🎾 Predicting Tennis Match Outcomes with ML

This project applies machine learning to predict the outcome of ATP Tour matches based purely on performance data, not by any measure of player identity. By transforming 20 years of ATP match data (pre-2023) into player-level statistical profiles, the goal was to see whether win/loss results could be explained — and predicted — through momentum, surface-specific performance, and match-level consistency.

Essentially an analytical dive into the dynamics of competition:

“If we strip away names and reputations, do the numbers still tell us who wins?”

---

Core Components

1. Data Engineering & Feature Design

• Ingested 20+ years of ATP Tour match data from Jeff Sackmann's tennis_atp dataset (thousands of matches).
• Aggregated and transformed raw match-by-match data into player-level statistical profiles (player_profiles.csv), tracking:
  • Surface-specific performance: Win %, break points saved/faced, serve/return stats for Clay, Grass, and Hard courts.
  • Momentum metrics: Rolling averages of recent form and consistency.
  • Contextual factors: Current ATP ranking, head-to-head records.
• Each match was reframed as a feature-based prediction problem:
  Player A (X_A stats) vs Player B (X_B stats) → Who wins?
• This abstraction allows the model to generalize beyond specific player identities.

2. Model Development

• Tested multiple classification models for match outcome prediction through iterative experimentation:
  • Logistic Regression – as a baseline interpretable model.
  • Random Forest – for capturing non-linear relationships and feature interactions.
• Performed hyperparameter tuning and cross-validation to optimize predictive accuracy.
• Final model saved as final_iter.pkl after multiple iterations of feature engineering and model refinement.
• Evaluated models on accuracy, precision, recall, and interpretability using historical ATP match data.

3. Interactive Bracket Predictions UI (Flask)

• Built toy Flask web application (app.py) that for demonstrating model capabilities.
• Users can simulate the Wimbledon 2023 tournament bracket (128 players) as an example.
• The app loads pre-trained models (models/final_iter.pkl) and player performance profiles to predict match outcomes round-by-round.
• Features an interactive UI showing bracket progression, demonstrating how player-level performance metrics drive probabilistic match predictions in a real tournament structure.

---

Running the Web App

To run the example interactive bracket simulator locally:

# Install dependencies
pip install -r requirements.txt

# Run the Flask app
python app.py

Then navigate to http://localhost:5000 to simulate the Wimbledon 2023 bracket and see the model predictions in action.

---

Key Insights

• Player momentum, game win %, and surface specialization were some of the stronger predictors of victory, beyond face-value measures like ranking alone.
• Models could reliably predict >70% of outcomes using pure statistical profiles.

---

Project Structure

tennis-ml/
 app.py                    # Flask web app for bracket simulation
 ml_tools.py               # Core model training, inference, and bracket logic
 models/                   # Trained models (final_iter.pkl, etc.)
 data/                     # ATP match data, player profiles, rankings
 notebooks/                # Exploratory data analysis & model experimentation
 templates/                # HTML templates for Flask app
 static/                   # CSS styling for web interface
 requirements.txt          # Python dependencies