How Video Game Matchmaking Algorithms Actually Work: The Mechanics of TrueSkill, Glicko-2, and EOMM

The moment you click "Find Match" in a multiplayer lobby, the outcome of your next game is already being negotiated by a mathematical model. Before the map loads, before you pick a character, and before the first shot is fired, an algorithm has evaluated your recent history, quantified your skill, and calculated the exact probability of your victory. This fraction of a second—the matchmaking phase—is the most consequential step in modern competitive gaming.

It is the invisible hand that dictates not just who you play against, but how you feel when you log off. For years, players have treated matchmaking as a simple sorting hat that pairs good players with good players and beginners with beginners. But the reality of modern network engineering is far more complex, blending Bayesian statistics with behavioral psychology to construct a lobby that serves the developer's ultimate goal.[5]

The gold standard for this calculation was originally a system borrowed from the analog world: the Elo rating. Invented by physics professor Arpad Elo for the United States Chess Federation, it assigned every competitor a single number. If a 1500-rated player beat a 1600-rated player, the system transferred points based on the mathematical unlikelihood of the upset. But Elo had a fatal flaw when applied to the digital era: it had no concept of uncertainty, and it could not handle teams.

A veteran with 1,000 matches and a beginner who got lucky in their first ten games might both sit at 1500, and the Elo algorithm would treat them identically. Furthermore, in a 4-vs-4 shooter, Elo required developers to invent clunky heuristics—like averaging the team's rating and updating everyone equally—which threw away vital information about who carried the team and who was merely along for the ride.[4]

The solution arrived in 2005, when Microsoft Research developed TrueSkill for the launch of the Xbox Live network. TrueSkill fundamentally changed the math by replacing the single rating number with a Bayesian probability distribution. Instead of one definitive digit, TrueSkill tracks two distinct variables: mu (μ), which represents your average skill, and sigma (σ), which represents the system's uncertainty about that skill.

By modeling every player as a Gaussian bell curve, the algorithm could mathematically express doubt. It also solved the team problem by assuming that a squad's overall performance is simply the sum of its individual players' distributions, allowing it to accurately rate massive 8-vs-8 matches in franchises like Halo or Gears of War.[1][2][6]

When a new player enters a game under the TrueSkill system, their sigma is massive. The algorithm knows it does not know how good they are, so their rating swings wildly after every match. As they play more games and establish a track record, the system gathers evidence, the sigma shrinks, and the rating locks in.

New players converge on their true rank rapidly, while veterans drift slowly—exactly the behavior required for a smooth matchmaking experience. TrueSkill runs expectation propagation, iteratively projecting each non-Gaussian message back onto the closest Gaussian by matching the first two moments, ensuring that the math remains computationally efficient even when processing millions of game outcomes per day.[1]

A parallel evolution happened in the open-source and academic communities with the Glicko-2 system. Developed by statistician Mark Glickman, this framework added a third crucial variable to the mix: volatility. If a player suddenly starts performing wildly above or below their established baseline—perhaps because they handed the controller to an older sibling, or finally mastered a crucial mechanical skill—the volatility metric spikes.

This tells the algorithm to temporarily increase their rating deviation, essentially unlocking their rank and allowing it to adjust rapidly to their new actual skill level. Systems like Glicko-2 became the backbone of massive online platforms, ensuring that dynamic changes in player ability were captured in real time.[4]

Historically, systems like TrueSkill and Glicko-2 operated on a single, unquestioned premise: the goal of matchmaking is to create a perfectly fair game. The algorithms were meticulously tuned to find matches where both teams had a mathematical 50% chance of winning.

But as the gaming industry shifted toward live-service models and free-to-play economies, developers began to realize that perfect fairness does not actually maximize player retention. A perfectly balanced match is often exhausting, requiring peak concentration and resulting in a coin-flip outcome. Data scientists began to ask a radical question: what if fair games actually make people want to stop playing?[5]

In 2017, researchers from Electronic Arts and UCLA published a peer-reviewed paper that quietly upended the industry's assumptions. They introduced a framework called Engagement Optimized Matchmaking (EOMM). By analyzing 36.9 million matches from a popular EA title, the researchers proved mathematically that equal-skill matchmaking is not the optimal way to keep players logged in.

As lead researchers Zhengxing Chen and Su Xue state in the paper's abstract, 'We prove that equal-skill based matchmaking is a special case of EOMM on a highly simplified assumption that rarely holds in reality.' The paper demonstrated that intentionally scheduling unbalanced matches can significantly enhance overall player engagement.[3]

The EOMM paper revealed the stark behavioral math of frustration. If a player suffers three consecutive losses, there is a 5.1% chance they will quit the game entirely. If they win three in a row, the churn risk drops to 3.7%, but they still might leave out of boredom or a lack of challenge.

The optimal sequence for retention—the pattern that drops the churn risk to a mere 2.6%—is a manufactured rollercoaster: a draw, followed by a loss, followed by a win. By feeding players this specific sequence of emotional peaks and valleys, the algorithm maximizes the probability that they will queue up for just one more match.[3]

This revelation shifted the objective function of matchmaking. Instead of merely calculating skill, an engagement-optimized system calculates the probability that you will quit after the next match. If the algorithm detects that you are on a losing streak and approaching the statistical threshold of frustration, it can intentionally place you in a lobby against lower-skilled opponents.

It feeds you a win to reset your emotional state and keep you in the ecosystem. Conversely, if you are winning too much, it can place you against highly skilled veterans to humble you and re-establish the game's challenge.[3][5]

The implications of this shift are profound for the economics of live-service gaming. The EOMM framework explicitly notes that its core components—the skill model, the churn model, and the graph pairing model—are completely uncoupled. This means the objective function can be swapped out at will.

A system designed to optimize for playtime can just as easily be tuned to optimize for in-game spending. The algorithm could theoretically pair players who haven't bought cosmetic items with highly skilled players who have, subtly marketing the premium gear through gameplay by associating it with victory.[3]

Unsurprisingly, the competitive gaming community has grown deeply suspicious of these invisible systems. Players frequently complain of forced 50% win rates and artificially punishing lobbies, sensing the algorithmic hand on the scale. While major publishers rarely confirm exactly which models are running under the hood of their flagship titles, the mathematical foundation is clear. The matchmaking screen is no longer just a waiting room where the server looks for a good connection. It is a sophisticated behavioral engine, calculating exactly what kind of experience you need to keep you playing.[5]