as well (also more robustness),
which allows to assert that we are psychologically more robust to changes in wealth than
implied from the distribution of such wealth, which happens to be extremely fat-tailed.
Accordingly, our approach relies on nonlinearity of exposure as detection of the vega-sensitivity,
not as a definition of fragility. And nonlinearity in a source of stress is necessarily associated
with fragility. Clearly, a coffee cup, a house or a bridge dont have psychological preferences,
subjective utility, etc. Yet they are concave in their reaction to harm: simply, taking z as a stress
level and (z) the harm function, it suffices to see that, with n > 1,
(nz) < n (z) for all 0 < n z < Z
where Z is the level (not necessarily specified) at which the item is broken. Such inequality
leads to (z) having a negative second derivative at the initial value z.
So if a coffee cup is less harmed by n times a stressor of intensity Z than once a stressor
of nZ, then harm (as a negative function) needs to be concave to stressors up to the point of
breaking; such stricture is imposed by the structure of survival probabilities and the distribution
of harmful events, and has nothing to do with subjective utility or some other figments. Just as
with a large stone hurting more than the equivalent weight in pebbles, if, for a human, jumping
one millimeter caused an exact linear fraction of the damage of, say, jumping to the ground from
thirty feet, then the person would be already dead from cumulative harm. Actually a simple
computation shows that he would have expired within hours from touching objects or pacing
in his living room, given the multitude of such stressors and their total effect. The fragility
that comes from linearity is immediately visible, so we rule it out because the object would
be already broken and the person already dead. The relative frequency of ordinary events
compared to extreme events is the determinant. In the financial markets, there are at least ten
thousand times more events of 0.1% deviations than events of 10%. There are close to 8,000
micro-earthquakes daily on planet earth, that is, those below 2 on the Richter scale about 3
million a year. These are totally harmless, and, with 3 million per year, you would need them
to be so. But shocks of intensity 6 and higher on the scale make the newspapers. Accordingly,
we are necessarily immune to the cumulative effect of small deviations, or shocks of very small
magnitude, which implies that these affect us disproportionally less (that is, nonlinearly less)
than larger ones.
Model error is not necessarily mean preserving. s- , the lower absolute semi-deviation does not
just express changes in overall dispersion in the distribution, such as for instance the scaling
case, but also changes in the mean, i.e. when the upper semi-deviation from to infinity
is invariant, or even decline in a compensatory manner to make the overall mean absolute
deviation unchanged. This would be the case when we shift the distribution instead of rescaling
it. Thus the same vega-sensitivity can also express sensitivity to a stressor (dose increase) in
medicine or other fields in its effect on either tail. Thus s (l) will allow us to express the
negative sensitivity to the disorder cluster (see Antifragile): i) uncertainty, ii) variability, iii)
imperfect, incomplete knowledge, iv) chance, v) chaos, vi) volatility, vii) disorder, viii) entropy,
ix) time, x) the unknown, xi) randomness, xii) turmoil, xiii) stressor, xiv) error, xv) dispersion of
outcomes.
Detection Heuristic Finally, thanks to the transfer function, this paper proposes a risk heuristic
that "works " in detecting fragility even if we use the wrong model/pricing method/probability
267
Figure 15.3: Disproportionate effect of tail events on nonlinear exposures, illustrating the necessary character of the nonlinearity of the harm function and showing how we can extrapolate outside the model to
probe unseen fragility.
268
15.2 introduction
Antifragility
It is not quite the mirror image of fragility, as it implies positive vega above some threshold
in the positive tail of the distribution and absence of fragility in the left tail, which leads to a
distribution that is skewed right. Table 24 introduces the Exhaustive Taxonomy of all Possible
Payoffs y = f (x)
Table 24: Payoffs and Mixed Nonlinearities
Condition
Left Tail
(Loss Domain)
Right Tail
(Gain Domain)
Nonlinear
Payoff Function y = f (x)
"derivative"
where x is
a
random
variable
Derivatives
Equivalent
Effect
of
fatailedness of f (x)
compared
to primitive
x.
Fragile
(type 1)
Fat (regular
or
absorbing
barrier)
Fat
Mixed
cave
convex
(fence)
Long
upvega, short
down-vega
Fragile
(type 2)
Robust
Thin
Thin
concave
Short vega
Thin
Thin
Antifragile
Thin
Fat
(thicker
than left)
Mixed convex
left, concave
right (digital,
sigmoid)
Convex
More
fragility
if
absorbing barrier,
neutral
otherwise
More
fragility
No effect
conleft,
right
Long vega
More
antifragility
The central Table, Table 1 introduces the exhaustive map of possible outcomes, with 4 mutually
exclusive categories of payoffs. Our steps in the rest of the paper are as follows: a. We provide
a mathematical definition of fragility, robustness and antifragility. b. We present the problem
of measuring tail risks and show the presence of severe biases attending the estimation of small
probability and its nonlinearity (convexity) to parametric (and other) perturbations. c. We
express the concept of model fragility in terms of left tail exposure, and show correspondence to
the concavity of the payoff from a random variable. d. Finally, we present our simple heuristic
to detect the possibility of both fragility and model error across a broad range of probabilistic
estimations.
Conceptually, fragility resides in the fact that a small or at least reasonable uncertainty
on the macro-parameter of a distribution may have dramatic consequences on the result of a
given stress test, or on some measure that depends on the left tail of the distribution, such as an
out-of-the-money option. This hypersensitivity of what we like to call an out of the money put
price to the macro-parameter, which is some measure of the volatility of the distribution of the
underlying source of randomness.
Formally, fragility is defined as the sensitivity of the left-tail shortfall (non-conditioned by
probability) below a certain threshold K to the overall left semi-deviation of the distribution.
Examples
269
How spreading risks are dangerous compared to limited one we need to weave into the
derivations the notion of risk spreading as a non-concave response to make links clearer.
Why error is a problem in the presence of nonlinearity.
Why polluting "a little" is qualitatively different from pollution "a lot".
Eventually, why fat tails arise from accelerating response.
270
16
Chapter Summary 16: Presents the fragility theorems and the transfer function between nonlinear response and the benefits and harm from increased uncertainty.
The following offers a formal definition of fragility as "vega", negative expected response from
uncertainty.
16.1
We construct a measure of "vega", that is, the sensitivity to uncertainty, in the left tails of the
distribution that depends on the variations of s the semi-deviation below a certain level W,
chosen in the L1 norm in order to ensure its existence under "fat tailed" distributions with finite
first semi-moment. In fact s would exist as a measure even in the case of undefined moments to
the right side of W.
Let X be a random variable, the distribution of which is one among a one-parameter family
of pdf, f , I R. We consider a fixed reference value and, from this reference, the "raw"
left-semi-absolute deviation:1
s () =
( x) f (x)dx
(16.1)
We assume that s () is continuous, strictly increasing and spans the whole range R+ =
[0, +), so that we may use the left-semi-absolute deviation s as a parameter by considering
the inverse function (s) : R+ I, defined by s ((s)) = s for s R+ .
This condition is for instance satisfied if, for any given x < , the probability is a continuous
and increasing function of . Indeed, denoting
F (x) = Pf (X < x) =
an integration by parts yields:
s () =
Z x
f (t) dt,
(16.2)
F (x) dx
This is the case when is a scaling parameter, i.e., X + (X1 ) indeed one has in this
case
x
,
F (x) = F1 +
1 We use a measure related to the left-semi absolute deviation, or roughly the half the mean absolute deviation (the part
in the negative domain) because 1) distributions are not symmetric and might have changes on the right of that are
not of interest to us, 2) standard deviation requires finite second moment.
Further, we do not adjust s by its probability with no loss of generality. Simply, probability in the negative domain
is close to 21 and would not change significantly in response to changes in parameters. Probabilities in the tails are
nonlinear to changes, not those in the body of the distribution.
271
2
It is also the case when is a shifting parameter, i.e. X X0 , indeed, in this case
Z K
( x) f (s ) (x)dx
(16.3)
V(X, f , K, s ) =
(K, s ) =
s
Z
( x)
f)
dx
ds
d
1
(16.4)
As in the many practical instances where threshold effects are involved, it may occur that
does not depend smoothly on s . We therefore also define a finite difference version of the
vega-sensitivity as follows:
1
(K, s + s) (K, s s)
2s
Z K
f (s + s)(x) f (s s)(x)
( x)
=
dx
2s
V(X, f , K, s ) =
(16.5)
P (K) =
Z K
(16.6)
(K x) f (x) dx
(16.7)
Where the first part ( K)F (K) is proportional to the probability of the variable being below
the stress level K and the second part P (K) is the expectation of the amount by which X is
below K (counting 0 when it is not). Making a parallel with financial options, while s () is a
put at-the-money, (K,s ) is the sum of a put struck at K and a digital put also struck at K with
amount K; it can equivalently be seen as a put struck at with a down-and-in European
barrier at K.
Letting = (s ) and integrating by part yields
Z
K, s () = ( K)F (K) +
F (x)dx =
Z
272
FK (x) dx
(16.8)
Figure 16.1: The different curves of F (K) and F0 (K) showing the difference in sensitivity to changes at
different levels of K.
(K, s )
s
FK
(x) dx
R F
(x) dx
R
=
(16.9)
1
2 s
K
F,s
(x)dx
(16.10)
K
K
where +s and
s are such that s(s ) = s + s, s(s ) = s s and F,s (x) = F + (x)
FK (x).
273
Definition of Fragility: The Intrinsic Case The local fragility of a random variable X depending
on parameter , at stress level K and semi-deviation level s () with pdf f is its K-left-tailed semi-vega
sensitivity V(X, f , K, s ).
The finite-difference fragility of X at stress level K and semi-deviation level s () s with pdf f
is its K-left-tailed finite-difference semi-vega sensitivity V(X, f , K, s , s).
In this definition, the fragility relies in the unsaid assumptions made when extrapolating the
distribution of X from areas used to estimate the semi-absolute deviation s (), around , to
areas around K on which the risk measure depends.
Definition of Fragility: The Inherited Case Next we consider the particular case where a random variable Y = (X) depends on another source of risk X, itself subject to a parameter . Let
us keep the above notations for X, while we denote by g the pdf of Y,Y = () and u () the
left-semi-deviation of Y. Given a strike level
L = (K), let us define, as in the case of X :
Z
L, u () =
(Y y)g (y) dy
(16.11)
The inherited fragility of Y with respect to X at stress level L = (K) and left-semi-deviation level
s () of X is the partial derivative:
VX Y, g , L, s () =
L, u () =
s
Z
g
(Y Y) (y)dy
ds
d
1
(16.12)
Note that the stress level and the pdf are defined for the variable Y, but the parameter which
is used for differentiation is the left-semi-absolute deviation of X, s (). Indeed, in this process,
one first measures the distribution of X and its left-semi-absolute deviation, then the function
is applied, using some mathematical model of Y with respect to X and the risk measure
is estimated. If an error is made when measuring s (), its impact on the risk measure of Y is
amplified by the ratio given by the inherited fragility.
Once again, one may use finite differences and define the finite-difference inherited fragility
of Y with respect to X, by replacing, in the above equation, differentiation by finite differences
between values + and , where s (+ ) = s + s and s ( ) = s s.
274
16.2
Let us study the case of a random variable Y = (X); the pdf g of which also depends on
parameter , related to a variable X by the nonlinear function . We are now interested in
comparing their intrinsic fragilities. We shall say, for instance, that Y is more fragilefragile at the
stress level L and left-semi-deviation level u () than the random variable X, at stress level K
and left-semi-deviation level s () if the L-left-tailed semi-vega sensitivity of Y is higher than
the K-left-tailed semi-vega sensitivity of X :
V(Y, g , L, ) > V(X, f , K, s )
(16.13)
One may use finite differences to compare the fragility of two random variables:V(Y, g , L, ) >
V(X, f , K, s). In this case, finite variations must be comparable in size, namely u/u = s/s .
Let us assume, to start, that is differentiable, strictly increasing and scaled so that Y =
(16.14)
Hence, if (L, u ) denotes the equivalent of (K), s with variable (Y, g ) instead of (X, f ), we
have:
Z
L, u () =
FK (x)
d
(x)dx
dx
(16.15)
FK (x)
d
(x) dx
dx
(16.16)
V Y, g , L, u () =
(16.17)
1
2u
K
F,u
(x)
d
(x)dx
dx
(16.18)
K (x) = F K (x)
Where +u and
are such that u(+u ) = u + u, u(+u ) = u u and F,u
+
u
FK (x).
u
Next, Theorem 1 proves how a concave transformation (x) of a random variable x produces
fragility.
Fragility Transfer Theorem
Theorem 16.1.
Let, with the above notations, : R R be a twice differentiable function such that () = and for
d
any x < , dx (x) > 0. The random variable Y = (X) is more fragile at level L = (K) and pdf g than
X at level K and pdf f if, and only if, one has:
275
HK (x)
d2
(x)dx < 0
dx2
Where
HK (x)
PK
(x)
PK
()
P
(x)
P
()
(16.19)
and where
P (x) =
Z x
F (t)dt
(16.20)
Z x
FK (t)dt
coincides
PK
()
in K
with the vertical axis. It increases with K, from 0 when K to a value above
276
PK
P
()
when
Figure 16.2: The Transfer function H for different portions of the distribution: its sign flips in the region
slightly below
PK
On the other hand, if K < then () < () and G (K) < GK (K), which implies that
G (x) < GK (x) for x K. An elementary convexity analysis shows that, in this case, the equation
G (x) = GK (x) has a unique solution with l ambda < < . The transfer function HK (x)
is positive for x < , in particular when x and negative for < x < .
Scaling Parameter
We assume here that is a scaling parameter, i.e. X = + (X1 ). In this case, as we saw
above, we have
1
x
x
f (x) = f 1 +
, F (x) = F1 +
x
P (x) = P1 +
and s () = s (1).
Hence
K
(K, s ()) = ( K)F1 +
K
+ P1 +
(16.21)
1
2
P
(K)
+
(
K)F
(K)
+
(
K)
f
(K)
s ()
(16.22)
1
(K, s ) =
(K, )
s
s (1)
=
277
Figure 16.3: The distribution of G and the various derivatives of the unconditional shortfalls
values of X. The scaling applies to small negative values of the transformed variable Y with a
d
d
coefficient dx (0), but large negative values are subject to a different coefficient dx (K), which can
potentially be very different.
16.3
fragility drift
To summarize, textitFragility is defined at as the sensitivity i.e. the first partial derivative of
the tail estimate with respect to the left semi-deviation s . Let us now define the fragility drift:
VK0 (X, f , K, s ) =
2
(K, s )
Ks
(16.23)
In practice, fragility always occurs as the result of fragility, indeed, by definition, we know that
(, s ) = s , hence V(X, f , , s ) = 1. The fragility drift measures the speed at which fragility
departs from its original value 1 when K departs from the center .
16.3.1 Second-order Fragility
The second-order fragility is the second order derivative of the tail estimate with respect to the
semi-absolute deviation s :
Vs0 (X, f , K, s ) =
(s )2
(K, s )
As we shall see later, the second-order fragility drives the bias in the estimation of stress tests when
the value of s is subject to uncertainty, through Jensens inequality.
278
16.4
Antifragility is not the simple opposite of fragility, as we saw in Table 1. Measuring antifragility,
on the one hand, consists of the flipside of fragility on the right-hand side, but on the other
hand requires a control on the robustness of the probability distribution on the left-hand side.
From that aspect, unlike fragility, antifragility cannot be summarized in one single figure but
necessitates at least two of them.
When a random variable depends on another source of randomness: Y = (X ), we shall
study the antifragility of Y with respect to that of X and to the properties of the function .
(16.24)
K1 6 K 0 6 K2
V(X , f , K 0 , s ()).
(16.25)
Note that the lower R, the tighter the control and the more robust the distribution f .
Once again, the definition of b-robustness can be transposed, using finite differences V(X , f , K, s (), s).
In practical situations, setting a material upper bound b to the fragility is particularly important: one need to be able to come with actual estimates of the impact of the error on the
estimate of the left-semi-deviation. However, when dealing with certain class of models, such as
Gaussian, exponential of stable distributions, we may be lead to consider asymptotic definitions
of robustness, related to certain classes.
For instance, for a given decay exponent a > 0, assuming that f (x) = O(eax ) when x ,
the a-exponential asymptotic robustness of X below the level K is:
Rexp (X , f , K, s (), a)
0
= max e a(K ) V(X , f , K 0 , s ())
K 0 6K
(16.26)
If one of the two quantities e a(K ) f (K 0 ) or e a(K ) V(X , f , K 0 , s ()) is not bounded from
above when K , then Rexp = + and X is considered as not a-exponentially robust.
279
Rpow (X , f , K, s (), a) =
2
max ( K 0 ) V(X , f , K 0 , s ())
K 0 6K
( K 0 ) f (K 0 )
( K 0 )2 V(X , f , K 0 , s ())
is not bounded from above when K 0 , then R pow = + and X is considered as not power robust. Note the exponent 2 used with the fragility, for homogeneity reasons, e.g. in
the case of stable distributions, when a random variable Y = (X ) depends on another source
of risk X .
Definition 16.1.
Left-Robustness (monomodal distribution). A payoff y = (x) is said (a, b)-robust below L = (K) for a
source of randomness X with pdf f assumed monomodal if, letting g be the pdf of Y = (X), one has,for
any K 0 K and L = (K):
VX Y, g , L0 , s () 6 aV X, f , K 0 , s () + b
(16.27)
The quantity b is of order deemed of negligible utility (subjectively), that is, does not exceed
some tolerance level in relation with the context, while a is a scaling parameter between variables
X and Y.
Note that robustness is in effect impervious to changes of probability distributions. Also
note that this measure of robustness ignores first order variations since owing to their higher
frequency, these are detected (and remedied) very early on.
Example of Robustness (Barbells):
a. trial and error with bounded error and open payoff
b. for a "barbell portfolio " with allocation to numeraire securities up to 80% of portfolio, no
perturbation below K set at 0.8 of valuation will represent any difference in result, i.e. q = 0.
The same for an insured house (assuming the risk of the insurance company is not a source of
variation), no perturbation for the value below K, equal to minus the insurance deductible, will
result in significant changes.
c. a bet of amount B (limited liability) is robust, as it does not have any sensitivity to perturbations below 0.
16.4.2 Antifragility
The second condition of antifragility regards the right hand side of the distribution. Let us define
the right-semi-deviation of X :
s+ () =
And, for H > L > :
280
Z +
(x ) f (x)dx
+ (L, H, s+ ()) =
W(X, f , L, H, s+ ) =
Z H
L
(x ) f (x)dx
+ (L, H, s+ )
s+
=
Z
H
L
f
(x ) (x)dx
Z
f
(x ) (x)dx
1
(H)
(L)
g
(y ()) (y)dy
Z
f
(x ) (x)dx
1
(16.28)
u+ ()
s+ ()
The scaling constant a provides homogeneity in the case where the relation between X and y is
linear. In particular, nonlinearity in the relation between X and Y impacts robustness.
The second condition can be replaced with finite differences u and s, as long as u/u = s/s.
16.4.3 Remarks
Fragility is K-specific We are only concerned with adverse events below a certain pre-specified
level, the breaking point. Exposures A can be more fragile than exposure B for K = 0, and
much less fragile if K is, say, 4 mean deviations below 0. We may need to use finite s to avoid
situations as we will see of vega-neutrality coupled with short left tail.
Effect of using the wrong distribution or model Comparing V(X, f , K, s , s) and the alternative distribution V(X, f , K, s , s), where f is the "true" distribution, the measure of fragility
provides an acceptable indication of the sensitivity of a given outcome such as a risk measure
to model error, provided no paradoxical effects perturbate the situation. Such paradoxical
effects are, for instance, a change in the direction in which certain distribution percentiles react
to model parameters, like s . It is indeed possible that nonlinearity appears between the core
part of the distribution and the tails such that when s increases, the left tail starts fattening
giving a large measured fragility then steps back implying that the real fragility is lower than
the measured one. The opposite may also happen, implying a dangerous under-estimate of the
fragility. These nonlinear effects can stay under control provided one makes some regularity assumptions on the actual distribution, as well as on the measured one. For instance, paradoxical
effects are typically avoided under at least one of the following three hypotheses:
1. The class of distributions in which both f and f* are picked are all monomodal, with
monotonous dependence of percentiles with respect to one another.
281
282
17
A P P L I C AT I O N S TO M O D E L E R R O R
In the cases where Y depends on X, among other variables, often x is treated as non-stochastic,
and the underestimation of the volatility of x maps immediately into the underestimation of the
left tail of Y under two conditions:
1. X is stochastic and its stochastic character is ignored (as if it had zero variance or mean
deviation)
2. Y is concave with respect to X in the negative part of the distribution, below
"Convexity Bias " or Jensens Inequality Effect: Further, missing the stochasticity under the
two conditions a) and b) , in the event of the concavity applying above leads to the negative
convexity bias from the lowering effect on the expectation of the dependent variable Y.
17.0.5 Example:Application to Budget Deficits
Example: A government estimates unemployment for the next three years as averaging 9%;
it uses its econometric models to issue a forecast balance B of 200 billion deficit in the local
currency. But it misses (like almost everything in economics) that unemployment is a stochastic
variable. Employment over 3 years periods has fluctuated by 1% on average. We can calculate
c Unemployment at 8% , Balance B(8%) = -75 bn
the effect of the error with the following: A
c Unemployment at 9%, Balance B(9%)= -200 bn A
c Unemployment
(improvement of 125bn) A
at 10%, Balance B(10%)= 550 bn (worsening of 350bn)
The convexity bias from underestimation of the deficit is by -112.5bn, since
B(8%) + B(10%)
= 312.5
2
Further look at the probability distribution caused by the missed variable (assuming to simplify
deficit is Gaussian with a Mean Deviation of 1% )
Adding Model Error and Metadistributions: Model error should be integrated in the distribution as a stochasticization of parameters. f and g should subsume the distribution of all
possible factors affecting the final outcome (including the metadistribution of each). The socalled "perturbation " is not necessarily a change in the parameter so much as it is a means to
verify whether f and g capture the full shape of the final probability distribution.
Any situation with a bounded payoff function that organically truncates the left tail at K will
be impervious to all perturbations affecting the probability distribution below K.
For K = 0, the measure equates to mean negative semi-deviation (more potent than negative
semi-variance or negative semi-standard deviation often used in financial analyses).
17.0.6 Model Error and Semi-Bias as Nonlinearity from Missed Stochasticity of Variables
Model error often comes from missing the existence of a random variable that is significant in
determining the outcome (say option pricing without credit risk). We cannot detect it using the
283
Figure 17.1: Histogram from simulation of government deficit as a left-tailed random variable as a result of
randomizing unemployment of which it is a convex function. The method of point estimate would assume
a Dirac stick at -200, thus underestimating both the expected deficit (-312) and the skewness (i.e., fragility)
of it.
heuristic presented in this paper but as mentioned earlier the error goes in the opposite direction
as model tend to be richer, not poorer, from overfitting.
But we can detect the model error from missing the stochasticity of a variable or underestimating its stochastic character (say option pricing with non-stochastic interest rates or ignoring
that the volatility can vary).
Missing Effects: The study of model error is not to question whether a model is precise or not,
whether or not it tracks reality; it is to ascertain the first and second order effect from missing
the variable, insuring that the errors from the model dont have missing higher order terms that
cause severe unexpected (and unseen) biases in one direction because of convexity or concavity,
in other words, whether or not the model error causes a change in z.
17.1
Having the right model (which is a very generous assumption), but being uncertain about the
parameters will invariably lead to an increase in model error in the presence of convexity and
nonlinearities.
As a generalization of the deficit/employment example used in the previous section, say we
are using a simple function:
f(x | )
284
() d
The mere fact that is uncertain (since it is estimated) might lead to a bias if we perturb
from the outside (of the integral), i.e. stochasticize the parameter deemed fixed. Accordingly,
the convexity bias is easily measured as the difference between a) f integrated across values of
potential and b) f estimated for a single value of deemed to be its average. The convexity
bias A becomes:
Z
Z Z
Z
( ) d )dx
(17.1)
A
f ( x | ) ( ) d dx
f (x
x
And B the missed fragility is assessed by comparing the two integrals below K, in order to
capture the effect on the left tail:
Z
Z K Z
Z K
B (K)
f ( x | ) ( ) d dx
f (x
( ) d )dx
(17.2)
Which can be approximated by an interpolated estimate obtained with two values of separated from a mid point by a mean deviation of and estimating
B (K)
Z K
1
( f ( x | + ) + f ( x | )) dx
Z K
f (x | ) dx
(17.3)
1
( f ( X | + ) + f ( X | )) f (X | )
2
(17.4)
So that
B (K) =
Z K
0 B (x)dx
(17.5)
0
which leads us to the fragility heuristic. In particular, if we assume that B (X) has a constant
sign for X K, then B (K) has the same sign.
The fragility heuristic is presented in the next Chapter.
285
18
Chapter Summary 17: Presents the IMF fragility heuristics, particularly in the improvement
of stress testing.
18.0.1 The Fragility/Model Error Detection Heuristic (detecting A and B when cogent)
18.1
Or detecting a and b when cogent. The famous firm Dexia went into financial distress a few
days after passing a stress test with flying colors.
If a bank issues a so-called "stress test" (something that has not proven very satisfactory), off
a parameter (say stock market) at -15%. We ask them to recompute at -10% and -20%. Should
the exposure show negative asymmetry (worse at -20% than it improves at -10%), we deem that
their risk increases in the tails. There are certainly hidden tail exposures and a definite higher
probability of blowup in addition to exposure to model error.
Note that it is somewhat more effective to use our measure of shortfall in Definition, but
the method here is effective enough to show hidden risks, particularly at wider increases (try
25% and 30% and see if exposure shows increase). Most effective would be to use power-law
distributions and perturb the tail exponent to see symmetry.
Example 2 (Detecting Tail Risk in Overoptimized System, B ). Raise airport traffic 10%,
lower 10%, take average expected traveling time from each, and check the asymmetry for nonlinearity. If asymmetry is significant, then declare the system as overoptimized. (Both A and
B as thus shown.
The same procedure uncovers both fragility and consequence of model error (potential harm
from having wrong probability distribution, a thin- tailed rather than a fat-tailed one). For
traders (and see Gigerenzers discussions, in Gigerenzer and Brighton (2009)[49], Gigerenzer
and Goldstein(1996)[50]) simple heuristics tools detecting the magnitude of second order effects
can be more effective than more complicated and harder to calibrate methods, particularly under
multi-dimensionality. See also the intuition of fast and frugal in Derman and Wilmott (2009)[23],
Haug and Taleb (2011)[56].
The Fragility Heuristic Applied to Model Error
1- First Step (first order). Take a valuation. Measure the sensitivity to all parameters p determining V over finite ranges p. If materially significant, check if stochasticity of parameter is taken
into account by risk assessment. If not, then stop and declare the risk as grossly mismeasured
(no need for further risk assessment). 2-Second Step (second order). For all parameters p compute the ratio of first to second order effects at the initial range p = estimated mean deviation.
H (p)
0
,
where
287
0 (p)
1
2
1
1
f p + p + f p p
2
2
2-Third Step. Note parameters for which H is significantly > or < 1. 3- Fourth Step: Keep
widening p to verify the stability of the second order effects.
18.2
288
18.3
stress tests
18.4
general methodology
289
19
19.1
Assume that someone tells you that the probability of an event is exactly zero. You ask him
where he got this from."Baal told me" is the answer. In such case, the person is coherent, but
would be deemed unrealistic by non-Baalists. But if on the other hand, the person tells you"I
estimated it to be zero," we have a problem. The person is both unrealistic and inconsistent.
Some- thing estimated needs to have an estimation error. So probability cannot be zero if it is
estimated, its lower bound is linked to the estimation error; the higher the estima- tion error, the
higher the probability, up to a point. As with Laplaces argument of total ignorance, an infinite
estimation error pushes the probability toward 21 .
We will return to the implication of the mistake; take for now that anything es- timating a
parameter and then putting it into an equation is different from estimat- ing the equation across
parameters (same story as the health of the grandmother, the average temperature, here "estimated" is irrelevant, what we need is average health across temperatures). And Markowitz
showed his incoherence by starting his "seminal" paper with "Assume you know E and V" (that
is, the expectation and the variance). At the end of the paper he accepts that they need to be
estimated, and what is worse, with a combination of statistical techniques and the "judgment
of practical men." Well, if these parameters need to be estimated, with an error, then the derivations need to be written differently and, of course, we would have no paperand no Markowitz
paper, no blowups, no modern finance, no fragilistas teaching junk to students. . . . Economic
models are extremely fragile to assumptions, in the sense that a slight alteration in these assumptions can, as we will see, lead to extremely consequential differences in the results. And,
to make matters worse, many of these models are "back-fit" to assumptions, in the sense that the
hypotheses are selected to make the math work, which makes them ultrafragile and ultrafragilizing.
19.2
For almost two hundred years, weve been talking about an idea by the economist David Ricardo
called "comparative advantage." In short, it says that a country should have a certain policy based
on its comparative advantage in wine or clothes. Say a country is good at both wine and clothes,
better than its neighbors with whom it can trade freely. Then the visible optimal strategy would
be to specialize in either wine or clothes, whichever fits the best and minimizes opportunity
costs. Everyone would then be happy. The analogy by the economist Paul Samuelson is that
if someone happens to be the best doctor in town and, at the same time, the best secretary,
then it would be preferable to be the higher earning doctor as it would minimize opportunity
lossesand let someone else be the secretary and buy secretarial ser- vices from him.
We agree that there are benefits in some form of specialization, but not from the models used
to prove it. The flaw with such reasoning is as follows. True, it would be inconceivable for a
291
Wine
Britain
100
110
Portugal
90
80
The logic The table above shows the cost of production, normalized to a selling price of one
unit each, that is, assuming that these trade at equal price (1 unit of cloth for 1 unit of wine).
What looks like the paradox is as follows: that Portugal produces cloth cheaper than Britain, but
should buy cloth from there instead, using the gains from the sales of wine. In the absence of
transaction and transportation costs, it is efficient for Britain to produce just cloth, and Portugal
to only produce wine.
The idea has always attracted economists because of its paradoxical and counterintuitive aspect. Clearly one cannot talk about returns and gains without discounting these benefits by
the offsetting risks. Many discussions fall into the critical and dangerous mistake of confusing function of average and average of function. Now consider the price of wine and clothes
variablewhich Ricardo did not assumewith the numbers above the unbiased average long-term
value. Further assume that they follow a fat-tailed distribution. Or consider that their costs of
production vary according to a fat-tailed distribution.
If the price of wine in the international markets rises by, say, 40 %, then there are clear benefits.
But should the price drop by an equal percentage, L40 %, then massive harm would ensue, in
magnitude larger than the benefits should there be an equal rise. There are concavities to the
exposuresevere concavities.
And clearly, should the price drop by 90 percent, the effect would be disastrous. Just imagine
what would happen to your household should you get an instant and unpredicted 40 percent
pay cut. Indeed, we have had problems in history with coun- tries specializing in some goods,
commodities, and crops that happen to be not just volatile, but extremely volatile. And disaster
does not necessarily come from varia- tion in price, but problems in production: suddenly, you
cant produce the crop be- cause of a germ, bad weather, or some other hindrance.
A bad crop, such as the one that caused the Irish potato famine in the decade around 1850,
caused the death of a million and the emigration of a million more (Irelands entire population
at the time of this writing is only about six million, if one includes the northern part). It is very
hard to reconvert resourcesunlike the case in the doctor-typist story, countries dont have the
ability to change. Indeed, monocul- ture (focus on a single crop) has turned out to be lethal in
historyone bad crop leads to devastating famines.
The other part missed in the doctor-secretary analogy is that countries dont have family and
friends. A doctor has a support community, a circle of friends, a collective that takes care of him,
a father-in-law to borrow from in the event that he needs to reconvert into some other profession,
292
293
Samuelson
mization
opti-
SOURCE OF FRAGILITY
Assuming knowledge of
the parameters, not integrating models across parameters, relying on (very
unstable) correlations. Assumes A (bias) and B
(fragility) = 0
Missing layer of randomness in the price of wine
may imply total reversal of allocation.
Assumes A (bias) and B
(fragility) = 0
Concentration of sources
of randomness under concavity of loss function. Assumes A (bias) and B
(fragility) = 0
Ludic fallacy: assumes
exhaustive knowledge of
outcomes and knowledge
of probabilities. Assumes
A (bias), B (fragility),
and C (antifragility) = 0
Missing
stochasticity
causing convexity effects.
Mostly considers LC
(antifragility) =0
REMEDY
1/n (spread as large a
number of exposures as
manageable),
barbells,
progressive and organic
construction, etc.
Distributed randomness
Use of metaprobabilities
changes entire model implications
Heuristics
model "stress correlations." Taleb (1997) deals with the effects of stochastic correlations: One is
only safe shorting a correlation at 1, and buying it at L1which seems to correspond to what
the 1/n heuristic does. Kelly Criterion vs. Markowitz: In order to implement a full Markowitzstyle optimi- zation, one needs to know the entire joint probability distribution of all assets for
the entire future, plus the exact utility function for wealth at all future times. And with- out
errors! (We saw that estimation errors make the system explode.) Kellys method, developed
around the same period, requires no joint distribution or utility function. In practice one needs
the ratio of expected profit to worst-case returndynamically adjusted to avoid ruin. In the case
of barbell transformations, the worst case is guar- anteed. And model error is much, much
milder under Kelly criterion. Thorp (1971, 1998), Haigh (2000).
The formidable Aaron Brown holds that Kellys ideas were rejected by economists in spite of
the practical appealbecause of their love of general theories for all asset prices.
Note that bounded trial and error is compatible with the Kelly criterion when one has an
idea of the potential returneven when one is ignorant of the returns, if losses are bounded, the
payoff will be robust and the method should outperform that of Fragilista Markowitz.
Corporate Finance: In short, corporate finance seems to be based on point projec- tions, not distributional projections; thus if one perturbates cash flow projections, say, in the Gordon valuation
model, replacing the fixedand knowngrowth (and other parameters) by continuously varying
294
I or
jumps (particularly under fat-tailed dis- tributions), companies deemed AIJexpensive,
A
those with high growth, but low earnings, could markedly increase in expected value, something the market prices heuristically but without explicit reason.
Conclusion and summary: Something the economics establishment has been missing is that
having the right model (which is a very generous assumption), but being un- certain about
the parameters will invariably lead to an increase in fragility in the presence of convexity and
nonlinearities.
19.2.1 Error and Probabilities
295
20
T H E O R I G I N O F T H I N -T A I L S
Chapter Summary 18: The literature of heavy tails starts with a random walk and finds
mechanisms that lead to fat tails under aggregation. We follow the inverse route and
show how starting with fat tails we get to thin-tails from the probability distribution of
the response to a random variable. We introduce a general dose-response curve show
how the left and right-boundedness of the reponse in natural things leads to thin-tails,
even when the underlying variable of the exposure is fat-tailed.
297
(n)
(x)
1 + exp (kbk x + ck )
i=1
(20.1)
2 S N
<
x2
2 S N
x2
The shapes at different calibrations are shown in Figure 1, in which we combined different
values of N=2 S2 ( x, a1 , a2 , b1 , b2 , c1 , c2 ) , and the standard sigmoid S1 ( x, a1 , b1 , c1 ), with a1 =1,
b1 =1 and c1 =0. As we can see, unlike the common sigmoid, the asymptotic response can be
lower than the maximum, as our curves are not monotonically increasing. The sigmoid shows
benefits increasing rapidly (the convex phase), then increasing at a slower and slower rate until
saturation. Our more general case starts by increasing, but the reponse can be actually negative
beyond the saturation phase, though in a convex manner. Harm slows down and becomes flat
when something is totally broken.
20.1
Now let x be a random variable with distributed according to a general fat tailed distribution,
with power laws at large negative and positive values, expressed (for clarity, without loss of
generality) as a Student T Distribution with scale and exponent , and support on the real line.
Its domain D f = (, ), and density f , (x):
! +1
2
x f ,
2
+ x 2
B
1
2, 2
(20.2)
R 1 a 1
where B(a, b) = (a)(b)
(1 t)b1 . The simulation effect of the convex-concave
(a+b) = 0 dtt
transformations of the terminal probability distribution is shown in Figure 2.
And the Kurtosis of the inherited distributions drops at higher thanks to the boundedness
of the payoff, making the truncation to the left and the right visible. Kurtosis for f .2,3 is infinite,
but in-sample will be extremely high, but, of course, finite. So we use it as a benchmark to see
the drop from the calibration of the response curves.
298
1.0
0.5
-5
10
15
20
S2Jx, 1, - , 2, 1, 1, 15N
2
S1Hx, 1, 1, 0L
-1.0
Figure 20.1: The Generalized Response Curve, S2 ( x, a1 , a2 , b1 , b2 , c1 , c2 ) , S1 ( x, a1 , b1 , c1 ) The convex part
with positive first derivative has been designated as "antifragile"
Distribution
Kurtosis
f .2,3 (x)
86.3988
S2 (1, 2, 1, 2, 1, 15)
8.77458
4.08643
S1 (1, 1, 0)
4.20523
a1
1 + exp(b1 x + c1 )
(20.3)
299
f.2,3 HxL
S2 H1,-2,1,2,1,15L
0.020
0.025
0.020
0.015
0.015
0.010
0.010
0.005
0.005
-20
-10
10
20
30
40
0.0
-0.5
S2 H1,-12,2,1,1,15L
0.5
1.0
S1 H1,1,0L
0.014
0.020
0.012
0.015
0.010
0.008
0.010
0.006
0.004
0.005
0.002
0.2
0.4
0.6
0.8
1.0
0.2
0.4
0.6
0.8
Figure 20.2: Histograms for the different inherited probability distributions (simulations,N = 106 )
300
1.0
a1
2
(log( a1xx )+c1)
b12 2
g(x) =
b1xB 2 , 12 (a1 x)
+
(20.4)
gHxL
gHx, 2, 0.1, 1, 1, 0L
gHx, 2, 0.1, 2, 1, 0L
gHx, 2, 0.1, 1, 1, 1L
3
gJx, 2, 0.1, 1, , 1N
2
x
0.5
1.0
1.5
2.0
Kurtosis
16
14
12
10
8
6
4
0.2
0.4
0.6
0.8
1.0
Remark 11: The inherited distribution from S(x) will have a compact support regardless of the
probability distribution of x.
20.2
We showed the dose-response as the neglected origin of the thin-tailedness of observed distributions in nature. This approach to the dose-response curve is quite general, and can be used
outside biology (say in the Kahneman-Tversky prospect theory, in which their version of the
301
302
21
Chapter Summary 19: We extract the effect of size on the degradation of the expectation
of a random variable, from nonlinear response. The method is general and allows to
show the "small is beautiful" or "decentralized is effective" or "a diverse ecology is safer"
effect from a response to a stochastic stressor and prove stochastic diseconomies of scale
and concentration (with as example the Irish potato famine and GMOs). We apply the
methodology to environmental harm using standard sigmoid dose-response to show the
need to split sources of pollution across independent
(nonsynergetic) pollutants.
21.1
Diseconomies and Harm of scale Where is small beautiful and how can we detect, even extract
its effect from nonlinear response? 1 Does getting larger makes an entity more vulnerable to
errors? Does polluting or subjecting the environment with a large quantity cause disproportional
"unseen" stochastic effects? We will consider different types of dose-response or harm-response
under different classes of probability distributions.
The situations convered include:
1. Size of items falling on your head (a large stone vs small pebbles).
2. Losses under strain.
3. Size of animals (The concavity stemming from size can be directly derived from the difference between allometic and isometric growth, as animals scale in a specific manner as they
grow, an idea initially detected by Haldane,[46] (on the "cube law"(TK)).
4. Quantity in a short squeeze
5. The effect of crop diversity
6. Large vs small structures (say the National Health Service vs local entities)
7. Centralized government vs municipalities
8. Large projects such as the concentration of health care in the U.K.
9. Stochastic environmental harm: when, say, polluting with K units is more than twice as
harmful than polluting with K /2 units.
1 The slogan "small is beautiful" originates with the works of Leonard Kohr [65] and his student Schumacher who thus
titled his influential book.
303
Figure 21.1: The Tower of Babel Effect: Nonlinear response to height, as taller towers are disproportionately
more vulnerable to, say, earthquakes, winds, or a collision. This illustrates the case of truncated harm
(limited losses).For some structures with unbounded harm the effect is even stronger.
304
305
Figure 21.2: Integrating the evolutionary explanation of the Irish potato famine into our fragility framework, courtesy http://evolution.berkeley.edu/evolibrary .
settings represented by the "sigmoid" S-curve (or inverted S-curve) which offers more complex
nonlinearities and spans a broader class of phenomena.
Unimodality as a general assumption Let the variable x, representing the stochastic stressor,
follow a certain class of continuous probability distributions (unimodal), with the density p(x)
satisfying: p(x) p(x + e) for all e > 0, and x > x and p(x) p(x e) for all x < x with
{ x : p(x ) = maxx p(x)}. The density p(x) is Lipschitz. This condition will be maintained
throughout the entire exercise.
21.2
In this section, we assume an unbounded harm function, where harm is a monotone (but nonlinear) function in C2 , with negative second derivative for all values of x in R+ ; so let h(x),
R+ R be the harm function. Let B be the size of the total unit subjected to stochastic stressor
x, with (B) = B + h(x).
We can prove by the inequalities from concave transformations that, the expectation of the
large units is lower or equal to that of the sum of the parts. Because of the monotonocity and
concavity of h(x),
!
N
i x
i=1
h(i x),
(21.1)
i=1
N
for all x in its domain (R+ ), where i are nonnegative normalized weights, that is, i=1
i = 1
and 0 i 1.
306
N
And taking expectations on both sides, E((B)) E i=1
(i B) : the mean of a large unit
under stochastic stressors degrades compared to a series of small ones.
Application
Let h(x) be the simplified harm function of the form
h(x) k x ,
(21.2)
k (0, ) , [0, ).
Table 25: Applications with unbounded convexity effects
h(x)
Environment Research
Liquidation
Costs
Toth
et
al.,[115],Bouchaud
et al. [14]
Bridges
Flyvbjerg et al
[42]
kx 2
x(
log(x)+7.1
)
10
Example 1: One-Tailed Standard Pareto Distribution Let the probability distribution of x (the
harm) be a simple Pareto (which matters little for the exercise, as any one-tailed distribution
does the job). The density:
p,L (x) = L x 1 for x L
(21.3)
The distribution of the response to the stressor will have the distribution g = (p h)(x).
Given that k the stressor is strictly positive, h(x) will be in the negative domain. Consider a
second change of variable, dividing x in N equal fragments, so that the unit becomes = x / N,
N N1 :
g,L,N () =
for k
L
N
/
N k
(21.4)
307
M (N)
Z kL
N
g,L,N () d
1 1
k L N
(21.5)
which leads to a simple ratio of the mean of the total losses (or damage) compared to a number
of its N fragments, allowing us to extract the "convexity effect" or the degradation of the mean
coming from size (or concentration):
M (N)
=
M (N)
1
1
(21.6)
With = 1, the convexity effect =1. With = 3/2 (what we observe in orderflow and many other
domains related to planning, Bouchaud et al., 2012, Flyvbjerg et al, 2012), the convexity effect is
shown in Figure 26.
Convexity Effects
Expected total loss for N units
1.0
0.8
0.6
0.4
0.2
N
2
10
Table 26: The mean harm in total as a result of concentration. Degradation of the mean for N=1 compared
to a large N, with = 3/2
Unseen Harm The skewness of g,L,N () shows effectively how losses have properties that hide
the mean in "small" samples (that is, large but insufficient number of observations), since, owing
to skewness, the observed mean loss with tend to be lower than the true value. As with the
classical Black Swan exposures, benefits are obvious and harm hidden.
21.3
Now the biological and physical domains (say animals, structures) do not incur unlimited harm,
when taken as single units. The losses terminate somewhere: what is broken is broken. From
ak
M
, a sum of single
the generalized sigmoid function of [? ], where S M (x) = k=1
1+exp(bk (ck x))
sigmoids. We assume as a special simplified case M = 1 and a1 = 1 so we focus on a single
stressor or source of harm S(x), R+ [1, 0] where x is a positive variable to simplify and the
response a negative one. S(0) = 0, so S(.) has the following form:
S(x) =
308
1
1
+
1 + e b (c x) 1 + eb c
(21.7)
10
-0.2
-0.4
-0.6
-0.8
-1.0
Higher scale
(dispersion or
variance)
Convex Response
Z c
0
Z
c
S() p() d,
S() p() d
309
We can therefore prove that the effect of the expectation for changes in N depends
R aexactly on
whether the mass to the left of a is greater than the mass to the right. Accordingly, if 0 p() d >
R
a p() d, the effect of the concentration ratio will be positive, and negative otherwise.
Application
Example of a simple distribution: Exponential Using the same notations as 21.2, we look for
the mean of the total (but without extracting the probability distribution of the transformed
variable, as it is harder with a sigmoid). Assume x follows a standard exponential distribution
with parameter , p(x) e( x)
Z
1
1
+
dx
(21.8)
M (N) = E (S()) =
e( x)
x
bc
0
eb(c N ) + 1 e + 1
N N
1
2 F1 1,
;
+ 1; ebc
M (N) = bc
b
b
e +1
where the Hypergeometric function 2 F1 (a, b; c; z) =
k=0
a k bk z k
k!ck .
M (N)
doesnt admit a reversal owing to the shape, as we can see in 21.4 but we can
The ratio M
(N)
see that high variance reduces the effect of the concentration. However high variance increases
the probability of breakage.
Example of a more complicated distribution: Pareto type IV Quasiconcave but neither convex
nor concave PDF: The second derivative of the PDF for the Exponential doesnt change sign,
2
( exp(x)) = 3 e( x) , so the distribution retains a convex shape. Further, it is not possible
x2
to move its mean beyond the point c where the sigmoid switches in the sign of the nonlinearity.
So we elect a broader one, the Pareto Distibution of Type IV, which is extremely flexible because,
unlike the simply convex shape (it has a skewed "bell" shape, mixed convex-concave-convex
shape) and accommodates tail exponents, hence has power law properties for large deviations.
It is quasiconcave but neither convex nor concave. A probability measure (hence PDF) p : D
[0, 1] is quasiconcave in domain D if for all x, y D and [0, 1] we have:
p(x + (1 )y) min ( p(x), p(y)).
Where x is the same harm as in Equation 21.7:
1
1
k1/ (x )
p,,,k (x) =
for x and 0 elsewhere.
310
k
x
1/
1
+1
(21.9)
M HL
M H1L
1.0
0
0.8
0.4
0.2
10
Figure 21.4: Exponential Distribution: The degradation coming from size at different values of .
311
The Four figures in 3.14 shows the different effects of the parameters on the distribution.
PDF
x
2
10
-0.2
-0.4
-0.6
PDF
0.35
0.30
0.25
0.20
0.15
0.10
0.05
-0.8
-1.0
PDF
0.7
0.6
0.5
0.4
0.3
0.2
0.1
10
12
14
10
12
14
PDF
0.5
0.4
0.3
0.2
0.1
x
2
4
6
8 10 12 14
The mean harm function, M,,,k (N) becomes:
M,,,k (N) =
k1/
Z
0
(x )
1
1
1
1
ebc + 1 eb(c Nx ) + 1
! 1
1/
k
+1
dx
x
(21.10)
M(.) needs to be evaluated numerically. Our concern is the "pathology" where the mixed
convexities of the sigmoid and the probability distributions produce locally opposite results
than 3.14 on the ratio
M,,,k (N)
M,,,k (N) .
effects, as in 21.7. However as shown in Figure 21.5, the total expected harm is quite large under
these conditions, and damage will be done regardless of the effect of scale.
Conclusion
This completes the math showing extracting the "small is beautiful" effect, as well as the effect
of dose on harm in natural and biological settings where the Sigmoid is in use. More verbal
discussions are in Antifragile.
312
-0.1
-0.2
-0.3
S''(x)=0
-0.4
-0.5
Mk,,, HL
Mk,,, H1L
1.5
1.0
0.5
10
M H2L
M H1L
0.6
0.5
0.4
0.3
313
314
22
Chapter Summary 20: Explains why the fragilefragile is necessarily in the nonlinear.
a
i
e ((+x))
+1
= 2a sinh
1
( + x)
2
csch3 (( + x)),
(22.1)
where sinh and csnh are the hyperbolic sine and cosine, respectively.
Next we subject all the families to a probability distribution of harm, f (z) being
a monomodal
distribution with the expectation E(z) ( x0 , x1 ] . We compose f Sto get f S,i , (x) . In this
case we pick a symmetric power law.
f , Sa,, (x) =,
with e (1, ) and (0, )
The objects will produce a probability distribution around [0, 1] since Sa,i , (x) is bounded at
these levels; we can see to the right a Dirac mass concentrating observations at 1. Clearly what
has survived is the nonlinear.
315
More
Concave
0.8
0.6
0.4
More Linear
0.2
Dose
n o
Figure 22.1: The different dose-response curves, at different values of i , corresponding to varying
levels of concavity.
100
80
60
40
20
0.2
316
0.4
0.6
0.8
1.0
2.5
2.0
1.5
1.0
0.5
0.2
0.4
0.6
0.8
1.0
317
23
Chapter Summary 21: Information is convex to noise. The paradox is that increase in
sample size magnifies the role of noise (or luck); it makes tail values even more extreme.
There are some problems associated with big data and the increase of variables available
for epidemiological and other "empirical" research.
23.1
To the observer, every day will seem weirder than the previous one. It has always been absolutely
silly to be exposed the news. Things are worse today thanks to the web.
Source
Effect
Weirder and weirder events reported on the front pages
News
Epidemiological
"Big Data"
Studies,
Track Records
Greater performance for (temporary) "star" traders
We are getting more information, but with constant consciouness, desk space, or visibility. Google News, Bloomberg News, etc. have space for, say, <100 items at any point
in time. But there are millions of events every day. As the world is more connected, with
the global dominating over the local, the number of sources of news is multiplying. But your
consciousness remains limited. So we are experiencing a winner-take-all effect in information:
like a large movie theatre with a small door.
Likewise we are getting more data. The size of the door is remaining constant, the theater is
getting larger.
The winner-take-all effects in information space corresponds to more noise, less signal. In
other words the spurious dominates.
Similarity with the Fooled by Randomness Bottleneck
This is similar to the idea that the more spurious returns dominate finance as the number of
players get large, and swamp the more solid ones. Start with the idea (see Taleb 2001), that as a
population of operators in a profession marked by a high degrees of randomness increases, the
number of stellar results, and stellar for completely random reasons, gets larger. The spurious
tail is therefore the number of persons who rise to the top for no reasons other than mere luck,
319
23.2
derivations
j
Let Z zi 1< j<m,1i<n be a (n m) sized population of variations, m population series and n
data points per distribution, with i, j N; assume noise or scale of the distribution R+ ,
signal 0 . Clearly can accommodate distributions with infinite variance, but we need the
expectation to be finite. Assume i.i.d. for a start.
320
23.2 derivations
Cross Sectional (n = 1) Special case n = 1: we are just considering news/data without historical
attributes.
Let F be the generalized inverse distribution, or the quantile,
F (w) = inf{t R : F(t) w},
for all nondecreasing distribution functions F(x) P(X < x). For distributions without compact support, w (0,1); otherwise w [0, 1]. In the case of continuous and increasing distributions, we can write F 1 instead.
The signal is in the expectaion, so E(z) is the signal, and the scale of the distribution determines the noise (which for a Gaussian corresponds to the standard deviation). Assume for now
that all noises are drawn from the same distribution.
Assume constant probability the threshold, = mk , where k is the size of the window of the
arrival. Since we assume that k is constant, it matters greatly that the quantile covered shrinks
with m.
Gaussian Noise
When we set as the reachable noise. The quantile becomes:
F 1 (w) =
2 erfc1 (2w) + ,
k
1
1
, () = 2erfc
2
+
m
Note that (noise) is multiplicative, when (signal) is additive.
As information increases, becomes smaller, and 1 moves away in standard deviations.
But nothing yet by comparison with Fat tails.
Fat Tailed Noise
Now we take a Student T Distribution as a substitute to the Gaussian.
! +1
2
f (x)
(x )2
2
1
2, 2
(23.1)
, ()
=+
v
u
u
sgn (1 2 ) t
1
1
I(1,(2
1)sgn(12))
1
2, 2
1
(23.2)
B(z
,z
) (a,b)
0 1
where I is the generalized regularized incomplete Beta function I(z0 ,z1 ) (a, b) = B(a,b)
, and
R z a 1
b
1
Bz (a, b) the incomplete Beta function Bz (a, b) = 0 t (1 t) dt. B(a, b) is the Euler Beta func-
321
=1
=2
=3
=4
20
15
Figure 23.2:
Gaussian,
={1,2,3,4}
10
0.02
h!
0.04
R1
0
0.06
0.08
0.10
t a1 (1 t)b1 dt.
As we can see in Figure 2, the explosion in the tails of noise, and noise only.
Fatter Tails: Alpha Stable Distribution
Part 2 of the discussion to come soon.
322
23.2 derivations
10 000
=1
=2
=3
=4
8000
6000
4000
2000
2. 10-7
4. 10-7
6. 10-7
8. 10-7
1. 10-6
60 000
=1
=2
=3
=4
50 000
40 000
30 000
20 000
10 000
2. 10-7
4. 10-7
6. 10-7
8. 10-7
1. 10-6
323
24
T H E C O N V E X I T Y O F W E A LT H T O
INEQUALITY
24.1
The one percent of the one percent of the population is vastly more sensitive to inequality than
total GDP growth (which explains why the superrich are doing well now, and should do better
under globalization, and why it is a segment that doesnt correlate well with the economy). For
the super-rich, one point of GINI causes an increase equivalent to 6-10% increase in total income
(say, GDP). More generally, the partial expectation in the tail is vastly more sensitive to changes
in scale of the distribution than in its centering.
Sellers of luxury goods and products for the superwealthy profit from dispersion more than
increase in total wealth or income. I looked at their case as a long optionality, benefit-fromvolatility type of industry.
From textitAntifragile[111]:
Another business that does not care about the average but rather the dispersion around the average
is the luxury goods industryjewelry, watches, art, expensive apartments in fancy locations, expensive
collec - tor wines, gourmet farm - raised probiotic dog food, etc. Such businesses only cares about the
pool of funds available to the very rich. If the population in the Western world had an average income
of fifty thousand dollars, with no inequality at all, the luxury goods sellers would not survive. But
if the average stays the same, with a high degree of inequality, with some incomes higher than two
million dollars, and potentially some incomes higher than ten million, then the business has plenty of
customerseven if such high incomes were offset with masses of people with lower incomes. The tails
of the distribution on the higher end of the income brackets, the extreme, are much more determined by
changes in inequality than changes in the average. It gains from dispersion, hence is antifragile.
This explains the bubble in real estate prices in Central London, determined by inequality in Russia
and the Arabian Gulf and totally independent of the real estate dynamics in Britain. Some apartments,
those for the very rich, sell for twenty times the average per square foot of a building a few blocks away.
Harvard s former president Larry Summers got in trouble explaining a version of the point and
lost his job in the aftermath of the uproar. He was trying to say that males and females have equal
intelligence, but the male population has more variations and dispersion (hence volatility), with more
highly unintelligent men, and more highly intelligent ones. For Summers, this explained why men
were overrepresented in the sci - entific and intellectual community (and also why men were overrepre
- sented in jails or failures). The number of successful scientists depends on the tails, the extremes,
rather than the average. Just as an option does not care about the adverse outcomes, or an author does
not care about the haters.
325
p(x) =
( 1) (m xmin )
We are dealing with a three parameter function, as the fatness of the tails is determined by
both and m xmin , with m xmin > 0 (since >1).
Note that with 7 billion humans, the one percent of the one percent represents 700,000 persons.
The same distribution applies to wealth and income (although with a different parametrization, including a lower as wealth is more unevenly distributed than income.)
Note that this analysis does not take into account the dynamics (and doesnt need to): over
time a different population will be at the top.
The Lorenz curve Where F(x), short for P(X < x) is the cumulative distribution function and
inverse F (z) : [0,1] [xmin , ), the Lorenz function for z L(z):[0, 1][0,1] is defined as:
Rz
L(z) R01
0
F (y)dy
F (y)dy
1
(1 z)1/ ((z 1) (m xmin )
m
1
+ (z 1) (m(z + z) + (z 1)xmin )
(24.1)
Which gives us different combination of and m xmin , producing different tail shapes: some
can have a strong middle class (or equivalent) while being top-heavy; others can have more
equal inequality throughout.
Gini and Tail Expectation
The GINI Coefficient, [0,1] is the difference between 1) the perfect equality,with a Lorenz L( f ) =
f and 2) the observed L (z, , m, xmin )
GINI (, m, xmin ) =
326
(m xmin )
(2 1)
m
24.1 the one percent of the one percent are divorced from the rest
Lorenz@zD
1.0
0.8
Figure 24.1:
Different
combinations
L(z, 3, .2, .1), L(z, 3, .95, .1), L(z, 1.31, .2, .1)
in addition to the perfect equality line L(
z)= z. We see the criss-crossing at higher
values of z.
0.6
0.4
0.2
z
0.2
0.4
0.6
0.8
1.0
Computing
the tail mass above a threshold K, that is, the unconditional partial expectation
R
E>K K xp(x) dx, which corresponds to the nominal share of the total pie for those with wealth
above K,
E>K = ( 1)
( (K + m xmin ) m)
m xmin
K + ( 1)m xmin
P>K = 1 +
( 1) (m xmin )
For the One Percent of the One Percent (or equivalent), we set the probability P>K and invert to
K P =( 1) (m xmin ) p1/ (1 + m + xmin ),
1
1
(m xmin ) + p (m m + xmin )
E>K = p
Now we can check the variations in GINI coefficient and the corresponding changes in E>K for
a constant m.
GINI
E>K
E>K /m
1.26
0.532895
0.33909
0.121103
1.23
0.541585
0.395617
0.141292
1.2
0.55102
0.465422
0.166222
1.17
0.561301
0.55248
0.197314
1.14
0.572545
0.662214
0.236505
1.11
0.584895
0.802126
0.286474
1.08
0.598522
0.982738
0.350978
327
25
25.1
Take health effect a function response from a single parameter, f: R ->R be a twice differentiable, the effect from dose x.
2 f (x)
If over a range x [a,b], over a set time period t, x2 > 0 or more heuristically, 21 (f(x+x)
+ f(x-x))> f(x), with x+x and x-x [a,b] then there are benefits from unevenness of distribution: episodic deprivation, intermittent fasting, variable pulmonary ventilation, uneven
distribution of proteins(autophagy), vitamins, high intensity training, etc.).
In other words, in place of a dose x, one can give 140% of x , then 60% of x, with a more
favorable outcome.
329
f Hx + DxL + f Hx - DxL
2
H
fHxL
Dose
Proof
Proof: Jensens Inequality.
This is a simplification here since dose response is rarely monotone in its nonlinearity, as we
will see further down.
Mixed Nonlinearities in Nature Nonlinearities are not monotone.
Nonlinearities in Biology
Biology- The shape convex-concave necessarily flows from anything increasing (monotone, i.e. never decreasing) and bounded, with a maximum and a minimum values,
i.e. never reached infinity from either side. At low levels, the dose response is convex (gradually
more and more effective). Additional doses tend to become gradually ineffective or hurt. The
same can apply to anything consumed in too much regularity. This type of graph necessarily
applies to any situation bounded on both sides, with a known minimum and maximum (saturation), which includes happiness.
For instance, If one considers that there exists a maximum level of happiness and unhappiness
then the general shape of this curve with convexity on the left and concavity on the right has
to hold for happiness (replace dose with wealth and response with happiness). KahnemanTversky Prospect theory models a similar one for utility of changes in wealth, which they
discovered empirically.
2 f (x)
Iatrogenics If x2 0 for all x (to simplify), and x is symmetrically distributed, then the
distribution of the outcome from administration of f (and only the effect of f ) will be leftskewed as shown in Figure 1. Further known limited upside, unknown downside to map the
effect of the next figure.
Probability
Hidden Iatrogenics
Benefits
Outcomes
330
Medical Breakeven
Iatrogenics zone
Condition
331
332
Part IV
D E E P E R M AT H E M AT I C A L P R O B L E M S I N F I N A N C E A N D
OPTION THEORY
26
26.1
central conflation
335
27
Chapter Summary 25: American Options have hidden optionalities. Using a European option
as a baseline we heuristically add the difference. We also show how these hidden options
are extremely significant compared to the focus of the research literature on insignificant
marginal improvements of the pricing equations but in the Black Scholes World.
27.1
this note
This is a paper in progress, not formatted for submission, but aiming at the development of
ideas and mathematical results around the problem. We start with the math, and end with the
explanations, much of which were aimed at updating Dynamic Hedging; the sequence will be
somewhat reversed in the final paper, and some comments will be added.
27.2
Optionality Only one payoff can be "accepted", which makes the maximal one the one of concern as we abandon the other, inferior ones. Define A as the new set of "events" i of concern
and ic the complentary event, that is that i does not take place, as follows:
337
0.10
0.10
0.08
0.08
0.06
0.06
0.04
0.04
0.02
0.02
0.00
0.00
Figure 27.1: the vector of weights under probabilities flat p1 = = p6 = 102 and 103 and m = 90
days. We can observe probabilities remain significant and cumulatively consequential. We also see how i
become equal, with a flat weights at small probabilities.
o nn
o
o
n
A = p cp n1
cp nc 1 n2
A=
p n
[
nn
o
o
cp nc 1 nc 2 n3 . . .
(27.2)
(27.3)
ij=0 nc j ni
i=0
(1 (1 pn )m )
n 1
(1 pni )m
i=1
!!
n p
(1 , 2 , . . . , p )
(1 , 2 , . . . , p )
We skip to the expectation without dealing with probability distribution:
E[X |A] = .T
(27.4)
Given that the events are disjoint,the expected value of the option on n draws over a sequence
of observations of length m (which could correspond to m time periods):
338
m = E[X |A]
=
n=1 n
(1 (1 pn ) )
n 1
(1 pni )
!
m
(27.5)
i=1
n=1 n (1 (1 pn )m )
(27.6)
Note on the difference between the heuristic Bernouilli and a set of full distributions
It will take a few lines to show whether the tractability of the Bernouilli simplification causes
losses in precision. We could use, say, the maximum of gamma/exponential family with different
calibrations but Monte Carlo shows no difference.
27.3
misplaced precision
So many "rigorous" research papers in derivatives have been involved in the "exact" pricing of
American options, though within model when in fact their most interesting attribute is that they
benefit from the breakdown of models, or they are convex to model errors.
Indeed an interesting test to see if someone has traded derivatives is to quiz him on American
options. If he answers by providing a "pasting boundary" story but using a Black-Scholes type
world, then you can safely make the conclusion that he has never gotten close to American
options.
Furthermore, with faster computers, a faster pricing algorithm does not carry large advantages.
The problem is in the hidden optionality... Major points:
An American option is always worth equally or more than the European option of the same
nominal maturity.
An American option has always a shorter or equal expected life than a European option.
Rule 27.1.
The value of the difference between an American and European option of same strike and maturity increases
with the following factors:
339
27.4
We can therefore show that, as of period t0 , for all periods to expiration t, where O A is the
"conventionally priced" American option (according to whichever method one choses), and OE
is the corresponding European option of the same maturity and strike,
OA = OE + E ((O A OE ) m )
(27.7)
the expectation of the maximum of two expectations, which allows the simplification:
OA = OE + ((O A OE ) m )
(27.8)
We now need to define the components entering m , namely the various probabilities pi and
associated payoff i .
NOTE: This part will need some derivations, a bit more clarity about the derivations, etc. Also
note that there is a need to prove iterated expectations...
27.5
war stories
340
Case 1 (Simplest, the bang comes from the convexity to changes in the carry of the premium)
Why do changes in interest rate carry always comparatively benefit the American option ? Take
a 1 year European and American options on a forward trading at 100, i.e. with a spot at 100.
The American option will be priced on the risk management system at exactly the same value
as the European one. S=100, F=100, where S is the spot and F is the forward. Assume that the
market rallies and the spot goes to 140. Both options will go to parity, and be worth $40.
Case 1 A Assume that interest rates are no longer 0, that both rates go to 10%. F stays equal to
S. Suddenly the European option will go from $40 to the present value of $40 in one year using
10%, i.e. $36.36. The American option will stay at $40, like a rock.
Case 1 B Assume the domestic rate goes up to 10%, spot unchanged. F will be worth approximately of S. It will go from 140 to 126, but the P/L should be neutral if the option still has no
gamma around 126 (i.e. the options trade at intrinsic value). The European option will still drop
to the PV of 26, i.e. 23.636, while the American will be at 26.
We can thus see that the changes in carry always work to the advantage of the American
option (assuming the trader is properly delta neutral in the forward). We saw in these two cases
the outperformance of the American option. We know the rule that :
If in all scenarios option A is worth at least the same as option B and, in some scenarios can
be worth more than option B, then it is not the greatest idea to sell option A and buy option B
at the exact same price.
This tells us something but not too much: we know we need to pay more, but how much
more?
341
27.6
Another non-trivial problem with American options lies in the fact that the forward hedge is
unknown. It resembles the problem with a barrier option except that the conditions of termination are unknown and depend on many parameters (such as volatility, base interest rate, interest
rate differential). The intuition of the stopping time problem is as follows: the smart option will
position itself on the point on the curve that fits it the best.
Note that the forward maturity ladder in a pricing and risk management system that puts the
forward delta in the terminal bucket is WRONG.
27.7
27.8
conclusion
342
28
Chapter Summary 26: Much of the work concerning martingales and Brownian motion
has been idealized; we look for holes and pockets of mismatch to reality, with consequences. Infinite (or undefined) higher moments are not compatible with Ito calculus
outside the asymptote. Path dependence as a measure of fragility.
28.1
S
140
120
j
ST
100
80
j
Path 1 , Smin
Time
0.0
0.2
0.4
0.6
0.8
1.0
j
Figure 28.1: Brownian Bridge Pinned at 100 and 120, with multiple realizations
{S0 , S1 , .., ST }, each indexed
j
by j ; the idea is to find the path j that satisfies the maximum distance D j = ST Smin
Let us examine the non-Markov property of antifragility. Something that incurred hard times
but did not fall apart is giving us information about its solidity, compared to something that has
not been subjected to such stressors.
343
Figure 28.2: The recovery theorem requires the pricing kernel to be transition
independent. So the forward kernel at S2
depends on the path. Implied vol at S2 via
S1b is much lower than implied vol at S2
via S1a.
Markov
Property
for,itsay,
a the
Brownian
X N |{ X1 ,X2 ,...X N 1 } = X N |{ X N 1 } , that is the
The(The
first time
I met Emanuel
Derman,
was in
summer ofMotion
1996, at Uncle
Nick's on 48th street and 9th Avenue. Stan Jonas paid, I remember (it is
last realization is the only one that matters. Now if we take fat tailed models, such as stochastic
sometimes easier to remember who paid than the exact conversation). Derman
volatility
processes,
thetheproperties
ofmodel
the system
Markov,
but the history of the past realizaand
Dupire had
come up with
local volatility
and I was are
burning
to
talk
to Emanuel
about it. I matter
was writing
Hedging the
and in
the middle
of
tions
of the process
in Dynamic
determining
present
variance.)
an intense intellectual period (I only experienced the same intellectual
intensity
2005-2006
as I was writing
The Black Swan).
I wasprocess
tortured with
one at St0 = 100 and ST = 120, sampled
Takein M
realizations
of a Brownian
Bridge
pinned
aspect to the notion of volatility surface. I could not explain it then. I will try
with N periods separated by t, with the sequence S, a collection of Brownian-looking paths
now.
withnote
single
realizations
indexeddoes
bynot
j , mean what you expect
First,
the following.
Local volatility
volatility to be along a stochastic sample path that delivers a future price
time pair. It is not necessarily the mean square variation
along a sample
path.
M
N
j
Nor is it the expected mean-square variation along ja sample path
that
allows
S
=
S
you to break-even on a dynamic hedge. It is the process
that it+t
would
a
i
0 provide
i=0 j=1
break even P/L for a strategy.
The resulting subtelty will take more than
n one post to explaino(or I may expand
j
j
= min
in Dynamic
2). But
I will
to explain
as much
Take mHedging
min
i try
and
j : min
S =asmIcan right here.
j
i
The first problem is that options are not priced off a imean-square variation in
Take 1) the sample path with the most direct route (Path 1) defined as its lowest minimum
, and 2) the one with the lowest minimum m (Path 2). The state of the system at period
T depends heavily on whether the process ST exceeds its minimum (Path 2), that is whether
arrived there thanks to a steady decline, or rose first, then declined.
If the properties of the process depend on (ST - m*), then there is path dependence. By properties of the process we mean the variance, projected variance in, say, stochastic volatility models,
or similar matters.
28.2
For time series sampled at (t0 , t0+t , ..., t t0+nt ), the minimum distance :
N N
S (t0 , t, t) min Sit+t0 min Sjt+t0 j=i+1
i=0
We have the stopping time { : S = S (t0 , t, t)} and the distance from the worst becomes
(t0 , t, t) St S
344
0.8
0.6
0.2
200
28.3
400
600
800
1000
We mentioned in the discussion of the Casanova problem that stochastic calculus requires a certain class of distributions, such as the Gaussian. It is not as we expect because of the convenience
of the smoothness in squares (finite x2 ), rather because the distribution conserves across time
scales. By central limit, a Gaussian remains a Gaussian under summation, that is sampling at
longer time scales. But it also remains a Gaussian at shorter time scales. The foundation is
infinite dividability.
The problems are as follows:
The results in the literature are subjected to the constaints that the Martingale M is member
of the subset (H2 ) of square integrable martingales, suptT E[M2 ] <
We know that the restriction does not work for lot or time series.
RT
We know that, with an adapted process, without 0 s2 ds < we cant get most of the
results of Itos lemma.
RT
Even with o dW 2 < , The situation is far from solved because of powerful, very powerful
presamptotics.
RT
Hint: Smoothness comes from o dW 2 becoming linear to T at the continuous limit Simply
dt is too small in front of dW
Take the normalized (i.e. sum=1) cumulative variance (see Bouchaud & Potters),
C(n) =
n
(W[it] W[(i 1)t])2
i=1
T /t
28.4
There is a difference between the Stratonovich and Itos integration of a functional of a stochastic
process. But there is another step missing in Ito: the gap between information and adjustment.
345
1.0
0.8
0.6
0.2
200
400
600
800
1000
1.0
0.8
0.6
0.4
0.2
200
400
600
800
1000
346
28.5 finite variance not necessary for anything ecological (incl. quant finance)
28.5
finite variance not necessary for anything ecological (incl. quant finance)
28.6
The recovery theorem require Markov conditions between transitional states, but data shows
that the path, particularly the position with respect to recent minimum matters for states,
significantly enough for equal negativemovement p between a price p and p p depend
severely on whether the latter price is above the recent minimum .
70
60
50
40
30
-0.35
-0.30
-0.25
-0.20
-0.15
-0.10
-0.35
-0.30
-0.25
-0.20
-0.15
-0.10
28.7
First, the recovery theorem hinges on the pricing kernel being Markov, and path independent, that is, at times t, the t + 1, , the state of the system (which we can simplify by
347
0.00006
0.00004
0.00002
-5.5
-5.0
-4.5
-4.0
focusing on , the "volatility" at which the distribution is priced, should not depend on the
path.
Second, more devastating for the argument, the recovery theorem is similar to variance
swaps with "capped" variance swaps, nice try but not the real McCoy. Alas, probabilities
deliver a single segment of the distribution, not full moments. Option prices are not
determined by just probabilities but by a combination of probability and payoff, hence the
extreme part of the distribution (in fat-tailed domains) may represent a disproportionate
segment of the valuation of the volatility smile. The current derivations by Ross (2013), and
in the continuous time version (Carr et al., 2012) are bounded.
{10.8776, 147.651},
DF SS
MS
F-Statistic P-Value
x
1
15664.8 15664.8 391.943
5.07638 1036
Error 98 3916.77 39.967
Total 99 19581.6
348
MS
F-Statistic
P-Value
549.093
8.11559 1065
21938.5
21938.5
Error
248
9908.63
39.9542
Total
249
31847.2
28.8
While accepting the entire Black Scholes framework, we set up the equations for dynamic hedging in "real life", that is how the limiting process works, and show that in real time t there is
an added variance term from the randomness of the exposure over the life of an option, which
we extract. The key problem is that we are not conditioning on specific hedge ratios.
First Problem, while theta= gamma on balance, the gamma level is stochastic, since a given
option will not be at the same strike all the time, which makes the variance of the package
much higher than usually calculated.
Second problem: the same applies to the stream of delta hedges, since delta is stochastic,
though uniformly distributed.
Severe Problem, am information-theoretic one, which makes the two previous ones relevant: transactions have an additional lag, causing a second effect in discretization of
stochastic integral.
28.9
We reexpress the dynamic hedging argument, which requires the ability to reduce the variance
of the portfolio through continuous revision of the holdings in the underlying asset.
Consider the Black and Scholes package, as a discretization of the stochastic integral: the
payoff of a call option corresponds to initial price at which it was sold, C0 , the expectation of
terminal "moniness" E ((St K)+ ), plus the performance of (t) the stream of dynamic hedges
between t0 and t (which includes the initial "delta" hedge). We break up the period (t0 , t) into
n increments of size t. Here, the hedge ratio C
S is computed as of time t0 + (i 1)t, but we
get the nonanticipating difference between the price at the time the hedge was initiated and the
resulting price at t0 + i t.
Hence the first equation:
n = C0 (St K)+ + (t)
where n =
t t0
t
(28.1)
and
n
(t) =
i=1
C
|St +(i1)t ,t0 +(i1)t St0 +it S(t0 +(i1)t
0
S
This operation is supposed to make the payoff deterministic at the limit of t 0, with
a net profit or loss of (t) (St K)+ that converges to equality with C0 . This net difference
349
-3
-2
-1
-0.01
-0.02
-0.03
-0.04
Stochastic
Scale
-0.05
-0.06
corresponds to an Ito-McKean stochastic integral, under a set of assumptions about the dynamics
of the process. More formally, under the right dynamics, by the law of large numbers, n
converges:
P
n
0
when n .
Over a single period t, the conditional mistracking in changes in the delta-hedged portfolio are
as follows:
The "gamma",
t0 +i t =
2
1 2 C
|St +(i1)t ,t0 +(i1)t St0 +it S(t0 +(i1)t
2
0
2 S
(28.2)
C
|
t
t St0 +(i1)t ,t0 +(i1)t
(28.3)
350
28.10
So here we have two random terms, the change in S and the distribution of the theta/gamma,
both independent since they address by the rules of stochastic calculussequential, nonoverlapping periods, as made clear in the equations. Now consider that the gamma of the
option is the scaled density of a Gaussian variable, hence:
We have the new random variable, = t0 +i t t0 +i t :
z St t
z (St St+t )2
2 t t0
2 St t t0
To separate the random variables:
z St2 2 t (St St+t )2
=
2 St t t0
Condensed:
z(z1 1)
c
=
where
z1 =
St St+t
St t
2
For z, it can be shown that its probability distribution and support are:
1
2
, z 0,
.
p(z) = p
2 log(z) log(2)
2
(and 0 elsewhere).
Note that
St St+t
St t
z2 =
z1
Gaussian(0, 1)
2 t t0
S t
z1 1
scaled Chi-square
c
2
e
, z1
2 z1
p(z2 ) =
c e 2 2
1
, z2 [ , ).
c
2 cz2 + 1
(28.4)
351
Asset Price
Revision Done
Order Placed
Figure 28.11: The Impossibility of Dynamic Hedging.
Additional informational lag in dynamic hedging prevents the smoothing of
S2 at the limit of t going
to 0, adding another term to
the Ito-McKean Integral.
Delta Observed
t1
t0
t3
t2
Period
We end up with:
1
2
c
2z 21
2
|c|e
2 log(z)log(2)
(
p() =
1
2
2 log(z)log(2)
V() =
dz
<0
(28.5)
c
2z 21
2
|c|e
28.11
c
z +1
c
z +1
dz
0<<
1
2c
1
3c2
Let us investigate the statistical properties of the hedge to examine what takes place at the
limit. Clearly we are dealing with the product of 1) the hedge ratio which follows a uniform
distribution on (0,1) and 2) the variations of lognormal St at the subsequent time window. The
time lag causes independence between the two r.v.s. Both have finite variance, so the sum of
the product converges to a Gaussian distribution for which the operation of the (weak) law of
large numbers is well known. The distribution of n is degenerate at the limit, a Gaussian with
a variance approaching 0 delivering a Dirac mass at 0.
Let s t be the standard deviation for St0 +it S(t0 +(i1)t , which we assume (in the BlackScholes world) is normally distributed for small increments (without any loss of generality).
352
to 0, since, by the properties of a Brownian motion the ratio |S|0 becomes small.
R
S N(0, s t)
2
0, (t)
2
2s t
(t) = h s
2 2 ts
(28.6)
The characteristic function of (t), C (), under n-convolutions and time increment t/n:
q ! n
s| |
erf
2
n
q
C ()n = 2n | |
s t
n
t
n
n
1
t(2) 2 s2
3
which in the Black-Scholes world would still converge to 0 (though less rapidly), except for the
problem of microstructure, as we see next.
28.11.1
The two previous problems set a case for difficulty in convergence. Let us see how in a true
dynamic environmnent the limit of a Dirac mass at 0 is never reached.
Let us consider specific periods t0 , t1 , ...tn over a given increment t. The problem, as we see in
Figure 1, is that the operator gets his information at period t1 about the corresponding value of
the underlying asset St1 , then makes an adjustment decision based on the hedge ratio at t1 , since
that was the last price he was aware of. The Black Scholes adjustment (Merton, 1973) assumes
that the operator is now hedged as of t2 , with the difference in hedging errors remaining of
order t2 , as we saw, converging to S2 , hence the variance of the asset (in expectation), as part
of the properties of stochastic integration.
But Black-Scholes misses another fundamental lag. The operator cannot instantaneously activate an order to buy at t1 , the period in which he gets the information; the order needs to be
placed at a subsequent time period > t1 , say t2 , with the added complication that the asset price
for the revision is obtained as of t3 , two periods later at the hedge ratio as of t1 , exposed to
C
the variation S
|St2 |St4 St2 |, an incompressible tracking error. The discrepancy is not trivial;
t
it is similar to the difference between the Stratonovich and Ito integrals. The additional incompressible variance prevents the option package from ever having, at the limit of t, a purely
deterministic payoff.
353
29
Chapter Summary 27: Proof that under simple assumptions, such as constraints of Put-Call
Parity, the probability measure for the valuation of a European option has the mean derived
from the forward price which can, but does not have to be the risk-neutral one, under any
general probability distribution, bypassing the Black-Scholes-Merton dynamic hedging argument, and without the requirement of complete markets and other strong assumptions. We
confirm that the heuristics used by traders for centuries are both more robust, more consistent, and more rigorous than held in the economics literature. We also show that options can
be priced using infinite variance (finite mean) distributions.
29.1
background
Option valuations methodologies have been used by traders for centuries, in an effective way
(Haug and Taleb, 2010). In addition, valuations by expectation of terminal payoff forces the mean
of the probability distribution used for option prices be be that of the forward, thanks to Put-Call
Parity and, should the forward be risk-neutrally priced, so will the option be. The Black Scholes
argument (Black and Scholes, 1973, Merton, 1973) is held to allow risk-neutral option pricing
thanks to dynamic hedging, as the option becomes redundant (since its payoff can be built as
a linear combination of cash and the underlying asset dynamically revised through time). This
is a puzzle, since: 1) Dynamic Hedging is not operationally feasible in financial markets owing
to the dominance of portfolio changes resulting from jumps, 2) The dynamic hedging argument
doesnt stand mathematically under fat tails; it requires a very specific "Black Scholes world"
with many impossible assumptions, one of which requires finite quadratic variations, 3) Traders
use the same Black-Scholes "risk neutral argument" for the valuation of options on assets that
do not allow dynamic replication, 4) Traders trade options consistently in domain where the
risk-neutral arguments do not apply 5) There are fundamental informational limits preventing
the convergence of the stochastic integral.1
There have been a couple of predecessors to the present thesis that Put-Call parity is sufficient
constraint to enforce some structure at the level of the mean of the underlying distribution,
such as Derman and Taleb (2005), Haug and Taleb (2010). These approaches were heuristic,
robust though deemed hand-waving (Ruffino and Treussard, 2006). In addition they showed
that operators need to use the risk-neutral mean. What this paper does is
355
unique option pricing measure with neither dynamic hedging nor complete markets
Beyond risk-neutrality, it establishes the case of a unique pricing distribution for option
prices in the absence of such argument. The forward (or future) price can embed expectations and deviate from the arbitrage price (owing to, say, regulatory or other limitations)
yet the options can still be priced at a distibution corresponding to the mean of such a
forward.
It shows how one can practically have an option market without "completeness" and without having the theorems of financial economics hold.
These are done with solely two constraints: "horizontal", i.e. put-call parity, and "vertical", i.e.
the different valuations across strike prices deliver a probability measure which is shown to be
unique. The only economic assumption made here is that the forward exits, is tradable in
the absence of such unique forward price it is futile to discuss standard option pricing. We also
require the probability measures to correspond to distributions with finite first moment.
Preceding works in that direction are as follows. Breeden and Litzenberger (1978) and Dupire(1994),
show how option spreads deliver a unique probability measure; there are papers establishing
broader set of arbitrage relations between options such as Carr and Madan (2001)2 .
However 1) none of these papers made the bridge between calls and puts via the forward, thus
translating the relationships from arbitrage relations between options delivering a probability
distribution into the necessity of lining up to the mean of the distribution of the forward, hence
the risk-neutral one (in case the forward is arbitraged.) 2) Nor did any paper show that in
the absence of second moment (say, infinite variance), we can price options very easily. Our
methodology and proofs make no use of the variance. 3) Our method is vastly simpler, more
direct, and robust to changes in assumptions.
We make no assumption of general market completeness. Options are not redundant securities
and remain so. Table 1 summarizes the gist of the paper.3 4
29.2
proof
Define C(St0 , K, t) and P(St0 , K, t) as European-style call and put with strike price K, respectively,
with expiration t, and S0 as an underlying security at times t0 , t t0 , and St the possible value
of the underlying security at time t.
29.2.1 Case 1: Forward as risk-neutral measure
Rt
Rt
Define r = t1t0 t rs ds, the return of a risk-free money market fund and = t1t0 t s ds the
0
0
payout of the asset (continuous dividend for a stock, foreign interest for a currency).
2 See also Green and Jarrow (1987) and Nachman(1988). We have known about the possibility of risk neutral pricing
without dynamic hedging since Harrison and Kreps (1979) but the theory necessitates extremely strong and severely
unrealistic assumptions, such as strictly complete markets and a multiperiod pricing kernel
3 The famed Hakkanson paradox is as follows: if markets are complete and options are redudant, why would someone
need them? If markets are incomplete, we may need options but how can we price them? This discussion may have
provided a solution to the paradox: markets are incomplete and we can price options.
4 Option prices are not unique in the absolute sense: the premium over intrinsic can take an entire spectrum of values; it
is just that the put-call parity constraints forces the measures used for puts and the calls to be the same and to have the
same expectation as the forward. As far as securities go, options are securities on their own; they just have a strong link
to the forward.
356
29.2 proof
Table 28: Main practical differences between the dynamic hedging argument and the static Put-Call parity
with speading across strikes.
Type
Market Assumptions
Black-Scholes
Merton
Continuous
rebalancing.
1) Continuous
Markets,
no
gaps, no jumps.
Market
Completeness
Realism of
Assumptions
Convergence
Fitness
Reality
to
Put-Call Parity
with Spreading
Interpolative
static hedge.
1) Gaps and
jumps
acceptable.
Continuous
Strikes,
or
acceptable
number
of
strikes.
2) Ability to borrow and lend
underlying asset
for single forward date.
3) Low transaction costs in
trading options.
Requires
all
moments to be
finite. Excludes
the
class
of
slowly varying
distributions
Achieved
through
dynamic completeness
Low
Requires finite
1st moment (infinite variance is
acceptable).
In
probability
(uncertain; one
large
jump
changes expectation)
Pointwise
Portmanteau,
using
specific
distribution
adapted
to
reality
357
unique option pricing measure with neither dynamic hedging nor complete markets
We have the arbitrage forward price FtQ :
FtQ = S0
(1 + r)(tt0 )
S0 e(r)(tt0 )
(1 + )(tt0 )
(29.1)
by arbitrage, see Keynes 1924. We thus call FtQ the future (or forward) price obtained by arbitrage,
at the risk-neutral rate. Let FtP be the future requiring a risk-associated "expected return" m, with
expected forward price:
FtP = S0 (1 + m)(tt0 ) S0 em (tt0 ) .
(29.2)
Remark: By arbitrage, all tradable values of the forward price given St0 need to be equal to FtQ .
"Tradable" here does not mean "traded", only subject to arbitrage replication by "cash and
carry", that is, borrowing cash and owning the secutity yielding d if the embedded forward
return diverges from r.
29.2.2 Derivations
In the following we take F as having dynamics on its own irrelevant to whether we are in case
1 or 2 hence a unique probability measure Q.
Define = [0, ) = AK AcK where AK = [0, K] and AcK = (K, ).
Consider a class of standard (simplified)
probability spaces (, i ) indexed by i, where i is a
R
probability measure, i.e., satisfying di = 1.
Theorem 29.1.
For a given maturity T, there is a unique measure Q that prices European puts and calls by expectation
of terminal payoff.
This measure can be risk-neutral in the sense that it prices the forward FtQ , but does not have
to be and imparts rate of return to the stock embedded in the forward.
For a given maturity T, there exist two measures 1 and 2 for European calls and puts of
the same maturity and same underlying security associated with the valuation by expectation of
terminal payoff, which are unique such that, for any call and put of strike K, we have:
C=
and
P=
respectively, and where f C and f P are (St
K)+
f C d1 ,
(29.3)
f P d2 ,
(29.4)
and (K St )+ respectively.
Proof. For clarity, set r and to 0 without a loss of generality. By Put-Call Parity Arbitrage, a
positive holding of a call ("long") and negative one of a put ("short") replicates a tradable forward;
because of P/L variations, using positive sign for long and negative sign for short:
C(St0 , K, t) P(St0 , K, t) + K = FtP
(29.5)
358
(29.6)
29.2 proof
for all K
Now a Call spread in quantities
1
K ,
expressed as
Z
AcK
d1 .
(29.7)
Z
AK
d2 = 1
Z
AcK
d2 .
(29.8)
As semi-closed intervals generate the whole Borel -algebra on , this shows that 1 and 2
are unique.
The probability measures of puts and calls are the same, namely for each Borel set A in ,
1 (A) = 2 (A).
Proof. Combining Equations 29.5 and 29.6, dividing by
1
K
and taking K 0:
C(St0 , K, t) P(St0 , K, t)
+
=1
K
K
(29.9)
d1 =
Z
AcK
d2 ,
(29.10)
hence 1 (AK ) = 2 (AK ) for all K [0, ). This equality being true for any semi-closed interval,
it extends to any Borel set.
Puts and calls are required, by static arbitrage, to be evaluated at same as risk-neutral measure
Q as the tradable forward.
Proof.
FtP =
Ft dQ ;
(29.11)
f C (K) d1
f P (K) d1 =
Ft dQ K
(29.12)
359
unique option pricing measure with neither dynamic hedging nor complete markets
for all values of K.
29.3
Consider the case where Ft is observable, tradable, and use it solely as an underlying security
with dynamics on its own. In such a case we can completely ignore the dynamics of thenominal
log
F
S
underlying S, or use a non-risk neutral "implied" rate linking cash to forward, m = tt00 . the
rate m can embed risk premium, difficulties in financing, structural or regulatory impediments
to borrowing, with no effect on the final result.
In that situation, it can be shown that the exact same results as before apply, by remplacing
the measure Q by another measure Q . Option prices remain unique 5 .
29.4
comment
We have replaced the complexity and intractability of dynamic hedging with a simple, more
benign interpolation problem, and explained the performance of pre-Black-Scholes option operators using simple heuristics and rules, bypassing the structure of the theorems of financial
economics.
Options can remain non-redundant and markets incomplete: we are just arguing here for a
form of arbitrage pricing (which includes risk-neutral pricing at the level of the expectation of
the probability measure), nothing more. But this is sufficient for us to use any probability distribution with finite first moment, which includes the Lognormal, which recovers Black Scholes.
A final comparison. In dynamic heding, missing a single hedge, or encountering a single gap
(a tail event) can be disastrous as we mentioned, it requires a series of assumptions beyond the
mathematical, in addition to severe and highly unrealistic constraints on the mathematical. Under the class of fat tailed distributions, increasing the frequency of the hedges does not guarantee
reduction of risk. Further, the standard dynamic hedging argument requires the exact specification of the risk-neutral stochastic process between t0 and t, something econometrically unwieldy,
and which is generally reverse engineered from the price of options, as an arbitrage-oriented
interpolation tool rather than as a representation of the process.
Here, in our Put-Call Parity based methodology, our ability to track the risk neutral distribution is guaranteed by adding strike prices, and since probabilities add up to 1, the degrees of
freedom that the recovered measure Q has in the gap area between a strike price K and the next
strike up, K
R c+ K, are
R cseverely reduced, since the measure in the interval is constrained by the
difference A d A
d. In other words, no single gap between strikes can significantly
K
K+K
affect the probability measure, even less the first moment, unlike with dynamic hedging. In fact
it is no different from standard kernel smoothing methods for statistical samples, but applied to
the distribution across strikes.6
5 We assumed 0 discount rate for the proofs; in case of nonzero rate, premia are discounted at the rate of the arbitrage
operator
6 For methods of interpolation of implied probability distribution between strikes, see Avellaneda et al.(1997).
360
29.4 comment
The assumption about the presence of strike prices constitutes a natural condition: conditional
on having a practical discussion about options, options strikes need to exist. Further, as it is the
experience of the author, market-makers can add over-the-counter strikes at will, should they
need to do so.
361
BIBLIOGRAPHY
[1] Martha Amram and Nalin Kulatilaka. Real options:: Managing strategic investment in an
uncertain world. OUP Catalogue, 1998.
[2] Kenneth J Arrow. Aspects of the theory of risk-bearing (yrjo jahnsson lectures). Yrjo
Jahnssonin Saatio, Helsinki, 1965.
[3] Kenneth Joseph Arrow. Essays in the theory of risk-bearing, volume 1. Markham Publishing
Company Chicago, 1971.
[4] Philippe Artzner, Freddy Delbaen, Jean-Marc Eber, and David Heath. Coherent measures
of risk. Mathematical finance, 9(3):203228, 1999.
[5] Louis Bachelier. Thorie de la spculation. Gauthier-Villars, 1900.
[6] Kevin P Balanda and HL MacGillivray. Kurtosis: a critical review. The American Statistician,
42(2):111119, 1988.
[7] Nicholas Barberis. The psychology of tail events: Progress and challenges. American
Economic Review, 103(3):61116, 2013.
[8] Shlomo Benartzi and Richard H Thaler. Myopic loss aversion and the equity premium
puzzle. The quarterly journal of Economics, 110(1):7392, 1995.
[9] George Bennett. Probability inequalities for the sum of independent random variables.
Journal of the American Statistical Association, 57(297):3345, 1962.
[10] Serge Bernstein. Sur lextension du thorme limite du calcul des probabilits aux sommes
de quantits dpendantes. Mathematische Annalen, 97(1):159, 1927.
[11] Fischer Black and Myron Scholes. The pricing of options and corporate liabilities. The
journal of political economy, pages 637654, 1973.
[12] Marvin Blum. On the sums of independently distributed pareto variates. SIAM Journal on
Applied Mathematics, 19(1):191198, 1970.
[13] mile Borel. Les probabilits et la vie, volume 91. Presses universitaires de France, 1943.
[14] Jean-Philippe Bouchaud, J Farmer, and Fabrizio Lillo. How markets slowly digest changes
in supply and demand. (September 11, 2008), 2008.
[15] Leo Breiman. Probability, classics in applied mathematics, vol. 7. Society for Industrial and
Applied Mathematics (SIAM), Pennsylvania, 1992.
[16] L Brennan, I Reed, and William Sollfrey. A comparison of average-likelihood and
maximum-likelihood ratio tests for detecting radar targets of unknown doppler frequency.
Information Theory, IEEE Transactions on, 14(1):104110, 1968.
[17] VV Buldygin and Yu V Kozachenko. Sub-gaussian random variables. Ukrainian Mathematical Journal, 32(6):483489, 1980.
363
BIBLIOGRAPHY
[18] Rmy Chicheportiche and Jean-Philippe Bouchaud. The joint distribution of stock returns
is not elliptical. International Journal of Theoretical and Applied Finance, 15(03), 2012.
[19] VP Chistyakov. A theorem on sums of independent positive random variables and its applications to branching random processes. Theory of Probability & Its Applications, 9(4):640648,
1964.
[20] George M Constantinides and Anastasios G Malliaris. Portfolio theory. Handbooks in
operations research and management science, 9:130, 1995.
[21] DA Darling. The influence of the maximum term in the addition of independent random
variables. Transactions of the American Mathematical Society, 73(1):95107, 1952.
[22] Bruno De Finetti. Theory of Probability, volumes I and 2. Wiley, 1977.
[23] Emanuel Derman and Paul Wilmott. The financial modelers manifesto.
http://ssrn. com/abstract, volume 1324878, 2009.
In SSRN:
[24] Persi Diaconis and David Freedman. On the consistency of bayes estimates. The Annals of
Statistics, pages 126, 1986.
[25] Wolfgang Doeblin. Sur certains mouvements alatoires discontinus. Scandinavian Actuarial
Journal, 1939(1):211222, 1939.
[26] Wolfgang Doeblin. Sur les sommes dun grand nombre de variables alatoires indpendantes. Bull. Sci. Math, 63(2):2332, 1939.
[27] Joseph L Doob. Heuristic approach to the kolmogorov-smirnov theorems. The Annals of
Mathematical Statistics, 20(3):393403, 1949.
[28] Bradley Efron. Bayes theorem in the 21st century. Science, 340(6137):11771178, 2013.
[29] Jon Elster. Hard and soft obscurantism in the humanities and social sciences. Diogenes,
58(1-2):159170, 2011.
[30] Edwin J Elton and Martin J Gruber. Modern portfolio theory, 1950 to date. Journal of
Banking & Finance, 21(11):17431759, 1997.
[31] Paul Embrechts. Modelling extremal events: for insurance and finance, volume 33. Springer,
1997.
[32] Paul Embrechts and Charles M Goldie. On convolution tails. Stochastic Processes and their
Applications, 13(3):263278, 1982.
[33] Paul Embrechts, Charles M Goldie, and Nol Veraverbeke. Subexponentiality and infinite
divisibility. Probability Theory and Related Fields, 49(3):335347, 1979.
[34] M mile Borel. Les probabilits dnombrables et leurs applications arithmtiques. Rendiconti del Circolo Matematico di Palermo (1884-1940), 27(1):247271, 1909.
[35] Robert Engle. Garch 101: The use of arch/garch models in applied econometrics. Journal
of economic perspectives, pages 157168, 2001.
[36] CG Esseen. On the concentration function of a sum of independent random variables.
Probability Theory and Related Fields, 9(4):290308, 1968.
364
BIBLIOGRAPHY
[37] Kai-Tai Fang. Elliptically contoured distributions. Encyclopedia of Statistical Sciences, 2006.
[38] Doyne James Farmer and John Geanakoplos. Hyperbolic discounting is rational: Valuing
the far future with uncertain discount rates. 2009.
[39] William Feller. 1971an introduction to probability theory and its applications, vol. 2.
[40] William Feller. An introduction to probability theory. 1968.
[41] Bent Flyvbjerg. Phronetic planning research: theoretical and methodological reflections.
Planning Theory & Practice, 5(3):283306, 2004.
[42] Bent Flyvbjerg. From nobel prize to project management: getting risks right. arXiv preprint
arXiv:1302.3642, 2013.
[43] Shane Frederick, George Loewenstein, and Ted Odonoghue. Time discounting and time
preference: A critical review. Journal of economic literature, 40(2):351401, 2002.
[44] DA Freedman and PB Stark. What is the chance of an earthquake? NATO Science Series
IV: Earth and Environmental Sciences, 32:201213, 2003.
[45] David Freedman. Statistical models: theory and practice. Cambridge University Press, 2009.
[46] Rainer Froese. Cube law, condition factor and weightlength relationships: history, metaanalysis and recommendations. Journal of Applied Ichthyology, 22(4):241253, 2006.
[47] Xavier Gabaix. Power laws in economics and finance. Technical report, National Bureau
of Economic Research, 2008.
[48] Gerd Gigerenzer. Adaptive thinking: rationality in the real world. Oxford University Press,
New York, 2000.
[49] Gerd Gigerenzer and Henry Brighton. Homo heuristicus: Why biased minds make better
inferences. Topics in Cognitive Science, 1(1):107143, 2009.
[50] Gerd Gigerenzer and Daniel G Goldstein. Reasoning the fast and frugal way: models of
bounded rationality. Psychological review, 103(4):650, 1996.
[51] BV Gnedenko and AN Kolmogorov. Limit distributions for sums of independent random
variables (1954). Cambridge, Mass.
[52] Charles M Goldie. Subexponential distributions and dominated-variation tails. Journal of
Applied Probability, pages 440442, 1978.
[53] Daniel Goldstein and Nassim Taleb. We dont quite know what we are talking about when
we talk about volatility. Journal of Portfolio Management, 33(4), 2007.
[54] Lars Peter Hansen and Thomas J Sargent. Robustness. Princeton university press, 2008.
[55] Clyde D Hardin Jr. Skewed stable variables and processes. Technical report, DTIC Document, 1984.
[56] Espen Gaarder Haug and Nassim Nicholas Taleb. Option traders use (very) sophisticated
heuristics, never the blackscholesmerton formula. Journal of Economic Behavior & Organization, 77(2):97106, 2011.
365
BIBLIOGRAPHY
[57] Martin Hlusek. On distribution of absolute values. 2011.
[58] Wassily Hoeffding. Probability inequalities for sums of bounded random variables. Journal
of the American statistical association, 58(301):1330, 1963.
[59] Chi-fu Huang and Robert H Litzenberger. Foundations for financial economics. Prentice Hall,
1988.
[60] IA Ibragimov and KE Chernin. On the unimodality of geometric stable laws. Theory of
Probability & Its Applications, 4(4):417419, 1959.
[61] Jean-Pierre Kahane. Some random series of functions. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 2nd edition, 1993.
[62] Daniel Kahneman and Amos Tversky. Prospect theory: An analysis of decision under risk.
Econometrica, 47(2):263291, 1979.
[63] Harry Kesten. A sharper form of the doeblin-lvy-kolmogorov-rogozin inequality for
concentration functions. Mathematica Scandinavica, 25:133144, 1969.
[64] John M Keynes. A treatise on probability. 1921.
[65] Leopold Kohr. Leopold kohr on the desirable scale of states. Population and Development
Review, 18(4):745750, 1992.
[66] A.N. Kolmogorov. Selected Works of AN Kolmogorov: Probability theory and mathematical
statistics, volume 26. Springer, 1992.
[67] David Laibson. Golden eggs and hyperbolic discounting. The Quarterly Journal of Economics,
112(2):443478, 1997.
[68] Paul Lvy and M mile Borel. Thorie de laddition des variables alatoires, volume 1. GauthierVillars Paris, 1954.
[69] Andrew Lo and Mark Mueller. Warning: physics envy may be hazardous to your wealth!
2010.
[70] Michel Love. Probability Theory. Foundations. Random Sequences. New York: D. Van Nostrand Company, 1955.
[71] Michel Loeve. Probability theory, vol. ii. Graduate texts in mathematics, 46:0387, 1978.
[72] HL MacGillivray and Kevin P Balanda. Mixtures, myths and kurtosis. Communications in
Statistics-Simulation and Computation, 17(3):789802, 1988.
[73] M Machina and M Rothschild. Risk.[in:] utility and probability. the new pal-grave. red. j.
eatwell, m. milgate, p. newman, 1987.
[74] Mark Machina and Michael Rothschild. Risk. in the new palgrave dictionary of economics,
edited by steven n. durlauf and lawrence e. blume, 2008.
[75] Harry Markowitz. Portfolio selection*. The journal of finance, 7(1):7791, 1952.
[76] Harry M Markowitz. Portfolio selection: efficient diversification of investments, volume 16.
Wiley, 1959.
366
BIBLIOGRAPHY
[77] T Mikosch and AV Nagaev. Large deviations of heavy-tailed sums with applications in
insurance. Extremes, 1(1):81110, 1998.
[78] Frederick Mosteller and John W Tukey. Data analysis and regression. a second course in
statistics. Addison-Wesley Series in Behavioral Science: Quantitative Methods, Reading, Mass.:
Addison-Wesley, 1977, 1, 1977.
[79] Aleksandr Viktorovich Nagaev. Integral limit theorems taking into account large deviations when cramrs condition does not hold. ii. Teoriya Veroyatnostei i ee Primeneniya,
14(2):203216, 1969.
[80] Sergey V Nagaev. Large deviations of sums of independent random variables. The Annals
of Probability, 7(5):745789, 1979.
[81] Sergey Victorovich Nagaev. Some limit theorems for large deviations. Theory of Probability
& Its Applications, 10(2):214235, 1965.
[82] SV Nagaev and IF Pinelis. Some inequalities for the distribution of sums of independent
random variables. Theory of Probability & Its Applications, 22(2):248256, 1978.
[83] Gloria Origgi. Is trust an epistemological notion? Episteme, 1(01):6172, 2004.
[84] Athanasios Papoulis. Probability, random variables, and stochastic processes, 1991.
[85] Giovanni Peccati and Murad S Taqqu. Wiener Chaos: Moments, Cumulants and Diagrams, a
Survey with Computer Implementation, volume 1. Springer, 2011.
[86] Valentin V Petrov. Limit theorems of probability theory. 1995.
[87] Iosif Pinelis. On the characteristic function of the positive part of a random variable. arXiv
preprint arXiv:1309.5928, 2013.
[88] Steven Pinker. The better angels of our nature: Why violence has declined. Penguin, 2011.
[89] EJG Pitman. Subexponential distribution functions. J. Austral. Math. Soc. Ser. A, 29(3):337
347, 1980.
[90] John W Pratt. Risk aversion in the small and in the large. Econometrica: Journal of the
Econometric Society, pages 122136, 1964.
[91] Yu V Prokhorov. An extremal problem in probability theory. Theory of Probability & Its
Applications, 4(2):201203, 1959.
[92] Yu V Prokhorov. Some remarks on the strong law of large numbers. Theory of Probability
& Its Applications, 4(2):204208, 1959.
[93] Colin M Ramsay. The distribution of sums of certain iid pareto variates. Communications
in StatisticsUTheory
and Methods, 35(3):395405, 2006.
[94] BA Rogozin. An estimate for concentration functions. Theory of Probability & Its Applications,
6(1):9497, 1961.
[95] BA Rogozin. The concentration functions of sums of independent random variables. In Proceedings of the Second Japan-USSR Symposium on Probability Theory, pages 370376. Springer,
1973.
367
BIBLIOGRAPHY
[96] Michael Rothschild and Joseph E Stiglitz. Increasing risk: I. a definition. Journal of Economic
theory, 2(3):225243, 1970.
[97] Michael Rothschild and Joseph E Stiglitz. Increasing risk ii: Its economic consequences.
Journal of Economic Theory, 3(1):6684, 1971.
[98] Gennady Samorodnitsky and Murad S Taqqu. Stable non-Gaussian random processes: stochastic models with infinite variance, volume 1. CRC Press, 1994.
[99] Leonard J Savage. The foundations of statistics. Courier Dover Publications, 1954.
[100] Mr Christian Schmieder, Mr Tidiane Kinda, Mr Nassim N Taleb, Elena Loukoianova, and
Mr Elie Canetti. A new heuristic measure of fragility and tail risks: application to stress testing.
Number 12-216. Andrews McMeel Publishing, 2012.
[101] Laurent Schwartz. Thorie des distributions. Bull. Amer. Math. Soc. 58 (1952), 78-85 DOI:
http://dx. doi. org/10.1090/S0002-9904-1952-09555-0 PII, pages 00029904, 1952.
[102] William F Sharpe. Mutual fund performance. Journal of business, pages 119138, 1966.
[103] Vernon L Smith. Rationality in economics: constructivist and ecological forms. Cambridge
University Press, Cambridge, 2008.
[104] Emre Soyer and Robin M Hogarth. The illusion of predictability: How regression statistics
mislead experts. International Journal of Forecasting, 28(3):695711, 2012.
[105] N N Taleb and R Douady. Mathematical definition, mapping, and detection of (anti)
fragility. Quantitative Finance, 2013.
[106] Nassim Taleb. Fooled by randomness: The hidden role of chance in life and in the markets.
Random House Trade Paperbacks, 2001/2005.
[107] Nassim N Taleb and Daniel G Goldstein. The problem is beyond psychology: The
real world is more random than regression analyses. International Journal of Forecasting,
28(3):715716, 2012.
[108] Nassim Nicholas Taleb. Dynamic Hedging: Managing Vanilla and Exotic Options. John Wiley
& Sons (Wiley Series in Financial Engineering), 1997.
[109] Nassim Nicholas Taleb. Errors, robustness, and the fourth quadrant. International Journal
of Forecasting, 25(4):744759, 2009.
[110] Nassim Nicholas Taleb. The Black Swan:: The Impact of the Highly Improbable Fragility. Random House Digital, Inc., 2010.
[111] Nassim Nicholas Taleb. Antifragile: things that gain from disorder. Random House and
Penguin, 2012.
[112] Albert Tarantola. Inverse problem theory: Methods for data fitting and model parameter estimation. Elsevier Science, 2002.
[113] Jozef L Teugels. The class of subexponential distributions.
3(6):10001011, 1975.
368
BIBLIOGRAPHY
[114] Peter M Todd and Gerd Gigerenzer. Ecological rationality: intelligence in the world. Evolution
and cognition series. Oxford University Press, Oxford, 2012.
[115] Bence Toth, Yves Lemperiere, Cyril Deremble, Joachim De Lataillade, Julien Kockelkoren,
and J-P Bouchaud. Anomalous price impact and the critical nature of liquidity in financial
markets. Physical Review X, 1(2):021006, 2011.
[116] Jack L Treynor. How to rate management of investment funds. Harvard business review,
43(1):6375, 1965.
[117] Lenos Trigeorgis. Real options: Managerial flexibility and strategy in resource allocation. MIT
press, 1996.
[118] Amos Tversky and Daniel Kahneman. Judgment under uncertainty: Heuristics and biases.
science, 185(4157):11241131, 1974.
[119] Vladimir V Uchaikin and Vladimir M Zolotarev. Chance and stability: stable distributions and
their applications. Walter de Gruyter, 1999.
[120] Willem Rutger van Zwet. Convex transformations of random variables, volume 7. Mathematisch centrum, 1964.
[121] Rafa Weron. Levy-stable distributions revisited: tail index> 2 does not exclude the levystable regime. International Journal of Modern Physics C, 12(02):209223, 2001.
[122] Stephen J Wolfe. On the local behavior of characteristic functions. The Annals of Probability,
pages 862866, 1973.
[123] Stephen James Wolfe. On the unimodality of infinitely divisible distribution functions.
Probability Theory and Related Fields, 45(4):329335, 1978.
[124] IV Zaliapin, Yan Y Kagan, and Federic P Schoenberg. Approximating the distribution of
pareto sums. Pure and Applied geophysics, 162(6-7):11871228, 2005.
[125] Vladimir M Zolotarev. One-dimensional stable distributions, volume 65. American Mathematical Soc., 1986.
369
INDEX
Skin in the Game, 57
Antifragility, 265, 269, 299, 325, 331
Black Swan, 2022, 47, 50, 52, 53, 61, 70, 90, 111,
195, 215, 227, 252, 308, 331
Citation Rings, 56
Concavity/Convexity, iii, iv, xvii, 6, 7, 45, 47,
48, 70, 72, 74, 86, 89, 113, 252254, 256,
258, 266, 267, 269, 275, 276, 283, 288,
298, 302, 305, 306, 309, 310, 313, 330,
331, 346
Econometrics, 54, 89, 186, 190, 195, 196, 283
Egomaniac as an rebuttal, 195
Empiricism (Naive Empiricism), 47
Finance Artists, 77
Fragility, 46, 49, 125, 252, 254, 275, 315
Law of Large Numbers, xvi, 20, 23, 39, 51, 61,
90, 115, 119, 120, 128, 131, 132, 145,
147, 148, 175, 177
Loss Function, 62, 63
Measure Theory, 51, 52, 156
Pinker Problem, 180
Pseudo-biases, 126
Skin in the Game, xvi, 9, 35, 58, 196, 239, 247
Triffat Fallacy, 57
Turkey Problem, 61, 90, 177
Via Negativa, 36
370