What is the quantitative predictive value of Elo rating for tournament performances? (Or, same, what is the natural strength variabitity of an average player, measured in Elo?)

A quantitative version, so to say, of this question, but the answers are unsuitable, although #3 points in the right direction.

The question has some practical value: assume you get 4.0 points where your Elo predicted 3.0 (D=H-E=+1.0) - are these to read as, on some arbitrary confidence level, rather as 3.0 ± 2.0 (you were just lucky) or 3.0 ± 0.5 (you played good)?

First of all, the Elo rating is a snapshot in time, averaged out a bit by past tournaments, and essentially assumed to be a constant measure of strength. We thus should expect that juniors and seniors in average have larger differences D.

Let's look at a (somewhat) random tournament (pre-Corona by suggestion of Brian), German championship 2015, and observe the last column, which measures the difference D of predicted and actual performance. Mean is ~0 (duh), standard deviation s=~13, interval [-32,+21]. I've seen many such columns and it looks pretty "normal" to me.

Thus two natural questions:

  1. How large is s for an "average" tournament? (I.e. pick a large number of random tournaments, compute s for each, compute the average of that.) My own estimate is around 20.
  2. Of course there are no "average" tournaments anyway - for example, I expect that s will be far higher in a lower Elo tier. Is that correct?
  • 1
    Are you asking for a confidence or a prediction interval (not the same)? What you provide as an example is not a "variance" mathematically. Commented Aug 1, 2022 at 20:18
  • USCF makes use of this for rating bonuses when a person does significantly better in a tournament then predicted. Their opponents are also related in an addational calculation phase using the fast improving players post tournament rating rather then published rating. Commented Jun 1 at 9:56


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.