A GA-based Polynomial FLANN with Exploration and Incorporation of Virtual Data Points for Financial Time Series Forecasting
Subhranginee Das1, Sarat Chandra Nayak2, Sanjib Kumar Nayak3, Biswajit Sahoo4

1Subhranginee Das, School of Computer Engineering, KIIT University, Bhubaneswar (Odisha), India.
2Sarat Chandra Nayak, Department of Computer Science and Engineering, CMR College of Engineering & Technology, Hyderabad (Telangana), India.
3Sanjib Kumar Nayak, Department of Computer Application, Veer Surendra Sai University of Technology, Burla (Odisha), India.
4Biswajit Sahoo, Telangana, India
Manuscript received on 23 March 2019 | Revised Manuscript received on 04 April 2019 | Manuscript Published on 18 April 2019 | PP: 422-430 | Volume-7 Issue-6S March 2019 | Retrieval Number: F02830376S19/2019©BEIESP
Open Access | Editorial and Publishing Policies | Cite | Mendeley | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Abstract: Global stock markets across the world continue to see a phenomenal increase in adoption and interest over the years, largely because of rapid expansion of international financial links, coupled with liberalization in stock markets. Individuals as well as financial institutions across the globe engage in trading stocks and derivatives, attempting to leverage benefits, associated with accurate prediction of the price trends and value. However, non-linear nature of the stock movements and their varied levels of volatility, adds to the challenges of prediction accuracy. While diverse artificial computational models, particularly multilayer perceptron (MLP) is arguably most frequently used forecasting model because of its good approximation and generalization abilities, the model’s computational complexities amplify manifold with increase in number of layers. Also, a greater neuron density in each layer and its black-box nature compels researchers to adopt computationally simpler models such as functional link artificial neural networks (FLANN). This article proposes a polynomial functional link artificial neural network (Poly-FLANN) model for stock movement forecasting. The proposed forecasting model is applied to forecast daily closing indices of BSE, NASDAQ, FTSE, TAIEX, and DJIA. Compared to the conventional FLANN, the process of functional expansion of input data is carried out after two data processing methods such as exploration and incorporation of virtual data points (VDP) to the original stock prices and combination of input data. Addition of VDPs to the original financial time series expands the volume of time series, where as taking combinations of data elements increases the dimension of the original financial time series. The performance of the proposed model is compared with trigonometric FLANN (TFLANN), Chebysheb FLANN (CFLANN), Lagurre FLANN (LFLANN), Legendre FLANN (LeFLANN), and a MLP model. All these models are trained by genetic algorithm (GA). Extensive simulation results prove the efficiency of the proposed model in terms of generating less forecasting error signals.
Keywords: Stock Market Forecasting; Multilayer Perceptron; Functional Artificial Neural Network; Virtual Data Position; Genetic Algorithm.
Scope of the Article: Virtual & Overlay Networks