Systematic evaluation of ETAS forecasting skill across 69 $$M_L \ge 6$$ earthquakes in Taiwan (1999–2024)

Abstract

Operational earthquake forecasting (OEF) increasingly relies on the Epidemic-Type Aftershock Sequence (ETAS) model to provide short-term, time-dependent seismicity rates, yet whether the model performs consistently across Taiwan’s complex tectonic settings has not been systematically evaluated. This study retrospectively assesses the forecasting skill of a 2-D ETAS model across 69 $$M_L \ge 6$$ earthquakes in Taiwan between 1999 and 2024, computing forecasts at three model times (immediately before the mainshock, immediately after, and one hour later) and evaluating them against observed earthquakes with $$M_L \ge 2$$ , $$M_L \ge 3$$ , and $$M_L \ge 4$$ using the Molchan diagram. Forecasting skill improves as the target-aftershock magnitude threshold is raised, because larger-magnitude events concentrate more strongly within the high-seismicity-rate region. Pre-event forecasting skill remains limited unless preceding activity has already concentrated the seismicity rate around the impending epicenter, as exemplified by the 2022 Taitung sequence, in which the $$M_L$$ 6.6 event on 17 September produced a concentrated hotspot that encompassed the $$M_L$$ 6.8 event on 18 September. Once a mainshock has occurred, forecasting skill improves within the first hour as each newly recorded aftershock refines the high-seismicity-rate region, although the $$M_L \ge 2$$ comparison remains subject to short-term aftershock incompleteness (STAI). These results establish the ETAS model as most valuable during the evolving stage of a sequence and provide a quantitative foundation for extending the framework toward OEF and time-dependent probabilistic seismic hazard analysis (PSHA) in Taiwan.