How the FluPulse score works

Last updated September 15, 2026.

What the score is

The FluPulse score is the probability, from 0 to 100 percent, that flu season begins in your area within the next four weeks. A score of 20 means a 20 percent chance; in the seasons we tested, places given that score went on to have flu season within four weeks about 20 times in 100, which is what makes the number a probability rather than a guess. Green is 0 to 24, amber 25 to 69, red 70 to 100.

As the season approaches, the card adds an onset window: the span of weeks, counted from today, within which flu season begins with 80 percent probability, for example “between 2 and 5 weeks from now”. The window comes from the same model fitted at each horizon from this week to twenty weeks out, and it is shown only when it is four weeks wide or narrower, since a wide window says little. It is set so that the statement holds: in the three replayed seasons, the season began inside a shown window 84 times in 100. Windows that narrow appeared in about one pre-season week in five, all of them close to onset; the rest of the time the four-week score is the forecast.

“Flu season begins” has a fixed, local meaning: the week a place’s own flu curve rises to a set fraction of what will be its seasonal peak, and it ends when the curve has fallen well below the peak for two weeks. A county’s season starts when its own emergency-department curve reaches that point, not when the national curve does.

What goes in (United States)

Every day, for every US county, the score is computed from public surveillance published by the CDC and others:

• Emergency department visits for flu, as a share of all visits (CDC NSSP, county where published, otherwise the state).
• Laboratory tests positive for flu, statewide (CDC NREVSS).
• Doctor visits for flu-like illness, statewide, compared with the same state’s summer level (CDC ILINet).
• Influenza A in wastewater at the nearest reporting sewershed (CDC NWSS).
• The mean outdoor temperature over the last seven days, and whether indoor heating season has begun.

Each input is read against the place’s own history, so that the same level of flu activity means more in a county that is normally quiet than in one that always runs high. Where an input is not published for a place that week, the score is computed by a version of the model built without it.

For cities outside the United States the inputs are the country’s weekly laboratory positivity and, where published, its consultation rate for influenza-like illness, from WHO FluNet and national health agencies; see “Other countries” below.

How it was built

The score is a statistical model fitted to three past seasons of CDC county data, 2022-23 to 2024-25, for 70 locations across 24 states: 210 location-seasons and 9,030 location-weeks, each week scored only with the data that had been published by that week. A separate state score was built the same way on eight seasons of state data, 2016-17 to 2024-25 (2020-21 excluded, when there was no flu season).

Every result below is out of sample: each season was scored by a model that had never seen that season. The model is refitted repeatedly on resampled sets of locations to produce the ± shown next to every score, which is one standard deviation of the estimate.

How well it works

County score, three seasons, all weeks. AUC 0.978. The AUC is the probability that, given one week that was followed by flu season within four weeks and one that was not, the score ranks the first above the second; 1.0 is perfect separation and 0.5 is a coin toss. The comparator throughout is the earlier three-band rule, which set the color from fixed thresholds on the same inputs; its AUC on the same weeks was 0.878. The Brier score, which measures how close the probabilities are to what happened, was 0.052 against 0.103 for the rule (lower is better). Per season, score against rule: 2022-23 AUC 0.989 against 0.916, Brier 0.053 against 0.077; 2023-24 AUC 0.989 against 0.879, Brier 0.042 against 0.106; 2024-25 AUC 0.970 against 0.854, Brier 0.061 against 0.125.

The harder test: weeks before onset only. Once a season is under way the score is easy; the question that matters is whether it warns in advance of onset. On the 3,933 weeks before local onset, the AUC of the score was 0.945 and the AUC of the earlier rule was 0.898: the score ranked a pre-onset week that led to a season above one that did not 94.5 times in 100, the rule 89.8 times.

Calibration. When the score says 20 to 30, the season followed 29 percent of the time; 40 to 50, 48 percent; 60 to 70, 69 percent; 80 to 90, 80 percent; 90 to 100, 98 percent. Across the eleven bins the predicted and observed rates are within a few points of each other.

Bands. In the replay, green weeks (0 to 24) were followed by the season within four weeks 4 percent of the time, amber weeks (25 to 69) 46 percent, red weeks (70 to 100) 95 percent. The first week a place crossed 25 came a median of five weeks before its season began; the first week above 70, three weeks before.

State score, eight seasons, 24 states. AUC 0.959 on all weeks and 0.916 before onset; the earlier rule 0.838 and 0.810. The gain held in every one of the eight seasons.

Other countries. For places outside the United States the state score is applied to each country’s national laboratory positivity from WHO FluNet, 2014 to 2025, and then recalibrated on that country’s own history. The transferred score ranks weeks well but its probabilities run too high; recalibration fixes that. Six countries pass the test (Denmark, Norway, France, Germany, Italy, China; recalibrated AUC before onset 0.81 to 0.98). Four do not (Singapore and Thailand, where flu runs year round; the Netherlands and Switzerland, on too few usable seasons) and carry a “not validated” label. Japan, Finland, Taiwan and Canada could not be tested.

What it cannot do

• It is a probability for an area, not for a person. It says nothing about any one person’s risk or illness, and an individual county can run early or late against it; the ± and the onset window show how much.
• It rests on three seasons of county data. The first live season is 2026-27; every weekly score is being recorded so that next spring the season can be scored against what actually happened.
• Where a state does not publish county emergency-department data (Alabama, Arizona, California, Florida, Ohio, Pennsylvania, Washington and New York City), the county score rests on the state series or a state health department file.
• The wording for a season already under way (rising, past the peak, winding down) follows simple rules that have not yet been tested the same way.
• Surveillance is published with a lag of about a week and is revised by its sources; a score can change when last week’s data are corrected.

Sources

CDC NSSP emergency department visits; CDC NREVSS clinical laboratory positivity; CDC ILINet outpatient visits; CDC NWSS wastewater; NOAA and Open-Meteo temperatures; Washington State Department of Health; NYC Department of Health; WHO FluNet for countries outside the United States. FluPulse is not affiliated with any of them.