ATP Match Predictions Model

Overview

The ATP Match Predictions Model predicts the outcomes (winner vs. loser) of tennis matches played by professional players on the ATP tour between 2018 and 2022, as well as on a separate unseen test set consisting of 2023 matches.

Data

The data was collected from January 2018 to December 2022, sourced from Jeff Sackman, who compiles comprehensive ATP match data. The dataset consists of 12,815 observations with 49 columns, covering match statistics such as match location, surface, points won, rankings, and player statistics.

Key Findings from Exploratory Data Analysis (EDA)

• Top Players: The top 20 players with the most match wins in the past five years include well-known names like Novak Djokovic, Rafael Nadal, and Roger Federer.
• Surface Impact: Win percentages vary by surface (hard court, grass, clay), with different players excelling on different surfaces.
• Player Height: Taller players tend to have stronger serves, though there is a tradeoff in terms of agility.

Preprocessing

• Data Removal: Matches from the NextGen Finals and Laver Cup were removed as these are exhibition matches and differ in scoring rules and player selection.
• Missing Data: Missing values were imputed for variables like seed and player height, and unhelpful columns were dropped, leaving 11,998 observations.
• Feature Extraction: Variables such as Elo ratings, serve statistics, break points, and match outcomes were created for each player.

Feature Selection

• Random Forest and Principal Component Analysis (PCA) were used to select the top 20 features for model training.
• PCA Scree Plot: An elbow was identified around 10 components, suggesting the optimal number of features.

Model Development

Preliminary Models Tested

• Random Forest with SVM (Radial Kernel)
• Random Forest with XGBoost

Hyperparameter Tuning with MLflow

Hyperparameters such as C, Gamma, Max Depth, and Eta were optimized for different models using MLflow.

Final Models

• Random Forest + SVM (Radial Kernel)
  • F1 Score: 0.738, AUC: 0.825, Accuracy: 0.75
• Random Forest + XGBoost
  • F1 Score: 0.817, AUC: 0.907, Accuracy: 0.828

Metrics Collected

The model's performance was evaluated using metrics such as F1-score, AUC, Precision, Recall, and Test Accuracy.

Results

• The model’s performance on the 2023 test set showed a F1 score range of 0.648 to 0.817 for different model configurations.
• Top performing model: Random Forest + XGBoost with optimal hyperparameters achieved an AUC of 0.907 and Test Accuracy of 0.828.

Model Limitations

• Data Drift: The model may face performance degradation over time due to the emergence of new players and the aging of older players, necessitating model retraining.

Deployment

The XGBoost model was deployed using the "plumber" package to create an API endpoint. The index.html file has the front end interface where two player names are entered and a winner is predicted.

Authors

• Mona Ascha (Author)

Last Updated

• November 7, 2024