Foundation Models Evaluated for Epidemic Forecasting; Mixture-of-Experts Excels

Accurate short-term forecasting of seasonal influenza is a critical public health need, informing vaccination schedules, hospital staffing, and resource allocation. Despite this, the comparative performance of modern forecasting architectures on infectious disease surveillance data has not been thoroughly characterized. This research addresses that gap by systematically evaluating various models for regional influenza forecasting, using influenza-like illness surveillance and hospitalization time series data. The study assessed classical neural networks, numerical transformer-based models, pretrained time series foundation models, and LLM-based forecasting approaches for 1-4-week-ahead predictions, considering both temporal and spatial generalization. A key finding was that a mixture-of-experts model, which fuses diverse pretrained forecasters, delivered the strongest overall performance, suggesting that heterogeneous pretrained representations offer complementary predictive information. Further results showed that numerical transformer-based models provide reliable forecasts, with pretraining yielding the most substantial improvements for longer prediction horizons, especially when the pretraining domain aligns mechanistically with influenza dynamics. Conversely, LLM-based time series methods performed less effectively than numerical forecasters in this specific application. The research also explored the utility of hospitalization information as an auxiliary covariate and pretraining source, finding that it can provide complementary improvements in certain scenarios and clarify when additional surveillance streams enhance multi-horizon forecasting robustness.