STOCK MARKET DIRECTIONAL CHANGE FORECASTING USING ARTIFICAL NEURAL NETWORKS AND DYNAMIC BINARY TIME SERIES MODELS.

Abstract

In this study, we forecast the daily direction of the S&P 500 index using static and dynamic probit models and Artificial Neural Networks (ANNs). The first two models are constructed using binary dependent variable regression, whereas the ANNs are estimated using White’s (2006) QuickNet algorithm to mitigate the problems inherent in nonlinear optimisation. We find that the dynamic probit models yield the greatest accuracy and trading returns. Surprisingly, additional explanatory variables do not improve the dynamic probit models’ forecast performance. While the ANNs’ are comparable with the dynamic probit models in terms of accuracy, the former produced much lower returns at shorter investment horizons due to their high sensitivity to transaction costs. We find that the ANNs yield higher returns at longer investment horizons and propose adjustments to QuickNet to further improve its out of-sample forecasting performance.