Leakage-audited machine learning versus ETAS for earthquake forecasting in the Sea of Marmara

Basri Kerem Alhan , Kenessary Khabat

The Sea of Marmara hosts a seismic gap directly beneath a metropolitan region of some 18 million people, and the 23 April 2025 Mw 6.2 Kumburgaz earthquake renewed attention on whether short-term forecasting can add value there. We report a leakage-audited forecasting experiment built on a strictly causal, homogenized KOERI catalogue (31,329 model-box events, 2003–2026). Motivated by the finding that machine-learning models rarely beat the Epidemic-Type Aftershock Sequence (ETAS) model, we designed the benchmark around a machine-checkable causality gate — a truncated-catalogue self-test
that recomputes every feature from a catalogue truncated to the forecast time and requires exact reproduction. Within this design we (i) replaced first-generation ETAS with a conditional cascade Monte-Carlo forecaster, which ranks best on M≥3.5/30-day targets (precision–recall AUC 0.130 vs 0.126 for first -generation ETAS) and produces roughly 1.8× more offspring inside active sequences; (ii) built an ETAS×ML hybrid that by construction cannot lose to the cascade but does not beat a properly-fit ETAS in information gain (−0.38 nats per event); (iii) trained a conditional "bigger-event-ahead" discriminator on renewal-timed synthetic sequences, which ranks the escalating 2025 Sındırgı sequence far above the decaying Kumburgaz sequence; and (iv) introduced an information-arrival analysis — a general procedure that freezes the forecast at successive lead times and measures when, and through which observable, a
target event first becomes visible above the background hazard. Applied to the 2025 Marmara events, it shows that spatial localization is available far in advance (the 23 April Mw 6.2 fault cell was the top ~1% seismicity cell throughout the preceding year) while the specific timing is bounded by foreshock physics: that event's cell was quiescent, and only the lone M4.5 foreshock lifted the 30-day M≥6 probability gain to 42×, ten minutes before the mainshock. We report all negative results (machine learning ≤ ETAS on rare targets, a null short-term aftershock-incompleteness correction, no measurable gain from static GNSS strain, and a spatial-coverage miss for the out-of-box Sındırgı sequence). The system issues probability *gain*, not alarms; the live 30-day regional M≥6 probability at 2026-07-05 is a fraction of a percent, the same order as the Poisson base rate.