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

Basri Kerem Alhan , Kenessary Khabat

The Sea of Marmara’s locked central fault lies beneath ~18 million people; the 2025 Mw 6.2 Kumburgaz earthquake renewed forecasting interest. We present a leakage-audited benchmark on a homogenized, strictly causal KOERI catalogue: causality is machine-checkable (a truncated-catalogue self-test), and every comparison is adjudicated by a pre-specified block bootstrap requiring both paired intervals to exclude zero. Four physics forecasters (first-generation, cascade Monte-Carlo, spatially-variable-background, and an independent ETAS inversion) form a stable cluster whose verdicts never move: the first three are mutually inseparable, and first-generation beats the inversion. The ETAS×ML hybrid has no robust standing against them: re-running across a narrow range of one calibration constant swings its verdict versus physics between inseparable and beaten, so the machine-learning stage adds no reliable value over a well-fit ETAS. Two engineered channels (GNSS, dense sub-Mc3 catalogue) gave interval-significant gains a placebo battery voided. The Mw 6.2 cell ranked top-1% all year, but temporal information arrived only with its ML 4.0 foreshock 36 minutes before rupture. CSEP tests show the top-ranked forecasters over-predict counts at the honest calibration; the live 30-day regional M≥6 probability is ~1%.