Machine learning versus ETAS for earthquake forecasting in the Sea of Marmara: a leakage-audited negative result and a closed-form scoring artifact

Basri Kerem Alhan , Kenessary Khabat

We ask whether machine learning adds 30-day earthquake-forecasting value beyond modern ETAS in the Sea of Marmara, a locked seismic gap beneath Istanbul, once leakage and scoring artifacts are controlled. On a homogenized KOERI catalog, feature causality is machine-checkable, and comparisons follow a pre-specified rule: paired block-bootstrap intervals on likelihood and ranking must agree. The ETAS×ML hybrid adds no ranking skill over a well-fit ETAS cascade. Its likelihood edge is inseparable from a scoring artifact with a closed form: count-scoring a clustered forecast against binary occurrence makes a(h) = h−1−ln h nats per positive recoverable by pure rescaling. Under a proper binary score, a small occurrence-recalibration edge survives, resolved in active cells and absent at M≥3.5; an ablation shows catalog features cannot reconstruct ETAS, and ranking information lives on the ETAS axis alone. The 2025 Mw 6.2 epicentral cell ranked in the top ~1–2% all year; the only escalation was a foreshock 36 minutes before rupture, and ~220 comparable alarms were followed by no M≥6.