Pretrained models show promise for infectious disease forecasting across six countries

A retrospective study in BMC Public Health compared pretrained time series foundation models with disease-specific neural forecasting approaches for infectious disease surveillance. Researchers assembled 289 disease–country series comprising 106,234 weekly or monthly observations from China, Singapore, Japan, South Korea, Ireland and the United States. Four foundation models—TimesFM, Moirai, Chronos and TimeMoE—were tested without task-specific training against five neural network baselines across short-, medium- and long-term forecasting horizons.

A foundation-model variant achieved the lowest observed error in 30 of 36 country–horizon–metric settings (83.3%), assessed using mean absolute error and root mean squared error. At the one-step-ahead horizon, foundation-model variants ranked first in every country on both metrics. Larger models did not consistently produce smaller errors, and prediction-interval performance varied by country and horizon. An exploratory analysis suggested potential gains from jointly forecasting co-circulating respiratory pathogens.

These findings support further evaluation of pretrained models for surveillance pipelines, including smaller variants suitable for modest hardware. However, the rankings were descriptive rather than evidence of statistical superiority, and comparator models used fixed default configurations. Each series had only one held-out evaluation year, limiting conclusions about reliability across years, epidemic cycles or changing surveillance conditions. The study received support from the Natural Science Foundation of Chongqing; the authors declared no competing interests.