About

Background. Seasonal influenza causes substantial morbidity and mortality worldwide, and timely activity forecasts support preparedness and resource planning. Reliable forecasting models exist for only a few well-resourced countries, requiring disease-specific design and local expertise; most countries have no operational forecasts. We assessed whether a time-series foundation model could be fine-tuned into a single influenza forecaster for many countries and compared it with purpose-built national models.

Methods. We fine-tuned Chronos-2, a pretrained time-series foundation model, on weekly influenza-like illness and acute respiratory infection records for 78 countries, selected from the WHO FluID global epidemiological surveillance database using predefined data-quality criteria. The model produced probabilistic four-week-ahead forecasts. We benchmarked it for the 2025/26 season against the models of Influcast, Italy's collaborative forecasting hub, using the pairwise relative weighted interval score (rWIS), where values below one beat the naive baseline. This season lay outside the fine-tuning data and post-dated Chronos-2's release, ensuring an out-of-sample test. We also evaluated Chronos-2 without fine-tuning (zero-shot).

Results. Across 16 weekly forecasting rounds, the fine-tuned model ranked second of all entries, with a pairwise rWIS of 0.71, ahead of the Influcast ensemble (0.74) and eight of nine participating models, behind only one (0.70). The zero-shot model achieved an rWIS of 1.01, no better than the naive baseline; fine-tuning reduced the rWIS by 30% and mean absolute percentage error from 23% to 17%. Prediction intervals were over-confident: 50% intervals covered 24% of observed values.

Conclusions. A fine-tuned foundation model produced influenza forecasts that matched the purpose-built models of the Italian forecasting hub, whereas the off-the-shelf model performed no better than a naive baseline. Because one model forecasts across 78 countries, the approach could extend influenza forecasting to countries lacking dedicated modelling capacity, supporting preparedness and resource allocation. Benchmarking against forecasting models in other countries is needed to confirm whether this generalises.