2 OCT 2026 — Each flu season, the US Centers for Disease Control and Prevention asks outside teams to forecast how many people will be admitted to hospital with flu, and combines their forecasts into one that it publishes. Last season that combined forecast beat every individual model. In the 2025–26 season it came seventh, behind a model built by Google, according to CDC's end-of-season evaluation published on 30 September.
CDC uses that combined forecast in its public messaging about how many flu admissions to expect in the coming weeks.
How the forecasts are judged
Through the programme, called FluSight, teams submitted forecasts each week from November 2025 to May 2026 for the current week and up to three weeks ahead, for every state and Washington DC. A total of 34 teams sent in 53 models, and 39 submitted enough forecasts to be scored.
Each forecast is a range, not a single number. The main score, the relative weighted interval score, rewards ranges that are both narrow and right. It is measured against a simple baseline that assumes next week will look like last week. A score below 1 beats the baseline, and 33 of the 39 models did.
The model that came first
The top model, Google_SAI-FluEns, scored 0.56, against 0.62 for CDC's combined forecast. CDC classes the Google model as an ensemble with AI, mechanistic and statistical components.
CDC cites two references for the model. One is a Nature paper on an AI system that writes expert-level scientific software. The other is a preprint on forecasting diseases with a search guided by a large language model, which suggests the model was developed with AI-generated code. Google's announcement says only that the model was "built with Google AI" and that it "best matched the season's observed hospital admissions."
Where the forecasts went wrong
The 2025–26 season was moderate, with weekly admissions peaking above 40,000 in the week ending 27 December. That week was the combined forecast's worst. Across states, fewer than a quarter of its two-week-ahead ranges contained the number of admissions that actually happened. Coverage dropped again in mid-January, during the steepest fall in hospitalisations, and only settled near the intended 95% from February.
CDC's summary is blunt. Forecast performance declined "around periods of rapidly changing influenza trends." Those turning points are when planners most need an accurate forecast.
Why the ranking changed
In the 2024–25 season, CDC's combined forecast outperformed all 36 submitted models, and the top ten individual models scored within 0.04 of each other. This season the combined forecast still did well. It was one of 12 models that beat the baseline in every jurisdiction, but six individual models finished ahead of it.
CDC builds the combined forecast by taking the median of the models that ask to be included. One season is not enough to show whether individual models will keep beating it.
What comes next
CDC says an evaluation of its forecasts for emergency-department visits due to flu is still to come. Whether Google's lead holds will be clear only after another season of forecasts is scored.
View author profile → · Editorial policy
