Understanding Key Features of Time Series Foundation Models from Epidemic Forecasting
A new research paper evaluates various time series forecasting models for predicting seasonal influenza in the United States. The study found that a mixture-of-experts model, which combines multiple pretrained forecasters, achieved the best performance. Transformer-based models also showed reliability, with pretraining offering the most significant benefits for longer-term predictions, especially when aligned with influenza dynamics. Large language model (LLM) based time series methods, however, underperformed compared to numerical forecasters in this specific application. AI
IMPACT Provides guidance on selecting and applying advanced time series models for public health forecasting.