Financial time series forecasting using conditional Restricted Boltzmann Machine

Abstract

Inspired by the success of deep learning in big data image recognition in Restricted Boltzmann Machine, Conditional Restricted Boltzmann Machine which was the original design to forecast human motion movement had been modified to forecast financial time series. As far as the author is aware of, this is the first attempt to apply deep learning in financial time series forecasting. Conventionally, deep learning is applied in image classification and several layers of deep learning in the huge dataset could increase its accuracy. The traditional forecasting method is using Euclidean distance to map the dataset into a higher dimension which facilitates to draw a hyperplane to separate the data. The more the cluster of the data in the hyperplane, the closer the distance of those neighbour data. As a result, those cluster data are the foundation to forecast. A new approach in Restricted Boltzmann Machine is to assign low energy based on probability concept to those connections that are relevant to each other while high energy is assigned to those that are irrelevant. The advantage of this method over Euclidean distance is that the probability energy assignment can be done one layer at a time and extend to many layers. Each layer information is retained and passed on to another layer to be trained again. As a consequence, all the information in the dataset is carefully scrutinized to obtain the best result. In this research, it has been demonstrated in the following Chapters that deep learning using modified Conditional Restricted Boltzmann Machine is able to handle high dimensionality data which is over 100 with the dataset array as big as 600000x100. This setup enables it to capture the information of the high dimensions in each layer. Eventually, it will improve the forecasting accuracy. This was not possible before as our previous research has experienced. Historical records are not as important as the dimension of the financial time series problem domain. 30 or 20 years of stock history may not have that much impact on the current stock price in one stock. As the financial market is closely related to other markets, the stock price of a particular security is heavily dependent on other stocks in the same market as well as the performance of other markets. Hence, it is more important to increase the dimensionality or features of the data. In other words, including more factors such as the price of others stocks, economic factors such as interest rates and GDP can enhance the performance. The algorithm based on Conditional Restricted Boltzmann Machine has demonstrated remarkable forecasting accuracy as reported in Chapter 4 and 5.