Rating football teams of all amateur levels based on performance, An approach using implementation specific knowledge.

Abstract

The Royal Dutch Football Association is responsible for dividing all amateur teams in the Netherlands into groups of similar strength each season. The teams in these groups play games against each other, which forms the basis for future placement in higher or lower level groups. This dividing of teams is currently mostly done by hand. To facilitate the partial automation of this process, the current study presents several ways of creating a rating for all standard football teams in the Netherlands. This includes analysis of the Elo, Glicko2 and Elo++ rating systems and their relevance in the context of rating amateur football teams. The influence of implementation specific parameters on the rating is investigated and the analysis shows that home field advantage and goal difference are relevant in this context. Thirdly the closeness factor stemming from graph theory was analyzed for all teams, which showed a positive correlation between a low closeness factor (and thus a high connectedness with all other teams) and the performance of teams. Parameters describing the population density, club density and average disposable income in the area where a team is located, were not found to correlate with the rating of teams. Various models were trained based on the Elo, Glicko2 and Elo++ algorithms, which all showed an accuracy of about 66%. The addition of factors for the three relevant implementation specific parameters to the models improved their accuracy by maximally 0.5%.