Conformal prediction for earthquake early warning: finite-sample coverage guarantees for prediction intervals of peak ground acceleration

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Earthquake early warning requires reliable uncertainty estimates from short initial waveforms. We evaluate split conformal prediction (CP) and conformalized quantile regression (CQR) for single-station peak ground acceleration using event-partitioned K-NET records. Each 3 s waveform is normalized by component while retaining amplitude scales measured before normalization. The convolutional neural network (CNN)-Transformer that retains amplitude information is more accurate and yields narrower intervals than the waveform-only ablation. At the nominal 90 per cent level, CQR attains 90.03 per cent coverage; raw quantile regression and split CP are also near nominal, whereas heteroscedastic deep ensemble and MC-dropout baselines are more conservative and wider. Event-cluster bootstrap and repeated event-level resplits support marginal-coverage stability, although performance degrades for rare large earthquakes. Upper-bound screening trades increased recall for more false alarms, so the intervals are an uncertainty-quantification and screening layer rather than a standalone alert system. Under exact exchangeability, conformal prediction supplies finite-sample marginal coverage; record-level exchangeability is not established for this event-clustered sample and event-conditional reliability is not guaranteed.

키워드

Machine learning; Probability distributions; Computational seismology; Earthquake early warning; Statistical seismology
제목
Conformal prediction for earthquake early warning: finite-sample coverage guarantees for prediction intervals of peak ground acceleration
저자
Shin, Youngjae; Jeong, Woodon
DOI
10.1093/gji/ggag415
발행일
2026-12
유형
Article
저널명
Geophysical Journal International
권
247
호
3