A Hybrid Ensemble Framework for Probabilistic Earthquake Forecasting in Northern California in Support of SDG 11: Sustainable and Resilient Cities

Forecasting earthquakes is still one of the most difficult problems in geophysics, mainly because seismic activity is irregular and often influenced by many factors that interact in complex ways. In this study, we develop a leakage-controlled hybrid ensemble model that combines CatBoost, LightGBM, XGBoost, and Gradient Boosting to predict five earthquake parameters: magnitude, depth, latitude, longitude, and a scaled inter-event interval in Northern California. These models were trained using USGS earthquake data ranging from 1900 to 2025 (M ≥ 4.0), with a process designed to prevent time leakage through strict time separation, a moving window feature, and prospective validation. Overall, the hybrid models produced consistently low MAE and RMSE values and very high R² values (above 0.99) for all target variables. While the estimates performed impressively, the results should be interpreted in a probabilistic context, with recognition of the inherent uncertainty of seismic processes. The framework proposed here provides a clear and replicable approach that can support the development of systems for more reliable short-term earthquake forecasting.