In ice hockey, like in most sports, games are more exciting if the contestants have about the same proficiency. To this end, considerate distribution of the registered teams to inter-matching groups (divisions) is essential. In professional sports, they usually have well-devised rules for promoting or relegating the teams, but in most amateur leagues these are not applicable, foremost due to lower continuity. Amateur players might skip seasons; teams get formed, reformed and disbanded more often, and statistics on their previous performance is not always available. In our paper, we describe and evaluate an algorithm for the distribution of teams to divisions in Novosibirsk Ice Hockey Amateur League, which is based on a specially devised rating calculation mechanism. It involves the set of the players’ statistics, with the relative weights for different positions obtained from the survey we ran with 25 experts. We further benchmark several machine learning models trained on the 1212 historic games and tested on 305 recent games data. The evaluation suggests that our algorithm yields 10–30% more “fair” team distribution compared to the manual one previously done by the League.