Testing a System for Discerning Blackjack-Playing Tendencies from Card-by-Card Hand Histories: An Initial Simulation Using Bots

Abstract

The game of blackjack involves in-game actions that can materially affect the possible outcomes and/or their probabilities. In theory, with data from enough hands and decision points, researchers and gambling operators should be able to develop player profiles and then use those profiles to detect increased risk of gambling harm. However, it is currently unclear how much blackjack play is needed to accurately classify players. As a preliminary experiment, we separated blackjack hands into nine classes, and then proposed a total of twenty play patterns (“heuristics”) across hand classes to operationalize players’ strategies. We used our heuristics as components to construct 506 blackjack-playing bots to simulate human play. We then created a program that would read a session’s data and attempt to determine which bot generated that data. We found that for more infrequently occurring hand classes (e.g., pairs, soft hands), even 30 shoes of play were not enough to accurately determine the identity of a bot. With more frequent hand classes (e.g., hard hands), we could only accurately identify bots that played consistent strategies within 30 shoes. Results suggest that efforts to use blackjack hand histories and profile-fitting to generate markers for gambling harm might require simpler classification systems, or otherwise be limited to highly involved blackjack players.