Accuracy

Our forecasts, graded against pilot reports

Most forecast products quote the accuracy of the model they build on. We grade our own forecasts instead: after a flight or route check completes, we automatically compare what we briefed against the official pilot reports (PIREPs) filed along that route while it was flying - and publish the running record here, including the misses.

This page is generated from our verification database and refreshes hourly. Nothing on it is hand-picked: every scored check since the loop went live is counted, and checks where no pilot report could be matched are reported as exactly that - not quietly dropped, and never counted as wins.

The record so far

964 completed checks since June 2026 - 277 had pilot reports to score against.

Coverage: 169 distinct US routes, 534 pilot-report matches (a report can be matched by more than one overlapping check), pooled from 23 paid post-flight verifications and 941 automated daily route checks - both scored by the same rules. Where the grading engine is recorded (951 checks), 873 used NOAA GTG guidance as the primary source and 78 used our other data sources.

How a check is scored

For each completed flight or route check we reconstruct the flight path and collect every pilot report filed within 45 minutes of overflight, within about 50 km of the path, and within 3,000 ft of cruise altitude (reports filed without an altitude are matched without that vertical filter). Each report is compared with the strongest intensity band we forecast anywhere within that window - deliberately generous to the report, because pilot positions carry real error - and the flight gets one verdict:

• Confirmed - every matched report landed within one intensity band of the briefing. Forecasts are bands rather than points, and pilot-report intensity is subjective, so one band is the honest tolerance.
• Bumpier in places - at least one report came in two or more bands above what we briefed. These are the misses we most want to count.
• Smoother than briefed - the air ran at least two bands calmer than the briefing, with nothing two or more bands above it. We'd rather brief a bump that never arrives than the reverse - but it still counts as a miss, not a win.
• No data - no pilot report could be matched to the flight window: none were filed nearby, or those filed were too far off the path, at a different altitude, or filed before our forecast was computed. Common on quieter corridors; reported honestly rather than scored.

Two kinds of checks feed the pool. Paid briefings are graded on the final pre-flight forecast we emailed - the like-for-like test of the product. Daily route checks grade a forecast computed about 30 minutes before departure on popular US routes, which measures short-lead accuracy - the easier end of the forecast problem, and we label it as such rather than blending it silently.

Since October 2026 each daily route check is also forecast about a day ahead and about 6 hours ahead, then graded on exactly the same pilot reports as the forecast made 30 minutes before departure. Those earlier forecasts are shown in their own table below and are never mixed into the record above.

One more rule keeps the grading fair: pilot reports filed before the forecast was computed are excluded from scoring. Our live forecasts incorporate recent pilot reports, so without that cut-off a report could inform a forecast and then be counted as evidence the forecast was right. This rule was added in August 2026; forecasts computed before then are graded without it (or with an approximate cut-off), so the older part of the corpus can, in a narrow window around departure, have matched a report the forecast had already seen - a small optimistic bias that washes out as new checks accrue.

The strict scores

Meteorologists score each matched report against the forecast with no tolerance bands, excluding light-intensity reports (pilots agree on those only about two-thirds of the time). Matching still uses the 50 km / 45-minute window above - the standard tolerance for pilot position error - but the intensity comparison is exact, which is deliberately tougher than the flight-level verdicts, so the numbers read lower. Publishing both is the point.

These are early numbers from 319 scored point-pairs, and they will move as the corpus grows - we publish them anyway, because a small honest sample beats a borrowed headline figure. A report matched by several overlapping checks is counted once (report identity is recorded since August 2026; the older share of the corpus lacks it, so a little correlation remains). For context on what the field achieves: NOAA's published verification for its operational guidance reports probability of detection around 70-85% for moderate-or-greater turbulence with a false-alarm ratio of 25-35%, and roughly 0.85 AUC is the practical ceiling nobody - human or model - meaningfully beats. Those are the upstream model's numbers, not ours; this page exists so you can see ours.

Compared with pilot reports alone

A score means more next to a simple reference. Since October 2026 each check also scores a second forecast, built only from pilot reports: whatever pilots had reported near the route in the 2 hours before our forecast was made, and calm everywhere else. It is scored on the same reports with the same rules, so the gap between the two shows what our forecast adds beyond the reports themselves.

309 checks and 182 scored point-pairs. The reference began in October 2026, so it covers fewer checks than the table above. A forecast that always says calm would score a TSS of zero.

How the forecast does at each lead time

Our briefings start a day or more before departure, so we grade the early forecasts too. Each daily route check is forecast about a day ahead, about 6 hours ahead and 30 minutes ahead, and all three are scored on the same pilot reports with the strict rules above. Each row is the forecast you would have received at that stage, including its data source: a day out it comes from global weather models, because NOAA's GTG guidance only looks about 18 hours ahead, and closer in it uses GTG where GTG covers the route. So the rows reflect both the shorter lead and the change of source, just as your own briefings do.

301 route checks graded at all three lead times, 180 scored point-pairs at each. NOAA GTG was the primary source for forecasts 0 of 301 made a day ahead, 211 of 301 made 6 hours ahead, 279 of 301 made 30 minutes ahead. "Reports alone" is the TSS of the pilot report reference, as of the time each forecast was made.

Read more

The full picture of where the forecast is strong and where it is genuinely hard is in our honest accuracy breakdown. For how this compares with other turbulence tools, see Turbli vs Turbcast. Or just check your flight - the forecast this page grades is free, with no signup.

Updated hourly from the verification database. Most recent check: 21 October 2026 (UTC).