See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/358235790

Machine Learning Models in Stock Market Prediction

Article in International Journal of Innovative Technology and Exploring Engineering · January 2022
DOI: 10.35940/ijitee.C9733.0111322

CITATIONS READS

25 1,224

1 author:

Gurjeet Singh

17 PUBLICATIONS 82 CITATIONS

SEE PROFILE

All content following this page was uploaded by Gurjeet Singh on 06 February 2022.

The user has requested enhancement of the downloaded file.

 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-11 Issue-3, January 2022

Machine Learning Models in Stock Market

Prediction
Gurjeet Singh

Abstract: The paper focuses on predicting the Nifty 50 Index by Very less researches in the literature are found to predict
using 8 Supervised Machine Learning Models. The techniques the movements of index in the Indian Stock Market
used for empirical study are Adaptive Boost (AdaBoost), k- especially Nifty 50 Index.
Nearest Neighbors (kNN), Linear Regression (LR), Artificial
Neural Network (ANN), Random Forest (RF), Stochastic My previous research works in the area of Stock Market
Gradient Descent (SGD), Support Vector Machine (SVM) and (Singh & Nagar, 2012), (Nagar & Singh, 2012) and (Nagar
Decision Trees (DT). Experiments are based on historical data of & Issar, 2013) motivated me to do further study as still there
Nifty 50 Index of Indian Stock Market from 22nd April, 1996 to are very limited researches have been done in the area of
16th April, 2021, which is time series data of around 25 years. Stock Market.
During the period there were 6220 trading days excluding all the The NIFTY 50 Index [39] consists of 50 stocks from 13
non trading days. The entire trading dataset was divided into 4
subsets of different size-25% of entire data, 50% of entire data, different sectors of the Indian economy.
75% of entire data and entire data. Each subset was further ANN, LR, SGD, SVM, AdaBoost, RF, kNN and DT
divided into 2 parts-training data and testing data. After applying Machine Learning techniques are used for modeling and
3 tests- Test on Training Data, Test on Testing Data and Cross predicting the movements of the index.
Validation Test on each subset, the prediction performance of the The main objective of this empirical study is to explain
used models were compared and after comparison, very and validate the expectedness of Stock Index movements
interesting results were found. The evaluation results indicate
that Adaptive Boost, k- Nearest Neighbors, Random Forest and using the above models and to make comparison of the
Decision Trees under performed with increase in the size of data performance of used techniques.
set. Linear Regression and Artificial Neural Network shown This research paper is divided into six sections. Section II
almost similar prediction results among all the models but provides the literature review of predicting Stock Market
Artificial Neural Network took more time in training and using Machine Learning. Section III explains the research
validating the model. Thereafter Support Vector Machine data. Section IV tells about the Prediction Models used in
performed better among rest of the models but with increase in
the size of data set, Stochastic Gradient Descent performed better this empirical study. Section V describes about the
than Support Vector Machine. experiments and empirical results from the comparative
Keywords: Artificial Neural Network, Stock, Market
analysis. And the last section, Section VI contains the
Prediction, Supervised Machine Learning Models, Time Series concluding remarks.
Data
II. LITERATURE REVIEW
I. INTRODUCTION Using Artificial Neural Networks and Support Vector
Machines, Y.Kara et al. (2011) attempted to predict the
S tock Markets are the most popular financial market movements of stock price in the Istanbul Stock Exchange.
instrument. Perfect prediction of Stock Market Indices and They prepared 2 prediction models and compared their
stocks price is very difficult due to its highly dynamic performances. The average prediction performance of the
nature. According to Abu-Mostafa & Atiya (1996), stock ANN model was 75.74% and SVM model was 71.52%.
market prediction is considered as a very difficult task in Alaa F et al. (2015) explored the use of Multiple Linear
financial time series prediction. According to Tan, Quek, & Regression and Support Vector Machine (SVM) to develop
See (2007), stock market is affected by many macro models for predicting the S&P 500 stock market index. 27
economical factors such as political events, firm’s policies, potential financial and economic variables which impact the
general economic conditions, investors’ expectations, stock movement were adopted to build a relationship
institutional investors’ decisions, other stock market between the stock index and these variables. The
movements, and investors psychology etc. Lots of empirical constructed SVM model with Radial Basis Function (RBF)
studies dealt with the predicting stock index movement for kernel model provided good prediction capabilities with
the developed financial markets. respect to the Regression and ANN models. Madge and
Bhatt (2015) predicted the movements of stock price using
SVM. 34 technology stocks and 4 parameters (technical
Manuscript received on January 15, 2022.
Revised Manuscript received on January 24, 2022. indicators) were taken into consideration in their study. The
Manuscript published on February 28, 2022. work concluded that in the short-term predictions there was
* Correspondence Author very low accuracy, on the other hand in the long-term
Dr. Gurjeet Singh, Associate Professor & Dean, Department of Lords prediction the prediction accuracy stood between 55 and
School of Computer Applications & IT, Lords University, Alwar,
Rajasthan, India. E-mail: research.gurjeet@gmail.com 60%. Aparna Nayak et al. (2016) used historical and social
media data.
© The Authors. Published by Blue Eyes Intelligence Engineering and
Sciences Publication (BEIESP). This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 18 ©Copyright: All rights reserved.
 Machine Learning Models in Stock Market Prediction

Prediction models were constructed using Decision III. RESEARCH DATA

Boosted Tree, Support Vector Machine and Logistic
Table 1. Data Sets
Regression to predict Indian Stock Market.
It was found that Decision Boosted Tree performed the
best among the 3 models. Dash and Dash (2016) constructed
a model which is called computational efficient functional
link artificial neural network (CEFLANN) which generates
the decisions of trading. CEFLANN provides profitable
decision making with technical analysis rules. CEFLANN
performance was compared with KNN, SVM, and DT, and
CEFLANN provided the highest percentage of profit. The
prediction of stock market behavior can depend on different
types of ANNs because different type of ANNs can give
different performance. It was observed that an input layer,
hidden layer, and output layer combination can affect the
performance of ANN. To get better performance, a right
combination of layers is necessary. Sujin Pyo et al. (2017)
employed an artificial neural networks (ANN), support
vector machines (SVMs) with polynomial kernels, and
radial basis function (RBF) kernels to predict the trend of
the Korea Stock Price Index 200 (KOSPI 200) prices. In this
study controversial issues were stated and hypotheses about
the issues were tested. The results were inconsistent in
comparison of the earlier research, which were assumed to
have higher prediction rate of performance. Further it was
proved by Google Trends that these were not effective Fig. 1. Data Sets of different size (12.5%, 25%, 50%
and 100%)
factors in the prediction of the KOSPI 200 index prices.
Next, the study also showed that ensemble methods do not This section describes the research data and the selection
improve the prediction accuracy. For prediction of the stock of Data Sets. The research data used in this study is the
prices of having different capitalizations and different direction of daily closing price movement in the Nifty 50
markets, Henrique et al. (2018) have used Support Vector Index. The entire data set covers the period from April 22,
Regression (SVR). The results obtained were indicating that 1996 to April 16, 2021. During the period, total numbers of
SVR has a better predictive capability. Indu Kumar et al. trading days were 6220.
(2018) have used Supervised Machine Learning Algorithms- As shown in the Table 1 and Fig.1, four subsets were
Random Forest, SVM, KNN, Softmax Algorithm and Naive derived from the entire data set and each subset was further
Bayes Algorithm for the prediction of stock price divided into 2 parts, Training Data Set (80% of the Data Set)
movements. The results of study revealed that in large and Testing Data Set (20% of the Data Set). The first subset
dataset, Random Forest Algorithm had outperformed in is 12.5% of the entire data set and contains 778 trading
comparison of all the other algorithms but with reduced size day’s data. This data set is called ‘‘Data Set-1’’. The Data
of the dataset (half of the original) Naïve Bayes Algorithm Set-1 covers the period from April 22, 1996 to June 8, 1999,
performed best accuracy wise. Next, decrease in the number containing 622 trading day’s data in Training Data Set and
of technical indicators will decrease the accuracy of each 156 trading day’s data in Testing Data Set. The second
model in prediction of the stock market trends. Sheikh subset is 25% of the entire data set and contains 1555
Mohammad Idrees et al. (2019) tried to analyze the time trading day’s data. This data set is called ‘‘Data Set-2’’. The
series data of Indian stock market. ARIMA model which is Data Set-2 covers the period from April 22, 1996 to July 10,
a statistical model was constructed to predict the Indian 2002, containing 1244 trading day’s data in Training Data
stock market volatility. Andrea Picasso et al. (2019) Set and 311 trading day’s data in Testing Data Set. The third
combined Technical and Fundamental analysis and subset is 50% of the entire data set and containing 3110
constructed models to predict the market using Random trading day’s data.
Forest, Support Vector Machines, and a feed forward Neural This data set is called ‘‘Data Set-3’’. The Data Set-3
Network applied to time series pre- diction and sentiment covers the period from April 22, 1996 to September 19,
analysis. Irfan Ramzan Parray et al. (2020) attempted to 2008, containing 2488 trading day’s data in Training Data
predict NIFTY 50 index using SVM, Perceptron, and Set and 622 trading day’s data in Testing Data Set. The
Logistic Regression. The SVM produced the best accuracy fourth subset is 100% of the entire data set and contains
results among all algorithms. 6220 trading day’s data. This data set is called ‘‘Data Set-
4’’. The Data Set-4 covers the period from April 22, 1996 to
April 16, 2021, containing 4976 trading day’s data in
Training Data Set and 1244 trading day’s
data in Testing Data Set.

Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 19 ©Copyright: All rights reserved.
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-11 Issue-3, January 2022

The performance comparison was performed on the Data Detection, Stock Market Prediction etc. Orange Data
Set-1, Data Set-2, Data Set-3 and Data Set-4 (entire Data Mining Tool (Demsar, J. et al., 2013) provide Neural
Set) using ANN, LR, SGD, SVM, AdaBoost, RF, KNN and Network widget which uses Multi-Layer Perceptron (MLP)
DT models. The results for each model and Data Set were with back propagation, algorithm that can learn non-linear
compared to each other. models as well as linear is used in this study. MLP is chosen
as it is well suited in stock market prediction.
IV. METHODOLOGY
E. Random Forest (RF)
A. Adaptive Boosting (AdaBoost) According to Breiman et al. (2001), in Random Forest, 2
Boosting method is an ensemble method which is mostly parameters are required: (i) ntree- the number of trees (ii)
used for predictions. This method is a group of algorithms mtry- the number of features in each split. Many studies
which converts weak learners to a powerful learner. By have declared that the suitable results could be obtained with
using boosting weak learners are trained sequentially to the default parameters (Duro et al. and Immitzer et al.
correct its past performance. According to R.E. Schapire (2012); Liaw et al. (2002); Zhang et al. (2017)). According
(2003) AdaBoost was the initial boosting algorithm which to Liaw et al. (2002), the more number of trees are giving
was successfully being used to meet the requirements of more stable results. According to Breiman et al. (2001) use
binary classification. AdaBoost algorithm is used to raise the of more than the required number of trees may not be
performance of any machine learning algorithm, and it is necessary, however this is not harmful to the model.
mostly used with weak learners. According to Feng et al. (2015) with ntree = 200, accurate
results can be obtained in Random Forest. Duro et al. (2012)
B. K-Nearest Neighbors (kNN)
states default value of mtry = √p, here p is denoted as
According to Duda et al. (2015), the k-Nearest Neighbors number of predictors.
technique is a non-parametric technique which was used in
the before 1970’s in the applications of statistics (Franco- F. Stochastic Gradient Descent (SGD)
Loez et al., 2001). The basic theory behind kNN is that in According to L.Bottou (2010), in convex loss functions,
the dataset of calibration, kNN is capable to find a group of Stochastic Gradient Descent is very much efficient approach
k samples that are more nearest to unknown samples. The for learning of unequal linear classifiers for example in
labels of unknown samples out of these k samples can be SVM and logistic regression. SGD is being used for a long
determined by finding the average of the response variables time in the Machine Learning society. In the context of
(Akbulut et al. and Wei et al., 2017). The k plays an large-scale learning, SGD is being widely used recently.
important role in the performance of the kNN, because it is Except large scale learning, SGD is being applied in sparse
key tuning parameter of kNN (Qian et al., 2015). By using a machine learning obstacle mostly in text classification and
bootstrap procedure, the parameter k can be determined. natural language processing.
C. Linear Regression (LR) G. Support Vector Machine (SVM)
Linear Regression is known as the linear relationship The Support Vector Machine algorithm can work well
between the input data and the output data. The formula for especially on the basis of classification. SVM is basically
a linear regression is as shown following: used due to its high generalization performance (Liu et al.,
y=c0 + c1x1 +…..+ cnxn 2017; Mohri et al., 2018). However the learning time of this
In the equation shown above there are n input variables; algorithm is long, although more optimized results can be
these are known as predictors or regressors and one output obtained by reducing the number of processes in learning.
variable which is denoted by y (variable to be predicted). SVM is successful on large-volume data sets. SVM is used
The constants c0, c1 ……… cn, are called the regression in prediction, image classification, medical diagnosis, text
coefficients which is computed by principle of Least Square analytic, outlier detection etc.
method. This is known Multiple Linear Regression because
H. Decision Tree (DT)
there are more than one predictor (Margaret H. Dunham,
2005). Regression Analysis is a methodology of statistics According to J.-J.Sheu et al. (2016) Decision Tree is a
that is mostly used in numeric predictions (Jiawei Han and data mining approach and it is based on the design of tree
Micheline Kamber, 2010). like structure. DT is a predictive modelling approach which
is used in Statistics, Data Mining and Machine Learning. DT
D. Artificial Neural Network (ANN) is mostly used in classification and regression based
According to T. Subramaniam et al. (2010), first of all problems.
McCulloch and Pitts introduced Artificial Neural network
(ANN) in 1943. Earlier ANN was widely used in words V. EXPERIMENTS AND RESULTS
classification. ANN emulates the capabilities of human
In this section, we conduct experimentation to evaluate
mind. In human mind neurons which are also called nerves
the performance of Machine Learning models for Nifty 50
cells, contact to each other through sending messages. ANN
Index prediction on four different datasets.
is highly capable and efficient data-driven model which is
widely used in complex non linear behavior of data. ANN
can be used in Hand Writing Recognition, Pattern
Recognition, Face Identification, Text Translation, Medical
Diagnosis, Speech Recognition, Credit Card Fraud
Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 20 ©Copyright: All rights reserved.
 Machine Learning Models in Stock Market Prediction

The version 3.26.0 of Orange Data Mining Tool (Demsar,

J. et al., 2013) was used in the experimentation and
evaluation of the used Machine Learning Models in the
study which can be downloaded from [37]. The Fig.2 shows
the work flow of used Supervised Machine Learning Models
in the study.

Fig.4. AdaBoost Parameters

Fig.2. Work Flow of Supervised Machine Learning

Models
A. Features Extraction
Feature extraction is very important as wrong selection of
features can result into wrong prediction. The historical data
of Nifty 50 Index was downloaded from the website
“niftyindices.com” [38] which is openly accessible, in CSV
(Comma Separated Value) format. The available attributes
in the historical data were S.No, Open, High, Low, Close
and Date. S.No and Date attributes were skipped as these
attributes are not important for prediction. As shown in the
Fg.3, Open, High and Low attributes are selected as feature
and Close attribute is selected as target because it is to be
predicted. Fig.5. kNN Parameters

Fig.3. Features Extraction

B. Parameter Settings for used Models
The setting of parameters for each model is important as
wrong parameters setting can result into wrong output. The
best combinations of parameters were set to get more
accurate results. Figs. 4, 5, 6, 7, 8, 9, 10 and 11 are showing Fig.6. Linear Regression Parameters
parameters setting of AdaBoost, kNN, LR, ANN, RF, SGD,
SVM and DT respectively.

Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 21 ©Copyright: All rights reserved.
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-11 Issue-3, January 2022

Fig.7. Artificial Neural Network Parameters

Fig.10. Support Vector Machine Parameters

Fig.8. Random Forest Parameters

Fig.11. Decision Tree Parameters

C. Testing Results on Data Set-1, Data Set-2, Data
Set-3 and Data Set-4
In a dataset, a training dataset is implemented to build up
a model, while a test (or validation) data set is used to
validate the constructed model. Data points in the training
datasets are excluded from the test dataset (validation). So,
the training data is used to fit the model and testing data is
used to test it. The ratio in each data set was taken 80% and
20% for Training Data and Testing Data respectively.
Repeated k-fold cross-validation is used to improve the
estimated performance of machine learning models.

Fig.9. Stochastic Gradient Parameters

Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 22 ©Copyright: All rights reserved.
 Machine Learning Models in Stock Market Prediction

Table 2. Cross Validation Test of Data Set-1

Rank Model Train time [s] Test time [s] MSE RMSE MAE R2 CVRMSE
1 Linear Regression 0.464 0.065 53.158 7.291 4.577 0.995 0.708
2 ANN 2.515 0.084 53.158 7.291 4.577 0.995 0.708
3 SVM 2.939 0.098 53.991 7.348 4.688 0.995 0.713
4 SGD 1.229 0.083 54.477 7.381 4.773 0.995 0.716
5 kNN 0.486 0.103 86.858 9.320 6.195 0.992 0.904
6 AdaBoost 10.653 0.532 92.244 9.604 6.327 0.992 0.932
7 Random Forest 2.548 0.156 95.146 9.754 6.614 0.992 0.947
8 Decision Tree 9.088 0.007 131.607 11.472 7.881 0.989 1.113
Table 3. Test on Training Data of Data Set-1
Train Test time
Rank Model MSE RMSE MAE R2 CVRMSE
time [s] [s]
1 AdaBoost 1.296 0.082 0.517 0.719 0.235 1.000 0.070
2 Random Forest 0.294 0.019 20.645 4.544 3.079 0.998 0.441
3 Decision Tree 1.125 0.001 38.821 6.231 3.433 0.997 0.605
4 kNN 0.063 0.032 41.023 6.405 4.196 0.996 0.622
5 Linear Regression 0.053 0.016 51.801 7.197 4.536 0.995 0.698
6 ANN 0.395 0.020 51.801 7.197 4.536 0.995 0.698
7 SVM 0.382 0.042 52.367 7.236 4.633 0.995 0.702
8 SGD 0.161 0.009 52.759 7.264 4.706 0.995 0.705
Table 4. Test on Test Data of Data Set-1
Train Test time
Rank Model MSE RMSE MAE R2 CVRMSE
time [s] [s]
1 Linear Regression 0.040 0.006 49.051 7.004 5.482 0.995 0.722
2 ANN 0.298 0.008 49.051 7.004 5.482 0.995 0.722
3 SVM 0.307 0.014 51.519 7.178 5.688 0.995 0.740
4 SGD 0.116 0.008 55.283 7.435 5.900 0.995 0.767
5 kNN 0.046 0.012 84.938 9.216 6.496 0.992 0.950
6 Random Forest 0.229 0.015 87.267 9.342 7.086 0.992 0.963
7 AdaBoost 0.984 0.046 98.286 9.914 7.316 0.991 1.022
8 Decision Tree 0.932 0.001 134.667 11.605 9.042 0.987 1.197

Table 5. Cross Validation Test of Data Set-2

Rank Model Train time [s] Test time [s] MSE RMSE MAE R2 CVRMSE
1 Linear Regression 0.472 0.053 123.563 11.116 6.444 0.997 0.964
2 ANN 2.868 0.089 123.563 11.116 6.444 0.997 0.964
3 SVM 6.214 0.173 129.862 11.396 6.929 0.997 0.989
4 SGD 1.763 0.073 132.654 11.518 7.004 0.997 0.999
5 kNN 0.533 0.116 175.334 13.241 7.930 0.996 1.149
6 Random Forest 3.053 0.137 182.252 13.500 8.312 0.996 1.171
7 AdaBoost 14.946 0.520 192.136 13.861 8.350 0.996 1.203
8 Decision Tree 17.061 0.005 220.225 14.840 9.167 0.995 1.288
Table 6. Test on Training Data of Data Set-2
Rank Model Train time [s] Test time [s] MSE RMSE MAE R2 CVRMSE
1 AdaBoost 1.813 0.107 0.948 0.974 0.360 1.000 0.084
2 Random Forest 0.364 0.022 47.379 6.883 3.933 0.999 0.597
3 kNN 0.058 0.043 78.852 8.880 5.288 0.998 0.770
4 Decision Tree 1.973 0.002 81.221 9.012 4.421 0.998 0.782
5 Linear Regression 0.054 0.006 122.394 11.063 6.428 0.997 0.960
6 ANN 0.459 0.021 122.394 11.063 6.428 0.997 0.960
7 SVM 0.721 0.124 124.990 11.180 6.699 0.997 0.970
8 SGD 0.214 0.008 126.222 11.235 6.761 0.997 0.975
Table 7. Test on Testing Data of Data Set-2
Rank Model Train time [s] Test time [s] MSE RMSE MAE R2 CVRMSE
1 Linear Regression 0.046 0.007 30.471 5.520 4.134 0.993 0.512
2 ANN 0.341 0.008 30.471 5.520 4.134 0.993 0.512
3 SVM 0.685 0.034 31.231 5.588 4.274 0.993 0.518
4 SGD 0.179 0.008 32.043 5.661 4.387 0.993 0.525
5 AdaBoost 1.818 0.066 42.352 6.508 4.806 0.991 0.603
6 Random Forest 0.320 0.016 51.749 7.194 5.154 0.989 0.667
7 Decision Tree 1.824 0.001 62.872 7.929 5.889 0.986 0.735
8 kNN 0.052 0.015 69.306 8.325 5.031 0.985 0.772

Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 23 ©Copyright: All rights reserved.
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-11 Issue-3, January 2022

Table 8. Cross Validation Test of Data Set-3

Train time
Rank Model Test time [s] MSE RMSE MAE R2 CVRMSE
[s]
1 Linear Regression 0.721 0.057 147.960 12.164 6.101 0.999 0.893
2 ANN 6.591 0.101 147.960 12.164 6.101 0.999 0.893
3 SGD 4.054 0.081 161.002 12.689 6.933 0.999 0.932
4 kNN 0.783 0.153 231.892 15.228 8.010 0.999 1.118
5 AdaBoost 30.680 0.947 240.668 15.513 8.335 0.999 1.139
6 Random Forest 5.639 0.152 241.731 15.548 8.658 0.999 1.142
7 Decision Tree 38.699 0.011 296.749 17.226 9.890 0.999 1.265
8 SVM 14.939 0.451 307.789 17.544 12.945 0.999 1.288

Table 9. Test on Training Data of Data Set-3

Train time
Rank Model Test time [s] MSE RMSE MAE R2 CVRMSE
[s]
1 AdaBoost 3.039 0.302 0.887 0.942 0.358 1.000 0.069
2 Random Forest 0.515 0.026 56.254 7.500 3.994 1.000 0.551
3 kNN 0.087 0.078 107.462 10.366 5.355 1.000 0.761
4 Decision Tree 4.131 0.003 117.247 10.828 4.494 1.000 0.795
5 Linear Regression 0.051 0.007 148.125 12.171 6.089 0.999 0.894
6 ANN 0.811 0.017 148.125 12.171 6.089 0.999 0.894
7 SGD 0.430 0.009 157.518 12.551 6.772 0.999 0.922
8 SVM 1.615 0.367 361.298 19.008 14.896 0.999 1.396

Table 10. Test on Testing Data of Data Set-3

Rank Model Train time [s] Test time [s] MSE RMSE MAE R2 CVRMSE
1 Linear Regression 0.050 0.007 1329.200 36.458 23.923 0.998 0.850
2 ANN 0.501 0.008 1329.200 36.458 23.923 0.998 0.850
3 SGD 0.443 0.007 1396.104 37.364 26.271 0.998 0.871
4 SVM 1.558 0.089 1770.128 42.073 31.593 0.997 0.981
5 AdaBoost 3.135 0.170 1579476.492 1256.772 997.343 -1.418 29.304
6 Random Forest 0.514 0.016 1642014.285 1281.411 1024.897 -1.513 29.878
7 kNN 0.081 0.021 1653818.787 1286.009 1030.458 -1.531 29.986
8 Decision Tree 3.741 0.001 1659037.594 1288.036 1033.104 -1.539 30.033

Table 11. Cross Validation Test of Data Set-4

Rank Model Train time [s] Test time [s] MSE RMSE MAE R2 CVRMSE
1 Linear Regression 0.963 0.063 1767.328 42.040 13.638 1.000 1.223
2 ANN 9.136 0.084 1767.328 42.040 13.638 1.000 1.223
3 SGD 10.261 0.080 1879.098 43.349 16.298 1.000 1.261
4 Random Forest 9.196 0.186 2151.396 46.383 17.688 1.000 1.349
5 AdaBoost 64.507 1.637 2209.671 47.007 17.535 1.000 1.367
6 kNN 1.349 0.235 2292.070 47.876 16.926 1.000 1.392
7 Decision Tree 71.682 0.014 2328.751 48.257 20.030 1.000 1.404
8 SVM 28.303 0.710 5188.280 72.030 57.253 0.999 2.095

Table 12. Test on Training Data of Data Set-4

Rank Model Train time [s] Test time [s] MSE RMSE MAE R2 CVRMSE
1 AdaBoost 7.840 0.462 3.561 1.887 0.780 1.000 0.055
2 Random Forest 0.966 0.038 443.957 21.070 8.215 1.000 0.613
3 kNN 0.157 0.136 1078.444 32.840 11.402 1.000 0.955
4 Decision Tree 7.838 0.008 1495.199 38.668 9.181 1.000 1.125
5 Linear Regression 0.096 0.006 1754.389 41.885 13.613 1.000 1.218
6 ANN 1.189 0.033 1754.389 41.885 13.613 1.000 1.218
7 SGD 0.980 0.009 1857.122 43.094 16.203 1.000 1.253
8 SVM 3.507 0.810 5129.207 71.618 56.908 0.999 2.083

Table 13. Test on Testing Data of Data Set-4

Rank Model Train time [s] Test time [s] MSE RMSE MAE R2 CVRMSE
1 Linear Regression 0.105 0.006 1358.994 36.865 25.722 0.999 0.348
2 ANN 0.796 0.008 1358.995 36.865 25.722 0.999 0.348
3 SGD 1.189 0.008 1802.542 42.456 29.093 0.999 0.400
4 SVM 3.123 0.175 4668.566 68.327 49.110 0.998 0.644
5 AdaBoost 7.222 0.206 5187963.087 2277.710 1737.004 -0.930 21.474
6 Random Forest 1.000 0.017 5248175.973 2290.890 1750.752 -0.952 21.598
7 kNN 0.142 0.039 5357927.984 2314.720 1772.651 -0.993 21.823
8 Decision Tree 8.049 0.006 5379403.563 2319.354 1780.252 -1.001 21.867

Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 24 ©Copyright: All rights reserved.
 Machine Learning Models in Stock Market Prediction

D. Prediction Results for each used model with Actual

Close Price of Nifty 50 Index

Fig.16. Actual and Predicted Close Price of NIFTY Index

using Adaptive Boost (AdaBoost) in Testing Data of Data
Set-4
Fig.12.Actual and Predicted Close Price of NIFTY Index
using Linear Regression (LR) in Testing Data of DataSet-4

Fig.13.Actual and Predicted Close Price of NIFTY Index

using Artificial Neural Network (ANN) in Testing Data of Fig.17. Actual and Predicted Close Price of NIFTY Index using
Data Set-4 Random Forest (RF) in Testing Data of Data Set-4

Fig.18. Actual and Predicted Close Price of NIFTY Index

using k-Nearest Neighbors (kNN) in Testing Data of Data
Set-4
Fig.14. Actual and Predicted Close Price of NIFTY Index
using Stochastic Gradient Descent (SGD) in Testing Data
of Data Set-4

Fig.19. Actual and Predicted Close Price of NIFTY

Index using Decision Tree (DT) in Testing Data of
Data Set-4
Fig.15. Actual and Predicted Close Price of NIFTY Index
using Support Vector Machine (SVM) in Testing Data of
Data Set-4

Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 25 ©Copyright: All rights reserved.
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-11 Issue-3, January 2022

E. Discussion of Experiments and Results Table 6 shows that AdaBoost, RF, kNN and DT
Three tests were applied- Cross Validation Test, Test performed well receptively in training even better than
on Training Data and Test on Testing Data on each Data LR, ANN, SVM and SGD but Table 7 shows that these
Set. These results are shown in the Tables 2, 3, 4, 5, 6, models did not performed as well in Validation Test.
7, 8, 9, 10, 11, 12 and 13 in the section V(C) Tables 2, 3 In Data Set-3, from Tables 8 and 10, it is observed
and 4 are showing the testing results of Data Set-1. that this time also LR and ANN results are almost
Tables 5, 6 and 7 are showing the testing results of Data similar and ranked 1 st and 2 nd respectively. This time
Set-2. Tables 8, 9 and 10 are showing the testing results SGD performed better than SVM. SVM stood last in the
of Data Set-3. Tables 11, 12 and 13 are showing the Cross Validation Test. Table 9 shows that AdaBoost,
testing results of Data Set-4. RF, kNN and DT performed well receptively in training
In the testing results the time taken to train the model even better than LR, ANN, SGD and SVM but Table 10
and the time taken to test the model in seconds, also shows that these models did not performed as well in
included to compare the execution time of the models. Validation Test and obtained negative value of R 2 .
The next measurements in the testing results are Negative value of R 2 on Test Data as shown in the Table
MSE, RMSE, MAE, R 2 and CVRMSE which stands for 10 denotes the models AdaBoost, RF, kNN and DT are
Mean Square Error, Root Mean Square Error, Mean overfitted.
Absolute Error, R-Square, Coefficient of the Variation In Data Set-4, from Table 11 and 13, it is observed
of the Root Mean Square Error respectively. The lower that LR and ANN results are almost similar and
value of MSE, RMSE, MAE and CVRMSE implies obtained rank 1 and 2 respectively. Thereafter SGD
higher accuracy of a model. However, a higher value of performed well compared to SVM. Table 12 shows that
R square is considered desirable. AdaBoost, RF, kNN and DT performed well
Over fitting is identified by checking validation respectively in training even better than LR, ANN, SGD
metrics such as accuracy and loss. The validation and SVM but Table 13 shows that these models did not
metrics are generally increased up to a certain point and performed as well in Validation Test and obtained
after that start declining when the model is affected negative value of R 2 . Negative value of R 2 on Test Data
by over fitting. As a result, over fitting fails to fit as shown in the Table 13 denotes the models AdaBoost,
additional data and this effect the accuracy of a RF, kNN and DT are overfitted.
predicting model. If the errors on the testing or The above comparison made among all the models
validation dataset are higher than the errors on training can also be understood by seeing the Figs. 12, 13, 14,
dataset, it is said over fitting. Under fitting is identified 15, 16, 17, 18 and 19.
when the model neither performs well with training data Fig.12. shows the Actual and Predicted Close Price of
nor generalizes to test data. When a model performs NIFTY 50 Index using Linear Regression (LR) in
well on Training Data as well as on Test Data then the Testing Data of Data Set-4. Both the lines of actual and
model is a best fit or a good model. predicted close price of Nifty 50 Index are almost
Before comparing obtained testing results, sorting is overlapping to each other; it means the prediction model
done on RMSE variable, in ascending order for each is good fit model for LR.
table. It means the models having lower RMSE values Fig.13. shows the Actual and Predicted Close Price of
are up in the tables in comparison to models having NIFTY 50 Index using Artificial Neural Network
higher values of RMSE. After sorting the tables in (ANN) in Testing Data of Data Set-4. Both the lines of
ascending order on the basis of RMSE, rank is given to actual and predicted close price of Nifty 50 Index are
each model in increasing order as lower RMSE value almost overlapping to each other; it means the
will be having higher accuracy than the higher RMSE prediction model is a good fit model for ANN.
value. Fig.14. shows the Actual and Predicted Close Price of
In Data Set-1, from Tables 2 and 4, it is observed that NIFTY 50 Index using Stochastic Gradient Descent
LR and ANN results are almost similar but ANN takes (SGD) in Testing Data of Data Set-4. Both the lines of
more time in training and testing, so rank 1 is given to actual and predicted close price of Nifty 50 Index are
LR and rank 2 is given to ANN. Thereafter SVM and almost overlapping to each other; it means the
SGD performed well respectively among rest of the prediction model is a good fit model for SGD.
models, but SVM takes more time in training and testing Fig.15. shows the Actual and Predicted Close Price of
in comparison to SGD. DT performed worst among all NIFTY 50 Index using Support Vector Machine (SVM)
the models. Table 3 shows that AdaBoost, RF, DT and in Testing Data of Data Set-4. Both the lines of actual
kNN performed well respectively in training even better and predicted close price of Nifty 50 Index are almost
than LR, ANN, SVM and SGD but Table 4 shows that overlapping to each other; it means the prediction model
these models did not performed as well in Validation is a good fit model for SVM. But comparing the
Test. Validation Test results of LR, ANN, SGD and SVM it is
In Data Set-2, from Tables 5 and 7, it is observed that observed that LR and ANN performed better among all.
once again LR and ANN results are almost similar and
ranked 1 st and 2 nd respectively. SVM and SGD were
ranked 3 rd and 4 th respectively. Table 5 shows that DT
performed worst in Cross Validation Test and Table 7
shows that kNN performed worst in Validation Test.
Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 26 ©Copyright: All rights reserved.
 Machine Learning Models in Stock Market Prediction

Thereafter SVM performed better than SGD in Data networks can deal better with non linear dependencies.
Set-1 and Data Set-2 but SGD performed better than So if the data will be having some non linear
SVM in Data Set-3 and Data Set-4. It indicates that with dependencies, Neural Networks should perform better
increase in the size of data set, SVM performance than regression. After LR and ANN, Support Vector
decreases. Machine performed well but with increase in the size of
Fig.16. shows the Actual and Predicted Close Price of data Stochastic Gradient Descent performed better than
NIFTY 50 Index using Adaptive Boosting (AdaBoost) SVM. Thereafter ensemble learning methods of
in Testing Data of Data Set-4. The line of predicted Decision Tree- AdaBoost and Random Forest performed
close price of Nifty 50 Index overlaps the line of Actual better than kNN and Decision Tree with increase in the
Close price up to a certain point- 231 days out of 1244 size of data.
days thereafter it remains constant as predicted close In this empirical study eight Supervised Machine
price in the output was also constant. It means the Learning Models were used, in future empirical study, more
prediction model is an over fit model for AdaBoost. ensemble methods for Supervised Machine Learning
Fig.17. shows the Actual and Predicted Close Price of Models can be taken. This empirical study used around 25
NIFTY 50 Index using Random Forest (RF) in Testing years historical data, which is good for machine learning
Data of Data Set-4. The line of predicted close price of because in such a long period many bull and bear phases of
Nifty 50 Index overlaps the line of Actual Close price stock market were included.
up to a certain point- 231 days out of 1244 days
thereafter it remains constant as predicted close price in REFERENCES
the output was also constant. It means the prediction 1. Abu-Mostafa, Y. S., & Atiya, A. F. (1996). Introduction to financial
model is an over fit model for Random Forest. forecasting. Applied Intelligence, 6(3), 205–213.
Fig.18. shows the Actual and Predicted Close Price of 2. Akbulut Y., Sengur A., Guo Y., Smarandache F. NS-k-NN:
Neutrosophic Set-Based k-Nearest Neighbors classifier. Symmetry.
NIFTY 50 Index using k Nearest Neighbors (kNN) in 2017;9:179. doi: 10.3390/sym9090179.
Testing Data of Data Set-4. The line of predicted close 3. Breiman L. Random forests. Mach. Learn. 2001;45:5–32. doi:
price of Nifty 50 Index overlaps the line of Actual Close 10.1023/A:1010933404324.
4. Dash, R., & Dash, P. K. (2016). A hybrid stock trading framework
price up to a certain point- 231 days out of 1244 days integrating technical analysis with machine learning techniques.
thereafter it remains constant as predicted close price in Journal of Finance and Data Science, 2(1), 42–57.
the output was also constant. It means the prediction https://doi.org/10.1016/j.jfds.2016.03.002
model is an over fit model for kNN. 5. Demsar, J. et al. Orange: data mining toolbox in python. J. Mach.
Learn. Res.14, 2349–2353 (2013).
Fig.19. shows the Actual and Predicted Close Price of 6. Duda R., Hart P. Pattern Classification and Scene Analysis. John
NIFTY 50 Index using Decision Tree (DT) in Testing Wiley & Sons; New York, NY, USA: 1973.
Data of Data Set-4. The line of predicted close price of 7. Duro D.C., Franklin S.E., Dubé M.G. A comparison of pixel-based
and object-based image analysis with selected machine learning
Nifty 50 Index overlaps the line of Actual Close price algorithms for the classification of agricultural landscapes using
up to a certain point- 218 days out of 1244 days SPOT-5 HRG imagery. Remote Sens. Environ. 2012;118:259–272.
thereafter it remains constant as predicted close price in doi: 10.1016/j.rse.2011.11.020.
8. F., A., Elsir, S., & Faris, H. (2015). A Comparison between
the output was also constant. It means the prediction Regression, Artificial Neural Networks and Support Vector Machines
model is an over fit model for Decision Tree. for Predicting Stock Market Index. International Journal of Advanced
Research in Artificial Intelligence, 4(7), 55–63.
https://doi.org/10.14569/ijarai.2015.040710
VI. CONCLUSION 9. Feng Q., Liu J., Gong J. UAV remote sensing for urban vegetation
mapping using random forest and texture analysis. Remote Sens.
Prediction of the movements of the stock market
2015;7:1074–1094. doi: 10.3390/rs70101074.
index is very important for developing the effective 10. Franco-Lopez H., Ek A.R., Bauer M.E. Estimation and mapping of
market trading strategies. Financial decision to buy or forest stand density, volume and cover type using the k-Nearest
sell an instrument may be made by the traders by Neighbors method. Remote Sens. Environ. 2001;77:251–274. doi:
10.1016/S0034-4257(01)00209-7.
choosing the effective predictive model. Successful 11. Henrique, B. M., Sobreiro, V. A., & Kimura, H. (2018). Stock price
prediction of Stock Market Index movements may be prediction using support vector regression on daily and up to the
beneficial for investors. The tasks of predicting the minute prices. Journal of Finance and Data Science, 4(3), 183–201.
https://doi.org/10.1016/j.jfds.2018.04.003
movements of the Stock Market Index are highly 12. Idrees, S. M., Alam, M. A., & Agarwal, P. (2019). A Prediction
complicated and very difficult. This empirical study Approach for Stock Market Volatility Based on Time Series Data.
attempted to predict the direction of Nifty 50 Index IEEE Access, 7, 17287–17298.
https://doi.org/10.1109/ACCESS.2019.2895252
movement in the Indian Stock Market. Eight prediction 13. Immitzer M., Atzberger C., Koukal T. Tree species classification with
models were constructed and their performances were random forest using very high spatial resolution 8-Band WorldView-2
compared on historical data from April 22, 1996 to satellite data. Remote Sens. 2012;4:2661–2693. doi:
10.3390/rs4092661.
April 16, 2021. Based on the experimental results 14. J.-J.Sheu, Y.-K.Chen, K.-T.Chu, J.-H.Tang, andW.-P.Yang, “An
obtained, some important conclusions can be drawn. intelligent three phase spam filtering method based on decision tree
Linear Regression and Artificial Neural Network data mining,” Security and Communication Networks, vol. 9, no. 17,
pp. 4013-4026,2016.
performed almost equal performance in all the segments
of different size of data. The reason behind good
performance by Linear Regression competing to
Artificial Neural Network was that regression method
deals better with linear dependencies whereas neural

Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 27 ©Copyright: All rights reserved.
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-11 Issue-3, January 2022

15. Jiawei Han and Micheline Kamber, Second Edition (2010), “Data 39. https://www.nseindia.com/products-services/indices-nifty50-index
Mining, Concepts and Techniques”, Morgan Kaufmann, An imprint
of Elsevier (San Francisco).
16. Kara, Y., Acar Boyacioglu, M., & Baykan, Ö. K. (2011). Predicting
AUTHOR PROFILE
direction of stock price index movement using artificial neural Dr. Gurjeet Singh, is working as Associate Professor
networks and support vector machines: The sample of the Istanbul & Dean in the department of Lords School of
Stock Exchange. Expert Systems with Applications, 38(5), 5311– Computer Applications & IT, Lords University,
5319. https://doi.org/10.1016/j.eswa.2010.10.027 Alwar, Rajasthan. He has rich experience of 22 years
17. Kumar, I., Dogra, K., Utreja, C., & Yadav, P. (2018). A Comparative in teaching in the field of Computer Applications. He
Study of Supervised Machine Learning Algorithms for Stock Market is awarded a Ph.D degree in Computer Science. His
Trend Prediction. Proceedings of the International Conference on area of specialization is Data Mining. His interested
Inventive Communication and Computational Technologies, ICICCT areas of research are Outlier Detection, Data
2018, 1003–1007. https://doi.org/10.1109/ICICCT.2018.8473214 Management, Knowledge Discovery and Data Mining, Artificial
18. L.Bottou, “Large-scale machine learning with stochastic gradient Intelligence, Machine Learning and Deep Learning. He is also reviewer of
descent,” in Proceedings of COMPSTAT’2010. Springer, 2010, pp. reputed journals like IEEE Access. Email: research.gurjeet@gmail.com,
177-186. gurjeet.singh@lordsuni.edu.in
19. Liaw A., Wiener M. Classification and regression by randomForest. R
News. 2002;2:18–22.
20. Liu Y, Bi JW, Fan ZP. Multi-class sentiment classification: The
experimental comparisons of feature selection and machine learning
algorithms. Expert Syst Appl 2017.
https://doi.org/10.1016/j.eswa.2017.03.042.
21. Madge, S., & Bhatt, S. (2015). Predicting Stock Price Direction using
Support Vector Machines.
22. Margaret H. Dunham (2005), “Data Mining, Introductory and
Advanced Topics”, Pearson Education(Singapore).
23. Mohri M, Rostamizadeh A, Talwalkar A. Foundations of Machine
Learning. MIT Press; 2018.
24. Nagar, P., & Issar, G. S. (2013). DETECTION OF OUTLIERS IN
STOCK MARKET USING REGRESSION ANALYSIS,
International Journal of Emerging Technologies in Computational and
Applied Sciences, 3(2), 176–181.
25. Nagar, P., & Singh, G. (2012). An Analysis of Outliers for Fraud
Detection in Indian Stock Market, Researchers World : Journal of
Arts, Science and Commerce, Vol.-III, Issue 4(4), 10-15.
https://www.researchgate.net/publication/351415274_An_Analysis_o
f_Outliers_for_Fraud_Detection_in_Indian_Stock_Market
26. Nayak, A., Pai, M. M. M., & Pai, R. M. (2016). Prediction Models for
Indian Stock Market. Procedia Computer Science, 89, 441–449.
https://doi.org/10.1016/j.procs.2016.06.096
27. Parray, I. R., Khurana, S. S., Kumar, M., & Altalbe, A. A. (2020).
Time series data analysis of stock price movement using machine
learning techniques. Soft Computing, 24(21), 16509–16517.
https://doi.org/10.1007/s00500-020-04957-x
28. Picasso, A., Merello, S., Ma, Y., Oneto, L., & Cambria, E. (2019).
Technical analysis and sentiment embeddings for market trend
prediction. Expert Systems with Applications, 135, 60–70.
https://doi.org/10.1016/j.eswa.2019.06.014
29. Pyo S, Lee J, Cha M, Jang H (2017) Predictability of machine
learning techniques to forecast the trends of market index prices:
Hypothesis testing for the Korean stock markets. PLoS ONE 12(11):
e0188107. https://doi.org/ 10.1371/journal.pone.0188107
30. Qian Y., Zhou W., Yan J., Li W., Han L. Comparing machine
learning classifiers for object-based land cover classification using
very high resolution imagery. Remote Sens. 2015;7:153–168. doi:
10.3390/rs70100153.
31. R.E.Schapire, “The boosting approach to machine learning: An
overview,” in Nonlinear estimation and classification. Springer, 2003,
pp. 149-171.
32. Singh, G., & Nagar, P. (2012). A Case Study on Nutek India Limited,
regarding Deep Fall in Share Price, Researchers World - Journal of
Arts, Science and Commerce, Vol.– III, Issue 2(3), 64–68.
33. T.Subramaniam, H.A.Jalab, and A.Y.Taqa, “Overview of textual
antispam filtering techniques,” International Journal of Physical
Sciences, vol.5, no. 12, pp. 1869-1882, 2010.
34. Tan, T. Z., Quek, C., & See, Ng. G. (2007). Biological brain-inspired
genetic complementary learning for stock market and bank failure
prediction. Computational Intelligence, 23(2), 236–261.
35. Wei C., Huang J., Mansaray L.R., Li Z., Liu W., Han J. Estimation
and mapping of winter oilseed rape LAI from high spatial resolution
satellite data based on a hybrid method. Remote Sens. 2017;9:488.
doi: 10.3390/rs9050488.
36. Zhang H.K., Roy D.P. Using the 500 m MODIS land cover product to
derive a consistent continental scale 30 m Landsat land cover
classification. Remote Sens. Environ. 2017;197:15–34. doi:
10.1016/j.rse.2017.05.024.
Web References
37. https://download.biolab.si/download/files/
38. https://www.niftyindices.com/reports/historical-data

Published By:
Retrieval Number: 100.1/ijitee.C97330111322 Blue Eyes Intelligence Engineering
DOI: 10.35940/ijitee.C9733.0111322 and Sciences Publication (BEIESP)
Journal Website: www.ijitee.org 28 ©Copyright: All rights reserved.
View publication stats