Accurately predicting the timing, location, and magnitude of earthquakes is a major global scientific challenge. While machine learning methods have been widely used in seismic time series analysis and outperform traditional statistical models, their predictive capability for strong earthquakes (Mw ≥ 6. 5) remains limited. This study develops a new prediction method by constructing multi-timescale seismic activity features and integrating them with machine learning models. The novelty of this approach is a parameterized framework based on a historical seismic window (Tₕistory), forecast window (Tforecast), and magnitude threshold (Mₜhreshold). It transforms earthquake prediction into an optimization problem, using machine learning to model the relationship between these parameters and performance metrics (e. g. , AUC (Area Under the (ROC) Curve) ), and searches the parameter space to maximize AUC. To improve spatial prediction, clustering algorithms divide the study area into subregions with consistent seismic characteristics, and an optimization algorithm determines the optimal time-magnitude parameters for each subregion. Using earthquake catalogs from Japan and New Zealand (1973–2024), K-means + + was selected to partition Japan into 4 and New Zealand into 3 subregions. With Tforecast set to 365 days and time-series cross-validation applied to reduce overfitting, the method achieved average AUC values of 0. 75 and 0. 81 for predicting strong earthquakes within one year in Japan and New Zealand, respectively, outperforming existing models. The main contribution is a flexible and extensible prediction framework that overcomes the limitations of fixed regions, time windows, and magnitude thresholds. Future work will focus on extensible improvements, overfitting control, and integrating multi-source features to enhance generalization.
Wang et al. (Fri,) studied this question.