Reading Skill from Behaviour: Inferring Chess Rating from Clock Usage and Opening Habits in Online Blitz

Purpose. Behavioural AI is often asked to infer hidden properties of people, such as how skilled they are, from how they act rather than from what they achieve. We ask how well the skill of an online chess player can be inferred from behaviour that does not require a chess engine: how the player spends time on the clock and how varied the openings and lengths of games are. Design and methods. We collected up to 40 recent rated 3+0 blitz games (move lists, per-move clocks, opening codes) for 540 public Lichess players with at least 15 such games (21,409 games), sampled from the result lists of three finished arena tournaments. For each player we computed 18 behavioural features (14 on time management, 4 on openings and game length) and predicted the player's mean rating over those games with ridge regression and gradient boosting under repeated five-fold cross-validation over players, comparing feature sets with a median-prediction baseline and examining how accuracy depends on the number of games observed. Findings. The best model, ridge regression on all behavioural features, reached a cross-validated R² of 0.37 (95% bootstrap CI 0.30–0.44) and a mean absolute error of 167 rating points, against 213.5 for predicting the median rating (a 21.8% reduction); 37.2% of players were predicted within 100 points and 64.8% within 200 points (ratings: mean 1706, SD 267.5). Time management carried somewhat more information (R² 0.32) than opening and game-length habits (0.27), although the intervals overlap. Gradient boosting did not improve on ridge regression (0.36). Accuracy rose with the number of games observed, from R² 0.18 with 5 games to 0.35 with 40, without a clear plateau. Originality. The study quantifies, on public data and with person-level cross-validation, how much of an individual's skill can be read from behaviour alone, without any engine analysis of move quality, and shows that this signal is informative at the group level but too noisy to be used as a measure of an individual player's strength.