#🏀 March Machine Learning Mania 2025 – NCAA Tournament Prediction

📌 Overview

This project is my submission for Kaggle's March Machine Learning Mania 2025 competition. The goal is to predict the probability that one NCAA basketball team will beat another in the tournament, using historical data and advanced machine learning techniques.

🛠️ Project Workflow

1️. Data Exploration

• The dataset consists of 35 CSV files, including detailed and compact results of regular seasons and tournaments, team seeds, and sample submission files.
• Key datasets used:
  • Regular Season Detailed Results (Men & Women)
  • NCAA Tournament Detailed Results (Men & Women)
  • NCAA Tournament Seeds

2️. Feature Engineering

We engineered meaningful features to improve model predictions, including:

• SeedNumDiff: Difference in tournament seed numbers between teams.
• WinRateDiff: Difference in win rates between teams.
• AvgPointDifferentialDiff: Difference in average point differential.
• AvgPointsScoredDiff & AvgPointsAllowedDiff: Differences in offensive & defensive performance.
• EloRatingDiff: Difference in team Elo ratings.
• TourneyWinPctDiff: Difference in past tournament win percentages.

3️. Feature Importance Analysis

We analyzed feature importance using permutation importance to identify the most influential features:

• EloRatingDiff and SeedNumDiff emerged as the most important predictors.

4️. Model Training & Evaluation

Approach 1: Baseline Models

• Logistic Regression
  • Brier Score: 0.2498
  • ROC AUC Score: 0.5355
• Random Forest
  • Brier Score: 0.2088
  • ROC AUC Score: 0.7386
• XGBoost Classifier
  • Brier Score: 0.1813
  • ROC AUC Score: 0.8024

Approach 2: Fine-Tuning XGBoost

• Hyperparameter tuning was applied to optimize XGBoost.
• Achieved:
  • Brier Score: 0.3304
  • ROC AUC Score: 0.7313

Approach 3: Fine-Tuning RandomForest (Best Model)

• Hyperparameter tuning was applied to optimize RandomForest.
• Achieved**:**
  • Brier Score: 0.2098
  • ROC AUC Score: 0.8763

5️. Final Submission Preparation

• Used the SampleSubmissionStage2.csv to prepare matchups.
• Applied the best trained model to predict win probabilities.
• Ensured submission format matches Kaggle’s requirements (ID format: 2025_TeamA_TeamB).

---

📈 Key Learnings & Challenges

✅ Importance of feature engineering in sports analytics. ✅ Season feature was problematic for predictions. ✅ Classification models vs. probability-based submissions.

🚀 Next Steps

• Try LightGBM & CatBoost for comparison.
• Experiment with deep learning approaches.
• Improve generalization to avoid overfitting.

---

🏆 Conclusion

By focusing on feature engineering, feature selection, and model optimization, I significantly improved my prediction accuracy. This project highlights how data-driven insights can be used in sports analytics & tournament predictions.

📂 Full Code & Notebooks available in this repo! 🔥