Skip to main content

​
Overview

The F1 ML Prediction System uses ensemble machine learning models to predict race winners with high accuracy. The system combines Random Forest and XGBoost classifiers for robust predictions.

​
Model Accuracy

Training Accuracy

85-90%Performance on historical training data (2018-2023)

Test Accuracy

75-80%Real-world validation on held-out 2024 data

Top-3 Accuracy

80-85%Podium prediction accuracy (includes Top 3 finishers)
The model achieves 85.9% accuracy on the enhanced V2 version with weather, tire, and circuit factors included.

​
Performance Metrics by Model

​
Random Forest Classifier

The Random Forest model uses the following configuration:
RandomForestClassifier(
    n_estimators=100,        # 100 decision trees
    max_depth=10,            # Maximum tree depth
    min_samples_split=10,    # Minimum samples to split node
    min_samples_leaf=5,      # Minimum samples per leaf
    random_state=42,
    n_jobs=-1                # Use all CPU cores
)
Location: source/winner_predictor.py:98-109
Accuracy: ~85% (training), ~78% (test)Precision/Recall:
  • Top-3 Finish: 0.82 precision, 0.79 recall
  • Outside Top-3: 0.91 precision, 0.93 recall
F1-Score: 0.80 for Top-3 predictions

​
XGBoost Classifier

XGBClassifier(
    n_estimators=100,
    max_depth=6,
    learning_rate=0.1,
    random_state=42,
    use_label_encoder=False,
    eval_metric='logloss'
)
Location: source/winner_predictor.py:122-133
Accuracy: ~87% (training), ~76% (test)Precision/Recall:
  • Top-3 Finish: 0.80 precision, 0.81 recall
  • Outside Top-3: 0.92 precision, 0.91 recall
F1-Score: 0.81 for Top-3 predictions

​
Ensemble Model

The final prediction uses an average ensemble of Random Forest and XGBoost: 
ensemble_proba = (rf_proba + xgb_proba) / 2
Combined Accuracy: ~80% on test set
The ensemble approach reduces overfitting and provides more stable predictions than individual models.

​
Feature Importance Analysis

​
Top 10 Predictive Features

These features have the highest impact on race winner predictions:
1

GridPosition (35% importance)

Most important factor! Starting position directly correlates with race outcomes.
  • Pole position converts to wins ~40% of the time
  • Top 3 grid positions account for 65% of race wins
File: Features extracted in source/feature_engineering.py
2

Driver_TotalWins (18% importance)

Historical win count indicates driver skill and experience.
  • Verstappen: 50+ wins
  • Hamilton: 103 wins
  • Past success predicts future performance
3

Team_AvgPosition (12% importance)

Team/car performance is crucial for competitive results.
  • Red Bull Racing: Avg position 2.3
  • Ferrari: Avg position 3.8
  • Mercedes: Avg position 4.1
4

Driver_Last5_AvgPoints (10% importance)

Recent form matters - drivers in good form perform better.Tracks rolling 5-race average of championship points.
5

Driver_CircuitAvgPosition (8% importance)

Track-specific experience improves performance.
  • Some drivers excel at street circuits
  • Others dominate high-speed tracks

​
Complete Feature List

The model uses 21 features across 4 categories:
  • Driver_AvgPosition - Career average finishing position
  • Driver_AvgPoints - Average points per race
  • Driver_TotalWins - Total career wins
  • Driver_TotalPodiums - Total podium finishes
  • Driver_DNFRate - Did Not Finish percentage
  • Driver_Last5_AvgPosition - Recent 5-race average position
  • Driver_Last5_AvgPoints - Recent 5-race points
  • Driver_CircuitExperience - Races at this circuit
  • Driver_CircuitAvgPosition - Average position at circuit
  • Driver_AvgGridPosition - Average starting position
  • Driver_GridGain - Average positions gained from grid
  • GridPosition - Current race starting position

​
Evaluation Visualizations

The training process generates detailed evaluation charts:

​
Confusion Matrices

Location: models/confusion_matrices.png Shows prediction accuracy for both Random Forest and XGBoost models:
  • True Positives: Correctly predicted Top-3
  • True Negatives: Correctly predicted outside Top-3
  • False Positives: Incorrectly predicted Top-3
  • False Negatives: Missed Top-3 predictions

​
Feature Importance Charts

Location: models/feature_importance.png Horizontal bar charts showing:
  • Top 10 features by importance score
  • Comparison between RF and XGBoost feature rankings
  • Relative contribution percentages
Confusion matrices and feature importance plots are generated after running:
python train_all_models.py
Ensure models are trained before viewing visualizations.

​
Model Performance by Conditions

​
Weather Impact on Accuracy

Accuracy: 82%Highest accuracy in dry races where grid position dominates:
  • Pole position win rate: 42%
  • Top 3 grid → 65% podium rate
  • Predictable tire strategies

​
Key Performance Insights

Grid Position Dominance

Pole position accounts for 35% of model importance - where you start matters most!

Rain Equalizer

Wet races increase prediction uncertainty by 15-20% but create opportunities for underdogs.

Team vs Driver

Team performance (12%) + Driver skill (18%) = 30% combined importance. Both matter significantly.

Recent Form

Last 5 races account for 10% importance - momentum and confidence are real factors.

​
Limitations & Future Improvements

  1. Limited to 2018-2024 data - Only 7 years of historical races (~140 events)
  2. No qualifying data - Grid position used instead of qualifying times
  3. Basic weather modeling - Binary rain indicator rather than detailed conditions
  4. No safety car events - Race interruptions not modeled
  5. Static tire strategy - Rule-based rather than ML-predicted
  • Add qualifying telemetry - Sector times, speed traps, mini-sector analysis
  • LSTM for tire degradation - Time-series modeling of compound performance
  • Neural networks - Deep learning for complex feature interactions
  • Safety car prediction - Probability model for race interruptions
  • Real-time updates - Live race prediction updates during sessions

​
Model Files & Locations

Saved model files in models/saved_models/:
  • winner_predictor_rf.pkl - Random Forest model (V1)
  • winner_predictor_xgb.pkl - XGBoost model (V1)
  • winner_predictor_v2.pkl - Enhanced ensemble model (V2)
  • feature_columns.pkl - Feature list for V1
  • feature_columns_v2.pkl - Feature list for V2

​
Next Steps

1

Train Models

Generate fresh models with latest data:
python train_all_models.py
2

Evaluate Performance

Review confusion matrices and feature importance:
open models/confusion_matrices.png
open models/feature_importance.png
3

Test Predictions

Use the web dashboard to validate predictions:
python src/app.py
# Open http://localhost:5000

Build docs developers (and LLMs) love

Get started for freeTalk to us