Foundation models capture human and environmental context to improve dengue forecasting

Accurate and timely probabilistic dengue forecasts can support public health preparedness, but producing reliable forecasts across locations and time horizons remains challenging. We evaluated five statistical, machine-learning, deep-learning, and foundation-model approaches for forecasting monthly reported dengue cases across 5,570 Brazilian municipalities at 1-, 3-, 6- and 12-month horizons, with forecasts backtested from January 2020 to August 2025. The time-series foundation model TimesFM maintained comparatively stable probabilistic forecast performance at longer horizons, while most other models showed declining performance. We then tested whether spatial representations of human and environmental context from a Population Dynamics Foundation Model (PDFM) could improve forecasts through post-hoc residual adjustment. PDFM adjustment improved mean probabilistic forecast performance, measured by the Weighted Interval Score (WIS), in 19 of 20 model-horizon combinations, although the magnitude and frequency of improvements varied across models, horizons and municipalities. The largest improvements were observed for N-BEATS, including a reduction in mean WIS of 1.13 at the six-month horizon (95% CI [0.84, 1.48]; Diebold-Mariano p < 0.0001), with improvements also observed for XGBoost. The proportion of forecasts improved by PDFM generally increased with locally reported dengue case counts, although this relationship varied between models. Covariate selection, spatial fitting and forecast aggregation strategies showed strong model-dependent effects, while monthly model retraining provided no consistent advantage in the settings evaluated. Together, these findings highlight the potential of pretrained representations for dengue forecasting, with benefits dependent on forecasting architecture, horizon and modelling configuration.