A Theory Driven Generalist Forecaster for Influenza: TFP Cuts 1 Week Ahead Flu MAE by 37 Percent Compared With FluSight and UMass Ensembles

This record contains a preprint on influenza forecasting using The First Pattern (TFP), a theory-driven generalist forecasting algorithm.

TFP v2.2 is a single, fixed-configuration forecaster that was tuned once for general use, not specifically for influenza. In this study it is evaluated on US state-level influenza hospitalization forecasts and compared with two operational ensemble systems used in the CDC FluSight contests: the official FluSight Ensemble and the UMass Flusion ensemble.

Using a rolling-origin design over two post-COVID flu seasons, TFP reduces 1-week-ahead mean absolute error (MAE) by about 37% relative to both ensembles (p < 0.001). At 2-week horizons, TFP retains an 8–12% MAE advantage. By 3 weeks ahead, performance is comparable to FluSight and slightly worse than UMass, which helps identify the limits of this configuration at longer horizons.

The paper includes forecast comparison tests, error decomposition, and an analysis of when and why a single model can outperform multi-model ensembles at short horizons. Key limitations include the small number of seasons, post-pandemic regime shifts in influenza dynamics, and the fact that the TFP configuration was selected based on prior cross-domain experiments before being frozen for this study.

TFP arises from a broader theory about recurring patterns in time series and is evaluated here as a candidate generalist forecaster rather than a flu-specific mechanistic model. Code for TFP is available from the author on request. Supplementary materials and related resources are linked at the end of the manuscript.