Football prediction model based on the teams' Elo ratings and scoring indicators

The study focuses on proposing a solution for the 2023 Soccer Prediction Challenge organized in conjunction with the Machine Learning Journal's special issue on Machine Learning for Soccer. The challenge aimed to predict the outcomes of future matches from various leagues worldwide within a specific timeframe. In this paper, we examine the solution provided by our team "Friends of Elo" in detail. We experimented experimented with using Elo ratings and scoring indicators (goals) as arguments for the linear regression method. Poisson distribution was employed to predict the match results. The performance was evaluated based on the root mean squared error and the ranked probability score. Submitted predictions (714 matches) ranked 7th among 11 contestants in Task 1 and 5th among 13 contestants in Task 2, respectively. Futhermore, we conducted additional tests, where our model performed even greater on the expanded dataset of over 6800 matches to predict. The most significant advantage of our approach is that it does not require advanced match data, making it applicable worldwide. Overall, our study provides an in-depth analysis of the solution proposed by the "Friends of Elo" team, and we offer a superior alternative that performs well and is highly practical.