How the Elo Rating System Adjusts Player Scores Based on the Logistic Function of Expected Win Probability

On October 1, 2015, the mathematical framework underlying competitive rankings shifted from a heuristic sports formula into a formal machine learning algorithm. Writing in The Angry Statistician, researchers demonstrated that the Elo rating system operates identically to a "forgetful logistic model"—specifically, a stochastic gradient descent optimizing for predictive accuracy. This reclassification stripped away the chess-specific terminology to reveal a pure probability engine.

The mechanism at the core of this engine is the logistic function. Before expected win probability can dictate how many points a player gains or loses, the system must translate the raw difference between two ratings into a percentage.

How that translation occurs determines the stability of the entire ranking ecosystem. If the mathematical function underestimates the likelihood of an upset, top-ranked competitors will bleed points too rapidly when they inevitably lose a match, causing systemic rating deflation.[2]

Arpad Elo’s initial 1978 specification assumed that a player’s performance in any given match followed a normal distribution—the classic bell curve. Under that strict assumption, extreme deviations from a player's average skill level were mathematically rare.[2]

However, empirical data from thousands of tournament matches revealed that human performance possesses "fat tails." Upsets happen significantly more frequently than a normal distribution predicts, meaning the bell curve was punishing top players too harshly for standard variance.[1][2]

To correct this structural flaw, the system was modified to use a logistic curve. As outlined by Davide Aversa, the modern expected score formula divides 1 by the sum of 1 plus 10 raised to the power of the rating difference divided by 400.

This specific divisor—400—acts as the scale parameter. It dictates that a player with a 400-point advantage over their opponent has exactly a 10 to 1 ratio of expected success, translating to a 90.9% win probability.

Nic Dobson's analysis of FiveThirtyEight's sports models highlights how this logistic approach perfectly mirrors logistic regression in statistics. The rating difference serves as the single independent variable predicting a binary outcome: win or loss.

When a match concludes, the system calculates the error between the expected outcome (e.g., 0.909) and the actual outcome. The actual outcome is coded as 1 for a win, 0 for a loss, and 0.5 for a draw.

This error term is then multiplied by the K-factor, a constant that determines the maximum possible rating change from a single event. If a heavy favorite wins, the error (1 minus 0.909) is small, resulting in a minimal rating increase of just 0.091 multiplied by K.[1]

Conversely, if the 400-point underdog wins, the error is massive (1 minus 0.091). The underdog absorbs a massive rating boost, while the favorite suffers an equal and opposite penalty.[1]

"The Elo system is essentially a forgetful logistic model," notes The Angry Statistician, pointing out that the K-factor acts exactly like a learning rate in machine learning. A high K-factor makes the system highly reactive to recent results, "forgetting" older data quickly.

A low K-factor, often used for established grandmasters, forces the system to prioritize historical performance over recent volatility. FIDE, the international chess federation, drops the K-factor from 40 to 10 once a player reaches a 2400 rating.[2]

The zero-sum nature of this point exchange ensures that the total number of rating points in the system remains constant, provided no new players enter or exit. This conservation of points allows ratings to function as a relative currency of skill across decades.

Yet, the logistic Elo model is not without structural limitations. Because it tracks only a single point estimate for each player, it cannot quantify its own mathematical uncertainty.

If a player has competed in 1,000 matches and holds a rating of 1500, and another player has competed in exactly one match and holds a rating of 1500, the logistic function treats them as identical entities.[1]

This inability to distinguish between established baselines and high-variance newcomers led to the development of systems like Glicko, which add a secondary variable—rating deviation—to track confidence.

The choice of the logistic function over the normal distribution remains the single most consequential design decision in competitive matchmaking. By mathematically accepting that human performance is prone to wild, unpredictable variance, the base-10 curve ensures that a massive upset recalibrates the system instantly, rather than breaking the scale entirely.[1]