Comparative evaluation of physics‐based and statistical forecasts in Northern California

Abstract

AbstractWe perform a retrospective forecast test using Northern California seismicity for the period between 1980 and 2009. We compare 7 realizations of the short‐term clustering epidemic‐type aftershock sequence (ETAS) model, and 21 models combining Coulomb stress change calculations and Rate/State theory (CRS) to forecast seismicity rates in 10 day time intervals. We employ a common learning phase (1974–1980) for CRS models to ensure consistency, and we evaluate the forecasts with log likelihood statistics to detect any spatial inconsistencies and compare the total numbers of forecasts versus observed events. We find that: (1) ETAS models are better forecasters of the spatial evolution in seismicity in the near‐source region, (2) CRS models can compete with ETAS models away from the mainshock rupture, and for short periods after mainshocks, (3) CRS models with optimally oriented receiver fault planes perform better in the first few days after mainshocks, whereas mapped fault planes should be implemented for longer‐term forecasting, and (4) CRS models based on shear stress change calculations have comparable performance with Coulomb stress change models, with the benefit of lesser parameters involved in stress calculations. We conclude that physics‐based and statistical forecast models are complimentary to each other and that future forecasts should be based on statistical models for near‐source regions, and physical models for longer periods and distances. However, the realization of the CRS models involves a number of critical parameters (reference seismicity rates, regional stress field, and loading rates), which should be retrospectively tested to improve the predictive power of physics‐based models.