gbó

“Box, box. The tyres are gone.”

A model that watches a stint the way a pit wall does and calls when it falls off, trained on every green-flag lap of 2025 and replayed against all 698 real stops.

Machine learning
Python
scikit-learn
FastF1 timing
0.50 s
ten laps out
the model0.50field’s slope0.54nothing changes0.58

mean error forecasting a stint's pace ten laps ahead, on races the model never saw.

26%
of every stop
180 of 698 stopscame almost free

came under a safety car or VSC. The cheapest stop beats any tyre model.

+9 → −1
laps
0earlylate2 s3 s4 s

late, then early: the teams' verdict as the fuel assumption moves from 2 s to 4 s a lap.

The question

A tyre loses grip as it wears, and at some point the lap time it costs is worth more than the 20-odd seconds of a pit stop. Stop too early and you run a longer stint later on worse rubber; too late and you bleed time every lap. The old picture is a cliff: steady, steady, then suddenly two seconds a lap. If a model can see that coming from the laps so far, it can call the stop. Gbó is Yoruba for worn out, the way cloth wears thin.

Everything here is from the public timing feed of the 2025 season, the same data FastF1 reads, in TracingInsights' archive: every lap time, tyre compound, tyre age, pit entry and exit and track status for 22 dry races. Australia and Silverstone were wet and are left out.

How it works

  1. Clean 20,669 laps
    Green flag only, no in- or out-laps, no first lap, nothing 3% off that lap's field median.
  2. Take the fuel out
    About 100 kg at 0.03 s a kilo: every lap goes on an empty-tank basis, 3 s at the start, burned to nothing.
  3. Seconds off the best
    A lap's three-lap median against the stint's best so far. The cliff is 1.0 s.
  4. Forecast
    Gradient boosting, from the stint so far and everyone else's laps so far, 1 to 20 laps ahead.

The model only sees what the pit wall had at that lap: the stint's own trend, its compound, age and fresh or used, track temperature, how far into the race, and how quickly every other car on the same compound has been fading in this race up to now. Each race is predicted by a model trained on the other 21, and since every track ran once in 2025, it is always a track it has never seen.

Key decisions

A second off, not a shape

A cliff should be a bend: the loss per lap suddenly jumping. A hinge fitted to every stint of a dozen laps or more finds a strong one in 31% of them, but a median 13 laps before the stop: a driver managing then pushing, more than rubber giving out. So the cliff here is plain, the first lap a stint runs 1.0 s off its own best.

Live, or it doesn’t count

The field's fade rate is recomputed at every lap from laps already run, and a stint's best is its best so far. A model allowed to know how a stint ends would forecast it beautifully and be useless on the wall.

Only the stop lap moves

Replaying a stop, the next stint keeps its compound and its end. At each lap the model adds its forecast for the old tyre to the typical fade of the new one and picks the stop lap that loses least. Its call is the first lap that answer is this lap.

Results

Fade · 892 stints
Tyres wear on a slope
0.00.51.01.5010203040laps on the tyres off bestthe “cliff”softmediumharda slope, not a cliff
Median loss by tyre age. A soft gives up about seven tenths in ten laps, a hard about four, and 66% of stints run a second off their best before they stop (72% of softs). Teams are not stopping before a fall; they are deciding how much fade a stint is worth.
Forecast · races it never saw
The fade can be forecast
0.000.250.500.7515101520laps aheaderror, sfield’s slopenothing changesthe model0.50 s, ten laps out
One lap out nothing beats assuming the next lap matches this one (0.17 s against 0.14). From five laps the model pulls ahead, and by twenty it is 20% better than standing still and 22% better than borrowing the field's rate.
Hungarian · NOR · medium
One stint, one call
0.00.51.01.51102030laps off bestall the wall has at lap 20the forecastmodel: box, lap 29team, lap 31

Here is how a call happens. At lap 20 of Norris's first stint in Hungary, the wall has the black line and nothing else. The model's forecast, in brown, crosses a second off nine laps later, so on lap 29 it says box. McLaren boxed on lap 31, and the tyre did fall away in between.

Most calls are not this tidy. Below is every stint of two races, with the lap the model would have called marked above it.

Replay · 70 laps
Hungarian Grand Prix
NORPIARUSLECALOBORSTRLAWVERANTHADHAMHULSAIALBOCOGASTSUCOLBEA110203040506070
Replay · 56 laps
Chinese Grand Prix
PIANORRUSVERLECHAMOCOANTALBBEAGASSTRSAIDOOHADLAWBORHULTSUALO11020304050
softmediumhardmodel calls itstill wantedsafety car
Calls · 485 green-flag stops
Who stops late depends on what you assume about fuel
2 s a lap
139median +9early0late
within 2 laps: 11%
3 s a lap · used throughout
median +3early0late
within 2 laps: 24%
4 s a lap
median −1early0late
within 2 laps: 23%
Laps between the model's call and the team's stop, at three guesses of what a full tank costs. Dark bars are within two laps of the model; the ends gather everything fifteen or more out. The ten-lap forecast barely moves (0.52, 0.50, 0.50 s); the verdict swings from nine laps late to a lap early.

At 3 s, teams stop a median 3 laps after the model, 6 on hards, and match it within two laps 24% of the time. It looks like the wall stays out too long. It isn't that simple. Inside a stint, every lap adds one lap of tyre age and burns one lap of fuel, so lap times alone can never say which one is slowing the car.

Where it stands

Half of it works. A model trained on public timing can forecast a stint's fade ten laps out to half a second at a track it has never seen, better than any simple rule. The other half, calling the stop, needs a number only the teams have: their fuel burn. And the model is blind to what the wall weighs most, the gap to the car behind, the undercut and the chance of a safety car, which delivered 26% of the year's stops anyway.

Three things would sharpen this: the real fuel loads, gaps to the cars around each driver so pit windows are part of the call, and 2026, when every track comes round a second time and the model can learn the circuit, not just the compound.

Laps are loaded and cleaned in research/gbo/laps.py, the model is trained and tested in research/gbo/model.py, the stops are replayed in research/gbo/replay.py, and the numbers above come from research/gbo/summarize.py and sketch_data.py. Data: the 2025 F1 live-timing feed as archived by TracingInsights from FastF1. The 3 s fuel effect, the 1.0 s cliff, the 3% traffic cut and the one-stop-moves replay are this study's choices, not the teams'.