Application of Feed-Forward Neural Networks Smoothing Transition Autoregressive Models in Stock Returns Forecasting

Abstract

In this paper we propose and examine new approaches in smoothing transition autoregressive (STAR) models. Firstly, a new STAR function is proposed, which is the hyperbolic tangent sigmoid function. Secondly, we propose Feed-Forward Neural Networks Smoothing Transition Autoregressive (FFNN-STAR) models. We examine the stock returns of US S&P 500, FTSE-100 in UK stock index, DAX index in Germany and CAC-40 in France and we apply bootstrapping ordinary least squares simulated regressions, while also GARCH models with bootstrapping simulations can be applied as well. The results are in favor of neural networks, while in almost all cases the forecasting performance of Feed-Forward Neural Networks STAR models is superior to conventional STAR models. This paper can be a guide and set up the fundamentals for further advanced research in econometrics and time-series analysis.

Keywords: Feed-Forward Neural Networks, Smoothing Transition Autoregressive Models, Stock Returns, Forecasting