A retrospective comparative forecast test on the 1992 Landers sequence

We perform a retrospective forecast experiment on the 1992 Landers sequence comparing the predictive power of commonly used model frameworks for short-term earthquake forecasting. We compare a modified short-term earthquake probability (STEP) model, six realizations of the epidemic-type aftershock sequence (ETAS) model, and four models that combine Coulomb stress changes calculations and rate-and-state theory to generate seismicity rates (CRS models). We perform the experiment under the premise of a controlled environment with predefined conditions for the testing region and data for all modelers. We evaluate the forecasts with likelihood tests to analyze spatial consistency and the total amount of forecasted events versus observed data. We find that (1) 9 of the 11 models perform superior compared to a simple reference model, (2) ETAS models forecast the spatial evolution of seismicity best and perform best in the entire test suite, (3) the modified STEP model matches best the total number of events, (4) CRS models can only compete with empirical statistical models by introducing stochasticity in these models considering uncertainties in the finite-fault source model, and (5) resolving Coulomb stress changes on 3-D optimally oriented planes is more adequate for forecasting purposes than using the specified receiver fault concept. We conclude that statistical models perform generally better than the tested physics-based models and parameter value updates using the occurrence of aftershocks generally improve the predictive power in particular for the purely statistical models in space and time.