2018 NCAA March Madness Men's Basketball Predictions

By Brice Walker

View full project on nbviewer

The Bracket (Go Bulldogs!)

Outline

• Getting started
• Introduction
• Feature extraction and engineering
• Classification analysis

Getting Started

I have provided a few handy scripts to allow quickly running this on your own. This project was built/tested on python 3.6 and it is recommended to use 3.6+. It is also recommended that you use Anaconda. This project requires Jupyter Notebook.

Simply run the following codes from a terminal in the repo directory:

pip install requirements.txt
python process_data.py
python train_model.py

Note: You may need to install XGBoost from source.

You may also need to run pip uninstall tensorflow to uninstall tensorflow and then run pip install tensorflow-gpu in order to take advantage of TensorFlow's GPU acceleration.

I have also provided a jupyter notebook that walks you through the iterative process.

Note: This project has been optimized for deployment on Clusterone.

To run this project on Clusterone, use clusterone.py

Introduction

This is a classification project completed for the 2018 March Madness Kaggle Competition. In this project, I have extracted 18 season-based, and 28 tournament-based team-level characteristics from several datasets using data from 1994-2017. I used datasets provided by kaggle, as well as data scraped from sports-reference.com. I then engineered several advanced measures and extracted Elo ratings. I used these characteristics to predict probabilities for each matchup in the 2018 March Madness Schedule. I then created a well calibrated soft voting classifier that used KNeighbors, Random Forest, Extra Trees, Logistic Regression, Gradient Boosting, and LightGBM classifiers to predict probabilities for each matchup. I also trained a Keras/TensorFlow based Neural Network. Finally, I developed predictions based on Microsoft's TrueSkill rating system and weighted them with the machine learning model predictions.

Feature Engineering and Extraction

This project attempts to predict outcomes of games based on the following team level characteristics for season games:

• A modified Elo rating (where new entrants are initialized at a score of 1500, and there is no reversion to the mean between seasons)
• Number of wins
• Avg points per game scored
• Avg points per game allowed
• Avg # of 3 pointers per game
• Avg turnovers per game
• Avg Assists per game
• Avg rebounds per game
• Avg steals per game
• Power 6 Conference
• Reg Season championships
• Strength of team's schedule
• Championship appearances
• Location of the game
• A simple rating system

And the following team level characteristics for tournament performance:

Note: If a team plays in more than one tourney in a year than these values are averaged over all tourneys they played that year.

• Tournament appearances
• Conference tournament championships
• Points scored for winning/losing team
• A measure of possession
• Offensive efficiency
• Defensive efficiency
• Net Rating (Offensive - Defensive efficiency)
• Assist Ratio
• Turnover Ratio
• Shooting Percentage
• Effective Field Goal Percentage adjusting for the fact that 3pt shots are more valuable
• FTA Rating : How good a team is at drawing fouls.
• Percentage of team offensive rebounds
• Percentage of team defensive rebounds
• Percentage of team total rebounds

Finally, predictions using these features are weighted and stacked with predictions made through team trueskill ratings.

Classification Analysis

Predictive binary classification statistical models explored in this project include:

• Logistic Regression
• K-Nearest Neighbors
• Random Forests
• Extra Trees
• Support Vector Machines
• Gradient Boosting
• TensorFlow/Keras Neural Networks
• Principal Component Analysis
• Ensembling/Stacking and Weighting Models

Libraries

Machine learning libraries used in this project include:

• Sci-Kit Learn
• XGBoost
• Light GBM
• Keras
• TensorFlow