Introduction

My project is about football and specifically my project is about betting on football. Football betting is perfectly situated to be solved by machine learning algorithms. Combine more data available than a human mind could possibly analyze with a discrete goal and you have the perfect machine learning problem.

Sports betting may seem complicated but it is actually quite simple. In football betting there is what’s referred to as a three way bet. You can bet on the home team to win, the away team to win, or a draw. That is the type of outcome I will attempt to predict. There is however another component to the bet, the most central component, odds. Odds are the means by which a sports better can calculate the take and risk a corresponding amount. A heavy favorite will payout considerably less for the same risk than an underdog or a team who is not expected to win. Therefore the confidence level needs to be much higher when taking a bet on a heavy favorite.

I chose this project because I am passionate about sports. Sports are the initial reason I became interested in data science. Data and sports is a marriage made in heaven. Data affects everything in sports from the game theory, to roster construction strategy and even on to handicapping and betting sports. The best part about this marriage is that the data is publicly available and mostly free.

The research question is what inputs need to be combined in order to create a model using machine learning or conventional programming in order to produce a profit. My hypothesis is that efficiency over average is a key predictor of outcomes and therefore can be leveraged to create a profitable model. Efficiency over average is a term coined by Football Outsiders a group name so because they took an approach to analyzing the game without the bias an insider would have. They state concerning their method “[It] is a method of evaluating teams, units, or players. It takes every single play during the NFL season and compares each one to a league-average baseline based on situation. “ [1] Applying their approach to Premiere League Football I could take two goals. One from Team A against Team B, another from Team C against Team D. Which was the better goal. That is the question they answer. Let’s assume Team D had a significantly better defense than Team B. Therefore Team C’s goal was better than Team A’s in that specific instance.

State of the Art

I researched the state of the art then I analyzed two approaches to the problem one using conventional programming and one using machine learning. They articles can be found on the links section of this page.

The best article I came across in my research for summarizing the state of the art is titled, A machine learning framework for sport result prediction.[2] In section 2 of the article 'Literature review and critical analysis' the authors Bunker and Thabtah, discuss different studies that approach sports predictions using classification through Artificial Neural Networks (ANNs). The oldest approach they discussed was done by Purucker in 1996. He used five features from 8 rounds of games and achieved 61% accuracy in predicting outcomes compared to 72% accuracy from subject matter experts. Purucker used an ANN with a backward-propagation (BP) algorithm to create his model. In 2003 Kahn was able to beat Purucker’s accuracy result with a 75% accuracy score which beat subject matter experts score of 63%. Khan used 6 features and used a longer training period going 14 rounds instead of 8. Mcabe and Trevathan built on that work. They used, “A multi-layer perceptron, trained with BP and conjugative-gradient algorithms was used. The ANN had 20 nodes in the input layer, 10 nodes in the hidden layer, and 1 node in the output layer (20-10-1).”

The most novel approach that Bunker and Thabtah discuss is one performed by Tax and Joustra. They used betting odds alone to predict outcomes rather than betting odds and past performance. They used a few different types of classifiers. The naive Bayes and ANN classifiers performed the best correctly predicting outcomes at a clip of 54.7%. So using only betting odds one can predict an outcome of an event at a better than random but not as well as the previously discussed approaches.

My approach will be to use ANN classification with BP and see if the model I create can match the performance of the models previously discussed.

I went on to analyze the code for two approaches:

The conventional programming approach was described thusly: “They propose that you can look at past results and scores between different teams within the same league system and from these past results be able to predict future scores and results.” [3]

They built a basic version of the model using csv, math, ast, and numpy in Python. The model returned a 5% return on investment in the first season tested but when taken in snippets or other seasons the return can be negative.

This is likely because the approach is a bit too simple.

The machine learning approach leveraged ANNs with Keras. That approach returned 5% in training sets, and 8% in testing sets. This approach works but there is always room for other profitable models. Especially in a landscape like handicapping where sharp betters and odds makers constantly have to adjust to minimize inefficiencies. A model that works for odds now may not work next year if the odds makers change their approach.

The Data I will use to train and test my model is game results from the last four seasons of English Premier league soccer. The data includes date, teams, scores, results, odds, and other match statistics. Each set contains information for all 381 games making the total set over 1,500 rows of data. The data as well as a notes document defining the values can be found here in my github repository.

Initially I did some simple exploration on the data that can be found in my github repository. Also as a proof of concept and a practice exercise I took a snippet of code from the conventional programming approach and replicated it using Pandas. This allowed me to test a very simple theory of taking home dogs. Or taking bets where the home team has higher paying odds than the away team. The theory did pay off with the test set that I used, but was very high in variance, so not a suitable long term approach.