This project was developed for the Kaggle March Machine Learning Mania 2025 competition.
It focuses on analyzing the data to gain the best insights and building predictive models that estimate the probability of one NCAA basketball team defeating another during the tournament.
The data used in this project is provided and pre-cleaned by Kaggle as part of the official competition.
It is structured into five major sections:
- Team Info: Historical teams and seasons (
MTeams.csv,MSeasons.csv) - Seedings: NCAA Tournament seedings for each team (
MNCAATourneySeeds.csv) - Compact Results: Regular and tournament win/loss summaries
- Detailed match results: Full box scores including rebounds, assists, steals, etc.
- Enables more granular statistical analysis of team and player performance
- Game locations: City-level data for where each game occurred
- Useful for modeling potential home-court or travel distance effects
- Massey Ordinals: Over 40 ranking systems, including Sagarin, Pomeroy, AP, etc.
- These rankings are tracked throughout the season and are a rich source of features
- Coaching records: Including season-by-season coach names and changes
- Conference affiliations, secondary tourneys, and tourney slots
- Extra depth for modeling team consistency, stability, and historical performance
✅ All datasets are cleaned and standardized for direct use — no additional preprocessing is necessary to begin analysis.
- Seeds & Upsets: Lower seeds dominate overall, but most upsets occur between seeds 9–13.
- Conference Trends: Certain conferences consistently under- or over-perform their seedings.
- Coach Analysis: Coaches with more career wins and clutch success reach deeper rounds more often.
- Ranking Systems: Aggregated system rankings (e.g., AP, Sagarin) offer predictive insight into team strength.
- Deep Runs: Teams with a history of late-stage appearances may carry hidden value beyond their seed.
- Engineered features from:
- Tournament history
- Elo ratings & ranking systems
- Seed difference, win margins, upset counts
- Coaching experience and clutch performance
- Boxscore aggregates (FG%, steals, blocks, etc.)
- Used LightGBM as the main classifier due to its speed and performance
- Trained on historical matchups and predicted probabilities for 2025 tournament games
Submissions are evaluated using the Brier score, which measures the accuracy of predicted probabilities — the lower, the better.
🎯 Our best submission achieved a Brier score of 0.126 on the men's bracket, beating the public benchmark with strong feature engineering and model design.
| File | Description |
|---|---|
ncaa_eda.ipynb |
Exploratory Data Analysis, visuals, and insights |
ncaa_feature_engineering_and_model.ipynb |
Feature engineering and LightGBM modeling |
README.md |
This documentation file |
This project was built by my amazing mate Dylan and me (Thao)!
We absolutely loved exploring this dataset — from the richness of the historical records to the excitement of seeing data turn into real March Madness insights.
🏀💥 We hope this work scores a three-pointer for you too.
We extend our gratitude to:
- Kenneth Massey for providing much of the historical ranking data
- Jeff Sonas of Sonas Consulting for his support in assembling the full competition dataset
And big thanks to the Kaggle competition organizers for making this incredible basketball dataset available to the public.