Academic journal article Journal of Management Information and Decision Sciences

Integration of Two Different Signal Processing Techniques with Artificial Neural Network for Stock Market Forecasting

Academic journal article Journal of Management Information and Decision Sciences

Integration of Two Different Signal Processing Techniques with Artificial Neural Network for Stock Market Forecasting

Article excerpt

INTRODUCTION

One of an active portfolio management's critical tasks is forecasting stock market variations with time. Although forecasting of the stock market is crucial, many techniques have been explored in this area but did not fully succeed to achieve satisfactory levels of prediction accuracy. Especially, during uncertain local and global circumstances and events, stock market indices behave with a high degree of uncertainty. In the challenging times of the economy, investors are very careful and conservative about their investments and their financial security. According to the Modern Portfolio Theory, non-systemic risk can be reduced by mutual funds and index funds. Whereas, the systematic risk due to the unstable economic and financial situations cannot be controlled. Technical trading techniques claim that by sending buying and selling signals on a regular basis, investors can make better investment decisions even in volatile market conditions. Given that claim, establishing a technical trading mode that can reliably predict stock market indices in very challenging and dynamic economic situations requires refined and advanced solutions.

One of the most popular models which has been investigated and used by many researchers in this area of research is Artificial Neural Networks (ANNs). ANNs are mathematical models that attempt to emulate the human brain neural network and its reasoning process to recognize patterns. The power of ANNs is represented in their learning ability from training on incomplete, imprecise, and partially incorrect examples. Its unique advantage makes it well suited to deal with unstructured problems, inconsistent information, and real-time output (Trippi & Turban, 1996). ANNs have many useful and reliable purposes, including: clustering, classification, and recognition in many fields of research. They are applied to forecast stock and commodity prices, bond ratings, foreign exchange rates, T-bills, bonds, and inflation (Aiken, 1999; Krishnaswamy et al., 2000; Sharma & Alade, 1999).

Several studies have been published in the last two decades that suggested that the ANNs is more reliable than other traditional forecasting techniques (Sharda & Patil, 1990; Tang, 1991; Trippi & Turban, 1993; Atsalakis & Valavanis, 2009). ANNs is a robust predictive model for challenging time series via training to approximate the hidden patterns in a time series. Although ANNs is a reliable predictive model, its effectiveness relies on a number of factors such as: learning algorithm, quality of training data sets, network architecture, etc. (Sharma & Alade, 1999). Further, one of the most important factors is reported to be noise content in the signal that significantly destabilizes the performance of ANNs (Kim, 2006). Therefore, it is of great interest and potential to investigate the benefits of signal processing techniques that can be used for signal de-noising as a preprocessing stage. A very interesting observation made by (Nair et al., 2010) that filtering techniques are widely used in general signals but rarely in financial signals. So, from the preceding discussion, we propose to utilize the reliability of the ANNs and the de -nosing capability of the signal processing techniques in an integrated predictive model for financial signals, especially stock market indices. In this proposed method, the feature vectors are modeled as k-entry days of the stock market close day data. Based on the multi-layer perceptron (MLP) BP-ANN architecture requirements, the classes of feature vectors are taken as the last entry of each vector. Since the stock market data is real data, the vectors are discretized into a suitable number of levels representing the range of classes. The training and testing data was collected from DOW30 and NASDAQ 100 over 12 years where the experimental results demonstrated the reliability of the proposed approach with average prediction accuracy of 98. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.