datamatters24/f1-race-predictor-model

Telemetry Chaos — F1 Race Prediction Model

XGBoost + LightGBM ensemble predicting Formula 1 race winners from 76 seasons of historical data. Tuned with Optuna hyperparameter optimization across 200 trials. Auto-retrains weekly during the active season.

Live demo: telemetrychaos.space

Performance

Metric	Score
Top-1 Accuracy	53%
Top-3 Accuracy	85%
Top-5 Accuracy	96%

Evaluated on 2024–2025 seasons with time-series split to prevent data leakage.

Features (21 per driver per race)

• Form: Rolling average points, recent podiums, win streak
• Pace: Practice session lap time delta vs. teammate and field
• Constructor: Team rolling performance, reliability score
• Track history: Driver-specific circuit win rate, podium rate
• Tyre: Degradation profile, pit stop speed, strategy tendency
• Conditions: Weather forecast, safety car probability
• Grid: Starting position, qualifying gap to pole

Architecture

XGBoost (GPU) + LightGBM
Optuna HPO: 200 trials, TPE sampler
Time-series split: train on seasons N-5 to N-1, evaluate on N
Final output: softmax win probabilities per driver

Dataset

• Coverage: 1950–2025, 76 seasons
• Records: 1,322,914 race records
• Telemetry laps: 470K+
• Sources: FastF1, Jolpica-F1, f1db, Kaggle

Usage

import joblib

model = joblib.load("f1_ensemble.joblib")
# Input: 21-feature vector per driver
# Output: win probability (0-1)
probs = model.predict_proba(X)

Auto-Update Pipeline

During the active F1 season the model retrains weekly:

1. Pull latest race results and telemetry via FastF1
2. Engineer features for upcoming race grid
3. Retrain ensemble with updated data
4. Publish updated predictions to telemetrychaos.space

Citation

@misc{rubin2026telemetrychaos,
  author = {Rubin, Theodore},
  title = {Telemetry Chaos: F1 Race Prediction with XGBoost/LightGBM Ensemble},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datamatters24/f1-race-predictor-model}
}