Anking Earthquake Forecasts Using Proper Scoring Rules Inary Events in A Low Probability Environment
Anking Earthquake Forecasts Using Proper Scoring Rules Inary Events in A Low Probability Environment
A P REPRINT
A BSTRACT
Operational earthquake forecasting for risk management and communication during seismic se-
quences depends on our ability to select an optimal forecasting model. To do this, we need to compare
the performance of competing models in prospective experiments, and to rank their performance
according to the outcome using a fair, reproducible, and reliable method, usually in a low-probability
environment. The Collaboratory for the Study of Earthquake Predictability (CSEP) conducts prospec-
tive earthquake forecasting experiments around the globe. One metric employed to rank competing
models is the Parimutuel Gambling score. We examine the suitability of this score for ranking
earthquake forecasts. We prove analytically that this score is in general ‘improper’, meaning that,
on average, it does not prefer the data generating model assuming this is known. In the special case
where it is proper, we show it can still be used improperly. We compare its performance with two
commonly-used proper scores (the Brier and logarithmic scores) using confidence intervals to account
for the uncertainty around the observed score difference. We think that using confidence intervals
enables a rigorous approach to distinguish between the predictive skills of candidate forecasts in
addition to their rankings. Our analysis shows that the Parimutuel Gambling score is biased, and the
direction of the bias depends on the forecasts taking part in the experiment. Our findings suggest
the Parimutuel Gambling score should not be used to distinguishing between multiple competing
forecasts, and for care to be taken in the case where only two are being compared.
Keywords Earthquakes forecasts comparison · Proper scoring rules · CSEP · Parimutuel gambling score
1 Introduction
Probabilistic earthquake forecasts are used to estimate the spatial and/or temporal evolution of seismicity and have
potential utility during earthquake sequences, including those following notable earthquakes. For example, they have
been applied to forecast (pseudoprospectively) the seismicity that followed the Darfield earthquake and in turn led to
the 2011 Christchurch earthquake [Rhoades et al., 2016], and to monitor induced seismicity at Groningen [Bourne et al.,
2018]. In Italy, earthquake probabilistic forecasts and ground-motion hazard forecasts are produced on a regular basis
by the Instituto Nazionale di Geofisica e Vulcanologia (INGV) to inform the Italian government on the risk associated
with natural hazard [Marzocchi et al., 2014]. INGV is working to use probabilistic forecasts as a basis for modelling
Ranking earthquake forecasts using proper scoring rules A P REPRINT
important quantities for operational loss forecasting such as the number of evacuated residents, the number of damaged
infrastructure, the number of fatalities [Iervolino et al., 2015]. A wider uptake requires further demonstrations of the
operational utility of the forecasts, and in presence of multiple alternative models, a fair and rigorous method to express
a preference for a specific approach is needed. The Collaboratory for the Study of Earthquake Predictability (CSEP,
see Jordan [2006], Zechar et al. [2010a], Schorlemmer et al. [2018]) is a global community initiative that seeks to
make earthquake research more rigorous and open-science. This is done by comparing forecasts against future data in
competition with those from other models through prospective testing in pre-defined testing regions. In this paper, we
focus on comparing different forecasts that can be made from such competing models in the light of observed data.
In statistics, a common approach to compare probabilistic forecasts is the use of scoring rules [Gneiting and Raftery,
2007]. Scoring rules have been widely applied in many fields of science to measure the quality of a forecasting model
and to rank competing models based on their consistency with the observed data and the degree of uncertainty around
the forecast itself. Much of the underlying methodology and concepts (such as what it means to be a "good" forecast)
have been developed for weather forecasts [Murphy, 1993, Jolliffe and Stephenson, 2003]. A positively oriented scoring
rule, to be effective, has to be proper, which simply means that the highest score is achieved, on average, by the
forecasting model "closer" to the distribution that has generated the observations. Various meaning of "closer" can be
used depending on the context and the use that will be made of the forecasting model under evaluation, thus, a variety
of proper scoring rules exists. Proper scoring rules are mathematically appealing for a range of different tasks: they can
be used as utility function tailored to the problem at hand, they can be used as loss functions in parametric estimation
problems and they can be used to rank competing models based on different aspects of the phenomenon under analysis
[Rosen, 1996, Hyvärinen and Dayan, 2005, Hernández-Orallo et al., 2012].
CSEP aims to compare the predictive performance of diverse earthquake forecasts in a rigorous and reproducible
fashion. The forecasts themselves are generated by underlying physical, stochastic or hybrid models using a variety of
input data such as past seismicity, deformation rates, fault maps, etc [Field et al., 2014, Steacy et al., 2014, Bayliss et al.,
2020]. The two most widely used types are alarm-based forecasts and probabilistic forecasts. The first class of forecasts
is usually expressed as a binary statement ("alarm" or "not alarm") based on the value of a precursory alarm function.
In contrast, probabilistic forecasts, as intended in past CSEP experiments [Schorlemmer and Gerstenberger, 2007a],
provide a distribution for the number of earthquakes. They can be expressed as grid-based forecasts (providing the
expected number of events in each space-time-magnitude bin) or as catalogue-based (providing a number of simulated
catalogues, Savran et al.,2020 ). The forecasts are variously compared using a suite of community-endorsed tests.
Depending upon the forecasts at hand, three common challenges are the need for a reference model, how to handle bins
(or regions) for which the forecaster didn’t provide a forecast and, the need to specify a likelihood. The latter has been
partially solved by the possibility of considering a pseudo-likelihood [Savran et al., 2020].
Molchan diagrams [Zechar and Jordan, 2008] and the area-skill score [Zechar and Jordan, 2010] do not need a likelihood
and can be used to compare both alarm-based and probabilistic forecasts together. However, they need a reference model
for assessing the significance of the results. This can be problematic because specifying a credible reference model is a
difficult task [Stark, 1997, Luen et al., 2008, Marzocchi and Zechar, 2011]. Likelihood-based tests [Schorlemmer et al.,
2007, Zechar et al., 2010b, Rhoades et al., 2011, Schneider et al., 2014] allow for pairwise comparison without the need
of a reference model, but can only be applied to probabilistic forecasts. Further, methods for grid-based forecasts rely on
the Poisson assumption, which has been observed to be not realistic [Werner and Sornette, 2008]. Moreover, pairwise
comparison may lead to paradoxical results like model A is preferred to model B which is preferred to model C which
is preferred to model A [Zechar et al., 2013]. Bayesian methods have been proposed [Marzocchi et al., 2012] but they
also rely on the Poisson assumption. Catalogue-based forecasts, can be evaluated using a pseudo-likelihood approach
[Savran et al., 2020] which does not rely on the Poisson assumption and enable information gains and likelihood ratios
to be used. However, the latter are unbounded and sensitive to low-probability events, meaning that they can be unduly
influenced by a few observations [Holliday et al., 2005, Zechar and Zhuang, 2014]. Lastly, in past experiments such as
the Regional Earthquake Likelihood Models (RELM) [Field, 2007], forecasters did not provide a forecast for all bins,
some of them were left as missing value; for the methods outlined above, making a comparison is complex, given that
considering only the overlapping space-time-magnitude volume may be too restrictive and introduce unfairness in the
evaluation.
Zhuang [2010], Zechar and Zhuang [2014] tried to overcome the difficulties outlined above by introducing the parimutuel
gambling score, which provides a framework to evaluate different types of forecasts, with no need to explicitly specify
a reference model or a likelihood, and with the ability to handle missing values in an intuitive way. This approach
is based on the idea that alarm-based forecasts could be imagined as gamblers engaged in a game called the seismic
roulette, where Nature controls the wheel [Main, 1997, Kossobokov, 2004, 2006]. In this framework, the forecasters
are the gamblers, a forecast consists of a collection of probabilities for observable events (bets) like observing at least
one earthquake in a specified space-time-magnitude bin. Each bin represents a bet and the probability assigned by the
forecaster represents the amount of money wagered. The observations consist of binary variables taking value 1 if the
2
Ranking earthquake forecasts using proper scoring rules A P REPRINT
event occurs and zero otherwise. The forecaster gets a reward depending on the forecasted probability and the actual
observation of an event or not. The forecasts are ranked based on their rewards. In this sense, the parimutuel gambling
score is a positively oriented score (the higher, the better) for binary probabilistic forecasts. In this paper, we prove that
the parimutuel gambling score is not proper in general but only in a specific situation and we compare its performance
with two proper alternatives: the Brier [Brier, 1950] and the logarithmic [Good, 1952] score.
To fairly compare the performance of the scores in a realistic framework, we use simulated data from a known model
and we compare it with alternative models. In doing that, it is crucial to account for the uncertainty in the observed
score difference. In fact, properness ensures that, at least on average, the scoring rule provides the correct ranking.
However, the score calculated from any finite set of observations could be far from its average and, therefore, we need to
account for uncertainty. In this paper, we show how to express a preference towards a model using confidence intervals
for the expected score difference. This method introduces the possibility of not expressing a preference. Considering
this outcome is potentially useful because it indicates that, for a scoring rule, the forecasts have similar performances,
or the data are not enough to distinguish between models, or the bins’ dimension is offset (too large or too small).
In summary, the main goal of this article is to present the notion of proper scoring rules for probabilistic forecasts of
binary events: why is it crucial for a scoring rule to be proper and what are the consequences of using an improper
one? In Section 2 we define a proper and a strictly proper scoring rule, introduce the well known Brier and log scores
as examples, and give a brief proof of their propriety. In Section 3, we introduce the parimutuel gambling score
and analytically explore its improperness in the context of a forecast for a single bin. If a score is proper for single
bins, then, the average score of different bins is also a proper score [Gneiting and Raftery, 2007]. In Section 4, we
generalise to the case where we have multiple bins but with the same probability, Multiple Bins Single Probability.
This case is equivalent to considering the activity rate in each bin as independent and identically distributed. This is a
significant assumption but allows us to calculate analytically the confidence intervals and the probability of expressing a
preference for a given model. We generalise further to the case in which we have multiple bins but with a different
probability for each bin, Multiple Bins Multiple Probabilities. In this case, we do not have analytical results and
we are required to use approximate confidence intervals and simulations to calculate the probability of expressing
a preference. These simulations are now close to a real forecast scenario. We illustrate this case using simulations
from the time-independent 5 year adaptively-smoothed forecast for Italy [Werner et al., 2010]. We choose this model
because the adaptively-smoothed approach performed well across multiple metrics in the RELM experiment [Zechar
et al., 2013] and, as a result, was incorporated into the California seismic hazard map produced by the third Uniform
California Earthquake Rupture Forecast model (UCERF3) [Field et al., 2014].
2 Proper Scores
Scoring rules quantify the quality of probabilistic forecasts, allowing them to be ranked. The quality depends on both
the predictive distribution, produced by the model in true prospective mode, and on the subsequent observations. A
scoring rule is a function of the forecast and the data measuring two factors: the consistency between predictions and
observations and the sharpness of the prediction. Consistency assesses the calibration of the model, how well the
forecast and the data agree, and is a joint property of the forecast and the data. Sharpness is a measure of the forecast
uncertainty and is a property of the forecast only. Different scoring rules measure the consistency and the sharpness of a
forecast differently. As in Gneiting and Raftery [2007], we call S(P |x) the score for forecast P given the observation x.
In general, we use capital letters for random variables, lowercase letters for scalar quantities such as realizations of a
random variable (everything that is not random) and bold letters represents vectors. The only exception is N which
represents the number of bins.
Thus, a scoring rule, given a forecast P , is a function of the observation only S(P |·) : X → [−∞, ∞] where X is the
set of all possible values of x. For consistency, we will use a positively orientated convention, where a larger score
indicates a better forecast. Assuming that the observations are samples from a random variable X ∈ X with true
distribution Q, the score S(P |X) is a random variable itself, since it is a function of the random variable X. We define
S E (P |Q) as the expected value of the scoring rule under the true distribution Q:
A positively oriented scoring rule S is said to be proper if, for any forecast P and any true distribution Q, S E (Q|Q) ≥
S E (P |Q) holds. It is said to be strictly proper if S E (Q|Q) = S E (P |Q) if and only if P = Q. Propriety is essential, as
it incentivises the assessor to be objective and to use the forecast P "closer" to the true distribution Q. Different scoring
rules rely on different meanings of closer. Also, proper scores can be used as loss functions in parameter estimation; in
fact, since the likelihood assigned by a model to the observations can be seen as a proper scoring rule, the maximum
likelihood estimator can be viewed as optimizing a score function [Huber, 1992]. Investigating the ability of a score
3
Ranking earthquake forecasts using proper scoring rules A P REPRINT
of distinguishing between different instances of the same model (with different parameters values) may bring insight
regarding parameters identifiability.
Here, we are interested in scoring rules for binary variables, in which the variable X can be only 0 or 1, namely
X ∈ {0, 1}. Grid-based earthquake forecasts divide the region of interest into regular space-time-magnitude bins (e.g.
the spatial region is divided in bins of 0.1 × 0.1 degrees, the magnitude by 0.1 magnitude units, and the time is one
5-year bin), and the forecasters estimate the expected number of earthquakes per bin. In this case, for example, the
binary variable might be 0 for empty bins and 1 if at least one event occurs. The forecasts may be ranked based on the
average score across different bins [Zechar et al., 2013].
Considering a single bin, for grid-based binary forecasts, where both the forecast P and the true distribution Q are
specified by just one number: the probability of X being 1. We call p the probability assigned to the event X = 1 by
the forecaster, and p∗ denotes the true probability. Thus, the expectation is given by
S E (P |Q) = S E (p|p∗ ) = p∗ S(p|1) + (1 − p∗ )S(p|0). (2)
A scoring rule of this type is proper if, for any p ∈ [0, 1] and any p ∈ [0, 1], we have
S E (Q|Q) ≥ S E (P |Q).
The properness of a score ensures that given two models p1 , p2 , the model with the greatest expected score S E (pi |Q)
is the closest to the true p∗ . This notion can be generalized to rank a set of k forecasts p1 , ..., pk according to their
expected scores.
Two of the most widely used strictly proper scoring rules, for binary data, are the Brier (or quadratic) score [Brier, 1950]
and the logarithmic score [Good, 1952]. These are good candidates for evaluating this class of earthquake forecasts.
Here we give the definitions of these two scores, including brief proofs of their propriety.
The positively oriented Brier score [Brier, 1950] for a categorical variable X (the binary case is obtained considering
only two possible outcomes) can be defined by:
X
SB (P |x) = − [p(z) − I(z = x)]2 , (3)
z∈X
where X is the set of possible outcomes, p(z) is the forecasted probability of the event X = z, and I(z = x) is an
indicator function assuming value 1 if z = x and 0 otherwise. This definition differs from the original only in the sign,
since the original Brier score is negatively oriented.
The ordinary Brier score for binary events is the special case X = {0, 1}, with p = p(1) and 1 − p = p(0):
−2(p − 1)2 , x = 1,
SB (p|x) = −[(1 − p) − (1 − x)]2 − (p − x)2 = −2(p − x)2 =
−2p2 , x = 0,
which has expectation
E
SB (p|p∗ ) = −2p∗ (p − 1)2 − 2(1 − p∗ )p2 (4)
∗
under the true event probability p . Taking the derivative with respect to p and imposing it equal zero, we find that the
value p = p∗ uniquely maximizes the function SB E
(p|p∗ ) which proves that the Brier score is strictly proper.
4
Ranking earthquake forecasts using proper scoring rules A P REPRINT
Given an observation x, to express a preference between two forecasts p1 and p2 , an important quantity is the score
difference ∆.
S(p1 |0) − S(p2 |0) with prob 1 − p∗ ,
∆(p1 , p2 , x) = S(p1 |x) − S(p2 |x) =
S(p1 |1) − S(p2 |1) with prob p∗ .
In principle, if the expected value of ∆ is positive we tend to prefer the first forecast, vice versa if it is negative.
Considering the observation as a Bernoulli random variable X ∼ Ber(p∗ ), the difference ∆(p1 , p2 , X) is also a binary
random variable, assuming the values ∆0 = ∆(p1 , p2 , 0), ∆1 = ∆(p1 , p2 , 1) with probabilities 1 − p∗ and p∗ . The
distribution of ∆(p1 , p2 , X) is therefore completely determined by the distribution of X:
∆(p1 , p2 , X) = X∆1 + (1 − X)∆0 = ∆0 + X(∆1 − ∆0 ). (9)
It follows that the expected value and variance of ∆(p1 , p2 , X) are determined by the properties of X:
E[∆(p1 , p2 , X)] = ∆0 + E[X](∆1 − ∆0 ) = ∆0 + p∗ (∆1 − ∆0 ) (10)
2 ∗ ∗ 2
V[∆(p1 , p2 , X)] = V[X](∆1 − ∆0 ) = p (1 − p )(∆1 − ∆0 ) (11)
We can give an alternative definition of the properness based on the random variable ∆(p1 , p2 , X). In fact, a scoring
rule S is said to be proper if EX [∆(p, p∗ , X)] ≤ 0 when p 6= p∗ , no forecast have an expected score higher than the
data generating model p∗ . However, they can achieve the same score. S is strictly proper if EX [∆] = 0 if and only if
p = p∗ , the highest score, on average, is achieved only by the data generating model. The definition implies, also, that
proper scoring rules are invariant under linear transformations, in the sense that, a linear transformation of a proper
score yields another proper score and the operation does not change the ranking.
Figure 1 reports the expected score difference between a candidate forecast p and the true value p∗ = 0.001 using the
Brier and the log score. The value p∗ = 0.001 was chosen to be comparable to the estimated probability of having
an event with magnitude greater than 5.5 calculated the days before the L’Aquila earthquake in the neighbourhood of
where it struck [Fig. 4 in Marzocchi and Lombardi [2009]]. To enable a visual comparison, the expected Brier score
values have been normalized to match the curvature of the log score when p = p∗ . This is done by multiplying the
expected Brier score values by the ratio of the second derivatives of the two expected scores calculated at p = p∗ . The
proper score scale invariance ensures that the ranking obtained using the original and normalized version of the Brier
score is unchanged.
Both expected score differences are uniquely maximized at p = p∗ which means that the forecast matching the true
probability has the highest expected score. This is an easy way to assess if a scoring rule for binary outcomes is proper
or not. Furthermore, Figure 1 offers an example of how different scores penalize differently the same forecasts. The log
score is asymmetric and takes into account the relative differences between the forecasts (equation 8), and if p∗ 6= 0 the
expected score for p = {0, 1} is −∞. This means that the log score heavily penalizes overconfidence in the forecast.
The Brier score, instead, considers the absolute difference between forecasts (equation 7) resulting in a symmetric
distribution. For example, using the Brier score, a forecast p = 0 will be preferred to any forecast in (2p∗ , 1), for any
p∗ < 1/2.
The choice of score, and consequently the style of penalty, should reflect the task at hand. Predicting p = 0, 1 means
that we are absolutely certain about the outcome of X. If the forecasts under evaluation are planned to be used in
an alarm based system, for which an alarm is broadcasted if the probability is above or below a certain threshold,
being overconfident may put lives at risk and perhaps the log score would be the right choice in this situation. On the
other hand, the Brier score may be suitable when being overconfident is less dangerous. This example illustrates the
flexibility of proper scores and how important it is to choose the right one depending on the purposes of the forecast
under evaluation.
5
Ranking earthquake forecasts using proper scoring rules A P REPRINT
0.000 0
Score Score
Brier Brier
−0.001
Log Log
−1
−0.002
E[∆]
E[∆]
−0.003
−2
−0.005 p∗ −3
Figure 1: Differences in the scores expected value for a generic value of the forecast p and the optimal forecast p = p∗ ,
namely E[∆] = S E (p|p∗ ) − S E (p∗ |p∗ ), in the case p∗ = 0.001. Panel (a) p ∈ (0, 0.006], Panel (b) p ∈ (0, 1). The
expected Brier score have been normalised to match the curvature of the log score when p = p∗ .
3 Improper scores
Scores which are not proper are called improper. Being improper means that a model may exist with expected score
greater than the data generating model. In the specific case of probabilistic forecasts for binary events, a score is
improper if it is biased towards models which systematically under/overestimate the true probability p∗ . In the context
of earthquake forecasting experiments we do not know the true value of p∗ . Therefore, it is crucial to use proper scoring
rules for which we are sure that, at least on average, they will prefer the closest model to the data generating one.
Improper scoring rules do not have this property, which implies that the smallest or the largest (or any other) forecast,
on average, could achieve the highest score. This is in clear contrast with the aim of any forecasting experiment. Below,
we demonstrate that the parimutuel gambling score [Zhuang, 2010, Zechar and Zhuang, 2014] is an example of a
scoring rule which is proper only in a specific situation and not in general.
The parimutuel gambling score was designed to rank forecasting models for binary events and was applied to rank
earthquake forecasting models in CSEP experiments [Taroni et al., 2018, Zechar and Zhuang, 2010]. Initially, it was
used to compare models against a reference model [Zhuang, 2010], which is improper. Later, it was generalized to
compare models against each other simultaneously [Zechar and Zhuang, 2014], the case with only two players is the
special case for which the score is proper, all the others are not. The score is based on a gambling scheme in which the
forecasting models play the role of the gamblers and, for each observation, they obtain a reward proportional to the
probability assigned by the gambler to the event occurring. In particular, it is a zero-sum game, in the sense that bids
6
Ranking earthquake forecasts using proper scoring rules A P REPRINT
and rewards in each bin sum to zero, which makes the parimutuel gambling score relative to one forecast dependent on
the other forecasts.
In contract to the Brier and log scores, it is not possible to define the parimutuel gambling score using the form
S(p|x) because it needs at least two forecasts to be evaluated and is a function of them all. Given a set of k forecasts
p = (p1 , ..., pk ), we define SG (p|x) as the vector such that the i-th component, SG,i (p|x), is given by the parimutuel
gambling score of the i-th forecast, given x has been observed. In the case of the Brier and log score the components
of the vector S(p|x) are defined independently, in the case of the parimutuel gambling score they have to be defined
Pk
jointly. Let p̄ be the average probability involved in the gambling scheme, namely p̄ = i=1 pi /k. The parimutuel
gambling score relative to the i-th forecast is defined as
(
pi
p − 1, x = 1,
SG,i (p|x) = 1−p
1−p − 1,
i
x = 0.
P
The above expression is a zero-sum game, meaning that i SG,i (p|x) = 0, therefore the rewards may be positive or
negative. Each gambler obtains a positive reward if and only if they assign a greater probability to the observed event
than the average gambler involved in the game. Vice versa, the reward is negative if the probability is smaller.
The expected value with respect the true probability p∗ is given by
pi 1 − pi
E
SG,i (p|p∗ ) = p∗ − 1 + (1 − p∗ ) −1 ,
p̄ 1 − p̄
p∗ pi (1 − p∗ )(1 − pi )
= + − 1,
p̄ 1 − p̄
(pi − p̄) (p∗ − p̄)
= . (12)
p̄(1 − p̄)
The denominator involves all the probabilities in the game which demonstrates the interdependence with all other
forecasts and complicates the study of the derivatives. However, it is still possible to prove that the gambling score is
strictly proper when k = 2. In this case, p = (p1 , p2 ), and
p1 + p2 p1 + p2
E
4p̄(1 − p̄)SG,1 (p|p∗ ) = 4 p1 − p∗ − ,
2 2
= (p1 − p2 ) (2p∗ − p1 − p2 ) ,
= − [(p1 − p∗ ) − (p2 − p∗ )] [(p1 − p∗ ) + (p2 − p∗ )] ,
= (p2 − p∗ )2 − (p1 − p∗ )2 .
The expected reward of the first modeler is non-negative when |p1 − p∗ | ≤ |p2 − p∗ |, implying that p1 is favoured over
p2 if it is closer to the true probability p∗ . In fact, if p2 = p∗ then, SG,1
E
(p|p∗ ) ≤ 0, with the equality verified only for
∗
p1 = p . Furthermore, the expected gambling score in this case is proportional to the expectation of the corresponding
Brier score differences ∆B = SB E
(p1 |p∗ ) − SBE
(p2 |p∗ ), thus, they produce, on average, the same rankings.
When comparing forecasting models, ensuring that the score is proper may not be sufficient. It also has to be used
properly. The gambling score with k = 2 offers a nice example of this situation. We have demonstrated that the
parimutuel gambling score is proper when k = 2, however, the dependence of the score value on all the forecasts
E
involved in the comparison is a source of bias. In fact, SG,1 (p|p∗ ) ≥ SG,2
E
(p|p∗ ) when p1 is closer to p∗ than p2 ,
∗ E ∗
however, p1 = p does not maximize SG,i (p|p ) as shown in Figure 2. This means that the score becomes biased when
we rank forecasts based on the score difference against a reference model.
Formally, we are considering pairwise vectors p1 = (p1 , p0 ), p2 = (p2 , p0 ), etc., where p0 is the reference model.
For each of these we can estimate pairwise comparison score vectors SG (p1 |x), SG (p2 |x), and so on. The first
component of each vector, namely SG,1 (p1 |x), SG,1 (p2 |x), etc, represents the score of p1 and, respectively, p2 against
the reference model p0 . At this point, one would be tempted to rank the models based on SG,1 (p1 |x) and SG,1 (p2 |x),
and this is the approach taken in Taroni et al. [2014] in which the official national time-independent model [Gruppo di
Lavoro, 2004] is used as reference model.
If the parimutuel gambling score is used to rank forecasts based on the score difference relative to a reference model,
it will not reliably favour the model closest to the true one, and the size of the bias will depend on the choice of the
7
Ranking earthquake forecasts using proper scoring rules A P REPRINT
1.0
p2 1.0
p2
0.001 0.001
0.002 3e−04
0.004 5e−04
0.5 0.5
SEG
SEG
0.0 0.0
(a) (b)
−0.5 −0.5
−1.0 −1.0
0.000 0.001 0.002 0.003 0.004 0.000 0.001 0.002 0.003 0.004
p1 p1
The expected scores have been normalized so that their minimum is equal to -1.
reference model. For example, in Figure 2a, for p2 = 0.004, the gambling score is maximized at p1 = 0. This means
that if the reference model is p0 = 0.004, the overconfident forecast p1 = 0 would be favoured by the ranking even if
another forecast is perfect, e.g. p3 = p∗ . This problem can be particularly relevant in operational seismology where it is
common for candidate forecasts to be compared against a reference model which is known to be based on simplistic
assumptions (for example a homogeneous Poisson process).
Hereon, the term pairwise gambling score refers to the comparison against a reference model as described in this section,
while the term full gambling score will refer to the case where the forecasts compete directly against each other as we
describe in the next section. Using this terminology, the full gambling score with k = 2 is the only proper score.
The generalized version of the full parimutuel gambling score, as presented in Zechar and Zhuang [2014], for k ≥ 3 is
improper. For example, when k = 3 and p2 = p∗ , following equation 12 the difference between the expected score for
p1 and p2 is given by
E
3p̄(1 − p̄)[SG,1 (p1 , p∗ , p3 |p∗ ) − SG,2
E
(p1 , p∗ , p3 |p∗ )] = 3(p1 − p∗ )(p∗ − p̄)
= (p1 − p∗ )(2p∗ − p1 − p3 ),
with both sides scaled by the common factor 3p̄(1 − p̄). This means that when 2p∗ − p3 ≥ p1 , the forecast p1 will have
a positive score for any p1 ≥ p∗ . Any value of p1 ∈ [p∗ , 2p∗ − p3 ] will be preferred to p2 that is equal to p∗ . When
2p∗ − p3 ≤ p1 , with the same reasoning, p1 is preferred over p2 = p∗ in the interval [2p∗ − p3 , p∗ ].
8
Ranking earthquake forecasts using proper scoring rules A P REPRINT
0.25 0.25
p3 p3
0.001 0.001
0.00 0.00
3e−04 0.002
5e−04 0.003
−0.25 −0.25
E[∆]
E[∆]
−0.50 −0.50
−1.00 −1.00
0.000 0.001 0.002 0.003 0.004 0.000 0.001 0.002 0.003 0.004
p1 p1
E
Figure 3: Expected gambling score differences (k = 3) between p1 and p2 , SG,1 (p|p∗ ) − SG,2
E
(p|p∗ ), as a function of
p1 ∈ (0, 0.004), p2 = p = 0.001 (vertical line), and p3 ∈ {p , p /2, p /3} (a) and p3 ∈ {p , 2p∗ , 3p∗ } (b) .
∗ ∗ ∗ ∗ ∗
In Figure 3 we consider k = 3, p∗ = p2 = 0.001 and report the difference between the expected scores of p1 and p2 ,
E
namely SG,1 (p|p∗ ) − SG,2
E
(p|p∗ ), for different values of p3 . The expected score difference is not maximize at p1 = p∗ ,
which means that the score is biased, and the "direction" of the bias depends on p3 being greater than or equal to p∗ .
Consider k > 3 gamblers who propose probabilities p = {p1 , ..., pk }. It is helpful to consider the vector of probabilities
that excludes the first component; we name this p−1 = p/{p1 } and its mean p̄−1 . Assuming p2 = p∗ , thus
p = (p1 , p∗ , ..., pk ), we have that
E
k p̄(1 − p̄)[SG,1 (p|p∗ ) − SG,2
E
(p|p∗ )] = k(p1 − p∗ )(p∗ − p̄),
= (p1 − p∗ )[kp∗ − p1 − (k − 1)p̄−1 ],
from which we conclude that the expected score difference is positive when p1 ∈ [kp∗ − (k − 1)p̄−1 , p∗ ] or p1 ∈
[p∗ , kp∗ − (k − 1)p̄−1 ], depending on if p∗ ≶ p̄−1 . Specifically, when p∗ > p̄−1 (Figure 3a) the first gambler is
encouraged to bet on a p1 > p∗ and vice versa when p∗ < p̄−1 (Figure 3b). Furthermore, p∗ is always an extreme
of the interval where the expected score difference is positive. Considering p̄−1 as fixed, the length of the interval
is an increasing function of the number of gamblers k (Figure 4) which means that the size of the set of forecasts
capable of obtaining a score value higher than the data generating model is an increasing function of the number of
forecasts involved in the comparison. It is clear that the multi-forecast parimutuel gambling score favours models that
are contrary to the average of the other forecasts. This could be particularly dangerous when evaluating the performance
of earthquake forecasting models. For example, the trigger for an alarm being broadcast (or not) is often defined when
the probability of having an earthquake above a certain magnitude exceeds a specified threshold. Using a model chosen
looking at the full parimutuel gambling score could therefore lead to broadcasting alarms when they are not needed
(p p∗ , ’crying wolf’, Figure 4a) or not broadcasting an alarm when needed (p p∗ , providing ’false reassurance’,
Figure 4b).
9
Ranking earthquake forecasts using proper scoring rules A P REPRINT
k k
3 3
0.001 0.001
5 5
10 10
20 20
0.000 0.000
E[∆]
E[∆]
−0.001 (a) −0.001 (b)
−0.002 −0.002
0.0000 0.0025 0.0050 0.0075 0.0100 0.000 0.001 0.002 0.003 0.004 0.005
p1 p1
Figure 4: Expected gambling score differences (k ∈ {3, 5, 10, 20}) between p1 and p2 as a function of p1 considering
p2 = p∗ = 0.001 (vertical line). The averageof the forecast probabilities (excluding the first forecast) is the constant
p̄−1 = p∗ /2 (a) and p̄−1 = 2p∗ (b).
The root of the problems with this score is that the score, relative to a candidate forecast, explicitly depends on the
other forecasts. This design brings two problems: (i) the score, even in the special case when it is proper, can be
used improperly and (ii) the score is never proper when considering more than two models. The Brier and log scores
do not suffer from the same problem since the score of a forecast depends only on the forecast and the observation.
Furthermore, the improperness demonstrated here can be expressed in terms that show that the gambling metaphor is
part of the problem: If the outcome x = 0 is likely (e.g. p∗ = 0.001) and the majority of the forecasts have too large
probabilities, then the expected gain is higher for an overconfident forecast, p p∗ , since that will give the forecaster a
larger share of the total payout.
Until now, we have analysed the expected score for a single bin, here we analyse the ability to express a preference
between two forecasts using the average score across multiple bins. We assume to have access only to one observation
xi ∈ {0, 1} per bin . We analyse extensively the case in which the probability of observing xi = 1, is the same for each
bin, p∗i = p∗ for any i = 1, ..., N . We refer to this as the Multiple Bins Single Probability case; the only quantity of
interest is p∗ and a forecast is represented by a single value p. Even though, the Multiple Bins Single Probability case
is clearly unrealistic in practice, it builds the basic concepts we will then use to explore the Multiple Bins Multiple
Probabilities case where the probability of observing xi = 1 is potentially different for each bin.
Considering multiple bins, we observe a realization of the random variable Xi ∼ Ber(p∗i ) for i = 1, ..., N . A forecast
is given by the vector p = (p1 , ..., pN ) specifying the probability of Xi = 1 for each bin. Following the terminology
10
Ranking earthquake forecasts using proper scoring rules A P REPRINT
in the literature
P regarding Bernoulli random variables, the event Xi = 1 is referred to as a success. The quantity
XS = i Xi is therefore referred as the sum of the observations or the number of successes or the number of active
bins.
Given an arbitrary scoring rule S(p|X), the average score associated with the forecast p is given by:
N
1 X
S(p|X) = S(pi |Xi ).
N i=1
The quantity S(p|X) is a random variable itself, because it is a function of random variables X. To compare two
forecasts p1 and p2 , we study their score difference:
N N
!
1 X X
∆(p1 , p2 , X) = S(p1i |Xi ) − S(p2i |Xi ) ,
N i=1 i=1
N
1 X
= ∆(p1i , p2i , Xi ).
N i=1
The quantity ∆(p1 , p2 , X) is also a random variable as it too depends on the vector of random variables X. If S(p|X)
is a proper scoring rule, and if the expected value of the score difference is positive, namely E[∆(p1 , p2 , X)] > 0, the
forecast p1 is "closer" to the true p∗ than the alternative forecast p2 . The expected value should be considered with
respect the distribution of the observations X. However, we do not observe the full distribution - we only observe a
sample (i.e. we observe the quantity ∆(p1 , p2 , x) which is a realization of the random variable ∆(p1 , p2 , X)). Even if
the expected score difference E[∆(p1 , p2 , X)] is positive, which means that we should express a preference for the first
forecast, the observed score difference ∆(p1 , p2 , x) may be negative and lead to the opposite conclusion. To avoid this
problem we need to account for the uncertainty around the observed ∆(p1 , p2 , x) which is the point estimate of the
expected score difference E[∆(p1 , p2 , X)].
For the Multiple Bins Single Probability case, the observation in each bin is a binary random variable Xi ∼ Ber(p∗ ),
i = 1, ..., N . Given an arbitrary scoring rule S(p|X) and two candidate forecasts p1 and p2 , the score difference for the
i-th bin is a discrete random variable with distribution:
with probability 1 − p∗ ,
∆0 = S(p1 |0) − S(p2 |0)
∆(p1 , p2 , Xi ) =
∆1 = S(p1 |1) − S(p2 |1) with probability p∗ .
The forecasts are ranked based on the average score difference across all bins:
N
1 X
∆(p1 , p2 , X) = ∆(p1 , p2 , Xi ),
N i=1
N
1 X
= (∆0 + Xi (∆1 − ∆0 )) ,
N i=1
XS
= ∆0 + (∆1 − ∆0 ),
N
P
where, XS = i Xi is the sum of all observations or, equivalently, the total number of successes. By definition, XS is
the sum of N (assumed to be) independent and identically distributed Bernoulli trials Xi . Therefore, XS has a Binomial
distribution with size parameter N , the number of bins, and probability parameter p∗ . When we observe a sample
x1 , ..., xN , the observed score difference is given by:
xS
∆(p1 , p2 , x) = ∆0 + (∆1 − ∆0 ),
N
where xS is a realization of the random variable XS . The observed score difference depends on the observations only
through the quantity xS /N . Thus, it is enough to study the quantity xS /N to make inference about the expected value
of the score difference. The quantity xS /N it is said to be sufficient [Fisher, 1922] with respect to the expected score
difference because it contains all the information provided by the observations x1 , ..., xN on the parameter of interest
(in this case, the expected score difference E[∆(p1 , p2 , X)]). For an introduction to statistical inference and the theory
behind we refer to [Schervish, 2012, Hastie et al., 2009].
11
Ranking earthquake forecasts using proper scoring rules A P REPRINT
A way to account for the uncertainty around the observed score difference is to consider an interval estimate of the
expected value of the score difference. Once a sample x = x1 , ..., xN has been observed and the confidence interval
calculated, if the entire interval lies above zero we express a preference towards p1 , alternatively if it lies below zero we
express a preference towards p2 . If the interval contains the value zero we conclude that the observed sample does not
contain enough information to express a preference. It is important to consider the latter case as a possible outcome
because it is an indication that we need to collect more data or that the forecasts perform similarly (as measured by the
score) and provide an additional information than the pure rankings.
We are considering the confidence interval for the expected value of the score difference:
E[XS ]
E[∆(p1 , p2 , X)] = ∆0 + (∆1 − ∆0 ),
N
= ∆0 + p∗ (∆1 − ∆0 ). (13)
Having an observation x = x1 , ..., xN per bin, the point estimate of E[∆(p1 , p2 , X)] is the observed score difference:
∆(p1 , p2 , x) = ∆0 + p̂(∆1 − ∆0 ), (14)
where p̂ = xS /N is the observed probability of success. Comparing the equations 13 and 14, the point estimate of the
score difference is retrieved plugging in the point estimate of the probability of success p̂ in place of p∗ . In the same
way, to retrieve an interval estimate of the expected score difference is sufficient to retrieve an interval estimate of the
probability of success p∗ .
Therefore, we need the confidence interval of level α for the true probability p∗ given observations x1 , ..., xN from
a Ber(p∗ ), namely CIp∗ (α) = (p̂L , p̂U ), and plug those values into expression 14 to obtain a confidence interval for
E[∆(p1 , p2 , X)], namely CI∆ (α) = (∆L , ∆U ). Various methods have been found to estimate p̂L and p̂U , most of
them relying on a Gaussian approximation. However, this approximation is not reliable for small sample sizes (number
of bins N ) and for values of p∗ close to zero or one, as in our case [Wallis, 2013].
Hereafter, we use the Clopper-Pearson confidence interval [Clopper and Pearson, 1934]. This method is referred to as
exact because it relies on cumulative binomial probabilities rather than an approximation and is therefore more efficient
and accurate than simulation based methods. The confidence interval with level α for p̂ is given by:
α
pL (α) = BetaQ( ; xS , N − xS + 1),
2
α
pU (α) = BetaQ(1 − ; xS + 1, N − xS ),
2
where the function BetaQ(q; a, b) is the q-th quantile of a Beta distribution with parameters a and b. We can construct
confidence intervals for E[∆(p1 , p2 , X)] as follows:
∆L = ∆0 + p̂L (∆1 − ∆0 ),
∆U = ∆0 + p̂U (∆1 − ∆0 ).
The obtained confidence interval for p∗ depends on the data only through the sum of the observations xS , which is a
sufficient statistic for the problem. Similarly, the confidence interval for E[∆(p1 , p2 , X)] depends on the data through
the value of the sufficient statistic, xS .
Figure 5 shows the confidence interval for the score difference as a function of the sum of observations xS considering
two competing forecasts p1 = 0.001, p2 = p1 /3, a reference model for the pairwise gambling score p0 = 5p1 and
N = 10, 000 bins. Here, we do not need to choose a value for p∗ . Indeed, the confidence interval is determined solely
by the forecast and observation. The Brier, log and full gambling score(Figure 5a, 5c, 5d) all express a preference for p1
if we observe xS > 12, while they express a preference for p2 when xS < 2. This result is expected because p1 > p2 ,
which means that xS > 12 is much more probable under p1 than p2 . In fact, the average number of successes using p1
is N p1 = 10 while N p2 = 3.34. The same reasoning applies when we express a preference for p2 (xS < 2).
The pairwise gambling score (Figure 5 (b)), instead, requires xS > 24 to express a preference for p1 and xS < 9 to
express a preference for p2 . It is heavily biased toward the forecast closer to zero. In fact, when p1 is the true probability,
the probability of observing xS > 24 is less than 0.0001 and the probability of observing xS < 9 is 0.33. Therefore,
we are more likely to express a preference for p2 than for p1 , even when p1 = p∗ . This reinforces the problems with
employ improper scores introduced in Section 3.
12
Ranking earthquake forecasts using proper scoring rules A P REPRINT
p2 p1
(a) p2 p1
4 no pref 4 no pref
E[∆]
E[∆]
2 2
(b)
0 0
0 10 20 0 10 20
Xs Xs
(c) p2 p1
(d)
4 p2 no pref p1 4 no pref
E[∆]
E[∆]
2 2
0 0
0 10 20 0 10 20
Xs XS
Figure 5: Confidence interval (shaded area) and point estimate (black solid line) for E[∆] as a function of the number
of observed successes xS considering p1 = 0.001, p2 = p1 /3, p0 = 5p1 and N = 10000. In each plot shows a different
score: (a) Brier score; (b) pairwise gambling score; (c) logarithmic score; (d) full gambling score. Black solid line
represents the observed score difference while the orange area represents the confidence interval. The black vertical
dashed lines represent the interval of values of xS for which we do not express a preference
The confidence interval for the expected score difference, CI∆ (α), is a function
Pof the competing forecasts, the scoring
rule and depends on the data only through the sum of the observations xS = i xi . In particular, there are a range of
values (between the dashed lines in Figure 5) of xS for which we are not able to express a preference. We refer to this
interval as (xmin , xmax ). With respect to the sum of the observations xS there are only three possible outcomes:
xS < xmin −→ preference for p2 ,
xmax ≤ xS ≤ xmax −→ no preference,
xS > xmax −→ preference for p1 .
The values xmin and xmax are determined solely by p1 , p2 , the number of bins N and the scoring rule. Table 1 reports
the values of xmin and xmax for the scoring rules depicted in Figure 5. These values can be used to compute the
preference probabilities once a value for p∗ is assumed. Indeed, in the Multiple Bins Single Probability case, the
distribution of XS is a Binomial distribution, XS ∼ Bin(N, p∗ ). Table 2 reports the probabilities of i) no preference; ii)
Preference for p1 ; iii) Preference for p2 . The probabilities are calculated considering alternatively p∗ equal to p1 (first
half of the table) or p2 (second half of the table).
The similarity among the values xmin and xmax for the proper scores lead to similar preference probabilities. The proper
scores always assign the greatest probability to the case in which we are not able to express a preference, however, when
13
Ranking earthquake forecasts using proper scoring rules A P REPRINT
Table 1: Multiple Bins Single Probability case: table reporting the values xmin and xmax for the Brier, log, pairwise
gambling (PG) and full gambling (FG) score. The reported values refers to the case where N = 10, 000, p1 = 0.001,
p2 = p1 /3 , and do not depend on p∗ .
Table 2: Multiple Bins Single Probability case: table reporting for each score (row) the probabilities of expressing (or
not) a preference using either the Brier, log-, pairwise gambling (PG) or full gambling (FG) score. The probabilities are
calculated considering N = 10, 000, p1 = 0.001, p2 = p1 /3 and considering two cases: p∗ = p1 and p∗ = p2 .
Score No pref p1 p2
∗
p = p1
Brier 0.7912 0.2083 0.0005
Log 0.6963 0.3032 0.0005
PG 0.6672 0.0000 0.3327
FG 0.7912 0.2083 0.0005
p∗ = p2
Brier 0.8454 0.0000 0.1545
Log 0.8453 0.0000 0.1545
PG 0.0073 0.2083 0.9927
FG 0.8454 0.0000 0.1545
p∗ = p1 it is unlikely to express a preference for p2 . Vice versa when p∗ = p2 . There is a slightly difference between
the Brier and the log score coming from the different penalty applied to forecasts close to zero. The log score penalises
more heavily forecasts close to zero and, in fact, when p1 = p∗ chances to express a preference for p1 are higher than
using the Brier score. The full gambling score for p1 against p2 is proportional to the Brier score difference between p1
and p2 (see Section 3.2) and thus, their preference probabilities coincide. Considering the pairwise gambling score the
probability of expressing a preference for p1 is always very close to zero, even when p∗ = p1 . This shows again that it
is possible to find a combination of p1 , p2 and p0 such that the model providing the smallest forecast obtains the highest
reward with probability over 0.9, even when the other forecast is equal to the true p∗ .
Figure 6 shows the preference probabilities as a function of p∗ ∈ (10−6 , 10−2 ) which is the range of values of the
5-year adaptively-smoothed forecast for Italy (aggregating over the magnitude bins) used later to illustrate the Multiple
Bins Multiple Probabilities case. The Brier score behaves as expected. The probability of expressing a preference for p2
increases as p∗ goes to zero, which is what we expect given p2 < p1 . On the other hand, the probability of preferring p1
increases when p∗ increases, because p1 > p2 . Finally, the probability of not being able to express a preference is higher
when p2 < p∗ < p1 (Figure 6a). The pairwise gambling score, instead, does not behave as expected. The probability of
preferring p1 is almost zero in the range of values of p∗ considered in the example. The two most probable outcomes
are: expressing a preference for p2 or not expressing a preference at all.
It is interesting to study how the probability for each case changes as a function of p∗ for different numbers of bins N
and different ratios between p1 and p2 . To do that it is useful to focus only on two possible outcomes: expressing a
preference and not expressing a preference. The probability of expressing a preference is given by the probability of
observing a sample such that the sum of the observations xS is greater than xmax or smaller than xmin . We refer to this
probability as β which is given by
β = 1 − Pr[xmin ≤ XS ≤ xmax ].
This probability depends on the scoring rule, the forecasts p1 and p2 ,the number of bins N and the true probability p∗ .
We study β as a function of p∗ for different numbers of bins. In this artificial case, to increase the number of bins we
14
Ranking earthquake forecasts using proper scoring rules A P REPRINT
1.00
0.75
Preferences
no pref
p2 p1 (a)
Pr
0.50
p1
0.25
p2
0.00
1.00
0.75
Preferences
no pref
p2 p1 (b)
Pr
0.50
p1
0.25
p2
0.00
Figure 6: Multiple Bins Single Probability case: each plot shows the probability of each possible outcome (solid lines
no preference, dotted lines preference for p1 , dashed lines preference for p2 ) as a function of p∗ using the Brier score (a)
and the pairwise gambling score (b) considering p1 = 0.001 and p2 = p1 /3 (vertical lines), p0 = 5p1 and N = 10, 000.
The true probability p∗ varies in (10−6 , 2 · 10−2 ) which is a realistic range of values in Italy.
are considering additional bins with the same probability, we are explicitly not splitting any bin; this is analogous to
increase the data at hand applying the model to a larger spatio-temporal region.
Figure 7a considers only the Brier score. The region of p∗ presenting low values for β shrinks when the number of bins
increase which simply means that the more data we have, the more chances of expressing a preference. Moreover, β is
at the minimum when p∗ ∈ (p2 , p1 ), which is reasonable because if the distances |p∗ − p1 | and |p∗ − p2 | are similar
the probability of no preference should be high. The N = 2000 can be explained considering p1 = p∗ . In this case,
the expected sum of observations is N p1 = 2 and it is more probable to observe XS < 2 than XS > 2. Given that
N p2 < N p1 , the probability of not expressing a preference is high.
Figure 7 presents the probability β as function of p∗ for different scores with a fixed number of bins N = 5000 (b) and
N = 20000 (c). For N = 5000, the proper scores (Brier, log and full gambling score) present the same values of β for
any value of p∗ . For N = 20000, the proper scores start to behave differently. The Brier and full gambling score still
coincide, while the log score is slightly different. Specifically, the log score presents higher β values when p∗ = p1 , and
lower when p∗ = p2 . This depends on the different penalties applied to forecasts close to zero. The log score presents
greater chances of expressing a preference for p1 when p∗ = p1 because the other forecast p2 is smaller than p1 and,
therefore, penalized. On the other hand, when p∗ = p2 the log score presents smaller β values than the Brier score.
In contrast to the proper scores, the pairwise gambling score reaches its minimum β value for p∗ > p1 . Here, the
pairwise gambling score tends to express a preference for the smaller forecast even when the other one is closer to
p∗ . This leads to higher values of β when p∗ ∈ (p2 , p1 ) because the pairwise gambling score will likely express a
preference for p2 . Only when p∗ > p1 the probability of no preference grows and the value of β decreases accordingly.
15
Ranking earthquake forecasts using proper scoring rules A P REPRINT
1.00
p2 p1 N
0.75 2000
5000
10000
β
0.50
20000
0.25
(a)
0.00
0.0000 0.0005 0.0010 0.0015 0.0020
∗
p
0.50 β 0.50
0.25 0.25
(b) (c)
0.00 0.00
0.0000 0.0005 0.0010 0.0015 0.0020 0.0000 0.0005 0.0010 0.0015 0.0020
∗ ∗
p p
Figure 7: Multiple Bins Single Probability case: (a) Brier score preference probability as a function of p∗ for different
numbers of bins N ∈ {2000, 5000, 10000, 20000}. (b-c) Probability of expressing a preference as a function of p∗ .
Colors represent the different scores: Brier, log , pairwise gambling (PG), and full gambling (FG) score. The Brier and
FG scores coincide. The number of bins is fixed to N = 5000 (b) and N = 20000 (c). We set p1 = 0.001, p2 = p1 /3
(vertical lines), p0 = 5p1 , and p∗ ∈ (10−6 , 2−3 ) which is a realistic range of values in Italy.
Given that p1 and p2 are scalars, we can consider β as a function of the ratio between p1 and p2 , ω = p2 /p1 , for a
fixed p∗ . In principle, we expect that β is an increasing function of ω. We assume that the first forecast and the true
probability are identical p1 = p∗ = 0.001. The reference model for the pairwise gambling score is p0 = 5p1 and
we consider different numbers of bins N ∈ {2000, 5000, 10000, 20000}. The ratio ω = p2 /p1 varies in the interval
(0.1, 4). We expect low β values when ω is around one (similar forecast) and high β values otherwise.
Figure 8a shows that, as expected, for N > 2000, β has its minimum when ω = 1. Considering ω as fixed, β is an
increasing function of the number of bins. Figure 8b-c compares the β values relative to different scores for a fixed
number of bins, N = 5000 (b) and N = 20000 (c). The Brier and full gambling score coincide, whilst the log score
presents slightly different β values. As before, this is due to the different penalties applied to forecasts close to zero.
The pairwise gambling score is not consistent with the trends in the proper scores. Using this score and considering
N = 20000 (Figure 8c), the probability β is consistently greater than 0.5 for the considered values of ω. Considering
that p1 = p∗ , the quantity ω is also the ratio between p2 and p∗ . This implies that regardless of ω, we will erroneously
express a preference for p2 with a probability above 0.5.
Importantly, these sanity checks of a proposed scoring procedure can be done before looking at the observations. It is
possible to check if forecasts can, in principle, be distinguished in light of the amount of expected data. We recommend
the use of such exploitative figures when introducing a new scoring rule whose performance have not been tested. If the
proposed scoring rule does not behave acceptably in this simple scenario, it is unlikely that it would behave acceptably
in a real application.
16
Ranking earthquake forecasts using proper scoring rules A P REPRINT
1.00
N
0.75
(a) 2000
5000
10000
β
0.50 20000
0.25
0.00
0 1 2 3 4
ω
0.50 β 0.50
0.25 0.25
(b) (c)
0.00 0.00
0 1 2 3 4 0 1 2 3 4
ω ω
Figure 8: Multiple Bins Single Probability case: (a) Probability of expressing a preference using the Brier score
as a function of ω = p2 /p1 ∈ (0.1, 4) for different numbers of bins N ∈ {2000, 5000, 10000, 20000}. We set
p1 = p∗ = 0.001, and the reference model is p0 = 5p1 . (Bottom) Probability of expressing a preference as a function
of ω. Colors represent the different scores: Brier, log, pairwise gambling (PG), and full gambling (FG) score. The
number of bins is fixed to N = 5000 (b) and N = 20000 (c).
The Multiple Bins Multiple Probabilities case generalizes the Multiple Bins Single Probability case, and is much more
similar to a real earthquake forecasting experiment. For example, the forecasts involved in the first CSEP experiments
[Field, 2007, Schorlemmer and Gerstenberger, 2007b, Zechar et al., 2013, Michael and Werner, 2018] were mostly
grid-based forecasts providing for each space-time-magnitude bin, the expected number of earthquakes. Then, the
number of events in each bin is modelled using a Poisson distribution with intensity equal to the number of events
provided by the forecasts and the probability of observing at least one event is calculated accordingly. In this scenario,
we do not have analytical results for the score difference distribution and we need to recur to simulations.
We now want to specify a true model which has more realistic probabilities. Since we do not actually know these in
reality, we choose to work with one of the CSEP models that was submitted to the 2010 Italy experiment [Taroni et al.,
2018]. We choose to simulate synthetic data from the 5-year adaptively-smoothed forecast for Italy [Werner et al.,
2010] and explore the ability of the scoring rules to discriminate between linearly scaled versions of this true model.
This means that we are considering only one time bin of size 5 years, while the space-magnitude domain is divided in
multiple regular bins. The spatial domain is represented by the coloured area in Figure 9 and it is divided in 0.1 × 0.1
longitude-latitude bins. The magnitude domain ranges from 4.95 to 9.05 magnitude units and is divided in bins of
length 0.1. The forecast is relative to the period from January 1, 2010, to December 31, 2014.
The adaptively-smoothed forecast provides the expected number of earthquakes in each space-magnitude bin. For each
bin, to calculate the probability of observing at least one earthquake, in accordance with the methodology in the 2010
17
Ranking earthquake forecasts using proper scoring rules A P REPRINT
48°N
46°N ln(p)
44°N
−4
latitude
42°N
−6
40°N
−8
38°N
−10
36°N
Figure 9: 5-year adaptively-smoothed forecast for Italy [Werner et al., 2010]. The figure shows for each spatial bin the
natural logarithm of the probability of observing at least one earthquake at or above magnitude 4.95 in the period from
January 1, 2010, to December 31, 2014.
Italy CSEP forecast experiment, we consider a Poisson distribution for the number of events with intensity given by
the predicted number of events. Assuming independence in the magnitude bins, we can aggregate the probabilities
over magnitude bins and, for each space bin, obtain the probability of observing at least an earthquake in the period of
interest with magnitude greater, or equal, to 4.95. Figure 9 shows the forecasted log-probability for each spatial bin
used as data generating model.
The Italian adaptively-smoothed forecast reported in Figure 9 is the vector of true probabilities p∗ = p∗1 , ..., p∗N , where
N = 8993. As in the previous sections, we compare two forecasts p1 = p∗ and p2 = ωp∗ . We will be ignoring the
18
Ranking earthquake forecasts using proper scoring rules A P REPRINT
Table 3: Expected score difference considering p∗ equal to the 5 year Italy adaptively-smoothed forecast, p1 = p∗ ,
p2 = ωp∗ and reference model for the pairwise gambling score p0 = 5p∗ . The scores considered are: the Brier score,
the log score, the pairwise gambling (PG) score and the full gambling (FG) score.
Score E[∆]
Brier 0.0000026
Log 0.0003137
PG -0.0000900
FG 0.0002422
where, ∆0,i = ∆(p1i , p2i , 0) and ∆1,i = ∆(p1i , p2i , 1) are, respectively, the score difference in the i-th bin in case we
observe Xi = 0 (no earthquake at or above magnitude 4.95 during the 5 years) or Xi = 1 (at least one earthquake
above magnitude 4.95 during the 5 years). The quantity ∆ ¯ 0 is the average ∆0,i . The observations Xi ∼ Ber(p∗ ) follow
i
a Bernoulli distribution, each bin has a potentially different parameter p∗i 6= p∗j for any i 6= j. The expected value of the
score difference is given by
N
¯0 + 1
X
E[∆(p1 , p2 , X)] = ∆ p∗ (∆1,i − ∆0,i ).
N i=1 i
Given that we are considering p1 = p∗ and p2 = ωp∗ , the expected score difference is non-negative if a proper scoring
rule is used while it could be negative if the scoring rule is improper. Specifically, we show that it is possible to find
a reference model p0 such that, if used in combination with the parimutuel gambling score to rank the forecasts, the
expected score difference is negative. As before, the Brier, log and full gambling score are used for comparison. In
Table 3 we report the expected score differences considering different scores. As expected, they are all positive except
for the pairwise gambling score.
In Figure 10 is showed the expected score difference as a function of the forecasts ratio ω ∈ [10−3 , 4]. The results are
similar to the ones reported in Figure 1 and 2. The Brier, log and full gambling scores behave suitably, while the pairwise
gambling score does not. The Brier and full gambling score are bounded and they prefer a forecast p = 10−3 p∗ to
p0 = 4p∗ . Indeed, in Figure 10 the left hand side is greater than the right hand side. That is because the penalty is
based on the absolute difference between a forecast and the data generating model, therefore, a forecast p = 10−3 p∗
is preferred to p0 = 4p∗ , because k10−3 p∗ − p∗ k ≤ k4p∗ − p∗ k. On the other hand, the log score is unbounded
and is based on the relative difference. With the log score, a forecast p0 = 4p∗ is preferred to p = 10−3 p∗ because
kp∗ /10−3 p∗ k > kp∗ /4p∗ k. The pairwise gambling score, instead, is heavily biased towards zero.
We can extend the comparison by considering k = 3 forecasts. In this case, we consider the reference model p0 = 5p∗
as third competitor. Figure 11, for each scoring rule, shows the expected score differences E[∆(p2 , p1 , X)] (dashed
blue) and E[∆(p2 , p0 , X)] (solid red), representing the expected score difference between p2 and p1 , and the expected
score difference between p2 and p0 . Given that p1 is equal to the true probabilities, the score differences have to be
negative for any value of ω 6= 1 in order for the scoring rule to be effective. Indeed, this is the case for the Brier and log
score (Figure 11 (a), (c)). On the other hand, both the pairwise and full gambling score (Figure 11 (b), (d)) are improper
and prefer p2 over p1 when ω ∈ (0, 1). Moreover, all the scores prefer p0 to p2 when ω > 5. However, the log score
prefers p0 to p2 also when ω approaches zero. This shows, again, how different scoring rules apply different penalties
to the forecasts.
We note that the pairwise and full gambling score present almost the same expected score difference between p2 and
p1 . This is because both scoring procedures implicitly assume a reference model given by the average forecast. If
the average forecast in a bin is greater than p∗i , a forecaster will obtain a positive reward each time they submit a
value smaller than p∗i and Xi = 0 occurs. Therefore, given that we are in a low probability environment for which
Pr[Xi = 0] > 0.99, the smallest forecast is likely to be preferred. The bias depends on the relationship between the
reference model and the true probabilities. In the gambling metaphore, the reference model plays the role of the house
19
Ranking earthquake forecasts using proper scoring rules A P REPRINT
score
Brier
FG
0.0
Log
PG
E[∆]
−0.4
−0.8
0 1 2 3 4
ω
Figure 10: Expected score difference between p1 and p2 as a function of ω = p1 /p2 for ω ∈ (10−3 , 4). We set p∗
equal to the 5-year adaptively-smoothed Italy forecast, p1 = p∗ , p2 = ωp∗ , and reference model for the pairwise
gambling score p0 = 5p∗ .
(or banker) which determines the returns, and against which all forecasts are competing. Considering equation 12, if
p∗ < p0 , the player has a positive reward forecasting p < p0 . If the number of forecasts is large enough that changing a
forecast does not affect significantly the average, then, the smaller the forecast the higher the reward, and the forecaster
is encouraged by the score to provide p = 0. The same reasoning applies if p∗ > p0 .
4.6 Confidence Interval and preference probabilities - Multiple Bins Multiple Probabilities
Also in the Multiple Bins Multiple Probabilities case it is crucial to account for the uncertainty around the observed
score difference. The binomial formulation used before to retrieve confidence intervals no longer holds, and we need an
alternative methodology. One approach to calculate confidence intervals for the expected score difference relies on a
Gaussian approximation of the score difference distribution [Rhoades et al., 2011]. The score difference in each bin,
∆(p1i , p2i , Xi ) for i = 1, ..., N , are assumed to be independent draws from a Gaussian distribution with expected value
E[∆(p1 , p2 , X)] and variance σ 2 . When we observe a sample x = x1 , ..., xN , the point estimate of the expected score
difference is the observed score difference ∆(p1 , p2 , x) and the (1 − α)% confidence interval is given by:
s
∆(p1 , p2 , x) ± t1−α/2,N −1 √ ,
N
where s2 is an estimate of the variance σ 2 and t1−α/2,N −1 is the 1 − α/2 percentile of a t-student distribution with
N − 1 degrees of freedom. The reliability of such interval estimates is determined by the accuracy of the Gaussian
approximation, which, in turns, depends on the amount of data (the more the better) and on the correlation between
the score difference in each bin (the more the worst). We analyse the reliability of this approximation in Appendix A:
20
Ranking earthquake forecasts using proper scoring rules A P REPRINT
0.5 0.5
(a) (b)
0.0 0.0
E[∆]
E[∆]
−0.5 −0.5
E[∆(p2,p0)] E[∆(p2,p0)]
E[∆(p2,p1)] E[∆(p2,p1)]
−1.0 −1.0
0 2 4 6 0 2 4 6
ω ω
0.5 0.5
(c) (d)
0.0 0.0
E[∆]
E[∆]
−0.5 −0.5
E[∆(p2,p0)] E[∆(p2,p0)]
E[∆(p2,p1)] E[∆(p2,p1)]
−1.0 −1.0
0 2 4 6 0 2 4 6
ω ω
Figure 11: Expected score difference between p2 and p1 (blue dashed) and p2 and p0 (red solid) as a function of
ω = p1 /p2 for ω ∈ (10−3 , 7). We set p∗ equal to 5-year adaptively-smoothed Italy forecast, p1 = p∗ , p2 = ωp∗ and
p0 = 5p∗ . Vertical lines represent ω = 1 and ω = 5. We consider the Brier (a), pairwise gambling (b), log (c), and full
gambling (d) scores.
Reliability of the gaussian confidence intervals and conclude that it can be used with the log, pairwise gambling and full
gambling score but not with the Brier score.
Figure 12 shows the evolution of the preference probabilities varying the forecasts ratio, ω = p2 /p1 . It is quite similar
to Figure 6 and the same problems with the pairwise gambling score are evident; i.e. it favours forecast smaller than
the true probability when the average forecast is greater than the latter. On the other hand, the log score probability of
preferring p1 increases rapidly when ω → 0, while the full gambling score is not able to distinguish between p1 and p2
for ω < 2.5. The latter remark suggests a potential problem with the use of the full gambling score given that, in real
forecasting experiments, the competing forecasts tends to be quite similar, in which case there is an high probability of
no preference.
This concludes our analysis on the use of proper scoring rules to rank earthquake forecasting models.
5 Discussion
The parimutuel gambling score was introduced as a general scoring rule to compare, within a unified framework,
earthquake forecasts of different kinds (e.g. alarm-based forecast and probabilistic forecast). It overcomes two
limitations common to other forecast comparison techniques: i) the need to define a reference model, and ii) to allow
forecasts defined on different space-time-magnitude regions to be compared. We showed that the parimutuel gambling
score is proper only when two forecasts are compared directly against each other. In the other cases (multi-forecast
comparison and comparison against a reference model), the parimutuel gambling score is improper. Consequently, we
21
Ranking earthquake forecasts using proper scoring rules A P REPRINT
1.00
Preferences
0.75
no pref
Pr
0.50 (a) p1
0.25
p2
0.00
0 2 4 6
ω
1.00
Preferences
0.75
no pref
Pr
0.50 (b) p1
0.25
p2
0.00
0 2 4 6
ω
1.00
Preferences
0.75
no pref
Pr
0.50 (c) p1
0.25
p2
0.00
0 2 4 6
ω
Figure 12: Multiple Bins Multiple Probabilities case: each plot shows the probability of each possible outcome (solid
no preference, dotted preference for p1 , dashed preference for p2 ) as a function of ω = p1 /p2 for ω ∈ (10−3 , 7).
The log (a), the pairwise gambling (b) and the full gambling (c) scores are considered. We set p∗ equal to 5-year
adaptively-smoothed Italy forecast, p1 = p∗ , p2 = ωp∗ , and reference model for the pairwise gambling score
p0 = 5p∗ .
discourage its use in multi-model comparisons such as CSEP and encourage researchers and practitioners to re-consider
rankings obtained using this score.
Specifically, the parimutuel gambling score tries to avoid the need to pre-define a reference model by using the average
forecast. Therefore, for each bin, a positive reward means that the model is better than the average forecast, vice versa
if the reward is negative. This allows to produce a map of the parimutuel gambling rewards from which to infer the bins
where the forecast is better than the average forecast, and the bins where it is not. Since the parimutuel gambling score
is proper only when k = 2, any map obtained by computing the comparisons for k > 2 may be biased. This difficulties
may be circumvented by using any proper scoring rule that allows for multi-forecast comparison. In fact, maps of this
kind may be produced by reporting the score difference between a forecast and the average one. Furthermore, given
that proper scores are scale invariant, we can re-scale the score values to be between −1 and 1. In this way, we can
visualize which bins has a positive or negative contribution to the average score difference.
The need to compare forecasts defined on different set of bins comes from the design of the forecasting experiment.
In the RELM experiment [Zechar et al., 2013] modelers were allowed to choose a subset of bins to include in their
forecast, referred to as masking. Modellers involved in the RELM experiment provided forecasts with very different
masks; some issued forecasts for the entire California region [Bird and Liu, 2007, Helmstetter et al., 2007, Holliday
et al., 2007], some for only Southern California [Ward, 2007, Shen et al., 2007, Kagan et al., 2007], while others used
irregular masks [Ebel et al., 2007]. The parimutuel gambling score addressed these differences using the gambling
metaphor. Each forecaster is a gambler which plays a certain number of rounds (bets) corresponding to the bins. A
22
Ranking earthquake forecasts using proper scoring rules A P REPRINT
forecaster does not have to make a forecast for every bin - they can just sit out this round. The forecasters are ranked by
their total reward (i.e. the sum of the rewards for each bin). We argue that this solution is still problematic. First, the
parimutuel gambling score needs at least two forecasts for each bin to be computed. If only one forecaster plays in a bin
we can not calculate the parimutuel gambling score for that bin. Second, consider two bins for which different sets of
forecasters provided a forecast; in this situation the models are rewarded with respect to different odds. This becomes
problematic when we attempt to interpret the observed result because in each bin the reference model is given by a
potentially different combination of models.
The Brier score can also be used to assess masked forecasts. The maximum Brier score value obtainable by a forecast is
zero and it is achieved by the perfect forecast which assumes p = 1 when x = 1 and p = 0 when x = 0. Any other
forecast obtains a negative Brier score. Therefore, the Brier score of a forecast can be seen as the Brier score difference
between the perfect forecast and the forecast under evaluation. Given two models, the one with the highest average
Brier score is the one closest (on average) to the perfect forecast. If the models provide forecasts on two different sets
of bins, B1 and B2 , we can still compare the forecasts in terms of their average Brier score. Suppose the first model
achieves an higher average Brier score, we can conclude that, on average, the first model in B1 is closer to the perfect
forecast than the second model in B2 . We can make the above comparison also when B1 and B2 have zero bins in
common because the Brier score requires only a forecast and the observation to be calculated.
In this paper, we have used confidence intervals to asses the statistical significance of observed score differences.
Analytically determining these confidence intervals may be too complex and a basic approximative Gaussian approach
may fail, as highlighted by the Brier score example. Problems of this type come from the fact that the score differences
per bin are treated as independent and identically distributed. This assumption is false, especially when considering
space-time bins which depend on each other both in time and space due to clustering of earthquakes. A possible solution
to relax the independence assumption is to consider a Diebold-Mariano test [Diebold and Mariano, 2002] on the score
differences which takes into account the correlation structure of the score differences sequence.
6 Conclusion
The parimutuel gambling score, commonly applied to compare earthquake forecasts, is improper when the number
of forecasts being tested is greater than two. In the special case of two competing forecasts, the score is proper, and
can return results similar to alternate proper scoring methods, but even then it can be used improperly. In the common
testing scenario of multiple forecasts being compared simultaneously, or when multiple forecasts are compared against
a reference model, the parimutuel gambling score provides a biased assessment of the skill of a forecast when it is
tested against a given outcome. This is fundamentally a problem of the gambling analogy itself; the betting strategy
of maximizing the expected reward (score) does not have to be consistent with the data generating model (in the case
where this is known) and, therefore, gamblers (modellers) are not encouraged to provide forecasts resembling the data
generating model. This is because the score for a given forecast is dependent on all the forecasts taking part in the
competition, not just on the observed data; one can therefore change the ranking of two models by changing one of the
other models in the pool. This introduces the undesirable property that one can potentially game the system to prefer a
specific model. Further, if we only have access to the forecasts and the data, it is impossible to know if the parimutuel
gambling score results will be biased or not. Moreover, the only case in which they are correct is when one of the
competing forecasts is the data generating model, which is highly unlikely. These findings are sufficiently clear for us
to discourage the use of the parimutuel gambling score in distinguishing between multiple competing forecasts, and for
care to be taken even in the case where only two are being compared.
We recommend that alternative scores that do not suffer from these shortcomings should be used instead to assess the
skill of prospective earthquake forecasts in a formal testing environment. The Brier and log scores are both proper, and
require no new information beyond what was used to calculate the parimutuel gambling score, so switching existing
analyses to a proper score should be simple to implement. We recommend testing for properness when introducing new
scoring rules, either analytically or via simulations using a known model to generate testing data.
7 Acknowledgments
All the code to produce the present results is written in the R programming language. We have used the package
ggplot2 [Wickham, 2016], the package rnaturalearth [South, 2017] for the Italy contour map presented in Figure
9 and the package bayesianETAS [Ross, 2016] for the Maximum Likelihood estimate used in Section 5. We thank
Kirsty Bayliss and for her useful feedback and constructive discussions which led to the final version of this work.
23
Ranking earthquake forecasts using proper scoring rules A P REPRINT
Declarations
Some journals require declarations to be submitted in a standardised format. Please check the Instructions for Authors
of the journal to which you are submitting to see if you need to complete this section. If yes, your manuscript must
contain the following sections under the heading ‘Declarations’:
• Funding: This work was founded by the Real-time Earthquake Risk Reduction for a Resilient Europe “RISE”
project, which has received funding from the European Union’s Horizon 2020 research and innovation program
under grant Agreement 821115.
• Conflict of interest/Competing interests : The authors have no financial or proprietary interests in any material
discussed in this article.
• Ethics Approval
• Availability of data and materials : Simulated data and figures can be found at https://github.com/
Serra314/Serra314.github.io/tree/master/Ranking_earthquake_forecast.
• Code availability : The code to generate all the figures in the manuscript can be found at https://github.
com/Serra314/Serra314.github.io/tree/master/Ranking_earthquake_forecast.
• Authors’ contributions : All authors contributed to the study design. Material preparation, analysis and first
draft were performed by Francesco Serafini. Maximilian Werner provided the 5-year adaptively-smoothed
Italy forecast. All authors commented and contributed to previous versions of the manuscript. All authors read
and approved the submitted manuscript.
We assess the reliability of the Gaussian approximation to calculate confidence intervals for the expected score
differences using simulated replicates from p∗ equal to the 5-year adaptively-smoothed Italy forecast. We set the first
forecasts p1 = p∗ , p2 = ωp∗ , and the reference model p0 = 5p∗ . The method is reliable if the probability that the
approximated confidence interval of level α contains the true expected value (coverage probability) is close to 1 − α.
To estimate the coverage probability, for a set of values of ω ∈ (10−3 , 7), we simulated x = x1 , ..., xN 10,000 times
and calculated the approximate 95% confidence intervals. The coverage probability is given by the fraction of times in
which the confidence intervals contains the true expected score difference.
24
Ranking earthquake forecasts using proper scoring rules A P REPRINT
1.0
0.9
coverage probability
0.8
score
Brier
Log
0.7 FG
PG
0 2 4 6
ω
Figure 13: Coverage probability for the Gaussian approximated confidence interval for the expected score difference
between p1 and p2 as a function of ω = p2 /p1 for ω ∈ (10−3 , 7). We consider the Brier, log, pairwise gambling (PG)
and full gambling (FG) scores. We set p∗ equal to the 5-year adaptively-smoothed Italy forecast, p1 = p∗ , p2 = ωp∗
and p0 = 5p∗ . Horizontal line represents the target coverage probability (0.95)
Figure 13 shows that the approximation is not reliable for the Brier score for which produces confidence intervals
which are too small (coverage below 0.95) or too wide (coverage above 0.95) depending on the value of ω. This is
because the Brier score differences have a asymmetrical bimodal distribution as shown in Figure 14a-c and therefore
are not normally distributed. Furthermore, Figure 13 shows also that the coverage for the log, pairwise gambling and
full gambling score is usually below 0.95 and reaches this value only for ω > 6. This is because, taking the log score as
example, the distribution of the score differences becomes smoother as ω grows (Figure 14d-f). The distributions of the
pairwise and full gambling score resemble the log score one.
From this analysis we conclude that it is possible to use the gaussianly approximated confidence intervals for the log,
pairwise gambling and full gambling because the coverage probability is always between 0.88 and 0.96. In this example,
the reliability of the approximation depends on how much different the forecasts are.
25
Ranking earthquake forecasts using proper scoring rules A P REPRINT
density
density
0.50 0.50 0.50
density
density
0.50 0.50 0.50
Figure 14: Average score difference distribution for the Brier score (a-c) and log score (d-f). We set p∗ equal to
the 5-year adaptively-smoothed Italy forecast, p1 = p∗ , p2 = ωp∗ , and reference model for the pairwise gambling
score p0 = 5p∗ . We consider ω = 0.001 (a,d); ω = 1.5 (b,e); ω = 4 (c,f). Black solid line represents the empirical
distribution obtained from 10000 simulations. Red dashed line represents the corresponding Gaussian approximation.
Vertical dotted line represents the true value of the expected score difference.
References
DA Rhoades, M Liukis, A Christophersen, and MC Gerstenberger. Retrospective tests of hybrid operational earthquake
forecasting models for canterbury. Geophysical Journal International, 204(1):440–456, 2016.
SJ Bourne, SJ Oates, and J Van Elk. The exponential rise of induced seismicity with increasing stress levels in the
groningen gas field and its implications for controlling seismic risk. Geophysical Journal International, 213(3):
1693–1700, 2018.
Warner Marzocchi, Anna Maria Lombardi, and Emanuele Casarotti. The establishment of an operational earthquake
forecasting system in italy. Seismological Research Letters, 85(5):961–969, 2014.
Iunio Iervolino, Eugenio Chioccarelli, Massimiliano Giorgio, Warner Marzocchi, Giulio Zuccaro, Mauro Dolce, and
Gaetano Manfredi. Operational (short-term) earthquake loss forecasting in italy. Bulletin of the Seismological Society
of America, 105(4):2286–2298, 2015.
Thomas H Jordan. Earthquake predictability, brick by brick. Seismological Research Letters, 77(1):3–6, 2006.
J Douglas Zechar, Danijel Schorlemmer, Maria Liukis, John Yu, Fabian Euchner, Philip J Maechling, and Thomas H
Jordan. The collaboratory for the study of earthquake predictability perspective on computational earthquake science.
Concurrency and Computation: Practice and Experience, 22(12):1836–1847, 2010a.
26
Ranking earthquake forecasts using proper scoring rules A P REPRINT
Danijel Schorlemmer, Maximilian J Werner, Warner Marzocchi, Thomas H Jordan, Yosihiko Ogata, David D Jackson,
Sum Mak, David A Rhoades, Matthew C Gerstenberger, Naoshi Hirata, et al. The collaboratory for the study of
earthquake predictability: achievements and priorities. Seismological Research Letters, 89(4):1305–1313, 2018.
Tilmann Gneiting and Adrian E Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the
American statistical Association, 102(477):359–378, 2007.
Allan H Murphy. What is a good forecast? an essay on the nature of goodness in weather forecasting. Weather and
forecasting, 8(2):281–293, 1993.
IT Jolliffe and DB Stephenson. Forecast verification. chichester, england and hoboken, 2003.
David B Rosen. How good were those probability predictions? the expected recommendation loss (erl) scoring rule. In
Maximum Entropy and Bayesian Methods, pages 401–408. Springer, 1996.
Aapo Hyvärinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of
Machine Learning Research, 6(4), 2005.
José Hernández-Orallo, Peter Flach, and César Ferri Ramírez. A unified view of performance metrics: Translating
threshold choice into expected classification loss. Journal of Machine Learning Research, 13:2813–2869, 2012.
Edward H Field, Ramon J Arrowsmith, Glenn P Biasi, Peter Bird, Timothy E Dawson, Karen R Felzer, David D
Jackson, Kaj M Johnson, Thomas H Jordan, Christopher Madden, et al. Uniform california earthquake rupture
forecast, version 3 (ucerf3)—the time-independent model. Bulletin of the Seismological Society of America, 104(3):
1122–1180, 2014.
Sandy Steacy, Matt Gerstenberger, Charles Williams, David Rhoades, and Annemarie Christophersen. A new hybrid
coulomb/statistical model for forecasting aftershock rates. Geophysical Journal International, 196(2):918–923, 2014.
Kirsty Bayliss, Mark Naylor, Janine Illian, and Ian G Main. Data-driven optimization of seismicity models using
diverse data sets: Generation, evaluation, and ranking using inlabru. Journal of Geophysical Research: Solid Earth,
125(11):e2020JB020226, 2020.
Danijel Schorlemmer and MC Gerstenberger. Relm testing center. Seismological Research Letters, 78(1):30–36, 2007a.
William H Savran, Maximilian J Werner, Warner Marzocchi, David A Rhoades, David D Jackson, Kevin Milner,
Edward Field, and Andrew Michael. Pseudoprospective evaluation of ucerf3-etas forecasts during the 2019 ridgecrest
sequence. Bulletin of the Seismological Society of America, 110(4):1799–1817, 2020.
J Douglas Zechar and Thomas H Jordan. Testing alarm-based earthquake predictions. Geophysical Journal International,
172(2):715–724, 2008.
J Douglas Zechar and Thomas H Jordan. The area skill score statistic for evaluating earthquake predictability
experiments. In Seismogenesis and Earthquake Forecasting: The Frank Evison Volume II, pages 39–52. Springer,
2010.
Philip B Stark. Earthquake prediction: the null hypothesis. Geophysical Journal International, 131(3):495–499, 1997.
Brad Luen, Philip B Stark, et al. Testing earthquake predictions. In Probability and Statistics: Essays in Honor of
David A. Freedman, pages 302–315. Institute of Mathematical Statistics, 2008.
Warner Marzocchi and J Douglas Zechar. Earthquake forecasting and earthquake prediction: different approaches for
obtaining the best model. Seismological Research Letters, 82(3):442–448, 2011.
D Schorlemmer, MC Gerstenberger, S Wiemer, DD Jackson, and DA Rhoades. Earthquake likelihood model testing.
Seismological Research Letters, 78(1):17–29, 2007.
J Douglas Zechar, Matthew C Gerstenberger, and David A Rhoades. Likelihood-based tests for evaluating space–rate–
magnitude earthquake forecasts. Bulletin of the Seismological Society of America, 100(3):1184–1195, 2010b.
David A Rhoades, Danijel Schorlemmer, Matthew C Gerstenberger, Annemarie Christophersen, J Douglas Zechar, and
Masajiro Imoto. Efficient testing of earthquake forecasting models. Acta Geophysica, 59(4):728–747, 2011.
Max Schneider, Robert Clements, David Rhoades, and Danijel Schorlemmer. Likelihood-and residual-based evaluation
of medium-term earthquake forecast models for california. Geophysical Journal International, 198(3):1307–1318,
2014.
Maximilian J Werner and Didier Sornette. Magnitude uncertainties impact seismic rate estimates, forecasts, and
predictability experiments. Journal of Geophysical Research: Solid Earth, 113(B8), 2008.
J Douglas Zechar, Danijel Schorlemmer, Maximilian J Werner, Matthew C Gerstenberger, David A Rhoades, and
Thomas H Jordan. Regional earthquake likelihood models i: First-order results. Bulletin of the Seismological Society
of America, 103(2A):787–798, 2013.
27
Ranking earthquake forecasts using proper scoring rules A P REPRINT
Warner Marzocchi, J Douglas Zechar, and Thomas H Jordan. Bayesian forecast evaluation and ensemble earthquake
forecasting. Bulletin of the Seismological Society of America, 102(6):2574–2584, 2012.
James R Holliday, Kazuyoshi Z Nanjo, Kristy F Tiampo, John B Rundle, and Donald L Turcotte. Earthquake forecasting
and its verification. arXiv preprint cond-mat/0508476, 2005.
J Douglas Zechar and Jiancang Zhuang. A parimutuel gambling perspective to compare probabilistic seismicity
forecasts. Geophysical Journal International, 199(1):60–68, 2014.
Edward H Field. Overview of the working group for the development of regional earthquake likelihood models (relm).
Seismological Research Letters, 78(1):7–16, 2007.
Jiancang Zhuang. Gambling scores for earthquake predictions and forecasts. Geophysical Journal International, 181
(1):382–390, 2010.
Ian Main. Long odds on prediction. Nature, 385(6611):19–20, 1997.
Kossobokov. Earthquake prediction: basics, achievements, perspectives. Acta Geodaetica et Geophysica Hungarica, 39
(2-3):205–221, 2004.
Kossobokov. Testing earthquake prediction methods: The west pacific short-term forecast of earthquakes with
magnitude mwhr ≥ 5.8. Tectonophysics, 413(1-2):25–31, 2006.
Glenn W Brier. Verification of forecasts expressed in terms of probability. Monthly weather review, 78(1):1–3, 1950.
I. J. Good. Rational decisions. Journal of the Royal Statistical Society, Ser. B, pages 107–114, 1952.
Maximilian J Werner, Agnès Helmstetter, David D Jackson, Yan Y Kagan, and Stefan Wiemer. Adaptively smoothed
seismicity earthquake forecasts for italy. arXiv preprint arXiv:1003.4374, 2010.
Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics, pages 492–518. Springer, 1992.
Warner Marzocchi and Anna Maria Lombardi. Real-time forecasting following a damaging earthquake. Geophysical
Research Letters, 36(21), 2009.
Matteo Taroni, Warner Marzocchi, Danijel Schorlemmer, Maximilian Jonas Werner, Stefan Wiemer, Jeremy Douglas
Zechar, Lukas Heiniger, and Fabian Euchner. Prospective csep evaluation of 1-day, 3-month, and 5-yr earthquake
forecasts for italy. Seismological Research Letters, 89(4):1251–1261, 2018.
J Douglas Zechar and Jiancang Zhuang. Risk and return: evaluating reverse tracing of precursors earthquake predictions.
Geophysical Journal International, 182(3):1319–1326, 2010.
M Taroni, JD Zechar, and W Marzocchi. Assessing annual global m 6+ seismicity forecasts. Geophysical Journal
International, 196(1):422–431, 2014.
MPS Gruppo di Lavoro. Redazione della mappa di pericolosità sismica prevista dall’ordinanza pcm 3274 del 20 marzo
2003. Rapporto Conclusivo per il Dipartimento della Protezione Civile, INGV, Milano-Roma, 5, 2004.
Ronald A Fisher. On the mathematical foundations of theoretical statistics. Philosophical Transactions of the Royal
Society of London. Series A, Containing Papers of a Mathematical or Physical Character, 222(594-604):309–368,
1922.
Mark J Schervish. Theory of statistics. Springer Science & Business Media, 2012.
Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The elements of statistical learning: data mining, inference,
and prediction. Springer Science & Business Media, 2009.
Sean Wallis. Binomial confidence intervals and contingency tests: mathematical fundamentals and the evaluation of
alternative methods. Journal of Quantitative Linguistics, 20(3):178–208, 2013.
Charles J Clopper and Egon S Pearson. The use of confidence or fiducial limits illustrated in the case of the binomial.
Biometrika, 26(4):404–413, 1934.
Danijel Schorlemmer and MC Gerstenberger. Relm testing center. Seismological Research Letters, 78(1):30–36, 2007b.
Andrew J Michael and Maximilian J Werner. Preface to the focus section on the collaboratory for the study of earthquake
predictability (csep): New results and future directions. Seismological Research Letters, 89(4):1226–1228, 2018.
Peter Bird and Zhen Liu. Seismic hazard inferred from tectonics: California. Seismological Research Letters, 78(1):
37–48, 2007.
Agnes Helmstetter, Yan Y Kagan, and David D Jackson. High-resolution time-independent grid-based forecast for m ≥
5 earthquakes in california. Seismological Research Letters, 78(1):78–86, 2007.
James R Holliday, Chien-chih Chen, Kristy F Tiampo, John B Rundle, Donald L Turcotte, and Andrea Donnellan. A
relm earthquake forecast based on pattern informatics. Seismological Research Letters, 78(1):87–93, 2007.
28
Ranking earthquake forecasts using proper scoring rules A P REPRINT
Steven N Ward. Methods for evaluating earthquake potential and likelihood in and around california. Seismological
Research Letters, 78(1):121–133, 2007.
Zheng-Kang Shen, David D Jackson, and Yan Y Kagan. Implications of geodetic strain rate for future earthquakes, with
a five-year forecast of m5 earthquakes in southern california. Seismological Research Letters, 78(1):116–120, 2007.
Yan Y Kagan, David D Jackson, and Yufang Rong. A testable five-year forecast of moderate and large earthquakes in
southern california based on smoothed seismicity. Seismological Research Letters, 78(1):94–98, 2007.
John E Ebel, Daniel W Chambers, Alan L Kafka, and Jenny A Baglivo. Non-poissonian earthquake clustering and the
hidden markov model as bases for earthquake forecasting in california. Seismological Research Letters, 78(1):57–65,
2007.
Francis X Diebold and Robert S Mariano. Comparing predictive accuracy. Journal of Business & economic statistics,
20(1):134–144, 2002.
Hadley Wickham. ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New York, 2016. ISBN 978-3-319-
24277-4. URL https://ggplot2.tidyverse.org.
Andy South. Rnaturalearth: world map data from natural earth. R package version 0.1. 0, 2017.
GJ Ross. Bayesian estimation of the etas model for earthquake occurrences. Preprint, 2016.
29