2019 7th Iranian Joint Congress on Fuzzy and Intelligent Systems (CFIS) Forecasting Stock Market Price Using Deep Neural Networks 1 st Nima Gozalpour dept. of Electrical Engineering K. N. Toosi University of Technology Tehran, Iran nimagozalpour@email.kntu.ac.ir 2 nd Mohammad Teshnehlab dept. of Electrical Engineering K. N. Toosi University of Technology Tehran, Iran teshnehlab@eetd.kntu.ac.ir Abstract—Investors in the stock market expect to make their profits, meaning they are looking for a profit by buying a share at low prices and selling it to the highest level, which requires stock price prediction. Significant challenges arise due to the non-linear and chaotic nature of stock markets when predicting prices in time series. In this research, unsupervised methods were used for the reduction of the input dimension and creating indicators. In the next step, the output of the unsupervised model was fed as input into the Jordan Recursive Neural Network(JRNN) to predict the closing price of the upcoming day. To illustrate the functionality and superiority of the proposed algorithm, data from three different and challenging stock symbols of Nasdaq were used. Ultimately, the model was evaluated with mean square error and mean absolute percentage error. Index Terms—Forecasting, Principal Component Analysis, Auto-encoder, Jordan Recursive Neural Network. I. I NTRODUCTION The necessity of investment for the development of each country is indisputable. Directing and administering funds, liquid assets and personal savings towards productive invest- ments will increase the production and growth of the gross national income and ultimately create jobs and increase per capita income and welfare. In this process, the stock market can play an important role. If we have a precise prediction of each company, then the market capital can be leveraged towards these productive companies, so that the investors get the best profit with the least risk. In recent decades, with the advent of statistics and artificial intelligence, there are many ways to predict the stock market. However, the problem with these methods is the high dimensionality of data due to the different methods of calculating technical indicators or finding the sufficent length of the days that effect on next day price. To solve this problem, decision tree [1], genetic algorithm [2], and delay coordinate embedding [3] methods are used, however, in recent years, with the introduction of deep neural networks [4] with its undeniably impressive results in the field of image and audio recognition [5], [6] as a dimensionality reduction 978-1-7281-0673-1/19/$31.00 ©2019 IEEE and feature extraction method, we sought to use deep neural networks to reduce the dimensionality and creating indicators without human intervention whatsoever. II. DATASET In order to test the proposed system, data of three NAS- DAQ symbols, Microsoft Corporation, Intel Corporation and National Bankshares Inc. were used between 2007 and 2017, where each day’s data includes five variables, such as closing price, opening price, highest Price, lowest price and volume of the stock transaction. Time series systems depend heavily on the data of previous days [7]. For this reason, after normalizing any of the variables between zero and one, the input data to the system should include the information of past days, it was concluded by trial and error that 30 previous days of data are sufficient enough. Due to having 5 variables for each day, the qualities of each sample will be 150 and according to the availability of the data for the last 10 years, we will have 2489 samples for each symbol, of which 80% will be given as training data, 10% Validation and the remaining 10% were used as test data. III. Methodology A. Principal Component Analysis Principal component analysis is a technique that is used extensively to reduce the dimension and noise of data [8]. This technique finds the input data model with maximal variance. The algorithm was originally introduced by Pearson [9]. However, the modern model was formulated by Hotelling [10]. The analysis of the principal components was followed by a linear combination of data that had the maximum variance and could be separated into the main components, so that the highest variance was in the highest component, in other words, the most critical information could be extracted from the data. The algorithm is used in a variety of areas, including gene analysis and face detection. Figure 1 shows the model process. As seen in Figure 1, PCA reduces the number of features from