Key points are not available for this paper at this time.
The main goal of this article is to compare the performance of team ratings and individual player ratings when trying to forecast match outcomes in association football. The well-known Elo rating system is used to calculate team ratings, whereas a variant of plus-minus ratings is used to rate individual players. For prediction purposes, two covariates are introduced. The first represents the pre-match difference in Elo ratings of the two teams competing, while the second is the average difference in individual ratings for the players in the starting line-ups of the two teams. Two different statistical models are used to generate forecasts. The first type is an ordered logit regression (OLR) model that directly outputs probabilities for each of the three possible match outcomes, namely home win, draw and away win. The second type is based on competing risk modelling and involves the estimation of scoring rates for the two competing teams. These scoring rates are used to derive match outcome probabilities using discrete event simulation. Both types of models can be used to generate pre-game forecasts, whereas the competing risk models can also be used for in-game predictions. Computational experiments indicate that there is no statistical difference in the prediction quality for pre-game forecasts between the OLR models and the competing risk models. It is also found that team ratings and player ratings perform about equally well when predicting match outcomes. However, forecasts made when using both team ratings and player ratings as covariates are significantly better than those based on only one of the ratings.
Arntzen et al. (2020) studied this question.