Article 2415

Title of the article

STUDY OF STOCK PRICE PREDICTION ERROR USING THE WANG-MENDEL’S FUZZY NEURAL NETWORK MODEL 

Authors

Soldatova Ol'ga Petrovna, Candidate of engineering sciences, associate professor, sub-department of information systems and technologies, Samara State Aerospace University named after S. P. Korolev (National Research University) (34 Moskovskoe highway, Samara, Russia), op-soldatova@yandex.ru

Index UDK

004.032.26

Abstract

Background. Stock prices forecasting is an important task, but the problem of identifying a correlation between a predicted sequence and exchange prices for other goods is important too. The goal of this study is to investigate the effectiveness of stock prices forecasting with a correlation of commodity prices by using the Wang-Mendel’s fuzzy neural network model.
Materials and methods. All studies were conducted in custom software implementing the Wang-Mendel’s network model; the network used three different training algorithms and displayed the results of learning and forecasting as graphic lines and error values. The article describes a method for determining correlations between training samples from graphs of correlation functions and time delay values, adduces calculations of the forecasting sequence correlation and Wang-Mendel’s network training. The research used prices of the following oil and gas companies: "LUKOIL", "Rosneft" and "Transneft". To investigate the correlation the author used prices of two oil markers – Brent and WTI, as well as fuel oil prices.
Results. Usage of the correlation between the forecasting sequence and the training samples improves the predictive ability of the Wang-Mendel’s network, significantly reducing the maximum relative error of the forecast. Analysis of the forecasting effectiveness shows the advantage of using at least two correlated samples with an average delay (not more than 60–70 days). Analysis of the training algorithms influence on the forecasting error shows the advantage of the adaptive algorithm and the steepest descent algorithm with k-averages initialization in comparison with the steepest descent algorithm with adaptive initialization.
Conclusions. The proposed methods, models and algorithms, as well as the conducted research have provided numerical estimates of the forecasting error with and without using the correlation between training samples, on this basis of which one can make a choice of correlating samples and network training algorithms.

Key words

stock prices, forecasting sequence, correlation, training sample, Wang-Mendel‘s fuzzy neural network, adaptive learning algorithm, steepest descent algorithm, k-averages algorithm, forecasting error.

Download PDF
References

1. Ossovskiy S. Neyronnye seti dlya obrabotki informatsii [Neural networks for information processing]. Transl. from Polish by I. D. Rudinskiy. Moscow: Finansy i statistika, 2002, 344 p.
2. Khashei M., Bijari M. Applied Soft Computing. 2011, vol. 11, no. 2, pp. 2664–2675.
3. Zhang G. P. Handbook of Natural Computing. Berlin: Springer Berlin Heidelberg, 2012, pp. 461–477.
4. Yan W. Neural Networks and Learning Systems, IEEE Transactions on. 2012, vol. 23, no. 7, pp. 1028–1039.
5. Khashei M., Bijari M. Journal of Intelligent and Fuzzy Systems. 2014, vol. 26, no. 2, pp. 831–845.
6. Soldatova O. P., Pankratova D. A., Litvinov V. G. Perspektivnye informatsionnye tekhnologii (PIT 2015): tr. Mezhdunar. nauch.-tekhn. konf. (Samara, SGAU, 28–30 aprelya 2015 g.) [Prospective information technologies (PIT 2015): proceedings of the International scientific and technical conference (Samara, SSAU, 28–30 April 2015)]. Samara: Izd-vo Samar. nauch. tsentra RAN, 2015, vol. 1, pp. 214–217.
7. Tseny syr'evykh tovarov [Primary product prices]. Available at: http://www.finanz.ru/birzhevyye-tovary.

 

Дата создания: 12.05.2016 10:49
Дата обновления: 12.05.2016 12:49