tennis-betting-odds-model

🎾 Predicting Tennis Betting Dynamics Using Machine Learning

Regression, Classification & Feature Engineering on WTA 2007–2023 Match Data

Video: https://docs.google.com/videos/d/1i5HiqWHHtO9vakVhAM8epUm_P1hSEDB4ifv0_Oaf_Gs/edit?scene=id.p#scene=id.p

This project explores whether machine learning models can predict betting odds and betting risk levels from historical WTA tennis statistics.
Using a large-scale dataset (2007–2023), we built regression and classification models to understand how ranking, performance gaps, match context, and engineered features shape betting market behavior.

The final goal:
Can historical match statistics help identify betting patterns and support decision-making in real time?

---

1. Dataset Overview

Dataset used: WTA Tennis 2007–2023 Daily Update
Size: ~50,000 matches

Features extracted:

• Player rankings (Rank_1, Rank_2)
• Player points (Pts_1, Pts_2)
• Betting odds (Odd_1, Odd_2)
• Match attributes (Surface, Round)

Engineered features:

These features proved essential for improving model performance.

---

2. Exploratory Data Analysis (EDA)

We started by exploring the distribution of betting odds and competitive gaps.

2.1 Distribution of Pre-Match Odds (Odd_1)

This plot shows that:

• Most matches have low odds for Player 1 (strong favorites).
• There is a long tail of higher odds representing clear underdogs.

---

2.2 Distribution of Ranking Difference (Rank_Diff)

Interpretation:

• Many matches are played between players with small rank differences.
• Large positive or negative gaps are less common but very influential in betting.

---

2.3 Relationship Between Ranking Gap and Odds

This scatter plot shows:

• A clear relationship between Rank_Diff and Odd_1.
• When Player 1 is much stronger (large negative Rank_Diff), Odd_1 tends to be lower.
• When Player 1 is weaker (positive Rank_Diff), Odd_1 increases.

This confirms that ranking gaps are a key driver of betting odds.

---

2.4 Odds by Surface

We see that:

• Odds distributions differ slightly across Hard, Clay, and Grass.
• Some surfaces exhibit higher variance in odds, reflecting different levels of uncertainty.

---

3. Feature Engineering

We engineered several features to improve predictive power:

✔ Rank_Diff & Pts_Diff

Capture performance gaps between players — strong predictors of market expectations.

✔ Favorite_Flag

Binary indicator:

• 1 → Player 1 is the pre-match favorite (Odd_1 < Odd_2)
• 0 → Player 1 is the underdog

✔ Cluster_Label (K-Means)

K-Means clustering on:

• Rank_1, Rank_2, Pts_1, Pts_2

This groups matches into tiers (balanced, uneven, elite matchups) and adds a latent “match type” signal to the model.

All engineered features were inspected visually and statistically before modeling.

---

4. Regression Modeling: Predicting Betting Odds (Odd_1)

We trained three regression models to predict Odd_1:

1. Linear Regression
2. Random Forest Regressor
3. Gradient Boosting Regressor (🏆 Winning regression model)

Evaluation metrics:

• MAE (Mean Absolute Error)
• RMSE (Root Mean Squared Error)
• R² (Goodness of fit)

4.1 Feature Importance – Regression (Gradient Boosting)

We see that:

• Rank_Diff and Pts_Diff are among the most important predictors.
• Cluster-based features (Cluster_Label) and favorite information (Favorite_Flag) also contribute significantly.

Why Gradient Boosting Won:

• Best R² score
• Lowest errors (MAE/RMSE)
• Handles nonlinear tennis behavior better than linear models

The final regression model is saved as:
winning_regression_model.pkl

---

5. Converting Regression to Classification

To create a classification problem, we transformed Odd_1 into three balanced classes using quantiles:

• Low_Odds — strong favorites
• Medium_Odds — balanced matches
• High_Odds — clear underdogs

This created a well-balanced classification setup where each class has a similar number of samples.

We also discussed:

• Recall vs Precision:
  Recall is more important, because missing a high-value opportunity (false negative) is more costly than marking a match as risky when it is not.

---

6. Classification Modeling: Predicting Odd_Class (Risk Level)

We trained three classification models:

1. Logistic Regression
2. Random Forest Classifier
3. Gradient Boosting Classifier

Main evaluation metric:

• Macro F1 — to balance performance across all three classes.

6.1 Confusion Matrix – Winning Classification Model (Random Forest)

The confusion matrix shows that:

• Random Forest handles all three classes reasonably well.
• It reduces false negatives in the High_Odds class, which is crucial for not missing risky underdog situations.

---

6.2 Feature Importance – Classification (Random Forest)

Key insights:

• The same engineered features (Rank_Diff, Pts_Diff, Favorite_Flag, Cluster_Label) are highly influential.
• Match context (Surface, Round) also plays a role in predicting risk levels.

Why Random Forest Won:

• Highest Macro F1 score
• Most stable performance across all odds classes
• Best trade-off between recall and precision in the risk classification task

The final classification model is saved as:
winning_classification_model.pkl

---

7. Model Deployment

Both models were exported as .pkl pipelines:

• winning_regression_model.pkl
• winning_classification_model.pkl

They can be:

• Loaded in external Python environments
• Used for batch or near real-time predictions on new matches
• Integrated into a future sports analytics & betting decision-support system

---

8. Final Insights & Conclusion

✔ Can historical statistics predict betting odds?

Yes.
The regression model successfully learned the structure of betting markets using ranking, point differences, and match-level context.

---

✔ Can we classify matches into betting risk levels?

Yes.
The Random Forest classifier reliably distinguishes between Low, Medium, and High odds categories.

---

✔ Can this support real-time betting decisions?

Partially.

• These models work with pre-match data only.
• For full real-time betting, additional information is needed (live score, momentum, injuries, etc.).

Still, the models provide a solid foundation for pre-match risk assessment.

---

✔ What this project demonstrates

• Machine learning can meaningfully model betting markets
• Tennis statistics provide strong predictive signals
• Feature engineering dramatically improves performance
• Combining regression + classification creates a foundation for an automated betting recommender system

---

9. Repository Structure

Copy_of_Assignment_2_Classification,_Regression,_Clustering,_Evaluation.ipynb
1_distribution_odd1.png
2_distribution_rank_diff.png
3_rank_diff_vs_odd1.png
4_odd1_by_surface.png
5_confusion_matrix_rf.png
6_feature_importance_regression_gb.png
7_feature_importance_classification_rf.png
winning_regression_model.pkl   (optional alias: winning_model.pkl)
winning_classification_model.pkl
wta (1).csv
README.md