<![CDATA[Indonesia experiences high seismic activity due to its location at the convergence of three major tectonic plates, making continuous monitoring of earthquake potential crucial. A fundamental parameter in seismic hazard analysis is the b-value, which reflects the stress conditions, structural heterogeneity, and magnitude distribution within the Earth’s crust. Predicting b-value fluctuations remains challenging due to its highly volatile nature. This study aims to analyze and forecast the seismic b-value in Southern Sumatra by comparing a classical statistical model, ARIMA , with two advanced machine learning architectures, LSTM and GRU. Historical earthquake catalogs from BMKG and NEIC-USGS spanning 1960–2025 were utilized. The data underwent a declustering process using the Reasenberg method to eliminate foreshocks and aftershocks, yielding 15,844 independent events. The monthly b-value was then calculated using Maximum Likelihood Estimation. Furthermore, PSO was applied to tune the hyperparameters of the deep learning models. Evaluation reveals that ARIMA yields the highest predictive accuracy, achieving a Mean Absolute Error (MAE) of 0.11, Root Mean Square Error (RMSE) of 0.16, and Mean Absolute Percentage Error (MAPE) of 13.85%. In contrast, GRU (MAPE 14.30%) and LSTM (MAPE 16.22%) produced smoother predictions but struggled to capture extreme short-term fluctuations. The findings conclude that for volatile and limited-size time series data like regional b-values, the linear approach of ARIMA remains significantly more effective than complex deep learning models.]]>
