How accurate are turbulence forecasts? What the checks show, including ours

Key facts

• In NOAA's assessment of the global Graphical Turbulence Guidance (July to October 2017), the forecast scored 0.784 against turbulence measured by aircraft sensors and 0.722 against pilot reports, on a scale where 0.5 means no skill. [1]
• The same assessment found every forecast it tested scored lower against pilot reports than against aircraft measurements. [1]
• One study found the location of turbulence in pilot reports differed from automated aircraft reports by an average of 25 miles, and another found data entry errors in 10% to 20% of turbulence reports. [2]
• Sorapuff's clear-air turbulence index scored 0.680 at telling moderate-or-worse reports from lighter ones, on 3,083 US pilot reports at 20,000 feet and above. [3]
• In that sample, 26% of reports were moderate or worse, and pilots called the ride severe only 5 times. [3]
• A 2026 preprint scored the same index at 0.834 against more than 247 million automated aircraft reports, using a reconstruction of past weather rather than a forecast. [4]

How is a turbulence forecast checked?

Forecast verification is the process of comparing what a forecast said with what actually happened. For turbulence, "what happened" comes from two kinds of observation [2]:

• Pilot reports (PIREPs): a pilot's own description of the ride, on a four-level scale of light, moderate, severe and extreme. They're real, but subjective. See what PIREPs and SIGMETs are.
• Automated aircraft reports: many airliners compute the eddy dissipation rate (EDR), an objective measure of the air's turbulence, and send it with an accurate position and time. Our guide on how turbulence is measured explains both.

Each forecast is then matched with each observation and sorted into four boxes: a hit, a false alarm, a miss or a correct "no" [1]. Moving the forecast's threshold up and down trades hits against false alarms. Plotting that trade-off gives a curve called a ROC curve, and the area under it, the AUC, is the standard single score [1].

What does an AUC score mean in plain words?

Take one report a pilot called moderate or worse and one they called lighter, at random. The AUC is the chance that the forecast gave the bumpier one the higher value. An AUC of 0.5 is a coin flip; 1.0 would never get the order wrong. NOAA describes it as a measure of a forecast's ability to tell events from non-events, where larger is better [1].

AUC is useful because it doesn't depend on where an app draws its "light" or "moderate" lines. It doesn't tell you how often a single forecast is "right," which is why we also show calibration below.

What published checks found

The most useful public numbers come from NOAA's assessment of the global Graphical Turbulence Guidance (GTG-G), the blend of many turbulence indexes developed at the National Center for Atmospheric Research [1]:

Those are for a ±1.5-hour window around the forecast time, from July to October 2017 [1]. Three of NOAA's findings help put any turbulence forecast in context:

1. Pilot reports make every forecast look worse. Scores against PIREPs were lower than against aircraft measurements for every product tested [1]. A pilot's report can be miles off in position, and controllers told the NTSB that some airlines uniformly report higher intensities than others [2].
2. Lead time mattered little. From 12 to 36 hours ahead, performance changed only a little, with the 12-hour lead best overall [1].
3. Season matters. The period was Northern Hemisphere summer and early autumn, when global model skill is lowest, and NOAA expected results to change in winter [1].

A 2026 preprint gives another reference point: the classic Ellrod index (TI1) scored an AUC of 0.834 for moderate-or-worse turbulence against more than 247 million automated aircraft reports from 2017 to 2024 [4]. That test used ERA5, a reconstruction of past weather, and objective aircraft data. A real forecast scored against pilot reports is a harder test.

Our own check: Sorapuff's index against 3,083 pilot reports

Sorapuff's forecast runs on your phone. It reads NOAA's free GFS winds along your route and computes the Ellrod index (TI1), a clear-air turbulence index developed in the early 1990s and in use at many aviation forecasting centers [5]. It multiplies vertical wind shear by deformation, how strongly the flow is being stretched [6]. We tested it before building the app around it.

How we tested it. We took 3,083 US pilot reports from 20,000 feet and above that gave a turbulence intensity, covering 48 hours in November 2026. We scored each one with the newest GFS run issued before the report, so the index never saw the future [3].

What we found [3]:

In plain words: given one report a pilot called moderate or worse and one they called lighter, our index ranked the bumpier one higher 68% of the time. For any bump at all versus a smooth ride, it was about 70%.

What didn't beat it. Nothing we tried did better than TI1 on its own. Adding temperature (the Richardson number) added no skill, so the app doesn't download it. A finer 0.25° grid scored 0.675 and 0.707 on 1,502 reports while needing four times the data; a coarser 1° grid scored 0.661 and 0.671, and the app uses it only for long flights [3].

What "light" and "moderate" mean in our forecast

AUC says how well the index ranks the air. Calibration says what a given value means. Here is the share of pilot reports that were moderate or worse at each index value [3]:

On the 0.5° grid, Sorapuff shows light from 2 and moderate from 8 [3], close to a published rough guide that puts light at 2 to 9 and moderate at 9 to 20 [6]. Even in the highest band, about 4 in 10 reports were lighter than moderate, and in the lowest band about 1 in 8 were still moderate or worse. That's what "likelihood, not a promise" looks like in numbers.

Pilots called it severe only 5 times in 3,083 reports [3], too few to calibrate, so Sorapuff never shows "severe" from its own index: only an official SIGMET can put severe on your route.

The honest limits of our test

• One short sample. 48 hours in November: a winter jet stream or a summer storm season could score differently.
• US only, cruise only. Public pilot reports come mainly from the US and the North Atlantic [7], and we tested 20,000 feet and up.
• Pilot reports are noisy. The location and intensity errors above [2] pull every score down, ours included [1].
• Clear air only. The index looks at wind shear, so storms and mountain waves reach your forecast through SIGMETs and recent pilot reports instead. Our score doesn't include those.
• Not an apples-to-apples ranking. NOAA's numbers come from a different period, region and method, so compare them only as a rough guide: the professional blend scored higher, as you'd expect.

Why a forecast can say smooth and the ride is bumpy

Turbulence happens in eddies far smaller than any weather model's grid, and new storms can form between model runs; how turbulence forecasts work explains why. For a step-by-step way to use a forecast before you fly, see will my flight be bumpy?

How Sorapuff helps

Sorapuff shows your route's forecast as a minute-by-minute timeline of smooth, light or moderate, with SIGMETs and recent pilot reports added, and it tells you it's a likelihood. After you land, "Help Sora learn" can ask whether you'd like to send that flight's bump readings, anonymously, so we can keep checking the forecast against real flights. It's optional, asked per flight, and you can see exactly what's sent. Always follow your crew.

FAQ

Can turbulence be predicted?

Partly. Forecasts can reliably show where rough air is more or less likely, using wind models, warnings and reports, but not the exact moment or the exact bump. In NOAA's 2017 check, global guidance put pairs of pilot reports in the right order about 72% of the time (AUC 0.722). Treat any forecast as odds, check again before you leave, and keep your seat belt fastened when seated.

How far ahead are turbulence forecasts accurate?

NOAA found the skill of global turbulence guidance changed only a little between 12 and 36 hours ahead, with 12 hours best. Storm warnings change faster: SIGMETs are valid for up to 4 hours. So a forecast a day or two out is a fair first look, and a check a few hours before departure gives your best picture.

Are turbulence apps accurate?

Apps use different data and methods, and few publish how they were tested. Ask any app what its forecast is built from, whether it includes storms and pilot reports, and whether it has checked itself against real reports. Sorapuff's index scored an AUC of 0.68 on 3,083 US pilot reports, and this page shows the full method and its limits. Our turbulence app comparison covers the rest.

Why was my flight bumpy when the forecast said smooth?

Because a forecast describes odds over an area, not a promise for your seat. Eddies are smaller than any model's grid, storms can pop up between updates, and even at our lowest index values about 1 in 8 pilot reports were moderate or worse. A bumpy ride after a smooth forecast is uncomfortable but normal, and a belted passenger is very well protected.

What does "AUC 0.68" mean?

It means that if you take one pilot report of moderate-or-worse turbulence and one lighter report at random, the forecast gives the bumpier one the higher value 68% of the time. A coin flip scores 0.5 and a perfect forecast 1.0. It measures how well a forecast ranks the air, not how often a single "light" or "moderate" label is exactly right.

Sources

1. Assessment of Graphical Turbulence Guidance, Global (GTG-G), Part 1, NOAA Technical Memorandum OAR GSD-60 (2018), NOAA
2. Preventing Turbulence-Related Injuries in Air Carrier Operations Conducted Under Title 14 CFR Part 121, Safety Research Report SS-21/01 (2021), NTSB
3. Sorapuff forecast check: the Ellrod index from NOAA GFS forecasts scored against 3,083 US pilot reports with the app's skycheck verify tool, Simple Products Labs (November 2026); method and results as described on this page
4. A Generalized Richardson Number Diagnostic for Turbulence in the Free Atmosphere, arXiv preprint 2602.21770 (2026)
5. Improvements to an Operational Clear-Air Turbulence Diagnostic Index, 90th AMS Annual Meeting (2010), American Meteorological Society
6. Ellrod Index, SKYbrary Aviation Safety
7. Data API, Aviation Weather Center