F1machinelearningproject.
A recruiter-facing machine learning project that doubles as a real product: historical race dossiers, calibrated evaluation, and a session-aware GitHub Actions pipeline that retrains the model after every F1 weekend and republishes predictions.
Current snapshot
model version
a9872a5
published
2026-04-23T10:31:09+00:00
held-out log-loss
0.483
seasons public
2024 · 2025 · 2026
Japanese Grand Prix.
Flip through every target the model predicts — race points, podium, win, qualifying, sprint, head-to-head. The ranking below shows the top ten probabilities, colour-coded by whether the model got it right.
Race date 2026-03-29. The pipeline will start pre-publishing predictions ahead of each session in a later pass — for now this shows the freshest race in the held-out set.
Full dossierTarget
Race — driver scores points (top 10)
- 01
PIA
McLaren
94% - 02
ANT
Mercedes
94% - 03
RUS
Mercedes
94% - 04
LEC
Ferrari
90% - 05
NOR
McLaren
88% - 06
VER
Red Bull Racing
78% - 07
HAM
Ferrari
74% - 08
HAD
Red Bull Racing
69% - 09
GAS
Alpine
66% - 10
LIN
Racing Bulls
54%
Every race, every prediction.
Compact per-race summary for the 2026 season using the primary target. Click any card for the full dossier with driver-level factors.
Seventargets,onemodelfamily.
Each target is trained independently with calibrated probabilities. Metrics are held-out across all public seasons, not cherry-picked.
Race — driver scores points (top 10)
best model: gbm
log-loss
0.483
brier
0.158
ECE
0.035
Qualifying — driver reaches Q3 (top 10)
best model: features_logit
log-loss
0.476
brier
0.157
ECE
0.065
Sprint — driver scores points (top 8)
best model: features_logit
log-loss
0.484
brier
0.136
ECE
0.127
Sprint Qualifying — driver starts top 5
best model: features_logit
log-loss
0.355
brier
0.103
ECE
0.079
Race — driver finishes on the podium (top 3)
best model: gbm
log-loss
0.191
brier
0.058
ECE
0.026
Race — driver WINS (1st place)
best model: gbm
log-loss
0.091
brier
0.030
ECE
0.024
Head-to-head — driver beats their teammate in the race
best model: gbm
log-loss
0.580
brier
0.198
ECE
0.051
Primary honesty check
The main public story stays anchored on the held-out 2025 season. Predictions for each race are regenerated after every session, but the historical evaluation stays honest by preserving the numbers as they were first tested.
Training lives in the ml/ folder — LightGBM per target, isotonic calibration, regulation-era weighting. The GitHub Actions pipeline rewrites itself weekly to schedule runs ~30 minutes after every F1 session.