Methodology

The model is trained on weekly influenza-like illness (ILI) and acute respiratory infection (ARI) case counts from the WHO's FluID / FluMart global surveillance database. Countries report either ILI (a clinical syndrome) or ARI (a broader respiratory category); countries that switched indicator mid-series (e.g. Italy, ILI to ARI in late 2025) are kept as a single continuous series with a marker for which indicator each week belongs to, so the model can treat the regimes differently.

Not every country's series is usable. Candidates were screened with fixed, predefined quality criteria, and 78 countries passed:

• at least 156 non-missing weeks (≈ 3 years of actual data)
• spanning at least 4 years of coverage
• at most 50% of weeks missing for irregularly reporting series
• median non-zero weekly count of at least 20 (excludes micro-series with near-zero counts)

Selection criteria are applied identically to all countries and are deliberately data-driven rather than judgment-based, to avoid cherry-picking easy-to-predict series.

The forecaster is Chronos-2, a pretrained time-series foundation model (patch-based transformer trained with a quantile-regression loss, so its outputs are probabilistic by construction). It is fine-tuned on the 78-country series set; the production configuration conditions on two covariates, the indicator type (ILI vs ARI) and hemisphere, selected by ablation. Additional covariates (week-of-year, historical weather from Open-Meteo, neighbouring-country activity) were evaluated and did not improve out-of-sample scores, so they are not used.

The design goal is one model, many countries: no disease-specific model, no local calibration loop, and no country-specific hyperparameters. Countries with sparse or low-quality series are simply not in the training set rather than given special treatment.

Every week, the model is given each country's historical series up to the most recent complete week and produces a probabilistic forecast for the next four weeks. The dashboard displays the central estimate on the map and 50%, 90% and 95% prediction intervals on the country pages. The pipeline updates automatically when the WHO publishes new weekly data.

Forecasts are scored with the weighted interval score (WIS), a proper scoring rule for interval forecasts that rewards accurate central estimates and appropriately wide intervals. Because WIS is hard to interpret in absolute terms, we report it relative to a naive baseline (rWIS): the ratio of our WIS to the WIS of a simple seasonal-naive baseline computed on the same weeks. rWIS < 1 means the model beats the baseline; 1.0 means it matches it.

The headline validation is a head-to-head benchmark against Influcast, Italy's collaborative forecasting hub: for the 2025/26 season we compared our forecasts against the hub's nine participating models, its ensemble, and its naive quantile baseline across 16 weekly forecasting rounds. The season lies outside the fine-tuning data and post-dates the release of Chronos-2, making this a true out-of-sample test.

Per-country performance varies with data quality: well-monitored countries with long stable series are predicted well, while countries with sparse or irregular surveillance are much harder — and in a few cases the model adds no value over a naive baseline. We show per-country forecasts regardless, because even rough activity levels are better than nothing for countries that have no other operational forecast at all.

• Over-confident intervals. In the Italy benchmark, our 50% prediction intervals covered only ~24% of observed values. The central estimates are well calibrated relative to competing models, but the intervals are too narrow; treat them as indicative, not literal probabilities.
• The target is not influenza-specific. ILI and ARI counts include all acute respiratory illness — SARS-CoV-2, RSV, rhinovirus and others — not just influenza virus. Co-circulating viruses can therefore distort both the target and the forecasts.
• Surveillance artefacts. Reporting lags, mid-series indicator changes, and missing weeks are baked into the training data; the model learns them rather than correcting them.
• Single model, no ensemble. Purpose-built national systems (including Influcast's) benefit from model diversity; we do not yet combine multiple models or apply post-hoc calibration.
• Generalisation. The rigorous head-to-head validation is against one hub, in one country. Whether the same pattern holds against forecasting systems in other countries is an open question we intend to test.

This is a research tool. Forecasts are provided for comparison and exploration, not for operational decision-making.