Escolar Documentos
Profissional Documentos
Cultura Documentos
4, DECEMBER 2009
37
I. I NTRODUCTION
38
k
P
i=1
!
b
b
(1)
(0)
x
(k) = x (1)
e a(k1) +
a
h
i
1 T
where a
, b = B T B
B Xn and
(1)
,
B= .
..
..
.
x
(0) = x
(0) (1), x
(0) (2), x
(0) (3), , x
(0) (n)
where x
(0) (1) = x(0) (1). Applying the inverse AGO yields
a
(0)
(0)
x
(k) = x (1)
1 e e a(k1)
a
where k = 2, 3, . . . , n, x
(0) (n + 1) , x
(0) (n + 2) , are the
so-called GM(1,1) forecast values.
2) GM(1,N ) Model: Consider a system described by
(0)
(0)
the sequences xi (k), i = 1, 2, 3, . . . , n, in which x1 (k)
describes the main factor of interest and sequences
(0)
(0)
(0)
x2 (k), x3 (k), . . . , xn (k) are the factors which influence
this main factor. Such a system can be analyzed using the
following multivariate GM(1,N ) Grey model:
(0)
N
X
(1)
bj xj (k), k = 2, 3, . . . , n,
j=2
(1)
in which xj (k) =
(1)
k
P
i=1
(0)
(1)
(1)
0.5x1 (k 1), k 2.
(1)
Substituting all possible xj (k) terms into the
above yields a matrix of the form
(0)
x1 (2)
(0)
x1 (3)
XN = .
..
(0)
x1 (n)
(1)
(1)
(1)
z1 (2)x2 (2) . . . xn (2)
a
(1)
(1)
(1)
z1 (3)x2 (3) . . . xn (3) b2
= .
.
..
..
(1)
(1)
(1)
z1 (n)x2 (n) . . . xn (n)
bn
equation
=B
a.
x(0) (i),
k = 2, 3, 4, . . . , n.
The GM(1,1) model is then constructed by establishing the first-order differential equation of x(1) (k) as
dx(1) (k)
+ ax(1) (k) = b.
dk
1 T
Applying the matrix operation a
= BT B
B XN , the
values of bj , j = 2, 3, . . . , N can be easily found. The relative
influence exerted on the major sequence by each influencing
sequence can then be determined simply by inspecting the bj
value of the corresponding sequence.
HUANG & LI, AN OPTIMIZED PORTFOLIO ALLOCATION SYSTEM BASED ON GREY THEORY AND A MODIFIED MARKOWITZ MV MODEL
3) Grey Relational Analysis: [27] In general, Grey Relational Analysis (GRA) methods provide an effective means
of solving multiple-criteria decision problems by ranking the
potential solutions in terms of their so-called Grey Relational Grade (GRG) such that the optimal solution can be
determined. Whereas traditional statistical-based techniques
for analyzing the relationships between variables rely upon
the availability of large volumes of data and require the
factors to be expressed functionally, GRA has the advantage that it can operate with significantly less data and is
applicable to classification or decision-making problems involving many different factors. GRA therefore provides an
ideal tool for analyzing the complicated inter-relationships
amongst the individual parameters in systems with multiple performance characteristics [51][34][47][49] and has
been applied to a wide variety of optimization, decisionmaking and classification problems in such diverse fields
as finance, business, economics, design, manufacturing and
production [30][20][15][50][14][28][35][24].
In the GRA method, data characterized by the same set of
features are regarded as belonging to the same series. The
relationship between any two series of data can be determined
by evaluating the differences between them and assigning an
appropriate Grey Relational Grade (GRG).
According to Huang etc. [27], the attribute impulse factor
|xi (k)|
has the form oi (k) = |x
, where x0 and xi are the
0 (k)|
reference object and the inspected object, respectively. Furthermore, the grey relational grade is defined as
0i =
0i min
max min
where
1, 2, . . . , m,
0i
m
Q
m1
0i (k) ,
k=1
|xi (k)|
oi (k) = |x
, x0 (k) is the reference value, and xi (k) is
0 (k)|
0i and
the comparative value. Furthermore, min = min
i
0i .
max = max
i
Having calculated the GRG for each sequence, the sequences are ranked using a so-called grey relational ranking
procedure. For example, for the case of a reference sequence
x0 (k), the grey relational rank of sequence xi (k) is greater
than that of xj (k) if (x0 , xi ) > (x0 , xj ). The corresponding
ranking is denoted as xi xj .
B. Rough Set Theory
RS theory was introduced by Pawlak [39] in 1982 and provides a powerful mathematical tool for handling the vagueness
and uncertainty inherent in many decision-making processes.
RS theory is underpinned by the assumption that every object
in the universe of discourse is associated with a particular
set of information (i.e., attributes). Objects characterized by
the same information are regarded as indiscernible. The indiscernibility relationships generated amongst all the objects in
the universe of discourse provide the basic mathematical basis
for RS theory. Typical problems amenable to RS processing
include the classification of sets of objects based upon their
attribute values, checking the dependencies (full or partial)
39
40
n P
n
P
P
minimize
wi wj ij + (1 )
ri wi , subject
i=1 j=1
i=1
n
P
i=1
wi = 1 and n
n P
n
P
P
minimize
wi wj ij + (1 )
ri wi , subject
to
n
P
i=1
i=1 j=1
i=1
wi = 1, i = 1, 2, . . . , n and
w2 w1 0, w3 w2 0, . . . , wn wn1 0.
HUANG & LI, AN OPTIMIZED PORTFOLIO ALLOCATION SYSTEM BASED ON GREY THEORY AND A MODIFIED MARKOWITZ MV MODEL
41
are then processed using RS theory and a hybrid GRA / MVUAC scheme to determine the optimal stock portfolio.
In the forecasting model developed in this study, each element (Xi ) in U is processed by the GM(1,1) prediction model
and assigned appropriate values of the conditional attributes
(C1 Cn ) and decision-making attributes (D1 Dm ) based
upon its trend over the previous quarter.
Start
Yes
To continue
investment
No
End
42
by using the Box Plots method [10] to establish an interquartile range such that any data falling outside this range
can be automatically assigned a default value depending on
the interval within which it is located.
3) Step 3: GM(1,1) grey prediction: A GM(1,1) forecasting
model is used to predict the future trends of the financial
variables of each of the selected companies. Note that in the
current GM(1,1) model, the forecasting process is deliberately
restricted to a one-step-ahead mode in order to prevent the
accumulation of errors from the four previous forecasting
periods.
4) Step 4: Information reduction using GM(1,N ) multivariate model: To improve the efficiency of the RS selection
process, certain conditional attributes are removed if they are
found to have little effect on the decision-making attributes.
In the system proposed in this study, this pruning operation is
performed by using a GM(1,N ) multivariate model to identify
the top ten influential sequences (i.e., conditional attributes).
5) Step 5: K-means clustering: Prior to submission to the
RS stock selection mechanism, the forecast conditional attribute values (C1 Cn ) are clustered into three groups using
a K-means clustering algorithm. The objective of clustering
is to partition the conditional attributes of stock companies
into disjoint nonempty subsets such that similar attributes
are grouped together and attributes in different subsets are
dissimilar. Each of the three clusters represent Excellent,
Common, and Poor respectively.
6) Step 6: Selection of approximate sets: Having clustered
the forecast data, the RS model is applied to determine the
lower approximate set. The generalized rules extracted by the
lower approximate set are all recognized rules or relationships
in the stock markets. And these decision rules might be helpful
to evaluate potential stocks. In the current system, the RS
model could output the potential stocks for inclusion in the
optimal portfolio.
7) Step 7: Optimized fund allocation: Having identified
suitable stocks using the RS model, an integrated GRA / MVUAC scheme is applied to determine the portfolio allocation
which maximizes the overall rate of return. As described in
Section III-C, the optimal portfolio is constructed subject to
two constraints, namely:
1) the number of assets within the portfolio;
2) the maximum allocation wi of each selected asset i.
Having completed Steps 17 above, a check is made on
the overall rate of return on the investment. If the rate of
return is judged to be acceptable, a decision is made as to
whether or not the model should be run for a further quarter
using the existing attributes. However, if the rate of return
is unacceptable, the suitability of the conditional attributes is
reviewed and the attributes are amended if necessary.
IV. E VALUATION OF P ROPOSED M ODEL U SING
E LECTRONIC S TOCK DATA
A. Data Extraction
In this study, the feasibility of the proposed forecasting and
stock selection model was evaluated using electronic stock
data extracted from the New Taiwan Economy database (TEJ).
HUANG & LI, AN OPTIMIZED PORTFOLIO ALLOCATION SYSTEM BASED ON GREY THEORY AND A MODIFIED MARKOWITZ MV MODEL
43
Fig. 2: MV and MV-UAC efficient frontiers for portfolio containing two stocks (2441
and 3017).
0.214
0.005
0.089
0.081
0.074
0.005
0.117
-0.006
0.005
0.004
0.089
-0.006
0.161
0.072
0.065
0.081
0.005
0.072
0.147
0.046
0.074
0.004
0.065
0.046
0.127
44
TABLE I: COMPARISON OF EXPECTED RISK AND RATE OF RETURN FOR PORTFOLIOS CONTAINING DIFFERENT NUMBERS OF SELECTED STOCKS AT MINIMUM
EXPECTED RISK POINT ON EFFICIENT FRONTIER
2 companies
3 companies
4 companies
5 companies
expected
risk
rate of
return
expected
risk
rate of
return
expected
risk
rate of
return
expected
risk
rate of
return
0.395
10.18
0.3415
-0.61
0.2514
-3.44
0.22851
-8.86
0.279
-3.99
0.2481
-6.14
0.2374
-5.71
0.2271
-7.57
0.4449
15.19
0.4323
15.26
0.3584
14.18
0.33387
20.90
21.39
8.51
5.03
4.48
0.3243
5.23
0.3062
2.43
0.3037
1.53
0.2806
5.17
0.2634
-0.49
0.2397
-5.59
0.1940
-2.78
0.1914
-2.10
0.2227
32.74
0.2122
30.19
0.1853
24.43
0.1831
20.87
37.47
27.04
23.19
23.94
0.3530
-1.53
0.3280
-1.80
0.2595
1.27
0.2391
-3.81
0.2700
0.85
0.2240
5.21
0.2264
3.39
0.1986
1.58
0.3562
50.14
0.3028
36.58
0.2503
37.91
0.2065
26.89
49.45
39.99
42.57
24.66
108.31
75.53
70.79
53.08
36.10
25.18
23.60
17.69
TABLE II: COMPARISON OF EXPECTED RISK AND RATE OF RETURN FOR PORTFOLIOS CONTAINING DIFFERENT NUMBERS OF SELECTED STOCKS AT
MAXIMUM EXPECTED RISK POINT ON EFFICIENT FRONTIER
Maximum Lowest expected return point of efficient frontier without allocation constrains
No. of selected stocks
2 companies
3 companies
4 companies
5 companies
expected
risk
rate of
return
expected
risk
rate of
return
expected
risk
rate of
return
expected
risk
rate of
return
0.4530
22.02
0.42303
19.01
0.28466
-3.87
0.25652
-5.50
0.4185
4.85
0.39297
3.72
0.3671
2.88
0.34055
1.13
0.4841
7.94
0.50135
20.72
0.46506
19.81
0.42817
20.93
34.81
43.45
18.82
16.56
0.3434
7.02
0.3525
-2.62
0.3402
-3.08
0.3003
8.43
0.2858
-3.25
0.2757
-12.20
0.2457
-10.35
0.2322
-8.64
0.2495
24.07
0.2732
20.40
0.2722
11.92
0.2426
10.69
27.84
5.58
-1.51
10.48
0.4006
11.98
0.3767
10.20
0.3469
-3.63
0.3180
6.67
0.2883
-3.37
0.2687
-2.01
0.3393
-6.49
0.2708
-3.47
0.3894
48.21
0.3626
46.60
0.3323
45.86
0.3006
42.59
56.82
54.78
35.74
45.79
119.47
103.81
53.05
72.83
39.82
34.60
17.68
24.28
HUANG & LI, AN OPTIMIZED PORTFOLIO ALLOCATION SYSTEM BASED ON GREY THEORY AND A MODIFIED MARKOWITZ MV MODEL
45
TABLE IV: COMPARISON OF RETURN RATE OBTAINED USING FOUR DIFFERENT STOCK SELECTION MODELS BASED ON THE MV AND MV-UAC EFFICIENT
FRONTIER CURVES RESPECITIVELY
rate of
return
the year
rate of
return
rate of return
rate of
return
the year
rate of
return
rate of return
-8.86
4.48
-3.85
11.42
-5.50
16.56
11.07
30.76
-7.57
20.90
5.17
-2.10
20.87
-3.81
1.58
0.85
-3.47
-0.29
26.89
34.57
42.59
42.59
Accumulated 3 years
rate of return
53.08
63.64
72.83
96.50
-4.37
1.13
19.64
23.94
1.90
20.93
20.71
-2.46
17.69
-3.92
8.43
18.57
10.48
-8.64
21.27
24.66
1.13
21.21
applied to select the stocks for inclusion within the portfolio. Finally, a hybrid GRA / MV-UAC Markowitz model is
employed to determine the optimal portfolio given a specified
number of stocks within the portfolio and constraints on the
proportion of the available capital to be allocated to each stock
within the portfolio. The major findings of this study can be
summarized as follows:
1) In theory, the expected risk of a portfolio will decrease
with an increasing number of selected stocks when
the average risk of each selected stock is the same
and the return rates of the individual selected stocks
are independent of one another. However, the results
presented in this study have shown that these conditions
do not hold in practice, and thus the expected risk does
not significantly reduce as the number of selected stocks
is increased.
2) The rate of return obtained using the constrained MVUAC portfolio allocation model is higher than that
obtained using the conventional MV model irrespective
of the point chosen on the efficient frontier (i.e., the
lowest expected risk or the maximum expected return)
when optimizing the portfolio.
3) The points lying on the efficient frontier obtained using
the MV-UAC model are a subset of those on the efficient frontier obtained using the MV model since the
former model differs from the latter only in terms of
its inclusion of some additional un-equivalent allocation
constraints.
Overall, the results presented in this study have confirmed
that the proposed stock forecasting and portfolio allocation
mechanism successfully constructs a portfolio which yields the
optimal rate of return subject to the pre-determined constraints
imposed by the efficient frontier of the MV-UAC model. In a
future study, an investigation will be performed to investigate
the relative efficacies of different GRA methods in enhancing
the ability of the MV-UAC portfolio allocation module to in-
6.67
16.76
-3.68
10.69
31.50
1.53
18.92
45.79
6.67
24.28
48.97
32.17
46