Using Recurrent Neural Networks and the Hilbert-Huang Transform for Stock Market Prediction

Abstract The prediction of the stock market is an important and critical issue in financial field. For that reason researchers never stopped examining and searching for new solutions and models. Goal of the current dissertation is the prediction of the FTSE100 stock market index, using Elman Neural Networks and the Hilbert-Huang Transform. Five variations of Elman Neural Networks were developed aiming to predict the price of FTSE100 for the following unseen day. The input data to the neural networks included a moving window of the historical prices of FTSE100 and the corresponding Hilbert-Huang coefficients. The variation for each neural network is the length of the moving window and some average prices of previous days that were added to the one of the variations. The Hilbert-Huang coefficients were obtained using a program that was given to it as input the price series of FTSE100 and gave as output the coefficients. For each variation of the neural networks, the optimal structure and the optimal characteristics were found in order to achieve the best results. Also, some error measurements, the number of successful predictions, the percentage of successful predictions and the percentage of direction accuracy were computed for each case so that to be able to derive some conclusions and make comparisons between them. By examining the results we concluded that the Hilbert-Huang coefficients have a positive effect to the stock market prediction and especially to the prediction of the prices’ direction, since the percentage of direction accuracy of the neural network included the Hilbert-Huang coefficients reached the percentage of 57.31% in contrast to the 50.28% of the neural network that excluded the Hilbert-Huang coefficients. Also, it is quite remarkable that the low percentages of direction accuracy that the neural networks attained were because of the period of 2007 to 2009 where the prices of FTSE100 were fluctuating sharply and were included in the testing data. The neural networks were not trained in such complicated data as a result to have a weak prediction. Finally, a synopsis for the current dissertation is composed and some methodologies are suggested for a future work