1 History and development

1.1 Origin in chess

The Elo rating system was conceived by Hungarian‑born American physicist Arpad Elo (1903–1992) as an improvement over the Harkness system, which was used by the United States Chess Federation (USCF) in the mid‑20th century. Elo’s original work, published in 1960, proposed a statistical model in which a player’s performance is treated as a normally distributed random variable, with the rating representing the mean of that distribution. The system was first implemented by the USCF in 1961 and gradually replaced earlier ad‑hoc rating methods. Elo’s key insight was to predict the outcome of a match based on the difference between two ratings, updating each player’s score after every game in a way that automatically adjusts for the strength of opponents.

1.2 Adoption by international federations

The World Chess Federation (FIDE) adopted the Elo system in 1970, initially using a variant with a logistic distribution rather than the normal distribution that Elo originally proposed. FIDE’s implementation standardized the scaling factor (400 points corresponding to a 10‑to‑1 odds ratio) and established a fixed K‑factor (initially 30 for most players). The success of the system in chess led to its rapid adoption by other two‑player games, including checkers, go, and shogi. National chess federations around the world followed suit, and the system became the de facto global standard for rating competitive chess players.

1.3 Modern variants and computer implementations

As computing power increased, the Elo system was adapted for digital environments. Online chess platforms (e.g., Chess.com, Lichess) implemented their own Elo‑based ratings, often using modified K‑factors and provisional rating periods. The rise of esports in the late 1990s and early 2000s saw Elo used for matchmaking in games such as *Starcraft*, *Counter‑Strike*, and *League of Legends*, though those implementations frequently deviated from the original formulas to handle team dynamics and large player pools. More sophisticated descendants—such as Glicko, Glicko‑2, and TrueSkill—emerged to address specific shortcomings, but the core Elo framework remains influential in both academic and applied statistics.

2 Mathematical foundation

2.1 Expected score formula

The expected score for player A against player B is given by:

\[ E_A = \frac{1}{1 + 10^{(R_B - R_A)/400}} \]

where \(R_A\) and \(R_B\) are the ratings of players A and B respectively. This formula ensures that a player with a higher rating has an expected score greater than 0.5, and the magnitude of the difference determines how much greater.

2.1.1 Logistic distribution assumption

The use of the base‑10 exponent of rating differences is equivalent to assuming that the difference in playing strengths follows a logistic distribution rather than a normal distribution. The logistic distribution has heavier tails, meaning it assigns a slightly higher probability to upsets than a normal model would. This choice simplifies computation and has been empirically validated in chess: the probability of a stronger player beating a weaker one aligns well with the logistic curve across a wide range of rating gaps.

2.1.2 Scale factor (400 points rule)

The constant 400 in the denominator originates from FIDE’s convention that a 400‑point rating difference corresponds to an expected score of approximately 0.91 (i.e., the stronger player is about ten times as likely to win as to lose). This scaling factor is arbitrary but consistent; different organisations may adjust it (e.g., 200 points in some esports implementations) without changing the underlying logic.

2.2 Rating update formula

After a game, each player’s rating is updated as:

\[ R_{\text{new}} = R_{\text{old}} + K (S - E) \]

where \(S\) is the actual outcome (1 for a win, 0 for a loss, 0.5 for a draw) and \(K\) is the K‑factor.

2.2.1 K‑factor and its variability

The K‑factor determines the magnitude of each rating adjustment. Higher K‑factors make ratings more volatile (faster to respond to new results), while lower values increase stability. In chess, FIDE uses a tiered system: K=40 for new players (first 30 games), K=20 for players with ratings under 2400, and K=10 for those above 2400. Many online games use a fixed K for all players, while others (e.g., provisional‑rating systems) start with a high K that decreases as the number of played games increases.

2.2.2 Example calculation

Suppose player A has a rating of 1500 and player B has 1400. The expected score for A is:

\[ E_A = \frac{1}{1 + 10^{(1400-1500)/400}} = \frac{1}{1 + 10^{-0.25}} \approx 0.640 \]

Thus A is expected to score 0.64 points. If A wins (S=1), and K=20, the new rating for A is:

\[ R_A^{\text{new}} = 1500 + 20 \times (1 - 0.640) = 1500 + 20 \times 0.36 = 1507.2 \]

B’s new rating would be \(1400 + 20 \times (0 - 0.360) = 1392.8\). Note that the sum of rating changes is zero because the total points gained by one player equal the points lost by the other.

2.3 Statistical properties

2.3.1 Stationarity and convergence

The Elo system is a Markov process: ratings evolve over time based on game outcomes. If players’ true strengths are constant, ratings will converge to a steady state where the expected update is zero. However, the presence of a non‑zero K‑factor means ratings fluctuate indefinitely around the true values, creating a stationary distribution. Convergence speed is inversely related to K: high K leads to rapid initial adjustments but greater long‑run variance.

2.3.2 Confidence intervals and reliability

Each Elo rating is an estimate, not an exact measure of strength. The variance of the estimate depends on the number of games played and the K‑factor. Roughly, a rating’s standard error (in points) is about \(K / \sqrt{n}\), where \(n\) is the number of games. For example, a player with K=20 and 100 games has a standard error of about 2 points. Many systems publish ratings only to the nearest integer or whole point, but statistical error is always present; rating differences of less than 10–20 points are rarely statistically significant.

3 Applications

3.1 Traditional board games

3.1.1 Chess (FIDE Elo)

FIDE has maintained a global Elo rating list since 1970. Ratings are calculated monthly and are used to determine titles (Grandmaster, International Master, etc.), seeding in tournaments, and qualification for high‑level events. FIDE’s implementation uses K‑factors that vary by rating and number of games, and it includes special rules for players with no prior rating (provisional ratings).

3.1.2 Go and other two‑player games

The game of Go adopted Elo‑based rankings in the late 20th century, particularly through the European Go Federation (EGF) and online servers (e.g., KGS, OGS). Because Go’s game outcomes are more deterministic (fewer draws), the logistic model remains a good fit. Other two‑player abstract games (checkers, othello, backgammon) use similar systems, often with adjusted scale factors to account for different game dynamics.

3.2 Digital gaming and esports

3.2.1 Matchmaking algorithms

Most competitive online games use Elo‑like ratings to pair players of similar skill. The essential process: after each match, each player’s rating is updated; the system then searches for opponents whose current ratings are close. League of Legends uses a variant called “MMR” (Matchmaking Rating), while *Counter‑Strike: Global Offensive* uses a modified Elo for its competitive ranking tiers. The transparency of the Elo system is appealing to developers: it is computationally cheap and easily understood.

3.2.2 Team‑based Elo variants (e.g., Glicko)

Pure Elo does not handle team games well because an individual’s contribution cannot be inferred from a single match outcome. The Glicko system (developed by Mark Glickman in 1995) adds a “rating deviation” (RD) that quantifies uncertainty. In team variants, a common approach is to treat each team’s rating as the average (or sum) of its members’ ratings, then update each player using a team‑based outcome. Glicko was adopted by the Internet Chess Club and by several esports titles such as *Team Fortress 2*.

3.3 Non‑gaming domains

3.3.1 Educational assessment

Elo ratings have been applied to testing scenarios where student ability and question difficulty are linked through the logistic model. For example, the “Item Response Theory” framework used in computer‑adaptive tests (e.g., the GRE) shares mathematical similarities with Elo. Some online learning platforms (e.g., Duolingo) use Elo‑like algorithms to estimate learner proficiency and adjust question difficulty in real time.

3.3.2 Customer service performance

Some call‑centre and help‑desk systems assign Elo ratings to support agents based on customer satisfaction scores or ticket resolution rates. The system can then match difficult customers to higher‑rated agents, while simpler queries go to newer staff. This application is informal but illustrates the broader versatility of the rating mechanism.

4 Criticisms and limitations

4.1 Inflation and deflation of ratings

Over time, rating pools can drift. Inflation occurs if weaker players enter the pool and lose to existing players, artificially raising the average; deflation happens if strong players retire without being replaced, pulling ratings down. In chess, FIDE periodically adjusts the average rating (e.g., by recalibrating the zero point) to counter these effects. In online games, inflation is often self‑limiting when ratings are used for matchmaking—players at the top and bottom of the distribution are matched with each other, preserving relative differences.

4.2 Sensitivity to K‑factor selection

Choosing an appropriate K‑factor is a trade‑off. Too high a K makes ratings volatile, causing large swings after a single upset; too low a K makes ratings sluggish, failing to reflect genuine improvement or decline. The optimal K depends on the frequency of play and the stability of player skill. In practice, fixed K for all players can unfairly penalise inconsistent performers or reward sandbagging (deliberate loss of rating to enter a weaker pool).

4.3 Inability to handle multi‑player dynamics

Classic Elo is designed for two players. In multiplayer games (e.g., a 10‑player free‑for‑all), assigning a single rating update becomes ambiguous. Linear extensions (treating each pairwise comparison equally) lead to computational complexity and do not capture strategic alliances or collusion. More modern systems (TrueSkill, Glicko‑2) include algorithms that handle multi‑player outcomes, but Elo itself is fundamentally a pairwise model.

5.1 Glicko and Glicko‑2

Developed by Mark Glickman, the Glicko system adds a “rating deviation” (RD) that measures the reliability of a rating. After each game, both the rating and the RD are updated. Glicko‑2 further introduces a “rating volatility” parameter to model long‑term changes in a player’s consistency. Both systems reduce the impact of very old games and are widely used in chess servers and online gaming.

5.2 TrueSkill

Created by Microsoft Research for the Xbox 360 gaming platform, TrueSkill extends Elo to multi‑player and team games using Bayesian inference. It estimates a probability distribution for each player’s skill, then updates it based on match outcomes. TrueSkill can model draws, differences in player performance, and variable numbers of players per team.

5.3 Elo‑based vs. Bayesian methods

Bayesian methods (e.g., TrueSkill, Glicko‑2) explicitly model uncertainty and incorporate prior knowledge, providing a more rigorous statistical framework. They are less sensitive to K‑factor choices and capable of automatically handling rating drift. However, they are computationally more intensive and harder to explain to end users. Elo remains popular precisely because of its simplicity and transparency—its update rule is a single line of arithmetic.

6 See also

6.1 Rating systems in competition

  • Chess rating system
  • World Chess Federation (FIDE) rating list
  • Greatest chess players by rating
  • Sports rating systems (e.g., ELO for football)
  • Comparative rating systems (e.g., ECF grading)

6.2 Probability and logistic regression

  • Logistic function
  • Bradley–Terry model (paired comparison)
  • Item response theory
  • Bayesian inference in rating systems