Who Drives the Market? Estimating a Heterogeneous Agent-based Financial Market Model Using a Neural Network Approach

We propose a method for estimating complex heterogeneous agent-based models, especially their time-varying micro data, based on time-varying real-world macro data . We estimate the model at high frequency without posing simplifying assumptions on the model or the estimation process. We estimate daily time series of market participants’ trading strategies, i.e., chartists and fundamentalists, at the S&P 500. For this context, heterogeneous agent-based models which explain macro market behavior by time-varying usage of strategies on the micro level have shown superiority to alternative models. Due to complexity, these agent-based models can hardly be directly estimated. As micro-level data from real stock markets are largely unobservable, model-free estimation methods cannot be applied to map macro to micro variables. Thus, we suggest a combination of both methods in terms of a model-free estimation of the inverse of an agent-based model, mapping macro to micro variables, which can then be applied to real-world macro data. Using an artificial neural network we estimate an inverse model of the heterogeneous agent-based financial market model introduced by Lux and Marchesi (1999) and apply it to S&P 500 data. Comparisons with previously estimated yearly time series and with historic events illustrate validity of the estimation results. Our results also contribute to the understanding of theoretical models.