Verifying rate was 10.96%
Precipitation reliability
Do precipitation probabilities match observed frequencies?
Raw NBM probabilities were higher than verifying frequencies in every probability band. The size of that mismatch changed little with lead time, but overall probability error grew, and predicting the amount remained difficult on days with substantial precipitation.
Brier score; lower is better
When more than 0.25 in was observed
A forecast probability of 54.3% corresponded to a 43.5% verifying frequency.The same high-side pattern appeared in every probability band, from very small chances through the 90–100% group.
The event being tested is specific
This analysis is not the probability of any rain at any time during a calendar day. It tests whether a fixed 24-hour total from 12:00 UTC one day to 12:00 UTC the next was strictly greater than 0.25 inches. A trace or a tenth of an inch counts as a non-event for this question.
The verifying value is NOAA’s radar- and gauge-based Multi-Radar/Multi-Sensor precipitation estimate averaged to the NBM grid. It is a gridded estimate rather than a rain-gauge observation at each forecast point.
The threshold is useful because it separates minor amounts from a wetter 24-hour period and matches the probability shown in ModelSpread’s precipitation outlook. It should be compared with the official chance of precipitation only with care: the National Weather Service probability on a forecast card can use a different time window and is generally the chance of measurable precipitation, not the chance of exceeding a quarter inch.
Raw probabilities were higher than verifying rates in every band
| Raw forecast band | Mean forecast | Verifying frequency | Share of cases |
|---|---|---|---|
| 0–9% | 1.7% | 1.1% | 65.5% |
| 10–19% | 14.0% | 10.6% | 11.1% |
| 20–29% | 24.1% | 19.0% | 6.6% |
| 30–39% | 34.2% | 27.9% | 4.5% |
| 40–49% | 44.3% | 36.2% | 3.4% |
| 50–59% | 54.3% | 43.5% | 2.6% |
| 60–69% | 64.3% | 52.1% | 2.0% |
| 70–79% | 74.3% | 61.4% | 1.6% |
| 80–89% | 84.3% | 73.3% | 1.3% |
| 90–100% | 95.4% | 89.2% | 1.4% |
Forecasts in the 10–19% band averaged 14.0%, while the event occurred in 10.6% of verifying estimates. The 30–39% band averaged 34.2%, compared with a 27.9% verifying frequency. At higher probabilities, the gap remained: 74.3% forecasts verified 61.4% of the time, and the 95.4% group verified 89.2% of the time.
About two-thirds of all cases fell in the 0–9% band. That large dry-background group matters when interpreting national averages: a probability system can produce a good overall score by handling common low-probability cases well even if its less common high-probability forecasts have visible reliability gaps.
Reliability is only one part of probability quality
A reliable 30% forecast should occur about 30% of the time over many comparable cases. Reliability does not measure whether the system successfully assigns higher probabilities to wetter cases than to dry ones. A forecast could be reliable but uninformative by issuing the same base-rate probability every day. The current table tests calibration of the probability scale; it does not provide a full discrimination or value analysis.
The Brier score combines probability error across events and non-events, with lower values indicating better performance. It rose from 0.0480 at day 1 to 0.0817 at day 7, showing that probability quality deteriorated with lead time even though the overall event rate stayed close to 11%.
Probability and amount errors grew farther ahead
| Day | Brier score | Mean forecast | Verifying rate | Range coverage when observed amount exceeded 0.25 in | Average error of the median forecast when observed amount exceeded 0.25 in |
|---|---|---|---|---|---|
| 1 | 0.0480 | 13.73% | 10.86% | 75.11% | 0.383 in |
| 3 | 0.0595 | 13.51% | 10.92% | 71.67% | 0.437 in |
| 5 | 0.0702 | 13.47% | 11.05% | 65.43% | 0.505 in |
| 7 | 0.0817 | 13.48% | 11.06% | 57.35% | 0.582 in |
The mean raw probability remained near 13.5% at every selected lead, and the verifying rate remained near 11%. The growing Brier score therefore came from weaker case-by-case probabilities rather than a major shift in the overall sample frequency.
Amount forecasts also became less precise on days that actually exceeded 0.25 inches. The median amount missed by 0.383 inches on day 1 and 0.582 inches on day 7. The share of observed amounts inside the raw P10–P90 range fell from 75.11% to 57.35% across the same leads.
Why the overall amount-range coverage can look deceptively strong
Across all cases, the raw amount range covered about 93% of outcomes at every lead. That number sounds excellent until the dry-day mixture is considered. Zero lies inside many precipitation ranges, and dry or nearly dry outcomes are common. Selecting only cases that ended above 0.25 inches reveals much lower coverage.
The figures for cases above 0.25 inches should not be compared directly with an 80% target. The NBM range applies to all forecast cases, while the subset is selected using what ultimately happened. The comparison serves a narrower purpose: it shows that an impressive all-case coverage number does not answer the user’s question about how accurate the amount will be when substantial precipitation occurs.
Historical calibration reduced the raw probability bias
In rolling retrospective tests, the calibrated mean probability moved from 13.55% to 10.99%, close to the 10.96% verifying rate, and Brier score improved by 2.1%. Detailed product-validation results, including amount-range width, interval score, monthly tests, and prospective status, are kept with the precipitation calibration methodology rather than treated as a separate research result.
How to use the precipitation outlook
The probability answers whether the quarter-inch threshold is likely to be exceeded during the stated 24-hour window. The likely amount range applies to all possible outcomes, including zero. It is not a conditional range for the amount assuming substantial precipitation occurs.
For a low-probability dry outlook, the threshold chance is usually the clearest summary. When the chance is larger, both the threshold probability and amount range matter. Farther ahead, users should allow more room because the probability score, amount error above 0.25 inches, and corresponding range coverage all worsened with lead.
Bottom line: raw quarter-inch probabilities ran high in this archive, and amount uncertainty grew with lead time. Historical calibration reduced the probability mismatch, but predicting the amount on genuinely wet days remained difficult.
Methods
The full archive contains 8,387,624 usable matches from 758 forecast dates and 1,709 contiguous-U.S. locations with at least one usable match. The primary recent-year analysis uses 344 dates from October 1, 2025 through September 13, 2026 and 3,820,069 cases. NOAA’s radar-based Multi-Radar/Multi-Sensor precipitation estimate is averaged to the NBM grid; cases with a Radar Accumulation Quality Index below 0.5 are excluded. That quality filter removes 2% to 5% of cases in the Southeast, South, Ohio Valley, and Northeast but 31% to 34% in the Southwest, Northwest, and Northern Rockies and Plains, where radar coverage is weaker. Winter exclusions are larger: 17% nationally and 26% to 40% in the Upper Midwest and those three western regions. Those regions and seasons are therefore represented by fewer cases. Reliability bands pool location cases within each probability band. Lead-time summaries first average locations within a date and then give each date equal weight.
This analysis uses the precipitation verification archive snapshot generated August 12, 2026. Unless noted, results describe past performance of archived, unadjusted NBM guidance. Nearby stations and adjacent dates often experience the same weather, so the observations are not fully independent.