Magnitude prediction remains one of the most critical yet challenging aspects of earthquake forecasting. Accurate magnitude prediction can enhance rescue efficiency and mitigate the impact of seismic disasters. This study develops a light gradient boosting machine (LightGBM)-based prediction model by analyzing historical earthquake catalogs spanning 1900—2024, covering spatial coordinates of 5°—50°N, 55°—150°E, and magnitudes ≥ M 5.0. Through rigorous data preprocessing and feature parameter selection, the model was optimized for temporal magnitude prediction. Comparative analyses were conducted against four benchmark methods: long short-term memory, random forest, temporal graph convolutional network, which combines graph convolutional network and gated recurrent unit, and support vector regression. Experimentalresults demonstrate that the LightGBM model outperforms all alternatives in capturing magnitude time-series trends of historical earthquakes, achieving a root mean square error of 0.101, a mean absolute error of 0.100, and a coefficient of determination (R²) of 0.707, all superior to the other four models. Conclusively, the LightGBM method effectively captures recent trends in earthquake magnitude, thus providing a promising approach for the prediction of earthquake magnitude characteristics.