Challengers

Predicting ATP tennis match outcomes with machine‑learning models so we (responsibly!) improve our odds when placing wagers.

---

Table of Contents

1. Project Motivation
2. Dataset
3. Problem Statement
4. Methods & Techniques
5. Project Plan
6. Team
7. Inspiration

---

Project Motivation

We enjoy betting on tennis but admit we are awful at picking winners. By building the most accurate predictive model we can, we hope to gain a data‑driven edge over gut feelings.

Dataset

• Source: Jeff Sackmann’s ATP match repository
• Data types: Categorical & quantitative match‑level records dating back to 1968.
• Key features under consideration:
  • Current ATP ranking of each player (pure skill indicator)
  • Career earnings & seasonal prize money
  • Historical win/loss counts
  • Game & set score differentials
• NOTE: You’ll need to make sure the years 2018‑2024 are in the directory named tennis_atp/..., or just clone the repo in the same folder.

Problem Statement

Can we accurately classify the winner of an upcoming tennis match using historical performance data? Success will be measured by overall accuracy and betting unit return on an unseen 2025 test set.

Methods & Techniques

This is a supervised classification task. We will experiment with:

• Logistic Regression
• K‑Nearest Neighbors (KNN)
• Decision Trees & Random Forests
• Gradient Boosting (e.g. XGBoost, AdaBoost)
• Regularized models (L1/L2)

For each algorithm we will perform hyper‑parameter tuning (grid/random search) and apply ensemble techniques such as bagging and boosting where applicable.

Project Plan

1. Data Cleaning & Pre‑processing
   ‑ Deduplicate matches, handle missing values, encode categoricals, remove or combine highly collinear variables.
2. Exploratory Data Analysis
   ‑ Summary stats & visualizations to understand distributions and relationships.
3. Model Training
   ‑ Train baseline models, evaluate with cross‑validation, iterate on feature engineering.
4. Hyper‑parameter Tuning
   ‑ Optimize promising models; evaluate on a 2024 hold‑out set.
5. Final Evaluation
   ‑ Test on 2025 matches and compare to the 83 % accuracy benchmark.
6. Deployment (Stretch Goal)
   ‑ Package the best model into a CLI or web app for real‑time predictions.

How Is It Run?

We use Jupyter Notebook to run this so just open the Challengers.ipynb through the notebook and run.

Team

Role	Member
Leader	Mathew Alangadan
Data Engineers	Ryder Keeney • Anton Salvador
ML Engineers	Dylan Tallon • Zachary Miguel Allarey

Inspiration

A video by AlgoBeans ("Building a Tennis Match Predictor – 83 % Accuracy", 2025) demonstrated an 83 % Random Forest model on 2025 data. We aim to beat that score and share our improvements.

---

Disclaimer: This project is for educational purposes only. Gamble responsibly.