Abstract
This study examines the application of machine learning models for short-term earthquake forecasting in Almaty, Kazakhstan, a region known for its high seismic activity. Using catalog data from the USGS, both regression and classification models were developed to predict earthquake magnitudes and exceedance of a 4.5 Mw threshold. A range of algorithms, including Gradient Boosting, Random Forest, and Support Vector Machines, were evaluated and optimized via grid and Bayesian search strategies. Gradient Boosting achieved the best regression performance with a mean squared error of 0.13, while Random Forest yielded the highest classification F1 score of 0.69. The results demonstrate that careful feature engineering and hyperparameter tuning significantly improve model performance. These findings suggest that data-driven models offer a promising complement to traditional forecasting methods in seismic risk mitigation.