How Matchmaking Ratings Work

A new player's rating in most competitive games does not start at zero. It starts at an assumed average, and every match afterward nudges that number based on one question the system keeps re-asking: did this result match what the rating predicted? That gap between prediction and outcome is the entire mechanism behind matchmaking ratings, from Arpad Elo's original chess system to the Bayesian models running Valorant, League of Legends, Overwatch, and Xbox Live matchmaking today.

Where matchmaking rating actually came from

Rating systems for competitive games did not begin as a game design tool. They began as a statistics problem in chess administration. In the early 20th century, chess federations ranked players using systems that were essentially win-loss tallies with ad hoc adjustments, methods that produced inconsistent and easily disputed results, especially when comparing players who rarely faced each other directly.

Arpad Elo, a physics professor and chess master, proposed a fix in the 1960s that treated chess performance as a statistical variable rather than a fixed quantity. His model assumed that a player's performance in any given game is normally distributed around their "true" skill, meaning even a strong player has off games and a weaker player occasionally overperforms. From that assumption, Elo derived a formula for the probability that one player beats another purely from the numerical gap between their ratings, and a companion formula for how much a rating should move after the actual result comes in. The United States Chess Federation adopted his system in 1960, and the World Chess Federation (FIDE) adopted it in 1970, and the core Elo formula has been essentially unchanged in chess ratings since.

Elo's genius was not the specific numbers he chose. It was reframing "skill" as something you never observe directly, only infer from outcomes, an idea that turned out to generalize far beyond chess. Any competition with a single winner and a single loser per encounter can, in principle, be scored by an Elo-style system: table tennis leagues, competitive Scrabble, esports ladders, even some sports prediction markets. The system spread into video games specifically because early online multiplayer titles needed a cheap, well-understood way to sort players into competitive brackets without a human tournament director, and Elo's chess formula was public, simple to implement, and already validated on decades of real tournament data.

That history matters for understanding modern matchmaking, because every system used in games today, including the far more sophisticated ones covered below, is still answering Elo's original question. Given what I know about these players, how surprising was this result, and how should that surprise update what I know?

The core mechanism: predicted outcome versus actual outcome

The Elo formula works from one core calculation: given two players' ratings, compute the probability that the higher-rated player wins, then adjust both ratings based on whether that prediction turned out right or wrong. If a 1,600-rated player beats a 1,400-rated player, the system barely reacts, since that outcome was expected under the model. If the 1,400-rated player wins instead, both ratings shift substantially, because the system just learned its prediction was wrong and needs to correct.

The exact adjustment comes from a compact equation. A player's new rating equals their old rating, plus a constant called the K-factor, multiplied by the difference between the actual result (1 for a win, 0 for a loss, 0.5 for a draw) and the expected result the formula predicted before the game. The expected result itself comes from a logistic curve based on the rating gap between the two players: a 400-point rating advantage translates to roughly a 91% predicted win probability under Elo's original formula, a specific number chosen so that a "class" difference in chess (roughly 200 rating points, historically) produced intuitive win-rate expectations.

The size of each adjustment depends on two things: how surprising the result was, and that K-factor, which controls how aggressively the system reacts to new information. A high K-factor makes ratings swing quickly, useful for new players whose true skill is still unknown. A low K-factor makes ratings more stable, appropriate for established players whose rating is likely already close to accurate. FIDE has historically used a K-factor around 40 for new players with fewer than 30 rated games, dropping toward 10 for players rated above 2400, precisely to stabilize elite rankings while letting newcomers find their true level faster.

Elo's original formula assumes a 400-point rating gap between two chess players corresponds to roughly a 91% win probability for the higher-rated player, a specific constant baked into the math rather than a rounded estimate, chosen so ratings would line up with the informal "class" system chess federations already used.

This is where a useful analogy helps: think of a matchmaking rating less like a report card grade and more like a weather forecaster's probability estimate. A forecaster who says "70% chance of rain" is not claiming certainty. They are making a calibrated bet based on everything known so far, and if it doesn't rain, that does not mean the forecast was "wrong" in the way a wrong math answer is wrong. It just means a less likely outcome happened. A rating system works the same way: a prediction that the 1,600-rated player wins 75% of the time is not falsified by one loss. It's only falsified, in the statistical sense, if the 1,400-rated player starts winning far more often than the model expects across many games.

Why does my rating change more after some wins than others?

Because rating systems weight surprising results more heavily than expected ones. Beating a much higher-rated opponent, or winning while your own rating carries high statistical uncertainty, produces a larger rating change than beating an opponent your rating already predicted you would beat comfortably. The size of the swing is proportional to how wrong the system's prior prediction turned out to be.

Why Elo alone was not enough for modern games

Chess Elo has a structural blind spot: it treats every rating as a single fixed number with no attached confidence level. A player who has completed 5 rated games and a player who has completed 5,000 rated games can carry the exact same numerical rating, even though the system should be far more confident about the second player's true skill than the first. Elo has no mechanism to express that difference.

Mark Glickman, then a statistics researcher, addressed this directly with the Glicko system, introduced in the mid-1990s and later refined into Glicko-2. Glickman's core addition was a second number attached to every player: the rating deviation, or RD, representing one standard deviation of uncertainty around the rating estimate. A player with a rating of 1500 and an RD of 350 (roughly where new players start under Glicko) might have a true skill anywhere from about 800 to 2200 with reasonable statistical confidence, an enormous range. A veteran player with an RD of 50 has a true skill the system is confident lies within a much narrower band, something like 1400 to 1600 [Glickman, "The Glicko System," Harvard University / Boston University, glicko.net].

Glicko-2 added a further refinement Glickman called volatility: a measure of how erratically a player's results are swinging relative to what their rating and RD would predict, used to detect players whose true skill may be genuinely changing rather than just being poorly estimated. A player on a genuine hot streak, actively improving, looks statistically different from a player whose results are simply noisy, and Glicko-2's volatility term is designed to tell those two situations apart.

Microsoft Research went a step further with TrueSkill, published by researchers Ralf Herbrich, Tom Minka, and Thore Graepel in 2007 and built specifically for Xbox Live matchmaking [Herbrich, Minka & Graepel, "TrueSkill: A Bayesian Skill Rating System," Advances in Neural Information Processing Systems 20, 2007]. TrueSkill models each player's skill as a full probability distribution, not just a mean and a single uncertainty number, and it uses a technique called factor graphs with approximate message passing to update those distributions after every match. Crucially, TrueSkill was designed from the start to handle team games with more than two participants and variable team sizes, a case the original two-player Elo formula was never built to handle, since Elo assumes exactly one winner and one loser per match. The researchers validated the system using beta-test match data from Bungie's Halo 2, a genuinely team-based, many-versus-many shooter, before it became the ranking backbone for Xbox Live matchmaking.

Comparing the three major systems

The three systems share a common ancestor in Elo's core insight, but they differ meaningfully in what they track and what kinds of games they were built to serve.

System	Core mechanism	Handles uncertainty	Handles teams / multiplayer	Known real-world use
Elo	Single rating updated by predicted-vs-actual result, weighted by a K-factor	No	No (assumes one winner, one loser)	Chess (USCF, FIDE), many early esports ladders
Glicko / Glicko-2	Rating plus rating deviation (RD); Glicko-2 adds a volatility term	Yes, explicit RD and volatility values	Limited, mainly designed for 1v1	Chess.com, various online chess and Go servers
TrueSkill	Full Bayesian probability distribution per player, updated via factor-graph message passing	Yes, full distribution over skill	Yes, built for team games and variable team sizes	Xbox Live (Halo and other titles), used as a basis for later matchmaking models in other studios

Elo established the core "predicted versus actual" logic; Glicko added a formal confidence measure; TrueSkill extended that logic to teams and variable-size matches.

Why uncertainty, not just skill, drives who you get matched against

Tracking uncertainty alongside skill lets these systems make smarter matchmaking decisions than raw Elo alone. TrueSkill, for instance, can deliberately match a highly uncertain new player against a wider range of opponents, specifically because doing so provides more information to narrow down that player's true skill level faster. This is comparable to how a doctor orders a broader panel of tests for an unfamiliar patient before narrowing in on a specific diagnosis: early, wide-net information gathering pays off later in more precise, targeted decisions.

A second useful analogy is estimating a stranger's basketball skill. Watching them make one shot from half-court tells you almost nothing certain, since a lucky shot and a genuinely skilled one look identical from the outside. You'd want several more attempts across different conditions before drawing confident conclusions. Watching a professional player who has thousands of recorded games behind them needs far less additional data to update your estimate, because the existing sample is already large and internally consistent. That is functionally what rating deviation and TrueSkill's variance term are measuring: how much a single new data point should be allowed to move the needle.

Why does my rating change so much in early games?

Because systems like Glicko and TrueSkill deliberately treat new players as high-uncertainty cases. A wide confidence interval around a new rating means each additional game carries more weight in narrowing that interval, so early wins and losses produce larger rating swings than the same results would for an established player with hundreds of tracked matches.

Why do some games avoid showing a raw numeric rating to players?

Many studios display tiers, ranks, or percentile bands instead of raw numbers specifically because the underlying number, especially one carrying a wide uncertainty interval for newer players, is not a stable enough single figure to present as a precise measurement. Showing a range or tier communicates the same information without implying false precision.

Why matchmaking still feels inconsistent even when the math is sound

Matchmaking ratings estimate a single number (or distribution) for overall skill, but actual performance varies match to match due to factors the rating cannot see: fatigue, team communication in team-based games, champion or hero pool familiarity on a given day, and the ordinary variance present in any human performance. A rating system converging on your statistically "true" skill level does not mean every individual match will feel perfectly balanced. It means balance holds on average across a large enough sample of games, which is a very different guarantee than "this specific match will feel fair."

Team-based games add a further layer of estimation difficulty. The system observes only the team's win or loss directly, not each individual's contribution to that result, so it must infer individual skill from a partial, noisier signal than the head-to-head data a system like chess Elo works with. Games such as League of Legends and Overwatch layer additional statistical models on top of a TrueSkill-style or Elo-derived foundation specifically to help separate individual performance from team outcome, but that inference is inherently less precise than direct one-on-one results.

There is also a deliberate design tradeoff most players never see: matchmaking systems weigh queue time against match quality. A theoretically perfect skill match that takes twenty minutes to find is often a worse experience than a slightly less balanced game found in two minutes, so most production matchmaking systems widen the acceptable skill range the longer a player waits in queue. This means the "opponent quality" of any given match is partly a function of server population and time of day, not just the rating math itself. A niche game mode with a small active player base will show more matchmaking variance purely because there simply aren't enough nearby-rated players online at any given moment to fill a tightly matched lobby.

Smurfing, meaning experienced players creating new accounts that start with default low-uncertainty assumptions, is a related real-world failure mode. A skilled player on a fresh account temporarily fools the rating system precisely because new-account uncertainty is supposed to let ratings move quickly, and a smurf exploits that fast-movement window to stomp lower-rated opponents before the system catches up. Anti-smurf detection, when studios implement it, typically looks for anomalous win rates and performance metrics that diverge sharply from what the account's current rating and uncertainty would predict.

Common misconceptions about matchmaking ratings

A common assumption is that a rating number represents a precise, unchanging measurement of skill, similar to a person's height. In reality, most modern systems track a probability distribution, a best estimate plus a margin of uncertainty, not a fixed value. That is precisely why the same visible rating can behave very differently for a player with 20 games played versus one with 2,000: the underlying confidence interval is not shown on screen, but it is doing real work behind the scenes.

Another misconception is that losing to a much lower-rated player should always tank a rating dramatically, regardless of context. The actual penalty depends on the opponent's rating uncertainty and the specific K-factor or comparable weighting term the system applies. An "upset" loss against a player whose true skill was still statistically uncertain produces a smaller rating hit than losing to someone confidently, reliably rated far below you, because the system has less confidence that the lower-rated player's true skill is actually lower.

A third error is assuming matchmaking always tries to create the closest possible skill match above all else. As covered above, many systems also weigh queue time as a competing priority, which is why matchmaking ranges intentionally widen the longer someone waits in queue, and why "close skill match" is a target the system optimizes toward on average, not a guarantee it delivers every single game.

A fourth misconception, worth naming directly, is that a rating system claims to measure innate talent or long-term potential. It does not. It measures recent, observed competitive performance under a specific game's specific ruleset, against a specific pool of opponents. A player's rating in one game mode, or one title, carries no formal statistical relationship to their rating in a different one, even if the two games feel similar to play.

What developers actually optimize for

It's worth being explicit about what a matchmaking system is designed to do, because it is not simply "find the most accurate skill number." Studios building these systems are typically optimizing a joint objective across several competing goals: predictive accuracy (does the rating correctly forecast who wins), convergence speed (how quickly does a new player's rating become trustworthy), queue time (how long does a player wait), and perceived fairness (does the match feel close, independent of whether the underlying probabilities were actually close).

These goals genuinely conflict with each other. A system tuned purely for predictive accuracy might produce longer queue times as it holds out for a tighter skill match. A system tuned purely for fast queues will occasionally produce lopsided matches that feel unfair even if the rating math behind them was technically sound. Because of this, published details on exactly how a specific commercial game tunes its matchmaking (its K-factor equivalents, queue-widening curves, or team-balancing weights) are usually proprietary and not publicly documented in the way Elo's original chess formula or the published TrueSkill and Glicko papers are. What is publicly verifiable is the underlying mathematical framework each system is built on, which is what this article has focused on.

The takeaway

Matchmaking ratings work by continuously comparing predicted outcomes to actual outcomes and adjusting a skill estimate based on how surprising each result was. Elo established that core logic for chess in the 1960s; Glicko added an explicit measure of statistical confidence in the 1990s; TrueSkill, published by Microsoft Research in 2007, extended the same underlying idea to team games with variable team sizes. A rating is a probabilistic estimate refined game by game, not a fixed score, and production matchmaking systems balance close skill matching against acceptable wait times rather than optimizing for either one alone. Understanding this explains both why ratings swing wildly early on, and why they settle down, and why no single match is ever a clean test of "who is actually better."

---

Read next:

• Why random loot feels rewarding
• Why lag compensation is controversial
• How pathfinding AI moves NPCs

FAQ

Q: What is the Elo rating system?

A rating system created by Arpad Elo for chess that calculates the probability of a win between two players based on their ratings, then adjusts both ratings after the match based on whether the actual result matched that prediction. It was adopted by the USCF in 1960 and FIDE in 1970.

Q: How is TrueSkill different from Elo?

TrueSkill, published by Microsoft Research in 2007 and built for Xbox Live, tracks a full probability distribution over each player's skill rather than a single number, and it was specifically designed to handle team games with variable team sizes, which the original two-player Elo formula does not natively support.

Q: What does rating deviation (RD) mean in the Glicko system?

Rating deviation, introduced by statistician Mark Glickman in the Glicko system, is a measure of statistical uncertainty around a player's rating, expressed as one standard deviation. A high RD means the system is not yet confident in the rating; a low RD means the rating is well established from many prior games.

Q: Why does a new player's rating change so much in early games?

New players have high uncertainty in their rating estimate, so systems like Glicko and TrueSkill let their rating shift more dramatically per game until enough matches have been played to narrow down a confident, stable estimate.

Q: Does matchmaking always prioritize the closest possible skill match?

No. Most matchmaking systems balance skill closeness against queue time, intentionally widening the acceptable skill range the longer a player waits, since a long search for a marginally closer match often hurts player experience more than a slightly wider skill gap would.

Q: Can a matchmaking rating measure long-term potential or innate talent?

No. A matchmaking rating measures recent, observed competitive performance in a specific game and ruleset against a specific pool of opponents. It carries no formal statistical relationship to a player's rating in a different game or mode, even a similar one.