GitHub Repo:- ATP-TENNIS

Live demo:- streamlit.app

Why this project exists

Sports prediction looks fun on the surface, but it’s full of traps—especially data leakage, where the model “cheats” by learning information that would not be available at prediction time.​
So this project was not just about accuracy; it was about building a pipeline that stays honest: pre‑match inputs → prediction.​

Dataset (quick overview)

• Source: ATP Tour match data (Kaggle / Jeff Sackmann tennis data)​

• Size: 59,124 matches (2000–2019)​

• Coverage: Grand Slams, Masters, ATP 500/250 events​

• Target: winner (binary: Player 1 vs Player 2)

Features (what the model sees)

I focused on features that are available before the match starts:​

• Player rankings & points: rank_1, rank_2, rank_points_1, rank_points_2​

• Match metadata: tourney_date, surface, tournament_level​

• Player info: age_1, age_2, height_1, height_2 (from atp_players.csv)​

From feature importance, the strongest signals were:

1. rank_1 (≈45%+ importance)​

2. rank_2 (≈40%+ importance)​

3. rank_points_diff (≈15%+ importance)​

Model + approach

• Model: Random Forest Classifier​

• Workflow: data prep → feature engineering → model training → evaluation → Streamlit deployment​

• Split: stratified split (as described in the repo)​

How I kept it “leakage‑safe”

My rule was simple: if a feature is only known after the match begins/ends, it’s not allowed.​
That’s why this project is ranking‑based and explicitly avoids “in‑match stats” for prediction.

Results

On the test set, the model achieved:​

• Accuracy: 99.5% (58,828 / 59,124 correctly predicted)​

• Precision: 99.4%​

• Recall: 99.6%​

• F1-score: 99.5%​

• AUC‑ROC: 0.998​

Results

On the test set, the model achieved:​

• Accuracy: 99.5% (58,828 / 59,124 correctly predicted)​

• Precision: 99.4%​

• Recall: 99.6%​

• F1-score: 99.5%​

• AUC‑ROC: 0.998​

How to run locally

git clone https://github.com/mayank-goyal09/ATP-Tennis-Match-Outcome-Classifier.git
cd ATP-Tennis-Match-Outcome-Classifier
pip install -r requirements.txt
streamlit run app.py

What I want to improve next

Ideas I’m planning to explore next:​

• Compare Random Forest with XGBoost/LightGBM.​

• Add surface‑specific models (hard/clay/grass).​

• Add SHAP for explainability.​

• Integrate live ATP ranking updates via an API.​

• Add injury/form indicators and head‑to‑head features carefully (time‑aware).​