Using Expected Goals to Forecast the English Premier League

Abstract

This work aims to evaluate the expected goals and post-shot expected goals statistics and see if they can be used to predict future match outcomes using data from the English Premier League from 2017 to 2022. First, an exploratory data analysis was performed to see if expected goals accurately measures goal probabilities, showing that in general expected goals is an approximately unbiased estimator of observed goals but skilled teams and individuals are able to outperform their expected goals over an extended period. We then show how to incorporate expected goal data into existing Poisson regression models to forecast future match outcomes and scores by modifying the training process from Poisson maximum likelihood estimation to Gaussian maximum likelihood estimation on expected scores. The results show that the expected goals model forecasts future matches with significantly better accuracy than the observed goals model when training on a smaller sample and when accounting for time-varying team strengths. These forecasting models were then used to implement basic sports betting strategies by placing wagers on match outcomes and total goals scored, and results indicate that the expected goals model is able to generate profit while the observed goals model fails to generate profit for both types of betting. Finally, we show how the expected goals model can be modified to evaluate different hypotheses about the game of football by adding additional parameters to test whether strong teams tend to underestimate weak teams or whether England's ``Big Six" has a stronger home-field advantage than other teams in the league.