Abstract:
This study investigates the capability of neural network-based approaches for
predicting the direction (up or down) of a stock market. The Australian Stock
market index (AORD) was selected as the target market. It includes several
aspects: univariate feature selection from the historical time series of the target
market, intermarket analysis for finding the most relevant influential markets,
investigations of the effect of time cycles on the target market and the
discovery of the optimal neural network architectures. It was found that the
relative return series of the Open, High, Low and Close prices of the target
market, show 6-day cycles during the study period of about fourteen years.
Multilayer feedforward neural networks trained with a backpropagation
algorithm were used for the experiments. The best neural network developed
in this study has achieved 87%, 81% 83% and 81% accuracy respectively in
predicting the next-day direction of the relative return of the Open, High, Low
and Close prices of the AORD. The architecture of this network consists of 33
input features, one hidden layer with three neurons and four output neurons.
The best input features set includes the relative returns from one to six days in
the past of the Open, High, Low and Close prices of the target market, the day
of the week, and the previous day’s relative return of the Close prices of the
US S&P 500 Index, US Dow Jones Industrial Average Index, US Gold/Silver
Index, and the US Oil Index.
