The fuel model lands within 5% of published figures it never saw
Fuel Horizon estimates how much fuel each plane in the sky has left. This page checks the model behind that estimate against 126 published fuel-burn figures for real airliners on real sector lengths. It compares the raw OpenAP performance model with the version Fuel Horizon uses, which is tuned per aircraft type and then tested on figures it wasn't tuned on.
Predicted vs. published fuel burn
Each dot is one published figure: fuel burned per kilometre for a given aircraft, sector length and seat count. Dots on the dashed line are exact matches.
How far off each prediction is
The same figures, sorted by signed error. Calibration pulls almost everything into the middle band, where the model is within 5% either way.
Where it still misses
The largest held-out errors mostly come from types whose published figures disagree with each other. A single correction per type can't match two sources that are 20% apart for nearly the same trip.
Correction per aircraft type
The multiplier applied to OpenAP's fuel flow for each type. Below 1 means raw OpenAP burned more than published figures; above 1 means it burned less. Types without published figures borrow a correction from a close relative.
Every figure
All published figures used, with both predictions, sortable. Errors are colour-coded: within 5%, 10%, 20%, and beyond.
How the check works
- Fly each published trip in OpenAP
The model climbs at 2,000 ft/min to its cruise height, cruises at its usual Mach number, and descends, getting lighter as it burns fuel. Payload is the published seat count at 82% full, 100 kg per passenger.
- Compare fuel per kilometre
Taxi plus trip fuel, divided by the sector length, is compared with the published figure in kg/km. Raw OpenAP is off by on average.
- Fit one correction per type
Each aircraft type gets a single multiplier on fuel flow: the geometric mean of the multipliers that would match each of its published figures exactly.
- Test on figures it didn't see
For every figure, the correction is refitted using only that type's other figures, then used to predict the one left out. That leave-one-out error is what this page reports.
- 19 figures can't be tested this way. Their types have only one published figure, so there is nothing left to fit on once it's held out. They are in the table but not in the charts.
- Sources use different assumptions. Seat layouts, load factors and reserve rules vary between the figures, which caps how close any single model can get.
- Live flights add more uncertainty. Real payload, winds and routing aren't known from the broadcast signal, so the globe shows each estimate as a range of at least ±12%, wider than the 5% here.
- Turboprops aren't simulated. The 9 turboprop and utility types use their published figures directly, so they aren't part of this check.