Fuzzy Expert System for Stock Portfolio Selection: An

Application to Bombay Stock Exchange

Gour Sundar Mitra Thakura , Rupak Bhattacharyyab , Seema Sarkar (Mondal)c

a Departmentof Artificial Intelligence and Machine Learning, Dr. B. C. Roy Engg. College,
Durgapur, West Bengal, India
b Department of Mathematics,Bijoy Krishna Girls’ College, Howrah, West Bengal, India
arXiv:2204.13385v1 [cs.AI] 28 Apr 2022

c Department of Mathematics, National Institute of Technology Durgapur, West Bengal,

India

Abstract
Selection of proper stocks, before allocating investment ratios, is always a cru-
cial task for the investors. Presence of many influencing factors in stock perfor-
mance have motivated researchers to adopt various Artificial Intelligence (AI)
techniques to make this challenging task easier. In this paper a novel fuzzy ex-
pert system model is proposed to evaluate and rank the stocks under Bombay
Stock Exchange (BSE). Dempster-Shafer (DS) evidence theory is used for the
first time to automatically generate the consequents of the fuzzy rule base to
reduce the effort in knowledge base development of the expert system. Later
a portfolio optimization model is constructed where the objective function is
considered as the ratio of the difference of fuzzy portfolio return and the risk
free return to the weighted mean semi-variance of the assets that has been used.
The model is solved by applying Ant Colony Optimization (ACO) algorithm
by giving preference to the top ranked stocks. The performance of the model
proved to be satisfactory for short-term investment period when compared with
the recent performance of the stocks.
Keywords: Ranking of stocks, Stock Portfolio Selection, Fuzzy rule-based
system, Dempster-Shafer Theory.

1. Introduction

Stock portfolio selection is basically an optimal allocation of investor’ assets

among certain number of stocks to provide maximum return with minimum risk.
There are different types of factors which directly or indirectly influence the
performance of stocks. This has made the task of stock selection and portfolio
construction very challenging and uncertain for the stock market researchers

Email addresses: [email protected] (Gour Sundar Mitra Thakur),

[email protected] (Rupak Bhattacharyya), [email protected] (Seema
Sarkar (Mondal))
since the very beginning. According to [1] portfolio selection may include two
stages: firstly, performances of different securities is observed with beliefs about
their future performances and secondly, a proper choice of portfolio is made with
relevant beliefs about future performances. The main aim in Modern Portfolio
Theory (MPT) of investment is to maximize the expected return of the portfolio
for a given amount of portfolio risk or equivalently minimize risk for a given level
of expected return, by carefully choosing the proportions of various assets. In
a ground-breaking work [1], Markowitz has quantified return as the mean and
risk as the variance of the portfolio of securities. The relationship between risk
and mean-variance are later examined in various researches like [2, 3, 4, 5]. The
twin objectives of investors’, profit maximization and risk minimization, are
thus quantified. Though this theory has been widely accepted and adopted, it
is being criticized since last few years. The main reason behind this is in MPT
the efficiency of market is considered to be the basic assumption and obtaining
information about markets in frequent intervals is costly and time consuming [6].
Computational burden caused by the quadratic utility functions and covariance
matrix when number of stock increases, is another major concern in MPT.
Another limitation of MPT is, it does not give importance to real investors’
preferences [7]. In practice, it is also found that investors prefer portfolios
that lie behind the efficient frontier of Markowitz model even though they are
dominated by other portfolios with respect to return and risk. From the above
discussion it is obvious that some additional criteria should be added to the
classical risk-return framework.
Thus portfolio selection problem is proved to be a multi-dimensional problem
and to solve this inherent multi-criteria nature of this problem Multi-criteria
Decision Making (MCDM) approach has been adopted in various researches
like [8, 9, 10, 11]. In spite of all these efforts researchers were facing difficulties
in bringing efficiency in portfolios specially in dynamic uncertain environment.
As a result the use of different AI techniques in stock selection and portfolio
construction became inevitable due to their well known capability of handling
uncertain and vague data. In literature we can find wide use of different AI and
Soft Computing techniques to analyse and predict the performance of stocks in
various stock exchanges around the world. In some researches [12, 13, 14, 15, 16]
efficient learning capability of Artificial Neural Network (ANN) is adopted to
select stocks and construct portfolios whereas in some other researches [17, 18,
19, 20, 21] optimization capability of Genetic Algorithm is used for portfolio
optimization. But the application of Fuzzy Logic and Fuzzy Set theory has
become the most popular among other soft computing techniques, due to its
uncertainty handling capability and the efficiency to incorporate vagueness in
investors’ preferences in portfolio construction [22, 23, 24, 25, 26].
Portfolio selection process can be divided into two stages [7]. In the first
stage, some stocks are selected to construct the portfolio and in the second
stage, the percentage of the total value for each stock is identified. Use of
fuzzy expert systems in the selection of stocks and equity have recently become
very effective and popular. In a work Fasanghari et al. [24] developed an
fuzzy expert system to select superior stocks in Tehran Stock Exchange (TSE)

2
by identifying 7 factors influencing the stock market. A rule base of total 932
rules was constructed for this purpose. Though no proper portfolio optimization
model was used in this work, the outcome of the model proved to be satisfactory.
But the major concern of this model can be the development time and cost due
to repetitive expert interactions for the development of the model. Authors
have used only fuzzy set theory to deal with the inherent uncertainty present in
the rule base though fuzzy set theory is more effective in dealing with vagueness
in the model than addressing inherent uncertainty of the model [27]. In another
significant work Xidonas et al.[7] developed an expert system for the selection
of equity securities but the concerns remained more or less same as the previous
one.
In this proposed work selection of stocks are made with the help of an ex-
pert system where the fuzzy rule base is automatically generated by applying
DS evidence theory. This has considerably minimized the development time
and cost as no cost is incurred here in terms of experts’ consultations for the
construction of the rule base. DS evidence theory is well known for its capa-
bility of dealing with incomplete and uncertain information. So another level
of uncertainty handling mechanism beside fuzzy set theory is incorporated in
the model. Though there are many factors which function as evidences of stock
performance, here four most popular performance indicators: Price to Earning
Ratio (P/E), Price to Book Value (P/B), Price to Sales ratio (P/S) and Long
Term Debt to Equity ratio (LTDER) are initially identified and their historical
data for all the registered stocks under BSE, are collected during the period of
FY 2003-04 to FY 2011-12. In the first phase of the design, these four factors
has functioned as input variables of a fuzzy inference system in the proposed
model and their historical data have been used to generate the consequents of
the fuzzy rules with the help of DS evidence theory. The model is tested with
the historical data of FY 2012-13, when values of these four factors for each
stocks are provided as input to the fuzzy inference system and then they are
ranked based on their defuzzified output. In the second phase, a portfolio op-
timization model is constructed and then it is solved with ACO by considering
top 10 stocks, as per the previous ranking, in the portfolio, to allocate invest-
ment ratios by giving preferences to the top ranked stocks. As mentioned earlier
the model is tested with the data of FY 2012-13 to predict the performance of
different stocks in FY 2013-14 and FY 2014-15, the performance of the model is
proved to be satisfactory when compared with their actual performance. Figure
1 depicts the brief structure of the article.
The rest of the paper is organized as follows: Section 2 gives brief description
about the four identified factors and their importance in stock performance.
In Section 3 design of the proposed fuzzy expert system model is elaborated.
Portfolio construction using top 10 stocks is discussed in Section 4. Result
analysis and comparison with recent performance of stocks is done in Section 5
and finally Section 6 concludes the discussion.

3
 Figure 1: Design diagram of the proposed inflation forecasting model

2.

There are many technical and fundamental factors which influence the stock
market directly or indirectly. In this work four fundamental factors Price to
Earning Ration (P/E), Price to Book Value (P/B), Price to Sales ratio (P/S)
and Long Term Debt to Equity ratio (LTDER) are identified and their brief
definitions are given below:

2.1. Price to Earnings (P/E):

The price to earnings ratio (P/E) is a valuation ratio of a stock that is
determined by Market Value per Share divided by earnings per share (EPS) of
the stock.
P Market Value per Share
= (1)
E EP S
In general, a high price to earnings ratio for any company indicates that investors
are expecting higher earnings growth in future from it compared to companies
with a lower price to earnings ratio.

2.2. Price to Book value (P/B):

Price to book value ration is used to compare market value of a stock to
its Book value. It is also known as price to equity ratio. It is calculated by
dividing the current closing price of the stock by the latest quarter’s book value
per share.
P Stock Price
= (2)
B Total Assets − Intangible Assets and Liabilities

4
In general, If P/B value is low then it suggests that the stock is undervalued
and it could also mean that something is going wrong with the company.

2.3. Price to Sales (P/S):

Price to Sales ratio is another valuation ratio that compares a company’s
stock price to its revenues. It can be calculated by dividing the company’s
market capitalization by its total sale. This ratio is also known as revenue
multiple or sales multiple.
P Market Capitalization
= (3)
S Total Sales

2.4. Long term debt/equity (LTDER):

Long term debt to equity ratio is the ratio of total liabilities of a stock to its
shareholders’ equity. It is a leverage ratio and it measures the degree to which
the assets of the stock are financed by the debts and the shareholders’ equity of
a stock. It is calculated by the total liabilities divided by shareholders’ equity.
Total Liabilities
LT DER = (4)
Shareholders’ Equity

3. DS-Fuzzy model for evaluation and ranking of stocks

DS evidence theory and fuzzy rule based expert system are hybridized in
this work to evaluate and rank the stocks under BSE. Brief introduction of DS
evidence theory and fuzzy rule based expert system are given here before their
application in the proposed model is elaborated.

3.1. DS evidence theory

A. P. Dempster in 1967 [28] proposed a multi-valued mapping from one space
to another which was used for statistical inference when we needed to identify
a single hypothesis from multiple sample information. The DS-theory was first
proposed by Dempster in 1968 [29] and later in 1976 it was extended by Shafer
[30] to deal with incomplete and uncertain information.
DS-Theory mainly deals with four components: Frame of discernment, Basic
Probability Assignment (BPA), plausibility function Pl and belief function Bel.
In evidence theory first a set of hypothesises Θ called frame of discernment,
is assumed and defined as follows:

Θ = {H1 , H2 , H3 , ..., HN }. (5)

The set is composed of N mutually exhaustive and exclusive hypothesises. From

the Θ, let us denote P (Θ), be the power set composed with 2N propositions Xof
Θ:
P (Θ) = {∅, {H1 }, {H2 }, ..., {HN }, {H1 ∪ H2 }, {H1 ∪ H3 }, ..., Θ} (6)

5
where ∅ denotes the empty set. The most important task in applying evidence
theory is Basic Probability Assignment (BPA). A BPA is a function from P (Θ)
to [0, 1] defined by:

m : P (Θ) → [0, 1]
(7)
X 7→ m(A)
and which satisfies the following criteria:
X
m(X) = 1;
X∈P (Θ) (8)
m(∅) = 1.

Dempster’s rule of combination is denoted by m = m1 ⊕ m2 , which can

combine two BPAs m1 and m2 to yield a new BPA:
P
m1 (Y )m2 (Z)
m(A) = Y ∩Z=X (9)
1−k
with X
k= m1 (Y )m2 (Z) (10)
Y ∩Z=∅

where k is a normalization constant, known as conflict because it measures the

degree of conflict between m1 and m2 . k = 0 corresponds to the absence of
conflict between m1 and m2 , whereas k = 1 implies total contradiction between
them. The resulting belief function from the combination of I information
sources can be defined as:

m = m1 ⊕ m2 ⊕ m3 ... ⊕ mI (11)

With the help of Equation (11) multi source information can be fused into a
single framework in belief theory very easily and it proved to be very effective
when it is applied in fuzzy rule generation discussed later in Section 3.3.3.

3.2. Fuzzy Expert System

For many real world problems, like strategic planning, medical diagnosis
and other decision making problems solutions are something more than simple
reasoning and require some expertise knowledge to solve. So the use of Expert
Systems (ES) became popular because it is capable of representing and reasoning
knowledge rich domain with a view to solve problems and giving advice [31].
The main advantage of ES over any other conventional software applications
is its effective reasoning capability as they process knowledge instead of data
or information [32]. On the other hand the application of fuzzy set theory
in uncertainty handling has become one of the strongest tool to design any
framework. As fuzzy set theory is now widely applied in information gathering,
modelling, analysis, optimization, control, decision-making and supervision [33],
fuzzy expert system is used in the construction of the proposed model.

6
 Fuzzy expert system is an expert system that uses fuzzy logic instead of
Boolean logics in its knowledge base and derives conclusion from user inputs
and fuzzy inference process [34]. Figure 2 depicts the basic architecture of a
fuzzy expert system.

Figure 2: Architecture of Fuzzy Rule-based Expert System

Various components of a fuzzy expert system are briefly discussed below:

• Fuzzy Knowledge Base: Fuzzy knowledge base is a collection of fuzzy
rules and it is built up by the knowledge engineer after eliciting knowledge
from domain experts. The basic structure of a fuzzy rule is as follows:
IF I1 is i1 AND I2 is i2 AND ... AND In is in THEN O is o.

where I1 , I2 ... , In are fuzzy linguistic input variables and i1 , i2 , ... , in

are possible linguistic values of I1 , I2 ... , In respectively. Similarly, O is
fuzzy linguistic output variable and o is its linguistic value. The terms I1 is
i1 and I2 is i2 ... are fuzzy propositions suggesting partial set membership
or partial truth. In this way modifying the fuzzy sets parameters of the
output fuzzy sets, the partial truth of the rule premise can be evaluated
[35].
• Fuzzy Inference Engine: To analyse and manipulate the rules in the
knowledge base an inference mechanism is adopted to conclude a logical
result. Various inference mechanisms can be adopted for fuzzy inference
engine depending on the aggregation, implications and operators used for
s-norms and t-norms [36]. Defuzzification of the results is also achieved in
this module. This modules functions as per the agenda. An agenda is a
prioritized list of rules identified and enlisted by inference engine. Patterns
of these rules are satisfied by facts or objects in the working memory.
• Working Memory: This module is contained of a global database of
facts those are going to be used by the fuzzy rule base.

7
 • Explanation Facility: This particular module explains every steps in
reasoning process of the ES.
• Knowledge Acquisition Facility: This module provides an automated
facility for the user to enter knowledge in the system.
• User Interface: All kinds of interactions of the user with the system is
facilitated by this module. This is used to accept inputs from the user
and display the outcomes of the inference process.

3.3. Design and Implementation of the proposed model

The designed diagram of the proposed model is introduced in Section 1 as
Figure 1. Here individual steps followed to construct and implement the whole
model are elaborated.

3.3.1. Collection of Historical Data

Total 30 stocks are registered under BSE. Initially, historical data of last 12
years (2003-04 to 2014-15) of four factors (P/E, P/B, P/S and LTDER) are col-
lected from different web sources like www.capitaline.com, www.bseindia.com,
www.nseindia.com etc. First 9 years’ data are used for the implementation of
the model and last 3 years’ data are used for testing and comparison. For the
simplicity of the model the data are normalized within the range of [0, 10] by
scaling up/down the maximum historical value of any factor in last 9 years to 10
and then scaling other data accordingly relative to the maximum value. Table
1 shows an example of actual data and transferred for Axis Bank Ltd.

Table 1: Sample normalized data for Axis Bank Ltd.

Actual P/E ratio of

Financial Year Transferred Data
Axis Bank Ltd.
2011-12 11.46 4.24
2010-11 17.5 6.48
2009-10 19.47 7.2
2008-09 8.49 3.14
2007-08 27.02 10
2006-07 21.67 8.02
2005-06 21.12 7.82
2004-05 20.5 7.59
2003-04 12.54 4.64

3.3.2. Fuzzification
Fuzzification is a process of taking a crisp value as input and transforming
it into the degree required by the terms. Uncertainty basically comes from
imprecision, vagueness or ambiguity and one of the best ways to represent these
in any variable can be in the form of fuzzy variable. Fuzzy membership functions

8
are used to represent the fuzzy variables. All four factors have been illustrated
in TABLE II in terms of fuzzy linguistic variables and each of them has three
linguistic values, Low, Standard and High. Trapezoidal membership function
has been used to represent the inputs within the range of [0, 10].

Table 2: Membership functions for the linguistic values of the Input Variables

Factors in terms of Fuzzy Trapezoidal

Linguistic Values
linguistic variables Membership
Low (0 0 1.8 2.8)
P/E Ratio (Range 0
Standard (1.7 3.5 4.6 5.8)
to 10)
High (5.3 7.5 10 10)
Low (0 0 2.2 3.5)
P/B Ratio (Range 0
Standard (2.5 4.6 5.6 7.9)
to 10)
High (6.4 9.6 10 10)
Low (0 0 2.4 3.6)
P/S Ratio (Range 0
Standard (1.8 4.4 5.7 8.2)
to 10)
High (6.4 8.7 10 10)
Low (0 0 2.3 3.6)
LTDER (Range 0 to
Standard (2.8 4.6 5.7 7.6)
10)
High (6.124 8.09 10 10)

3.3.3. Fuzzy Rule Construction using DS-Theory

The most significant novelty of this research work is the application of DS
evidence theory in fuzzy rule construction. The knowledge base of any fuzzy
inference system is developed using collection of fuzzy rules which determines
how the output will be generated based on the given input. The proposed system
has four input parameters namely, P/E, P/B, P/S and LTDER, and based on
these input parameter the DS-Fuzzy system is expected to determine whether
the selection of any stock will be Highly Favourable, Moderately Favourable or
Not Favourable, as output. The frame of discernment is considered as Θ =
{High Performance, Average Performance, and Poor Performance}.
Risk/Return ratio is one of the most commonly used measure to evaluate
the performance of stocks. Risk is generally measured by semi-variance(S) of
the stocks’ previous returns and Return(R) is actually the mathematical mean
of previous returns. Obviously lower value of this ratio indicate better perfor-
mance of stocks. Based on linguistic values of each input factors and comparing
those with S/R values as a measure of the performance of various stocks BPAs
for different hypothesises are achieved. The historical data of FY 2012-13 are
used to identify the stocks. Consulting various domain experts and analysing
historical data initially standard values for each factors for all the stocks are
decided. By analysing last 9 years’ historical data it is found that when P/E
ratio was around its standard value then 75% of stocks under BSE performed
better and S/R value was also found to be satisfactory. So a belief of 0.75 is as-
signed towards the hypothesis {High Performance} when the P/E ratio is near

9
to its standard value. Again, when P/E value found to be much lower than
its standard value, 65% of total stocks under BSE performed average and S/R
value also found to be average. So 0.6 degree of belief is assigned towards the
hypothesis {Average Performance} when P/E ratio is lower than its standard
value. Similarly, when the value of P/E ratio was much higher than the stan-
dard value then it was found that 70% of stocks under BSE performed poor
and S/R value was also very high. So 0.7 degree of belief is assigned towards
the hypothesis {Poor Performance}. In the same way initial believes or Basic
Probabilities are assigned for all other factors as mentioned in Table 3.

Table 3: Basic Probability Assignment

High Average Poor

Linguistic
Factors Performance Performance Performance
Values
(H P) (A P) (P P)
Low 0.6
P/E Standard 0.75
High 0.7
Low 0.8
P/B Standard 0.6
High 0.65
Low 0.7
P/S Standard 0.65
High 0.75
Low 0.6
LTDER Standard 0.75
High 0.65

Now to elaborate further let us consider the following four sample rules with
their initial belief:
Rule 1. IF P/E is Low THEN Performance will be Average (m1 (A P ) =
0.6).
Rule 2. IF P/B is Standard THEN Performance will be Average (m2 (A P ) =
0.6).
Rule 3. IF P/S is High THEN Performance will be High (m4 (H P ) = 0.75).

Rule 4. IF LTDER is High THEN Performance will be Poor (m6 (P P ) =

0.65).

The above four rules represent the belief towards the performance of stocks with
respect to the values of four input factors as the presence of evidences. Now
consider the IF part of Rule 1 and Rule 2 as the first two evidences and m1 and
m2 as mass functions for them respectively.

10
 Now from Rule 1, m1 (A P ) = 0.6 and m1 (Θ) = (1 − 0.6) = 0.4, where,
m1 (Θ) represents the degree of belief in the rest of the hypotheses present
in the hypothesis set. Similarly, from Rule 2, m2 (A P ) = 0.6 and m2(Θ) =
(1 − 0.6) = 0.4. Now these two mass functions are combined using Dempster’s
rule of combination (Equation 9) to obtain a new function m3 as mentioned in
Table

Table 4: Combination of mass considering first two evidences

Combining m1 and m2 m2 (A P ) = 0.6 m2 (Θ) = 0.4

m1 (A P ) = 0.6 A P = 0.36 A P = 0.24
m1 (Θ) = 0.4 A P = 0.24 Θ = 0.16

The new mass value m3 for the same hypothesis set is calculated as:

m3 (A P ) = 0.36 + 0.24 + 0.24 = 0.84 
(12)
m3 (Θ) = 0.16 

Now consider the IF part of the Rule 3 as the new evidence and m4 be
the new mass function. From Rule 3, m4 (H P ) = 0.75 and m4 (Θ) = (1 −
0.75) = 0.25. Now again, m3 and m4 are combined to generate new mass m5
as mentioned in Table 5.

Table 5: Combination of mass considering first three evidences

Combining m3 and m4 m4 (H P ) = 0.75 m4 (Θ) = 0.25

m3 (A P ) = 0.84 Φ = 0.63 A P = 0.21
m3 (Θ) = 0.16 H P = 0.12 Θ = 0.04

New mass values m5 for the hypothesis set are calculated as below:
0.21
m5 (A P ) = = 0.567 

1−0.63 

0.12
m5 (H P ) = 1−0.63 = 0.324 (13)


0.04 
m5 (Θ) = 1−0.63 = 0.108

Now if we consider the IF part of the Rule 4 as the last evidence and m6
as its mass function then we get m6 (H P ) = 0.6 and m6 (Θ) = (1 − 0.6) = 0.4.
Now finally, m5 and m6 are combined as mentioned in Table 6 to generate the
final mass values mf for the hypothesis set (Equation 14).

11
 Table 6: Combination of mass considering all four evidences

Combining m5 and m6 m6 (P P ) = 0.65 m6 (Θ) = 0.35

m5 (A P ) = 0.567 Φ = 0.368 A P = 0.198
m5 (H P ) = 0.324 Φ = 0.21 H P = 0.113
m5 (Θ) = 0.108 P P = 0.07 Θ = 0.037

0.113

mf (H P ) = 1−0.578 = 0.267 



0.198 
mf (A P ) = 1−0.578 = 0.469 


0.07 (14)
mf (P P ) = 1−0.578 = 0.165 



0.37 
mf (Θ) = 1−0.578 = 0.087 



Following the same procedure final mass values for the rest 80 rules are
calculated. For the hypothesis High Performance(H P) maximum and minimum
mass values (mf ) are found to be 0.9916 and 0 respectively. As mass values for
any hypothesis can lie between 0 and 1 we have divided the favourability of
any stock, based on the mf (H P ) of the rules, into three categories: Highly
Favourable (0.76 ≤ mf (H P ) ≤ 1), Moderately Favourable (0.46 ≤ mf (H P ) ≤
0.75) and Not Favourable (mf (H P ) ≤ 0.45).
So the combined version of the four sample rules mentioned earlier is formed
as:
“IF P/E is Low AND P/B is Standard AND P/S is High AND LTDER
is High THEN the stock is Not Favourable”.
In this way total 81 fuzzy rules are formulated using DS evidence theory to
develop the knowledge base of the proposed DS-fuzzy inference system and the
above mentioned three selection categories are converted into fuzzy linguistic
variables with trapezoidal membership values as shown in Table 7 for output of
the proposed inference system.

Table 7: Degree of membership of linguistic values for the output variable

Fuzzy Trapezoidal
Output Variable Linguistic Values
Membership
Not Favourable (0 0 0.172 0.448)
Selection (Range: 0
Moderately
to 1) (0.34 0.46 0.57 0.75)
Favourable
Highly Favourable (0.64 0.88 1 1)

Figure 3 shows the snapshot of the fuzzy rule base and Figure 4 shows the
sensitive analysis for the same when designed in MATLAB.

12
 Figure 3: Fuzzy rule base of the DS-fuzzy system for Stock Selection

Figure 4: The Sensitive analysis of the DS-fuzzy system for Stock Selection

13
3.3.4. Defuzzification
The process of obtaining a single value that represent the best outcome of the
fuzzy set evaluation is known as defuzzification [37]. Though there are many
other popular defuzzification methods namely, Bi-sector, Mean of Maximum
(MOM), Smallest of Maximum (SOM), Largest of Maximum (LOM), in this
proposed model ‘Centroid’ is used for the defuzzification purpose due to its
wide range of acceptability. This method was introduced by Sugeno in 1985
and it can be expressed as:
R
µC (z) · zdz
∗
z = R (15)
µC (z)dz
e
e
Where z is the output variable, z ∗ is the defuzzified output value and µC (z) is
the aggregated membership function. e

3.3.5. Ranking of stocks

The proposed DS-fuzzy system is used to rank all 30 registered stocks under
BSE based on the defuzzified outputs when the values corresponding to each of
the four input factors for 2012-13 for each stocks are provided as input to the
system. Top 10 stocks from this ranking are mentioned in Table 8

Table 8: Top 10 Stocks based on the input data for FY 2012-13

Defuzzified
Sl. No Name of the Stocks
Values(FY 2012-13)
1 Hindustan Unilever Ltd. 0.8644
2 Sun Pharmaceutical Inds. Ltd. 0.8457
3 I T C Ltd. 0.5369
4 Coal India Ltd. 0.534
5 Tata Consultancy Services Ltd. 0.4936
6 Infosys Ltd. 0.3842
7 Dr. Reddy’S Laboratories Ltd. 0.3553
8 Bajaj Auto Ltd. 0.3292
9 Hero Motocorp Ltd. 0.3002
10 Cipla Ltd. 0.285

As the consequent of every rule indicates the favourability of the stocks,

higher defuzzified value indicate higher favorability. The highest defuzzified
value of 0.8691 is found for Hindustan Unilever Ltd. among all 30 registered
stocks and it has topped the ranking. These 10 stocks are used for portfolio
construction as discussed in the next section.

4. Portfolio Construction
The first phase of stock portfolio selection, i.e. the selection of suitable
stocks, is achieved in the previous section. In this section a portfolio construction

14
model is discussed to determine optimum investment ratios for the selected
securities such that the overall return is maximized under a tolerable risk.

4.1. Objective Function

The ratio of the difference of fuzzy portfolio return and the risk free return to
weighted mean semi-variance of the assets is used as the objective function in this
work. Certainly, higher value of this ratio will indicate the better investment;
so the optimization target will be to maximize the ratio.
The notations used for the construction of this objective function are men-
tioned below:
• xi : Fraction of the total investment allotted to the ith asset;

• r˜i : Fuzzy return of the ith asset;

• rf : Risk free return rate;
• µs : Weighted mean of asset semi-variances;
• rp : Portfolio return;

• vp : Variance of the portfolio;

• sp : Skewness of the portfolio;
Thus the objective function is formed as:
P
E( r˜i xi ) − rf
M aximize (16)
µs

xi si , i.e. xi is the ith weight

P
Where, after descending sort of portfolio, µs =
in the descending order and si is the semi-variance of the ith ranked asset.
A fuzzy aggregation is used to calculate the fuzzy returns of the securities
from the statistical database of past five years(2008-09 to 2012-13). If ti means
ith position of the data, the fuzzy return can be calculated as:
 P 
ti ri
r˜i = min(ri ), P , max(ri ) (17)
ti

The constraints included in this model are:


rp > α, vp > β, sp > γ 

n
P (18)
xi = 1, xi ≤ M, xi > 0, ∀i 

i=1

Values for α,β, γ, M and m are decided based on the investor’s preferences.

15
 Thus the final portfolio optimization model can be summarized as below:
P 
E( r˜i xi )−rf
M aximize µs





Subject to,




rp > α, vp > β, sp > γ

(19)
Pn 

xi = 1, xi ≤ M, xi > 0, ∀i 



i=1 




4.2. Optimization of the proposed model using ACO

An algorithm using Ant Colony Optimization (ACO) is proposed and imple-
mented in this section to solve the optimization model as mentioned in Equation
(19). ACO is a very popular meta-heuristic optimization technique inspired by
the foraging behaviour of biological ants [38, 39]. Pseudo code of the proposed
algorithm is given as below:

Algorithm 1 ACO algorithm for portfolio optimization

1: procedure ACO-Portfolio
2: Generate N random solution nodes based on Equation (19);
3: Initialization of the ACO;
4: for ITERATION=1 to I do
5: for ANT=1 to Γ do
6: Select the start node randomly;
7: for LIFETIME=2 to L do
8: Select the next node based on the heuristic information and
pheromone concentration in the path. Move to the next node
only if it is better than the current node;
9: Update pheromone on the selected path;
10: end for
11: Store the details of the final node reached by each ants;
12: end for
13: Identify the solution node where maximum number of ants reached
and consider that to be the optimum solution for the current solution;
14: Update the pheromone on the path of each ants who have reached
this optimum solution;
15: Evaporate the pheromone from all paths.
16: end for
17: end procedure

To execute the above algorithm in this work, initially 2000 random solutions
are generated for the ant colony of 50 ants. Total 400 iterations are used for the
optimization purpose while lifetime for each ant is considered as 20.

16
 In Equation (19), r˜i is represented as a triangular fuzzy number. The Ex-
pected Return (E), Variance (Var), Skewness (Skew) and Semi-variances of these
top 10 securities for past five years (2008-09 to 2012-13) as used in the above
algorithm are calculated by the following theorem and are mentioned in Table
9

Theorem 4.1. Let à = (a, b, c) be a triangular fuzzy number. The the weighted
possibilistic mean, variance and skewness can be calculated as [40]:

= 16 (a + 4b + c)

E(Ã) 



1 2
V ar(Ã) = 18 (a + b2 + c2 − ab − bc − ca)

(20)
19(a3 +c3 )−8b3 −42b(a2 +c2 )+12b2 (a+c)−15(a2 c+ac2 )+60abc


Skew(Ã) = √ √
10 2( a2 +b2 +c2 −ab−bc−ca)3




Table 9: Expected Return, Variance, Skewness and Semivariance of stocks

Rank Name of the Stocks Return Variance Skewness Semivariance

Hindustan Unilever
1 0.1443 0.0001 0.2391 0.00004
Ltd.
Sun Pharmaceutical
2 0.0529 0.0006 -0.3269 0.00254
Inds. Ltd.
3 I T C Ltd. 0.2801 0.0035 0.7523 0.00314
4 Coal India Ltd. 0.1275 0.0028 0.6046 0.0373
Tata Consultancy
5 0.1228 0.0013 -0.5029 0.00557
Services Ltd.
6 Infosys Ltd. 0.0356 0.0003 0.8648 0.0002
Dr. Reddy’S
7 0.0484 0.0006 -0.0065 0.00098
Laboratories Ltd.
8 Bajaj Auto Ltd. 0.0579 0.001 0.7262 0.0013
9 Hero Motocorp Ltd. 0.1379 0.0004 0.8057 0.00049
10 Cipla Ltd. 0.0437 0.0002 0.8258 0.00018

When the algorithm is executed in MATLAB with the above dataset and
considering other parameters as rf = 0.01, β = 0.5, α = 0.05, γ = 0.001,
M = 0.8 and µs = 0.0016, the maximum return is found as 0.1317. The
proposed ratio allocation for this return is given in Table 10 .
The convergence of the objective values as per the proposed model is depicted
in Figure 5 and the accumulation of ants to the optimum objective values in
each iteration is depicted in Figure 6.

17
 Table 10: Ratio allocation for the proposed portfolio

Sun Dr. Hero

Hindustan Coal Bajaj
Pharma. ITC TCS Infosys Reddy’s Moto- Cipla
Unilever India Auto
Inds. Ltd. Ltd. Ltd. Lab. corp Ltd.
Ltd. Ltd. Ltd.
Ltd. Ltd. Ltd.
0.2403 0.2235 0.1970 0.0764 0.0574 0.0525 0.0452 0.0413 0.0407 0.0255

28

27

26

25

24

23

22

21
0 100 200 300 400 500

Figure 5: Convergence of objective values based on proposed ranking

50

45

40

35

30

25

20

15

10

0
0 100 200 300 400 500

Figure 6: Ant accumulation at optimum solutions

18
5. Result Analysis

To justify the applicability of the proposed model its performance is analyzed

further in three aspects. First, top 15 companies are ranked based of their S/R
values for the year of 2013-14 and 2014-15. Details of these rankings along with
the ranking of the proposed model is given in Table 11.

Table 11: Top 15 Stocks

Stocks Ranking Stocks Ranking Stocks Ranking
Rank based on the based on the S/R based on the S/R
proposed model values (2013-14) values (2014-15)
Hindustan Unilever State Bank Of
1 Sesa Sterlite Ltd.
Ltd. India
Sun
2 Pharmaceutical NTPC Ltd. Infosys Ltd.
Inds. Ltd
Hero Motocorp
3 ITC Ltd. ITC Ltd.
Ltd.
Maruti Suzuki
4 Coal India Ltd. HDFC Bank Ltd.
India Ltd.
Hindustan Unilever Hero Motocorp
5 TCS Ltd.
Ltd. Ltd.
Hindustan Unilever
6 Infosys Ltd. Cipla Ltd.
Ltd.
Dr. Reddy’s State Bank Of
7 Wipro Ltd.
Laboratories Ltd. India
Bharat Heavy Maruti Suzuki
8 Bajaj Auto Ltd.
Electricals Ltd. India Ltd.
Hero Motocorp Tata Power Co.
9 Wipro Ltd.
Ltd. Ltd.
10 Cipla Ltd. ITC Ltd. Tata Motors Ltd.
Tata Power Co.
11 HDFC Bank Ltd. TCS Ltd.
Ltd.
12 Wipro Ltd. Infosys Ltd. Coal India Ltd.
Larsen & Toubro Hindalco Industries Mahindra &
13
Ltd. Ltd. Mahindra Ltd.
Dr. Reddy’S
14 Bharti Airtel Ltd. Cipla Ltd.
Laboratories Ltd.
Sun Sun
Tata Power Co.
15 Pharmaceutical Pharmaceutical
Ltd.
Inds. Ltd. Inds. Ltd.

From the above table we can find a match of 9 companies for FY 2013-14 and a
match of 11 companies for FY 2014-15. As S/R values are effective performance
indicator of stocks, it can be concluded that the proposed model will give better

19
return in short-term investment period.
In the second stage by considering top 10 stocks under BSE, based on their
S/R values for FY 2012-13, a portfolio is constructed using the same ACO
algorithm and the objective function (Eq. 19). The optimum return is found to
be 0.0740, which is much less than the expected return (0.1317) of the proposed
model.
In the third stage another portfolio is constructed using the same ACO
algorithm, where all 30 registered companies were under BSE are considered.
But the optimum return found to be 0.0624 which is much lower than the return
calculated by the proposed model. This again proves that the proposed model is
more effective in portfolio construction. Figure 7 compares the returns of these
three portfolios.

Figure 7: Portfolio returns

6. Conclusion

In this research a novel fuzzy expert system model is designed and implemented
to evaluate and rank the stocks under BSE. DS-evidence theory is used for
the development of the fuzzy rule base to reduce the overall implementation
time and cost of the system. A portfolio optimization model is constructed
by considering the ratio of the difference of fuzzy portfolio return and risk free
return to the weighted mean semi-variance of the assets as the objective function.
ACO is used to obtain the optimum value of this portfolio.
As the outcome of this model found to be satisfactory, this can be imple-
mented for any stock exchanges around the world but the selection of critical
factors may vary over different stock exchanges. This fuzzy expert system model
can be used to rank any set of alternatives based on the factors influencing them.
The portfolio optimization model and the related ACO algorithm can be used
to optimize any rank-preference based portfolios.

20
References

[1] H. Markowitz, Portfolio selection, Journal of Finance, 7 (1952) 77–91.

[2] X. Y. Zhou, D. Li, Continuous-time mean-variance portfolio selection: A
stochastic lq framework, Applied Mathematics Optimization 42 (2000) 19–
33.
[3] Y. Fang, S. Wang, An interval semi-absolute deviation model for portfolio
selection, in: Springer, LNAI 4223, 2006.

[4] W. Ogryczak, Multiple criteria linear programming model, Annals of Op-

erations Research 97 (2000) 143–162.
[5] X. Huang, Risk curve and fuzzy portfolio selection, Computers and Math-
ematics with Applications 55 (2008) 1102–1112.
[6] S. J. Grossman, J. E. Stiglitz, On the impossibility of informationally effi-
cient markets, The American economic review 70 (3) (1980) 393–408.
[7] P. Xidonas, E. Ergazakis, K. Ergazakis, K. Metaxiotis, D. Askounis,
G. Mavrotas, J. Psarras, On the selection of equity securities: An expert
systems methodology and an application on the athens stock exchange,
Expert Systems with Applications 36 (2009) 11966—-11980.

[8] P. Xidonas, G. Mavrotas, C. Zopounidis, J. Psarras, Ipssis: An integrated

multicriteria decision support system for equity portfolio construction and
selection, European Journal of Operational Research 210 (2) (2011) 398–
409.
[9] R. D. Edwards, J. Magee, W. Bassetti, Technical analysis of stock trends,
CRC Press, 2007.
[10] F. Abdollahzadeh, Investment management and tehran stock exchange
(2002).
[11] J. Siskos, D. Despotis, A dss oriented method for multiobjective linear
programming problems, Decision Support Systems 5 (1) (1989) 47–55.
[12] A. Adebiyi, C. Ayo, M. O. Adebiyi, S. Otokiti, Stock price prediction using
neural network with hybridized market indicators, Journal of Emerging
Trends in Computing and Information Sciences 3 (1) (2012) 1–9.
[13] A. Fernández, S. Gómez, Portfolio selection using neural networks, Com-
puters & Operations Research 34 (4) (2007) 1177–1191.
[14] P.-C. Ko, P.-C. Lin, Resource allocation neural network in portfolio selec-
tion, Expert Systems with Applications 35 (1) (2008) 330–337.

21
[15] S. O. Olatunji, M. S. Al-Ahmadi, M. Elshafei, Y. A. Fallatah, Saudi arabia
stock prices forecasting using artificial neural networks, in: Applications
of Digital Information and Web Technologies (ICADIWT), 2011 Fourth
International Conference on the, IEEE, 2011, pp. 81–86.
[16] F. A. De Oliveira, L. E. Zárate, M. de Azevedo Reis, C. N. Nobre, The
use of artificial neural networks in the analysis and prediction of stock
prices, in: Systems, Man, and Cybernetics (SMC), 2011 IEEE International
Conference on, IEEE, 2011, pp. 2151–2155.
[17] J.-S. Chen, Y.-T. Lin, A partitioned portfolio insurance strategy by a rela-
tional genetic algorithm, Expert Systems with Applications 36 (2) (2009)
2727–2734.
[18] T.-J. Chang, S.-C. Yang, K.-J. Chang, Portfolio optimization problems
in different risk measures using genetic algorithm, Expert Systems with
Applications 36 (7) (2009) 10529–10537.

[19] J.-S. Chen, J.-L. Hou, S.-M. Wu, Y.-W. Chang-Chien, Constructing in-
vestment strategy portfolios by combination genetic algorithms, Expert
Systems with Applications 36 (2) (2009) 3824–3828.
[20] J. R. Jiao, Y. Zhang, Y. Wang, A heuristic genetic algorithm for product
portfolio planning, Computers & Operations Research 34 (6) (2007) 1777–
1799.

[21] Y. Chen, S. Mabu, K. Hirasawa, A model of portfolio optimization us-

ing time adapting genetic network programming, Computers & operations
research 37 (10) (2010) 1697–1707.
[22] J. D. Bermudez, J. V. Segura, E. Vercher, A fuzzy ranking strategy for
portfolio selection applied to the spanish stock market, in: Fuzzy Systems
Conference, 2007. FUZZ-IEEE 2007., 2007, pp. 1–4.
[23] A. Bilbao-Terol, B. M. Pér-Gladish, Arenas-Parra, M. V. Rodrı́guez-Uria,
Fuzzy compromise programming for portfolio selection, Applied Mathemat-
ics and Computation 173 (2006) 251–264.

[24] M. Fasanghari, G. A. Montazer, Design and implementation of fuzzy ex-

pert system for tehran stock exchange portfolio recommendation, Expert
Systems with Applications 37 (9).
[25] X. Huang, Risk curve and fuzzy portfolio selection, Computers & Mathe-
matics with Applications 55 (6) (2008) 1102–1112.

[26] F. Tiryaki, M. Ahlatcioglu, Fuzzy stock selection using a new fuzzy rank-
ing and weighting algorithm, Applied Mathematics and Computation 170
(2005) 144–157.

22
[27] R. M. Rodrı́guez, L. Martı́nez, F. Herrera, Hesitant fuzzy linguistic term
sets for decision making, Fuzzy Systems, IEEE Transactions on 20 (1)
(2012) 109–119.
[28] A. P. Dempster, Upper and lower probabilities induced by a multivalued
mapping, The annals of mathematical statistics (1967) 325–339.

[29] A. P. Dempster, A generalization of bayesian inference, Journal of the Royal

Statistical Society. Series B (Methodological) (1968) 205–247.
[30] G. Shafer, et al., A mathematical theory of evidence, Vol. 1, Princeton
university press Princeton, 1976.

[31] P. Jackson, Introduction to expert systems, Addison-Wesley Pub. Co.,

Reading, MA, 1986.
[32] K. Darlington, The essence of expert systems, Prentice Hall, 2000.
[33] R. E. Bellman, L. A. Zadeh, Decision-making in a fuzzy environment, Man-
agement science 17 (4) (1970) B–141.
[34] A. Kandel, Fuzzy expert systems, CRC press, 1991.
[35] C. Matthews, A formal specification for a fuzzy expert system, Information
and Software Technology 45 (7) (2003) 419–429.

[36] L.-X. Wang, A course in fuzzy systems, Prentice-Hall press, USA, 1999.
[37] E. W. T. Ngai, F. K. T. Wat, Design and development of a fuzzy expert
system for hotel selection, Omega 31 (4).
[38] M. Dorigo, M. Birattari, T. Stützle, Ant colony optimization, Computa-
tional Intelligence Magazine, IEEE 1 (4) (2006) 28–39.

[39] J.-L. Deneubourg, S. Aron, S. Goss, J. M. Pasteels, The self-organizing

exploratory pattern of the argentine ant, Journal of insect behavior 3 (2)
(1990) 159–168.
[40] R. Bhattacharyya, Possibilistic sharpe ratio based novice portfolio selec-
tion models, in: National Conference on Advancement of Computing in
Engineering Research (ACER 13) Krishnagar, West Bengal, India, 2013,
pp. 33–45.

23