Short-Term Earthquake Forecasting Using Early Aftershock Statistics

by Peter Shebalin, Clment Narteau, Matthias Holschneider, and Danijel Schorlemmer

Abstract We present an alarm-based earthquake forecast model that uses the early
aftershock statistics (EAST). This model is based on the hypothesis that the time delay
before the onset of the power-law aftershock decay rate decreases as the level of stress
and the seismogenic potential increase. Here, we estimate this time delay from (t
g
), the
time constant of the OmoriUtsu law. To isolate spacetime regions with a relative high
level of stress, the single local variable of our forecast model is the E
a
value, the ratio
between the long-termand short-termestimations of (t
g
). When and where the E
a
value
exceeds a given threshold (i.e., the c value is abnormally small), an alarmis issued, and
an earthquake is expected to occur during the next time step. Retrospective tests show
that the EASTmodel has better predictive power than a stationary reference model based
on smoothed extrapolation of past seismicity. The official prospective test for California
started on 1 July 2009 in the testing center of the Collaboratory for the Study of
Earthquake Predictability (CSEP). During the first nine months, 44 M 4 earthquakes
occurred in the testing area. For this time period, the EAST model has better predictive
power than the reference model at a 1% level of significance. Because the EAST model
has also a better predictive power than several time-varying clustering models tested in
CSEP at a 1%level of significance, we suggest that our successful prospective results are
not due only to the spacetime clustering of aftershocks.
Online Material: The EAST model; retrospective test and nine-month prospective
evaluation versus time-dependent rate-based models.
Introduction
The aftershock rate, , is often well-described by the
OmoriUtsu law,
(t) =
K
(t c)
p
; (1)
where t is the elapsed time since the mainshock, K the after-
shock productivity, p a power-law exponent, and c the time
delay before the onset of the power-law aftershock decay rate
(Utsu, 1961). The cvalue is often interpreted, at least partially,
as an artifact (Kagan, 2004, Lolli and Gasperini, 2006) related
to difficulties in detecting events shortly after a mainshock.
The possible reasons are aftershocks buried in coda waves,
overlapping aftershock records, catalog compiler overload,
absence or malfunction of seismic stations close to the source
zone, and other technical or administrative factors. For exam-
ple, the analysis of high-resolution aftershock records at
Parkfield, California, and in Japan shows that a careful iden-
tification of microearthquakes directly after mainshocks leads
to lower c values than the ones obtained from standard earth-
quake catalogs (Vidale et al., 2004; Peng et al., 2006, 2007;
Enescu et al., 2007, 2009). Nevertheless, even in these
studies, the c value is never equal to zero (Nanjo et al.,
2007). Furthermore, after analyzing catalogs of seismicity
above completeness magnitude thresholds, several studies
have also pointed out that the c value may change according
to the magnitude range of aftershocks (Shcherbakov et al.,
2004) and/or the recurrence time of major earthquakes in
the area (Narteau et al., 2002). From these findings, it can
be argued that many properties of the aftershock decay rate
over short times remain partly unknown, despite evidence that
they can provide a useful source of information about the
physical mechanisms of earthquake triggering.
Theoretical models for aftershock production may
explain systematic variations of the c value. Considering that
aftershocks result from a steplike perturbation of the stress
field in the neighborhood of a triggering event, the vast
majority of these models assume that the amplitude of this
perturbation is inversely proportional to the time delay before
the onset of the power-law aftershock decay rate. For exam-
ple, this is the case in rate-and-state friction models (Dieter-
ich, 1994), in damage mechanics models (Shcherbakov and
Turcotte, 2004; Ben-Zion and Lyakhovsky, 2006), and in
static fatigue models (Scholz, 1968; Narteau et al., 2002).
Then, assuming that the amplitude of the near-field stress
redistribution is directly proportional to the level of stress
imposed by the large-scale tectonic forcing (Atkinson,
297
Bulletin of the Seismological Society of America, Vol. 101, No. 1, pp. 297312, February 2011, doi: 10.1785/0120100119
1987), a decreasing c value may be directly related to the
accumulation of stress within the seismogenic crust.
To test this idea, Narteau et al. (2005) studied stacks
of aftershocks sequences in southern California from the
previous 20 years. To avoid artifacts arising from overlap-
ping records, they did not consider large earthquakes and
their aftershock sequences (Shebalin, 2004); they analyzed
only the largest aftershocks of mainshocks in the magnitude
range of 2:5 < M < 4:5 and disregarded data during the
first minutes after the mainshocks. Then, they observed that
the c value suddenly increases after large earthquakes (e.g.,
Landers, Hector Mine) and slowly decreases at a constant
rate during periods of low seismicity. Using the same ap-
proach along the San Andreas fault (SAF) system in central
California, Narteau et al. (2008) also analyzed the evolution
of the c value during aseismic deformation and transient slip
events. Combining all these observations, they showed that
in creep-slip and stick-slip zones the rate of change of the c
value is proportional to the deficit of slip rate along the SAF.
For a linear elastic rheology, it follows that the rate of change
of the c value can be directly related to the level of stress.
Inspired by this result and the systematic variations of b,
the slope of the earthquake-size distribution (Schorlemmer
et al., 2005), Narteau et al. (2009) used Californian and
Japanese earthquake catalogs with focal mechanisms to
study the evolution of the c value across different faulting
styles. They observed that c values for thrust earthquakes are
on average three times smaller than c values for normal ones,
taking intermediate values for strike-slip events. Assuming
that thrust, strike-slip, and normal faulting are associated
with a decreasing value of differential shear stress (Sibson,
1974), the time delay before the onset of the power-law after-
shock decay rate can be related once again to the level of
stress in the seismogenic crust.
With these observations, we see a growing body of
theoretical and observational evidence that suggests a depen-
dency of the early aftershock decay rate on stress. Here, we
build on this dependence for developing an alarm-based fore-
cast model that uses early aftershock statistics at much more
local scales. This alarm-based model, called EAST for its use
of early aftershock statistics, is designed to detect places more
prone to moderate and large earthquakes within an active fault
zone. It satisfies all the requirements for being tested in testing
center of the Collaboratory for the Study of Earthquake
Predictability (CSEP; Jordan, 2006; Zechar et al., 2010): fore-
casts are provided on a regular predefined spatial grid and for a
predefined time interval and magnitude ranges.
In this paper, we introduce the alarm function of the
EAST model, compare it to the c value, and show how these
two quantities are related to each other. Then, we present the
EAST model (The East Forecast Model section) and a retro-
spective evaluation of its performance using Molchan dia-
grams (Retrospective Evaluation of the EAST Model in
California section). In the The EAST Model in the CSEP
Testing Center for California section, we present the results
of the first nine months of official CSEP testing, which started
on 1 July 2009. Finally, we discuss the stability of our results
and, with respect to the magnitude of the target earthquakes,
the duration for which prospective tests have to be conducted
to determine the predictive power of the model.
Early Aftershocks Statistics: Geometric Mean of
Aftershock Times versus c Value
To compute the time delay before the onset of the
power-law aftershock decay rate, the standard approach is
the maximum-likelihood estimate of the c value in equa-
tion (1). If only short time series of aftershocks are available,
the maximum-likelihood estimator has a large variance.
Moreover, an implicit equation has to be solved for each
estimation. Finally, it is not clear that the OmoriUtsu law
is a good description of the aftershock decay rate over short
time. Therefore, given the difficulty in precisely estimating
the aftershock decay rate over a short time, we propose the
geometric mean of elapsed times from the mainshock to
aftershocks, (t
g
), as a simplified proxy for c. Considering a
single aftershock sequence, we define
(t
g
)

Y
n
i=1
t
i
n
s
= exp

1
n
X
n
i=1
ln(t
i
)

; (2)
where t
i
, i[1; 2; ; n|, is the time between the ith aftershock
in a fixed time interval [t
start
; t
stop
| and the mainshock.
Considering that the aftershock decay rate follows the
OmoriUtsu law, there is a relation between the (t
g
) value
and the c value for a given p value and a given time interval
after a mainshock. In more general terms, let us suppose that
aftershock production follows a nonhomogeneous Poisson
process with rate (t). In this case, considering a time inter-
val [t
start
; t
stop
| after a mainshock, the distribution density of
the elapsed time from mainshock to aftershocks is
f(t) =
(t)
R
t
stop
t
start
(u)du
; t[t
start
; t
stop
|: (3)
In the same time interval, the expectation value of the loga-
rithm of the elapsed time from mainshock to aftershocks is
E(ln(t)) =
Z
t
stop
t
start
ln(t)f(t)dt: (4)
We now suppose that obeys the OmoriUtsu law.
Because the observed p values are always close to 1, we fix
the value to p = 1 in equation (1). Then, we have
(t) =
K
(c t)
: (5)
It follows that
E(ln(t)) =
R
t
stop
t
start
ln(u)
uc
du
ln(t
stop
c) ln(t
start
c)
h(c): (6)
For a specific aftershock sequence, an estimator of this
expectation value is simply the arithmetic mean of the
298 P. Shebalin, C. Narteau, M. Holschneider, and D. Schorlemmer
logarithms of the elapsed times from the mainshock to
aftershocks
ln((t
g
)) =
1
n
X
n
i=1
ln(t
i
): (7)
Therefore, we have an estimator, ^ c, for c by solving the im-
plicit equation
h( ^ c) = ln((t
g
)): (8)
Note that because the function h does not depend on the
aftershock productivity K, this estimator does not depend
on K either. An explicit expression for equation (6) can
be given in terms of a polylogarithm special function
h(c) =
ln(t
stop
) ln(1
t
stop
c
) ln(t
start
) ln(1
t
start
c
) Li
2
(
t
stop
c
) Li
2
(
t
start
c
)
ln(1
t
stop
c
) ln(1
t
start
c
)
; (9)
where
Li
2
(z) =
Z
z
0
ln(1 u)
u
du =
X

k=1
z
k
k
2
: (10)
Figure 1 shows two numerical solutions for t
stop
=
10
1
days: t
start
= 10
7
days (solid line) and t
start
=
10
3
days (dotted line). In both cases, we observe that
(t
g
) is a monotonic function of c with slopes that are slightly
lower than 0.45 in bilogarithmic scale. In what follows, we
only use the (t
g
) value to quantify the characteristic time
delay before the onset of the power-law aftershock decay
rate, keeping in mind this relation to the c value.
The EAST Forecast Model
The EAST model is an alarm-based model with a single
alarm function derived from (t
g
), the geometric mean of
elapsed times from mainshocks to aftershocks (equation 2).
This alarm function is estimated locally in space and time to
determine whether or not an alarm is issued. Where the alarm
function exceeds a given threshold, an alarm is issued, and at
least one target earthquake with a magnitude M M
target
is
expected to occur during the next time step t. Here, the
magnitude threshold M
target
is a free parameter that can be
adjusted to account for different factors (e.g., duration of
the analysis, level of seismic activity). To follow the rules
of the CSEP testing center (Schorlemmer and Gerstenberger,
2007), we choose a time step t = 3 months, and we discre-
tize the area covered by the test into a square grid of side
length = 0:1. Thus, we decompose the entire spacetime
region into a three-dimensional grid of spacetime cells
noted c(x; y; t). For each of these cells, we identify a set of
mainshocks to stack the corresponding aftershock sequences
with respect to the mainshock times.
Mainshock and Aftershock Selection
Our declustering method uses a set of parameters
{M
target
, r
1
, r
2
, T, D, T
short
, T
long
, t
start
, t
stop
, R
0
, N
min
,
M
M
min
, M
M
max
, M
A
min
} that is described in Table 1. First, we
eliminate all earthquakes that are aftershocks of events with
a magnitude M M
target
by removing earthquakes of magni-
tude smaller than M that are within a r
1
10
r
2
M
-km radius dur-
ing the first T days after a magnitude M M
target
earthquake.
Second, an event is not considered a mainshock if there is at
least one earthquake of the same or higher magnitude in the
time interval [t
stop
; t
stop
|. All remaining events are consid-
ered mainshocks. Their respective aftershocks are selected
within R
0
-km radius in the time interval [t
start
; t
stop
| after
the mainshocks focal time. In the model, we always take
t
stop
< 1 days to concentrate on early aftershocks. Further-
more, we always take t
start
> 10 s to limit artifacts related
to aftershock catalog incompleteness immediately after
mainshocks.
Another critical issue of our declustering method is
that the selected earthquakes are classified in ranges of
Figure 1. The geometric mean of elapsed times from
mainshocks to aftershocks, (t
g
), with respect to the c value.
Numerical solutions of equation (9) for t
stop
= 10
1
days with
t
start
= 10
6
days (solid line) and t
start
= 10
4
days (dotted line).
Short-Term Earthquake Forecasting Using Early Aftershock Statistics 299
mainshocks, [M
M
min
; M
M
max
|, and ranges of aftershocks,
[M
A
min
; M
A
max
|. Thus, we investigate only earthquakes from
particular magnitude ranges that can be shown to be
complete even in early times of the aftershock sequence.
Mainshocks should be sufficiently small to have shorter coda
waves such that subsequent aftershocks can be reliably
detected. Aftershocks should be sufficiently large to ensure
completeness at that particular magnitude level from the
early times of an aftershock sequence. These magnitude
ranges are determined from seismological and statistical
constraints following Narteau et al. (2009) (Table 1): M
M
min
is determined by the global catalog completeness; M
M
max
is
the mainshock magnitude below which the mean magnitude
of a given range of aftershocks remain stable; M
A
min
is the
aftershock magnitude for which the estimation of the c value
remains stable as we increase the time at which we start the
fit. On the other hand, M
A
max
is not a parameter of the model
any more because here we simply consider the magnitude of
the mainshock to use the maximum number of events for
each aftershock sequence.
The Alarm Function
Within the study area, each spacetime cell c(x; y; t) is
associated with two spacetime volumes with the same spa-
tial extent, a circle with diameter D, but two different time
intervals T
short
T
long
. The first volume covers the recent
period from t T
short
to t. The second one covers the pre-
ceding period from t T
long
to t T
short
. In these two
spacetime volumes, we identify mainshocks and stack their
aftershocks by sorting them according to the elapsed time
from their respective mainshocks. Let us consider that
N
short
and N
long
are the numbers of aftershocks in the two
stacks. If N
short
< N
min
, we do not define the value of the
alarm function of the corresponding spacetime cells
c(x; y; t). If N
short
N
min
and N
long
N
min
, we use the
stacks of aftershocks as individual sequences to estimate
the (t
g
)
short
and (t
g
)
long
values from equation (2). If N
short

N
min
> N
long
, we take the same equation to calculate the
(t
g
)
short
value but consider that (t
g
)
long
= t
stop
. Then, we
use these short-term and long-term estimations of (t
g
) to cal-
culate E
a
, the alarm function of the EAST model:
E
a
(x; y; t) =
(t
g
)
long
(t
g
)
short
: (11)
The (t
g
)
long
value is used as a reference level to identify
abnormally small values of (t
g
)
short
(i.e., abnormally small
c values). Thus, we try to isolate spacetime regions with
a relative high level of stress that are currently more prone
to earthquakes. In practice, E
a
values above a given thresh-
old, E
0
a
, define the spacetime cells occupied by alarms. In
these cells, earthquakes with magnitude M M
target
are
expected to occur during the next time step t. In addition,
larger E
a
values should correspond to sites that are currently
more vulnerable to target earthquakes.
Retrospective Evaluation of the EAST
Model in California
Before starting prospective testing of the EAST model
in the CSEP testing center, we evaluated the models
performance in retrospective tests for the testing region of
California (Schorlemmer and Gerstenberger, 2007) for the
period from 1984 to 2008. Thus, we test our underlying
hypothesis and analyze the stability of the EAST model with
respect to the model parameters and the choice of the space
time region under consideration. According to the CSEP ex-
periment, we always use the same time step t = 3 months
and the same spatial mesh of 0:1 0:1.
Table 1 shows the values of the parameters of the EAST
model and, for each of them, the range of values for which
the results are stable. During the testing period, approxi-
mately 50% of M 5:0 earthquakes occurred near Cape
Mendocino (mostly offshore), in the CaliforniaNevada bor-
der zone, and in Mexico. In these zones, the density of seis-
mic stations is lower and the number of aftershocks in the
stacks is smaller. Hence, we prefer not to consider the entire
CSEP California testing region (dashed line in Fig. 2) in the
retrospective analysis. Instead, we concentrate on a more
central region where the catalogs of early M >M
A
min
after-
shocks are likely to be more complete (solid line in Fig. 2).
Nevertheless, even in this case, the E
a
value cannot be
defined everywhere because of an insufficient number of
aftershocks, and some target earthquakes occurred in cells
Table 1
Parameter Values of the EAST Model and Their Ranges*
Parameter Units
Value in the
Main Test Minimum Value Maximum Value
M
M
min
2.5 1.8 3.0
M
M
max
4.5 3.4 8.0
M
A
min
1.8 0 1.9
r
1
km 1 10
2
0 3 10
2
r
2
0.725 0 0.8
T d 40 0 1 10
3
D km 25 21 35
T
short
a 5 4.1 7.9
T
long
a 25 8 50
R
0
km 4 1.3 16
t
start
d 1 10
4
0 1 10
3
t
stop
d 1 10
1
3 10
2
3 10
1
N
min
3 2 5
*To determine these ranges, one parameter is changed at a time, the
results are compared to the main retrospective results presented in
Figure 3a (M
target
= 5), and the difference of the integral of the Molchan
trajectory from
RI
= 0 to
RI
= 0:3 should not be larger than 5%.
Because the r
1
, r
2
, and T values can be equal to 0, it is not always
necessary to eliminate mainshocks that are themselves aftershocks of
large earthquakes, as this improves the predicting accuracy of the EAST
model only slightly. The value of M
A
min
and t
start
depend only on
aftershock catalog completeness, and smaller values may be used. This
is also the case for higher M
M
max
and T
long
values. The M
A
max
value is
given by the magnitude of the mainshock. Finally, only six parameters
cannot be eliminated fromthe procedure: M
M
min
, D, T
short
, R
0
, t
stop
, and N
min
.
300 P. Shebalin, C. Narteau, M. Holschneider, and D. Schorlemmer
with undefined E
a
value (see the distributions of gray cells
and open circles in Fig. 2, respectively).
CSEP testing centers host various classes of forecast mod-
els. Most of themare rate based. Their outputs are estimates of
the expected rate of earthquakes in prespecified spacetime
magnitude bins. In contrast, alarm-based forecast models do
not try to quantify earthquake rates. Instead, they concentrate
on the identification of spacetime regions with high seismo-
genic potential. Alarm-based models have at least one control
parameter, the so-called alarm function (Zechar and Jordan,
2008). This alarmfunction is evaluated at every point in space
and time. Its higher values correspond to a higher, but usually
unknown, probability of earthquakes. Then, because it is
impossible to estimate this probability, a threshold value is
used to determine if an alarm is issued locally. By adjusting
the amplitude of this threshold, it is possible to cover any frac-
tion between 0 and 1 of the spacetime region with alarms.
Then, the quality of the prediction has to be quantified with
respect to the entire range of threshold values.
Molchan diagrams (Molchan, 1990, 1991) are usually
used to evaluate the performance of the spacetime predic-
tion of an alarm-based model with respect to a chosen refer-
ence-model or null hypothesis (Molchan and Keilis-Borok,
2008, Zechar and Jordan, 2008). A binary outcome of the
forecasts is considered. If a target earthquake occurs within
an alarm, then a successful prediction is scored. Otherwise it
is a failure to predict or a miss. Molchan diagrams plot the
miss rate with respect to
RI
, a variable that is often de-
scribed as the fraction of the spacetime region occupied
by alarms. More exactly, it is the ratio of the Poissonian
earthquake frequency of the region occupied by alarms to
the Poissonian earthquake frequency of the entire spacetime
region. These frequencies are given by the chosen reference
model. Therefore, 1
RI
is the expected miss rate of the
reference model. In a Molchan diagram, each threshold value
of the alarm function corresponds to a specific point. If the
corresponding miss rate of the model is smaller than 1
RI
,
the model has better predictive power than the reference
model. This comparison can be systematically implemented
from
RI
= 0 (no alarm) to
RI
= 1 (permanent alarm) by
decreasing the threshold value of the alarm function. Then,
we can visually estimate from the so-called Molchan trajec-
tory the predictive power of the model with respect to the
reference model. An important property of this approach
is the higher weight given to a successful prediction in a zone
of low seismicity (small earthquake frequency) than in a zone
of high seismicity (high earthquake frequency). Another
property is that zones without seismicity are not taken into
account.
As a null hypothesis, we here employ a relative intensity
(RI) reference model, a time-independent model of earth-
quake frequencies. This reference model is obtained by
smoothing the location of earthquakes in the past (Kossobo-
kov and Shebalin, 2003; Helmstetter et al., 2006; Molchan
and Keilis-Borok, 2008; Zechar and Jordan, 2008).
Obviously, it would be preferable to use a time-dependent
reference model of seismicity, especially in a prospective
test. Such a comparison of models will be done in the CSEP
testing center during a regular experiment to which our mod-
el is submitted. Hence, at this stage, we prefer to keep the
most commonly accepted RI-type time-independent model.
Different models may have different spatial resolution,
different smoothing length, and different time scales. Here,
we use the same spatial resolution and smoothing length as in
the EAST model in order to assess the gain in prediction
power that may be attributed to the temporal aspect of our
forecast model. Furthermore, mainshocks and aftershocks
are identified using the declustering method of Gardner
and Knopoff (1974), and we use the longest time period
available with respect to catalog completeness in California
before the beginning of the retrospective test. In practice, we
use M 3 earthquakes from 1960 to 1984. We suppose that
this magnitude threshold is sufficiently low to guarantee a
correct evaluation of the RI reference model in all spatial
cells in which an earthquake may occur. Then, for each
spatial cell c(x; y) of the EAST model, the RI reference model
is obtained by calculating the rate of M >3 mainshocks in
circles of diameter D = 25 km.
Figure 2. Testing regions and past seismicity in California.
Dashed lines limit the CSEP California testing region. The dark
gray zone shows the region in which we perform the retrospective
analysis of the EAST model. Solid lines limit three subregions
chosen for additional tests: southern, central, and northwest
California. Gray cells c(x; y) show where the E
a
value is defined
on 1 January 2009. For a period from 1984 to 2008, black circles
show epicenters of M 5 earthquakes that occur in zones where the
E
a
value can be defined. For the same time period, open circles
show epicenters of M 5 earthquakes that occur in zones where
the E
a
value cannot be defined.
Short-Term Earthquake Forecasting Using Early Aftershock Statistics 301
In what follows, Molchan diagrams are always con-
structed according to the same convention:
The solid line is the Molchan trajectory calculated from the
highest (
RI
= 0) to the smallest (
RI
= 1) threshold value
E
0
a
of the alarm function.
The dotted line is the Molchan trajectory that incorporates
zones in which the E
a
value cannot be defined because of
an insufficient number of aftershocks in the stacks. To ob-
tain this curve, we replace the alarm function of the EAST
model by the seismic rate of the reference model in all the
cells c(x; y; t) in which the E
a
value has not been defined.
Then, when the threshold E
0
a
of the alarm function reaches
its lowest value, we complement the spacetime region
occupied by the alarm with the cell c(x; y; t), which has
a decreasing seismic rate in the reference model. Thus,
we have an increasing
RI
value, and we can calculate
the corresponding miss rate .
The dashed diagonal line = 1
RI
corresponds to an
unskilled forecast model with respect to the reference mod-
el, considering an infinite number of target earthquakes
(see Zechar and Jordan, 2010, for a finite number of target
earthquakes).
The shaded area is the zone of the Molchan diagram in
which the prediction of the EAST model is better than
the prediction of the RI reference model at a level of sig-
nificance of = 1% (Kossobokov and Shebalin, 2003;
Shebalin et al. 2006; Zechar et al., 2007). In other words,
only 1% of the best predictions of the RI reference model
fall in the shaded region. To find the limit of this area at a
given
RI
value, we consider that the number of earth-
quakes predicted by the reference model is a random vari-
able that follows a binomial distribution with parameters n
and P. n is the number of M >M
target
earthquakes in the
spacetime region under consideration. P is the probability
to predict a M >M
target
in this region according to the RI
reference model. From this binomial distribution, we can
calculate the miss rate distribution of mean value 1
RI
.
Then, the upper limit of the shaded region is simply the
quantiles of this miss-rate distribution.
Figure 3 shows a comparison between the prediction of
the EAST model and the prediction of the RI reference model
for three different M
target
values, M
target
{5; 5:5; 6]. (The
same comparisons for the entire CSEP testing region are
available as an electronic supplement to this paper.)
In all three cases, the Molchan trajectories show that
the EAST model has better predictive power than the RI
reference model at a significance level of 1%. In addition,
steeper Molchan trajectories with increasing M
target
values
suggest that the EAST model works better in zones where
large events did occur. Equivalently, we may say that the
apparent slope of the expected earthquake-size distribution
is likely to decrease in the spacetime region occupied by
alarms.
In Figure 4 and Figure 5, we analyze in more detail the
spacetime structures of alarms for two threshold values of
the alarm function: E
0
a
= 2:0 (
RI
= 0:1 in Fig. 3) and E
0
a
=
1:0 (
RI
= 0:22 in Fig. 3). Figure 4 shows the histogram of
the duration of successful alarms for four different M
target
values, M
target
{4; 5; 5:5; 6]. Using the time step t =
3 months, we obtain the duration of a successful alarm by
counting the number of consecutive times the E
a
value ex-
ceeded the threshold value E
0
a
before a predicted M >M
target
earthquake. The median of all distributions is approximately
one year; and, in about 80% of the cases, the alarm duration
is less than two years. These results suggest a characteristic
time scale of less than a year for the forecast of the EAST
model. In this case, the time step t = 3 months used in
the CSEP testing center is small enough to be implemented
in the EAST model.
Figure 5 shows the evolution of
RI
with respect to time
for the entire region of the retrospective test (solid lines
in Fig. 2) and three subregions (dotted lines in Fig. 2).
Figure 3. Retrospective evaluation of the EAST model in California from 1984 to 2008 for three M
target
values: (a) M
target
= 5, (b) M
target
= 5.5, (c) M
target
= 6. Using a Molchan diagram, we compare the prediction of the EAST model to the prediction of the RI reference model.
The solid line is the Molchan trajectory calculated from the highest to the lowest threshold value E
0
a
of the alarm function. The dotted line is
the Molchan trajectory that incorporates zones where the E
a
value cannot be defined (see text). The dashed diagonal line corresponds to an
unskilled forecast model with respect to the reference model. The shaded area indicates the zone of the Molchan diagram in which the
prediction of the EAST model is better than the prediction of the RI reference model at a level of significance = 1%.
302 P. Shebalin, C. Narteau, M. Holschneider, and D. Schorlemmer
The
RI
values are systematically larger at the beginning of
the retrospective test. Interestingly, the number of M >
M
target
earthquakes is also larger in this earlier period. To
check the stability of the prediction algorithm on two periods
with a different level of seismicity, we perform two indepen-
dent analyses from 1 January 1984 to 30 June 1992 and from
1 July 1992 to 31 December 2008 (Fig. 6). In both cases,
Molchan trajectories are quite similar to each other and
Figure 4. Distributions of the duration of successful alarms: (a) M
target
= 4, (b) M
target
= 5, (c) M
target
= 5.5, and (d) M
target
= 6. Histograms
in light and dark gray correspond to E
0
a
= 1 and E
0
a
= 2, respectively. Solid and dashed lines mark the 50% and 80% quantiles, respectively.
0.0
0.1
0.2
0.3

R
I
0.0
0.1
0.2
0.3

R
I
0.0
0.1
0.2
0.3

R
I
0.0
0.1
0.2
0.3

R
I
1985 1990 1995 2000 2005 1985 1990 1995 2000 2005
1985 1990 1995 2000 2005 1985 1990 1995 2000 2005
(a) (c)
(b) (d)
Figure 5. Temporal evolution of
RI
in (a) the entire region covered by the test and in (b) southern, (c) central, and (d) northwest
California (see the solid and dotted lines in Fig. 2). Dashed and solid lines correspond to E
0
a
= 1 and E
0
a
= 2, respectively.
Short-Term Earthquake Forecasting Using Early Aftershock Statistics 303
do not exhibit any substantial changes when compared to the
main retrospective result that covers the entire period (Fig. 3).
We conclude that the EAST model is stable in time despite
the permanent evolution of seismic activity and the asso-
ciated variation of alarm distributions.
We believe that the EAST model can be applied world-
wide in a large diversity of active tectonic settings. To verify
that the EAST model can be generalized to different seismic
zones, we use the same set of parameters to perform a retro-
spective analysis of three subregions in California (dotted
lines in Fig. 2). Namely, we consider southern, central, and
northeast California. Figure 7 shows the corresponding Mol-
chan diagrams for three M
target
values, M
target
{5; 5:5; 6]. We
see that the EAST model has a good performance in two of the
three subregions. In northeast California, despite the small
number of targets, the result is even better than for entire
California (Fig. 3). About 50% of M 5:0 earthquakes,
60% of M 5:5 earthquakes, and both M 6:0 earthquakes
are forecast with
RI
value as small as 4%. Nevertheless,
results are worse in central California because we cannot dis-
tinguish between the miss rate of the EASTmodel and the miss
rate of the RI reference model at a significance level of 1%.
The main reason is that the epicenters of the Loma Prieta
(17 October 1989, M
w
6.9) and Parkfield (28 September
2004, M
w
6.0) events are in the spacetime regions where
the E
a
value cannot be defined. However, high E
a
values are
observed nearby. Actually, the results in central California
become as good as in other subregions if we decrease the
minimum aftershock magnitude threshold from M
A
min
= 1:8
to M
A
min
= 1:7 and take the centroid locations instead of
the epicenter locations for the Loma Prieta and Parkfield
events.
Short-term and long-term estimations of (t
g
) values
range from 5 10
2
to 5 10
1
day in more than 99%
of the cases. These values correspond to (1) E
a
values close
to unity, with extreme values in the range of [10
2
; 10
2
|, and
(2) c values that range from 10
4
to 10
2
day, taking t
start
=
10
7
day and t
stop
= 10
1
day in equation (9). These obser-
vations are in good agreement with reported c values in
California for the same aftershock magnitude ranges (Nar-
teau et al., 2002; Shcherbakov et al., 2004). Figure 8 shows
EAST forecast maps before the three largest earthquakes in
southern California during the period of the retrospective
test: Landers (28 June 1992, M
w
7.3), Northridge (17 Jan-
uary 1994, M
w
6.7), Hector Mine (16 October 1999,
M
w
7.1). These maps also give an idea of the temporal
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0

RI
61 targets of M5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0

RI
28 targets of M5.5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0

RI
9 targets of M6
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0

RI
66 targets of M 5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0

RI
14 targets of M 5.5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0

RI
4 targets of M 6
=0.01
(a) (b) (c)
(d) (e) (f)
Figure 6. Retrospective evaluation of the EAST model in California for two time periods and three M
target
values: (a and d) M
target
= 5,
(b and e) M
target
= 5.5, and (c and f) M
target
= 6. Using a Molchan diagram, we compare the prediction of the EAST model to the prediction of
the RI reference model. Arbitrarily, we choose two time periods with approximately the same number of M 5:0 earthquake from January
1984 to June 1992 (top diagrams) and from July 1992 to December 2008 (bottom diagrams). For comparison, a thin dashed line shows the
result obtained in Figure 3 from January 1984 to December 2008.
304 P. Shebalin, C. Narteau, M. Holschneider, and D. Schorlemmer
variation of the forecasts from our model. For comparison,
Figure 8d shows the forecast of the RI model that we used as
a reference model in the Molchan diagrams. We use the same
smoothing procedure for all these maps, but the RI map looks
more smoothed than the EAST maps. The main reason is the
absence of E
a
values in about 90% of the spacetime regions
c(x; y; t) due to the restrictive aftershock selection process
(see the Mainshock and Aftershock Selection section).
The RI reference model has nonzero values in many of these
cells because of the long time period over which it has been
calculated; this is in contrast to the EAST maps, which are
more fragmented and with sharp peaks.
The EAST Model in the CSEP Testing Center
for California
We submitted the EAST model to the CSEP testing center
for California in April 2009. To entirely satisfy the CSEP
requirements, we use M
target
= 3.95 and expand the predic-
tion to the entire state of California (dashed lines in Fig. 2).
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0

RI

RI

RI
RI

RI

RI

RI

RI
RI
23 targets of M5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
11 targets of M5.5
Northeast California
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
2 targets of M6
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
18 targets of M5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
7 targets of M5.5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
3 targets of M6
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
83 targets of M5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
23 targets of M5.5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
7 targets of M6
=0.01
(a) (b) (c)
(d)
(e) (f)
(g)
(h)
(i)
Figure 7. Retrospective evaluation of the EAST model in (top) northeast, (center) central, and (bottom) southern California from 1984 to
2008 for three M
target
values: (left) M
target
= 5, (central) M
target
= 5.5, and (right) M
target
= 6. Using a Molchan diagram, we compare the
prediction of the EAST model to the prediction of the RI reference model. For comparison, thin dashed lines show the result obtained in
Figure 3 for California. For central California, thin solid lines show the Molchan trajectory using M
A
min
= 1:7 and centroid locations instead
of the epicenter locations for Loma Prieta and Parkfield (see text).
Short-Term Earthquake Forecasting Using Early Aftershock Statistics 305
Note that we expect an increasing number of misses outside
the region of the retrospective test (solid lines in Fig. 2)
because of an increase of the completeness threshold of
the network, especially in offshore areas.
The prospective test of the EAST forecast model for
California started officially on 1 July 2009. To date, three
three-month forecast periods have ended; and, based on these
nine months of seismicity, we can perform a comparison
between the predictive power of the EAST model and the
predictive power of the RI reference model. To produce
the RI maps, we use the same smoothing procedure as pre-
viously described here but count the number of M 3:0
events in the period from 1960 to 2008. Nevertheless, in
the retrospective analysis, we have observed that target earth-
quakes often occur in cells of high E
a
value but of moderate
RI frequency. Hence, we currently use a modified E
a
value
E
a1
(x; y; t) =
E
a
(x; y; t)

g
(x; y)
; (12)
Figure 8. Examples of EAST forecast maps and RI map. EAST forecast maps show the distribution of E
a
values before the three largest
events in southern California from 1984 to 2008: (a) Landers forecast map from 1 April to 30 June 1992. (b) Northridge forecast map from 1
January to 31 March 1994. (c) Hector-Mine forecast map from 1 October to 31 December 1999. (a) and (b) have the same color scale as (c).
Circles show epicenters of M >5 earthquakes that occurred during the corresponding forecasts. For the RI map (d), shading represents a
number of earthquakes. For each cell c(x; y; t), we count M 3:0 earthquakes in circles of diameter D = 25 km from 1960 to 1984. The
color version of this figure is available only in the electronic edition.
306 P. Shebalin, C. Narteau, M. Holschneider, and D. Schorlemmer
where
g
(x; y) is the smoothed frequency of M 4 earth-
quakes since 1960, calculated using a two-dimensional
Gauss filter with standard deviation of 10 km. All the other
model parameters have the same values as in the third col-
umn of Table 1.
During the first nine months of the test, 44 target earth-
quakes have occurred in the region under consideration, 21 of
them in cells c(x; y; t) with relatively high E
a1
value (Fig. 9).
Figure 10a shows that the EAST model has better predictive
power than the RI reference model at a significance level
Figure 9. EAST forecast maps for the periods from (a) 1July30 September 2009, (b) 1 October31 December 2009, and (c) 1 January
30March 2010. For this last period, the E
a1
value is calculated from the number of aftershocks shown in (d). Circles show epicenters of M
M
target
earthquakes that occurred during the corresponding forecasts. The color version of this figure is available only in the electronic edition.
Short-Term Earthquake Forecasting Using Early Aftershock Statistics 307
of 1%. For example, approximately 40% of the target
earthquakes are predicted with a
RI
value that is less than
0.05. In addition, the current performance of the EAST
model is better than that obtained retrospectively in the
whole CSEP testing region. Not surprisingly, the predictive
power of the EAST model is even higher within the reduced
area of the retrospective test (Fig. 10b). These stimulating
results are mostly due to the successful forecast of a swarm
of 11 M M
target
earthquakes in central California (Fig. 9b).
However, 10 other events occurred in cells with high E
a1
values in northern California and near the United States
Mexico border (Fig. 9b,c). Even if we exclude the 11 events
of the swarm from the comparison, the EAST model still
has better predictive power than the RI reference model.
Accordingly, we suggest that our successful prospective
results are not due only to the spacetime clustering of
aftershocks.
To test the influence of earthquake clustering, we take
advantage of the CSEP experiment by comparing the
prediction of the EAST model with the prediction of four
time-dependent rate-based clustering models. Two are three-
months models proposed by Rhoades (2007): Every Earth-
quake a Precursor According to Scale (EEPAS; 5 versions)
and Parkfield Prediction Experiment (PPE; 2 versions). Two
others are one-day models: epidemic-type aftershock se-
quence (ETAS; Ogata, 1998, prepared for the test in California
by Zhuang and Liukis) and short-term earthquake probability
(STEP; Gerstenberger et al., 2005). The three-months models
are tested in California for M
target
= 4.95. Only six M >4:95
earthquakes have occurred during the first nine months of
the test. This number is obviously too small to obtain statis-
tically significant results. Then, we have extrapolated the
EEPAS and PPE forecasts to M
target
= 3.95 and performed
the joint analysis for both M
target
values. The Molchan
diagrams for all these tests are available as an electronic
supplement to this paper (Figs. S2S7). In all cases, the EAST
model outperformed the time-dependent reference models.
For M
target
= 3.95, this result is obtained with a 1% signifi-
cance level. Furthermore, the three-month EAST model shows
much better performance than one-day ETAS and STEP mod-
els. This may be partially the result of a lower spatial resolu-
tion in those models. Nevertheless, the final evaluation of the
EAST model, as well as a reliable estimation of its relative ac-
curacy compared with time-independent and time-dependent
models, obviously requires a longer-term analysis of the
prospective test.
Discussion and Perspectives
The working hypothesis of the EAST forecast model is
that the time delay before the onset of the aftershock decay
rate is anticorrelated with the level of stress in the seismo-
genic crust. To characterize this time delay, we work at small
spatial (<10 km) and temporal (<12 hr) scales to concen-
trate on the effect of coseismic stress perturbation and to re-
duce the influence of other factors that may control the
distribution of aftershocks (e.g., afterslip [Peng and Zhao,
2009] or fluids [Mori et al., 2008]). In addition, we introduce
a new quantity (t
g
), the geometric mean of the elapsed times
0.0
0.2
0.4
0.6
0.8
1.0
(a)
(b)
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0

RI

RI
44 targets of M4
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
34 targets of M4
=0.01
Figure 10. Evaluation of the EAST model in California from 1 July to 31 December 2009 for M
target
= 3.95 in (a) the CSEP testing region
and (b) the region used in the retrospective test (Fig. 2). Using a Molchan diagram, we compare the prediction of the EAST model to the
prediction of the RI reference model. The solid line is the Molchan trajectory calculated from the highest to the lowest threshold value E
0
a
of
the alarm function. The dotted line is the Molchan trajectory that incorporate zones where the E
a
value cannot be defined (see text).
The dashed diagonal line corresponds to an unskilled forecast model with respect to the reference model. The shaded area indicates
the zone of the Molchan diagram in which the prediction of the EAST model is better than the prediction of the RI reference model
at a level of significance = 1%
308 P. Shebalin, C. Narteau, M. Holschneider, and D. Schorlemmer
from mainshocks to aftershocks. We could have continued
on to estimate the c value of the OmoriUtsu law (Utsu,
1961) or the
b
value of the limited power-law model (Nar-
teau et al., 2002). However, any model explaining the onset
of the power-law aftershock decay rate is an idealization of a
real behavior that is more complex. For this reason, we con-
sider a nonparametric value such as the geometric mean (t
g
)
as a more meaningful quantity than the c value and the
b
value to assess the time delay before the onset of the after-
shock decay rate.
In both, retrospective and prospective tests, the EAST
model shows better performance than the RI reference model.
Nevertheless, we complement our analysis by using this
reference model in spacetime regions in which we cannot
identify at least N
min
aftershocks. The results of the alarm-
based model therefore may be biased by aftershock produc-
tivity, and we have to test that the E
a
value is the dominant
contribution to the model. With this objective in mind, we
repeat the retrospective analysis using only the spacetime
regions c(x; y; t) in which the E
a
value is defined. Accord-
ingly, we only consider M >M
target
earthquakes that oc-
curred in these regions from t to t t. Thus, we keep 91
events out of 124 for M
target
= 5, 31 out of 42 for M
target
= 5.5,
and 10 out of 13 for M
target
= 6. Figure 11 also shows that in
these cases the predictive power of the EAST model is better
that the predictive power of the RI reference model at a 1%
significance level. Consequently, there is no clear influence
of the fragmented distribution of aftershocks on the perfor-
mance of the forecast of the EAST model.
From the Molchan diagrams, we can systematically
estimate the probability gain (Aki, 1981):
=
1

RI
: (13)
In a vast majority of cases, the comparisons between the
EAST model and the RI reference model show values close
to 10. In other forecast models, similar probability gain
values are mostly due to short-term forecasts of triggered
events (Helmstetter et al., 2006, Kagan and Jackson, 2000).
Considering only mainshocks, the prediction power of these
models is significantly lower (Kossobokov, 2006; Schorlem-
mer et al., 2010). This may be also the case for the EAST
model, especially for small M
target
values for which an in-
creasing number of target earthquakes may be aftershocks.
For this reason, we now estimate the predictive power of
the EAST model when all M >M
target
aftershocks are re-
moved. To identify mainshocks and aftershocks, we use
the declustering algorithm of Gardner and Knopoff (1974)
and repeat the retrospective evaluation of the EAST model
using only M >M
target
mainshocks as the target. Using
E
a1
values, Figure 12 shows the corresponding probability
gain curves for three M
target
values, M
target
{4; 4:5; 5], and
a comparison with the probability gain curves obtained with
M >M
target
aftershocks. In all cases, the EAST model has
better predictive power than the RI reference model. For
M
target
<5, the EAST model has a higher predictive power
with M >M
target
aftershocks for a small threshold value
of the alarm function (high rate of hits). For high threshold
values of the alarm function (low rate of hits) the EAST mod-
el works better if the M >M
target
aftershocks are removed.
For M
target
= 5, the performance of the EAST model is always
better without than with aftershocks. Combined with the
fact that the number of M >M
target
aftershock decrease with
an increasing M
target
value, this result may also explain why
the EAST model works better for higher M
target
values
(see Fig. 3).
This paper has two goals. First, it proposes a new alarm-
based model that may improve seismic hazard assessment.
Second, it tests whether the time delay before the onset of
the power-law decay rate can be used to characterize the
level of stress in the seismogenic crust. We think that the
results presented here extend the set of evidence that allows
us to answer positively to this question. In Figure 12, the
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0

RI
RI

RI
(a)
93 targets of M5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
(b)
32 targets of M5.5
=0.01
0.0
0.2
0.4
0.6
0.8
1.0
M
i
s
s

R
a
t
e
0.0 0.2 0.4 0.6 0.8 1.0
(c)
10 targets of M6
=0.01
Figure 11. Retrospective evaluation of the EAST model in the spacetime regions of California in which the E
a
value can be defined.
This evaluation covers a period from 1984 to 2008 for three M
target
values: (a) M
target
= 5, (b) M
target
= 5.5, and (c) M
target
= 6. Using a
Molchan diagram, we compare the prediction of the EAST model to the prediction of the RI reference model. The solid line is the Molchan
trajectory calculated from the highest to the lowest threshold value E
0
a
of the alarm function. The dashed diagonal line corresponds to an
unskilled forecast model with respect to the reference model. The thin line limits the zone of the Molchan diagram in which the prediction of
the EAST model is better than the prediction of the RI reference model at a level of significance = 1%.
Short-Term Earthquake Forecasting Using Early Aftershock Statistics 309
difference between the probability gain curves with and
without M >M
target
aftershocks also support this hypothesis.
In fact, a vast majority of M >M
target
aftershocks are likely
to result from a perturbation of the state of stress in zones
where the level of stress may be significantly lower before
the mainshock. At the three-month time step of the EAST
model, it is impossible to capture such abrupt variations,
and the model performs better when we remove aftershocks.
A possible direction for further investigations is to refine the
spacetime mesh of the EAST model to forecast large after-
shocks with the same accuracy.
Another perspective is to develop a frequency-based
forecast model using retrospective results of the EAST model.
In fact, for different ranges of E
a
value, the long-term aver-
aged expected frequency of events may be calculated from
the rate of hits per spacetime cell c(x; y; t). In this case, the
probability of having an M >M
target
earthquake during the
next time step is likely to decrease extremely rapidly with
respect to the M
target
value. As a result, for practical applica-
tions, it may be more useful to detect peak areas and peak
periods in the probability map. Additional development of
the EAST model in combination with other methods might
further resolve this problem, too (Schorlemmer and Wiemer,
2005).
A definitive evaluation of the EAST model in the CSEP
testing center requires long-term analysis. Using the Mol-
chan diagrams of the retrospective evaluation, we can try to
estimate this time for different M
target
values. In practice, we
take the Molchan trajectories obtained in California from
1984 to 2008 for M
target
= {4; 5; 6] (two of them are plotted
in Fig. 3). Then, we recalculate the = 1% quantiles of the
miss-rate distribution for an increasing number of events. As
this number increases, the standard deviation of this binomial
distribution is decreasing, and the zone in which the null-
hypothesis cannot be rejected narrows and focuses around
the diagonal 1
RI
. When the Molchan trajectory is entirely
below this critical level , we note the corresponding number
of events. We obtain values of 65 for M
target
= 4, 40 for M
target
= 5, and 8 for M
target
= 6. Using the corresponding average
seismic rate, we conclude that the EAST model need to be
tested for 1 year for M
target
= 4, 5 years for M
target
= 5,
and 10 years for M
target
= 6. Taking into account that, in a
forward test, results may be slightly worse, the duration of
the test may become a little longer.
Conclusions
In a new alarm-based model, called the EAST model, the
time delay before the onset of the aftershock decay rate is
used to identify spacetime regions with a higher level of
stress and, consequently, a higher seismogenic potential.
Retrospective analysis and the current evaluation in the CSEP
testing center show some promising results that indicate the
feasibility of the approach and that support our working
hypothesis. Hence, we conclude that early aftershock decay
rate may be a powerful way to estimate the evolution of the
level of stress within the seismogenic crust and, more
generally, a diagnostic tool for earthquake activity at both
regional and local scales.
Data and Resources
The prospective test of the EAST model is carried out in
the framework of the Collaboratory for the Study of Earth-
quake Predictability (CSEP; http://www.cseptesting.org/).
The Advanced National Seismic System (ANSS) earthquake
catalog was searched using http://quake.geo.berkeley.edu/
cnss/catalog-search.html. Most of the plots were made
using the Generic Mapping Tools version 4.2.1 (www
.soest.hawaii.edu/gmt; Wessel and Smith, 1998).
0
5
10
15
P
r
o
b
a
b
i
l
i
t
y

G
a
i
n
,

(a)
(b)
(c)
0.0 0.2 0.4 0.6
Rate of Hits, 1-
=0.01
1279 targets of M4
(456 mainshocks)
0
5
10
15
20
25
P
r
o
b
a
b
i
l
i
t
y

G
a
i
n
,

0.0 0.2 0.4 0.6
Rate of Hits, 1-
=0.01
404 targets of M4.5
(150 mainshocks)
0
5
10
15
20
25
30
35
40
45
50
P
r
o
b
a
b
i
l
i
t
y

G
a
i
n
,

0.0 0.2 0.4 0.6
Rate of Hits, 1-
128 targets of M5
(57 mainshocks)
Figure 12. Probability gain plots for the retrospective evalua-
tion of the EAST model: (black line) with and (gray line) without
M >M
target
aftershocks . This evaluation covers a period from 1984
to 2008 for three M
target
values: (a) M
target
= 4, (b) M
target
= 4.5, and
(c) M
target
= 5. The values are calculated from the RI reference
model. Circles indicate the limit at which we start to incorporate
spacetime regions in which the E
a
value is not defined. As in a
Molchan diagram, the shaded areas show the = 1% quantiles
of the miss-rate distribution.
310 P. Shebalin, C. Narteau, M. Holschneider, and D. Schorlemmer
Acknowledgments
This paper has been improved by the constructive comments of
J. Zechar and two anonymous reviewers. The authors are also grateful to
M. Liukis for invaluable help in installing the EAST model in CSEP and
the evaluation of the prospective forecasts. This work has been partially
supported by the French Ministry of Research (ANR-09-RISK-02-001/
CASAVA).
References
Aki, K. (1981). A probabilistic synthesis of precursory phenomena, in
Earthquake Prediction: An International Review, Maurice Ewing
Series 4, D. Simpson and P. Richards (Editors), American Geophysical
Union, Washington, D.C., 566574.
Atkinson, B. K. (Editor) (1987). Fracture Mechanics of Rocks, Academic
Press, New York, 534 pp.
Ben-Zion, Y., and V. Lyakhovsky (2006). Analysis of aftershocks in a litho-
spheric model with seismogenic zone governed by damage rheology,
Geophys. J. Int. 165, 197210.
Dieterich, J. (1994). A constitutive law for rate of earthquake production and
its application to earthquake clustering, J. Geophys. Res. 99, 2601
2618.
Enescu, B., J. Mori, M. Miyazawa, and Y. Kano (2009). Omori-Utsu law c-
values associated with recent moderate earthquakes in Japan, Bull.
Seismol. Soc. Am. 99, 884891.
Gardner, J., and L. Knopoff (1974). Is the sequence of earthquakes in south-
ern California with aftershocks removed Poissonian?, Bull. Seismol.
Soc. Am. 5, 13631367.
Gerstenberger, M. C., S. Wiemer, L. M. Jones, and P. A. Reasenberg (2005).
Real-time forecasts of tomorrows earthquakes in California, Nature
435, 328331.
Helmstetter, A., Y. Y. Kagan, and D. D. Jackson (2006). Comparison of
short-termand time-independent earthquake forecast models for south-
ern California, Bull. Seismol. Soc. Am. 96, 90106.
Jordan, T. H. (2006). Earthquake predictability, brick by brick, Seismol. Res.
Lett. 77, 36.
Kagan, Y. Y. (2004). Short-term properties of earthquake catalogs and
models of earthquake source, Bull. Seismol. Soc. Am. 94, 12071228.
Kagan, Y. Y., and D. D. Jackson (2000). Probabilistic forecasting of earth-
quakes, Geophys. J. Int. 143, 438454.
Kossobokov, V. (2006). Testing earthquake prediction methods: The west
Pacific short-term forecast of earthquakes with magnitude m
w
5:8,
Tectonophysics 413, 2531.
Kossobokov, V., and P. Shebalin (2003). Earthquake prediction, chap. 4
in Nonlinear Dynamics of the Lithosphere and Earthquake
Prediction, V. I. Keilis-Borok and A. A. Soloviev (Editors), 141
205, Springer-Verlag, New York, 141207.
Lolli, B., and P. Gasperini (2006). Comparing different models of aftershock
rate decay: The role of catalog incompleteness in the first times after
main shock, Tectonophysics 423, 4359.
Molchan, G. (1990). Strategies in strong earthquake prediction, Phys. Earth
Planet. In. 61, 8498.
Molchan, G. (1991). Structure of optimal strategies in earthquake prediction,
Tectonophysics 193, 267276.
Molchan, G., and V. Keilis-Borok (2008). Earthquake prediction: Probabil-
istic aspect, Geophys. J. Int. 173, 10121017.
Mori, J., Y. Kano, and B. Enescu (2008). Comparison of early aftershock
sequences for the 2004 Mid-Niigata and 2007 Noto Hanto earthquakes
in central Japan, Earth Planets Space 60, 151154.
Nanjo, K. Z., B. Enescu, R. Shcherbakov, D. L. Turcotte, T. Iwata, and
Y. Ogata (2007). Decay of aftershock activity for Japanese earth-
quakes, J. Geophys. Res. 112, no. B08309, 12 pp., doi 10.1029/
2006JB004754.
Narteau, C., P. Shebalin, and M. Holschneider (2002). Temporal limits of the
power law aftershock decay rate, J. Geophys. Res. 107, no. B122359,
14 pp., doi 10.1029/2002JB001868.
Narteau, C., P. Shebalin, and M. Holschneider (2005). Onset of power law
aftershock decay rates in southern California, Geophys. Res. Lett. 32,
no. B03306, 5 pp., doi 10.1029/2005GL023951.
Narteau, C., P. Shebalin, and M. Holschneider (2008). Loading rates in
California inferred from aftershocks, Nonlin. Proc. Geophys. 15,
245263.
Narteau, C., S. Byrdina, P. Shebalin, and D. Schorlemmer (2009). Common
dependence on stress for the two fundamental laws of statistical
seismology, Nature 462, 642645.
Ogata, Y. (1998). Space-time point-process models for earthquake occur-
rences, Ann. Inst. Statist. Math. 50, 379402.
Peng, Z., and P. Zhao (2009). Migration of early aftershocks following the
2004 Parkfield earthquake, Nature Geosci. 28, 877881.
Peng, Z. G., J. E. Vidale, and H. Houston (2006). Anomalous
early aftershock decay rate of the 2004 M
w
6.0 Parkfield, California,
earthquake, Geophys. Res. Lett. 33, L17307, doi 10.1029/
2006GL026744.
Peng, Z. G., J. E. Vidale, M. Ishii, and A. Helmstetter (2007). Seismicity rate
immediately before and after main shock rupture from high frequency
waveforms in Japan, J. Geophys. Res. 112, no. B03306, doi 10.1029/
2006JB004386.
Rhoades, D. A. (2007). Application of the EEPAS model to forecasting
earthquakes of moderate magnitude in southern California, Seismol.
Res. Lett. 78, 110115.
Scholz, C. (1968). Microfractures, aftershocks and seismicity, Bull. Seismol.
Soc. Am. 58, 11171130.
Schorlemmer, D., and M. Gerstenberger (2007). RELM Testing Center,
Seismol. Res. Letts. 78, 3036.
Schorlemmer, D., and S. Wiemer (2005). Microseismicity data forecast
rupture area, Nature 434, 10861086.
Schorlemmer, D., S. Wiemer, and M. Wyss (2005). Variations in
earthquake-size distribution across different stress regimes, Nature
437, 539542.
Schorlemmer, D., J. D. Zechar, M. Werner, D. D. Jackson, E. H. Field,
T. H. Jordan (, and The RELM Working Group). First results of
the regional earthquake likelihood models experiment, Pure Appl.
Geophys. 167, no. 89, 859876.
Shcherbakov, R., and D. L. Turcotte (2004). A damage mechanics model for
aftershocks, Pure Appl. Geophys. 161, 23792391.
Shcherbakov, R., D. L. Turcotte, and J. B. Rundle (2004). A generalized
Omoris law for earthquake aftershock decay, Geophys. Res. Lett.
31, L11613, 5 pp., doi 10.1029/2004GL019808.
Shebalin, P. (2004). Aftershocks as indicators of the state of stress in a fault
system, Dokl. Earth Sci. 398, 978982.
Shebalin, P., V. Kellis-Borok, A. Gabrielov, I. Zaliapin, and D. Turcotte
(2006). Short-term earthquake prediction by reverse analysis of
lithosphere dynamics, Tectonophysics 413, 6375.
Sibson, R. H. (1974). Frictional constraints on thrusts, wrench and normal
faults, Nature 249, 542544.
Utsu, T. (1961). A statistical study on the occurrence of aftershocks,
Geophys. Mag. 30, 521605.
Vidale, J. E., Z. Peng, and M. Ishii (2004). Anomalous aftershock decay
rates in the first hundred seconds revealed from the Hi-net borehole
data, Eos Trans. AGU 85, no. 47, Fall Meet. Suppl., Abstract
S23C-07.
Wessel, P., and W. Smith (1998). New, improved version of Generic
Mapping Tools released, Eos Trans. AGU 79, no. 47, 579.
Zechar, J. D., and T. H. Jordan (2008). Testing alarm-based earthquake
predictions, Geophys. J. Int. 172, 715724.
Zechar, J., and T. Jordan (2010). The area skill score statistic for evaluating
earthquake predictability experiments, Pure Appl. Geophys. 167,
doi 10.1007/s00024-010-0086-0.
Zechar, J. D., T. H. Jordan, D. Schorlemmer, and M. Liukis (2007).
Comparison of two earthquake predictability evaluation approaches,
Molchan error trajectory and likelihood, Seismol. Res. Lett. 78, 250.
Zechar, J. D., D. Schorlemmer, M. Liukis, J. Yu, F. Euchner, P. J. Maechling,
and T. H. Jordan (2010). The Collaboratory for the Study of Earth-
Short-Term Earthquake Forecasting Using Early Aftershock Statistics 311
quake Predictability perspective on computational earthquake science,
Concurrency Comput. Pract. Ex. 22, 18361847. doi 10.1002/
cpe.1519.
International Institute of Earthquake Prediction Theory and Mathematical
Geophysics
Moscow, 84/32 Profsouznaya
Moscow 117997, Russia
(P.S.)
Laboratoire de Dynamique des Fluides Geologiques
Institut de Physique du Globe de Paris
75252 Paris Cedex 05, France
(C.N.)
Institutes of Applied and Industrial Mathematics
Universitt Potsdam
POB 601553, 14115
Potsdam, Germany
(M.H.)
Department of Earth Sciences
University of Southern California
3651 Trousdale Parkway
Los Angeles, California 90089
(D.S.)
Manuscript received 1 May 2010
312 P. Shebalin, C. Narteau, M. Holschneider, and D. Schorlemmer