Forecasting financial time-series using data mining models: A simulation study

Abstract

In this paper, we examine the static and dynamic predictive ability of artificial neural networks and random forests for financial time series within a simulation context. Our simulation design, in which we generate data from an AR(1)-GARCH(1,1) model, allows for several degrees of persistence in the mean equation to mimic the behavior of short and long-horizon asset returns. While the true data generating process beats the data mining techniques in terms of static forecasting, the novelty in this paper is to demonstrate that the data mining techniques outperform the true model under a dynamic forecasting scheme for moderate to highly persistent time series. We provide an empirical application using one-day and long-horizon returns on two exchange rates. Our empirical findings corroborate our simulation results in that the data mining models exhibit superior predictive ability for highly persistent time series. We discuss the importance of our findings for asset allocation and portfolio management.