Foundation models capture human and environmental context to improve dengue forecasting

TimesFM, a time-series foundation model, maintained comparatively stable probabilistic dengue forecast performance at longer horizons across 5,570 Brazilian municipalities (1–12 months).

• Why it matters: Reliable probabilistic dengue forecasts are needed for preparedness, yet performance degrades across locations and time horizons, limiting dependable planning.
• What they did: The study backtested five model classes and TimesFM on monthly reported dengue cases from January 2020 to August 2025, then applied post-hoc residual adjustment using PDFM spatial context.
• The result: PDFM adjustment improved mean forecast quality (WIS) in 19 of 20 model–horizon combinations, with the largest gain for N-BEATS at 6 months: mean WIS reduced by 1.13 (95% CI [0.84, 1.48]; Diebold-Mariano p < 0.0001).

AI summary of the abstract · may contain errors · not medical advice · read the paper.