temporal decay

How forecast skill changes with how much of the sequence a model has seen, and with how far ahead it looks. The zero line is the Omori–Utsu law refitted at every origin.

skill by forecast origin

Pooled across sequences and horizons. The hypothesis that learned models help most early — when a single sequence offers little data and other sequences can lend it — is read off the left end of this chart.

  • ETAS (fixed μ, subcritical)
  • Omori–Utsu + NB spread
  • GBM-hybrid · cross/pooled
-0.100.000.100.200.300.401101001,000forecast origin (days since mainshock)nats per event vs Omori
Table view
forecast origin (days since mainshock)ETAS (fixed μ, subcritical)Omori–Utsu + NB spreadGBM-hybrid · cross/pooled
10.360.00-0.05
20.210.00-0.03
30.270.000.06
50.050.000.01
70.050.060.01
100.070.080.00
140.030.03-0.02
210.010.02-0.01
300.010.02-0.03
450.010.020.01
600.110.120.02
90-0.000.020.01
1200.010.030.03
1800.150.15-0.03
2400.020.060.01

skill by horizon

The same pooling, split by forecast length.

  • ETAS (fixed μ, subcritical)
  • Omori–Utsu + NB spread
  • GBM-hybrid · cross/pooled
-0.050.000.050.100.150.200.25125102050100forecast horizon (days)nats per event vs Omori
Table view
forecast horizon (days)ETAS (fixed μ, subcritical)Omori–Utsu + NB spreadGBM-hybrid · cross/pooled
10.020.01-0.02
70.100.01-0.00
300.220.02-0.01

decay exponents

Omori p and ETAS parameters as refitted at three origins. ETAS values on a search bound are marked with an asterisk; see failures.

sequenceOmori p · 7 d· 30 d· 120 dETAS α · 30 dETAS p · 30 dETAS n₃₆₅ · 30 dETAS α · 120 dETAS n₃₆₅ · 120 d
Landers 19925.001.631.441.161.370.991.350.96
Hector Mine 19990.891.341.242.341.150.602.380.55
Ridgecrest 20190.961.041.222.441.100.992.460.95
Northridge 19941.311.271.282.331.180.952.300.79
El Mayor–Cucapah 20101.131.171.032.431.090.832.400.92
Loma Prieta 19891.371.040.932.031.010.571.990.65
San Simeon 20030.851.221.002.131.000.992.060.90
South Napa 20140.830.790.631.660.810.581.540.67
Parkfield 20040.820.790.664.190.810.151.130.91
L’Aquila 20090.921.221.181.211.490.901.440.85
Amatrice 20161.000.851.301.280.98
Norcia 20161.010.940.792.441.080.632.420.61
Darfield 20105.001.061.080.881.670.991.500.99
Kaikōura 20162.181.411.241.831.310.572.150.44

a year from one month

One forecast issued at day 30 for the rest of the window. Omori extrapolates a curve; ETAS simulates a cascade, and near-critical cascades grow long tails.

sequenceobservedOmori meanOmori 90%ETAS meanETAS 90%free ETAS mean
Landers 19921,1767937478409,3814,01016,57633,874
Hector Mine 1999285224199248729402,156597
Ridgecrest 20191,3931,8331,7631,90414,5015555,80112,426
Northridge 19942542241992483,486018,5112,876
El Mayor–Cucapah 20103892782513061,58556,1721,544
Loma Prieta 19892,1381,2681,2101,3274,1731,6107,7052,820
San Simeon 20038476015616424,58616214,4277,722
South Napa 20144222592332868485321,224945
Parkfield 20041,8569609101,0121,3477792,0421,311
L’Aquila 20095846355946771,7441,0482,5835,497
Amatrice 20167268387918861,5259742,1761,462
Norcia 20161,8411,4291,3671,4922,3418564,4174,930
Darfield 20103,1221,0761,0231,1314,8061,05710,76611,868
Kaikōura 201623315613617738395825336