Using Big Data and Machine Learning to Uncover How Players Choose Mixed Strategies (by Toshihiko Hirasawa, Michihiro Kandori, Akira Matsushita)
Toshihiko Hirasawa, Michihiro Kandori, Akira Matsushita
How do humans behave in a situation where (i) one needs to make one’s own behavior unpredictable and (ii) one needs to predict an opponent’s behavior? This is an important class of strategic situations, formulated as games with a mixed strategy equilibrium. If humans are put in such a situation, it is obvious that, rather than calculating the mixed equilibrium strategy, they use their hunches and some heuristics to achieve the aforementioned goals (i) and (ii). Exactly what kind of mechanisms are employed has not been fully understood. To address this issue, we use our unique big experimental data set about a game with a mixed strategy equilibrium, which has about 75,000 observations, and compare conventional behavioral economics models with some leading machine learning models. The use of big data enables us to examine the external validities of those models, i.e., compare the predictive powers of those models in data sets that are not used for parameter estimation. We found that machine learning models, most notably a version of the deep learning model LSTM, substantially outperform the leading behavioral model (EWA), and this happens only when the size of the data set for parameter estimation is sufficiently large. Finally, we try to improve the EWA model by incorporating the insights gained from the machine learning models.