Nonlinear Prediction of The Standard & Poor's 500 and The Hang Seng Index under A Dynamic Increasing Sample

Abstract

This study attempts to forecast the next day’s returns of two time series in the Hang Seng Index (HSI) and Standard & Poor’s (S&P) 500 indices using Artificial Neural Networks (ANN) with past returns as input variables. Results from ANN are compared with those from the autoregressive integrated moving average (ARIMA) model. This study uses a longer time period than ARIMA (i.e., daily data of 80 and 35 years for the S&P 500 and HSI, respectively) to develop and test the models. The two competing models are rigorously evaluated in terms of widely-used penalty-based criteria, such as directional accuracy, as well as in terms of trading performance criteria like annualised return, the Sharpe ratio and annualised volatility via a simple trading strategy. Moreover, the robustness of the two models is tested for 36 test periods. Empirical results show that ANN works better than ARIMA and delivers consistent results across the periods tested. These results support ANN’s robustness and its use in formulating a strategy for trading in the S&P 500 and HSI time series.