Beyond Elo: How Glicko-2 and TrueSkill Are Rewriting the Rules of Competitive Ranking

The quest to mathematically quantify human skill is as old as competition itself. For decades, the gold standard has been the Elo rating system, a formula devised in 1960 by physics professor Arpad Elo to rank chess players. By assigning a single number to represent a competitor's strength, Elo created a zero-sum economy where points are exchanged after every match. If a grandmaster defeats a novice, the rating exchange is minimal; if the novice pulls off an upset, the point transfer is massive. This elegant logic made Elo the foundational algorithm for everything from international chess to early video games and even dating apps.[5]

But the explosion of online gaming has pushed Arpad Elo's mathematical model to its breaking point. Modern matchmaking systems must process millions of games a day, dealing with new accounts, returning veterans, and complex team dynamics. A single number is no longer enough to capture the nuances of digital competition. In response, data scientists and engineers have developed sophisticated successors. Today, the competitive landscape is dominated by a fascinating three-way philosophical battle between the traditional Elo framework, Mark Glickman's highly adaptive Glicko-2, and Microsoft's proprietary TrueSkill algorithm.[1][3]

When evaluating the case for Elo, its primary advantage lies in its absolute transparency and historical continuity. The math is straightforward enough to be calculated on a napkin. Players understand exactly what is at stake before a match begins, and the system requires minimal computational overhead to maintain. This simplicity fosters a sense of trust; there are no hidden variables suppressing a player's climb up the ladder. For traditional over-the-board tournaments where players compete in a closed, highly regulated environment, Elo remains a perfectly calibrated instrument.[4][5]

The case against Elo centers on its inability to measure its own ignorance. In the Elo system, a player who has maintained a 1500 rating over five hundred games is treated exactly the same as a rookie who was just assigned a 1500 rating upon account creation. Elo has no mechanism to track uncertainty. Furthermore, it is notoriously slow to adapt. If a player takes a five-year hiatus and returns with diminished skills, Elo will stubbornly drain their rating one painful loss at a time, ruining the matchmaking experience for everyone involved in those corrective matches.[3][5]

The evidence regarding Elo's modern viability is mixed. The International Chess Federation (FIDE) continues to use it, preserving decades of historical records and title qualifications. However, almost every major online platform has abandoned it. Data from massive multiplayer environments shows that Elo requires an impractical number of matches to converge on a player's true skill level. In fast-paced digital ecosystems where players expect fair matches immediately, Elo's slow convergence rate is a fatal flaw that leads to rampant mismatched lobbies and frustrated user bases.[3][4]

To solve these exact problems, the Glicko system—and its successor, Glicko-2—was introduced. The case for Glicko-2 rests on a concept called Rating Deviation (RD). Instead of just tracking a player's skill, Glicko-2 tracks how confident the algorithm is in that skill. A player's true strength is represented as a 95 percent confidence interval. If a player is new or returning from a long break, their RD is high, allowing their rating to swing wildly to quickly find their correct placement. As they play more consistently, the RD shrinks, stabilizing their rank.[3][5]

The case against Glicko-2 is largely psychological rather than mathematical. Because the system locks in on a player's true skill and reduces their RD, highly active players often find themselves in a state of rating stagnation. A hard-fought victory might yield only a fraction of a point if the system is already highly confident in the player's placement. This can feel unrewarding to users who view ratings as a progression system rather than a strict measurement tool. Additionally, like Elo, Glicko-2 is fundamentally designed for one-on-one competition and struggles to parse team dynamics.[3][5]

The evidence supporting Glicko-2 is overwhelming in the realm of one-on-one digital competition. It is the engine powering the world's largest chess servers, including Lichess and Chess.com, processing millions of matches daily. Empirical analyses demonstrate that Glicko-2 predicts match outcomes with significantly higher accuracy than Elo, particularly in the crucial first fifty games of a new user's lifecycle. By aggressively adjusting the ratings of uncertain players, it effectively neutralizes highly skilled players on new accounts far faster than traditional methods, ensuring that the broader competitive ecosystem remains balanced and fair for genuine beginners.[2][3]

But what happens when competition moves beyond one-on-one duels? The case for TrueSkill, developed by Microsoft Research in 2005, is built entirely around the complexities of team-based multiplayer games. TrueSkill utilizes Bayesian inference and a Gaussian factor graph to evaluate matches where multiple players form heterogeneous teams. It can look at a four-versus-four match, analyze the individual skill bands of all eight participants, and accurately distribute rating updates based on the specific composition of the teams, even handling scenarios where players quit mid-match.[1][2]

The case against TrueSkill is rooted in its opacity and its proprietary nature. The mathematics behind TrueSkill are intensely complex, making it impossible for a player to manually verify why their rating moved the way it did. This black-box nature can breed community resentment. Furthermore, TrueSkill is patented by Microsoft. While it powers massive franchises like Halo and Gears of War, independent developers and open-source projects cannot legally use it without licensing agreements, forcing them to rely on public domain alternatives or build their own Bayesian approximations.[1][2]

The evidence for TrueSkill's efficacy in team environments is heavily documented. Microsoft's internal data from the Halo 2 beta testing phase showed that TrueSkill substantially outperformed Elo in predicting multiplayer outcomes. Independent researchers analyzing games like Counter-Strike: Global Offensive observed a 62 percent accuracy rate in match prediction using TrueSkill models. In 2018, Microsoft pushed the envelope further with TrueSkill 2, which incorporates granular in-game metrics like individual kills, deaths, and quit rates to accelerate skill convergence even faster than the original algorithm.[1][2]

When comparing these systems directly, the trade-offs become a matter of application rather than absolute superiority. Elo and Glicko-2 treat every match as a monolithic zero-sum event between two entities. If a team of five plays another team of five, these systems must treat the teams as single players, failing to recognize that one player might have carried four weaker teammates. TrueSkill fundamentally deconstructs the team, updating each player's individual Gaussian distribution based on the collective outcome, making it the undisputed leader in multiplayer environments.[1][4]

Conversely, when comparing transparency and computational load, the hierarchy flips. Elo requires almost zero processing power, making it ideal for lightweight applications and physical tournaments. Glicko-2 introduces moderate computational complexity to track volatility and time-based decay. TrueSkill requires heavy statistical processing, running complex message-passing algorithms across factor graphs for every single match outcome. For a solo developer building a simple web game, implementing TrueSkill is akin to using a supercomputer to balance a checkbook, representing a massive over-engineering of a simple problem.[2][4]

Ultimately, the traditional Elo system fits well when transparency is the highest priority, matches are strictly one-on-one, and the player pool is relatively small and highly active. It is the system of choice for physical, over-the-board tournaments where prestige and historical continuity matter more than rapid mathematical convergence. It does not fit when a platform experiences high player churn, massive influxes of new accounts, or sporadic participation, as its inability to track uncertainty leads to prolonged periods of mismatched games that can alienate a modern digital audience.[4][5]

Glicko-2 fits well when running a massive, high-volume one-on-one ladder. It is the perfect algorithm for online chess, fighting games, and digital card games. Its Rating Deviation mechanic ensures that new players are quickly sorted into their appropriate skill brackets, protecting the integrity of the broader ecosystem. It does not fit when the core gameplay loop involves varying team sizes, free-for-all deathmatches, or asymmetrical competition, as its mathematical foundation is strictly bound to pairwise comparisons that cannot untangle individual contributions from a group effort.[2][3]

TrueSkill fits well when matchmaking involves complex team dynamics, heterogeneous party sizes, and the need to extract individual skill data from collective outcomes. It is the gold standard for modern blockbuster esports and team shooters. It does not fit when building an open-source project, a low-budget indie game, or a platform where players demand absolute transparency in how their points are calculated. In the modern era of matchmaking, choosing the right algorithm is no longer just a math problem; it is the foundational design choice that dictates the entire player experience.[1][2][6]