INTRODUCTION:
Malaria remains a leading cause of under-five mortalities in the Democratic Republic of Congo (DRC). Understanding its spatiotemporal dynamics and forecasting trends are essential for precision control and resource targeting. This study described provincial patterns of under-five malaria deaths and developed forecasts using classical and neural network enhanced time series models.

METHODS:
Monthly under five malaria deaths (January 2020–December 2024) were harmonized for 26 provinces. Descriptive mapping and time series diagnostics assessed variability and seasonality. Eight forecasting methods (ARIMA, ETS, BATS, TBATS, Theta, NNAR, ARFIMA, and ANN MLP) were trained on 2020–2023 data and validated on 2024 using sMAPE. The best model for each province was refitted on full data to project January 2025–December 2026. Spatial autocorrelation was tested with queen contiguity Local Moran’s I and Getis Ord Gi*.

RESULTS:
Average monthly deaths were 53.2 (SD 36.9), ranging from 9 to 312 across provinces. ARFIMA, BATS, and Theta were most often optimal; Sankuru (ARIMA, sMAPE = 0.106) and Tanganyika (ARFIMA, sMAPE = 0.110) achieved the highest accuracy. ANN MLP failed in several provinces due to data limitations. Forecasts for 2025–2026 retained historical peaks and highlighted consistently high burden regions. Spatial analysis revealed only localized clustering.

CONCLUSIONS:
Integrating diverse forecasting models with spatial analysis produced robust province specific predictions. Results emphasize persistent regional disparities and the need for geographically focused malaria control. Future improvements should combine climatic and socioeconomic covariates and explore hybrid ARIMA deep learning models to enhance operational forecasting capacity.