A Portfolio Optimization Algorithm Using Fuzzy Granularity Based Clustering

S. M. Aqil Burney
Institute of Business Management
Korangi Creek, Karachi, Karachi City, Sindh 75190, Pakistan
Phone: +92 21 111 002 004
[email protected]

Tahseen Jilani
University of Nottingham
Nottingham NG7 2RD, UK
Phone: +44 115 951 5151
University of Karachi
Main University Rd, Karachi, Karachi City, Sindh 75270, Pakistan
Phone: +92 21 99261300
[email protected]

Humera Tariq
University of Karachi
Main University Rd, Karachi, Karachi City, Sindh 75270, Pakistan
Phone: +92 21 99261300
[email protected]

Zeeshan Asim
Virtual University, Pakistan
M. A. Jinnah Campus, Defence Road, Off Raiwand Rd, Lda Avenue Phase 1 Lda Avenue, Lahore, Punjab,
Pakistan
Phone: +92 42 111 880 880
[email protected]

Usman Amjad
University of Karachi
Main University Rd, Karachi, Karachi City, Sindh 75270, Pakistan
Phone: +92 21 99261300
[email protected]

Syed Shah Mohammad

University of Karachi
Main University Rd, Karachi, Karachi City, Sindh 75270, Pakistan
Phone: +92 21 99261300
[email protected]

Abstract
Clustering algorithms are applied to numerous problems in multiple domains including historic
data analysis, financial markets analysis for portfolio optimization and image processing. Recent years
have witnessed a surge in use of nature inspired computing (NIC) techniques for data clustering to solve
various real world optimization problems. Granular Computing (GC) is an emerging technique to
handle pieces of information, known as information granules. In this paper, an ensemble of fuzzy
clustering using Particle Swarm Optimization and Granular computing for stock market portfolio
optimization. The model is then tested on stocks listed in Hong Kong Stock Exchange. Experimental
results suggested that clusters formed through Fuzzy Particle Swarm Optimization (FPSO) with
Granular computing are well suited and efficient for portfolio optimization. For comparison, we have
used a benchmark index of Hong Kong Stock Exchange called as Hang Sang Composite Index (HSCI).
Results proved that results of proposed approach are better in comparison to benchmark results of HSCI.
Keywords: Hybrid Approach for Portfolio Selection; Fuzzy C-mean Clustering (FCM); Fuzzy
Particle Swarm Optimisation (FPSO); Granular Computing; Hong Kong Composite Index.

159
BRAIN – Broad Research in Artificial Intelligence and Neuroscience
Volume 10, Issue 2 (April, 2019), ISSN 2067-3957

1. Introduction
Data Clustering is the mathematical method designed to identify relevant data within a
collection of data (Nerurkar et. al., 2018). It can be described as a methodology for assignment of data
into groups in a manner that the data points in same group or cluster are analogous to each other and
unrelated with objects of other clusters or groups (Hammouda and Karray, 2000). It is being used
efficiently in various domains for identification of natural groups present in large datasets. Data
Clustering can be used by businesses for identification of potential customers of a product by analyzing
or collecting buying patterns of customers, so as to design marketing strategies based on those behaviors
(Ravi, Pradeepkumar and Deb, 2017). Finding out clusters in a large dataset is challenging task and
usually require some data mining tool. Clustering tools usually assign data elements to a clusters based
on their similarities to the group.
Clustering remained an area of interest for researchers in last few decades, thus various
clustering techniques were developed. Clustering techniques can be generally divided into two types
(Suganya and Shanthi, 2011). Based on classical set theory there is a type of clustering called hard
clustering algorithms in which data items can only be assigned to only one group at a point of time. A
widely used hard or crisp clustering algorithm is k-means. But for real datasets where there are no
definite boundaries, this technique is not useful. (Izakian and Abraham, 2009).
Soon after the introduction of the fuzzy theory, the researchers applied fuzzy set theory on
clustering algorithms (Izakian and Abraham, 2009). There is no sharp boundary in real world data, so
Fuzzy clustering algorithms remained fruitful in those applications. It can handle real world
uncertainties efficiently by assigning membership degree to items. Membership degree in such clusters
relies on the proximity of values to the cluster centers. Widely used and famous fuzzy clustering
algorithm is Fuzzy C-Means (FCM) introduced by Bezdek in 1974 and is being applied at large
(Bezdek, 1984).
Swarm intelligence (SI) is an area of computational intelligence which comprise of algorithms
getting inspiration from population based natural phenomenon working on the basis of decentralized
control and self-organization (Shandilya et. al., 2017). It can be said that SI is “collective behavior of
decentralized and self-organized systems” (Zhang et. al., 2013). On the other hand Granular Computing
(GC) is a computation theory for efficiently using granules such as clusters, groups and subsets to build
a computational model for complicated applications that contains huge amounts of data and
information. A granule can be described as one of the various small data points or particles combining
to form a larger unit.
In this paper we have used Fuzzy Particle Swarm Optimization (FPSO) using the concept of
granular computing to divide the information granules into different clusters to build a portfolio that can
optimize the weekly investor’s returns. The experimental results using the Hong Kong Stock Exchange
data indicate that our proposed method provides better returns than the benchmark index for the Hong
Kong Stock exchange.

2. Literature Review
2.1. Fuzzy Data Clustering
Fuzzy logic concepts are based on degree of membership so imprecision concepts are dealt
with fuzzy logic in better way. Fuzzy logic can be used in data clustering so as to deal with partial
membership of data points. Fuzzy logic based data clustering algorithms assign data object partly to
more than one cluster. FCM proposed by Bezdek (Bezdek, 1984), divides the collection of n data
objects denoted as o = {o1,o2 ,...,on} in R dimensional space into c fuzzy clusters, where (1 < c < n)
with centroids or cluster centers Z = {z1, z2,..., zc}. Fuzzy clustering can be represented by using
fuzzy matrix μ with dimensions n x c. Here n is count of data objects whereas c is count of data
clusters. Data item present at ith row and jth column is represented by μij. Degree of membership of
ith and jth object is represented by μ. Degree of membership μ has following properties:

[ ] (1)

160
S. M. A. Burney, T. Jilani, H. Tariq, Z. Asim, U. Amjad, S. S. Mohammad - A Portfolio Optimization Algorithm Using Fuzzy
Granularity Based Clustering

∑ (2)

∑ (3)

Fuzzy C-Means has the objective function to minimize the following equation:

∑ ∑ (4)

s.t,

| | (5)

where m; (m > 1) is a scalar constant value called as “weighting exponent”, which manages the
fuzziness of clusters whereas dij is Euclidian distance between object oi and the cluster center zj.
Where zj indicates cluster center of jth cluster and it is obtained using equation (6)

i1 ijmoi
n
z j (6)
i1 ijm
n

FCM is an iterative algorithm and described in below steps:

i) Select the weighting component m where (m>1) and initiate μij, the membership function
values where , i = 1,2,..., n; j = 1,2,..., c .
ii) Using above mentioned Eq. 6 find out the cluster centers zj , where j = 1,2,..., c.
iii) Calculate the Euclidian distance dij, where i = 1,2,..., n; j = 1,2,..., c
iv) Using below mentioned Eq. 7 update μij, the membership function, where i = 1,2,..., n; j
= 1,2,..., c.

ij  1 2
(7)

c  dij  ( m 1)
 k 1 dik 

v) If not converged, go to step 2.

There are many conditions that can be used to stop the execution of this loop. One of them is
to stop iterations of the algorithm when the change in the cluster center values becomes negligible
or the objective function as specified in the equation (4), cannot be minimized more. One problem
of FCM algorithm is that it is very much dependent on initial values and likely to fall in local
optima problem.

2.2. Fuzzy Particle Swarm Optimization

(Peng et al., 2004) suggested a variant of PSO based on fuzzy logic, for Travelling Salesman
Problem (TSP), known as “Fuzzy particle swarm optimization (FPSO)”. In this algorithm the
velocity and position of individuals are re-defined to characterize the fuzzy relation within
variables. In Fuzzy PSO algorithm, X represents fuzzy relationship between collection of data
objects, o = {o1,o2,..., on} , and the collection of cluster centers, Z = {z1, z2,..., zc} . Fuzzy
relationship X is represented as:

161
BRAIN – Broad Research in Artificial Intelligence and Neuroscience
Volume 10, Issue 2 (April, 2019), ISSN 2067-3957

 11 1c 
 
X   (8)
 nc 
 n1

In the above mentioned matrix X, μij denotes the membership value of object i to cluster j
with constraints specified in the eq. 9 and 10

[ ] (9)

∑ (10)

The position matrix specified in the above mentioned equation of each individual is similar
to fuzzy matrix μ specified in Fuzzy C-Means algorithm. Velocity of each individual is specified by
a matrix of dimension [n, c] where n denotes number of rows and c is the number of columns.
Elements of the matrix are within the range of [-1, 1]. Eq. 11 and 12 are used for changing the
velocities and positions of every particle on the basis of matrix operations.

V (t  1)  w  V (t )  (c1r1)  (pbest(t) X(t))  (c 2 r 2)  (gbest(t) X(t)) (11)

X (t  1)  X (t )  V (t  1) (12)

Here  denotes the matrix addition and  represents the matrix multiplication. It is
important to note here that constraints stated in eq. 9 and 10 may be violated after update of position
matrix. Thus normalizing position matrix is necessary here. For normalization purpose, all the
negative values in matrix are made zero. And if all matrix elements turn out to be zero then the
matrix is evaluated again using random numbers within range of [0, 1] and then matrix is
transformed without violating the conditions.

   c  1c   1 j 
c

 11 j 1 1 j j 1
 
Xnormal    (13)
 c c 
  n1    nj  nc    nj 
 j 1 j 1 

Similar to other nature inspired algorithms Fuzzy PSO algorithm uses a fitness function for
assessing the general solution. Following equation will be used for evaluation of the solutions.

K
f (X )  (14)
Jm
In equation 14, K is a constant while Jm is objective function for Fuzzy C-Means algorithm given in
eq. 15.
c n
Jm    ijmdij (15)
j 1 i1

As the value of Jm is smaller, clustering results will be better fitness value f(X) being higher. Fuzzy
PSO algorithm for fuzzy clustering is described as under:

162
S. M. A. Burney, T. Jilani, H. Tariq, Z. Asim, U. Amjad, S. S. Mohammad - A Portfolio Optimization Algorithm Using Fuzzy
Granularity Based Clustering

1. Instantiate the following parameters: P the population size, w, c1, c2 and maximum number of
iterations.
2. Initialize a swarm with P number of individuals. Here X, gbest, pbest and V are matrices of size
n, c.
3. Instantiate X, V and pbest values for every individual and gbest for the whole population.
4. Determine the centers of cluster for every individual using eq. 16.

i1 ijmoi
n
z j
i1 ijm
n (16)

5. Evaluate the objective value of every individual using eq. 15.

6. Evaluate the value of pbest for every individual.
7. Evaluate the value of gbest for the whole population.
8. Change velocity matrix for every individual by using eq. 11 and 12.
9. Change position matrix for every individual by using eq. 13.
10. Go back to step 4 until stopping criteria is met.
Stopping criteria is either the predefined maximum number of iterations or no progress in
global best fitness for a specified number of generations.
In this paper we have used Fuzzy PSO clustering algorithm on the granules created from the
original dataset. To divide the dataset into granules on the basis of Market Capital Value of the
company, the companies with similar market value are placed into same group. For the clustering of
these granules Fuzzy PSO algorithm is used which gives the benefit of lower computational time
and offer better results than Fuzzy C-Means clustering algorithm (Mehdizadeh, 2009).

3. Methodology Used
From the literature review it is revealed that lot of work was done for data clustering and
portfolio management but not much work is done on clustering stock data for the portfolio
optimization using Fuzzy PSO. The use of data clustering for stock data helps in segmenting
different stocks in a way that all stocks having similar characteristics are grouped together (Cheng,
Chen and Jian, 2015). A method for creating efficient portfolios with Markowitz model by using the
clustering method to select stocks, called clustering based selection was designed by (Nanda,
Mahanty and Tiwari, 2010). To classify stocks into clusters they used Fuzzy C-Mean data clustering
algorithm. After classification of stocks, some stocks were selected from clusters for building an
optimized portfolio to minimize the risk by diversifying the portfolio. According to them, the
problem of efficient frontier can be solved more efficiently by clustering the stocks. Although fuzzy
c-means (FCM) algorithm is considered one of the most popular and widely used fuzzy clustering
techniques because of its efficiency, straightforwardness, and convenience of implementation. But
problem is that fuzzy c-means is very sensitive to initialization and it can easily be trapped in local
optima. On the other hand Particle swarm optimization (PSO) is a stochastic global optimization
tool which is used for various optimization problems. (Li, Liu and Xu, 2007) proposed a fuzzy PSO
based data clustering algorithm to overcome the shortcomings of FCM. Their suggested method
uses the power of global search in PSO algorithm to overcome the shortcomings of Fuzzy C-Means.
In our methodology the focus is on how stocks can be divided into granules and how Fuzzy
PSO based data clustering algorithm can be applied on these granules to further divide data into
small clusters and how to design a diversified portfolio using stocks from different cluster to
maximize portfolio returns. So Fuzzy PSO algorithm is applied to create clusters for each granule.
The dataset used for the experiment contains the information about financial ratios of companies
listed in Hong Kong Stock Exchange. This dataset is divided into six different sub groups known as

163
BRAIN – Broad Research in Artificial Intelligence and Neuroscience
Volume 10, Issue 2 (April, 2019), ISSN 2067-3957

information granule. Information granules are collections of entities that are arranged together due
to their similarity, functional or physical adjacency, coherence etc. A granulation criterion deals
with the question of why two objects are put into the same granule. We divided the dataset into 6
different partitions or granules based on their market capitalization value.
This is followed by calculating the optimal number of clusters in each group. Then Fuzzy
PSO data clustering algorithm is applied on each granule to divide it into optimal number of clusters
as calculated in the previous step. After that 1 to 3 stocks are selected from every cluster according
to the important fields as indicated in the Principal Component Analysis. Then average weekly
return of each stock selected is calculated in the last step on the basis of their market value during
January- 2012 to June-2012. The stocks having good positive average weekly returns are selected
for the portfolio creation.
Finally, Variance – Covariance matrix for the selected stocks is calculated in the last step.
MATLAB is used for the development of efficient portfolio against the efficient frontier. For this
purpose, we have used the MATLAB financial tool box command frontcon. This command returns
optimized portfolios as per the provided input parameters. We have taken 3 portfolios from the
given set of portfolios and calculated the actual weekly portfolio returns for each portfolio on the
basis of market value during July 2012 to December 2012. Then we calculated the Hang Seng
Composite Index weekly performance from July 2012 to December 2012 from the Bloomberg
website. Hang Seng Composite Index is benchmark index for Hong Kong Stock exchange. HSCI is
a comprehensive benchmark index and covers about 95% of total listed companies on main board
of stock exchange of Hong Kong (“SEHK”). HSCI is used as a basis for performance benchmarks.
In next step we compared these portfolio results against the HSCI and the comparison
showed that these portfolio returns are better than the HSCI. Flow chart of proposed model is
shown in Figure 1.

4. Data Description
Dataset of the Hong Kong Stock Market companies’ data for the financial year 2011 was
taken from the New York University Dataset page. In data preprocessing step, it was checked for
missing values by removing the instances with missing data from dataset. There are 774 companies’
data present in the dataset after the removal of missing values. This dataset contains companies’
data from 77 different industry groups and represents almost all industry groups of the Hong Kong
Stock Exchange. There are 42 fields for each company which includes many different types of
financial ratios to represent the financial position of that company at the end year 2011. Some of the
financial ratios include Market Capital (in US$), Total Debt (in US$), Firm Value (in US$), Cash,
Enterprise Value (in US$), Cash Firm Value, Liquidity Ratio, Book Debt to Capital Ratio, Market
Debt to Capital Ratio, Book Debt to Equity Ratio, Market Debt to Equity Ratio, Beta, Correlation
with Market, PBV, PS, Return on Equity and Return on Capital etc.
In this dataset variety of data values are used. Some fields contain very large values like
Market Value, Enterprise Value, Market Capitalization and some very small fields like Beta, Debit
to Equity Ratio. Data is first pre-processed to deal with large values. Initially, the data was
transformed to z- scores to get similar variability of the values.
Another problem is that there are 42 fields in the dataset which are difficult to handle for
calculations while performing data clustering; therefore we used Principal Component Analysis
(PCA) for this purpose. PCA uses a mathematical procedure to transforms a number of correlated
variables into a smaller number of un-correlated variables called principal components. This
process is also known as Dimension Reduction. The transformation in PCA is done in such a way
that the 1st principal component has the largest possible variance, and each succeeding component
in turn has the highest variance possible under the constraint that it will be uncorrelated with the
preceding components. Before performing PCA the data must be standardized to remove the
influence of different measurement scales and to give approximately equal weightage to all the
values. We have used SPSS tool for this purpose.

164
S. M. A. Burney, T. Jilani, H. Tariq, Z. Asim, U. Amjad, S. S. Mohammad - A Portfolio Optimization Algorithm Using Fuzzy
Granularity Based Clustering

Figure 1. Flow chart of proposed model for portfolio optimization

After performing PCA on our data, 12 principle components cover 94.7% variation of the
full dataset with 42 variables. The identified fields sorted by respective Eigen values are,
1. Firm Value (in US$),
2. Book Debt to Capital Ratio,
3. Price to Sale Ratio (PS),
4. Free Cash Flow to Firm (FCFF),
5. Beta (A measure of the volatility of a portfolio in comparison to the market as a whole),
6. Liquidity Ratio,
7. Correlation with Market,
8. Return on Capital,
9. Net Profit Margin,
10. Net Debt Issued,
11. Cash Firm Value,
12. EV Invested Capital.

165
BRAIN – Broad Research in Artificial Intelligence and Neuroscience
Volume 10, Issue 2 (April, 2019), ISSN 2067-3957

First four fields represent around 62% of the dataset variation as per the PCA so we will use
these four fields for selection of stocks from different clusters.

5. Experimental Results
5.1. Granules Formation
The dataset was divided into six different sub groups also known as information granules.
Information granules are a group of objects that are organized together based on their similarity,
coherence or physical adjacency. A granulation criterion describes the rules for dividing data
objects into different granules. The categorization of companies into different partitions is made
based on market capitalization. The companies were divided into groups namely Mega, Large, Mid,
Small, Micro and Nano. There is no formal definition of the exact cutoff values. Therefore,
following market capitalization values as granule criterion are used:

1. Mega Companies: Over $10000 Million 2. Large Companies: $5000 $10000 Million
3. Mid Companies: $1000 $5000 Million 4. Small Companies: $250 $1000 Million
5. Micro Companies: Below $250 Million 6. Nano Companies: Below $50 Million

After the granules formation the number of companies in each granule is shown in the table 1:

Table 1. Granules Frequency Table

Granule Name Frequency Percent Cumulative Percent
Nano 182 23.5 23.5
Micro 312 40.3 63.8
Small 151 19.5 83.3
Medium 95 12.3 95.6
Large 17 2.2 97.8
Mega 17 2.2 100

Total 774 100

5.2. Granules Formation Optimal Number of Clusters Estimation in Each Granule

Number of records of companies’ data in each granule is different, with wide range of
values so there would be different number of clusters in each granule. To divide each granule into
clusters we have used K-means data clustering algorithm and to identify optimal number of clusters
Internal Validity Indices. For the cluster estimation we have used available tool for this purpose
designed by Mr. Kaijun Wang MATLAB code:
(http://www.mathworks.com/matlabcentral/fileexchange/13916). This tool is designed in MATLAB
release 7.2 (R2006a). The tool is suitable for the performance comparison of different indices on the
estimation of the number of clusters, algorithm design for applications by using or improving part
codes, etc. This tool also provides internal validity indices which we have used to estimate the
optimal number of clusters between 2 and 14. Following internal validity indices are used in this
method for the cluster estimation.
 Silhouette (SIL)
 Davies-Bouldin (DB)
 Calinski-Harabasz (CH)
 Krzanowski-Lai (KL)

The results of optimal number of clusters estimation for each granule using above
mentioned tool are described in table 2:

166
S. M. A. Burney, T. Jilani, H. Tariq, Z. Asim, U. Amjad, S. S. Mohammad - A Portfolio Optimization Algorithm Using Fuzzy
Granularity Based Clustering

Table 2. Optimal number of Clusters or each Granule Table

Granule Davies-Bouldin Calinski- Krzanowski-Lai (KL) Optimal no. of
Silhouette (Sil)
(DB) Harabasz (CH) Cluster
Nano 6 6 6 8 6
Micro 5 5 5 7 5
Small 2 2 2 2 2
Medium 3 3 3 3 3
Large 2 2 2 2 2
Mega 2 2 2 - 2

5.3. Clustering Granules using Fuzzy PSO

After the optimal numbers of cluster estimation for every granule we have performed data
clustering using Fuzzy Particle Swarm Optimization algorithm. The code for this data clustering
technique is written in MATLAB 2010 using the Fuzzy PSO algorithm. The code is run for 300
times for each granule and the best clustering result are used for further processing. To evaluate the
best clustering result objective function value and Internal validity indices are used. Following
internal validity indices are used for the evaluation of clustering results.
 Partition Coefficient (PC)
 Classification Entropy (CE)
 Partition Index (SC)
 Separation Index (S)
 Xie and Beni’s index (XB)
5.4. Selection of Companies from Each Cluster
After applying the Fuzzy PSO (FPSO) data clustering algorithm on all granules we have
developed clusters in each granules. In our next step we selected 1-3 companies from each cluster
based on the performance of financial ratios during the year 2011. The ratios used for the selection
of companies to build a portfolio are those fields which are identified in the principal component
analysis. The fields & ratios used for selection of companies are Firm Value in US$, Book Debt to
Capital Ratio, PS (Price Sale), FCFF (Free Cash Flow to Firm), Beta, Liquidity Ratio, Correlation
with Market, Return on Capital , Net Profit Margin, Net Debt Issued/Repaid, Cash Firm Value and
EV Invested Capital. As a result of this process 33 companies were selected for further processing.
The detail of number of companies selected from each granule is given in Table 3:

Table 3. No of companies selected from each granule

Granule Name No of Companies Selected Total no of Companies No of Clusters
Nano 4 182 6
Micro 6 312 5
Small 5 151 2
Medium 8 95 3
Large 5 17 2
Mega 5 17 2
Total 33 774 20

The number of companies selected for hybrid optimal portfolio from each cluster of every granule
is given in Table 4:
Table 4. No of companies selected from each cluster
Cluster No. 1 2 3 4 5 6
Granule Name
Nano 1 1 0 0 1 1
Micro 1 1 1 1 2 --
Small 3 2 -- -- -- --
Medium 1 4 3 -- -- --
Large 3 2 -- -- -- --
Mega 3 2 -- -- -- --

167
 BRAIN – Broad Research in Artificial Intelligence and Neuroscience
Volume 10, Issue 2 (April, 2019), ISSN 2067-3957

5.5. Selection of Companies for Portfolio Management

For the selection of companies for the portfolio management we calculated the average weekly
earnings of above mentioned companies for the period of January 2012 to June 2012 (26 weeks).
For the calculation of the average weekly earnings we have downloaded the company historical
share prices from yahoo finance. For further processing we selected only those companies that have
weekly earnings of 0.08%. There were 15 companies out of 33 that have weekly earning greater
than or equal to 0.08%. The names of those companies are as under:
Table 5. List of companies selected for portfolio creation
Sr # Company Name Industry Group Granule Cluster
1 VS International Group Ltd. (SEHK:1002) Machinery Nano 2
2 Chun Wo Development Holdings Ltd. (SEHK:711) Engineering Nano 5
3 Huafeng Group Holdings Limited (SEHK:364) Apparel Nano 6
4 National Electronics Holdings Ltd. (SEHK:213) Apparel Micro 2
5 Hon Kwok Land Investment Co. Ltd. (SEHK:160) Real Estate (Development) Micro 3
6 Convenience Retail Asia Ltd. (SEHK:831) Retail (Grocery and Food) Small 2
7 Shimao Property Holdings Ltd. (SEHK:813) Real Estate (Development) Medium 2
8 Kerry Properties Ltd. (SEHK:683) Real Estate Medium 2
9 Cafe de Coral Holdings Ltd. (SEHK:341) Restaurant Medium 2
10 Franshion Properties (China) Ltd. (SEHK:817) Real Estate Medium 3
11 Haier Electronics Group Co., Ltd. (SEHK:1169) Furn./Home Furnishings Medium 3
12 Galaxy Entertainment Group Limited (SEHK:27) Hotel/Gaming Large 1
13 China Resources Land Ltd. (SEHK:1109) Real Estate (Development) Large 2
14 China Overseas Land & Investment Ltd. (SEHK:688) Real Estate (Development) Mega 1
15 BOC Hong Kong Holdings Ltd. (SEHK:2388) Bank Mega 2

5.6. Design Portfolios using the FrontCon

To design different portfolios, we have used frontcon function from MATLAB’s financial
toolbox. This function returns the mean-variance efficient frontier with user specified covariance
and returns. FrontCon is a MATLAB 2010 function that helps us to design portfolios of asset
investment weights which minimize the risk for given values of the expected return. The portfolio
risk is minimized subject to constraints on the asset weights or on groups of asset weights. To use
the frontcon we need a Variance-Co Variance matrix of given companies, expected return and
number of portfolios required to be designed. We have calculated the Var-Co Variance matrix using
the average weekly earnings calculated before, for the expected returns we have used the average
weekly earnings of these companies between January 2012 and June 2012. To calculate the average
weekly earnings we downloaded historical share price of these companies from the yahoo finance.
The frontcon function returns the portfolios with specified number of shares of each company. The
frontcon also returns the estimated portfolio returns and the risk associated with that portfolio. The
syntax of frontcon is as under:

[PortRisk, PortReturn, PortWts] = frontcon (ExpReturn, ExpCovariance, NumPorts,

PortReturn, AssetBounds, Groups, GroupBounds, varargin)

We have used this function for 20 portfolios. The portfolio 1 gives a portfolio weekly return
of 1.1127% at the risk of 1.3694% and the portfolio comprised of 8 companies. This portfolio has
the lowest risk and the risk is lowered by diversifying investment in 8 companies. The portfolio 20
gives a portfolio return at the risk of 9.0925% and the portfolio comprised of only one company.
This portfolio gives highest return but also contains the highest risk. The efficient frontier for our
given values is as under. The associated return and risk of each portfolio is described in Table 6:

168
S. M. A. Burney, T. Jilani, H. Tariq, Z. Asim, U. Amjad, S. S. Mohammad - A Portfolio Optimization Algorithm Using Fuzzy
Granularity Based Clustering

Table 6. Risk associated with each portfolio

Portfolio No Portfolio Return Portfolio Risk
1 1.1127% 1.3694%
2 1.1903% 1.4578%
3 1.2678% 1.6697%
4 1.3454% 1.9324%
5 1.4230% 2.2264%
6 1.5005% 2.5407%
7 1.5781% 2.8711%
8 1.6556% 3.2147%
9 1.7332% 3.5679%
10 1.8108% 3.9389%
11 1.8883% 4.3340%
12 1.9659% 4.7513%
13 2.0434% 5.1856%
14 2.1210% 5.6427%
15 2.1986% 6.1287%
16 2.2761% 6.6375%
17 2.3537% 7.1710%
18 2.4312% 7.7630%
19 2.5088% 8.4072%
20 2.5864% 9.0925%

Figure 2. Risk and expected return of portfolio

A portfolio that offers maximum expected return for a given level of risk, or conversely the
lowest level of risk for a given expected return is known as optimal portfolio. Efficient frontier is a set of
optimal portfolios that suggests highest expected return for a defined level of risk or in other words
lowest risk for a specified level of expected return. So we can say that the set of all efficient portfolios is
called the efficient frontier, shown in graph presented in Figure 2.
5.7. Portfolio Performance
To assess the efficacy of our portfolios we measured the actual weekly performance of these
stocks from July 2012 to Dec 2012 and compared it with the standard index of the Hong Kong Stock
exchange for the same duration. For this we have used the Hang Seng Composite Index (HSCI) which
is one of benchmark index for the Hong Kong Stock Exchange. The Hang Seng Composite Index
(HSCI) offers a comprehensive Hong Kong market benchmark that covers about 95% of the total
market capitalization of companies listed on the main board of the stock exchange of Hong Kong
(SEHK). HSCI uses free float adjusted market capitalization methodology, and can be used as a basis
for performance benchmarks. So to compare the portfolio performance weekly performance of HSCI is
calculated and compared with the portfolio performance. Top three portfolios having least risk for the
invested capital are used. The portfolio details of 3 portfolios formed are given in Tables 7, 8 and 9:

169
BRAIN – Broad Research in Artificial Intelligence and Neuroscience
Volume 10, Issue 2 (April, 2019), ISSN 2067-3957

Table 7. Portfolio number 1 composition

Sr # Company Name Weight Granule Cluster Membership
1 Cr Asia - (831) 0.22186 Small 2
2 Franshion Ppt - (817) 0.00418 Medium 3
3 Haier Elec - (1169) 0.15436 Medium 3
4 Galaxy Ent - (27) 0.05250 Large 1
5 China Res Land - (1109) 0.02305 Large 2
6 National Elec H - (213) 0.16128 Micro 2
7 Hon Kwok Land - (160) 0.38278 Micro 3

Table 8. Portfolio number 2 composition

Sr # Company Name Weight Granule Cluster Membership
1 V.S. Intl - (1002) 0.01817 Nano 2
2 Cr Asia - (831) 0.24439 Small 2
3 Franshion Ppt - (817) 0.06214 Medium 3
4 Haier Elec - (1169) 0.15718 Medium 3
5 Galaxy Ent - (27) 0.03945 Large 1
6 National Elec H - (213) 0.14326 Micro 2
7 Hon Kwok Land - (160) 0.33540 Micro 3

Table 9. Portfolio number 3 composition

Sr # Company Name Weight Granule Cluster Membership
1 V.S. Intl - (1002) 0.040 Nano 2
2 Cr Asia - (831) 0.263 Small 2
3 Franshion Ppt - (817) 0.114 Medium 3
4 Haier Elec - (1169) 0.156 Medium 3
5 Galaxy Ent - (27) 0.024 Large 1
6 National Elec H - (213) 0.115 Micro 2
7 Hon Kwok Land - (160) 0.288 Micro 3

The actual weekly performance of our portfolios and the benchmark index for the July 2012 to
December 2012 is given in Table 10:

Table 10. Comparison between Benchmark Index & Our Portfolios

Week No HSCI Portfolio # 1 Portfolio # 2 Portfolio # 3
Week # 1 Return 0.83% 0.26% 0.25% 0.27%
Week # 2 Return -2.73% -1.08% -1.24% -1.48%
Week # 3 Return 1.09% 0.01% 0.17% 0.26%
Week # 4 Return -0.25% 0.23% 0.49% 0.82%
Week # 5 Return 0.85% 0.62% 0.48% 0.22%
Week # 6 Return 0.79% 3.06% 2.46% 1.86%
Week # 7 Return -0.60% -1.84% -2.02% -2.22%
Week # 8 Return -0.65% 2.65% 2.63% 2.58%
Week # 9 Return -2.58% -1.42% -1.36% -1.40%
Week # 10 Return 1.90% 2.77% 2.62% 2.61%
Week # 11 Return 3.94% 3.15% 2.79% 2.59%
Week # 12 Return 0.52% 3.33% 2.66% 1.84%
Week # 13 Return 0.75% 1.06% 1.01% 1.05%
Week # 14 Return 0.60% 0.08% 0.30% 0.58%
Week # 15 Return 0.79% -0.84% -0.58% -0.33%
Week # 16 Return 2.26% 1.41% 1.28% 1.17%
Week # 17 Return 0.12% 1.25% 1.18% 1.08%
Week # 18 Return 3.54% 3.73% 3.38% 3.17%
Week # 19 Return -2.62% -0.95% -0.53% -0.11%
Week # 20 Return -1.15% -0.31% -0.41% -0.44%
Week # 21 Return 3.15% 2.80% 2.79% 2.91%
Week # 22 Return 0.39% 3.21% 4.06% 4.94%
Week # 23 Return 0.91% 1.79% 2.03% 2.20%

170
S. M. A. Burney, T. Jilani, H. Tariq, Z. Asim, U. Amjad, S. S. Mohammad - A Portfolio Optimization Algorithm Using Fuzzy
Granularity Based Clustering

The performance of our three portfolios and the HSCI benchmark index from July 2012 to
December 2012 is shown above. The graphical view of weekly performance of our portfolios and
benchmark index are shown in Figure 3:

Figure 3. Performance comparison of three portfolios

The graph of Figure 3 clearly suggests that the performance of our three portfolios is greater
than the Hong Kong Stock Exchange benchmark index HSCI. Another important point to note here is
that in all the three portfolios, stocks belongs to different set of Granule and clusters. As a first step we
divided the stocks into 5 different granules based on their market capitalization value and in the next
step the granules are further sub divided into clusters. As we know that cluster helps in grouping similar
records so we can say that clustering will divide stocks into similar groups. Therefore, selecting stocks
from different groups will diversify our portfolio and as a result it will also reduce our risk, as we know
that diversification reduces the risk. In the Portfolio 1 there are two stocks from Micro granule. First
stock belongs to cluster 2 and second belong to cluster 3. There is one stock from the Small granule that
belongs to cluster 2. There are two stocks from the Medium granule and both of them belong to cluster 3
of that granule. There are two stocks from the large granule that belong to cluster 1 and 2 of that
granule. In the same way the stocks are also diversified from different granules and clusters in the
portfolio 2 and 3 which helps in reduction of portfolio risk.
6. Conclusion
In this research a method for portfolio management by using granule based fuzzy data
clustering is proposed. Granular Computing is integrated with Fuzzy based Particle Swarm
Optimization technique to design portfolios that can offer returns matching the benchmark index.
This method is also useful in selecting the stocks for the investment as initially there were 774
companies’ data from 77 different industry groups of the Hong Kong Stock Exchange. There are 42
fields for every company present in the dataset which includes many different types of financial
ratios to represent the financial position of these companies at the year-end 2011. This method
reduces time for the selection of stocks from different categories and also helps in convenient
grouping of stocks into a cluster and thus best performing stocks from those groups can be selected.
The selection of stocks from different granules and then from different clusters will diversify the
portfolio and as a result our portfolio risk will be reduced. The aim was to maximize return by
investing in different groups of stocks that would individually react in a different ways for the same
event. An important point to note here is that although diversification does not guarantee against
loss, but diversification can be used as an important factor for maximizing returns while minimizing
risk. We can reduce risk associated with a stock, but general market risks influence almost every
stock. That’s why for building an optimal portfolio we selected stocks from different granules and
clusters so that each stock has its own characteristics and would react differently for the same event.
To design optimal portfolios, MATLAB function frontcon from financial toolbox is used.
This helps us to design portfolios that maximize the return for the given value of risk. The results
of designed portfolios are better than the benchmark index of the Hong Kong Stock Exchange
which further validates the view. To summarize this work, it can be said that we have demonstrated
171
BRAIN – Broad Research in Artificial Intelligence and Neuroscience
Volume 10, Issue 2 (April, 2019), ISSN 2067-3957

a granule based FPSO data clustering approach for the selection of stocks, portfolio management
and designing portfolios on the efficient frontier.

References
Bezdek, J. C., Ehrlich, R., & Full, W. (1984). FCM: The fuzzy c-means clustering algorithm.
Computers & Geosciences, 10(2-3), 191-203. doi:10.1016/0098-3004(84)90020-7.
Cheng, S., Chen, S., & Jian, W. (2015). A Novel Fuzzy Time Series Forecasting Method Based on
Fuzzy Logical Relationships and Similarity Measures. IEEE International Conference on
Systems, Man, and Cybernetics, doi:10.1109/smc.2015.393.
Hammouda, K., & Karray, F. (2000). A Comparative Study of Data Clustering Algorithms.
Retrieved from http://www.pami.uwaterloo.ca/pub/hammouda/sde625-paper.pdf.
Izakian, H., Abraham, A., & Snasel, V. (2009). Fuzzy clustering using hybrid fuzzy c-means and
fuzzy particle swarm optimization. World Congress on Nature & Biologically Inspired
Computing (NaBIC). doi:10.1109/nabic.2009.5393618.
Li, L., Liu, X., & Xu, M. (2007). A Novel Fuzzy Clustering Based on Particle Swarm Optimization.
IEEE International Symposium on Information Technologies and Applications in Education.
doi:10.1109/isitae.2007.4409243.
Li, X. Y., Sun, J. X., Gao, G. H., & Fu, J. H. (2011). Research of Hierarchical Clustering Based on
Dynamic Granular Computing. JOURNAL OF COMPUTERS, 6(12), 2526-2533.
Maciel, L., Gomide, F., & Ballini, R. (2013). Forecasting Exchange Rates with Fuzzy Granular
Evolving Modeling for Trading Strategies. Proceedings of the 8th conference of the
European Society for Fuzzy Logic and Technology. doi:10.2991/eusflat.2013.40
Nanda, S., Mahanty, B., & Tiwari, M. (2010). Clustering Indian stock market data for portfolio
management. Expert Systems with Applications, 37(12), 8793-8798.
doi:10.1016/j.eswa.2010.06.026.
Nerurkar, P., Shirke, A., Chandane, M., & Bhirud, S. (2018). Empirical Analysis of Data Clustering
Algorithms. Procedia Computer Science,125, 770-779. doi:10.1016/j.procs.2017.12.099
Östermark, R. (1996). A fuzzy control model (FCM) for dynamic portfolio management. Fuzzy
Sets and Systems. doi:10.1016/0165-0114(96)84605-7.
Rajagopal, S. (2011). Customer Data Clustering Using Data Mining Technique. International
Journal of Database Management Systems, AIRCC Publishing Corporation, 3(4).
Ravi, V., Pradeepkumar, D., & Deb, K. (2017). Financial time series prediction using hybrids of
chaos theory, multi-layer perceptron and multi-objective evolutionary algorithms. Swarm
and Evolutionary Computation, 36, 136-149. doi:10.1016/j.swevo.2017.05.003.
Shandilya, S. K., Shandilya, S., Deep, K., & Nagar, A. K. (2017). Handbook of research on soft
computing and nature-inspired algorithms. Hershey, PA: Information Science Reference.
Shi, Y., & Eberhart, R. (1998). A modified particle swarm optimizer. IEEE International
Conference on Evolutionary Computation Proceedings. IEEE World Congress on
Computational Intelligence (Cat. No.98TH8360). doi:10.1109/icec.1998.699146.
Suganya, R., & Shanthi, R. (2012). Fuzzy C- Means Algorithm- A Review. International Journal of
Scientific and Research Publications, IJSRP Inc, 2(11).
Zhang, Y., Agarwal, P., Bhatnagar, V., Balochian, S., & Yan, J. (2013). Swarm Intelligence and Its
Applications. The Scientific World Journal, 2013, 3.
Zhu, Q., & Azar, A. (Eds.). (2015). Complex System Modelling and Control Through Intelligent
Soft Computations (Vol. 319). Springer, Cham. doi: https://doi.org/10.1007/978-3-319-
12883-2.
Zhu, H., Wang, Y., Wang, K., & Chen, Y. (2011). Particle Swarm Optimization (PSO) for the
constrained portfolio optimization problem. Expert Systems with Applications, 38(8),
10161-10169. doi:10.1016/j.eswa.2011.02.075.

172
S. M. A. Burney, T. Jilani, H. Tariq, Z. Asim, U. Amjad, S. S. Mohammad - A Portfolio Optimization Algorithm Using Fuzzy
Granularity Based Clustering

Professor Dr. S. M. Aqil Burney is the Head of Actuarial Sciences, Risk Management &
Mathematics at Institute of Business Management (IoBM) Karachi. He holds
M.Sc.(Statistics), M.Phil( Risk Theory and Insurance -Statistics) from University of
Karachi (UoK) and Ph.D.(Mathematics) from Strathclyde University, Glasgow-UK along
with many courses in Population Studies of UN, Computing. He has taught for more than
40 years at UoK and extensively delivered lectures at other institutions and universities of
Pakistan and abroad. He also holds extensive experience of academic management and
organization and Provost, Registrar, Project Director Development of Dept. of Computer
Science and a Institute of Information technology and founding director of Main Communication Network of
University of Karachi. Dr. Aqil Burney was Meritorious Professor at Dept of Computer Science University
of Karachi prior to joining at IoBM. He has published more than 135 research papers and 7 books nationally
and internationally in ICT, Mathematics, Statistics and Computer Science. He has supervised more than 10
PhD and 5 MS/M.Phil in Mathematics/Computer Science/Statistics and approved HEC Supervisor .Dr. Aqil
Burney is Chairman (elect) National ICT Committee for Standard PSQCA- Ministry of Science &
technology Govt. of Pakistan and member National Computing Education Accreditation Council (NCEAC),
Member IEEE(USA), Member ACM(USA) was Fellow Royal Statistical Society UK) for 30 years or so. His
fields of interests are algorithmic analysis & design of Multivariate Time series, Stochastic Simulation and
Modeling, Software engineering, computer science, soft computing, risk theory and insurance e-health
management and Data Sciences and Fuzzy and other logical systems.

Tahseen Jilani received the B.Sc. degree in Computer Science from Government Science
Degree College, in 1998, and the M.Sc. (Statistics) and Ph.D. (Computer Science) from
University of Karachi, Pakistan, in 2001 and 2007, respectively.
He is working as Associate Professor since 2014, in the Department of Computer Science,
University of Karachi. Since January, 2016, he is engaged with the School of Computer
Science and School of Medicine as post doc data scientist, University of Nottingham- UK.
His current research interests include data sciences, machine learning in medical sciences,
Statistical techniques for big data analytics, imprecise and uncertainty data modelling. Dr.
Jilani is a member of Rough set society (RSS) and Association for Professional Health Analysts (APHA).
He is serving as member of technical committee and active reviewers for many national and international
research activities. He was the recipient of the HEC Indigenous 5000 scholarship in 2003, the National
Science Foundation grant 2010, the Nottingham University fellowship and honorary postdoc at University of
Sterling.

Usman Amjad received BS. Degree in Computer Science from University of Karachi, in
2008. He recently completed his PhD in Computer Science from University of Karachi.
His research interests include soft computing, machine learning, artificial intelligence and
programming languages. He was the recipient of the HEC Indigenous 5000 scholarship in
2013. Currently, he is working as AI solution architect at Datics.ai Solutions.

Humera Tariq received B.E (Electrical) from NED University of Engineering and
Technology in 1999. She joined MS leading to PhD program at University of Karachi in
2009 and completed her PhD in 2015. Currently she is working as Assistant Professor at
Department of Computer Science, University of Karachi. Her research interest includes
image processing, biomedical imaging, Modeling, Simulation and Machine Learning.

Mr. Syed Shah Muhammad is working as a Lecturer in the department of computer

science in Virtual University of Pakistan. He obtained the degree of MS in computer
sciences (MSCS) with the specialization of Computer Networks from the University of
Agriculture, Faisalabad in 2005. Presently, he is Ph.D. Computer Science scholar at
University of Engineering and Technology (UET) Lahore.

173