Time-series forecasting of mRDT-positive malaria counts in Tanzania’s lake zone: a comparative evaluation of statistical and machine-learning models using routine surveillance data, 2018–2025

Monthly malaria forecasts can support diagnostic and treatment planning when the surveillance outcome and evaluation protocol are clearly defined. We analysed 4,128 council-month records from 43 councils in six Lake Zone regions during January 2018–December 2025. Complete mRDT-positive counts were aggregated to 576 region-month observations and one 96-month Lake Zone series. Six fitted models, an equal-weight ensemble and a seasonal-naive benchmark were evaluated on a fixed-origin 2025 holdout and 37 expanding-window origins with 1–12-month horizons. The dataset contained 11,874,519 mRDT-positive counts. Annual counts declined from 1.63 million in 2018 to 1.06 million in 2025, with the highest calendar-month averages in June–July. ARIMA (2, 1, 2) had the lowest 2025 holdout error (MAPE 12.1%; RMSE 11,709), reducing RMSE by 69.2% relative to seasonal naive. Across all rolling origin–horizon pairs, Gradient Boosting had the lowest RMSE (22,007), while the ensemble had the lowest MAE (16,687) and MAPE (16.1%). Model rankings varied with forecast origin, horizon and metric. The findings favour routinely updated, benchmarked forecasts over reliance on a single terminal-year ranking. The outcome represents positive diagnostic counts rather than population-standardized incidence or confirmed unique outpatient episodes.