Loc Vol in Multi Dimension
Loc Vol in Multi Dimension
Jesper Andreasen
Saxo Bank, Copenhagen
[email protected]
Outline
Prelude.
Conclusion.
2
References
Austing, P (2011): “Repricing the Cross Smile: An Analytic Joint Density.” Risk July.
3
Piterbarg, V (2011): “Farka’s Lemma, Spread Options, and Linear Programming“. Risk
August.
Shelton, D (2015): “Interpolating the Smile with Path-Dependent Local Volatility.” ICBI
Global Derivatives.
4
Prelude
To simplify the exposition and save time I will work with the model in its simplest form.
… but interest rates are more complicated and you will have to consult future material about
that.
5
Multi Asset Arbitrage
Assume interest rates and dividends are zero, and set the start prices to be si (0) 0 .
6
Absence of arbitrage implies that the covariance matrix
( wi si (t ))2 0 (5)
i
7
The arbitrage portfolio is in this case given by
8
Multi Asset Arbitrage -- Notes
10
Minimal Multi Asset Models
… where the local correlation is given from the volatility of the spread
11
So the model is parameterised from the local spread volatilities { ij (si s j )} which are given as
function of the spread levels.
cij 1 2cij
0 ij (t,k )2 2 (13)
t 2 k
{ (t, si , s j )} (15)
The construction through spread volatility rather than correlation is similar to Austing (2011).
13
Minimal Model and Arbitrage
15
Obviously, a quite specific case but it suggests that if the volatility smiles are more
pronounced in the spread directions than in the primal directions, then it may be that the
spread is the overall volatility driver …
16
Discrete Time
Secondly, in real applications the model setup will have to be somewhat modified relative to
what we have outlined so far.
Thirdly, it would be nice to be able to handle various model extensions such as stochastic
volatility and stochastic interest rates.
It turns out that these modifications and extensions are relatively straightforward to handle in
discrete time.
17
Discrete Time Minimal Model
18
As in the continuous time case, the model is specified through spread volatility rather than
correlation.
19
Monte-Carlo Pricing
In a Monte-Carlo simulation over samples {}, the value of an option that expiries as time th1
can be written as a sum over Bachelier’s formula
20
b( ,k;s,v) (s k )( x) v ( x) , x svk (20)
This is so because over each time step, si s j has a conditional normal distribution – due to the
Euler discretisation.
21
Monte-Carlo Calibration
If we wish to calibrate the model to the strikes {kij1 , ,kijL} at expiry th1 then we parameterise
the volatility function ij (th; si s j ) with L parameters.
(t ,)
inf ( c(th1,kijl ) cˆ(th1,kijl ) )2 (21)
ij h l
mc model price market price
Note that the calibration of { ij (th,)} is independent for different pairs (i, j) .
22
After calibration to the options for each spread pair (i, j) then can we construct the correlation
matrix P {ij}.
The methodology can also be used for correlation structures that are not minimal.
ij ij (aij s) (22)
… for constant vectors aij , then the calibration problem is still independent over the different
pairs (i, j) .
We do, however, not yet have a methodology for optimal choice of directional vectors {aij} .
23
Positive Definiteness and Bootstrap
To make it positive definite, decompose into the product P OO ' , chop negative eigenvalues
and rescale to obtain units along the diagonal.
Once done with calibration of the time step th th1 , we simulate forward to calibrate the
model to the time step th1 th2 .
24
Catch-Up and Discrete Quotes
If fiddling with the correlation (or covariance) matrix is necessary then the model will not hit
the option prices at the particular expiry.
However, the bootstrap methodology will attempt to catch-up at the next expiry.
This is so because the Monte-Carlo pricing/calibration (19) works a bit like updating local
volatility according to
cˆ(th1) c(th )
(th )2 2 th1 th [ kk
ˆ
c(th 1
)]1 (23)
maturity spread
computed as
market mod el
25
Equation (23) is a trick that has been used with success in finite difference implementation of
local volatility models.
Hence, the model fit will only be broken at the expiry with positive definiteness problems --
not necessarily at subsequent expiries.
Hence, we do not rely on perfectly smooth and arbitrage free volatility surfaces in all
directions.
26
Timelines
But we can insert extra simulation time points as we wish inside each calibration time bucket
[th,th1] .
As long as we keep the volatilities and correlations constant over these extra time points.
In that sense, the model looks a bit like the model in Shelton (2015).
Here, we use Monte-Carlo rather than numerical integration and this makes our model
applicable to high dimensions.
27
28
Applications and Extensions
Foreign exchange: Note that log-normal form and currency translations are necessary.
Equities: Calibrate to basket rather than spread options. Potentially, using notions of average
correlation.
Note that non-trivial dividend models can also be handled this way.
Interest rates: Non-trivial but interesting. Both multifactor Cheyette and LMM type models
can be constructed.
The interest rate models can potentially calibrate simultaneously to cap/swaption smiles and
smiles of spread and/or mid-curve options.
29
Stochastic volatility and even rough volatility is straightforward.
It is also possible to do models that simultaneously calibrate to SP500 and VIX smiles.
… and more.
30
Numerical Implementation
So far, we have implemented a multi factor Cheyette model for interest rates and a multi price
model for FX and equities.
The intention is to combine the two model types to a Next Gen Beast.
Adjoint differention (AAD) risk has been implemented for the interest rate model.
31
Numerical Performance
5 Ccy FX model calibration to 5 strikes in all crosses on expiries 1m, 2m, … 12m:
4 factor interest rate model. Calibration to 5 strikes in all crosses (spread options and mid
curves) in tenors 3m, 2y, 10y, 30y on expiries 1y, 2y, … , 10y.
32
6 factor interest rate model. Calibration to 5 strikes in all crosses in tenors 3m, 2y, 5y, 10y,
20y, 30y on expiries 1y, 2y, … , 10y.
Numerical performance is definitely ok, but a tad slower than my audience is used to.
33
Discrete Model Summary
The method does not require full continuous surfaces of arbitrage free option prices. Discrete
points are sufficient.
If arbitrage or positive definiteness is broken at a particular time point, the methodology will
attempt to catch-up at subsequent time-steps.
It applies to cases without non-trivial forward equations such as interest rates and
commodities.
34
Non-minimal correlation can also be handled but we don’t know yet how to optimally choose
the directions {aij}.
35
Conclusion
We have presented an approach to multi factor local volatility with associated Monte-Carlo
calibration methodology that is performing, flexible and general.
Next steps:
- GPU/TensorFlow acceleration.
36