How Skill-Based Matchmaking Actually Works: The Mathematics of Elo, Glicko, and TrueSkill

Deep inside a climate-controlled data center, a matchmaking server sits waiting for its next instruction. It is the ultimate arbiter of digital fate. When a player clicks the "Find Match" button, this server has exactly a few seconds to evaluate millions of online competitors, calculate the precise probability of every possible outcome, and assemble a perfectly balanced lobby. It gets to make this decision every single time a queue pops, and its mathematical judgment is absolute. The server does not care about a player's current mood, their recent losing streak, or their desire to reach the next competitive tier. It only cares about minimizing the mathematical distance between two competing teams.[1]

Yet, if you ask the players subjected to its rulings, the server is a malicious entity. Competitive gaming communities are rife with conspiracy theories about "Elo Hell"—the pervasive belief that the matchmaking algorithm is intentionally rigging games to force a 50 percent win rate and keep players trapped in their current rank. The reality, however, is entirely devoid of malice. It is just Bayesian inference operating at scale. To understand why climbing the ranks feels the way it does, you have to look at the underlying mathematics that govern these digital arenas.[3]

The foundation of all modern matchmaking is the Elo rating system, invented in the 1960s by Hungarian-American physics professor Arpad Elo to rank chess players. Elo's brilliance was its simplicity: it assigns every competitor a single number. When two players face off, the system compares their numbers to predict the winner. If the favorite wins, they gain a few points. If the underdog pulls off an upset, they gain a massive sum. It is a zero-sum exchange that naturally sorts a population by skill over time.

But standard Elo breaks down completely when applied to modern video games. A classic Elo system assumes a one-on-one match, like chess. It has no native concept of teams, meaning it cannot easily distribute rating changes across a four-player squad where one person carried the game and another barely contributed. Furthermore, Elo treats a veteran with 1,000 matches and a rating of 1500 exactly the same as a rookie who just calibrated to 1500 after five games. The server knows the veteran's rating is vastly more reliable, but standard Elo lacks the mathematical vocabulary to express that certainty.

This limitation led statistician Mark Glickman to develop the Glicko system in 1995, and later its successor, Glicko-2. Glickman introduced a crucial new concept: Rating Deviation (RD). Instead of a single number, a player's skill is represented as a range. A rating of 1500 with an RD of 50 means the system is 95 percent confident the player's true skill lies between 1400 and 1600. The more games a player completes, the lower their RD drops, narrowing the confidence interval as the system zeroes in on their actual ability.[2]

Glicko-2 added a third variable to the equation: volatility. According to Mark Glickman's documentation, "The volatility measure indicates the degree of expected fluctuation in a player's rating." If a player suddenly starts playing erratically—winning games they should lose and losing games they should win—their volatility spikes. This spike widens their confidence interval and allows their rating to swing more dramatically, ensuring the system can quickly adapt if a player fundamentally changes their playstyle or takes a long break.[2]

When Microsoft Research needed a matchmaking engine for the launch of Xbox Live and Halo 2 in 2004, they took these probabilistic concepts and built TrueSkill. TrueSkill is a fully Bayesian rating system that models every player as a Gaussian distribution, defined by a mean (μ, the best guess of their skill) and a standard deviation (σ, how unsure the system is). TrueSkill starts players at a mean of 25 and a standard deviation of 8.333. A player's true skill is estimated to be between μ ± 2σ, meaning a new player's baseline is anywhere from 8.334 to 41.666.[1]

TrueSkill solved the team problem through a mathematical assumption of additivity. When a four-player team forms, the server simply adds their means and variances together to create a single, massive Gaussian distribution for the squad. It then compares that team distribution against the opposing team's distribution using a factor-graph message passing algorithm. Before the match even begins, the server has calculated the exact percentage chance of either side winning, standardizing the gap into a predictable metric.[1]

Every match provides a Bayesian update. When the game ends, the server updates its beliefs. The mean moves up or down based on the result, and the standard deviation shrinks, because every data point makes the system slightly more confident in its assessment. As researchers at MetricGate note in their analysis of the algorithm, "New players converge fast; veterans drift slowly — exactly the matchmaking behavior you want." New players see massive rating swings because their standard deviation is huge; veterans see their rank barely budge because the system is already mathematically certain of where they belong.

This is where the friction between human psychology and matchmaking mathematics occurs. Players perceive the shrinking standard deviation as the game artificially "trapping" them. In reality, the algorithm has simply gathered enough data to accurately predict their performance. To climb out of this state, a player cannot just play more games; they have to fundamentally improve their actual skill to consistently beat the server's predictions.[3]

When a player does genuinely improve, the math actually works in their favor. Because systems like Glicko-2 and TrueSkill track volatility and uncertainty, a sudden streak of unexpected wins against higher-rated opponents forces the algorithm to doubt its previous certainty. The standard deviation expands, the update multiplier increases, and the player's rating climbs significantly faster than it would under a rigid Elo system. The server isn't holding anyone back; it is just waiting for mathematical proof that a player has evolved.[2]