Empirical Study on Stock Market Prediction Using

Machine Learning
Dr. Shivani Goel Dr. Pradeep Chatterjee
Rachna Sable
Bennett University, Bennett University, Head Digital Transformation
Department of Computer Science Department of Computer Science Change Management & Customer
and Engineering, and Engineering, Experience
Greater Noida, U.P. India Greater Noida, U.P. India GDC,Tata Motors,Pune
[email protected] [email protected] [email protected]

Abstract — Stock market prediction is a crucial and nothing to do with today’s worth, but tomorrow’s information
challenging task due to its nonlinear, evolutionary, complex, [2][26]. The objective of this paper is to compare the
and dynamic nature. Research on the stock market has been traditional , Machine and Deep learning algorithms and assist
an important issue for researchers in recent years. the researchers for further analysis.
Companies invest in trading the stock market. Predicting the
stock market trend accurately will minimize the risk and The rest of the paper is organized as follows. In part 2,
bring a maximum amount of profit for all the stakeholders. similar work is discussed based on various Traditional,
During the last several years, a lot of studies have been done Machine learning and Deep learning algorithms used for
to predict stock market trends using Traditional, Machine stock prediction. In the second part of related work, details
learning and deep learning techniques. This survey will assist about widely used datasets, various evaluation metrics , and
the readers & researchers in selecting algorithms that can be features used in the stock market are discussed. In part 3, the
useful for a predicting the stock market. A survey of various number of technical Indicators used for stock market
algorithms and its parameters for stock market prediction is prediction over different periods are highlighted. Conclusion
presented in this paper. ise presented in section 4.

Keywords— Stock market prediction, machine learning,

SVM, ARIMA, DAN2 Naïve Bayes, KNN, RBF I. RELATED WORK
I. INTRODUCTION
The Stock market plays a vital role in the country’s
economic growth as well as the individual economy to a large This section discusses the similar work for stock market
extent. Finding the right time to buy and sell the shares is prediction, implemented by various authors using different
dependent on predicting the trends in the stock market. The algorithms. There are many datasets that can be used for data
technique for most accurate prediction is to learn from past related to stocks. For example, S&P 500, NASDAQ, Karachi
instances and design a model to do this by using traditional & Stock Exchange (KSE), London Stock Exchange (LSE),
machine learning algorithms[1]. The Stock market trend varies NewYork Stock Exchange (NSE). Cao and Tay [1] used the
due to several factors such as political, economics, S&P 500 dataset. Time series forecasting is an essential area in
environment, society, etc. [2][3][18][19]. There are two types which variables past observations are collected and analyzed
of stock analysis. One is a fundamental analysis, which for development of the model, to describe the underlying
requires study of the company’s basics such as balance sheet, relationships, so that the model can be used to predict the
expenses and revenues, annual returns, company’s profile, and future [10]. ARIMA is one of the widely used time series
position, etc. The other one is a technical analysis ,which deals models due to its statistical properties, and it can represent
with studying the statistics generated by market activities such various types of time series like Pure Autoregressive (AR),
as historical data, past price, and volumes [4]. There are two Pure Moving Average (MA) and combined AR and MA
essential theories used in conventional approach for stock (ARMA) series [17]. Univariate model ARIMA was
market prediction namely Efficient Market Hypothesis (EMH) parametric and based on the assumption that the nature of time
introduced by Fama in 1964 which states that stock price future series is linear and stationary. ARIMA failed to capture the
is unpredictable based on the historical data [5] and Random nonlinear patterns because it is a linear model. To overcome
Walk Hypothesis (RWH) which states that stock’s future price this limitation , ANN was used due to its flexible nonlinear
is independent of its history. Tomorrow’s stock price has capabilities [17]. ANN’s performance was found to be
satisfactory because it could rely on more considerable
information, technical indicators, fundamental indicators, etc.
to make the predictions. But ANN had some critical issues like

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

overfitting , which lead to poor generalization and out of better with SVM. but the numbers of free parameters were
sample data. To resolve the mentioned issues the procedures less in SVMs as compared to BP.
such as Cross -Validation, Non-parametric Probability Density
Osman and Mustafa [9] used the same dataset to predict
Estimation [6], Modifying Training Algorithms and Pruning
parameters or hidden notes[7] were used. the stock market price using Particle Swarm Optimization
(PSO) and Least Square Support Vector Machines (LS-
To take advantage of the unique strength of both the SVM). Traditional SVM was reformulated, and the LS-SVM
models, Zhang proposed a hybrid model comprising of algorithm was presented. LS-SVM comprised of least square
combination of both ARIMA and ANN[10]. In hybrid model, function with equality constraints, leading to a linear system
ARIMA was used to analyze the linear part of the problem, which meets the KKT conditions to get an optimal solution.
and ANN was used to model the residuals provided by The methodology to accurately select the LS-SVM must be
ARIMA. By using a hybrid model , one could model linear robust against noise and will not need prior user knowledge
and nonlinear patterns separately and then combine the about the free parameter influence. The proposed model was
forecast with improving the model’s forecast and performance. based on the study of historical data, technical indicators and
Three datasets namely, Wolf’s sunspot data, the Canadian lynx optimizing LS-SVM with PSO algorithm to make daily stock
data, and British Pound/US dollar exchange rate data were price prediction. For comparison of LS-SVM and LS-SVM
used. To forecast the accuracy performance, MSE and MAD PSO model LM algorithm was used a benchmark. The
were selected. Based on the evaluation measures, it was architecture model consisted of six input vectors representing
concluded that the hybrid model outer performed each the historical data and derived technical indicators and one
component model used separately. Another version of ANN, output representing the next price. The algorithm was tested
Dynamic ANN (DAN2), was used by Guresen et al.[12] using for many companies that cover all stock sectors in the S&P
Multilayer perceptron (MLP) and Hybrid Neural Network- 500 stock market. LS-SVM PSO model converged into a
which used Generalized Auto-Regressive Conditional global minimum and was able to overcome the overfitting
Heteroscedasticity (GARCH) to extract new features. The found in ANN. The parameters of LS-SVM PSO could be
dataset used as NASDAQ, and the evaluation of the models tuned quickly and, the performance of the model was better
was done by using two evaluation metrics namely, MSE and than LS-SVM and various algorithms compared. It also
MAD. The paper concluded that the overall results showed achieved the lowest error value.
that the ANN model MLP gave the best and reliable result in
SVM was used with Radial Basis Function(RBF) to
time series forecasting , whereas the hybrid model failed to
predict the stock prices for large and small capitalizations, in
improve the outcome.
three different markets, using price feature for predicting both
Amin Hedayati, Moein[27]in their study investigated the daily and up-to-date minute frequencies[16]. SVM outputted
ANNs ability to forecast the stock market. The daily NASDAQ the optimal hyperplane , which categorized new examples.
Dataset was used for investigation. Two types of input dataset Weka and YALE data mining environments were used for
was used one was four prior days and another was nine prior experimenting . The dataset used was IBM Inc , and the
days. The Paper discussed the various Training and Transfer features used were price volatility, price momentum, sector
functions and 16 different network architectures using the same volatility, and sector momentum. SVM was found to work
were built and the two datasets were tested. The author well on massive datasets and did not give an overfitting
concluded that there is no difference between the prediction problem and generated better results.
ability of the four and nine prior working days as input
Zhu et al. [11] tried to investigate whether the trading
parameters.
volume features could improve the neural network
SVM improved the generalization property of neural performance under short, medium , and long-term forecasting
networks with the introduction of insensitive loss function scale. Component- based three-layer Feed Forward Neural
and cost penalty[8]. So SVM was modified to solve non- Network (FFNN) was employed , and trading volume under
linear regression problems. The objective of the paper was to different input selection schemes with the basic models was
examine SVM’s feasibility and functional characteristics in set up as three augmented network models to test if trading
financial forecasting. The comparison of SVM with MLP volume could significantly improve the performance. Three
trained by Back Propagation (BP) algorithms was done. The datasets namely, NASDAQ, DJIA , and STI, were used to
dataset S&P 500 daily index in the Chicago mercantile was investigate the impact of trading volume on daily, weekly and
used. Two groups of variables were taken as inputs to the monthly predictions. The observations of the experiment
neural network. The first group G1 was constructed from four were that trading volume could not fundamentally improve
lagged Relative Difference in Percentage (RDP) value based the forecasting performance for the short term, in contrast it
on five-day periods, and the transformed closing price was modestly improved for the medium term, but the augmented
obtained by subtracting a fifteen-day exponential moving model gave the best prediction for the long term.
average to eliminate the trend in price. The second group G2
Khan et al. [13][23][24][25] studied the effect of applying
was obtained by adding another three technical indicators, the
the Principal Component Analysis (PCA) to various machine
Moving Average Convergence Divergence (MACD), On
learning classifiers and they checked the errors and accuracy of
Balance Volume (OBV) and Volatility. The test was
the classifiers at both times i.e., before and after applying PCA.
conducted for both the groups of datasets, and the results
The prediction of the future trends was performed on three
were compared. The authors concluded that the forecast was
different stock markets, namely the Karachi stock market,
London stock market , and New York stock market. The three prices available. Two datasets NSE and NYSE were used ,
machine learning classifiers namely , SVM, K Nearest and the day wise closing price was recorded. The models
Neighbor (KNN), Naïve Bayes(NB), were employed. SVM were trained on Tata Motors company’s data from NSE and
increased the margin between the data and the hyperplane and for testing five companies from NSE, such as automobile,
tried to reduce the generalization error. It provided sparseness financial, IT sector were selected, and financial, petroleum
as it used a very less number of support vectors for finding sectors were chosen from NYSE for prediction. The Mean
class labels. KNN is the simple machine learning classifier in Absolute Percentage Error (MAPE) was used as an
which the stock prediction problem is mapped into a evaluation metric for the result obtained. The performance of
classification based on similarity. Training and Testing data is ARIMA and Neural network for a specific time period was
mapped into vectors, which represent dimensions of features. compared , and it was stated that neural network architecture
The Euclidian distance was used to find the distance, and the performance was better than that of ARIMA.
clusters were made based on similarity and decisions were
Perwej et al. [15] used two models namely Particle
made. Naïve Bayes is a statistical classifier as it can make the
Swarm Optimization (PSO) and Least Mean Square (LMS)
prediction of class membership-based on probabilities. It is
and comparison of both the models was made using historical
based on Bayes theorem and makes use of class conditional
data. The dataset used was the Bombay stock exchange
independence in which attributes are independent of each other
(BSE). The features of the dataset used were ending price,
given the class. After applying classifiers on all the given
opening price, the lowest and highest price in the day , and
features, PCA was used for feature reduction , and again, all
total volume of stock traded each day. The model was used to
the classifiers were employed for testing. The evaluation
predict the ending price on each day of the prediction period.
measures used were RMSE and MAE. From the experiment,
The models were compared using the MAPE evaluation
the conclusion drawn was that KNN had the highest accuracy
metric. The paper concluded that deep learning has excellent
and lowest MAE as compared to SVM and NB.
capability to extract features from a large set of raw data
Tang and Chen [2][18][26] proposed a hybrid method of without trusting the prior knowledge of predictors , and this
combining both historical prices and news as input to the makes them convenient for stock market forecast. PSO
model for predicting stock prices. RNN-LSTM for time series algorithm works optimal during the higher number of days
and CNN for high dimensional data were combined to make ahead prediction.
prediction. RNN is a form of neural network which has a
Omer, Murat [28] used Genetic algorithm and DNN to
looping/feedback structure, which gives it the ability to work
propose a stock trading system for creating buy-Sell points
with temporal data. RNN comprises of three layers: one input
based on Optimized technical parameters. In GA the best RSI
layer, the recurrent layer, and one output layer. The output of
value for buy-sell points in downtrend & uptrend are found
the first state is passed to the next state in combination with
and passed to MLP and SMA is used to show the trend. Dow
new input. The network was trained based on the comparison
Jones 30 stocks are evaluated using financial evaluation. The
between the prediction and, target due to which RNN suffers
author concluded that more technical parameters can be
from vanishing gradient problem and the training process took
combined to get better results.
a very long time. To overcome this, LSTM was used where the
hidden layers were replaced by special units/individual units Tian Xia, Qibo Sun [29] focused on feature selection
called memory blocks also called memory cells. These were problem and proposed a feature selection algorithm by
used to store the temporal state and multiplicative unit called extending GA. Daily data of CSI 300 index is used. GA is
gates. The input, output , and forget gate were used to control used for feature selection and SVM is used for classification.
the information flow. In normal FFNN, the adjacent layers are With proper feature selection the classification accuracy can
fully connected, and if the input vector has a huge dimension, definitely be improved.
it could cause an over fitting problem. To overcome this
problem, CNN was used. It had a special unit called filter or Xi Zhang, Siyu Qu [30], focused on the importance of
kernel to replace this part , and hence , each unit of output using multiple data source which can help in improving the
layer was connected to small portion of the input. Therefore, it prediction of stock market. They proposed to combine the
required less memory and was faster to train, as it focused on a events, sentiments & quantitative data and state the
smaller region. The DJIA was used as a dataset ,which was importance of each data source their correlation and effect on
divided into two parts Historical Data, and News every day’s prediction accuracy. Multi source Multiple Instance model
world news headlines crawled from Reddit World News was proposed for the prediction and concluded that the model
Channel[https://www.reddit.com/r/worldnews]. not only integrates the heterogeneous data source but their
effect on prediction of stock market is improved.
Top 25 headlines were considered for a single date. The
evaluation metrics used were Mean Accuracy. The paper Xu Jiawei, Tomohiro [31], these author proposed the
concluded that the performance of the hybrid model outer feature selection for selecting useful stock indexes. The
performed the other models that take historical data as the model worked on Quantitative & Qualitative data by using
input feature only. Long Short Term Memory (LSTM).

Hiransha et al.[14][21][22]implemented four deep A. Datasets, evaluation matrix and features

learning architecture namely, MLP, RNN, LSTM, and CNN The evaluation metrics used for stock prediction are
,for predicting the company’s stock price based on historical NMSE,MAE,DS,CP,CD, MSE and MAD. Stock prediction is
made based on multiple features like closing price, days,
relative frequency, trading volume, price volatility, stock It is clear from Fig 1., MSE is the most widely used
momentum, index volatility, index momentum. Various metrics. Mean accuracy, MAPE and MAE are also used by
algorithms, evaluation metrics , and features used by different most of the researchers for measuring the performance of
datasets are summarized in table 1. stock prediction.
TABLE I. COMPARATIVE OF DATASET, ALGORITHM, FEATURES
AND EVALUATION METRICS USED

Evaluation
Sr.
Dataset Used Algorithms used Metrics Features
No
Used
NMSE
MAE Days,
MLP-BP
1 S&P 500 DS Relative
SVM
CP Frequency
CD
NN-BP Days,
2 S&P 500 LS-SVM MSE Closing
LS-SVM PSO Price
The Wolf’s
sunspot data the ARIMA
3 Canadian lynx ANN MSE Days , Price
data British Hybrid MAD
Pound/US dollar
exchange rate data Fig 1. Percentage of Evaluation Metrics Usage from 2010 to 2018
One-step
NASDAQ
Component- based
Sign Trading B. Technical Indicators Used
4 three- layer feed
DJIA prediction Volume,
forward neural
STI rate Price
network TABLE 3: YEARWISE TECHNICAL INDICATORS USED
MSE
MLP
MSE
5 NASDAQ DAN2 Day, Price
MAD
GARCH Year Technical Indicators
Date, Open,
Low, High, 2019 SMA,EMA,OBV,ADO
Karachi, London KNN MAE Close, MA,SMA,WMA,EMA,SD,ROC,M
6 and New York SVM RMSE Volume, 2018
ACD, RSI,ADO,TR,ATR
stock exchange Naïve Bayes Accuracy Trend,
Sentiment 2017 R,RSI,SMA,SO
& Future
trend value 2016 MA,RSI,BB,OBV
RNN-LSTM Mean Price and 2015 MA,EMA,RSI,BIAS,MACD,SO,BB,ROC,TR,OBV
7 DJIA
CNN Accuracy Date , News

MLP
8 NSE,NYSE RNN MAPE Closing
Price , Days
CONCLUSION
LSTM
CNN This paper aims to study the stock market prediction using
multiple Traditional, Machine learning, and Deep learning
PSO Closing
9 BSE LSM MAPE Values, algorithms. Along with the algorithms, the survey has
Days focused on various datasets used for stock market prediction,
features of these datasets selected as input parameters and
Price
volatility,
the evaluation metrics used for comparing the results of
Stock predictions.
SVM
10 IBM Inc Log2c Momentum,
Radial Basis Log2g
Function Index
volatility,
Index
Momentum
 REFERENCES
[1] Lijuan Cao and Francis E.H. Tay, “Financial Forecasting Using [17] G.E.P. Box, G. Jenkins, “Time Series Analysis, Forecasting and
Support Vector Machines”, 2001 Springer-Verlag London Limited. Control, Holden-Day”, San Francisco, CA, 1970.
[2] Jinqi Tang, Xiong Chen, “Stock Market Prediction Based on [18] Chen. K. Zhou. Y., & Dai, F. 2015. A LSTM-based method for
Historic Prices and News Titles”, ACM ISBN 978-1-4503-6432- stock returns prediction: A case study of China stock market. IEEE
4/18/05. International Conference on Big Data. 2823-2824. IEEE.
[3] Billah, M., and Waheed, S. 2017. “Stock market prediction using an [19] Mahalakshmi, G., Sridevi, S,. & Rajaram, S. 2016. A Survey on
improved training algorithm of neural network”, International forecasting of time series data. International Conference on
Conference on Electrical, Computer and Telecommunication Computing Technologies and Intelligent Data Engineering. 1-8.
Engineering. pp. 1-4. IEEE. IEEE.
[4] Nischal Puri, Avinash Agarwal, Prakash Prasad, “A Survey on [20] S. SElvin, R. Vinayakumar, E.A. Gopalakrishnan, V.K. Menon
Machine Learning Approach for Stock Market Prediction”, Helix and K.P. Soman. (2017) “Stock Price Prediction using LSTM,
Vol. 8(5): 3705- 3709. RNN and CNN- sliding window model.” International Conference
on Advances in Computing, Communications and Informatics;
[5] Chang Sim Vui, Gan Kim Soon, Chin Kim On, and Rayner Alfred, 1643-1647.
“A Review of Stock Market Prediction with Artificial Neural
Network (ANN)”, 2013 IEEE International Conference on Control [21] Rather A. M., Agarwal A., and Sastry V. N. (2015). :Recurrent
System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, neural network and a hybrid model for prediction of stock
Penang, Malaysia.. returns.” Expert System with Applications 42(6): 3234-3241.
[6] Pan ZH, Wang XD. “Wavelet-based density estimator model for [22] Zhang G., Patuwo B. E., and Hu M/ Y. (1998). “Forecasting with
forecasting”, J Computational Intelligence in Finance 1998; 6: 6–13. artificial neral networks: The State of Art.” International Journal of
Forecasting 14(1): 35-62.
[7] Dorsey R, Sexton R. “The use of parsimonious neural networks for
forecasting financial time series”, J Computational Intelligence in [23] Mittermayer, Marc-Andre. “Forecasting intraday stock price trends
Finance 1998; 6: 24–30. with text mining techniques.” System Sciences, 2004. Proceedings
of the 37th Annual Hawaii International Conference on IEEE,
[8] Vapnik VN. “The Nature of Statistical Learning Theory”, New 2004.
York, Springer-Verlag, 1995.
[24] Schumaker, Robert P., H. Chen. “Textual analysis of stock market
[9] Osman Hegazy , Omar S. Soliman , and Mustafa Abdul Salam, “A prediction using breaking financial news: The AZFin Txt System.”
Machine Learning Model for Stock Market Prediction”, ACM Transactions on Information Systems (TOIS) Volume 27,
International Journal of Computer Science and Telecommunications Issue 2, pp:12, 2009.
volume 4, Issue 12, December 2013.
[25] Shen, Shunrong, H. Jiang, T. Zhang. “Stock Market Forecasting
[10] G. Peter Zhang, “Time series forecasting using a hybrid ARIMA using machine learning algorithms.” , 2012.
and neural network model”,2002 Elsevier Science B.V. All rights
reserved. [26] O, Bernal. S, Fok. R, Pidaparthi. 2012. Financial Market Time
Series Prediction with Recurrent Neural Networks. Citeseer.
[11] Xiaotian Zhu, Hong Wang , Li Xu , Huaizu Li, “Predicting stock
index increments by neural networks: The role of trading volume [27] Amin Hedayati Moghaddam, Moein Hedayati Moghaddam,
under different horizons”, 2007 Elsevier Ltd. All rights reserved. Morteza Esfandyari ,” Stock Market Index Prediction Using
Artificial Neural Network”,2077-1886© 2016 Universidad ESAN,
[12] Erkam Guresen, Gulgun Kayakutlu ,Tugrul U. Daim , “Using Elsevier.
artificial neural network models in stock market index prediction”,
2011 Elsevier Ltd. All rights reserved. [28] Omer Berat Sezera,c, Murat Ozbayoglua1, Erdogan Dogdub ,” A
Deep Neural-Network Based Stock Trading System Based on
[13] Khan, W. Ghazanfar, M. A. Asam, M, Iqbal, A. ,Ahmad, S. ,Javed Evolutionary Optimized Technical Analysis Parameters “,
Ali Khan, “Predicting trend in stock market exchange using Procedia Computer Science 114 (2017) 473–480.
machine learning classifiers”, Sci.Int.(Lahore),28(2),1363-
1367,2016. [29] Tian Xia, Qibo Sun, Ao Zhou, Shangguang Wang, Shilong Xiong,
Siyi Gao, Jinglin Li, Quan Yuan, «Improving the Performance of
[14] Hiransha M, Gopalakrishnan E.Ab, Vijay Krishna Menon, Soman Stock Trend Prediction by Applying GA to Feature Selection”,
K.P, “NSE Stock Market Prediction Using Deep-Learning Models”, 2018 IEEE 8th International Symposium on Cloud and Service
1877- 0509 © 2018 The Authors. Published by Elsevier Ltd. and Computing (SC2).
Data Science (ICCIDS 2018).
[30] Xi Zhang 1, (Member, Ieee), Siyu Qu1, Jieyun Huang 1, Binxing
[15] Asif Perwej, K. P. Yadav, Vishal Sood, Yusuf Perwej, “An Fang1, And Philip Yu2, “Stock Market Prediction via Multi-
Evolutionary Approach to Bombay Stock Exchange Prediction with Source Multiple Instance Learning”, 2169-35362018 IEEE.
Deep Learning Technique”, IOSR Journal of Business and Translations
Management (IOSR-JBM)
[16] V Kranthi Sai Reddy, “ Stock Market Prediction Using Machine
Learning”, International Research Journal of Engineering and
Technology, Volume: 05 Issue: 10 | Oct 2018