Elo Rating
A skill number built up from head-to-head results
- Elo rating is a skill number built by stacking up head-to-head results, one match at a time.
- The winner takes points from the loser. How much changes with the gap between them.
- An unexpected result moves the score a lot. An expected one barely moves it.
- The number by itself means nothing. The gap between two ratings is what tells you the odds of the next match.
- Who you've played shapes your score. A skewed set of opponents skews the rating too.
Contents
1The analogy
A neighborhood spinning-top contest can't have everyone play everyone. Ten kids show up and the pairs needed explode, and who turns up changes every day. So instead, each kid carries around a single number.
Two kids play, the match ends, and the winner takes a few points from the loser. How many changes every time. Beat someone with a much higher number and you take a lot; beat someone lower and you take almost nothing. Lose, and you give up the same amount. An expected result barely shakes the numbers; a surprise shakes them hard.
Stack up enough matches this way, and everyone's number settles into place without anyone having to play everyone else. Elo rating is a number kept this way.
2In detail
The winner takes points from the loser
There's no separate pool handing out points. Points only move between the two players in a match, so whatever one side gains, the other side loses. Add up everyone's number, and the total never changes.
A newcomer starts with a number somewhere in the middle. It doesn't mean anything yet, and it settles into place over a few matches. Start lower than your real skill and a streak of wins climbs it fast; start higher and losses bring it back down.
That's why a rating built on few matches isn't trustworthy yet. It's why a leaderboard always shows the number of matches played alongside the rating.
An unexpected result moves the score more
The core of this system is that it sets an expectation before every match. It looks at the gap between two ratings and decides ahead of time who's likely to win. Then it moves points based on how far the actual result strayed from that expectation.
Two closely matched players play, and either winning is hard to call a surprise, so the points barely move. Two players far apart play, and the lower-rated one wins — that's a big miss on the expectation, so points swing hard.
So this number isn't the same as counting wins. Beating weak opponents over and over barely raises it; beating a strong opponent once raises it a lot. It's not how many you beat, but who you beat, that the number carries.
The gap carries meaning, not the number itself
The rating on its own is an arbitrary number. Set the starting value differently, and the whole scale shifts up or down. What carries meaning is the gap between two ratings. A wide gap means the higher side is likely to win; no gap, and it's close to even.
That's why ratings from different competitions can't be compared directly. The number only means something within the group that played each other. A high score from the neighborhood top contest and one from the next town over share a name but aren't the same ruler.
Looking at a single match, the higher-rated side doesn't always win either. The gap tells you the odds, not a guaranteed outcome. A much lower-rated side pulling off a win now and then is actually what you'd expect.
Who you've played shapes your score
Two players with the same real skill can end up with different ratings depending on who they've faced. Someone who's only played strong opponents gets pushed down; someone who's cherry-picked easy ones gets inflated. That's why matchmaking matters. Pairing similarly-rated players together often gives each match more information, and ratings settle into place faster.
If lots of new players are joining, or skill shifts over time, the numbers can lag behind. Someone's old high rating might stick around long after their play has dropped off. In those cases, older matches get weighted less, or ratings get recalculated on a schedule.
It's the same reason this method is so widely used to compare AI models. Have people pick the better of two answers, and matches pile up without anyone needing to build an answer key. There's no need to write a new test — a new model just gets dropped into the contest.
3More precisely
Elo rating sets a win probability from the gap between two ratings, then shifts points by however far the actual result missed that expectation. A single tuning value controls how big the shift is. Set it high, and the rating tracks recent results fast but swings around; set it low, and it stays stable but lags behind. A leaderboard built from human votes usually shows a margin of error alongside the number, since a small sample of matches leaves plenty of room for wobble.
The analogy breaks down in one place. A spinning-top match has a clear winner and loser, but which of two answers is better depends on who's judging. Someone who likes short answers, someone who likes a table — that taste bleeds into the verdict. So this leaderboard is closer to an order of what people preferred than a strict order of ability.
The method itself was built for matches between people, on the assumption that outcomes are clear-cut and skill changes slowly. A model's ability can shift overnight, and it can play unlimited matches, so that assumption doesn't hold perfectly here.
4Try it yourself
5Common misconceptions
It's easy to think a higher score means better at everything, but actually it's a tally of what the judges happened to prefer, so there can be specific areas it does worse in.
It's easy to think the rating number itself carries meaning, but actually the whole scale shifts with the starting value, and only the gap between two ratings means anything.
It's easy to think the higher-rated side always wins, but actually the gap only tells you the odds, so an upset now and then is normal.
7One-line summary
In shortElo rating is a number built by trading points across head-to-head matches, and it's the gap between two ratings, not the number itself, that tells you the odds of the next one.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02