Proceedings of the 2nd International Conference on Information Technology for Application (ICITA 2004)
A Hybrid Swarm Intelligence Based Mechanism
for Earning Forecast
Po-Chang (P.C.) Ko, and Ping-Chen (P.C.) Lin
Abstract-- Earning forecast plays an important role in the
investment decision because it really reflects future growth of
business. Conventionally, the statistical tools, such as the
Univariate Time Series Methods and Multivariate Methods, are
broadly used in the earning forecast analysis. However, the
inferred forecasting accuracy is seriously limited without
considering sufficient critical financial ratios. Furthermore, the
forecasting accuracy is also influenced by its linear searching
sequence. In this paper, we propose a hybrid swarm intelligence
based mechanism, which combines the advantages of statistical
methodologies and Particle Swarm Optimization (PSO) method,
efficiently. The PSO algorithm provides the nonlinear searching
methodology and evolutionary computation characteristics by
adjusting the trajectories of a population of particles through a
problem spaces on the basis of information about each particles
previous best performance and the best previous performance of
its neighbors. Using the discrete binary version of PSO, the
introduced mechanism can intelligently select the minimal critical
financial ratios set to predict more precise earning.
Simultaneously, the earning forecast can be more accurate by
introducing the Sliding Window Process in our proposed
mechanism. From the empirical results, its evident that the
introduced mechanism is superior in the earning forecasting
accuracy with minimal critical financial ratios consideration. In
addition, the execution efficiency and convergence speed are
excellent in our proposed mechanism..
Index TermsParticle Swarm Optimization, Genetic
Algorithm, Evolutionary Computation, Multiple Regression,
Sliding Window, Earning Forecast.
I. INTRODUCTION
arning forecast plays an important role in the investment
decision because it really reflects the growth of business
[1]. More precise earning prediction effectively generates
abnormal investment returns [2]-[5]. It means that the
investment returns deeply depend on the earning forecast
accuracy. Furthermore, the earning forecast accuracy also
provides recommendations for the individual investor and
firms insider trader to make the
best investment decision [6]-[7]. However, making the
investment decision is greatly difficult because it is influenced
by the investors cognition, investment behavior, psychological
limitation and even the global economical and political
environment.
P. C. Ko is with the Information Management Department, National
Kaohsiung University of Applied Sciences, Kaohsiung, Taiwan 807, R.O.C.
(telephone: 886-7-381-4526ext7508, e-mail: cobol@ [Link]).
P. C. Lin is with the Information Management Department, Van Nung
Institute of Technology, Jung-Li, Taiwan 320, R.O.C. (telephone:
886-3-451-5811ext799, e-mail: lety@ [Link]).
Conventionally, the statistical tools, such as the Univariate
Time Series Methods [8] and Multivariate Methods [9], are
broadly used in the earning forecast analysis. The Univariate
Time Series Methods, such as the Simple Exponential Model,
Moving Average Model and the Simple Auto Regressive
Model, predict future earning, depending on the previous
earning statistics. These methods usually produce considerably
obvious predicting error, because the intrinsic quality and
competition of business changes. The Multivariate Methods
first select a suitable financial ratios set, pass them to each
analysis models, and finally produce the future earning forecast.
Among them, the process of selecting financial ratios set, such
as Factor Analysis Model, Stepwise Model, Forward Model,
Backward Elimination Model and ANOVA Analysis [10], is
critical and difficult in the final prediction results. Based on the
linear principle, these processes can not guarantee to produce
better prediction results because they are deeply affected by the
financial ratios selection sequence. Therefore, we will propose
an efficient and intelligent earning forecasting mechanism
when considering the smallest variables set.
Through cooperation and competition among the population,
the evolutionary computation algorithms often provide an
efficient optimization and nonlinear searching approach. Four
well-known examples are Genetic Algorithms (GA) [10],
Evolutionary Programming (EP) [16], Evolutionary Strategies
(ES) [17] and Genetic Programming (GP) [18]. Among them,
GA is suitable for relative feature selection problems [11]-[15].
Like other evolutionary technique, the Particle Swarm
Optimization (PSO) is a newer evolutionary technique, which
was first introduced by Eberhart and Kennedy [19]-[21]. It is
motivated from the simulation of social behavior. The discrete
binary version of PSO solves the problems in a space featuring
discrete and qualitative distinctions between variables [22].
Each particle flies in the D-dimensional search space and
dynamically adjusts its velocity based on its best flying
experience and global best flying experience. The best flying
experience is evaluated by the fitness function.
In this paper, we propose a hybrid swarm intelligence based
mechanism to combine the advantages of statistical
methodologies and Particle Swarm Optimization (PSO)
method, efficiently. The PSO algorithm provides the nonlinear
searching methodology and evolutionary computation
characteristics by adjusting the trajectories of a population of
particles through a problem spaces. Using the discrete binary
version of PSO, the introduced mechanism can intelligently
select the minimal critical financial ratios set, which produces
ICITA2004 ISBN 0-646-42313-4
E
193
Proceedings of the 2nd International Conference on Information Technology for Application (ICITA 2004)
near-accurate earning forecast. Our proposed mechanism can
more precisely predict the future trend by introducing the
Sliding Window Process. From the empirical results, its
proven that our introduced mechanism is superior in the
earning forecasting accuracy. Furthermore, the execution
efficiency and convergence speed are excellent if comparing
with other statistical methods and SWMR-MOGA, which is a
GA-based mechanism [23].
This paper is organized as follows. In Section II, we
introduce and overview the hybrid swarm intelligence based
mechanism in details. The design of proper fitness function
and encoding of particles position in D-dimension are
illustrated in Section III. The evaluation and prediction
processes in our proposed mechanism are also explained. In
Section IV, we will compare the results from our proposed
mechanism, other statistical methodology and GA-based
mechanism. Finally, conclusions and further research
directions are made to our proposed mechanism about the
earning forecast.
II. HYBRIDSWARMINTELLIGENCE BASED
MECHANISM
Figure 1. A Hybrid Swarm Intelligence Based Mechanism
The PSO is a population based optimization and search
method. It performs efficient nonlinear search in the problem
space. In this paper, we introduce a hybrid swarm intelligence
based mechanism shown in Figure 1 for earning forecast. For
combining the advantages of statistical methodology and the
evolutionary computation (PSO), our proposed mechanism are
divided into three processes: PSO Process, Sliding Window
Process and Multiple Regression Process. These processes are
described as follows.
A. PSO Process
The primary function of PSO process is to generate next
population. Each particle in the population is represented by
various combinations of selected financial ratios. The
generated population is passed to the Sliding Window Process
and Multiple Regression Process to evaluate the fitness value.
For satisfying more accurate earning prediction with minimal
financial ratios consideration, let F(MSSE, N) denote the
fitness function, where MSSE is the mean of SSE (Sum of
Square Error) of earning forecast generated from Multiple
Regression Process and N is the number of selected financial
ratios in each particle. Based on the fitness value, Adjusting
algorithm adjusts the velocity and position of each particle to
generate more suitable population.
B. Sliding Window Process
Each sliding window consists of Training Period and Test
Period shown in Figure 2. In this paper, the Training Period
length is 16Q (Quarter) and the Test Period length is 4Q. As
implied by the name, the data in the Training Period is used to
evaluate the coefficients in the Multiple Regression Process,
and the data in the Test Period is compared with the forecast
earning value. Applying the Sliding Window Process can
effectively enhance the prediction accuracy, because it
achieves the future trend
Figure 2. Sliding Window Simulation Process
C. Multiple Regression Process
Equation (1) is the multiple regression function used in the
Multiple Regression Process, where
^
ij
Y
denotes the forecasting
earning value in the j-th Quarter of Test Period in Window i.
Assume there are M Quarters in the Test Period and T sliding
window in Figure 2, sse
i
and MSSE is shown in Equation (2)
and (3), respectively.
D D ij
X b X b X b b Y + + + + =
2 2 1 1 0
( 1 )
_
=
=
M
j
ij
ij
i
Y Y sse
1
2
^
) (
( 2 )
_
=
=
T
i
i
sse
T
MSSE
1
1
( 3 )
III. PARTICLE SWARMALGORITHM
The Particle Swarm algorithm is an adaptive algorithm based
on the swarm cooperation and competition in the
social-psychological science. Each particle in the population is
encoded into D-dimensional space and each dimension is
represented as one binary bit shown in Figure 3. Each bit
represents a financial ratio, such as R308 (Book Value Per
Share), R408 (Total Growth Rate), R432 (Operating Income
Growth Rate), R612 (Fixed Asset Turnover), R835 (Operation
Income Per Employee) and so on shown in Table 1. In each bit,
i- th Generation
Populations
Sliding Window Process
Window Size
Multiple
Regression
Multiple
Regression
sse i
sse i +
Fitness
Function
Window Size
Selected
Financial
Ratio
+
sse: Sum of Square Error
MSSE: Mean of sse
Multiple Regression Process
MSSE
Adjusting
Algorithm
PSO Process
Test Training
16Q 4Q
Window 1:
Window 2:
Window T:
Test Training
16Q 4Q
Test Training
16Q 4Q
194
Proceedings of the 2nd International Conference on Information Technology for Application (ICITA 2004)
'1' and '0' means selected and non-selected to the corresponding
financial ratio, respectively.
Particle i R100 R308 R432 R835
Figure 3. Representation of a particle
Table 1. Particle Bit Description
Particle Bit Financial Ratio Description
0 R100 Return On Total Assets
1 R308 Book Value Per Share
2 R432 Operating Income Growth Rate
64 R835 Operation Income Per Employee
Applying the discrete binary version of PSO algorithm, the
particle swarm formula is shown as Equation (4), where c is an
arbitrary real number and rand() is a random number. A
logistic transformation S(vid) is used to accomplish the
modification, where S(vid) is a sigmoid limitation
transformation function, because S(vid) denotes the probability
of bit xid taking value 1. Therefore, the position adjustment is
defined in Equation (5).
( ) ( )
( ) x rand c x
x p x p v v
id gd id id id id
* () * =
+ + =
( 4 )
0
1 )) ( () (
=
= <
id
id id
x else
x then v S rand if
( 5 )
F(MSSE, N) is the fitness function in the particle swarm
algorithm, which is shown in Equation (6). Both and are
positive real number to represent the weights in the fitness
function. From Equation (6), it is obvious that the smaller
MSSE and N are preferred. It means that more precise earning
forecast with minimum financial ratio consideration are the
best.
( ) ( ) D N D MSSE N MSSE F + = 1 ) , (
( 6 )
IV. PERFORMANCES EVALUATION
In our proposed mechanism, the financial ratio selection is
based on the discrete binary version of PSO. To specify the
superiority of PSO, we compare the simulation results with
them from other selection techniques, such as the Stepwise and
Genetic Algorithm (GA) in Figure 4 - 7 and Table 5 - 9. The
Stepwise method is commonly used in the statistical
methodology and GA is a familiar evolutionary computation
methodology. We also compare the results with them from
univariate time series model, such as Simple Exponential
Model, Moving Average Model and Simple Auto Regressive
Model. The simulation program related our proposed
mechanism was written in Borland C++ Builder 5.0. The used
library related GA programming is Sugal 2.1. The stepwise
method was simulated in SPSS 8.0 package. All of the
financial data sources are from Taiwan Economic Journal Data
Bank (TEJ) and stored in Microsoft SQL 2000 server system.
All of the programs and packages are running on Windows
2000 server platform.
Currently, there are 531 companies are listed in Taiwan
Stock Exchange (TSE). After removing companies, which are
listed less than 10 years or have missing financial ratios, we
obtain 109 companies distributed among 17 industries. The
used parameters in GA and PSO are listed in Table 2 and Table
3, respectively.
Table 2. Parameters in GA Table 3. Parameters in PSO
Parameter Value Parameter Valu
e
Population Size 50 Population Size 50
Chromosome Size 65 Dimension Size 65
No. of Generation 2000 No. of Generation 2000
Selection Method Roulette Wheel Maximum Velocity (Vmax) 6.0
Crossover Method One-Point Acceleration Constant (c1) 2.0
Crossover Rate 0.6 Acceleration Constant (c2) 2.0
Mutation Rate 0.001
Table 4. The Descriptive Statistics of Sample Companies
|zq, Mean No. > Mean Max Min
Net sales (millions) 89.96 26 1035.65 1.07
Total assets (millions) 694.67 27 7410.03 23.38
Market value of equity (millions) 589.76 18 17493.45 6.63
Sales growth rate (%) 0.11 32 4.33 -0.52
Return on total assets (%) 0.16 50 12.16 -9.41
Current ratio (%) 1.63 34 7.93 0.20
Levarage (%) 0.45 50 0.45 0.22
Turnover (%) 2.05 34 10.99 0.25
The descriptive statistics of sample companies are shown in
Table 4. The simulation period is from the first quarter of 1991
to the last quarter of 1999. It means that 36 observations are
collected in each company. The training period is 16 quarters
and the test period is 4 quarters in each sliding window shown
in Figure 2. Finally, four quarters in 2000 (unseen data) is the
validation phrase, which verifies the forecasting accuracy by
each simulation method, respectively.
The partial results of financial ratios selected from Stepwise,
GA and PSO methods are shown in Table 5 - 7. In general, the
number of selected financial ratios is minimal, if using PSO
method, because their mean (
PSO
) and standard deviation (o
PSO
)
are 2.27 and 1.26, respectively. Let E
x
denote the
normalization of mean difference between real earning and
forecast earning shown in Equation (7), where x is the
predicting methods, such as Simple Exponential (Exp), Auto
Regressive (AR), Moving Average (MA), Stepwise (S),
Genetic Algorithm (GA), and Particle Swarm Optimization
(PSO). The E
Exp
, E
AR
, E
MA
, E
S
, E
GA
and E
PSO
are illustrated in
Table 8 and Figure 4 - 5. It is obvious that the mean and
195
Proceedings of the 2nd International Conference on Information Technology for Application (ICITA 2004)
standard deviation of E
PSO
(1.31, 4.30) are also minimal among
them. Furthermore, we define E/Nx shown in Equation (8) to
denote the E
x
to N
x
ratio, where N
x
is the number of financial
ratios selected by x method. The E/N
S
, E/N
GA
and E/N
PSO
are
illustrated in Table 9 and Figure 6. It is demonstrated again that
our proposed hybrid swarm intelligence based mechanism can
provide more accurate earning forecast with minimal number
of selected financial ratios consideration, because the mean and
standard deviation of E/N
PSO
is minimal (0.45, 0.97). Besides,
the average execution time using PSO and GA selection
method is 20.6 and 203 second. It means that the former is 10
times of the latter, approximately. The normalized fitness value
restricted in [0,1] is shown in Figure 7. It is evident that the
convergence speed and final fitness value after 2000
generations of using PSO algorithm is better than that of using
GA algorithm.
Earning Real
Earning Real - Earning Forecast
E
x
=
( 7 )
x
x
x
N
E
N E = /
( 8 )
Table 5. The financial ratios selected by Stepwise (partial)
Companies Ratios Numbers
1 R609 R614 2
2 R513 R108 R411 R403 4
3 R613 R206 2
4 R608 R303 2
5 R405 R308 R410 3
109 R502 R304 R608 R201 4
Mean (S) 3.1927
StdDev (oS) 1.7872
Maximum (MaxS) 9
Minimum (MinS) 1
Table 6. The financial ratios selected by GA (partial)
Companies Ratios Numbers
1 R107 R201 R205 3
2 R306 R507 R534 R537 4
3 R103 R302 R401 3
4 R531 R534 2
5 R205 R206 R302 R410 4
109 R308 R401 2
Mean (GA) 3.0000
StdDev (oGA) 1.2981
Maximum (MaxGA) 9
Minimum (MinGA) 1
Table 7. The financial ratios selected PSO (partial)
Companies Ratios Numbers
1 R411 R422 2
2 R507 R512 R514 3
3 R302 R433 2
4 R306 1
5 R303 R431 2
109 R100 R108 2
Mean (PSO) 2.2752
StdDev (oPSO) 1.2610
Maximum (MaxPSO) 9
Minimum (MinPSO) 1
Table 8. The normalization of mean difference between real earning and
forecast earning
Industries (companies) EExp EMA EAR ES EGA EPSO
Cement (6) 2.35 1.64 14.13 2.63 0.94 1.85
Foods (8) 0.65 1.19 2.93 1.23 0.61 0.53
Plastics (9) 2.98 3.44 3.76 5.71 1.02 0.87
Textile (22) 3.07 3.66 11.29 2.31 2.13 1.85
Electric and Engineering (4) 7.00 4.43 8.54 3.95 2.98 7.43
Electric Wire and Cable (8) 3.02 3.55 5.34 2.26 1.80 0.53
Chemical (11) 0.42 1.29 3.22 0.97 0.33 0.33
Glass (2) 0.11 0.14 1.36 0.85 0.09 0.07
Pulp and Paper (4) 1.70 1.82 1.71 1.43 0.51 0.55
Iron and Steel (5) 14.76 17.14 25.47 3.25 6.71 4.76
Rubber (5) 0.67 0.92 2.51 2.18 0.76 0.22
Motor (2) 0.35 0.21 1.16 0.62 0.16 0.06
Electronics (14) 2.18 1.92 4.21 4.71 0.69 0.92
Construction (2) 3.03 2.56 5.37 1.40 0.35 0.03
Transportation (1) 1.48 1.58 1.85 2.82 0.97 1.38
Hotel (3) 12.22 7.34 41.11 13.05 9.64 1.82
Merchandise (3) 1.07 1.44 2.89 0.92 0.42 0.19
StdDev 8.32 8.74 23.68 6.53 4.95 4.31
Mean 2.92 3.06 7.71 2.91 1.57 1.31
Maximum 53.60 66.00 189.21 53.50 33.78 29.26
Minimum 0.01 0.04 0.03 0.00 0.00 0.00
Table 9. E/Nx ratio of Stepwise, GA, PSO in each industries
196
Proceedings of the 2nd International Conference on Information Technology for Application (ICITA 2004)
Industries (companies) E/NS E/NGA E/NPSO
Cement (6) 0.95 0.28 1.11
Foods (8) 0.91 0.21 0.75
Plastics (9) 2.00 0.55 0.39
Textile (22) 0.96 0.97 0.52
Electric and Engineering (4) 1.06 0.43 1.38
Electric Wire and Cable (8) 0.68 0.47 0.34
Chemical (11) 0.38 0.13 0.21
Glass (2) 0.15 0.09 0.02
Pulp and Paper (4) 0.48 0.21 0.54
Iron and Steel (5) 1.40 2.23 1.26
Rubber (5) 0.65 0.29 0.10
Motor (2) 0.11 0.04 0.02
Electronics (14) 0.99 0.32 0.22
Construction (2) 0.42 0.15 0.02
Transportation (1) 1.20 0.20 0.03
Hotel (3) 6.50 2.45 0.51
Merchandise (3) 0.41 0.29 0.10
StdDev 2.16 1.94 0.97
Mean 1.06 0.56 0.45
Maximum 11.79 16.89 5.85
Minimum 0.00 0.00 0.00
0
2
4
6
8
10
12
14
C
e
m
e
n
t (6
)
F
o
o
d
s
(8
)
P
la
st
ic
s
(
9
)
T
e
x
t
ile
(
2
2
)
E
le
c
tr
ic
a
n
d
E
n
g
in
e
e
r
in
g
(
4
)
E
le
c
tr
ic
W
ir
e
a
n
d
C
a
b
le
(
8
)
C
h
e
m
ic
a
l (1
1
)
G
la
s
s (2
)
P
u
lp
a
n
d
P
a
p
e
r
(
4
)
I
r
o
n
a
n
d
S
te
e
l
(5
)
R
u
b
b
e
r
(5
)
M
o
to
r
(2
)
E
le
c
tr
o
n
ic
s
(
1
4
)
C
o
n
s
t
r
u
c
tio
n
(2
)
T
r
a
n
s
p
o
r
t
a
t
io
n
(1
)
H
o
te
l (3
)
M
e
r
c
h
a
n
d
is
e
(3
)
Industries
[$
[(
[[$( E
E
S
E
GA
E
PSO
Figure 4. ES, EGA and EPSO among each industry.
0
5
10
15
20
25
30
35
40
45
C
e
m
e
n
t (6
)
F
o
o
d
s
(8
)
P
la
s
tic
s
(9
)
T
e
x
tile
(2
2
)
E
le
c
tr
ic
a
n
d
E
n
g
in
e
e
r
in
g
(4
)
E
le
c
tr
ic
W
ir
e
a
n
d
C
a
b
le
(8
)
C
h
e
m
ic
a
l
(1
1
)
G
la
s
s
(2
)
P
u
lp
a
n
d
P
a
p
e
r
(4
)
I
r
o
n
a
n
d
S
te
e
l (5
)
R
u
b
b
e
r
(5
)
M
o
to
r
(2
)
E
le
c
tr
o
n
ic
s
(1
4
)
C
o
n
s
tr
u
c
tio
n
(2
)
T
r
a
n
s
p
o
r
ta
tio
n
(1
)
H
o
te
l (3
)
M
e
r
c
h
a
n
d
is
e
(3
)
Industries
E
[[p
[,
[,[
Figure 5. EExp, EAR and EMA among each industry.
0
1
2
3
4
5
6
7
C
e
m
e
n
t (
6
)
F
o
o
d
s
(
8
)
P
la
s
tic
s
(9
)
T
e
x
t
ile
(2
2
)
E
le
c
t
r
ic
a
n
d
E
n
g
in
e
e
r
in
g
(4
)
E
le
c
t
r
ic
W
ir
e
a
n
d
C
a
b
le
(
8
)
C
h
e
m
ic
a
l (
1
1
)
G
la
s
s
(
2
)
P
u
lp
a
n
d
P
a
p
e
r
(4
)
I
r
o
n
a
n
d
S
t
e
e
l (
5
)
R
u
b
b
e
r
(5
)
M
o
to
r
(2
)
E
le
c
t
r
o
n
ic
s
(1
4
)
C
o
n
s
tr
u
c
tio
n
(
2
)
T
r
a
n
s
p
o
r
t
a
tio
n
(
1
)
H
o
te
l (
3
)
M
e
r
c
h
a
n
d
is
e
(
3
)
Industries
E/N
[ $
[ (
[ $[(
Figure 1. E/NS, E/NGA and E/NPSO among each industry.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 51 101 151 201 251 301 351 401 451 501
Generations
N
o
r
m
a
l
i
z
a
t
i
o
n
o
f
F
i
t
n
e
s
s
V
a
l
u
eGA Fitness
PSO Fitness
Figure 2. The Normalization of Fitness Value by using GA and PSO selection
method in our mechanism.
V. CONCLUSIONS
The earning forecast is very important in the investment
decision, because it reflects future growth of business. The
conventionally broadly used statistic tools, such as the
Univariate Time Series Methods and Multivariate Method, can
not make more precise prediction without considering
sufficient critical financial ratios. Simultaneously, the
forecasting accuracy is seriously influenced by the linear
searching sequence. On the other hand, it is very hard to
predict future earning trend if only using past information,
because many situations change everyday, such as global
economic environment, industry prospects, and etc.
Combining the advantages of evolutionary computation and
statistics tools, we propose a hybrid swarm intelligence based
E
EXP
E
MA
E
AR
E/Ns
E/N
GA
E/N
PSO
197
Proceedings of the 2nd International Conference on Information Technology for Application (ICITA 2004)
mechanism to forecast a near-optimal earning accuracy when
considering minimal financial ratios.
In this paper, we compare the proposed mechanism with purely
statistic tools, such as Simple Exponential (Exp) model,
Moving Average (MA) model, Simple Auto Regressive (AR)
model and Stepwise (S) method. Furthermore, we also
compare the performances with SWMR-MOGA mechanism,
which combines the advantages of GA and statistic model [23].
Currently, GA is a familiar evolutionary computation technique,
which is commonly used in various nonlinear search problems.
From Table 8 and Figure 4 - 5, the mean and standard deviation
of earning forecast error using evolutionary computation is
better than that using statistic methodology, generally. In
particular, our proposed hybrid swarm intelligence based
mechanism can provide the most accurate earning forecast.
From Table 5 - 7, using PSO method also provides minimal
number of the selected critical financial ratios, when comparing
with other methodologies. Therefore, its obvious that our
proposed mechanism provides superior earning forecast
accuracy, minimal critical financial ratios consideration and
minimum E/N ratio, which is shown in Table 9 and Figure 6.
Besides, the average execution time in our proposed
mechanism is approximately 10 times faster than that in
SWMR-MOGA mechanism. From the normalized fitness
value restricted in [0,1] in each generation shown in Figure 7, it
is evident that the convergence speed of our mechanism is
faster than that of SWMR-MOGA mechanism. After 2000
generations, the final fitness value of using PSO algorithm is
better than that of using GA algorithm.
REFERENCES
[1] K. Schipper, Commentary on Analysts's Forecasts, Accounting
Horizons, pp. 105-121, 1991.
[2] J. L. Callen, Neural network forecasting of quarterly accounting
earnings, International Journal of Forecasting, pp. 475-482, 1996.
[3] L. D. Brown, Earnings Forecasting Research: Its Implications for Capital
Markets Research, International Journal of Forecasting, pp. 295-320,
1993.
[4] B. Mark, D. B. Messod and G. W. Susan, Whisper Forecasts of Quarterly
Earnings Per Share, Journal of Accounting and Economics, pp. 27-50,
1999.
[5] S. P. Bandyopadhyay, L. D. Brown and G. D. Richardson, Analysts's
Use of Earnings Forecasts in Predicting Stock Returns: Forecast Horiaon
Effects, International Journal of Forecasting, pp. 429-445, 1995.
[6] L. Steven and M. Vivek, Financial Analysts' Earnings Forecasts and
Insider Trading, J. of Accounting and Public Policy, pp. 233-261, 1995.
[7] P. A. Williams, G. D. Moyes and K. Park, Factors Affecting Earnings
Forecast Revisions for the Buy-side and Sell-side Analsyst, Accounting
Horizons, pp. 112-121, 1996.
[8] J. F. Frank, Investment Management, Prentice Hall, 1999.
[9] D. E. Johnson, Applied Multivariate Methods for Data Analysts,
Duxbury Press, 1998.
[10] J. Holland, Adaptation in Natural and Artificial Systems: An Introductory
Analysis with Applications to Biology, Control and Artificial Intelligence,
MIT Press, 1975.
[11] H. Aluallim and T. G. Dietterich, Efficient Algorithms for Identifying
Relevant Features, proc. of Canadian Conference on Artificial
Intelligence, pp. 38-45, 1992.
[12] A. Jain and D. Zongker, Feature Selection: Evaluation, Application, and
Small sample Performance, IEEE Transaction on Pattern Analysis and
Machine Intelligence, pp. 153-158, 1997.
[13] H. Liu and R. Setiono A Probabilistic Approach to Feature Selection: A
Filter Solution, Proc. of 13th International Conference on Machine
Learning, pp. 319-327, 1996.
[14] C. Emmanouilidis, A. Hunter, J. MacIntyre and C. Cox,
Multiple-Criteria Genetic Algorithms for Feature Selection in Neurofuzzy
Modeling, International Joint Conf. on Neural Networks, pp. 4387-4392,
1999.
[15] M. Srinivas and M. P. Lalit, Genetic Algorithms A Survey, IEEE
Computer, pp. 18-20, 1994.
[16] L. J. Fogel, Evolutionary Programming in Perspective: the Top-down
View, in Computational Intelligence: Imitating Life, J. M. Zurada, R. J.
Marks, and C. J. Robinson, Eds., IEEE Press, Piscataway, NJ, 1994.
[17] I. Rechenberg, Evolutionary Strategy, in Computational Intelligence:
Imitating Life, J. M. Zurada, R. J. Marks II, and C. J. Robinson, Eds.,
IEEE Press, Piscataway, NJ, 1994.
[18] J. R. Koza, Genetic Programming: On the Programming of Computers
by Means of Natural Selection, MIT Press, Cambridge, MA, 1992.
[19] R. C. Eberhart and J. Kennedy, A New Optimizer Using Particle Swarm
Theory, Proc. 6th International Symposium on Micro Machine and
Human Science, Nagoya, Japan, IEEE Service Center, pp. 39-43, 1995.
[20] J. Kennedy and R. C. Eberhart, Particle Swarm Optimization, Proc. of
IEEE International Conf. on Neural Networks, Perth, Austrila, IEEE
Service Center, pp. 1942-1948, 1995.
[21] J. Kennedy, The Particle Swarm: Social Adaptation of Knowledge,
Proc. of IEEE International Conf. on Evolutionary Computation,
Indianapolis, Indiana, IEEE Service Center, pp. 303-308, 1997.
[22] J. Kennedy and R. C. Eberhart, A Discrete Binary Version of the Particle
Swarm Algorithm, Proc. of International Conf. on Systems, Man, and
Cybernetics, IEEE Service Center, 1997.
[23] J. S. Chen and P. C. Lin, An Intelligent Financial Ratio Selection
Mechanism for Earning Forecast, accepted by J. of Oper. Res. Social
Japan, 2002.
198
