ArticlePDF Available

Efficient Machine Learning Techniques for Stock Market Prediction

Authors:
  • National University of Computer and Emerging Sciences, Islamabad, Pakistan

Abstract and Figures

Stock market prediction is forever important issue for investor. Computer science plays vital role to solve this problem. From the evolution of machine learning, people from this area are busy to solve this problem effectively. Many different techniques are used to build predicting system. This research describes different state of the art techniques used for stock forecasting and compare them w.r.t. their pros and cons. We have classified different techniques categorically; Time Series, Neural Network and its different variation (RNN, ESN, MLP, LRNN etc.) and different hybrid techniques (combination of neural network with different machine learning techniques) (ANFIS, GA/ATNN, GA/TDNN, ICA-BPN). By extensive study of different techniques, it was analyzed that Neural Network is the best technique till time to predict stock prices especially when some de- noising schemes are applied with neural network. We, also, have implemented and compared different neural network techniques like Layered Recurrent Neural Network (LRNN), Wsmpca-NN and Feed forward Neural Network (NN). By comparing said techniques, it was observed that LRNN performs better than feed forward NN and Wsmpca-NN performs better than LRNN and NN. We have applied said techniques on PSO (Pakistan State Oil), S&P500 data sets.
Content may be subject to copyright.
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 855 | P a g e
Efficient Machine Learning Techniques for Stock Market
Prediction
Zahid Iqbal, R. Ilyas, W. Shahzad, Z. Mahmood and J. Anjum, Member, IEEE
Abstract
Stock market prediction is forever important issue for investor. Computer science plays vital role to solve this
problem. From the evolution of machine learning, people from this area are busy to solve this problem
effectively. Many different techniques are used to build predicting system. This research describes different state
of the art techniques used for stock forecasting and compare them w.r.t. their pros and cons. We have classified
different techniques categorically; Time Series, Neural Network and its different variation (RNN, ESN, MLP,
LRNN etc.) and different hybrid techniques (combination of neural network with different machine learning
techniques) (ANFIS, GA/ATNN, GA/TDNN, ICA-BPN). By extensive study of different techniques, it was
analyzed that Neural Network is the best technique till time to predict stock prices especially when some de-
noising schemes are applied with neural network. We, also, have implemented and compared different neural
network techniques like Layered Recurrent Neural Network (LRNN), Wsmpca-NN and Feed forward Neural
Network (NN). By comparing said techniques, it was observed that LRNN performs better than feed forward NN
and Wsmpca-NN performs better than LRNN and NN. We have applied said techniques on PSO (Pakistan State
Oil), S&P500 data sets.
Index Terms Feed Forward Neural Network, Recurrent Neural Network, Stock Market, Machine Learning
I. INTRODUCTION
Stock market prediction is one of challenging
issue, catches attention of many researchers and
investors. Investors all across the globe showed their
great interest in stock predicting systems. Investors
start relying on predictions systems to make their
important business decisions. A lot of research is made
in this domain but still no complete solution is found.
Is stock market prediction fully possible is still debate
of the hour. This is due to difficulty in predicting stock
market with full accuracy because there is a great
influence of external entities (social, political,
psychological, and economic) [1] [20]. Secondly, it
requires huge amount of historical data. Various
machine learning techniques are being used to predict
market states. This research covers many state of the
art techniques to find out some optimal solution to the
problem.
Several machine learning techniques are
being in use for stock market prediction. There is no
specification made by which we can choose the
optimal solution for stock market prediction. This
research will study different machine learning
techniques being in use for stock market prediction.
This research will follow a comparative approach to
find out optimal technique for stock market prediction.
Comparison will be made on the basis of their
performance. Every technique has some advantages
and disadvantages. This research will analyze
advantages and limitations of these techniques. And
find that which technique is comparatively better for
stock market prediction.
Fundamental and Technical analysis are two famous
methods being in use for risk analysis from very old
age. When there was no computational method to
analyze risk these methods were used.
To predict stock prices (by analyzing past data) there
are many conventional methods. Generally used two
methods are [1][20][10][18].
i. Fundamental and
ii. Technical
In Fundamental analysis, accurate and
reliable information of company’s financial report,
competitive strength and economic conditions are
required to find out the accurate value of product in
which they have interest. This value is used for making
decision for investment. It is based on this idea If the
intrinsic is higher than the value it holds in the
market, invest, else it will be considered a bad
investment and avoid it”.
Fundamental analysts believe that 90 percent
logical factors and 10 percent physiological factors
define a market. Fundamental analysis is useful in
long-term predictions. The advantages of fundamental
analysis are its systematic approach and its ability to
predict changes before they show up on the charts. [1]
Technical analysis requires history of market. The
idea behind technical analysis is that constantly
changing attributes of investors in response to
different forces/factors make trends/move of stock
prices. Different technical factors like volume and
maximum and minimum prices per trading period are
used for analysis. Rules are extracted from data. On
the basis of these rules investor take decisions for
RESEARCH ARTICLE OPEN ACCESS
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 856 | P a g e
future. According to some chartists, market trend is 90
percent psychological and only 10 percent logical. The
major disadvantage of technical analysis is that trading
rule extraction from charts study is very subjective so
different analysts might derive different rules from
same charts [1]. Technical analysis is used for both
long and short term analysis.
The basic purpose of this research is to find
the optimal among many state of the art techniques.
What is the optimal technique for stock prediction? By
combining multiple techniques can we get a better
system or we have already optimal solution? On which
basis one prediction system is preferred on other? For
this purpose we compare and analyze some current an
efficient techniques on the basis of their produced
results.
II. TECHNIQUES USED FOR STOCK MARKET
PREDICTION AND COMPARISON
Prediction is making decision on the basis of
past known data (knowledge). Predicting stock market
trends is challenging task. It is because of non-linear
behavior of stock data. Many other factors are also
present here, which create hurdles in the way to
predict stock (economic conditions, trader’s behavior,
investor psychology, politics etc.). To some extent,
artificial intelligence makes it possible to predict stock
trends. The main purpose of any model built for
prediction of stock is to reduce risk (minimize error).
Models built for prediction by using these (Neural
Networks, Genetic algorithms etc.) techniques showed
successful For many past years different techniques
are being used for predicting systems. Few of
techniques will be defined here are Time series,
Neural Networks, Hybrid techniques.
2.1 Time Series
Ordered lists of values of one parameter or
variable, which are provided in equal time intervals
called time series. The prediction is the continuous of
pattern over time like growth in sale, stock market
analysis or gross national product (Gosasang et al.
2010). The common time series methods are
i. ARMA (Auto Regression Moving Average)
ii. ARIMA.(Auto Regression Integrated Moving
Average)
ARMA is combination of AR (auto
regression) and MA (moving average) models. Used to
predict future values. ARIMA is variation of ARMA.
Any Time series is stationary if mean and variance is
constant. Otherwise, it’s non-stationary. Non
stationary time series are difficult to predict these are
full of noise. Stock prices are characterized as non-
stationary and for that purpose de-noising techniques
are used for removing noise from data. [4]
2.2 Artificial Neural Networks (ANN)
In 1943, W.S.Mcculoch and W. Pitts
established Neural Network (NN) and its mathematical
model. The established model was named MP model.
Than MP model was used to put forward the network
construction method and neuron’s formalization
mathematical description and proved that each single
neuron performs logic function, so a new time of
Neural Network research began. Neural network are
used in pattern recognition, prediction etc.
Network is a set of interconnected nodes. A node is a
computational unit which receives inputs and after
processing produces output. The flow of information
between nodes can be determined by the connections
between the nodes. [2] The detailed introduction to
ANNs is given by [12]. The basic structure of ANN
with one hidden layer is given in fig . 1.
Input layer
Hidden layer
Output layer
Fig . 1 Artificial Neural Network
The basic architecture of ANN is Multi- layer
feed forward used by [2]. In this architecture,
information flows in one direction only (from input to
output). It consists of one input , one or more hidden
and one output layer. Inputs are sent into units in input
layer than weighted output from these units is taken as
input in next hidden layer, weighted output of this
layer is sent as input in next hidden layer and so on.
Until output of last hidden layer is send to output
layer. Output layer gives the predicted output. Back
propagation algorithm is used for learning process in
NN. In this algorithm, network is trained by repeatedly
comparing the output and target output. And
minimizing error. log-sigmoid is used as an activation
function. To check the error, mean square error is
used. A very simple approach is given in [2]. The
author used very general and simple architecture of
ANN. Author performed pre-processing on data.
Author used Relevance Attribute Analysis method to
remove unwanted attributes and then applied min-max
normalization for normalizing data. That decrease risk
of error or producing in sufficient answer.
2.2.1 Multi-Layer perceptron with back
propagation learning algorithm
Artificial Neural networks are applied in
many different ways for stock prediction. Multi-layer
perceptron (MLP) is used with supervised learning
algorithm (back-propagation). It has ability to solve
non-linear problems; stock prices usually used are
non-linear. MLP works as it first initialize weights of
all network and train the training pattern then get
output then error is propagated by using back-
propagation. Basic model of MLP with back
propagation is used in [10] to analyze the importance
of Neural Networks for stock prediction. They used
five attributes as input previous close price, close
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 857 | P a g e
price, open price, low price, and high price.
Correlation (R), Average Absolute Error (avg abs),
Max Absolute Error (max Abs), Root Mean Square
Error (RMS), Accuracy, Confidence internal are the
performance measures used to measure error. The
model was applied on Indian stock Market. TCS
company data was taken to test this model. This Model
gives 20% of Accuracy between target output and
calculated output. This model is made to predict in
short-term. This model can be used further for error
percentage reduction. [10].
In ANN, a large number of historical data
required. Accuracy and result network becomes
complex. Results statical relevance is required. Best
topology of network is unknown. NN requires a
careful data design and systematically analyzed. There
are many different models of Neural Networks made
to remove/cover these flaws. To overcome these flaws
advance/new approaches of NN are become in use.
2.2.2 Recurrent Neural Network with wavelet
transforms
Recurrent neural network (RNN) is type of
Neural Networks with back-loop in it. Recurrent
neural network has advantage over ANN that it saves
patterns with time.
RNN are applied for stock market prediction
by [4] with Wavelet transform as pre-processing
technique to remove noise from data. Recurrent neural
networks are available in two categories, Elman and
Hopfield [4].Hopfield networks are less used in
practice. There can be spurious stable points in
Hopfield network that may lead to incorrect results.
The Elman network commonly is a two-layer
network with feedback from the first-layer output to
the first layer input. This recurrent connection allows
the Elman network to both detect and generate time-
varying patterns. The Elman network has ‘tansig’
neurons in its hidden (recurrent) layer, and ‘purelin’
neurons in its output layer. Algorithms used for
training data .Least Mean Squared (LMS) algorithm
which involves Gradient descent with momentum and
batch learning rate back-propagation. Least Mean
Squared (LMS) algorithm which involves Gradient
descent without momentum and batch learning rate
back-propagation. It is also observed that Least Mean
Squared (LMS) algorithm which involves Gradient
descent with momentum and batch learning rate back-
propagation performs better. It is observed that RNN
gave different results when applied on different data.
From this one can say that performance of network
relay on data being used for it.
2.2.3 Layered Recurrent Neural Network
LRNN consist of two parts
i. Recurrent layers (for temporal patterns)
ii. Feed-forward part
In LRNN, input passes through the net more
than once so it helps to learn temporal patterns. It is
due to recurrent layers and new input and previous
temporal patterns are given to feed-forward part as
input. Basic structure of LRNN is given in fig. 2
Different variations of ANN are tested by
[16] for stock price prediction. Models compared by
author are Back-Propagation Neural Network (BPN),
Radial Basis Function Neural Network (RBFNN),
Generalized Regression Neural Networks (GRNN),
and Layered Recurrent Neural Networks (LRNN). It is
observed from results that LRNN outer-perform than
other three. And BPN performs better than other two.
LRNN is type of RNN. They have dynamic behavior.
Due to the dynamic behavior of stock market, RNN
has been found an attractive tool for performing
nonlinear time series prediction
Input Input
Hidden Layer
Output layer
Fig. 2 Layered Recurrent Neural Network
2.2.3 Echo State Network
Echo state networks, Subset of RNN. It can
be trained comfortably. This network is developed in
the emerging field known as reservoir computing. In
reservoir computing, the recurrent connections of the
network are viewed as a fixed reservoir used to map
inputs into a high dimensional, dynamical spacea
similar idea to the support vector machine. For RNN,
architecture and supervised learning is provided by
ESN. Finding random large, fixed RNN is the main
task of ESN.in ESN, input travels through the network
more than once. This improves the output results. Only
output weights are trained in echo state network. With
a sufficiently high dimensional space, a simple linear
decode can be used to approximate any function
varying with time. ESN are used for financial time
series forecasting by [13]. Author applied ESN for
forecasting and compared its performance with
Kalman filter (also used in forecasting process).
It is observed from [13] that ESN outer
perform than Kalman filter. The Kalman filter uses
linear dynamical model. That model estimates the
process state recursively by reducing error (Mean
square error). Quick changes in the stock price can be
captured by ESN. Whereas, Kalman filter is unable to
capture. Powerful black box method is used by ESN
for modeling time dependent phenomena. Black box
methods are appealing because they require no prior
assumptions or knowledge of the underlying dynamics
of the time series data. The Kalman filter does not
have enough features to predict prices and capture
rapid movement in the stock price.
This model is checked on S&P500 indexes. Input
features used in this model are current price, trading
volume, and the S&P500 price.
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 858 | P a g e
Different models of ANNs are evaluated to
test the effectiveness of neural networks for stock price
prediction. MLP, Dynamic Artificial Neural Networks
(DAN2), and hybrid neural networks using
generalized auto regression conditional
heteroscedastticity(GARCH) were analyzed by[15]. It
is observed that MLP performs better than other two
techniques.
Gradient descent for stock prediction is
applied by [7] and compared with Recurrent Neural
Network for stock market prediction. Gradient descent
back-propagation learning algorithm is used for
supervised training of Neural Network. Learning rate
is required for gradient descent algorithm. To check
whether algorithm is working properly convergence
test is taken. If learning rate is too small then
convergence is slow, if learning rate is too large than
may not convergence. To make his model work
appropriately, [7] performed pre-processing using R/S
analysis method. As quality of input data effects
output. In output, model gives newer open price.
Sigmoid transfer function is used in it. Mean square
error is calculated to check the model. This Model is
applied on NASQAD stock. Two models (Multi-Layer
feed forward and Recurrent Neural Network (RNN))
are built by using Gradient descent back-propagation
Algorithm [7]. It is observed that RNN outer performs
than ANN.
2.3 Hybrid Techniques
From last few year researchers move their
attention to merge different techniques to get more
optimal results. Our next section covers some of
hybrid techniques used for stock prediction. Here we
analyze few hybrid techniques used for stock
prediction. Fuzzy logic and genetic programming are
widely used with integration of other techniques for
stock price prediction. Fuzzy logics are used to build
rules for prediction on the basis of past knowledge
available.
2.3.1 Using GA with different ANN techniques
Genetic Programming (GP) first introduced
by Koza (Koza 1992) by developing symbolic
regression. It is a computational optimization tool.
This tool is used to derive optimal model from time
series data. Reproduction, crossover, and mutation are
main operation of GP. Fitness Function is the main
factor on which final population based. The process of
switching nodes in a population is known as cross
over. For making new generation based on fitness
function GP reproduces. The task of mutation is to get
and substitute the information of one node with those
individuals. To evaluate new generation fitness
function used (Langdon & Poli 2002).
Genetic Algorithms (GA) are used with two
different types of NN to improve performance of
network. GA’s are used here to identify input variables
and weights for these variables [28][31]. Time delay
Neural Networks (TDNN) and Adaptive time delay
neural network (ATNN) are used here for their ability
for saving temporal patterns. GA-TDNN and GA-
ATNN are proposed by [28] for forecasting stock. It is
observed from result given by author that GA-TDNN
and GA-ATNN outer performs than individual TDNN,
ATNN and RNN.
2.3.2 Integration of genetic fuzzy systems and
artificial neural networks
Integration of genetic fuzzy systems and
artificial neural networks for stock price forecasting is
an example of such hybrid technique developed by
combining Neural Network, fuzzy logic and genetic
programing. In this model three techniques are used to
build better prediction system. There are three main
stages of this model. Variable selection is first phase
of this model. Stepwise regression analysis (SRA) is
applied for key variable selection. Second phase is to
divide data, self-organization map (SOM) neural
network is used for this purpose. SOM reduce
complexity of data by dividing it into useful sub part.
Last phase is to build GFS for stock price prediction,
data clusters are sent to GFS for forecasting purpose
[24]. Stepwise regression analysis (SRA) finds out
the set of independent factors. SRA is recursive
function. On every step, a variable enters or remove
from model. SOM clustering is being used to combine
related data. From previous literature made on
clustering methods it is observed that SOM clustering
performs better than other techniques of clustering
(hierarchical). Next phase of this model is to build, a
genetic-fuzzy system. Mamdani-type fuzzy rule based
system to deal with stock price forecasting problems.
There are two general steps of this evolutionary
process used to make knowledge based (KB) of fuzzy
rule based system. First step is to evolve rules (through
genetic algorithms). And second step is to tune
database of fuzzy system.
Fig. 3 illustrates main steps of this model.
Data
collecti
on
Stepwise regression
Data Clustering
Construct GFS
Generate Forecast
Fig. 3. Integration of GF systems and ANN
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 859 | P a g e
2.3.3 Rough Set-Neural Network
Another technique used for making stock
market decisions is a hybrid Artificial Neural Network
and Rough Set model. Used to find best sell and buy in
Dhaka Exchange. In this model, Neural Networks are
used for its ability to predict (learn from knowledge)
and Rough set is used for its ability of Powerful Rule
Extraction.
In Rough Set (RS), Decision Table is used to
organize data. This table contains attributes and data
elements. Where attributes are placed in columns and
data elements are placed in rows. Analysis of limit
Discernibility is Main theme of Rough set Model.
Three regions are defined by RST. These regions base
on equalent classes induces by attribute values. These
regions are lower approximation, upper
approximation, and boundary. Rough set is here to
reduct data and extract rules from data for prediction.
Johnson’s reducer algorithm is used by [20] to find all
reducts and Levenberg-Marquardt Back-Propagation
algorithm is selected to train ANN. Confusion matrix
is used to analyze performance of this model.
This technique is applied on Dhaka Stock Exchange.
And its Performance is compared with individual RS
and ANN. It is observed that Hybrid RS-ANN is better
than alone ANN and RS.
2.3.4 Independent component analysis (ICA) with
Back Propagation Neural Network (BPN)
As noticed in above mentioned techniques
that data is full of noise. Noise in data could lead over-
fitting and under-fitting problems. Another method to
remove noise from data with combination of Neural
Networks is Independent component analysis with
Neural network (ICA-BPN). There are two basic steps
of this technique. One is to get independent
components from data and second is to input these
data elements to neural network architecture as input.
Author compared this technique with wavelet-BPN
and single BPN and observed that ICA-BPN outer
performs than other two. From all above mentioned
methods it is concluded that using ANNs alone have
flaws in it.
2.3.5 Adaptive Neuro-Fuzzy Inference system
(ANFIS) for Stock Prediction
Another technique being in use is ANFIS. A
neuro-fuzzy system is created by using this method.
These systems are built by using Adaptive neuro-fuzzy
based system controller. This controller is further used
to control the stock market process. In Basic
architecture of the model, Input is mapped to input
membership function. Then that Input membership
functions are mapped to rules. These rules are then
further used to set output. And then like input, this
output is mapped to output membership functions.
ANFIS are applied for stock market short
trends by [24]. Gaussian-2 shaped membership
functions are chosen over bell shaped Gaussian and
triangular ones to fuzzily the system inputs due to the
lowest RMSE [24]
The ANFIS model is a Sugeno first order
model with two inputs and one output. Five Gaussian-
2 membership functions correspond to each input, very
small, small, medium, big and very big. The
combination of two inputs and five membership
functions creates twenty five rules (5)2
The PR-ANFIS model is a Sugeno first order
model with three inputs and one output. Three
Gaussian-2 membership functions small, medium, and
big correspond to each input, for a total of twenty
seven (3)3 rules [24].
2.3.6 ANFIS with indirect Approach TKS-fuzzy
based
ANFIS is used for stock prediction, by
applying all this method on a rule based. To analyze
stock market a complex system is built. Neuro-Fuzzy
Inference System followed by a TakagiSugenoKang
(TSK) type Fuzzy Rule Based System is developed for
stock price prediction. Technical indexes are used as
input and output of this model is linear combination of
these indexes. To identify number of rules Fuzzy C-
Mean (FCM) Clustering is used. Gaussian function is
used in premise section. FCM algorithm is used to
define number of rules. Membership degree is
assigned to the output of last step. At last step, neuro-
fuzzy inference system is used to tune parameters of
system. The forecasting capability of the system is
greatly improved by applying this technique [25]. This
model is applied on Tehran stock exchange indexes.
After all above literature it is observed that
(Feed-Forward Neural Network) NN are used mostly
for prediction. Different models are built using Neural
Networks. NN are preferred for stock prediction
because they can learn from past data. And they map
the input to output. From a survey in 2012 it is
concluded that Neural Networks are used 80% for
prediction system. It is analyzed from above literature
that every hybrid technique used Neural Network. It is
concluded from previous study that for prediction
recurrent type of Neural Networks are preferred.
Comparisons and results of different models shows
that Recurrent Neural Network (RNN) is
comparatively better then (Feed-Forward Neural
Network) NN. But RNN are less used because it has
complex structure and it takes more memory. Major
point analyzed from the study is importance of data
pre-processing. Different models used different data
preprocessing methods to normalize their data. After
above literature, we took three models (Feed-Forward
Neural Network (NN), Layered Recurrent Neural
Network (LRNN), wmspca-feed forward neural
networks and apply them. Firstly, we apply simple
feed-forward neural network and then we apply a de-
noising method (wmspca) on simple feed-forward
neural network and compare their performance.
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 860 | P a g e
III. METHODOLOGY
This research compares different models of
Neural Networks being in use for stock price
prediction. Here we will apply few methods and
compare their results. Neural Networks are chosen for
this purpose. Neural Networks are selected to
implement on the basis of previous literature. From
year 2000-2012 Neural Networks are used 80% for
stock prediction In above literature it is also observed
that performance of Recurrent Neural Network (RNN)
in better for prediction when compared with Feed-
Forward Neural Network. From above literature, it is
observed that data pre-processing (de-noising) has
better impact on output. So, we will use de-noising
technique with feed-forward neural network. All this
working is performed by using tool MATLAB, Neural
Network toolbox is used for simulation. Predefined
methods/functions are used for this purpose.
1. Data collection
2. Built different models (Feed-Forward Neural
Network (NN) , wmspca- feed-forward neural
network , Layered Recurrent Neural Network
(LRNN))
3. Check on different data sets
4. Compare their output performance on the basis
of their performance measure and time
consumed.
Fig. 4 shows flow chart of main steps of this
methodology.
At first, data sets are collected for
experiment. Data is collected from yahoo.finance.com
and google.finance.com, PSO (Pakistan State Oil)
and S&P500 data is collected for these experiments.
After collecting data, we applied three models using
MATLAB. For checking performance of these
techniques, we took one month data of PSO as input to
check models. After that we increase the amount of
data with 6 month data and again test models.
After getting results, techniques will be
compared on the basis of performance measure used
and time taken by a technique to give output.
3.1 Implementation
As, basic aim of applying these models is to
compare their performance. So some features will
remain same for all models. These elements and their
values are defined in tab.1
From research, it is observed that levenberg-
Marquad. It is better training function when compared
with other training function (gradient, resilient back-
propagation). Whenever we use any model for
prediction, our aim is to minimize error.
Performance measure used for checking
output is normalized mean squared error with
regularization performance function.
Applied models are feed-forward neural
network, layered recurrent neural network and
wmspca- feed-forward neural network.
3.1.1 Feed-Forward Neural Network
Firstly, feed-forward net is created. This
function built a feed forward net. Then we set input
and target data. and train network. Training method
takes input data and network used. Here data is
divided into three parts (train, test, validate). We use
60% data to train the network, 20 % for testing and
20% for validation. Fig. 5 shows flow chart of feed-
forward neural network.
Fig. 4 Methodology
3.1.2 Layered Recurrent Neural Network
After applying feed-forward Neural Network,
Layered Recurrent Neural Network (LRNN) is
implemented. LRNN function creates a layered
recurrent net by taking input data, target data, no. of
neurons to create the net. (NOTE: we can also set
Factors
Values
Input data
PSO/S&P500
Training Function
Levenberg-Marquardlt
Performance measure
Training data
Validating data
Testing data
Msereg
60%
20%
20%
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 861 | P a g e
other values of the network at the time of creating this
net). After creating network we use same methodology
used in feed-forward neural network to train network.
Fig. 6 shows flow chart of layered recurrent neural
network.
3.1.3 Wmspca-Feed-Forward Neural Network
We used a wsmpca (multiscale principal
component analysis) to pre -process data. This method
is use for pre-processing data. Now input this pre-
processed data into Feed-Forward Neural Network.
Fig7 shows flow chart of feed-forward neural network.
This method takes input matrix, levels of
wavelet decomposition name of wavelet and principle
components. And returns a simplified version of input
matrix X obtained from the wavelet-based multi-scale
Principal Component Analysis (PCA). This wavelet
decomposition is performed using the decomposition
level ‘LEVEL’ and the wavelet WNAME’. Now input
this simplified data into Feed-Forward Neural
Networks.
Start
Input financial
data
Create Network Set parameters of
Network
Train Network
Output
End
Fig. 5 Feed-Forward Neural Network
Start
Input financial
data
Create Network Set parameters of
Network
Train Network
Output
End
Fig. 6 Layered Recurrent Neural Network
Start
Input financial
data
Create Network Set parameters of Network
Train Network
Output
End
Wmspca De-noising
Fig. 7 wmspca-Feed-Forward Neural Network
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 862 | P a g e
IV. EXPERIMENTS AND RESULTS:
After creating models, we get their results in
form of graphs and values. Results are sub-divided
according to the data sets used to check above
mentioned models.
Our result holds different parameters while
training. Here is short introduction of these
parameters. Parameters used are epoch, time, gradient,
performance.
Epoch: is number of iterations taken by a network to
train data. Training method perform training, testing
and validation of data in single method
Time: shows time interval taken by a network while
training data.
Performance: shows performance measure.
Performance measure shows the error value given by
any network. If the value of this performance is small
it shows that network works properly. If value of this
performance measure is large then more training is
required.
Gradient: is learning rate. It is speed of learning. If
learning rate is small than training will be slow. If
learning rate is large than learning will be fast but may
not give better results.
4.1 Results on PSO (PAKISTAN STATE OIL)
dataset:
Fig. 8 and 10 holds results of Feed-Forward Neural
Network,
Fig. 11 holds results of Layered Recurrent Neural
Network,
Fig. 14 and 15 holds results of wmspca-Feed-Forward
Neural Network when applied on PSO data.
Fig. 8. Show performance measure at y- axis and
number of epochs on x-axis. This figure shows that 49
epochs are required to train network without over
training.
Fig.9 plots the data fitting, on 4 different stages of
data,
i. When data is training
ii. When testing data
iii. When validating data
iv. When catching target
v.
Fig. 9 This plot show values of ‘R’ (regression). Basically,
regression shows relation between target and selected input
data. From provides most probable target value. This shows
the state of data fitting while training, testing and validating.
Value of ‘R’ near to ‘1’ shows that data fitting is better
performed
Fig. 10. Performance measure at y- axis and number
of epochs on x-axis. This figure shows that how many
epochs are required to train network without over
training
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 863 | P a g e
Fig. 11. Show performance measure at y- axis and
number of epochs on x-axis. This figure shows that
how many epochs are required to train network
without over training
Fig. 12. This plot show values of ‘R’ (regression).
Basically, regression shows relation between target and
selected input data. From provides most probable target
value. This shows the state of data fitting while training,
testing and validating. Value of ‘R’ near to ‘1’ shows that
data fitting is better performed
Fig. 13 Show performance measure at y- axis and
number of epochs on x-axis. This figure shows that
how many epochs are required to train network
without over training
Fig. 14. Show performance measure at y- axis and
number of epochs on x-axis. This figure shows that
how many epochs are required to train network
without over training
Fig. 15 Show performance measure at y- axis and
number of epochs on x-axis. This figure shows that
how many epochs are required to train network
without over training
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 864 | P a g e
Fig. 16 This plot show values of ‘R’ (regression).
Basically, regression shows relation between target and
selected input data. From provides most probable target
value. This shows the state of data fitting while training,
testing and validating. Value of ‘R’ near to ‘1’ shows
that data fitting is better performed.
4.2 Comparison
TABLE2 contain results of Feed-Forward Neural
Network. In this experiment we took one month data
of PSO and 6 month data of S&P500, we observed
that when amount of data increased , number of
iteration and time of execution also increases.
TABLE 3
LAYERED RECURRENT NEURAL NETWORK (LRNN)
Dataset
Epochs
Time
MSEREG
(performance
Measure)
ggg
PSO
36
00:06:22
0.05
S&P500
263
01:09:57
151
TABLE 3 holds results of LRNN own both data
sets. It is observed that LRNN takes more time but
less number of iterations.
TABLE 4
WMSPCA-FEED-FORWARD NEURAL NETWORK
Dataset
Epochs
Time
MSEREG
(performance
Measure)
ggg
PSO
11
00:00:07
0.02
S&P500
19
00:00:35
9.64
TABLE 4 holds results of wmspca-NN results. Visible
changes can be observed in time, epochs and error in
comparison with TABLE 2 and 3.
TABLE 5
RESULTS OF DIFFERENT TECHNIQUES ON S&P500 DATASET
Techniques
Epochs
Time
MSEREG
(performance
Measure)
ggg
Feed-Forward
Neural Network
1000
00:31:53
21.9
Layered
Recurrent
Neural Network
263
01:09:57
151
Wmspca- feed
forward Neural
Network
19
00:00:35
19
TABLE5 holds different techniques results on S&P500 data.
It shows time, performance and epochs used by different
techniques on same data set. It is observed from results that
data pre-processing methods enhance performance of NN
and give better results.
TABLE 6
RESULTS OF DIFFERENT TECHNIQUES ON PSO DATASET
Techniques
Epochs
Time
MSEREG
(performance
Measure)
ggg
Feed-Forward
Neural Network
48
00:00:19
0.15
Layered
Recurrent
Neural Network
36
00:06:22
0.05
Wmspca- feed
forward Meural
Network
11
00:00:07
0.02
TABLE6 holds results on PSO data set. This table holds
values of time , msereg and epochs to make the comparison
easy. From this table one can easly conclude that wmspca
outer performs than other two methods.
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 865 | P a g e
Fig. 17. Error measure by three mentioned techniques
on both data sets. It is observed that wmspca-feed-
forward neural network provides smallest error for
both data sets.
It is observed that wmspca-feed-forward
neural network use Fewer epochs for both data sets.
From all above results, it is observed that wmspca-
feed-forward neural networks outer perform than other
two techniques. Wmspca-feed-forward neural network
used less iteration, less time and small error measure.
It is analyzed from results that using appropriate de-
noising scheme gives better results.
V. CONCLUSION
Different machine learning techniques are
available and being used for stock market prediction.
It was observed from the comprehensive literature
survey and results of applied techniques that, although,
many state of the art techniques are available but data
pre-processing and post-processing have great effect
on results. It is also observed that Recurrent Neural
Network (RNN) performs better than Artificial Neural
Network (ANN) for prediction and Layered Recurrent
Neural Network (LRNN) performs better than Feed-
Forward Neural Network (NN). Layered Recurrent
Neural Network (LRNN) takes less iteration, but more
time. Time consumption factor of LRNN makes it odd
to use. Wsmpca was applied as pre-processing on
Feed-Forward NN and it was observed that pre-
processing methods enhance results of Feed-Forward
Neural Network (NN).
Applying Independent Component Analysis
or Rough Set on Recurrent Neural Network (RNN)
may improve the performance of RNN.
Similarly, Kalman filter may also be used with Echo
State Network (ESN) to improve its performance.
REFERENCES
[1] R.K. Dase, D. D. Pawar and D.S.
Daspute(2011) Methodologies for Prediction
of Stock Market: An Artificial Neural
TABLE 7
COMPARISON OF PERFORMANCE MEASURE (MSEREG)
Dataset
Msereg
in feed-
forward
NN
Msereg in LRNN
Msereg in
Wmspca-feed-
forward neural
network
ggg
PSO
0.15
0.05
0.02
S&P500
21.9
151
9.64
TABLE7 holds performance measure values on both
data sets when applied on above mentioned models. It is
clear that wmspca-feed-forward give small value of
performance measure.
TABLE 8
COMPARISON OF TIME CONSUMED
Dataset
Time
in feed-
forward
NN
Time
in LRNN
Time in
Wmspca-feed-
forward neural
network
ggg
PSO
00:00:19
00:06:22
00:00:07
S&P500
00:31:53
01:09:57
00:00:35
TABLE8 shows time consumption on both data sets in
three applied techniques.
TABLE 9
COMPARISON OF NO. OF EPOCHS
Dataset
epochs
in feed-
forword
NN
epochs
in LRNN
epochs in
Wmspca-feed-
forward neural
network
ggg
PSO
48
36
11
S&P500
1000
263
19
TABLE9 holds number of epochs required for training
data sets in three applied techniques.
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 866 | P a g e
Network. International Journal of Statistika
and Mathematika, volume (1)
[2] Parkash Ramani , Dr. P.D.Murarka (2013)
Stock Market Prediction Using Artificial
Neural Network. International Journal of
Advanced Research in Computer Science and
Software Engineering, Volume(3)
[3] Dase R.K. and Pawar D.D (2010)
Application of artificial Neural Network for
stock market predictions: A literature review.
International Journal of Machine Intelligence.
Volume (2), issue(2)
[4] Asim Maqbool, Umer Malik, Mohsin Fazal,
MutteAhmed(2005), Stock Market
Prediction Software using Recurrent Neural
Networks.
[5] Hykine S(1994) Neural networks: A
comprehensive foundation, Macmillan
college Publication, New york.
[6] Clarence N.W.Tan and Gerhard
E.Witting(1993), A study of parameters of
Back-Propagation for Stock Price Prediction
Modelin proc, first New ZealandInt. Two
Stream conf. Ann and Expert systems.
[7] Arka Ghosh()Comparative study of Financial
Time Series Prediction By Artificial Neural
Network with Gradient Descent Learning
[8] Kim, K-J & Han.I (2000), genetic algorithm
approach to feacture discretization inANN for
the prediction of Price index, Expert System
with Applications
[9] Dynamic financial forcasting, with
automatically induced fuzzy associations, In
Proceeding of the 9th International conference
on Fuzzy Systems.
[10] K.K.Suresh Kumar, Dr.N.M.Elango, kongu
Art and Science College (2012) performance
analysis of stock prediction using ANN,
Global journal of Computer Science and
Technology, Volume(12), issue 1
[11] Leonardo Noriega, school of computing,
staffordshire University,(2005), Multi-layer
perceptron Tutorial
[12] Andrej Krenker, Janez Bešter and Andrej
Kos Consalta d.o.o.Faculty of Electrical
Engineering, University of Ljubljana
Slovenia, introduction to Artifical Neural
Networks www.intechopen.com
[13] Armando bernal, Sam Fok, Rohit
Pidaparthi(2012) financial time series
prediction with Recurrent Neural Netwoks
[14] Chi-Jei-Lu(2010) integrating independent
component analysis based de-noising scheme
with Neural Network for Stock Price
Prediction, Expert System with Applications
[15] Erkam Guresen, Gilgun Kayakutlu, Turgrul
U.Daim(2011) , using Artificial Neural
Network Models in stock market index
prediction, Expert Systems with Applications
[16] Mrs. Simta Agrawal, Dr.P.D.Murarka(2013),
Stock price Prediction: Comparison of Short
Tem and Long Term Stock Price Forcasting
using Various Techniques of Artifical Neural
Networks,International journal of Advanced
Research in Computer Science and Software
Engineering volume(3), issue 6(june 2013)
[17] Emad W. Saad, Donald.C.Wunsch(1998),
Comparative Study of Stock Trend Prediction
Using Time Delay, Reccurent and
Probablistic Neural Nertworks, IEEE
transaction on Neural Networks, volume(9)
issue 6
[18] J. G. Agrawal1, Dr. V. S. Chourasia2,Dr. A.
K. Mittra (2013)State-of-the-Art in Stock
Prediction Techniques, International Journal
of Advanced Research in Electrical,
Electronics and Instrumentation Engineering
[19] Kryzysztof Pawlas, Davood Zall(2012),
Analysis of ForeCasting Methods and
Applications of Sysytem Dynamics and
Genetic Progrsmming: Case Studies on
Country Throughput,School of Computing at
Blekinge Institute Of Technology, sweden
[20] A.F.M.Khodadad Khan, Mohammad Anwer,
Shirpa Banik(2013), Market timing Decisions
by Hybird Machine Learning Technique: A
case Study For Dhaka Stock Market, Journal
of Computations & Modelling Volume(3)
no.2
[21] http://www.mathwork.com
[22] Paolo tenti (1996) , forcasting foreign
exchange rates using recurrent neural
networks, Applied Artificial intelligence
[23] Rough Sets : A Tutorial
[24] Esmaeil Hadavandi, Hassan Shavandi, Arash
Ghanbari (2010)Integration of genetic fuzzy
systems and artificial neural networks for
stockprice forecasting, Knowledge Based
Systems
[25] Akbar Esfahanipour, Werya Aghamiri(2010),
Adapted Neuro-Fuzzy Inference System on
Indirect approach TSK Fuzzy Rule Base for
Stock Market Analysis , volume(37)
[26] George S. Atsalakis, Kimon P.
Valavanis(2009)Forecasting stock market
short-term trends using a neuro-fuzzy based
methodology, Expert System With
Applications
[27] A.P.Papli´nski(2005), Neuro-Fuzzy
Computing.
[28] Hyun-jung Kim , Kyung-shik Shin(2007),
Applied Soft Computing”, A hybrid
approach based on neural networks and
genetic algorithms for detecting temporal
patterns in stock markets, 7 (2007) 569576
[29] Mohammad Ali Farid, Faiza Ather, Omer
Sheikh, Umair Siddiqui(2003), Modelling
and Forecasting Financial Time series,
Zahid Iqbal et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 3, Issue 6, Nov-Dec 2013, pp.855-867
www.ijera.com 867 | P a g e
Comparison of Forecasting ability of Neueral
Networks, Genetic Programming and
Econometric Methods
[30] Howard Demuth, Mark Beale, Neural
Network Toolbox for use with MATLAB
(2002) by The MathWorks
[31] Safdar Ali, Farrukh Aslam. Khan and
Waseem Shahzad, "Intrusion detection using
bi-classifier-based genetic algorithm," ICIC
Express Letters, Vol. 7, No. 6(A), pp. 2147-
2153, December 2010. 1881-803x.
Zahid iqbal received the B.S.
degree in Computer Science from
University of Punjab and M.S.
degree in Computer Science from
NUCES, FAST University, Pakistan
in 2010 and 2012 respectively.
He was a lecturer with Department
of IT, University of Punjab, PK. Since 2012, he is a
Lecturer with the Faculty of Computing and IT,
University of Gujrat, Punjab, PK. His research
interests include evolutionary algorithms, swarm
intelligence, artificial neural networks, computational
intelligence in dynamic and uncertain environments
and real-world applications.
Waseem Shahzad received the M.S.
and Ph.D. degrees in computer
science from the National University
of Computer and Emerging Sciences,
Islamabad, Pakistan, in 2007 and
2010, respectively. Since 2010, he
has been an Assistant Professor with the National
University of Computer and Emerging Sciences. His
current research interests include data mining,
computational intelligence, machine learning, theory
of computation, and soft computing.
He has several publications to his credit.
Javed Anjum received the Phd
degree in Computer Science from
Middlesex University U.K. His
research interests include HCI and
wireless network. He has many
research publications.
Zafar Mahmood received the MSc degree in
Computer Science from Kohat University of Science
and Technology in 2007. Now he is doing MS in
computer science from University of Gujrat. He is
doing research in computational intelligence and
wireless network.
Rafia Ilyas received the MSc degree in Computer
Science from University of Gujrat in 2013. She is
doing research in the field of Neural Network.
... Reference [19] offers the use in the form of exchange rates of repeating neural networks. Additional models used in exchange rates by ANN incorporate [20] Neural networks are more suitable for problems with optimization. It is therefore utilised to optimise the neural network parameters for more accurate predictions. ...
... The discoveries of the investigation are introduced utilizing the most generally utilized exhibition measurements: Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Mean Square Error (MSE), exactness, Sharpe proportion, and return rate are largely instances of root mean square mistake. Iqbal et al [20] purposed another strategy for stock market proficient machine learning. To address the issue of stock market prediction, software engineering is fundamental. ...
Article
Full-text available
Accurate stock market returns are quite difficult for the company because of the unpredictable and non-linear nature of the financial stock markets. With the development of artificial intelligence and increased computer power, programmed prediction approaches have demonstrated that they are increasingly effective in predicting stock values. In this study, the Artificial Neural Network, LSTM, and LR techniques were used to predict the closing price for the following day for five companies belonging to different business sectors. In today's economy, the stock market or equity market has a profound influence. The prediction of stock prices is quite complex, chaotic, and it is a big challenge to have a dynamic environment. Behavioural finance means that investors' decision-making processes are affected by emotions and attitudes in response to particular news. In order to help investors' judgements, we have supplied a technology for the analysis of the stock exchange. The method combines historical price prediction. For predicting, LSTM (Long Short-Term Memory), ANN and LR are employed. It includes the latest information on trade and analytical indicators. Financial data: Open, high, low and close stock prices are used to build new variables needed for model input. The models are validated with standard strategic indicators: RMSE and MAPE. The low values of these two variables indicate that the models are cost-effective.
... The study successfully achieves excess returns as compared to the market. Iqbal (2013) undertakes the application of several neural network models to predict the returns of a single stock in Pakistan stock exchange and suggests that the neural networks have successfully predicted the returns with more accuracy. All these studies have relied on the use of technical analysis but have ignored the established finance theory. ...
... The present study is an endeavor to benefit from the massive computational power of the current times and applies a pure engineering and computer science based tools and pave the way for interdisciplinary research. The existing studies on Pakistan's equity market including Iqbal et al. (2013), Haider and Nishat (2010), and Fatima and Hussain, (2008) have looked at the limited aspects of this subject. The present study has very broad testing horizons regarding the network testing parameters, time span, input out selection, investment theory, and designing of various programs. ...
Article
Full-text available
Artificial Neural Networks (ANN) approach is a relatively new and promising field of the prediction of stock price behavior. Neural networks approach is a mathematical model, flexible enough to accommodate both linear and non-linear aspect of stock returns. This paper applies the ANN to asset pricing models. It is found that the optimum number of neurons does not follow some mathematical rule rather it is based on the presentiment of the researcher to apply an exhaustive search for the number of optimum neurons. Another important finding is that the difference of errors between the testing and training dataset is minimum and the networks are not suffering from the over-fitting phenomenon. The predicted value of high beta portfolios is better than the low beta and mid beta portfolios. This finding reinforces the investment principle that the market compensates the high-risk portfolios more than other classes. The paper concludes that the proposed model achieves a significant improvement in the return on investment and the investors can magnify their profitability.
... There is a lot of research done in this field, but no complete solution has yet been found. Difficulty in predicting the stock market depends not only on the influence of social, political and economic reasons but also on a vast amount of historical data about the stocks and currency (Iqbal et al., 2013). This study consists of many states of the art machine learning techniques to find optimum solutions for predicting stock values. ...
Chapter
In this study, the last two years' hourly opening and closing prices of the banks' stocks traded on BIST-30 were used as the dataset. The research is aimed to predict the closing prices of these stocks in the light of machine learning. In this context, the authors propose a new method containing ensemble learning algorithms and fuzzy clustering technics for predicting stock prices. With this method, authors aim to find stocks which are similar characteristics with test sets and model them together. Thanks to this method, authors aim to improve modelling success. For comparing the results, authors also create models with classical machine learning methods such as support vector machines, random forest, and boosting type new generation algorithms such as extreme gradient boosting and catboost.
... Data that has been collected for the further process must undergo pre-processing and post-processing since it produces a high impact on the results [5]. The authors worked a model which is implemented to reduce the risk. ...
Article
Stock market prediction is considered as one of the most promising research area that is attainning the attention of various researchers. The vital information which is available for access is assumed to have predictive relationships to the future stock returns. The present work gives information to the investors so that the decision could be made better during the purchase of stocks. The factors that contribute towards the decision are the historical prices of stocks and tweet comments regarding the same. The proposed method uses four methods for predicting the stock market status, namely, Linear Regression (LR), Support Vector Machine (SVM), Naïve Bayes (NB), and Random Forest (RF) approaches. When evaluated with standard datasets, experimental results concluded that the SVM based prediction has significant predicting performance than the other methods. The proposed work gives a comparison of factors in order to decide the purchase.
... Short-term investment methods can differ significantly from medium-term and long-term strategies, such as fundamental macroeconomic analysis [9]. Day traders often use technical analyses [12,14,15,21,22]. Machine learning applications are also popular in stock index prediction [2,9,22]. ...
Article
Full-text available
Many day traders focus on forecasts of stock index futures. These securities are suitable for frequent and time-sensitive trading as well as for short-term investments. However, most day traders’ strategies are based on their experiences or news headlines. Combined with a pool trading policy, this may lead to unsatisfactory average monthly profit, particularly when compared to the opportunity cost of the traders’ full-time employment in other non-trading jobs. This paper represents multiple investment strategies for day traders based on visual trend bands on short-term stock index futures. This study uses sequential minimal optimization and other machine learning algorithms to evaluate the performance of visual trend bands and derive strategies for better predictions. This study also applies empirical methods on short-term stock index futures datasets to explore the impact of visual trend bands on short-term stock index trading. The accuracy of our proposed visual trend bands reaches 82%, which is not only an objectively high forecasting accuracy rate but also substantially higher than other visual trend bands. The proposed visual trend bands can support day traders in realizing higher profits in their day trades and short-term investments.
... Laser sensors integrated in the helmets for hurdle detection that a blind can wear, high definition image processing cameras with embedded circuits trained along some particular path that work by processing and comparing the images of the path being followed with the already saved ones [2], line follower applications [3] and sensors integrated within the cane [4] are among some solutions presented recently by the technology for the blinds. Apart from it, solutions to many other problems have also been provided by Artificial Intelligence in the form of; Swarm optimizer for locating and tracking multiple optima [5], hybrid approaches used for UCTP [6], machine learning techniques for stock price prediction [7] and systematic study on OC techniques [8]. The solution proposed by our research aims to offer eyes-free technological support for blind people to walk around and visit specified locations with proper guidance along the generated path indoors. ...
Preprint
Blind navigation is an accessibility application that enables blind to use an android Smartphone in an easy way for indoor navigation with instructions in audio form. We have proposed a prototype which is an indoor navigation application for blinds that uses QR codes. It is developed for android Smart phones and does not require any additional hardware for navigation. It provides automatic navigational assistance on pre-defined paths for blind. QR codes are placed on the floor sections after specific distance that acts as an input for current location detection and navigation. Whenever a QR code is scanned it provides the user with the information of the current location and asks the user to select the destination and then offers optimal and shortest path using path finding algorithms. During navigation whenever the deviation from the proposed path is detected it prompts the user and guides back to the right path by comparing the current path with the generated path. All of the instructions throughout the application are provided in audio form to the user. The interface of the application is well built for blinds which makes the smart phones user-friendly and useable for blind people. The user interacts with the application through a specific set of user-friendly gestures for specific inputs and operations. At the end, we have performed comparison between different state of art approaches and concluded that our approach is more user friendly, cost effective and produced more accurate results.
Preprint
Full-text available
Forecasting Portfolio returns is a challenging task, and conventional forecasting models have partially succeeded in dealing with the nonlinear and complex nature of Equity Markets. Artificial neural networks are a mathematical modelling approach that are resilient enough to forecast portfolio returns in volatile and nonvolatile markets and act like the human brain to simulate the behaviour of stock prices. This research documents the predictive ability of Artificial Neural Networks (ANN) by using the constructs of Fama and French three-factor and five-factor models. A comprehensive methodology of neural networks is applied to achieve the purpose of forecasting. The methodology includes the declaration of the three layers, the hidden layer neurons for processing, and varying parameters for an effective ANN system. The study employs 48-month rolling windows to calculate and compare forecasting errors among competing asset pricing models. The predictive performance of ANN is measured by mean squared, and the accuracy of ANNs under both the pricing models and the accuracy level is evaluated by the Diebold Mariano test. The significant findings of the study include the identification of the optimum architecture of the ANN under both asset pricing models, the nonappearance of the overfitting phenomenon of the networks, and the investor’s compensation for holding high-risk portfolios. JEL Classification: C45, D53, E37, G11, G17
Preprint
Full-text available
In real life, mostly problems are dynamic. Many algorithms have been proposed to handle the static problems, but these algorithms do not handle or poorly handle the dynamic environment problems. Although, many algorithms have been proposed to handle dynamic problems but still, there are some limitations or drawbacks in every algorithm regarding diversity of particles and tracking of already found optima. To overcome these limitations/drawbacks, we have proposed a new efficient algorithm to handle the dynamic environment effectively by tracking and locating multiple optima and by improving the diversity and convergence speed of algorithm. In this algorithm, a new method has been proposed which explore the undiscovered areas of search space to increase the diversity of algorithm. This algorithm also uses a method to effectively handle the overlapped and overcrowded particles. Branke has proposed moving peak benchmark which is commonly used MBP in literature. We also have performed different experiments on Moving Peak Benchmark. After comparing the experimental results with different state of art algorithms, it was seen that our algorithm performed more efficiently.
Article
Full-text available
Forecasting stock market has been a difficult job for applied researchers owing to nature of facts which is very noisy and time varying. However, this hypothesis has been featured by several empirical experiential studies and a number of researchers have efficiently applied machine learning techniques to forecast stock market. This paper studied stock prediction for the use of investors. It is always true that investors typically obtain loss because of uncertain investment purposes and unsighted assets. This paper proposes a rough set model, a neural network model, and a hybrid neural network and rough set model to find optimal buy and sell of a share on Dhaka stock exchange. Investigational findings demonstrate that our proposed hybrid model has higher precision than the single rough set model and the neural network model. We believe this paper findings will help stock investors to decide about optimal buy and/or sell time on Dhaka stock exchange.
Article
Full-text available
A neuro-fuzzy system composed of an Adaptive Neuro Fuzzy Inference System (ANFIS) controller used to control the stock market process model, also identified using an adaptive neuro-fuzzy technique, is derived and evaluated for a variety of stocks. Obtained results challenge the weak form of the Efficient Market Hypothesis (EMH) by demonstrating much improved and better predictions, compared to other approaches, of short-term stock market trends, and in particular the next day’s trend of chosen stocks. The ANFIS controller and the stock market process model inputs are chosen based on a comparative study of fifteen different combinations of past stock prices performed to determine the stock market process model inputs that return the best stock trend prediction for the next day in terms of the minimum Root Mean Square Error (RMSE). Gaussian-2 shaped membership functions are chosen over bell shaped Gaussian and triangular ones to fuzzify the system inputs due to the lowest RMSE. Real case studies using data from emerging and well developed stock markets – the Athens and the New York Stock Exchange (NYSE) – to train and evaluate the proposed system illustrate that compared to the “buy and hold” strategy and several other reported methods, the proposed approach and the forecasting trade accuracy are by far superior.
Conference Paper
Stock market prediction has been a challenging task due to the nature of the data which is very noisy and time varying. However, this theory has been faced by many empirical studies and a number of researchers have successfully applied machine learning approaches to predict stock market. The problem studied here is about stock prediction for the use of investors. It is true investors usually get loss because of unclear investment objective and blind investment. This paper proposes to investigate the rough set model, the artificial neural network model and the hybrid artificial neural network model and the rough set model for determining the optimal buy and sell of a share on a Dhaka stock exchange. Confusion matrix is used to evaluate the performance of the observed and predicted classes for selected models. Our experimental result shows that the proposed hybrid model has higher accuracy than the single rough set model and the artificial neural network model. We believe this paper will be useful to stock investors to determine the optimal buy and sell time on Dhaka Stock Exchange.
Thesis
Objectives. In this study we review previous attempts in forecasting country seaborne container throughput, analyze them and then classify in form of table to provide a concrete base for researchers in this field. Another aim of this study is to provide a Decision Support System (DSS) to assist experts in port management and forecast their country seaborne container demand. It will lead to reasonable decisions so as to provide sufficient supply which handles containers demand. This DSS, is a global forecasting model which can be applied to every country, independently of their specific parameters. Methods. In theoretical phase a number of scientific databases such as: Google Scholar, ACM, SCOPUS, IEEE, SpringerLink and some other are used to collect previous studies. After review and analysis, selected papers are classified in a form of table to provide a complete resource for us as well as future researchers in this field. In order to provide appropriate model, we combine System Dynamics modeling with Genetic Programming to provide an accurate and reliable model. This model is the result of the analysis of previous studies and applied in this study for the first time. Results. Our final model was applied to two cases (Sweden and China) and provides provided reliable results for both countries. To analyze the uncertain variables in the model, Monte Carlo simulation was used to assess the sensitivity of our model. In order to compare with other methods, we conducted a case study with Artificial Neural Network (ANN) and compared the results of our model and ANN. The results show the disadvantages of statistical methods to system dynamics. Additionally to compare with other attempts, our model was confronted with another study which provided a model for Finland. By comparing and considering their advantages and disadvantages we found out that our simplified model could be applied as a global model to other countries. Conclusions. We conclude that our model is an appropriate DSS to assist experts, forecast their country throughput and make appropriate decisions so as to invest, extending their ports in right time. The application of Genetic Programming in our model provides accurate mathematical equations for the influencing variables which even may not need to calibrate the model. It is a global model which can be applied to different countries but still requires more experiments to prove this claim.
Article
Nowadays because of the complicated nature of making decision in stock market and making real-time strategy for buying and selling stock via portfolio selection and maintenance, many research papers has involved stock price prediction issue. Low accuracy resulted by models may increase trade cost such as commission cost in more sequenced buy and sell signals because of insignificant alarms and otherwise bad diagnosis in price trend do not satisfy trader’s expectation and may involved him/her in irrecoverable cost. Therefore, in this paper, Neuro-Fuzzy Inference System adopted on a Takagi–Sugeno–Kang (TSK) type Fuzzy Rule Based System is developed for stock price prediction. The TSK fuzzy model applies the technical index as the input variables and the consequent part is a linear combination of the input variables. Fuzzy C-Mean clustering implemented for identifying number of rules. Initial membership function of the premise part approximately defined as Gaussian function. TSK parameters tuned by Adaptive Nero-Fuzzy Inference System (ANFIS). Proposed model is tested on the Tehran Stock Exchange Indexes (TEPIX). This index with high accuracy near by 97.8% has successfully forecasted with several experimental tests from different sectors.
Article
Stock market prediction is regarded as a challenging task in financial time-series forecasting. The central idea to successful stock market prediction is achieving best results using minimum required input data and the least complex stock market model. To achieve these purposes this article presents an integrated approach based on genetic fuzzy systems (GFS) and artificial neural networks (ANN) for constructing a stock price forecasting expert system. At first, we use stepwise regression analysis (SRA) to determine factors which have most influence on stock prices. At the next stage we divide our raw data into k clusters by means of self-organizing map (SOM) neural networks. Finally, all clusters will be fed into independent GFS models with the ability of rule base extraction and data base tuning. We evaluate capability of the proposed approach by applying it on stock price data gathered from IT and Airlines sectors, and compare the outcomes with previous stock price forecasting methods using mean absolute percentage error (MAPE). Results show that the proposed approach outperforms all previous methods, so it can be considered as a suitable tool for stock price forecasting problems.
Article
This study investigates the effectiveness of a hybrid approach based on the artificial neural networks (ANNs) for time series properties, such as the adaptive time delay neural networks (ATNNs) and the time delay neural networks (TDNNs), with the genetic algorithms (GAs) in detecting temporal patterns for stock market prediction tasks. Since ATNN and TDNN use time-delayed links of the network into a multi-layer feed-forward network, the topology of which grows by on layer at every time step, it has one more estimate of the number of time delays in addition to several control variables of the ANN design. To estimate these many aspects of the ATNN and TDNN design, a general method based on trial and error along with various heuristics or statistical techniques is proposed. However, for the reason that determining the number of time delays or network architectural factors in a stand-alone mode does not guarantee the illuminating improvement of the performance for building the ATNN and TDNN model, we apply GAs to support optimization of the number of time delays and network architectural factors simultaneously for the ATNN and TDNN model. The results show that the accuracy of the integrated approach proposed for this study is higher than that of the standard ATNN, TDNN and the recurrent neural network (RNN).
Article
This article proposes the use of recurrent neural networks in order to forecast foreign exchange rates. Artificial neural networks have proven to be efficient and profitable in fore- casting financial time series. In particular, recurrent networks, in which activity patterns pass through the network more than once before they generate an output pattern, can learn ex- tremely complex temporal sequences. Three recurrent architectures are compared in terms of prediction accuracy of futures forecast for Deutsche mark currency. A trading strategy is then devised and optimized. The profitability of the trading strategy, taking into account trans- action costs, is shown for the different architectures. The methods described here, which have obtained promising results in real-time trading, are applicable to other markets. For years, opposing views have existed between the trading and academic communities about the statistical properties of foreign exchange rates. Traders considered exchange rates to have persistent trends that permitted mechanical trading systems (systematic methods for repeatedly buying and selling on the basis of past prices and technical indicators) to consistently generate profits with relatively low risk. Researchers, on the other hand, presented evidence supporting the random walk hypothesis, which implies that exchange rate changes are independent and have identical statistical distributions. When prices follow a random walk, the only relevant information in the historical series of prices, for traders, is the most recent price. The presence of a random walk in a currency market is a sufficient, but not necessary, condition to the existence of a weak form of the efficient market hypoth- esis, i.e., that past movements in exchange rates could not be used to foretell future movements.
Article
Financial forecasting is an example of a signal processing problem which is challenging due to Small sample sizes, high noise, non-stationarity, and non-linearity,but fast forecasting of stock market price is very important for strategic business planning.Present study is aimed to develop a comparative predictive model with Feedforward Multilayer Artificial Neural Network & Recurrent Time Delay Neural Network for the Financial Timeseries Prediction.This study is developed with the help of historical stockprice dataset made available by GoogleFinance.To develop this prediction model Backpropagation method with Gradient Descent learning has been implemented.Finally the Neural Net, learned with said algorithm is found to be skillful predictor for non-stationary noisy Financial Timeseries.