Exploratory Analysis of Temporal Patterns in Seismic Time Series Using Translation-Invariant Machine Learning

This exploratory study examines whether a translation-invariant machine learningmethodology can identify non-random temporal patterns in seismic time series. Theframework extracts 63 spatiotemporal features from background activity (M≥3.0) usinggradient-boosted decision trees (XGBoost) to analyze signatures preceding significant events(M≥4.5). Focusing on California’s transform fault environment, strict temporal holdoutvalidation demonstrates discriminative power with AUC scores of 0.79 (regional, 500 km)and 0.85 (local, 200 km). The system achieves 75.6% coverage at 48.9 false alarms per yearin balanced mode (threshold 0.70). Retrospective applications include the 2019 Ridgecrestsequence (M7.1), flagged 14+ days in advance, and the 2026 Indio Hills event (M4.9),detected 13 days prior. A preliminary evaluation on New Zealand's Alpine Fault demonstratedsimilar performance (AUC 0.829), suggesting that similar signals may be detectable intectonically analogous settings. SHAP analysis highlights temporal acceleration and eventrecency as dominant features. This work demonstrates a research pathway for interpretablemonitoring approaches without claiming operational reliability. Keywords: time series analysis, gradient boosting, exploratory study, translation-invariant machinelearning, temporal validation, seismic monitoring, California seismicity, transform fault seismicity,interpretable AI.