An investigation into the use of neural networks for the prediction of the stock exchange of Thailand

Stock markets are affected by many interrelated factors such as economics and politics at both national and international levels. Predicting stock indices and determining the set of relevant factors for making accurate predictions are complicated tasks. Neural networks are one of the popular approaches used for research on stock market forecast. This study developed neural networks to predict the movement direction of the next trading day of the Stock Exchange of Thailand (SET) index. The SET has yet to be studied extensively and research focused on the SET will contribute to understanding its unique characteristics and will lead to identifying relevant information to assist investment in this stock market. Experiments were carried out to determine the best network architecture, training method, and input data to use for this task. With regards network architecture, feedforward networks with three layers were used - an input layer, a hidden layer and an output layer - and networks with different numbers of nodes in the hidden layers were tested and compared. With regards training method, neural networks were trained with back-propagation and with genetic algorithms. With regards input data, three set of inputs, namely internal indicators, external indicators and a combination of both were used. The internal indicators are based on calculations derived from the SET while the external indicators are deemed to be factors beyond the control of the Thailand such as the Down Jones Index.