Development of a Long Short-Term Memory (LSTM)-Based Statistical Model for Earthquake Forecasting in Central Asia

Earthquake forecasting using traditional methods remains a complex task due to the inherent nonlinearity and stochastic nature of seismic activity. Therefore, this study examines the application of deep learning methods, particularly a hybrid model combining Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), to create a robust prediction model that outperforms traditional methods. The study comprises six primary phases: data acquisition, pre-processing, exploratory analysis, seismicity characterization, model implementation, and effectiveness assessment. The dataset contained 17,565 earthquake events (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">M</i> ≥ 3.0) in Central Asia, collected from 1973–2024 via the USGS API, enriched with geological features such as distance to active faults (from GEM datasets) and eight key seismicity indicators. Compared with baseline architectures (standard LSTM and fully connected neural network), the proposed CNN-LSTM hybrid achieved superior performance, particularly in recall (0.80 compared with 0.67) and F1-score (0.77 compared with 0.71), demonstrating its ability to capture both spatial and temporal dependencies. The results indicate that using a hybrid CNN-LSTM model leads to a significant improvement in forecasting strong earthquakes. This is validated by ROC-AUC of 0.78 and F1-score of 0.77, supporting a good balance between model completeness and accuracy. These improvements confirm that the regional adaptation of CNN-LSTM, combined with seismic indicators and clustering-based subregionalization, provides a novel methodological contribution to earthquake forecasting in Central Asia. The model reliably distinguishes periods of elevated seismic risk from quiet intervals, supporting practical applications in regional forecasting and seismic risk visualization.