الگوی ترکیبی پیش بینی رفتارهای متغیر قیمت در بازار سهام (مقاله علمی وزارت علوم)
درجه علمی: نشریه علمی (وزارت علوم)
آرشیو
چکیده
هدف پژوهش حاضر، ارزیابی روش های فرا ابتکاری جهت پیش بینی رفتار قیمت سهام و معرفی کارآمدترین روش در بازار سهام ایران است. بدلیل نااطمینانی در زمینه سرمایه گذاری و کثرت متغیرها، سرمایه گذاران به روش پیش بینی روی می آورند که به واسطه آن ها، تخمین هایشان به واقعیت نزدیک و خطایشان کم شود. در این پژوهش، به پیش بینی قیمت سهام 5 شرکت پذیرفته شده در شاخص فلزات اساسی بورس اوراق بهادار تهران در یک بازه زمانی سه ساله، با شرط فقدان توقف معاملاتی پیوسته برای مدت بیش از 3 ماه پرداخته شد. بدین منظور، متغیرهای بهینه از بین 9 متغیر اولیه و پرکاربرد با استفاده از روش های انتخاب ویژگی، الگوریتم های فرا ابتکاری شاهین هریس و وال انتخاب و سپس با استفاده از شبکه های عصبی پس انتشار خطا، شبکه عصبی پایه شعاعی و شبکه عصبی با تأخیر زمان به پیش بینی قیمت سهام پرداخته شد. نتایج نشان داد که در پیش بینی قیمت سهام فملی، زنگان، فرآور، فاسمین و فولاد به ترتیب WOATD، HHOTD، HHOTD، HHOTD و HHORBF مدل برتر هستند. همچنین نتایج نشان می دهد که روش تکاملی شاهین هریس در یافتن ویژگی ها نسبت به روش تکاملی وال بهتر عمل کرده است. با توجه به نتایج، مدل HHOTD نسبت به بقیه مدل ها از دقت و کارایی بالاتری برخوردار است.A hybrid model for predicting variable price behaviors in the stock market
IntroductionThe research’s aim is to evaluate meta-innovative methods to predict stock price behavior and introduce the most efficient method in the Iranian stock market. Due to the uncertainty in the field of investment and the multitude of variables, investors turn to the forecasting method, by which their estimates are close to reality and their errors are reduced. Forecasting the stock market has many complexities as it is affected by a number of economic and non-economic factors such as market news, political events, social chaos and so on (B. Labiad, 2016). Due to the volatility of the stock market, predicting stock price changes in daily trading is a challenging task and it is one of the most important issues in the financial world that has attracted the attention of most financial analysts and researchers. Financial markets are an attractive field for investment in order to achieve high profits but obtaining high profits in this market is not easy, because some data are dynamic, non-linear, and variable, and unstable in nature, and therefore there are many risks in this way (Lahmiri & Boukadoum, 2015). Recently, computational intelligence (CI) techniques such as artificial neural network (ANN), nature-inspired and fuzzy algorithms have been developed for stock price prediction due to their ability to handle noisy data in the financial market(Lahmiri & Boukadoum, 2015). In many cases, ANNs suffer from limitations such as local minima, slow convergence, long training time, and fitting difficulty due to the large number of parameters to adjust(Chandar, 2018). These problems have been neglected in much previous research. Therefore, the main goal of this research is to develop a forecasting model consisting of artificial neural networks and nature-inspired algorithms to investigate the above issues and improve forecasting accuracy. In this regard, this research seeks to answer this question: Can the hybrid models of neural networks and meta-heuristic algorithms provide accurate stock predictions?Methodology In this study, the stock prices of 5 listed companies in the base metals index of Tehran Stock Exchange in the period 1396 to 1398 were predicted. For this purpose, the optimal variables are selected from among 9 primary and widely used variables using feature selection methods, super-innovative algorithms of Harris and Wall's algorithms and then using Back propagation neural networks, radial base neural network and neural network with time delay Stock prices were predicted. Focusing on basic metal companies in Tehran Stock Exchange as a statistical population, 21 companies with the highest correlation coefficient were finally selected as a sample. To generalize the sample as much as possible to the society from among the companies, 5 companies that included small, medium, and large companies were selected as samples. The two symbols Zangan and Faravar were among the smallest companies and the two companies Femli and Foulad were among the largest companies, which were included in the sample as representatives of small and large companies, respectively, and the symbol of Fasmin was included in the sample as a representative of medium-sized companies. The required data: opening price, closing price, highest price, lowest price, closing price, volume of transactions, value of transactions, number of transactions and range of changes were collected from Tehran Stock Exchange Technology Management Company (www.tsetmc.com) and Codal website. MATLAB software was used to perform calculations and implement the proposed method.Results and Discussion The results show that in the symbol of Zangan, the combined models HHOBP, WOARBF and HHOTD have the highest R2 coefficient, respectively. In terms of MSE and MAE criteria, the HHOTD model has the lowest error rate and the lowest absolute average of the difference between the actual and predicted data, which shows that in this model, the predicted values are closer to the actual values. According to the Hitrate criterion, the HHOTD model has the highest correct estimation (89%) compared to the rest of the models. With these results, it can be concluded that Harris Hawk algorithm has the highest accuracy and the neural network with time delay has high efficiency.Conclusion Since linear models do not have the ability to understand and extract non-linear patterns such as stock price time series, the goal in this research was to combine artificial neural networks and meta-heuristic algorithms to accurately predict stock price trends. In this research, first, the optimal variables were selected using feature selection methods, meta-heuristic algorithms of Harris hawk and Whale. Then, by using neural networks after propagation of error, radial basis neural network and neural network with time delay, the stock price was predicted. Considering the findings, it is suggested that the Harris's hawk evolutionary method is superior to the whale's evolutionary algorithm. Proper navigation of the problem space and finding features is the reason for the superiority of Harris's hawk results compared to whale's algorithm. It was also shown in this research that networks with time delay have higher efficiency. Findings indicate that the hybrid method of Harris hawk and neural network with time delay is the most effective. It is suggested to use a hybrid model of post-error propagation neural network, radial base neural network and time delay neural network with evolutionary algorithms such as genetics in future research. Also, it is suggested to use other predictive algorithms such as support vector regression in combination with neural network with time delay.
تبلیغات
