Comparative Study of Machine Learning Technologies for Stock Prediction

Transcription

1 Comparative Study of Machine Learning Technologies for Stock Prediction Mr. Lohith N #1, Prof. Raghavendra T.S *2 # Computer-Science Department, CiTech, Bangalore, India Abstract When we consider trading market we come across majorly three types of it. ETF (Exchange Traded Funds), Mutual Funds and Hedge Funds. In this paper we concentrate on ETF, which works with buying and selling of different securities. Someway it can be considering as a bond between Investors and brokers. Which is very much transparent and shows liquidity in data all the time. we mainly consider GOOGL security & BSE Stock index for analysis. We make use of yahoo finance Historical time-series data-set for this research work. We have considered Bollinger bands technical indicator as a main indicator with Regression technique to forecast the BSE index. For GOOGL security we have chosen Neural Network, Random Forest and Regression algorithms as machine learning strategies along with multiple technical indicators as prediction methods for forecasting the stock price trends. Keywords Machine Learning, Prediction Model, Regression Technique, Neural Network, Random Forest, BSE, GOOGL. I. INTRODUCTION T HE Google stock market is one of the world s largest stock market security on the basis of investors consideration. Stock Market exchanges are financial institutions which allow exchangeability of different goods between buyer and seller with the stock-broker as a mediator among them. It is the only considerable market trillions of dollars business happens. Currently many prediction mechanisms can be used in multiple domains and are useful to find out different states of the future. If consider the traditional system, the value of a stock is afforded by its entry on the stock exchange and the volume of its stock transactions. The ability to predict future price forecasting, can result in either profits or losses for low or higher volumes of share transactions. deciding the moment of buying/entry or selling/exit the securities for generating revenue can be conceived as our usual problem. Currently by making use of statistics, different predictors in the technological realm have been formed with the aim of forecasting the future price values and stock market trends, most of them apply machine learning strategies to solve the profitable problem of correctly indicating the future price of the markets. Machine learning algorithms such as hybrid PCASVM [2], Artificial Neural Networks (ANN) [3], Support Vector Machines (SVM) [4] and Sentiment analysis on internet-based data sources [5][6][7][8][9], Regression analysis and so on have the need of past and present events with different features to predict and forecast the stock trends. In this paper we are making use of Regression model to analyse the Indian BSE stock index and Neural- Network, Random-Forest, Linear Regression and few technical indicators for analysing the GOOGL security. All sort of prediction techniques can be classified into three broad categories: Fundamental analysis. Technical Analysis (also called charting) Technological methods. A. Fundamental Analysis In the fundamental analysis analysts are concerned about the company, where they assess a company's past performance along with the believability of its accounts. B. Technical Analysis(Charting) In the technical analysis analysts try to find out the future price value of a stock solely based on the 31 Mr. Lohith N, Prof. Raghavendra T.S

2 potential trends of the past price. C. Technological Methods It includes the machine learning strategies for predicting the stock price. Artificial neural networks (ANNs) and Ensemble Algorithms along with Support vector machine are the most prominent techniques which are used majorly. Further sections of this paper comprises the following, Section II as Literature Survey, Section III includes machine learning models and technical indicators used for prediction. Section IV comprises the Implementation results and finally Section V includes the Conclusion part with References. II. LITERATURE SURVEY Xiaotian Jin et al [22] states that the stock market is a major part of the country s economic development and also the capital market of the country as well. They have predicted the stock price by three statistical methods i.e. SVM regression model, least regression model and ridge regression. SVM and least regression models can have nonlinear functions whereas ridge regression model has a co-linear function for the estimation of the stock price. These methods can help to obtain the independent variable which governs the stock prices based on the historical data. They used N- gram algorithm to analyse the sentiments of the people from the social media data which is obtained from Twitter. They have used Ling-Pipe [24], which is an open source software for NLP. They used a bullishness index [25] in order to specify the emotional index along with distinguished sentiment index [26]. They note that there is a very much correlation can be found between the results obtained in analysis of tweets to the actual stock price of the dates. Ayodele et al [23] have used both fundamental analysis along with technical analysis to forecast the future stock prices and thus, created a hybrid approach which combines both fundamental and technical approaches. They used 18 input variable for analysis using ANN along with the multilayer perceptron model. Technical input variables are opening price, closing price, day high and day low price whereas the input variable for the fundamental analysis included the rumour to buy/sell, the financial status of the company and so on. The results of the hybrid approach which they used was found out to be an improved one compared to the result of just the technical analysis and the predictions were considered to be adequate to be used as a reference for the investors. [27] proposes usage of predictive analytics to forecast future stock prices based on Twitter data and Yahoo Finance data to predict daily stock prices. They also performed sentiment analysis on Yahoo Finance message data and on twitter data as well. They used the first few hundred tweets to train the model and hence to add the words such as buy, long, call which provide the positive sentiment whereas words such as sell, short, could give out the negative sentiment towards the stock. They were able to predict 60 percent accurately. They could have considered over 800 virtual trades and ended up with a positive return of investment of up to 0.49 percent. In [28] have used a combination of financial news along with the stock price quotes in order to forecast the stock prices. The financial online news articles which are textual in nature are analysed for the specific keywords. They specified that the bag of words approach could be the easiest to implement and meanwhile it was found to be the least effective [29]. They obtained the proper nouns from the text in news articles to perform the analysis. With the obtained text and the stock quotes, they could build a machine-learning algorithm with SVR. They could predict the stock price after 20 minutes and gathered about 9,211 news articles. They only considered only the articles that were published during the time when the stock market was open. They could obtain a result of 8.5 % return on trades. They state that their prediction was mainly due to the financial news article analysis. Thus, their research in the direction of financial text mining is can greatly be considered. In [30] the events were taken out of the financial news in the form of textual data. Their approach does not capture the entire meaning of the sentence at all times. For example, when Company A sues Company B which is analysed word by word is not as considerable as analysing the sentence by its subject and the object such as Company A which is has been the subject matter and also the Company B being the object and the action performed is suing. This could be performed with meaningful analysis of the text sentence with deep learning 32 Mr. Lohith N, Prof. Raghavendra T.S

3 method. They specified that predicting stock trends on a daily basis has proven to be more accurate compared to predicting stock on a weekly or a monthly basis [31, 32 and 33]. They used historical news data treated as event sequences and performed semantic analysis over the sentences with the convolution neural network (CNN). The market simulations presents that their model can provide higher profit compared to the existing methods. The results obtained shows that deep learning methods used with financial news obtained from Reuters and Bloomberg have shown with simulations a return of net profit of $16,785 with the investment of $10,000. YunusYETIS [10] uses a neural networks based ANN model to predict the stock market movements. They specifically performed for NASDAQ's stock value using ANNs with a given input parameters of share market. This paper makes use of generalized feed forward networks. Zen, Zhu and Ken Tse published the paper [11] with the base paper [12] is used for presenting the result as a demo. In their paper, they use the SVMs strategy for stock prediction. First they consider four factors which are specific to the company and six macro-economic factors which may influence the stock trends were selected for further stock multivariate analysis. In their Second part, SVM was used to analyze the relationship of those factors and predicting the stock performance. Finally, all of their results suggests that SVM can be considered as a powerful predictive tool for stock predictions. Prasanna [15] proposed methodology which makes use of hybrid McNiven s formula for evaluating the truthful information of the stock by incorporating new policies based on the historic financial data of the stock. As soon as the truthfully considerable value of the stock is calculated, the stocks that were undervalued. This particular predictive data mining makes use of the neural network technique to extract the information from the stocks by making use of the hybrid McNiven s stock value for each stock in its past. The results from this stock selection shows that the proposed formula can be considered as more impactful and trustable than the stocks selected by 3D subspace clustering method. This helps the investors to select better stocks for their investment and get a better yield than using other stock selection methods. Regarding the SVR presented in [16] and [17], and using the NMSE error for an accurate prediction, the SVR can help the traders with some information about the price and also about the trend. The only problem with the regression algorithms is that, in general, they tend to have a sample delay. In a market where time and accuracy are very important to be efficient, a small delay is not allowed as a viable error, so the SVR algorithm can be used just as a filter to eliminate the wrong buy and sell signals and having an approximate idea of how big is the difference from one day to another. This is helpful to detect the big fluctuations, to prevent the large losses and to increase the large winnings. The final experimental results, obtained with Taiwan Capitalization Weighted Stock Index datasets, shows that their method can perform better compared to other methods. Mehpare Timor, Hasan Dincer and Şenol Emir [13], have used different SVM and ANN models for Stock prediction which provide maximum returns, and these models have been applied on different combinations of data sets which obtained from the balance sheets, stocks prices and the results of a comparative analysis has been presented. The findings show that both the models including financial ratios can give meaningful performance results for the stock selection. Radu Iacomin[14] The outcomes of his study shows that the algorithm of classification SVM along with of feature selection PCA would bring the success of making a reasonable profit for traders. DepeiBao [18] makes use of the time-series dataset. Professional investors gather knowledge from technical indicators which can generally depict the aggregation of market on particular time period with customizable preference. Lay-Ki Soon [19] does the comparison of both the numeric as well as symbolic stock data considering similarity aspect. Kelvin Sim [20] have used graham s rules and propose cluster approach with 3D subspace for generating rules to select the possible depreciable stocks. Here the results are not influenced by human emotions and biases but promises 60% more profits than simply applying graham rules alone. [21] have used time series data as well as sentiments obtained from Twitter data in order to predict the stock prices. The have used NLTK to do sentiment analysis process and it includes naïve Bayes classifier with inbuilt training methods as well. Data-set is from Yahoo-Finance and twitter 33 Mr. Lohith N, Prof. Raghavendra T.S

4 data regarding financial information with for specific keywords such as sell. They trained the model with SVM, logistic regression and ANN and found out that the SVM outperforms the other two in terms of accuracy to predict stock. The results with combination of both Twitter data along with the time series data yielded the same result where the SVM provided the better results out of the three methods. The sentiment analysis method they used was based on just the keywords without the analysis of the context of the entire tweet data. III. PREDICTION MODELS USED FOR GOOGL AND BSE A. Neural Network model Neural networks consist of parallelly operable simpler elementary structures. These elements are rooted with biologic nervous systems. Neural network can be trained to do a specific functionality by setting up the underlying values of the connections (weights) among layers. B. Random Forest model Random forest ensemble comprises multiple trees, which ranges from one tree to n, where n is a user specified input value. Considering python as implementation language in sk-learn library by default, it would evaluate with 10 trees. C. Linear Regression model Linear regression analysis is a basic statistical regression technique, where prediction is done among dependent and independent feature values. This method has been considered for analyzing both BSE and GOOGL stocks. D. Technical Indicators for prediction Technical indicators are statistical methods which are used for predicting the future stock price trend by analyzing the past closing price trend of a security. In this paper we have considered mainly three indicators they are RSI, EMA and SMA. implementation purpose, graphs are plotted with the help of matplotlib-library. Sklearn has been used as a machine learning library for models building. Yahoo- Finance historical data as dataset for GOOGL security and BSE index. Dataset from has been used for training the model. From has been used for testing. Table-1 list out the net profitability outcomes of all the methods. Which mainly highlights that Random forest with 62.5% profitability as a better strategy compare to other methods. Figure-1 shows dataflow diagram for the stock prediction. Figures from 2-6 shows the implementation results in graphical form. TABLE I NET PROFITABILITY OF ALL THE USED METHODS Symbol Method Net profitability in % GOOG Random Forest 62.5 GOOG Linear Regression GOOG Neural Network GOOG Technical Indicators BSE Linear Regression In the analysis part we first collect the data-set from yahoo-finance and then we preprocess with necessary techniques such as normalization. Once the data is preprocessed, we select the closing price as target feature & other features such as volume, high, low and technical indicators as input features. For faster access we store our data in the mongodb database so that we can easily experiment without being online. In the list of Technical Indicators RSI prediction is better compared to other Indicators. Once the data is predicted we then simulate the strategy to find out its net-profitability, to do that we manually define entry and exit groups for simulation, to get the final results. IV. IMPLEMENTATION RESULTS This section mainly comprises analysis and implementation results of stock market prediction. we have chosen 4 strategies for comparative study. They are Neural Network(NN), Random Forest ensemble method, Linear Regression and Technical Indicators as predictors. Python has been used as programming language for Fig. 1. Data-Flow Diagram used for stock prediction 34 Mr. Lohith N, Prof. Raghavendra T.S

5 Fig. 2. Neural Network for GOOG security Fig. 6. Technical Indicators prediction for GOOG security Fig. 3. Random Forest ensemble method for GOOG security V. CONCLUSION For stock prediction, the time series data are readily available in the variants of daily, weekly and monthly information on stock. The daily time series historical data has been chosen as data-set for forecasting the stock price. Totally four strategies are chosen for predicting the stock price. Among them random-forest ensemble method gave better accuracy compared to other models such as neural network and regression method. In the list of technical indicators RSI strategy gave better predictions compared to other indicators such as EMA, SMA_50, SMA_100, STOCH and STOCHF. Fig. 4. Linear Regression for GOOG security Fig. 5. Linear Regression for BSE Index REFERENCES [1] Stock Market Prediction. [Online]. Available: on. [2] Radu Iacomin., Feature optimization on stock market predictor, Proc. of IEEE International Conference on Development and Application Systems (DAS), pp , [3] Yeh C, Huang C., Lee S., Foreign-exchange-rate forecasting with Artificial Neural Networks, Book [4] K. Kim, Financial time series forecasting using support vector machines, Neuro-computing, pp [5] R. E. Sorace, V. S. Reinhardt, and S. A. Vaughn, Analyzing Stock Market Movements Using Twitter Sentiment Analysis, Proc. of ASONAM Advances in Social Networks Analysis and Mining, pp [6] Rajat Ahuja, Harshil Rastogi, Arpita Choudhuri and Bindu Garg, Stock market forecast using sentiment analysis, Proc. of IEEE Computing for Sustainable Global Development, pp , [7] Li Bing, Keith C. C. Chan and Carol Ou, Public Sentiment Analysis in Twitter Data for Prediction of a Company's Stock Price Movements, Proc. of 35 Mr. Lohith N, Prof. Raghavendra T.S

6 IEEE e-business Engineering (ICEBE), pp [8] Alexander Porshnev, Ilya Redkin and Alexey Shevchenko, Machine Learning in Prediction of Stock Market Indicators Based on Historical Data and Data from Twitter Sentiment Analysis, Proc. of IEEE International Conference on Data Mining Workshops, pp: , [9] Yahya Eru Cakra and Bayu Distiawan Trisedya, Stock price prediction using linear regression based on sentiment analysis, Proc. of IEEE Advanced Computer Science and Information Systems (ICACSIS), pp , [10] YunusYETIS, Halid KAPLAN, and Mo JAMSHIDI, Stock Market Prediction by Using Artificial Neural Network, Department of Electrical and Computer Engineering University of Texas at San Antonio San Antonio, Texas, USA, World Automation Congress IS I Press. [11] Zen Hu, Jie Zhu and Ken Tse, Stock Market Prediction Using Support Vector Machine, Proc. of IEEE 6th international conference on Information, Innovation Management and Industrial Engineering pp.no , [12] Wei Huang, Yoshiteru Nakamori and Shou-Yang Wang, Forecasting stock market movement direction with support vector machine, Computers & Operations Research, Volume 32, Issue 10, pp.no , oct [13] Mehpare Timor, Hasan Dincer and Şenol Emir, Performance comparison of artificial neural network (ANN) and support vector achines (SVM) models for the stock selection problem: An application on the Istanbul Stock Exchange (ISE) African Journal of Business Management Vol. 6(3), pp. no , Jan [14] Radu Iacomin, Stock Market Prediction, Proc. of IEEE International Conference on System Theory, Control and Computing (ICSTCC), pp. no , [15] S. Prasanna and D. Ezhilmaran, Estimation of True Stock Value: A Hybrid McNiven Approach with Predictive Data Mining Concepts, EJSR Vol.110_Issue1_No [16] Clark R., Support Vector Regression, Mathworks [17] Yeh C, Huang C., Lee S., A multiple-kernel support vector regression approach for stock market price forecasting, Expert Systems with Application [18] DepeiBao, and Zehong Yang, Intelligent stock trading system by turning point confirming and probabilistic reasoning, Expert Systems with Applications 34, pp.no , [19] Lay-Ki Soon and Sang Ho Lee, An Empirical Study of Similarity Search in Stock Data, Second International Workshop on Integrating AI and Data Mining (AIDM 2007). [20] Kelvin Sim, Vivekanand Gopalkrishnan, Clifton Phua and Gao Cong, 3D Subspace Clustering for Value Investing, IEEE Intelligent Systems [21] Tina, Fang and Daniel Zuo, Stock Market Prediction based on Time Series Data and Market Sentiment,. [22] Xiaotian Jin, Defung Guo and Hongjian Liu, Enhanced Stock Prediction using Social Network and Statistical Model, 2014 IEEE Workshop on Advanced Research and Technology in Industry Applications, September 2014, pp [23] Adebiyi Ayodele A, Ayo Charles K, Adebiyi Marion O and Otokiti Sunday O, Stock Price Prediction using Neural Network with Hybridized Market Indicators, Journal of Emerging Trends in Computing and Information Sciences, Vol. 3, No. 1, Jan [24] NLP Toolkit [Online]. Available: [25] Oh C and Sheng O. Investigating predictive power of stock micro blog sentiment in forecasting future stock price directional movement [26] Baker M and Wurgler J. Investor sentiment in the stock market [27] Stefan Nann, Jonas Krauss and Detlef Schoder, Predictive analytics on public data the case of stock markets, Proc. of the 21st European Conference on Information Systems, [28] Robert P. Schumaker and Hsinchun Chen, A Discrete Stock Price Prediction Engine Based on Financial News, Computer, vol.43, no.1, pp , Jan [29] Schumaker, R.P. and Chen, H., Textual Analysis of Stock Market Prediction Using Financial News Articles. in 12 th Americas Conference on Information Systems (AMCIS), [30] Trip Kucera and David White, Predictive Analytics for Sales and Marketing, Aberdeen Group, Jan [31] Boyi Xie, Rebecca J. Passonneau, Leon Wu, and Germ an G. Creamer. Semantic frames to predict stock price movement In Proc. of ACL, pages , [32] Xiao Ding, Yue Zhang, Ting Liu, and Junwen Duan. Using structured events to predict stock price movement: An empirical investigation In Proc. Of EMNLP, pp.no , Oct [33] Paul C Tetlock, Maytal Saar-Tsechansky, and Sofus Macskassy. More than words: Quantifying language to measure firms fundamentals, The Journal of Finance, 63(3): , Mr. Lohith N, Prof. Raghavendra T.S