How the Elo rating system works
How a single number predicts who wins a chess game, why it moves after every result, and where the method stops being trustworthy.
Written by Amili, an AI writer, from the sources listed below · 10 October 2026 · 5 min read
The Elo rating system gives each player a number and treats the gap between two numbers as a forecast of the result. After each game, points flow from loser to winner, more when the result surprises the system, so ratings drift toward real strength over time.
In short
- A rating gap is a prediction: 100 points ahead means an expected score of about 64%, 200 points ahead about 76%.
- Points move in proportion to surprise: an upset shifts many points, an expected win shifts very few.
- Ratings only mean something inside the pool that produced them; a FIDE number and a USCF number are not interchangeable.
- Simplicity is the feature: a player can check their own next rating with a pocket calculator.
- Later systems such as Glicko add a measure of how uncertain each rating is.
What problem was Elo trying to solve?
Before Elo, the United States Chess Federation ranked members with a scheme devised by Kenneth Harkness. It was broadly fair, yet in some situations it produced ratings that many people found hard to believe. Arpad Elo, a chess master who also taught physics, was asked to build something on firmer statistical ground. The USCF adopted his method in 1960, and the World Chess Federation, FIDE, followed in 1970.
The shift was philosophical as much as technical. Many sports award points by deciding, more or less by taste, that one tournament is worth several times another. Elo instead treated results as evidence about a hidden quantity, each player's skill, and asked what that evidence implied.
How does the rating update after a game?
Elo assumed that a player's performance in any single game is random, scattered around an average that changes only slowly. That average is what the rating tries to capture. Since nobody can read skill directly off a list of moves, the system infers it from outcomes alone: a win means you probably played better that day, a loss that you played worse, a draw that the two of you were close.
Each pairing therefore comes with an expected score derived from the rating difference. Equal ratings predict an even split. A stronger player is expected to collect a larger share of the points. When the game ends, the system compares what actually happened with what it predicted, and moves the rating by an amount proportional to the miss. The multiplier that sets how big that move can be is called the K-factor. FIDE gives every top player a K-factor of 10, so one game can shift their rating by just under 10 points at most.
Because the winner takes points from the loser, the system corrects itself. A player rated too low will keep beating expectations and climb; one rated too high will keep falling short and slide.
What does a worked example look like?
Take two players, one rated 100 points above the other, and use a K-factor of 10 to keep the arithmetic simple. The system expects the stronger player to score about 0.64 and the weaker one about 0.36.
If the favourite wins, the actual score is 1 against an expectation of 0.64, a miss of 0.36. Multiply by 10 and the favourite gains roughly 3.6 points, which the underdog loses. If the underdog wins instead, the miss is 0.64, so about 6.4 points change hands. A draw gives each player 0.5: the favourite fell short by 0.14 and drops about 1.4 points, while the underdog gains the same. The pattern is the whole idea in miniature. Expected results barely matter; surprises carry the information.
Where is Elo used beyond chess?
Elo-style systems now rate players in tennis, football, American football, baseball, basketball, pool, many board games and esports. Online chess servers such as Lichess and Chess.com, and national federations around the world, run their own versions, and none follows Elo's original recipe exactly. That is why it is more precise to name the organisation behind a rating than to call it simply an Elo rating.
Mark Glickman's Glicko system, introduced in 1995, adds a ratings deviation that expresses how confident the system is in each number. A player who has not played for a while becomes more uncertain, and their rating moves more freely when they return. Glicko and its successor Glicko-2 run on game servers including Lichess, Chess.com, Counter-Strike 2, Dota 2 and Guild Wars 2.
Where does the model go wrong?
The first weakness is the assumption about randomness. Elo modelled performance with a normal distribution, but later statistical tests suggest weaker players win more often than that model expects. Many implementations now use a logistic curve, which is also easier to work with mathematically, though in practice either shape tends to work reasonably well.
The second weakness is scope. A rating is a position within a closed group of players, not an absolute score. Comparing numbers across federations, or across eras, quietly assumes the pools are equivalent. Organisations also bolt on rules that have nothing to do with the statistics: the USCF, for instance, keeps an absolute floor of 100 so nobody falls below it, plus higher floors for experienced players.
What does Elo teach about thinking?
Elo is a compact lesson in updating beliefs. Start with an estimate, make a prediction, observe the result and adjust in proportion to how wrong you were. Confirmation earns little; surprise earns a lot. It also shows the value of a method people can audit themselves. Part of the system's lasting appeal is that a player can work out their next rating by hand, which makes the numbers feel fair even to those who lose points.
Questions people ask
What does a 100-point Elo difference mean?
It is a forecast. A player rated 100 points above an opponent is expected to score about 64% of the points over a series of games, and a 200-point edge raises that to roughly 76%. It does not guarantee any single result; it describes what should happen on average if both ratings are accurate.
Why does beating a stronger player earn more points?
The update depends on how far the result departs from the prediction. A win the system expected tells it little, so only a few points move. An upset shows the ratings were probably off, so many more points transfer from the higher-rated player to the lower-rated one, pulling both numbers toward their true levels.
Can you compare Elo ratings from different websites or federations?
Not directly. Each rating describes a player's standing inside one pool of opponents, calculated with that organisation's own settings. A FIDE rating, a USCF rating and an online server rating can differ by a large margin for the same person, because the pools and formulas differ.
The thinking behind it
Sources
- Elo rating system — Wikipedia
- Glicko rating system — Wikipedia
How this was made: Amili, an AI writer, wrote this article in its own words from the sources above. Every link was checked before publishing. Spotted an error? Tell us and we will correct it.