Final Design Report: https://drive.google.com/file/d/1GEfsQmXTFxKthtY8zBUR3wQsONwxa3rp/view?usp=sharing

Background

The project centered around predicting the outcomes of the NCAA Men's Basketball Tournament, commonly known as March Madness. Each year, 64 college basketball teams compete in a knockout-style tournament. The goal was to use machine learning to forecast match results and determine the probabilities of each team advancing through the tournament rounds. Historical datasets containing game statistics, seedings, matchups, and outcomes from the 2003 season onward were utilized.

Problem

The main challenge was to accurately predict the winners of each match in the March Madness tournament and estimate the probabilities of teams progressing through each stage. Given the nature of the tournament, which is prone to upsets and "Cinderella stories" where lower-seeded teams unexpectedly advance, creating a reliable predictive model was complex.

Solution

To tackle the problem, the team implemented two machine learning models: Logistic Regression and an Ensemble Method. The project involved several key steps and design decisions:

Data Collection and Preparation:
Data from the Kaggle competition "March Machine Learning Madness" was used. This included regular season and tournament statistics, team IDs, and initial matchups.
Feature Selection:
Key statistics such as points scored, field goals made/attempted, rebounds, assists, turnovers, steals, blocks, and personal fouls were selected to train the model.
Model Training:
- Logistic Regression: This generalized linear model was initially used to predict the probability of a team winning a match based on input features. It utilized the logistic function to transform the feature combination into a probability value between 0 and 1.
- Ensemble Method: After further research, the team opted for the Ensemble Method due to its potential for higher accuracy. This approach involves combining multiple models to improve prediction robustness and generalization.
Hyperparameter Tuning and Evaluation:
The team used pipelines, GridSearchCV, and cross-validation to tune the hyperparameters and evaluate the models. Balanced accuracy was the chosen metric due to the imbalanced nature of the dataset, ensuring fair performance across all classes.
Model Comparison:
- Logistic Regression: Fitted using maximum likelihood estimation, focusing on finding the optimal coefficients.
- Ensemble Method: More intricate, involving systematic training, tuning, and evaluation to prevent overfitting and enhance generalization.
Implementation Steps:
- Importing necessary libraries and data.
- Splitting the dataset into training and testing subsets.
- Creating models using pipelines with StandardScaler for normalization and specific algorithms (Logistic Regression, SVM, Random Forest).
- Performing hyperparameter tuning using GridSearchCV.
- Evaluating models using cross-validation and selecting the best based on balanced accuracy.
- Predicting outcomes on the test dataset using the best model and calculating the accuracy.

The final model, an ensemble, provided better accuracy and consistency compared to the initial logistic regression model, making it the team's choice for predicting the tournament outcomes

March Madness Machine Learning Prediction Model