Benchmarking time series foundation models for multi-country infectious disease forecasting: a retrospective evaluation across over 70 diseases

Timely forecasting of infectious disease activity supports outbreak alerts, vaccination scheduling and resource allocation in public health surveillance. Modern deep-learning models can be accurate but typically require disease-specific training data and tuning, conditions often unavailable for rare or newly notified pathogens or in settings with limited computational resources. Time series foundation models (TSFMs), pre-trained on large heterogeneous corpora and deployable in a zero-shot manner, offer a candidate solution, but their value for routine multi-disease surveillance has not been systematically assessed. We assembled 289 disease–country time series across over 70 notifiable infectious diseases from six national surveillance systems (China, Singapore, Japan, South Korea, Ireland and the United States; 106,234 weekly or monthly observations). Four TSFMs (TimesFM, Moirai, Chronos and TimeMoE; 14–314 million parameters) were evaluated in zero-shot mode and compared with five neural network baselines (DLinear, LSTM, PatchTST, TimeMixer and TimesNet) across one-step-ahead, medium-term and long-term horizons. Forecasts were assessed using MAE, RMSE, MASE and R²; prediction intervals from Moirai variants were additionally compared using the Winkler score. Across 36 settings (6 countries × 3 horizons × 2 metrics, MAE and RMSE), a TSFM variant achieved the lowest observed error in 30 (83.3%). This is a descriptive ranking summary rather than evidence of statistical superiority. The pattern was most consistent at the one-step-ahead horizon, where a TSFM variant ranked first in every country on both metrics, and TSFM variants also ranked highest on MASE and R². Errors did not decrease monotonically with model size in Moirai, Chronos or TimeMoE, and the Moirai variant with the lowest Winkler score varied by country and forecasting horizon. An exploratory multivariate analysis using Moirai indicated that jointly forecasting co-circulating respiratory pathogens may yield additional performance gains. Within the evaluated test periods of this multi-country study, zero-shot TSFMs were competitive with task-specific neural forecasting models trained with fixed default configurations. Because each series was evaluated over a single held-out year, these results do not establish consistent performance across years, epidemic cycles or changing surveillance conditions. Pre-trained TSFMs, particularly small variants that run on modest hardware, are nonetheless a promising starting point for surveillance forecasting pipelines.