The Importance of Time Series Data Filtering for Predicting the Direction of Stock Market Movement Using Neural Networks

Abstract

Predicting future trends in the stock market from time-series data is a challenging task due to its high non-linear nature caused by the complexity involved in the trading process. This paper emphasizes the importance of time-series data filtering when neural network models are used for stock market direction forecasting. Performances of three different neural network models are compared on raw data, processed data with simple moving average, and data filtered with discrete wavelet transformation. Applying wavelet transformation on input financial data as a processing step shows better results than the use of raw financial data or simple moving average. Also, among tested neural network models, the better results are obtained by using long short-term neural network then by using other neural network models.