Data-Driven Agent-Based Models of Investors with Machine Learning

Image recognition or self-driving cars are just a few among many applications of machine learning (ML) methods. Given that we can train a cobot to mimic human behaviour, why not train a computer to mimic and simulate investor behaviour in stock markets? This would not only improve understanding about investor decision making and their interaction, but provide effective tools to predict investor behaviour on the microscopic level and simulate stock markets on the macroscopic level. The main objective is to create a data-driven Agent-Based Model (ABM), where agents' behaviour is governed by ML. Such models need appropriate data to be trained, which is possible thanks to a unique, big data set on investor level data accessible through the host. The objectives are: i) framework for data-driven ABM, ii) interpretable ML for ABM, iii) verification of the interpretability of data-driven ABMs using synthetic data, iv) training the data-driven ABMs using actual shareholder registration data, and finally v) analysis of investors’ decision-making mechanism. The objectives will be reached by using ML methods that achieve intrinsic interpretability with and without deep supervised learning. This research requires: a) strong numerical skills and experience with simulations, b) computer infrastructure allowing to carry out large scale numerical analysis for which the fellow and the host have complementary experience. The results will bring us closer to understanding the behavioural mechanism of market participants. The project does not just gain understanding, but introduces a data-driven approach to more realistic agent-based modelling, which is completely new. The outcome should focus the attention of regulators and policy makers, who are often unable to realistically predict the effects of considered economic measures. Finally, the project contributes to the ML literature on verification of interpretable methods with extensive data sets.