Towards earthquake forecasting in the bucaramanga nest, a machine learning approach

Abstract The Bucaramanga Seismic Nest (BSN), located in northwestern Colombia, is one of the world’s three most active volumes of concentrated intermediate-depth seismicity. Owing to its proximity to major urban centers, earthquakes with magnitudes 4. 5 ≥ 4. 5 in the BSN have repeatedly raised concern and caused infrastructure damage. Continuous monitoring has been carried out for decades, yet the prospect of anticipating the occurrence of moderate to large earthquakes in this region remains an open challenge. To address this problem, we develop a supervised learning approach that frames the task as a binary classification: a neural network issues a positive alert if at least one event with local magnitude (Ml) 4. 5 ≥ 4. 5 is likely to occur within the next six days. As inputs, the network uses daily event counts binned by magnitude over the preceding 30, 60, 180 or 365 days. The dataset comprises 148, 910 earthquakes recorded between January 1994 and February 2024 by the Servicio Geológico Colombiano. We established a robust baseline using three classical classifiers—Linear Discriminant Analysis, Logistic Regression, and Random Forests. These models provided interpretable reference points but consistently failed to achieve ROC–AUC values above 0. 60 and were strongly biased toward the majority class. In contrast, neural network models trained across multiple configurations consistently outperformed these baselines, with the best achieving ROC–AUC values >0. 60 > 0. 60, non-negligible true-positive rates, and correct positive classifications in advance of observed events. Our results demonstrate that seismicity in the Bucaramanga Nest is not entirely random but contains learnable patterns, which will serve as a basis for future research efforts aimed at seismicity forecasting in the BSN.