🏆 NCAA Tournament Prediction Pipeline

An end-to-end Machine Learning pipeline designed to predict game outcomes for both the Men's (MNCAA) and Women's (WNCAA) Basketball Tournaments. By combining historical tournament seeds, regular-season performance metrics, and advanced scoring differentials, this project trains an optimized XGBoost classifier to estimate the probability of match victories.

---

📌 Project Overview & Purpose

March Madness is famous for its unpredictability. The goal of this project is to model and predict the probability of team victory in tournament matchups by:

1. Unifying Data: Merging historical regular-season and tournament game logs for both men's and women's divisions.
2. Feature Engineering: Designing strong baseline estimators based on seed advantages, scoring power, and defensive resilience.
3. Optimized ML Modeling: Implementing an XGBoost pipeline utilizing imputation and scaling, validated via robust cross-validation.

---

📊 Feature Engineering System

Rather than passing raw scores which are unavailable at prediction time, the model relies on seasonal aggregated stats. Three primary features are calculated for each matchup ($Team_1$ vs $Team_2$):

Feature	Description	Strategic Representation
SeedDiff	$Team_1\text{ Seed} - Team_2\text{ Seed}$	Captures tournament-ranking disparity.
AvgPointsDiff	$Team_1\text{ Avg Points Scored} - Team_2\text{ Avg Points Scored}$	Captures offensive strength and scoring efficiency.
AvgPointsAllowedDiff	$Team_1\text{ Avg Points Allowed} - Team_2\text{ Avg Points Allowed}$	Captures defensive resilience and suppression ability.

Note: Team order is normalized ($Team_1 < Team_2$ based on ID) to ensure order-invariant predictions.

---

⚙️ Model & Evaluation Architecture

The model utilizes a Scikit-Learn Pipeline for clean preprocessing and training isolation:

1. Mean Imputer: Handles missing or new team stats with baseline averages.
2. Standard Scaler: Normalizes feature scales for reliable learning.
3. XGBoost Classifier: Configured with light regularization:
   • max_depth=3 (to prevent overfitting on high seed/point variance)
   • learning_rate=0.1
   • n_estimators=100

Validation Strategy: 5-Fold Stratified K-Fold

The pipeline evaluates model generalizability across 5 folds, preserving target class distributions.

Performance Metrics Obtained:

• 📉 Average Log Loss: 0.5027
• 🎯 Average Brier Score: 0.1677
• 📉 Average Mean Squared Error (MSE): 0.1677
• 📈 Average ROC AUC: 0.8312

---

📁 Repository Structure

.
├── NTLML_NACC_Tournment.ipynb  # The core Jupyter Notebook / pipeline runner
├── README.md                   # Project documentation (You are here!)
└── submission1.csv             # Final probability predictions for submission

---

🚀 Setup & Execution Guide

Prerequisite Dependencies

Install the required packages using pip:

pip install pandas numpy scikit-learn matplotlib xgboost

Dataset Structure

The notebook expects data to be organized in a directory structure (typically in Google Drive for Colab runtimes). Ensure files from the Kaggle March Madness dataset are present:

• MNCAATourneySeeds.csv & WNCAATourneySeeds.csv
• MRegularSeasonDetailedResults.csv & WRegularSeasonDetailedResults.csv
• MNCAATourneyDetailedResults.csv & WNCAATourneyDetailedResults.csv
• SampleSubmissionStage1.csv

Running the Predictor

You can execute the pipeline directly in the notebook:

1. Open NTLML_NACC_Tournment.ipynb in your notebook editor (Google Colab or Local Jupyter).
2. Configure the directory path:

   predictor = TournamentPredictor(data_dir="/path/to/your/NACC-DATA_2025/**")

3. Run all cells. The pipeline will automatically load datasets, extract features, train with cross-validation, plot evaluation graphs, and generate submission1.csv.