ChessMimic: Per-Rating Transformer Models for Human Move, Clock, and Outcome Prediction in Online Blitz Chess

2026-06-03

AI Summary

This work addresses the prediction of human move selection, thinking time, and game outcomes in rapid chess by proposing a multi-task modeling approach stratified by Elo rating. Three lightweight encoder-only Transformer models are developed to independently predict moves, deliberation time, and match results within 100-point Elo bands. Each model takes as input a fusion of board state, recent move history, player Elo ratings, and clock information. Without relying on complex attention mechanisms, the approach outperforms Maia-2 in move prediction across all Elo segments. It achieves an AUC of 0.78 for outcome prediction—significantly surpassing both Maia-2 and a logistic regression baseline—and attains Pearson and Spearman correlation coefficients of 0.41 and 0.50, respectively, for thinking time prediction, marking the first successful joint modeling of Elo ratings and clock dynamics.

Abstract

We present ChessMimic, a system of three small encoder-only transformers - for move, thinking-time, and outcome prediction - conditioned on the position, recent move history, player rating, and clock state. We fit a separate instance of each model per 100-Elo rating band, trading parameter efficiency for sharper per-skill calibration. On a held-out month-wide slice of Lichess Rated Blitz games ChessMimic's human move prediction accuracy outperforms Maia-2 in every Elo band. Compared to Maia-3, our 9M parameter model's accuracy sits between Maia-3-5M and Maia-3-23M without the additional complexity of Geometric Attention Bias. In addition to the move matching model, we also train a game outcome model that conditions not only on the position, but also player ratings, time control, and remaining clock times. The outcome model achieves an AUC of 0.78 out of sample, beating Maia-2 as well as logistic regressions based on material, ratings, and clock time. Finally, we train a clock model that predicts human thinking times. The clock model provides a usable but non-SOTA per-ply think-time signal under ALLIE-style filters (Pearson r = 0.41, Spearman rho = 0.50, MAE 4.10 s, against ALLIE's reported r = 0.70), with the residual gap concentrated in per-position bucket sharpness rather than bucket-marginal calibration. A public demo is at 1e4.ai and we release code, per-band weights, and the C++ data-filter pipeline code in GitHub.

Problem

• human move prediction
• thinking time prediction
• game outcome prediction
• blitz chess
• player rating

Innovation

• per-rating modeling
• human move prediction
• thinking-time estimation
• outcome prediction
• encoder-only transformers