Forecast accuracy
Checked every night against pilot reports, which are densest over North America.
- Bumps likely
- 1.8×
- Where we warn of bumps, pilots report them 1.8 times as often.
- Calm
- Nearly 8 in 10
- When we say calm, there’s no notable turbulence.
- vs NOAA GTG
- Comparable to NOAA
- Checked against NOAA GTG on US reports: comparable skill, and it’s worldwide.
- Flights behind the routes
- 16M flights
- Routes follow the paths real flights flew.
What a forecast can and can’t do
Turbulence is small-scale: a patch of rough air can be a few kilometres wide and last minutes. A forecast can’t say exactly where it will be. It can say where bumps are more or less likely, and that is what we check.
How we measure it
The sample
17 October – 15 November 2026 (30 days): 31,244 pilot reports at cruise levels, about 1,041 a day. Each report counts once and is checked against the forecasts we had published before its time; later forecast steps are not scored. Intervals (in grey) are 95% and resample whole days, because one day’s reports share the weather. Pilots report rough air far more often than smooth air: 32% of these reports were moderate or worse, far more than the sky as a whole, so the shares below describe reported spots, not the chance on a given flight.
The app’s levels against the reports
- Where we said bumps likely: 59% 55%–62% moderate or worse, against 32% overall (1.8× 1.7–2.0×).
- Where we said calm: 78% 76%–80% without moderate or worse; 29% completely smooth, against 22% overall.
- Of the moderate-or-worse reports, 60% 56%–64% were in patches or bumps likely, 17% 14%–21% in bumps likely.
| Ahead | Moderate+ in patches or bumps | Moderate+ in bumps | Calm without moderate+ |
|---|---|---|---|
| Within 24 hours | 64% 60%–68% | 22% 19%–26% | 79% 78%–81% |
| 24 to 48 hours ahead | 59% 55%–63% | 16% 13%–18% | 78% 76%–80% |
| 48 to 72 hours ahead | 55% 50%–58% | 11% 9.2%–13% | 76% 74%–79% |
Levels as the app shows them: calm below a 3.5% chance of moderate and 30.5% of light turbulence, bumps likely from 9.5% of moderate, patches between.
Ranking skill (AUC)
Given one moderate-or-worse report and one smooth one, how often the forecast gave the rough one the higher chance: 0.5 is a coin flip, 1.0 perfect. Light reports are left out, so these tables count fewer reports.
| Slice | AUC | 95% interval | Reports |
|---|---|---|---|
| All | 0.76 | 0.73–0.78 | 16,740 |
| 6 h ahead | 0.78 | 0.76–0.80 | 8,145 |
| 12 h ahead | 0.78 | 0.76–0.80 | 16,740 |
| 18 h ahead | 0.77 | 0.75–0.79 | 8,153 |
| 24 h ahead | 0.77 | 0.74–0.79 | 16,308 |
| 30 h ahead | 0.76 | 0.74–0.78 | 7,889 |
| 36 h ahead | 0.75 | 0.73–0.78 | 16,001 |
| 42 h ahead | 0.74 | 0.72–0.76 | 7,778 |
| 48 h ahead | 0.74 | 0.71–0.76 | 15,719 |
| 54 h ahead | 0.73 | 0.70–0.75 | 7,559 |
| 60 h ahead | 0.73 | 0.71–0.76 | 14,156 |
| 66 h ahead | 0.72 | 0.69–0.74 | 5,892 |
| 72 h ahead | 0.72 | 0.69–0.75 | 10,121 |
| Contiguous US | 0.76 | 0.73–0.78 | 15,167 |
| Alaska, Canada and ocean edges | 0.75 | 0.71–0.79 | 1,204 |
| Elsewhere | 0.77 | 0.71–0.82 | 369 |
Does the chance match the reports?
| We said moderate+ at | Reported moderate+ | vs all reports |
|---|---|---|
| under 3.5% | 23% 21%–24% | 0.7× |
| 3.5%–9.5% | 43% 41%–46% | 1.4× |
| 9.5%–20.0% | 57% 54%–59% | 1.8× |
| 20.0%+ | 66% 61%–69% | 2.1× |
Reported shares sit above the forecast chances because pilots mostly report rough air; what counts is that they rise together.
Allowing for how often pilots report, the largest gap between implied and reported rates across ten groups of forecasts was 3 percentage points. Most of these reports come from the US. Outside the US the percentages are not calibrated yet: a higher one does mean rough air is more likely, but don’t read it as exact odds.
Against NOAA GTG
NOAA’s Graphical Turbulence Guidance is the US government’s high-resolution turbulence forecast for aviation, for the contiguous US only. On the 13,517 US reports (of 15,167) where we hold GTG’s 6-hour forecast, each counted once against our latest forecast published before it (typically 9 h ahead): Jetbrief 0.78 0.75–0.80, GTG 0.76 0.75–0.78. The difference, +0.01 −0.00 to +0.03, is within the uncertainty. Flagging the same share of spots, Jetbrief caught 21% of the moderate-or-worse reports and GTG 20%. We hold GTG only for reports scored while NOAA still served the file.
An example day
2026-11-11: 536 of 1,513 reports were moderate or worse, and 134 of them (25%) fell where our latest forecast, typically 9 h ahead, said bumps likely; there 54% of reports were moderate or worse, elsewhere 32%. Picked as the day of the last 7 with the most such reports, so the best of the week, not a typical day (median day: 23%).
Behind the forecast
- Each forecast: 1,038,240 points of a global 0.25° grid × 10 flight levels × 25 steps over 72 hours, 260 million values, blended from ECMWF and GFS; 8 published in the last 24 hours.
- Routes: 16,436,065 airline flights seen by public ADS-B receivers (adsb.lol), 2025-10-12 to 2026-11-14, 34% with an inferred airport. They shape the routes; they don’t measure accuracy.
Model blend-v1.9 · September to November · generated 2026-11-16. Pilot reports: Iowa Environmental Mesonet. Yardstick: NOAA GTG. Forecast data: NOAA GFS and ECMWF open data. Flight paths: adsb.lol.
The figures on this page and the data file accuracy.json may be reused with a link to this page (CC BY 4.0).