How accurate is the ENSI.Score rating

Five to twenty minutes before a match starts, the probability that team 1 wins is computed from the two ENSI.Score ratings shown on the site at that moment and written down. It is never edited afterwards, and it is the number the match page prints.

p1 = 1 / (1 + 10 ^ ((r2 − r1 − 0) / 400))

r1 and r2 are the two ENSI.Score ratings at the moment the probability is locked. The published model is elo-v1: an offset of 0 and a scale of 400.

A match will count towards the figure when all of this is true of it, inside the 365 days the window covers:

• The match was played and ended inside the window, and the score names a winner.
• Draws and matches decided by a forfeit are left out: there was nothing on the server to predict.
• The prediction still describes the match. If the sides were swapped or the start moved after it was locked, it is marked and stops counting — and stays readable as the record of what was claimed.
• Nothing is measured on a match that was already over. The accuracy of a model is what it said beforehand, and only that.

The calibration bars are wide on purpose. Telling a group that says 70 % apart from one that wins 75 % of the time needs some 16 500 matches — about a year of them — and until there are that many the honest thing to print is the interval rather than a verdict.