Do climate covariates improve national-scale dengue forecasting beyond autoregressive and seasonal structure? A 15-year time-series analysis from Bangladesh

Objective: To evaluate whether climate covariates materially improve national-scale dengue forecasting beyond autoregressive and seasonal structure under strict rolling-origin validation. Results: I analyzed 15 years (2010–2025) of national monthly dengue surveillance data from Bangladesh combined with climate variables (temperature, rainfall, humidity). Autoregressive regression and SARIMA models were evaluated using rolling-origin validation to simulate real-time forecasting. Strong autoregressive dependence (Lag-1 ≈ 0.95) and consistent seasonal structure were observed. Autoregressive frameworks substantially outperformed naive persistence under rolling-origin validation. The Random Forest benchmark did not demonstrate performance gains relative to autoregressive models. Classical SARIMA achieved the strongest predictive performance (RMSE = 0.736; MAE = 0.559). Under national monthly aggregation, lagged climate covariates did not demonstrate statistically significant incremental predictive contribution beyond autoregressive and seasonal components. Forecast interval calibration showed 94% empirical coverage for nominal 95% intervals. A structural deviation was observed in 2020, likely reflecting surveillance disruption. These findings suggest that national-scale dengue transmission dynamics in Bangladesh are strongly characterized by autoregressive persistence and seasonal structure.