Predicting Thinking Time in Chess: An Empirical Study Based on Magnus Carlsen's Blitz Games

Artificial intelligence systems have reached superhuman performance in chess, yet they largely fail to reproduce essential aspects of human behavior, notably the temporal dynamics of decision-making. While recent research has focused on modeling human move selection and playing style, the time humans spend thinking before playing a move remains largely unexplored. This temporal dimension is central to human cognition and plays a key role in both strategic decision-making and the subjective experience of playing chess. In this work, we investigate whether human thinking time can be empirically modeled from real game data. Focusing on blitz games played by Magnus Carlsen, we construct a dataset from publicly available PGN files in which each move is enriched with temporal, positional, and game-dynamic features. Using this dataset, we evaluate several machine learning approaches, ranging from linear models to non-linear methods such as Random Forests and neural networks, to predict the time spent on individual moves. Our results show that non-linear models substantially outperform linear baselines, with a Random Forest model explaining up to 50% of the variance in human thinking time. These findings suggest that thinking time is not purely stochastic, but depends on observable features of the position and game context that are accessible in real time. Although limited in scope, this study demonstrates the feasibility of predicting human reflection time and highlights the importance of temporal modeling for human-aligned chess AI. More broadly, this work opens the door to integrating realistic time management into human-like chess engines and imitation-based systems such as Maia. By enriching move prediction with a temporal component, future AI systems could achieve more coherent human-AI alignment, improved pedagogical value, and a more immersive playing experience. Our implementation and experimental pipeline are enabled here: https://github.com/elijahlieb/magnus-time-prediction-ml/tree/main