Forecasting NCAA basketball tournament outcomes using machine learning models to support bracket predictions and sports analytics.
The aim of this competition was to predict the outcomes of all potential matchups in the 2025 NCAA Division I men’s and women’s basketball tournaments. Each submission included predicted probabilities for every possible matchup that could occur in the tournament brackets.
Hosted on Kaggle as part of the annual March Machine Learning Mania, submissions were evaluated using the Brier score, which penalizes inaccurate probability predictions. Lower Brier scores indicate better predictions.
You can explore the complete methodology in this notebook: 🔗 MMLM25 - EDA and Prediction
Key steps followed:
-
📊 Exploratory Data Analysis (EDA):
- Understood tournament structure, team seeding, and regional splits.
- Analyzed match histories and evaluated metrics such as
ScoreDiff,SeedDiff, and team-level historical performance. - Assessed other features such as win-loss records, and scoring statistics.
-
🧠 GOTO Model:
- Utilized the goto_conversion method to convert betting odds into outcome probabilities.
- This method adjusts inverse odds by the same units of standard error, effectively addressing the favourite-longshot bias by considering the proportionately wider standard errors implied for inverses of longshot odds and vice versa.
-
🧪 Prediction & Submission:
- Mapped test data using team identifiers from the sample submission file.
- Applied the converted probabilities to predict outcomes for every possible tournament matchup.
- Ensured that the predictions accounted for the inherent biases in betting markets, leading to more accurate forecasts.
-
✅ Public Leaderboard Scores:
- Achieved 0.12453 and 0.10921 Brier scores.
-
🏁 Private Leaderboard Score:
- Final score: 0.10921 on hidden test data.
-
🥇 Rank Achieved:
- Placed 53rd out of 1,986 participants and 1,727 teams as a solo participant.
- 🏆 Kaggle Competition: March Machine Learning Mania 2025
- 📁 Dataset: Competition Data
- 🗒️ Helpful Notebooks:
- 📦
goto_conversionPackage: GitHub repo
- Language: Python 🐍
- Libraries:
pandas,numpyfor data manipulationmatplotlib,seabornfor visualization
- Tools:
- Jupyter Notebook 📓 for development
- Kaggle Kernels / Google Colab for experimentation and submission
📌 This project highlights the effectiveness of advanced odds conversion techniques, like goto_conversion, in improving the accuracy of probabilistic forecasts for sports tournaments.