Information Sciences 177 (2007) 1193–1203

www.elsevier.com/locate/ins

q
FuzzyTree crossover for multi-valued stock valuation
a,* b
Ping-Chen Lin , Jiah-Shing Chen
a
Institute of Finance and Information, National Kaohsiung University of Applied Sciences, Kaohsiung, 807 Taiwan, ROC
b
Department of Information Management, National Central University, Jungli, 320 Taiwan, ROC

Abstract

Stock valuation is very important for fundamental investors in order to select undervalued stocks so as to earn excess
profits. However, it may be difficult to use stock valuation results, because different models generate different estimates for
the same stock. This suggests that the value of a stock should be multi-valued rather than single-valued. We therefore
develop a multi-valued stock valuation model based on fuzzy genetic programming (GP). In our fuzzy GP model the value
of a stock is represented as a fuzzy expression tree whose terminal nodes are allowed to be fuzzy numbers. There is scant
literature available on the crossover operator for our fuzzy trees, except for the vanilla subtree crossover. This study gen-
eralizes the subtree crossover in order to design a new crossover operator for the fuzzy trees. Since the stock value is esti-
mated by a fuzzy expression tree which calculates to a fuzzy number, the stock value becomes multi-valued. In addition,
the resulting fuzzy stock value induces a natural trading strategy which can readily be executed and evaluated. These exper-
imental results indicate that the fuzzy tree (FuzzyTree) crossover is more effective than a subtree (SubTree) crossover in
terms of expression tree complexity and run time. Secondly, shorter training periods produce a better return of investment
(ROI), indicating that long-term financial statements may distort the intrinsic value of a stock. Finally, the return of a
multi-valued fuzzy trading strategy is better than that of single-valued and buy-and-hold strategies.
 2006 Elsevier Inc. All rights reserved.

Keywords: Multi-valued stock valuation; Intrinsic value; Fuzzy number; Genetic programming

1. Introduction

Stock valuation sums up the activities of estimating the intrinsic value of a business entity. It is important to
securities analysis, loan decisions, and leveraged buyout analysis, and so on. Investors could suffer vast losses
if they were to make improper decisions based on an inaccurate business value. To determine the value of busi-
ness, it is necessary to understand the activities disclosed in financial statements. Conventional methods to
stock valuation using financial statements are divided into three categories: asset appraisal [11], discounted

q
This work was supported by the National Science Council, Taiwan, ROC, under contracts NSC 92-2626-H-238-001.
*
Corresponding author. Tel.: +886 73814526; fax: +886 73831544.
E-mail address: lety@cc.kuas.edu.tw (P.-C. Lin).

0020-0255/$ - see front matter  2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.ins.2006.08.017
1194 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203

present value [1,12], and price/earnings multiples price [5]. According to these valuation methods, using
different critical input variables gives rise to different outcomes even for the same stock. This suggests that
the value of a business may be multi-valued rather than single-valued. Another drawback of these approaches
is their use of pre-determined functions to design valuation models, which try to estimate the value of business
from linear functions under specific assumptions and limitations. They always fail to fully capture flexibility
and uncertainty.
The various soft computing technologies do provide alternative solutions to financial problems. For exam-
ple, fuzzy logic is used as a possibility distribution of portfolios [7,17,19], or for the credit analysis of loans
[5,8]. Neural networks are used to predict financial distress [2,3,6]. One of the evolutionary computational
techniques, genetic programming (GP), is applied to a stock market [10], futures or options pricing [9], and
a foreign exchange market [4]. However, few studies have used soft computing methods for stock valuation.
In this article we apply both fuzzy numbers to manifest multi-valued uncertainty and genetic programming to
optimize an effective stock valuation model.
It is well recognized that crossover and selection operators play the major role in terms of generating
solutions in GP [15]. A subtree crossover operator usually destroys building blocks (i.e., an effective partial
trees), due to randomly and blindly choosing crossover points. Hence, many investigators propose new cross-
over methods to obtain more effective building blocks by reserving crucial schemata. For example, a hierar-
chical crossover can be combined with simulated annealing and hill climbing to find correct solutions via
shrinking, growth, or internal substitution while preserving syntactic correctness [22]. Depth-dependent cross-
over accumulates building blocks according to the depth of a node. The depth selection ratio is higher for
nodes closer to a root node [20]. A directing crossover reduces the amount of unviable code (bloat) in individ-
uals while searching for a parsimonious solution [21]. It involves the identification of highly fit nodes to be
used as crossover points during the operator application. The island model crossover applies a subtree cross-
over to aborigines and the depth-dependent crossover is for immigrants based on their ages, which demon-
strates how long they can survive in the deme [16]. This can integrate many schemata to form a bigger
building block of different deme. The dynamic page-based crossover is described in terms of the number of
pages for all individuals. Pages are expressed as a number of instructions, which represents the dynamic
change for all individuals in the population. These give rise to a succinct solution without penalizing optimi-
zation ability [13,14]. These crossover operators only exchange constant schemata for all individuals in the
population. No new genotype is generated even when swapping partial trees in the dedicated population.
These are few studies available for exploring the new crossover operator in fuzzy genetic programming except
for the subtree.
The objective of the present study is to develop a FuzzyTree crossover for a multi-valued stock valuation
model which improves the convergence phenomenon. We generalize crisp expression trees evolved through
GP to fuzzy ones by introducing fuzzy numbers and fuzzy arithmetic operators in the trees. FuzzyTree uses
a subtree crossover operator if the selected crossover point is an internal node; otherwise, the selected terminal
nodes are snipped into pieces and interchanged with each other. This could improve convergence via protect-
ing the building block and increasing the variety of genotypes. Since the stock value is estimated by a fuzzy
expression tree which calculates to a fuzzy number, the real stock values become multi-valued. In addition,
the resulting trapezoidal fuzzy stock value induces a natural trading strategy which can readily be executed
and evaluated.

2. Fuzzy genetic programming approach

Our FuzzyTree model generalizes the subtree crossover to produce the next generation’s fuzzy GP individ-
ual. The fuzzy GP individual is represented by a fuzzy GP tree as shown in Fig. 1, which contains terminal
nodes (ni or vi) and internal nodes (fi). Each terminal node is represented by a trapezoidal fuzzy number.
The detailed FuzzyTree crossover algorithm is described in Section 2.1. Section 2.2 introduces the encoding
for each fuzzy node. The related arithmetic processing of our fuzzy GP tree is shown in Section 2.3. The eval-
uation of an individual’s ability to survive relies on the fitness function illustrated in Section 2.4. Finally, we
propose a fuzzy trading strategy to obtain better investment returns in Section 2.5.
 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203 1195

f1

f2 f3

v1 n1 n2 f4

v3 n3
μ A(x)

x1 x2 x3 x4
x1<=x2<=x3<=x4
Fig. 1. Fuzzy GP tree.

procedure EA
Initialize population
Evaluate population
while (termination condition is false)
Evaluation
Selection
Crossover
Mutation
endwhile
Report the best solution found
endprocedure
Fig. 2. Evolutionary algorithm.

2.1. FuzzyTree crossover

A basic evolutionary algorithm (EA) introduces a simple evaluation–selection-crossover-mutation loop as

outlined in Fig. 2.
As a general framework of our proposed FuzzyTree crossover algorithms (Fig. 3), Tree1 and Tree2 denote
the selected trees from the selection method, and NewTree1 and NewTree2 are the offspring generated by this
crossover function. Ratec is the crossover rate. The selection process is to choose relatively good solutions and
eliminate those not-so-good solutions from the parent population. In our model we use a well-known tourna-
ment selection methodology to pick up relatively good offspring, because it achieves better performance
[10,16,21]. The terms n1 and n2 are randomly selected crossover points from Tree1 and Tree2, respectively.
1196 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203

Fig. 3. FuzzyTree crossover algorithm.

f1 Tree1 f1 NewTree1

f2 f3 f2 f3

n1 v1 n2 f4 n1 v1 n2 f4

2 5 7 9 v3 n3 2 5 4 7 v3 n3
2 4 5 7
f1 Tree2 f1 NewTree2

f2 f3 f2 f3

n1 v1 n2 n1 v1 n2

1 3 4 7 1 3 7 9

Fig. 4. FuzzyTree crossover: (a) before crossover and (b) after crossover.

If both n1 and n2 belong to terminal nodes, then our proposed FuzzyTree(Æ) crossover method is employed;
otherwise, the conventional GP crossover method Subtree(Æ) is used.

• FuzzyTree(Æ) Crossover Function

In this section, the FuzzyTree(Æ) function is performed only when both terminal nodes (n1 and n2) are
selected to be crossover points simultaneously, as shown in Fig. 4. This means that both n1 and n2 should
be represented by trapezoidal fuzzy numbers, x1 6 x2 6 x3 6 x4. For example, in Fig. 4(a), before the cross-
over, n1 = [2, 5, 7, 9] and n2 = [1, 3, 4, 7]. Both n1 and n2 should be snipped into two pieces and the pieces then
interchanged with each other, where the snipped point is randomly selected. Assume that the newly-gener-
ated fuzzy node is ½x01 ; x02 ; x03 ; x04 . To maintain the order of the newly-generated fuzzy numbers, they are sorted
increasingly to ½x001 ; x002 ; x003 ; x004 , where x001 6 x002 6 x003 6 x004 . Assuming the snipped point lies between x2 and x3,
our crossover operator produces [2, 5, 4, 7] and [1, 3, 7, 9] after the interchanging takes place. These two fuzzy
 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203 1197

s1 s2 s3 s4 s5

x1 x2 x3 x4

Fig. 5. The snipped point selection in a trapezoidal fuzzy number.

numbers that are produced should be sorted increasingly to [2, 4, 5, 7] and [1, 3, 7, 9], as shown in Fig. 4(b).
Finally, the newly-generated offspring NewTree1 and NewTree2 are obtained. In short, the FuzzyTree cross-
over method is used when two leaf nodes are chosen as the selected crossover points on the two trees, whereas
the SubTree crossover method is a conventional GP-like method, except that each leaf node is a trapezoidal
fuzzy number. Each trapezoidal fuzzy number is represented by four floating numbers. If the fuzzy number is
described as [x1, x2, x3, x4], where x1 6 x2 6 x3 6 x4, then the snipped point (s1, s2, s3, s4, s5) is randomly
selected among them as shown in Fig. 5. In the current implementation, the numbers themselves are not sub-
ject to the crossover operator. They are only changed by mutation as in standard GP.

2.2. Encoding

Each individual in a GP population is a fuzzy expression tree, which represents a valuation model. An
expression tree consists of terminal nodes and internal nodes. A terminal node can be a financial variable
or a constant, while an internal node can be an allowed fuzzy arithmetic operator. The expression tree is
shown in Fig. 1, where v1, v2, . . . , vm 2 {R1, R2, . . . , R39} are financial variables, f1, f2, . . . , fk 2 {+, , ·} are fuzzy
operators, and n1, n2, . . . , ni are trapezoidal fuzzy numbers (constants). There are 39 available financial vari-
ables used in our study as listed in Table 1. Each fuzzy number is encoded by four floating numbers. A trap-
ezoidal fuzzy number (constant) can be denoted as a 4-tuple [x1, x2, x3, x4], as depicted in Fig. 1. Each financial
variable is a non-fuzzy number (exact value) which is represented as a degenerated trapezoidal fuzzy number
with x1 = x2 = x3 = x4.

2.3. Fuzzy arithmetic

To evaluate an expression tree with trapezoidal fuzzy numbers as terminal nodes, we define several fuzzy
arithmetic operations on trapezoidal fuzzy numbers so that the resulting fuzzy numbers will also be trapezoi-
dal. Currently, the supported fuzzy operations in our model are +,  and ·. Let X = [x1, x2, x3, x4],
Y = [y1, y2, y3, y4] be two trapezoidal fuzzy operands. We define the fuzzy +, , · as follows (Eqs. (1)–(3))
fX þ Y g ¼ fðx1 þ y 1 Þ; ðx2 þ y 2 Þ; ðx3 þ y 3 Þ; ðx4 þ y 4 Þg ð1Þ
fX  Y g ¼ fðx1  y 4 Þ; ðx2  y 3 Þ; ðx3  y 2 Þ; ðx4  y 1 Þg ð2Þ

Table 1
Some of the financial variables
Ratio Description
R1 Return on total assets (%)
R2 Book value per share
R3
.. .Current liabilities (%)
. ..
R39 Operating income per employee
1198 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203

fX  Y g ¼ minðx1 y 1 ; x4 y 1 ; x1 y 4 ; x4 y 4 Þ;
minðx2 y 2 ; x2 y 3 ; x3 y 2 ; x3 y 3 Þ;
maxðx2 y 2 ; x2 y 3 ; x3 y 2 ; x3 y 3 Þ;
maxðx1 y 1 ; x4 y 1 ; x1 y 4 ; x4 y 4 Þ ð3Þ
It is not difficult to see that the above definitions result in well-formed trapezoidal fuzzy numbers, i.e.
z1 6 z2 6 z3 6 z4. Since every operator produces a trapezoidal fuzzy number, an expression tree also yields
a trapezoidal fuzzy number, [z1, z2, z3, z4], which represents the fuzzy value of a stock under consideration.
In addition, the trapezoidal fuzzy stock value also induces a natural trading rule where the two sides of the
trapezoid are the buying range ([z1, z2]) and the selling range ([z3, z4]), respectively, and their slopes are used
as the investment weights.

2.4. Fitness function

In order to find effective valuation rules under maximum return and minimum risk, the fitness function
evaluates the return of the trading strategy induced by the fuzzy expression tree. Given a fuzzy expression tree
E, its fitness f(E) is derived from Eq. (4), where ROI is the return on the investment from all trades and r is the
standard deviation from the net value of all of the trading days. Let NVi be the net value of the ith trade, NV
be the mean, and N be the number of trades; r is evaluated from Eq. (5).
ROI
F ¼ ð4Þ
srffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn 2
t¼1 ðNVt  NVÞ
r¼ ð5Þ
N

2.5. Fuzzy trading strategy

We propose two fuzzy trading strategies to analyze the performance of the ROI. They are the multi-valued
trading strategy based on the trapezoidal fuzzy number and the single-valued trading strategy based on the
triangle fuzzy number.

• Multi-valued trading strategy

The trapezoidal fuzzy number is used for stock valuation, giving rise to a multi-valued trading strategy
as shown in Fig. 6. Each trapezoidal fuzzy number on a multi-valued stock price is divided into three
ranges: the buying range, selling range, and intrinsic value range. Our trading strategies apply buying
actions, selling actions, and doing nothing, respectively. If the market price of stock (p) enters the buying

μ A(x)

Invested
Capital
Ratio

Buying Intrinsic Value Selling

Range Range Range
x
0 x1 x2 x3 x4
Stock Price
x1 ≤ x2 ≤ x3 ≤ x4

Fig. 6. Multi-valued trading strategy.

 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203 1199

μ A(x)

Invested
Capital
Ratio
Buying Selling
Range Range

x
0 x1 x2 x3
Stock Price
x1 ≤ x2≤ x3

Fig. 7. Single-valued trading strategy.

range, x1 6 p 6 x2, then the invested capital ratio is proportional to the degree of membership. On the con-
trary, if p enters the selling range, x3 6 p 6 x4, then the sell shares ratio depends on the degree of member-
ship. Finally, neither buying nor selling actions are used if p falls within the intrinsic value range,
x2 6 p 6 x3.
• Single-valued trading strategy
We also use a triangle fuzzy number for stock valuation thus inducing a single-valued trading strategy as
shown in Fig. 7. The single-valued (triangle) strategy is a special case of a multi-valued (trapezoidal) trading
strategy. The difference between them is that the intrinsic value of a stock is only a single value in this strat-
egy. Each triangle fuzzy number is divided into two ranges: a buying range and a selling range. Similar buy-
ing actions and selling actions are used if the market price of stock (p) enters the buying range, x1 6 p 6 x2,
and the selling range, x2 6 p 6 x3, respectively.

3. Experimental results

The simulation environment, sample data, and experimental results are described in this section. Our
FuzzyTree-based program is written in Borland C++ Builder 6.0.

3.1. Fuzzy GP parameters

The parameters used in our fuzzy GP runs are shown in Table 2. The population size is 5000. The number
of generations is set to 500. The tournament selection method is used with the size of the tournament being 2.
The crossover method is our proposed FuzzyTree crossover. The crossover rate is set to 0.9 and the mutation
rate is set to 0.05. The parameters used in our experiments are based on Srinivas’s and Lalit’s suggestions [18].
Minor changes in these parameters do not seem to have a major effect on performance in our pilot experiments
except for crossover and selection.

Table 2
Parameters of Fuzzy GP
Number of generations 500
Population size 5000
Arithmetic operators +, , ·
Maximum tree depth 5
Selection Tournament
Crossover FuzzyTree crossover
Crossover rate 0.9
Mutation Replacing a subtree
Mutation rate 0.05
Reserved elitists Three best individuals of each population
1200 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203

Table 3
Descriptive statistics for our target companies
Company Market value Total assets
Lite-on Electronics Co., Ltd. (LOE) 869.86 309.71
United Micro Electronics Co., Ltd. (UME) 4519.41 3005.28
Microtek Electronics Inc. (Microtek) 21.8 117.38
Delta Electronics, Inc. (Delta) 666.12 457.86
Advanced Semiconductor Engineering, Inc. (ASE) 948.77 669.25
Kinpo Electronics, Inc. (Kinpo) 250.21 199.89
Compeq Manufacturing Co., Ltd. (Compeq) 115.28 282.48
Hon Hai Precision Co., Ltd. (Hon Hai) 3552.04 1264.22

3.2. Sample data

Eight electronics companies are arbitrarily selected to be our testing targets as listed in Table 3. These com-
panies have been listed and traded on the Taiwan Stock Exchange (TSE) since 1995 or earlier. The relevant
data are collected from the Taiwan Economic Journal Database (TEJ).
The data encompass the entire period from January 1, 1992 to June 31, 2003. The training phase (k), the test
phase (l), and the validation phase (v) are one period in each sliding window (SW) as shown in Fig. 8. The
terms k, l, and v could be one month, one quarter (quarterly report), half a year, or one year (annual report)
according to the financial statement reporting period. In this study we use financial annual reports as our
experimental data. The sliding window shifts one validation period gradually until June 31, 2003. The term
SWklv denotes the total sliding window size: k + l + v. Among them, the training phase data are used to
learn a stock valuation model; the test phase data are used to obtain multi-valued stock prices from the
learned model; and the validation phase data are used for calculating the ROI according to the trading strat-
egies. It is noted that the financial statements announced to the public are generally delayed half a year in Tai-
wan’s stock market. Thus, the testing phase data in reality are delayed half a year until the validation phase
data in our experiments.

3.3. Analysis of results

For brevity, we summarize the performance of the FuzzyTree crossover by mean of three parts: (1) the
comparisons of execution performances between FuzzyTree and subtree crossovers, (2) the relationship
between the size of the sliding window and the ROI, and (3) the comparisons of the ROIs between multi-val-
ued and single-valued trading strategies. To avoid outlier results in our experiments, each experiment is per-
formed five times and their mean values are taken.

SW1
Training Test Validation
k l v

SW2
Training Test Validation
k l v

…
SWi
Training Test Validation
k l v

Fig. 8. Sliding windows simulation process.

 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203 1201

Table 4
Execution performances (FuzzyTree vs. subtree crossover)
Company FuzzyTree crossover Subtree crossover
No. of nodes Exec. time No. of nodes Exec. time
LOE 15.8 03:51 18.4 15:22
UME 14.3 04:12 25.8 15:21
Microtek 16.2 04:23 23.3 15:34
Delta 14.6 04:22 27.9 15:19
ASE 13.7 04:08 19.3 15:33
Kinpo 11.1 03:53 22.2 15:11
Compeq 12.3 04:02 23.3 15:22
Hon Hai 13.4 04:05 24.4 15:24
Mean 13.93 04:37 23.08 15:23

• Execution performance
The objective of this study is to find a simple precise valuation model. The mean execution time and
number of nodes are compared using the FuzzyTree and subtree crossover methods individually, which
are shown in Table 4. The mean number of nodes used in the FuzzyTree crossover (13.93) is less than that
used in the subtree crossover (23.08). In addition, the mean execution time of the FuzzyTree crossover
(4 min) is shorter than that of the subtree crossover (15 min). It is obvious that our proposed FuzzyTree
crossover method could find an effective and succinct valuation model more quickly than the subtree
method.
• Sliding window
Due to over-fit learning, the size of the training phase (k) in the sliding window may deeply influence the
final ROI. We also try different sliding window sizes. In Table 5 we set the test phase (l) = 1 and the val-
idation phase (v) = 1, and choose k = 10, 3, 1, respectively, to compare their ROIs. The mean ROIs of
SW1011, SW311, SW111 are 4.39, 10.7, and 15.8 as shown in Table 5. By implementing the sliding
window size, we prove that a shorter training phase supports a better ROI. It appears that the long-term
financial statement information could distort the intrinsic value of a stock.

Table 5
Sliding windows size vs. ROI
Company SW1011 SW311 SW111
LOE 7.25 4.01 10.92
UME 10.04 1.58 1.1
Microtek 0.62 76.92 35.65
Delta 24.29 20.41 34.49
ASE 6.34 1.73 9.37
Kinpo 2.39 2.62 9.52
Compeq 1.15 49.37 6.58
Hon Hai 15.83 30.88 18.79
Mean 4.39 10.7 15.8

Table 6
ROIs of multi-valued, single-valued, and buy-and-hold strategies
Company Multi-valued Single-valued Buy-and-hold
LOE 22.75 8.59 44.36
UME 2.05 11 53.83
Microtek 15.03 10.32 13.51
Delta 27.56 4.49 21.64
ASE 12.42 8.18 35.27
Kinpo 13.25 2.09 28.57
Compeq 17.26 7.26 72.65
Hon Hai 19.25 4.18 25.28
Mean 16.2 4.26 36.89
1202 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203

40
ROI-MultiValued
ROI-SingleValued
ROI-BuyHold
20

0
LOE UME Microtek Delta ASE Kinpo Compeq Hon Hai
ROI
-20

-40

-60

-80
Companies

Fig. 9. Comparison of ROI performances of multi-valued, single-valued, and buy-and-hold strategies.

• Trading strategy
Two trading strategies: a multi-valued strategy using the FuzzyTree crossover and a single-valued strat-
egy using SubTree crossover are proposed to achieve better ROIs than using a buy-and-hold strategy. The
buy-and-hold strategy is a general comparison benchmark, which implies buying stocks at the beginning,
holding (doing nothing) until the end of the investment period, and selling them, regardless of rational
stock prices.
Table 6 lists the ROI obtained using a buy-and-hold trading strategy (buy-and-hold), a single-valued
trading strategy (single-valued), and a multi-valued trading strategy (multi-valued), respectively. Because
long-term statements distort the intrinsic value of a stock, the selected training phase, test phase, and val-
idation phase are the years 2000, 2001, and (2002.7.1–2003.6.30), respectively.
From Table 6, it is shown that the ROI of a multi-valued strategy (16.20) is greater than those of single-
valued (4.26) and buy-and-hold (-36.89) strategies on for all eight companies. Similar results are also shown
in Fig. 9. This means that multi-valued prices are more suitable for stock valuation and trading strategy
than single-valued prices.

4. Conclusions

Stock valuation is important to investment policymakers. The conventional methods, such as asset apprai-
sal, discounted present value, and price/earnings multiples, only use simple estimation functions to evaluate a
stock’s value, lack flexibility and give rise to uncertainty. Therefore, we apply a fuzzy number to represent
multi-valued uncertainty and GP to optimize the valuation model. However, the conventional crossover
method is not available, because each leaf node is a trapezoidal fuzzy number in our model. In this paper
we use fuzzy numbers to describe multi-valued business values and two operators (FuzzyTree, SubTree) to
solve crossover problems. The results for eight-arbitrarily-selected electronics companies demonstrate the fea-
sibility of a multi-valued stock valuation model and the superiority of the FuzzyTree crossover operator. First,
the number of tree nodes of the FuzzyTree operator is easier to determine than that of the subtree. Therefore,
the run time of the FuzzyTree is also shorter than that of the subtree. The FuzzyTree crossover could find a
succinct stock valuation model more effectively than the subtree model. The FuzzyTree crossover seems to
improve convergence during evolution. Second, from the results of different sliding window sizes, it seems that
a shorter training period produces better ROI. This indicates that long-term financial statements could distort
the intrinsic value of stocks. Finally, the ROIs obtained using a multi-valued strategy are better than those
 P.-C. Lin, J.-S. Chen / Information Sciences 177 (2007) 1193–1203 1203

obtained single-valued and buy-and-hold strategies. This shows that the reasonable business value should be
multi-valued rather than single-valued. The FuzzyTree crossover is suitable for a multi-valued stock valuation
model. Clearly, this technique is a promising tool in multi-valued stock valuation. Based on these results, a
dynamic multi-valued valuation model could be further discussed in the future, where the business value is
adjusted according to a distinct bear or bull market.

Acknowledgement

We are extremely grateful to two anonymous referees and the editor, Witold Pedrycz, for many useful com-
ments and suggestions on this paper.

References

[1] L. Andersen, How options analysis can enhance managerial performance, European Management Journal 20 (2002) 505–511.
[2] B.S. Ahn, S.S. Cho, C.Y. Kim, The integrated methodology of rough set theory and artificial neural network for business failure
prediction, Expert Systems with Applications 18 (2000) 65–74.
[3] B. Back, T. Laitinen, K. Sere, Neural networks and genetic algorithms for bankruptcy predictions, Expert Systems with Applications
11 (1996) 407–413.
[4] S. Bhattacharyya, O.V. Pictet, G. Zumbach, Knowledge-intensive genetic discovery in foreign exchange markets, IEEE Transactions
on Evolutionary Computation 6 (2002) 169–181.
[5] F.E. Block, A study of price to book relationship, Financial Analysts Journal (1995) 63–73.
[6] J.E. Boritz, D.B. Kennedy, Effectiveness of neural network types for prediction of business failure, Expert Systems with Applications
9 (1995) 503–512.
[7] C. Carlsson, R. Fuller, P.A. Majlender, Possibilistic approach to selecting portfolios with highest utility score, Fuzzy Sets and Systems
131 (2002) 13–21.
[8] L.H. Chen, T.W. Chiou, A fuzzy credit-rating approach for commercial loans: a Taiwan case, The International Journal of
Management Science 27 (1999) 407–419.
[9] S.H. Chen, W.C. Lee, C.H. YehChiou, Hedging derivative securities with genetic programming, International Journal of Intelligent
Systems in Accounting, Finance and Management 8 (1999) 237–251.
[10] S.H. Chen, C.H. Yeh, Toward a computable approach to the efficient market hypothesis: an application of genetic programming,
Journal of Economic Dynamics and Control 21 (1997) 1043–1063.
[11] J.R. Dietrich, M.S. Harris, K.A. Muller, The reliability of investment property fair value estimates, Journal of Accounting and
Economics 30 (2001) 125–158.
[12] J. Francis, P. Olsson, D.R. Oswald, Comparing the accuracy and explainability of dividend, free cash flow, and abnormal earnings
equity value estimates, Journal of Accounting Research 38 (2000) 45–70.
[13] M.I. Heywood, A.N. Zincir-Heywood, Dynamic page based crossover in linear genetic programming, IEEE Transactions on Systems,
Man, and Cybernetics – Part B: Cybernetics 32 (2002) 380–388.
[14] M.I. Heywood, A.N. Zincir-Heywood, Page-based linear genetic programming, IEEE International Conference on Systems, Man,
and Cybernetics (2000) 3823–3828.
[15] J. Koza, Genetic Programming, MIT Press, Cambridge, MA, 1995.
[16] I. Makoto, I. Hitoshi, Island model GP with immigrants aging and depth-dependent crossover, in: Proceedings of IEEE 2002
International Conference on Evolutionary Computation, 2002, pp. 267–272.
[17] T.H. Payne, J.H. Finch, Effective teaching and use of the constant growth dividend discount model, Financial Services Review 8
(1999) 283–291.
[18] M. Srinivas, M.P. Lalit, Genetic algorithms: a survey, IEEE Computer 27 (1994) 18–20.
[19] H. Tanaka, H. Nakayama A. Yanagimoto, Possibility portfolio selection, in: Proceedings of 1995 IEEE International Joint
Conference on Fuzzy Systems, 1995, pp. 813–818.
[20] I.T.O. Takuya, I.B.A Hitoshi, S.A.T.O. Satoshi, Depth-dependent crossover for genetic programming, in: Proceedings of IEEE 1998
International Conference on Evolutionary Computation, 1998, pp. 775–780.
[21] M.D. Terrio, M.I. Heywood, Directing crossover for reduction of bloat in GP, in: Proceedings of the IEEE 2002 Canadian
Conference of Electrical and Computer Engineering, 2002, pp. 1111–1115.
[22] O. Una-May, O. Franz, Hybridized crossover-based search techniques for program discovery, IEEE 1995 International Conference
on Evolutionary Computation (1995) 575–582.