FEMA's hail risk score and actual hail events agree less than you would expect

FEMA's National Risk Index is the most widely cited public measure of natural-hazard risk in the United States. It appears in lender risk models, municipal planning documents and consumer-facing tools. For hail it produces a 0–100 score per census tract.

We compared that score against what actually happened. Matching NOAA Storm Events data to counties, 1,354 counties above 50,000 residents have both a FEMA hail risk score and a record of observed hail events between 2021-01-01 and 2026-08-01. The Pearson correlation between the two is 0.243.

That is weaker agreement than most readers would assume, and the reason is important: the two are not measuring the same quantity. FEMA's score is an expected-annual-loss model that folds in building value, population and social vulnerability. NOAA's record is a count of reported events. A rural county with frequent hail and little to damage scores low on the first and high on the second. Neither is wrong.

The practical problem is interpretive. A homeowner who looks up their address, sees a low hail risk rating and concludes hail is unlikely has misread the score — it tells them expected loss is low, not that hail is rare. Williamson County, Tennessee carries a score of 26.3 and reported 84 hail events in five and a half years. Butler County, Pennsylvania scores 31.2 with 46 events.

Normalising by population to remove the reporting-density bias in NOAA data — more residents means more people to file a report — the same pattern holds. Williamson County reports 33.9 hail events per 100,000 residents against a risk score of 26.3.

• This is not a claim that FEMA's model is wrong. The National Risk Index measures expected annual loss, not event frequency; a low correlation with raw event counts is partly expected by design. The finding concerns how the score is likely to be read, not whether it is fit for its stated purpose.

• NOAA storm reports depend on human observation. Counties with more people, more storm spotters and more insured property generate more reports for the same weather. Population normalisation reduces this bias but does not eliminate it.

• Five and a half years is a short window for hail climatology. FEMA's model is built on far longer records, and some of the divergence is simply that a short observation period is noisy.

• Counties were joined on name rather than FIPS code, which drops parishes, boroughs and independent cities whose naming differs between the two sources. 1,354 counties matched; the unmatched remainder is excluded rather than estimated.

• County-level averaging of tract scores treats all tracts equally regardless of population.

The studies above are national. This is your specific roof — size, pitch, and the storm history recorded at your address. Free, about six seconds.