Forecasting cold wave in Bangladesh: a validated machine learning approach for early warning and vulnerability reduction

Background Cold waves are significant health hazards in northern Bangladesh, yet validated forecasting tools to support early warning remain limited. This study develops and validates machine learning models for predicting daily minimum temperatures and cold wave events in Mymensingh district. Methods Daily minimum temperature data (1985–2022) were analyzed (38 years). ARIMA and ETS, as well as machine learning models (SVR, RF, LSTM), and hybrid models were tested by means of 5-fold Time Series Cross-Validation. The most effective LSTM model (32 units, 1 layer) was chosen to be used in 2027 forecasting with prediction intervals. Results LSTM performed better than any other model, and the test RMSE of 1.395 °C, MAE of 1.053 °C, and MASE of 0.906 were obtained. Statistical models were not very good (RMSE > 6.7 °C). Hybrid models were not statistically significant compared to standalone LSTM. To detect cold waves (threshold = less than 10 °C), LSTM showed a high level of discriminative power (ROC-AUC = 0.975) but low recall (0.215), detecting only 14 out of 65 actual cold days. The 2027 projection showed no cold waves, and all the temperatures were above threshold. Findings are specific to the Mymensingh station; generalization to other regions requires further validation. Conclusions LSTM is effective for temperature prediction, but it’s very low recall (0.215) severely limits any direct application for operational early warning. These findings support operational guidelines that integrate model outputs with careful interpretation during winter months. Future work should incorporate multiple stations and health outcome data to enhance predictive power for vulnerability reduction.