Shared Learning for Nationwide Municipal Dengue Incidence Forecasting in Brazil: Evidence on History, Climate, and Model Trade-Offs
- Publié
- Serveur de preprints
- Preprints.org
- DOI
- 10.20944/preprints202609.0499.v1
Municipality-level dengue forecasting is challenging because each time series is short. Additional predictors also do not always provide useful information. We evaluated one-month-ahead monthly incidence across all 5,570 Brazilian municipalities. The evaluation used 60 rolling targets and history windows of 12–48 months. We first used a 50-municipality diagnostic sample to test history, shared versus local learning, additional information, and neural models. We then conducted nationwide internal validation. In the diagnostic sample, the shared gated recurrent unit (GRU) reduced pooled MAE by 23.0%–24.5% compared with municipality-specific GRUs. Climate improved local LightGBM (MAE, 3.9%–5.1%; RMSE, 10.3%–11.5%). However, climate increased the shared GRU’s MAE. The predictive value of climate therefore depended on the model. Nationwide, the history-only shared GRU had the lowest RMSE at every lookback. An adaptive ensemble slightly improved MAE but increased RMSE.