מרק לסט

This study aims at predicting whether an earthquake of magnitude greater than the regional median of maximum yearly magnitudes will occur during the next year. Prediction is performed by training various machine learning algorithms, such as AdaBoost, XGBoost, Random Forest, Logistic Regression, and Info-Fuzzy Network. The models are induced using a combination of seismic indicators used in the earthquake literature as well as various time-series features, such as features based on the moving averages of the number of earthquakes in each area, features that record the number of events above and below the mean in a time period, and features based on lagged values of the mean and median magnitude. Feature selection is performed using a forward search algorithm that chooses the most effective features for prediction. The models are trained and evaluated using earthquake catalog data obtained for California, Japan, and Israel. In addition, models trained on either California or Japan datasets are evaluated using the remaining data. Models trained on Japan data achieve AUC scores up to 0.825; models trained on California data achieve AUC scores up to 0.738; and models trained on Israel data achieve AUC scores up to 0.710.