Multi-agent-based VaR forecasting

We analyze the predictive power of value-at-risk forecasts generated by agent-based models. Specifically, we choose variants of the agent-based models proposed by Brock & Hommes (1998) and Franke & Westerhoff (2012) and calibrate them on the S&P 500 price and return series using a two-step process that enables the models to describe time series dynamics. To obtain a general approximation of the model parameters, our first estimation is conducted with the method of simulated moments. Following this, we apply a rolling window maximum likelihood estimation to obtain the state of the agent-based models at the current time step. The value-at-risk forecasts are then generated by iterating the models forward in time. Our results reveal that agent-based models are not only suitable for value-at-risk forecasting but are also capable of outperforming common benchmark models such as GARCH models. Most notably, we find that agent-based models outperform GARCH models in highly volatile recession periods, making agent-based models the superior choice for value-at-risk forecasting in periods with a high risk of loss.