failures

Where the models — classical and learned — go wrong, and where the benchmark’s own design choices are fragile. Nothing on this page is filtered for flattery.

fits on their bounds

Origins at which a fitted decay model finished on its search bound. Early in a sequence the catalog is still incomplete, the observed rate looks flat, and the likelihood does not identify the decay parameters — for Omori–Utsu as much as for ETAS. Forecasts from those fits can still be serviceable; the parameters are not.

sequenceoriginsOmori on a boundprimary ETAS on a boundparametersfree ETAS on a boundfree ETAS supercritical
Landers 19921543p, alpha08
Hector Mine 1999151021
Ridgecrest 2019150012
Northridge 1994151011
El Mayor–Cucapah 20101522p30
Loma Prieta 1989150001
San Simeon 20031511p16
South Napa 2014155000
Parkfield 20041501alpha30
L’Aquila 2009151010
Amatrice 2016111001
Norcia 2016150000
Darfield 20101535c, p77
Kaikōura 20161521p10

overconfidence

Share of windows where the observed count fell outside the 90% interval, and where the number test rejected the forecast, averaged across sequences. Poisson intervals are too narrow for clustered counts; the “+ NB spread” variants fit a negative-binomial spread to each model’s own closed forecasts.

modeloutside 90% intervalnumber-test rejections
Omori–Utsu44%39%
Omori–Utsu + NB spread6%3%
ETAS (fixed μ, subcritical)5%2%
ETAS (free)10%8%
Persistence64%59%
GBM-hybrid · within51%43%
GBM-hybrid · within + NB24%20%
GBM-hybrid · cross/pooled45%39%
GBM-hybrid · cross/pooled + NB7%3%

the worst windows

The fifteen largest errors in log₁₀(count + 1) across all sequences for Omori, ETAS and the learned hybrids.

sequencemodelorigin (d)horizon (d)observedforecast meanlog₁₀ error
San Simeon 2003GBM-hybrid · cross/pooled1301,17855.7−1.32
San Simeon 2003Omori–Utsu1301,178129.2−0.96
El Mayor–Cucapah 2010ETAS (fixed μ, subcritical)2407225.7+0.95
El Mayor–Cucapah 2010ETAS (fixed μ, subcritical)240107.5+0.93
Ridgecrest 2019ETAS (fixed μ, subcritical)60301741,471.9+0.93
San Simeon 2003GBM-hybrid · cross/pooled1768183.1−0.91
Ridgecrest 2019ETAS (fixed μ, subcritical)9030125991.7+0.90
San Simeon 2003ETAS (fixed μ, subcritical)1301,178159.4−0.87
Hector Mine 1999ETAS (fixed μ, subcritical)240705.9+0.84
Ridgecrest 2019ETAS (fixed μ, subcritical)45302771,910.9+0.84
Kaikōura 2016GBM-hybrid · cross/pooled13058389.4−0.81
Darfield 2010GBM-hybrid · within103064199.7−0.80
Ridgecrest 2019ETAS (fixed μ, subcritical)21305513,374.5+0.79
Darfield 2010GBM-hybrid · cross/pooled1301,3988,060.1+0.76
Northridge 1994ETAS (fixed μ, subcritical)213091525.5+0.76

design choices under pressure

Each robustness variant changes one choice — completeness cutoff, window start, zone radius or rule, window length, truncation, foreshock history — and reruns the benchmark at lighter settings.

sequencevariantMceventsETAS vs OmoriGBM-hybrid vs Omoriwindows
Amatrice 2016mc+0.32.11,652+0.001−0.0157
Amatrice 2016mc+0.62.4730+0.003−0.0187
Amatrice 2016no_lookback1.83,912+0.030−0.1497
Amatrice 2016primary_light1.83,912+0.030−0.1497
Amatrice 2016r0.51.83,845+0.035−0.0987
Amatrice 2016r1.51.83,922+0.030−0.1277
Amatrice 2016r2.01.83,936+0.030−0.0857
Amatrice 2016t0=11.65,883+0.031−0.2127
Amatrice 2016t0=21.83,065+0.029−0.1134
Amatrice 2016t_end=1801.83,912+0.030−0.1497
Amatrice 2016wc_radius1.82,616+0.040+0.0018
Darfield 2010mc+0.33.12,441−0.011−0.12815
Darfield 2010mc+0.63.41,242+0.005−0.08815
Darfield 2010no_lookback2.84,641+0.006−0.10215
Darfield 2010no_truncation2.84,641+0.005−0.10215
Darfield 2010primary_light2.84,641+0.005−0.10215
Darfield 2010r0.52.82,945−0.005−0.11615
Darfield 2010r1.52.84,712+0.004−0.08715
Darfield 2010r2.02.84,769+0.004−0.10015
Darfield 2010t0=12.84,507+0.004−0.09315
Darfield 2010t0=22.84,295+0.011−0.08712
Darfield 2010t_end=1802.83,217−0.002−0.11412
Darfield 2010wc_radius2.84,645+0.006−0.09315
Kaikōura 2016primary_light3.2941−0.029−0.02315
Kaikōura 2016r1.53.31,021−0.046−0.01415
Kaikōura 2016r2.03.31,544−0.023−0.00115
Kaikōura 2016wc_radius3.4802−0.048−0.02715
Landers 1992mc+0.32.81,644+0.011+0.01315
Landers 1992mc+0.63.1680+0.047+0.03815
Landers 1992no_lookback2.53,679+0.003−0.03915
Landers 1992no_truncation2.53,679+0.002−0.03915
Landers 1992primary_light2.53,679+0.002−0.03915
Landers 1992r0.52.52,353−0.019−0.02315
Landers 1992r1.52.54,403+0.005+0.01015
Landers 1992r2.02.54,487+0.011−0.01115
Landers 1992t0=12.53,465+0.002+0.01715
Landers 1992t0=22.53,153−0.012+0.00612
Landers 1992t_end=1803.0844+0.021+0.03112
Landers 1992wc_radius2.53,795+0.009−0.02315
Parkfield 2004mc+0.31.51,408+0.059+0.02915
Parkfield 2004mc+0.61.8760+0.060−0.02615
Parkfield 2004no_lookback1.22,666+0.034−0.00215
Parkfield 2004no_truncation1.22,666+0.034−0.00215
Parkfield 2004primary_light1.22,666+0.034−0.00215
Parkfield 2004r0.51.5573+0.009+0.00215
Parkfield 2004r1.51.24,485+0.046+0.01115
Parkfield 2004r2.01.24,805+0.044+0.00115
Parkfield 2004t0=11.22,566+0.041+0.00915
Parkfield 2004t0=21.22,483+0.048+0.02812
Parkfield 2004t_end=1801.21,885+0.021−0.00712
Parkfield 2004wc_radius1.2590+0.053+0.03115