v0.8.0 — Strategy ML Suite: Pit Stop Prediction & Undercut Intelligence

What's new

N15 — Pit Stop Duration (Quantile Regression)

• HistGradientBoostingRegressor predicting physical stop duration at P05/P50/P95
• Filters out penalty/failure stops — focuses on normal [2.0–4.5s] physical window
• P50 MAE: 0.487s vs baseline 0.555s (team×circuit median)
• Coverage P05–P95: 70.5% on test set
• Features: team, year, compound, tyre_life_in, under_sc, tight_pit_box, team_year_median
• Exports: data/models/pit_prediction/hist_pit_p05/p50/p95_v1.pkl

N16 — Undercut Success Predictor

• LightGBM binary classifier: driver X pits ≤5 laps before rival Y → X gains position?
• 1,032 labeled pairs (2023–2025) with DRY_COMPOUNDS filter
• AUC-PR: 0.6739 | AUC-ROC: 0.7708 | Threshold: 0.522 (Platt calibrated)
• SHAP top features: pos_gap_at_pit, pace_delta, circuit_undercut_rate, tyre_life_diff
• Exports: data/models/pit_prediction/lgbm_undercut_v1.pkl + calibrator + config

N12B — Causal TCN Overtake (Negative Result, Archived)

• Attempted temporal sequence modeling as alternative to N12 LightGBM
• AUC-PR ~0.10 vs N12's 0.5491 — feature-engineered LightGBM wins
• N12 remains production model. N12B archived as valid negative result.

Roadmap

• NLP Radio Processing pipeline planned: N17–N24 (Whisper ASR, BERT sentiment, intent classification, custom F1 NER, model merging, Race Control Messages)

Strategy ML models — full status

Model	Notebook	Metric
Delta Lap Time	N06	MAE 0.392s
Tire Degradation TCN	N09/N10	MC Dropout calibrated
Overtake Probability	N12	AUC-PR 0.549 · AUC-ROC 0.876
Safety Car Probability	N14	AUC-PR 0.072 · lift 1.67×
Pit Stop Duration	N15	P50 MAE 0.487s
Undercut Predictor	N16	AUC-PR 0.674 · AUC-ROC 0.771