BOX BOX BOX · S6E5
LIVE PROJECTNext-lap classification
F1 STRATEGY INTELLIGENCE
Predict the next pit stop. Before the radio call.
A lap-level machine-learning project that turns tyre, timing, position and race context into an actionable pit probability.
01Built from the full notebook workflow—not a generic dashboard.
CURRENT LAP31 / 78Monaco · 2025
PIT PROBABILITY78%Prepare the crew
CLICK · TO · PUSH
TYRE LIFE18 lapsInside pit window
BUILT ON REAL PROJECT OUTPUTS
THE STRATEGY LOOP
From race state to pit-wall decision.
The project follows the same question a race engineer asks every lap: what has changed, and is it enough to stop?
READ THE RACE
Every lap becomes a strategy signal.
Tyre age, race progress, stint, position and timing context are assembled into one lap-level view.FIND THE WINDOW
Patterns become pit-stop pressure.
Exploratory analysis reveals the windows, circuit effects and synthetic-data anomalies that shape the prediction.MAKE THE CALL
A probability arrives before the next lap.
The final blend ranks the chance of a stop so the result reads like a pit-wall decision, not a black-box score.THE BEST MODEL
Two models. One sharper strategy call.
LightGBM + RealMLP combine nonlinear race context with a different decision boundary. Their out-of-fold blend delivers the strongest saved submission.
Open the interactive model labLightGBM42.5%
RealMLP57.5%
DEPLOYMENT DECISIONBest score ≠ deployed model
The production API serves the baseline LightGBM rather than the higher-scoring LightGBM + RealMLP blend. Loading and running both models would exceed the available backend memory and compute budget, putting the server process at risk of a crash. Without those infrastructure constraints, the clear deployment choice would be the two-model blend.
EXPERIMENT SCORECARD
Every run moved the strategy forward.
COMPETITION METRICROC AUCHigher is better
Eight saved experiments show how stronger validation, feature engineering and model diversity lifted the leaderboard result.
Score progression
Ordered from the earliest saved run to the winning blend.
OOF / validationPublicPrivate
Public +0.00327Private +0.00344OOF —
Public ±0.00000Private ±0.00000OOF +0.00204
Public +0.01258Private +0.01269OOF +0.03224
Public +0.00198Private +0.00190OOF +0.02097
Public +0.00205Private +0.00188OOF +0.00176
Public +0.00022Private +0.00059OOF +0.00059
Public +0.00069Private +0.00054OOF +0.00077
Lines use a focused 0.89–0.96 axis. The private leaderboard is the primary score and receives the strongest visual emphasis.
| Experiment / Submission | Related file | Validation / OOF | Public | Private | Notes |
|---|---|---|---|---|---|
| BESTLGBM + RealMLP blend | submission_blend_lgbm_realmlp - lb - 0.95328.csv | 0.953876 | 0.95328 | 0.95369 | Best saved submission; OOF blend used 42.5% LGBM and 57.5% RealMLP. |
| RealMLP with feature engineering | submission_realmlp_fe - lb-0.95259.csv | 0.953105 | 0.95259 | 0.95315 | 6-fold StratifiedKFold OOF; mean fold AUC 0.953109. |
| LGBM with feature engineering | submission_lgbm_fe.csv | 0.952511 | 0.95237 | 0.95256 | OOF score loaded in the blend notebook. |
| XGBoost with feature engineering | submission_xgb_stratified_fe - lb 0.95032.csv | 0.950752 | 0.95032 | 0.95068 | 5-fold StratifiedKFold OOF; mean fold AUC 0.950761. |
| XGBoost OOF GroupKFold | submission_xgb_oof_groupkfolds - 0.94834.csv | 0.929780 | 0.94834 | 0.94878 | 5-fold GroupKFold by race; mean fold AUC 0.929450. |
| Baseline EDA LightGBM notebook | submission-baseline-lgbm-0.93576.csv | 0.897540 | 0.93576 | 0.93609 | Year-holdout validation using 2025 as validation. |
| Baseline LGBM submission | submission.csv | 0.895500 | 0.93576 | 0.93609 | Baseline predictions; experiment notebook year-holdout validation. |
| LGBM without early stopping | submission (1).csv | N/A | 0.93249 | 0.93265 | No validation score found in the checked notebooks. |
HOW THE SIGNAL MOVES
Five stages · one next-lap probabilityREADY ON THE PIT WALL