How chess ratings actually work

One number, one job

A chess rating has exactly one purpose: to predict the result of a game between two players. That is the whole design. It is not a score, not a measure of how much you know, and not a running total of anything — it is a prediction, and it is updated when the prediction turns out to be wrong.

Arpad Elo's insight was that only the difference between two ratings matters. A 200-point gap means the stronger player is expected to score about 76% across many games, whether the pair are rated 1000 and 1200 or 2400 and 2600. The absolute numbers are arbitrary; the distance between them is the claim.

So a rating cannot be earned or lost in the way points are. It is re-estimated after every game, and the size of the change depends entirely on how surprising the result was. Beating someone far above you moves it a long way, because the system had been confidently wrong about you. Beating someone far below you moves it barely at all, because it had already predicted that.

Why your rating moves by different amounts

Two things govern the size of a change: how unexpected the result was, and how confident the system is about you. The first is the rating difference. The second is what most people never see, and it is where the modern systems part company with Elo's original.

Lichess uses Glicko-2, which tracks not just your rating but its uncertainty — a rating deviation, alongside a volatility figure. A player the system is unsure about moves in large steps. A player it has watched for a thousand games moves in small ones, because it would take a great deal of evidence to overturn that much history.

This is why a long break can make your rating feel unstable when you return. The deviation grows while you are away, on the reasonable assumption that the system knows less about you than it did. It is not punishing you for the absence; it is admitting it has stale information, and the larger swings are it catching up.

Chess.com uses a Glicko variant of its own, and FIDE still uses a K-factor system much closer to Elo's original — a fixed multiplier that steps down as a player becomes established and higher rated. Different machinery, same idea: be less certain about newcomers, more certain about veterans.

Provisional ratings, and the doubt inside them

A new rating is a guess with almost nothing behind it. Both major sites mark this — Lichess with a question mark beside the number, Chess.com by withholding some features until enough games are played — and the mark means the deviation is still too large for the number to be treated as settled.

It is worth being precise about what provisional does and does not mean. It does not mean the rating is wrong; it means the system is not yet willing to defend it. The estimate may already be close. What is missing is evidence, and only games supply that.

The practical consequence is that a provisional rating should never be compared with an established one, in either direction. This site refuses to give a percentile for a provisional rating for that reason. Placing an unsettled number inside a distribution of settled ones produces a figure that looks precise and means nothing, and a number that looks precise is harder to disbelieve than one that admits it is unsure.

What a rating is not

It is not transferable. A rating is a statement about your results against one specific pool of opponents, and moving pools invalidates it — which is why the conversion between two sites is an approximation with real error bars rather than a formula.

It is not a measure of knowledge. Players who can explain far more about chess than their opponents lose to them regularly, because the rating measures results and nothing else. This is a feature. Any system that tried to rate understanding would have to decide whose understanding counted.

And it is not stable enough to read day to day. Ordinary variance in a system like this produces swings of fifty points or more in either direction over a short run of games, with no change in strength whatever. If you want to know whether you are improving, look at a few hundred games, or look at where you sit in the population rather than at the number itself.