COMPARISON OF WAVELET NETWORK AND LOGISTIC REGRESSION IN PREDICTING ENTERPRISE FINANCIAL DISTRESS

International Journal of Computer Science & Information Technology (IJCSIT) Vol 7, No 3, June 2015
DOI:10.5121/ijcsit.2015.7307 83
COMPARISON OF WAVELET NETWORK AND
LOGISTIC REGRESSION IN PREDICTING
ENTERPRISE FINANCIAL DISTRESS
Ming-Chang Lee1
and Li-Er Su2
1
National Kaohsiung University of Applied Sciences
2
Shih Chien University, Kaohsiung Campus, Kaohsiung Taiwan
ABSTRACT
Enterprise financial distress or failure includes bankruptcy prediction, financial distress, corporate
performance prediction and credit risk estimation. The aim of this paper is that using wavelet networks in
non-linear combination prediction to solve ARMA (Auto-Regressive and Moving Average) model problem.
ARMA model need estimate the value of all parameters in the model, it has a large amount of computation.
Under this aim, the paper provides an extensive review of Wavelet networks and Logistic regression. It
discussed the Wavelet neural network structure, Wavelet network model training algorithm, Accuracy rate
and error rate (accuracy of classification, Type I error, and Type II error). The main research opportunity
exist a proposed of business failure prediction model (wavelet network model and logistic regression
model). The empirical research which is comparison of Wavelet Network and Logistic Regression on
training and forecasting sample, the result shows that this wavelet network model is high accurate and the
overall prediction accuracy, Type Ⅰerror and Type Ⅱ error, wavelet networks model is better than
logistic regression model.
KEYWORDS
Wavelet Networks; Logistic Regression; Business Failure Prediction; Type I error; Type Ⅱerror.
1. INTRODUCTION
Business Failure Prediction (BFP) can help to avoid investing in business likely to fail.
Bankruptcy prediction model enhance the decision support tool and improve decision making
[33]. It uses statistical analysis and data mining technique. Statistical business failure prediction
models attempt to predict the business’ failure or success. The Multiple discriminant analysis
(MDA) has been the most popular approaches [6, 13]. There have a large number of alternative
techniques available in this MAD model [2, 9, 22, 33]. The data mining techniques include
decision tree, support vector machine (SVM) [10], neural networks (NNs) [11], fuzzy system,
rough set theory and genetic algorithm (GA) [33]. Various researches have demonstrated the
artificial intelligence (AI) techniques serve as a useful tool bankruptcy prediction such as artificial
neural networks (ANNs) [30].
The following method or model using for prediction in time series are Box-Jenkins method [28],
Grey forecasting model [11], artificial neural networks model [1, 4, 16], Logistic Regression [25,
30], ARMA Model [21]. In Box-Jenkins method, it is difficult to determine the model features;
it is also difficult to identify no-steady state. It is not need to calculate the statistical characters in
grey forecasting model or neural network model. But, grey forecasting model is suitable for
solve exponential growth practical problems. The network modelling approach in artificial neural

84
networks model is difficult to determine the network structure by using scientific application. In
training and learning state, BP model calculate the optimal weight. This model trapped in a local
minimum which will affect the reliability and accuracy of the model. ARMA Model provides
one of the basic tools in time series modelling. This method is valid in pure AR model and pure
MA model. But, it is difficult for identification in mixed ARMA model. Therefore, using
wavelet network clustering capabilities, the input time series sample autocorrelation function
(SACF) value and partial autocorrelation function (SPACF) value, the output is the identification
of ARMA model. Many researchers [20, 29, 36] discussed the ability of nonlinear approximation
of wavelets and neural network models. Some researches [31, 32, 34] studied fuzzy wavelet
neural network models for prediction and identification of dynamical systems.
The main contribution of this paper is to propose a financial distress prediction method based on
Wavelet Network model and logistic regression model. The financial and non-financial ratios
were used to enhance the accuracy of the financial distress prediction model. The dimension of
the inputs was reduced using a principal component analysis (PCA) method, and then Wavelet
Network model were used to predict the financial distress.
2. LOGISTIC REGRESSION MODEL AND WAVELET NETWORK
MODEL
2.1. Logistic regression model
Defines the probability )(xπ and 1- )(xπ , these probabilities are written in the following form:
),...,,1()( 21 nXXXYPx ==π
),...,,0()(1 21 nXXXYPx ==−π (1)
This model of the ln
)(1
)(
x
x
π
π
−
is:
∑+=
−
=
=−
=
=
n
j
jj
n
n
X
x
x
XXXYP
XXXYP
1
0
21
21
)(1
)(
ln
),...,,1(1
),...,,1(
ln ββ
π
π
(2)
Using the inverse of the Logit transformation of (2), it obtains at the following:
∑
=
+−∑
=
+
∑
=
+
+
=
+
== n
j
jXj
n
j
jXj
n
j
jXj
n
ee
e
XXXYP
1
)0(
1
0
1
0
21
1
1
1
),...,,1(
ββββ
ββ
(3)
(3) is a logistic regression model, the conditional mean is between 0 and 1.
2.2. Maximum Likelihood Estimate (MLE)
The maximum likelihood is the method of parameter estimation in logistic regression model. For
a set of observations in the data ),( ii yx , the following equation results for the contribution to the
likelihood function for the observation ),( ii yx is )( ixα :
]1)(1[)()( iy
ixiy
ixix −−= ππα (4)
The observations are assumed to be independent of each other so it can multiply their likelihood
contributions to obtain the complete likelihood function )(βl . The result is given in (5).
Where β is the collection of parameters ( nβββ ,...,, 10 ) and )(βl is the likelihood function of β .

85
∏
=
=
n
i
ixl
1
)()( αβ (5)
)(BL is denoted the log likelihood expression.
)])(1ln[)
1
1()](ln[()](ln[)( ixiy
n
i
ixiylL ππββ −∑
=
−+== (6)
It employs the techniques of calculus to determine the value of β based on maximum of )(βL . It
takes the derivate of )(βL with respect to nβββ ,...,, 10 and setting the resulting derivatives equal
to zero. These equations are called likelihood estimations, and there is n +1 equation.
2.3. Hypothesis testing
The hypothesis testing that a coefficient on an independent variable is significantly different from
zero use Wald statistic. The Wald statistic for the β coefficient is:
=2χ Wald = 2]/[ βσβ (7)
This is distributed chi-square with 1 degree of freedom.
2.4. Wavelet network model t
Wavelet analysis is a new technique designed for multi scale analysis in the time series
modelling. Wavelet analysis utilizes the wavelet basis function to approximate and extract the
features of interests from the original data. If the function )(tϕ satisfies the following condition
is called wavelet basis function.
∫
∝ <∝0
)(
dw
w
wϕ
(8)
Where )(wϕ is the Fourier transform of )(tϕ in the frequency domain. Variable t can actually
have other unit [12].
The wavelet basis function )(
1
)(,
a
bt
a
tba
−
= ϕϕ (9)
Where ‘a’ refers the scale (dilated) coefficient, ‘b’ refers the translation coefficient.
Wavelet networks employ activation function that are dilated and translated coefficient of a single
function RdR →:ϕ , where d is the input dimension [36]. This function called the ‘mother
wavelet’ is localized both in the space and frequency domains [7]. In order to increase
convergence speed, the wavelet neural network shows surprising effectiveness in solving the
conventional problem of poor convergence or even divergence encountered in other kind of
neural network [38].
Wavelet network model is feed forward network model; the wavelet basis function is neuronal
activation function. The basic strategy is changing the shape and scale wavelet basis, adjusting
the network weights and threshold value to take the error function minimization. Figure 1 is
Wavelet neural network structure.
Three layers of the wavelet neural network are input layer, hidden layer and output layer. Each
layer is fully connected to the nodes in the next layer.

86
}{ ix : The Input value of the sample, Mi ,..,2,1= .
}{ py : The output value of the sample p.
kw : The weight of linking hidden layer with output layer, Kk ,..,2,1=
The output value of the sample p in network:
]
1 1
)([ ∑
=
∑
=
−
=
K
k
M
i ka
kbi
ixkwpy ϕσ : ))exp(1(1)( tt −+=σ (10)



<
>
=
5.0)(0
5.0)(1
t
t
y
σ
σ
x1 ○ )
1
12
(
a
b−
ϕ
x2 ○ )
1
1(
a
bM −
ϕ w1
y
)
1
(
ka
kb−
ϕ wk
xM ○ )
2
(
ka
kb−
ϕ
)(
ka
kbM −
ϕ
Figure 1: Wavelet neural network structure
The minimum mean error function:
∑
=
−=
P
p
pypdE
1
2){
2
1
(11)
Where
:pd The output of expected classification of the sample p, :pd = 1 or 0
M: The number of input layer.
K: The number of hidden layer
P: The number of sample
2.5. Wavelet network model training algorithm
The training algorithm for a wavelet neural network is as follows [35].
(1) Select the no of hidden nodes required. Initialize parameters kwandkbka ,, with random
number in ].1,1[− Give objective mean error maxE . The Mother wavelet is has the following
form
)22exp()75.1cos()( ttt −=ϕ (12)
]22exp()]75.1sin(75.1)75.1cos([ tttt
dt
d
−+−=
ϕ
)
1
11
(
a
b−
ϕ
Σ
Σ
σ
Σ

87
When taken as Gaussian wavelet it become
)22exp()( tt −=ϕ (13)
(2) Input training sample }{ ix and the objective output }{ pd . Calculate the output }{ py and he
minimum mean error function E.
(3) Adjust kwandkbka ,, for reducing the error of production. Use the conjugate gradient
descend algorithm is employed:
)()()1( tkw
kw
E
tkwtkw ∆+
∂
∂
+=+ αη
)()()1( tka
ka
E
tkatka ∆+
∂
∂
+=+ αη (14)
)()()1( tkb
kb
E
tkbtkb ∆+
∂
∂
+=+ αη
ka
kbi
t
−
='
Where
∑
=
∑
=
−−=
∂
∂ P
p
M
i
tixtpypd
kw
E
1 1
)'()'(')( ϕσ
=
∂
∂
ka
E
)'
1 1
()'()( t
P
p
M
i ka
t
ka
ixkwpypd∑
=
∑
= ∂
∂
∂
∂
−− σϕ (15)
−=
∂
∂
kb
E
)'(
1 1
)'()( t
P
p
M
k kb
t
kb
kw
p
ky
p
kd σϕ∑
=
∑
= ∂
∂
∂
∂
−
Where η and α are the learning and the momentum rates respectively.
(4) Return to step (2) the process is continued until E satisfies the give error criteria, and the
whole training of wavelet neural network is completed.
2.6. Accuracy rate and error rate
Since Type I error only creates a lost opportunity cost from not dealing with a successful
business, for example, missed potential investment gains [17], therefore Type I error is more
important than Type II error. The objectives of predictive of accuracy should be to reduced Type I
error while keep Type II error. Table 1 denotes as robustness of model. The measure used for
calculating the accuracy of classifying distressed companies and no-distressed companies have:
Type I error =
BA
B
+
Type II error =
DC
C
+
E )/( BAA +=
F )/( DCD +=
G )/()( DCBADA ++++=

88
Where
Type I error: refer to the situation when actual failure company is classified as non
failures company
Type II error: refer to the situation when actual non failure company is classified as
Failures Company.
C: the number of Type I error that is the number of distressed companies in the sample
based on actual observation that was misclassified as a non-distressed company.
B: the number of Type II error that is the number of non-distressed companies in the
sample based on actual observation that was misclassified as a distressed company.
A: the number of non-distressed accurately classified by the models
D: the number of distressed accurately classified by the models
E: The accuracy of classification of non-distressed company
F: The accuracy of classification of distressed company
G: The oval accuracy of classification
Table 1: Robustness of model
No-distressed
Observed value
Non-distressed
company
Distressed company Accuracy of
Classification
Non-distressed company A B E
Distressed company C D F
Overall accuracy of
classification
G
3. EMPRICAL RESEARCH
We random select in the stock market listed company's traditional manufacturing in Taiwan
to analysis the wavelet network model prediction. The steps of business failure prediction model
(Wavelet network model and logistic regression model) are:
Step 1: Financial ratio of dataset
Step 2: Identification of independent variables
Step 3: Reduced the number of financial ratio
(a) Kolmogorov-Smirnov test (K-S test)
(b) Wilcoxon test
(c) Principal component analysis
Step 4: Building the wavelet networks model
(a) Wavelet networks model
(b) Training for a wavelet neural network
(c) Robustness of model in prediction accuracy
Step5: Building Logistic regression model
(a) Logistic regression model
(b) Training for Logistic regression
(c) Robustness of model in prediction accuracy
Step 6: Comparison of Wavelet Network and Logistic Regression
3.1. Financial Ratio of Dataset
A dataset of covariates used in this study includes a combination of financial ratios and market
variables [15]. Some researches such as [3, 8, 9, 23~24, 26~27, 39] have been widely used
financial ratios in explaining the possibility of business financial distress. Gibson [14] suggests
five factors (Profitability, Liquidity, Leverage, Efficiency, and Valuation ratio) for evaluation enterprise
financial failure. Table 3 is the 15 rations selected in this study.

89
Table 3: The 15 rations selected in this study
Category Covariate Code Definition
x1 Profitability EBIT margin EBT EBIT/operating revenue
x2 Return on Equity ROE Net income/ Total equity
x3 Return on Assets ROA Net income/ Total assets
x4 Liquidity Current ratio CUR Current assets/ current liabilities
x5 Quick ration QUK Quick assets/ current liabilities
x6 Leverage Debt ratio DET Total liabilities/ Total assets
x7 Debt to Equity ratio DER Total liabilities/ Total equity
x8 Efficiency Fixed Asset turnover FAT Revenue/Asset
x9 Capital turnover CAT Operating revenue/ operating invest
capital
x10
Valuation ratio
Price to Sales ratio PSR Stock price per share/ Sales per share
x11 Price earnings ratio PER Stock price per share/Earnings per
share
x12 Price to book value PBV Stock price per share/Equity per
share
x13 Growth ability Operating Profit grew OPG Operating Profit grew rate
x14 Net profit grow rate NPG Net profit grow rate
x15 Inventory turning rate ITR Sale/ Inventory
3.2. Data collection and Sample
This paper random select sample listed companies from 2006 to 2009 for estimating sample.
3.3. Reduced the number of financial ratio
There are three ways to reduce the large number of financial indicators (1) Financial indicators
normal distribution test (Kolmogorov-Smirnov test (K-S test)) (2) Paired samples’ significant test
in financial indicators (Wilcoxon test) (3) Principal component analysis (PCA). PCA is a
statistical procedure that uses an orthogonal transformation to convert a set of observations of
possibly correlated variables into a set of values of linearly uncorrelated variables called principal
components. The number of principal components is less than or equal to the number of original
variables.
(1) Kolmogorov-Smirnov test (K-S test)
The K-S test has the advantage of making no assumption about the distribution of data. Suppose
that an Independent and Identically Distributed (i.i.d.) sample nxxx ,...2,1 with some unknown
distribution p and we would like to test the hypothesis that p is equal to a particular (normal)
distribution 0
p , i.e. decide between the following hypotheses:
0
H : 0
pp = , :1H 0
pp ≠ . We know that testing this hypothesis use chi-squared goodness-of-
fit test.
(2) Wilcoxon test (Wilcoxon signed-rank test)
The Wilcoxon signed-rank test is a non-parametric statistical hypothesis test used when
comparing two related samples, matched samples, or repeated measurements on a single sample
to assess whether their population mean ranks differ (i.e. it is a paired difference test). Wilcoxon
test used as (1) an alternative to the paired Student's t-test, (2) t-test for matched pairs, (3) the t-
test for dependent samples when the population cannot be assumed to be normally distributed.
(3) Principal component analysis (PCA)

90
The main purposes of a PCA are (1) the analysis of data to identify patterns (2) finding patterns to
reduce the dimensions of the dataset with minimal loss of information. In the table of principal
component variance, it select no of components in this data testing, according to the cumulative
variance above to 95%.
3.4. Building the wavelet networks model
In wavelet network model, the numbers of neurons of the input and output layers are respectively
equal to the number of variables presented simultaneously to the network and to the number of
variables that the neural network must estimate. Wavelet network model the number of principal
components is equal to the number of input layer neuron. The number of output layer neuron is
1. The output neuron takes the value 0 if the company is healthy (no-financial distress) and 1 if it
is bankrupt (financial distress). To determine the number of hidden layer node is by researcher.
After determining the network architecture, its implementation requires two data samples. The
first set of data is used for the training, the second for the validation. The neural network is been
realized using Matlab software. Model input values of paired samples are the score value of
principal components.
3.5 Building Logistic regression model
Logistic regression model input values of paired samples are the score value of principal
components. The dependent variable in logistic regression is usually dichotomous, that is, the
dependent variable can take the value 1 with a probability of success )( iyp , or the value 0 with
probability of failure )(1 iyp− .
The Logistic regression is been realized using Spss21.0 software and the Wald statistic for test the
βˆ coefficient.
The predict probability of occurrence )( iyp given as:
)]}ˆ....11
ˆ
0
ˆ(exp[1{1)( mxmxiyp βββ +++−+=
The output 5.0)( >iyp takes the value 0, the company is healthy (no-financial distress) and
5.0)( <iyp takes the value 1, and the company is bankrupt.
3.6. Robustness of model in prediction accuracy
It easily calculates the accuracy rate and Error rate. Type I error is
BA
B
+
% and Type II error is
DC
C
+
% (See Table 1).
4. EMPIRICAL ANALYSIS
4.1. Data collection and Sample
This paper select sample listed companies from 2006 to 2009 for training sample. There select 40
paired samples for analysis. It randomly selected 20 paired sample listed companies from 2009 to
2012 for forecasting sample.

91
Through the kolmogorov- Siminor test (using SPSS20.0 software), no financial indicator on
paired sample is normal distribution. Through wilcoxon test, ten variables are selected as
potential predictor variables: EBIT margin (x1), Return on Equity (x2), Return on Assets (x3),
Current ratio (x4), Quick ration (x5), Fixed Asset turnover (x8), Price earnings ratio (x11), Price to
book value (x12), Operating Profit grew (x13), Net profit grow rate (x14). The paired financial
indicators sample wilcoxon test denotes as Table 4.
Table 4: The paired financial indicators sample wilcoxon test
Financial
indicator
Statistic Z-
value
2-taied test p-value Result
X1 (EBIT) -2.012 0.042 *
X2 (ROE) -3.325 0.001 *
X3 (ROA) -4. 855 0.000 *
X4 (CUR) -4.121 0.000 *
X5 (QUR) -1.968 0.045 *
X6 (DET) -1.763 0.084
X7 (DER) -0.213 0.812
X8 (FAT) -2.164 0.032 *
X9 (CAT) -1.357 0.160
X10 (PSR) -1. 656 0.085
X11 (PER) -3.902 0.002 *
X12 (PBV) -1.993 0.047 *
X13 (OPG) -2.133 0.034 *
X14 (NPG) -3.258 0.001 *
X15 (ITR) -1.885 0.068
* Significant at 5 percent level
4.2. Principal component analyse
PCA is statistical procedure that uses an orthogonal transformation to convert a set of
observations of possibly correlated variables into a set of values of linearly uncorrelated variables.
It uses SPSS 21.0 software on PCA. It gets 8 principal components (z1, z2, z3, z4, z5, z6, z7, z8), it
contains the original 10 financial indicators 97.82% (see table 5).
Tab 5: The principal component Variance
Principal
components
Eigenvalue Variance (%) Cumulated Variance
(%)
z1 3.165 31.642 31.642
z2 1.315 13.043 44.685
z3 1.088 10.863 55.548
z4 0.935 9.347 64.895
z5 0.921 9.225 74.120
z6 0.834 8.336 82.456
z7 0.781 7.712 90.168
z8 0.753 7.652 97.820
Z9 0.155 1.540 99.760
Z10 0.065 0.640 100.00
The component score express as:
*
1414
*
1313
*
1212
*
1111
*
88
*
55
*
44
*
33
*
22
*
11 xiaxiaxiaxiaxiaxiaxiaxiaxiaxiaiz +++++++++=

92
Where ,8....,,2,1=i
is
ixix
x
−
=*
1 , :ix the mean of financial indicator, :is the standard
deviation of financial indicator. Table 6 is denoted as the component scored matrix.
For example,
*
14176.0*
13085.0*
12052.0*
11034.0*
8042.0*
5056.0*
4615.0*
3538.0*
203.0*
1068.01 xxxxxxxxxxz −−−−−−++−−=
Table 6: The component scored matrix.
Principal
component
1z 2z 3z 4z 5z 6z 7z 8z
*
1x (EBIT)
-0.068 0.008 0.054 -0.013 0.046 0.017 0.072 0.027
*
2x (ROE)
-0.03 -0.011 -0.112 -0.042 -0.006 0.003 -
0.003-
-0.041
*
3x (ROA)
0.538 -0.036 -0.035 -0.058 -0.024 -0.036 -0.012 -0.124
*
4x (CUR)
0.615 -0.010 -0.042 -0.042 -0.038 -0.035 -0.003 -0.127
*
5x (QUR)
-0.056 0.040 0.019 0.016 0.011 0. 986 0.006 -0.011
*
8x (FAT)
-0.042 0.032 0.012 -0.043 0.0031 0.036 0.052 0.053
*
11x (PER)
-0.034 0.056 0.079 0.009 -0.006 0.002 0.912 -0.019
*
12x (PBV)
-0.052 0.031 0.048 -0.007 0.589 0.017 -0.054 -0.078
*
13x (OPG)
-0.085 -0. 042 -0.012 0.831 -0.006 0. 015 0.008 -00.94
*
14x (NPG)
-0.176 0.055 0.024 -0.017 -0.068 -0.011 -0.017 5.015
4.3. Wavelet network model training and forecasting
The wavelet network training and forecasting use Mat lab software. Vector
)8,7,6,5,4,3,1,1( zzzzzzzz is input of this model. Model input values of paired samples are 40,
the input layer neuron is 8. The hidden layer neuron is 6 and the output layer neuron is 1. Set
minimum mean error function 001.0=E , 1.0=η , 01.0=α . The paired samples are 40 for training
test. Predicted results denoted as table 7.
Table 7: Robustness of model (training sample)
No-distressed
Observed value
Non-distressed
company
Classification (%)
Non-distressed company 38 2 95
Distressed company 1 39 97.5
Overall accuracy of
classification
96.25
Type I error is
BA
B
+
% = 5% and Type II error is =
DC
C
+
% = 2.5%

93
In order to test the predictive capability of model, paired samples are 20 for forecasting test.
Predicted results denoted as table 8.
Table 8: Robustness of model (forecasting sample)
No-distressed
Observed value
Non-distressed
company
Classification (%)
Distressed company 3 17 85
Overall accuracy of
classification
82.25
Type I error is
BA
B
+
DC
C
+
% = 15%
4.4. Logistics regression model training and forecasting
Model input values of paired samples are 40. Vector )8,7,6,5,4,3,1,1( zzzzzzzz is input of this
model. The Logistic regression is been realized using Spss21.0 software and the Wald statistic
for test the βˆ coefficient. Since
=2χ Wald = 2]/[ βσβ =24.123
It has 0.1% significant effective. The logistic regression model is:
]}8675.07985.0655.25308.04879.03295.02654.01286.10448.0exp[1{1)( zzzzzzzziyp −+−+−−−−−+=
Table 9 is denoted as the results of logistic regression.
Table 9: The results of Logistic regression model
Variable coefficient Standard
deviation
Wald statistic Degree
freedom
probability
z1 -1.286 0.413 10.356 1 0.001
z2 -0.654 0.346 3.378 1 0.056
z3 --0.295 0.276 1.211 1 0.294
z4 -0.879 1.421 0.389 1 0.531
z5 0.308 0.324 0.926 1 0.334
z6 -2.555 1.692 1.467 1 0.224
z7 0.985 0.268 0.152 1 0.716
z8 -0.675 0.311 4.971 1 0.025
constant -0.448 0.388 1.273 1 0.251
Table 10 and table 11 are robustness of model in training and forecasting samples respectively.
Table 10: Robustness of model (training sample)
No-distressed
Observed value
Non-distressed
company
Classification (%)
Overall accuracy of
classification
77.5

94
Type I error is
BA
B
+
DC
C
+
% = 20%
Table 11: Robustness of model (forecasting sample)
No-distressed
Observed value
Non-distressed
company
Classification (%)
Overall accuracy of
classification
77.5
Type I error is
BA
B
+
DC
C
+
% = 15%
4.5. Comparison of Wavelet Network and Logistic Regression
In comparison of Wavelet Network and Logistic Regression on training and forecasting sample,
from the table 7, table 8 and table 10, table 11, no matter how the overall prediction accuracy,
Type Ⅰerror and Type Ⅱerror, wavelet networks model is better than logistic regression model.
5. CONCLUSION
This paper select sample listed companies from 2006 to 2009 for training sample. There select 40
paired samples for analysis. It randomly selected 20 paired sample listed companies from 2009 to
2012 for forecasting sample.
Wavelet method and Logistic regression are introduced to suppress the financial distress
prediction accuracy degradation due to potential accounting noise. In this paper, it use
Kolmogorov-Smirnov test (K-S test), Wilcoxon test, Principal component analysis (PCA) to
select potential predictor variables that are informative and closely related to companies’ financial
condition.
In this paper, it has two experiments to compare the prediction accuracy between Wavelet
methods with Logistic regression method. From the results of training and forecasting sample
test, the wavelet networks model in this paper on enterprise’s financial distress early warning is
valid. The result of comparison of Wavelet Network and Logistic Regression denoted as table 12.
Table 12: Comparison of Wavelet Network and Logistic Regression
Model Training Forecasting
Wavelet
Network model
Type errorⅠ 0.05 0.2
Type errorⅡ 0.025 0.15
Overall accuracy of
classification
0.9625 0.8225
Logistic
Regression
Type errorⅠ 0.25 0.30
Type errorⅡ 0.2 0.15
Overall accuracy of
classification
0.775 0.775

95
In Table 12, the comparison of Wavelet Network and Logistic Regression on training and
forecasting sample, no matter how the overall prediction accuracy, Type Ⅰerror and Type Ⅱ
error, wavelet networks model is better than logistic regression model.
ACKNOWLEDGEMENTS
I would like to thank the anonymous reviewers for their constructive comments on this paper.
REFERENCES
[1] Adiana, NHA. Halim, A., Ahmad, H., and Rohani, M. R. (2008), “Predicting corporate failure of
Malaysia’s listed companies: Comparing multiple disrtriminant analysis, logistic regression and the
Hazard model”. International Research Journal of Finance and Economic, Vol. 5, pp. 202-217.
[2] Altman, E.I. (1968), “Financial Ratios, Discriminant Analysis and the Prediction of Corporate
Bankruptcy”, Journal of Finance, Vol. 23, September, 189-209.
[3] Antonakis, A. C. and Sfakianakis, M. E., (2009), “Assessing naïve Bayesian as a method of screening
credit applications”, Journal of applied Statistics, Vol. 36, No. 5, pp. 537-545
[4] Banakar, A., Azeem, M. F., (2006), “A New Artificial Neural Network and its Application in Wavelet
Neural Network and Wavelet Neural-Fuzzy Case study: Time Series Prediction”, 3rd International
IEEE Conference Intelligent Systems.
[5] Banakar, A., Azeem, M. F., (2008), “Artificial Neural Network and its Application in Neural fuzzy
models”, Applied Soft Computing, Vol. 8, No. 4, pp. 1463-1485.
[6] Beaver, W. (1967).” Financial Ratios as Predictors of Failure”, Journal of Accounting Research 4, pp.
71-111.
[7] Becerra, V., Roberta, K. H. G. and Magda, A. S. (2005),” Neural and wavelet models for financial
distress classification”, Data mining and Knowledge Discovery, Vol. 11. No. 1, pp. 35-55.
[8] Bongini, P., Ferri, G. and Hahm, H. (2000), “corporate bankruptcy in Korea: Only the strong
survival?” Financial Review, Vol. 35, pp. 71-111.
[9] Catanach, A. H. and Perry, S. H. (2001), “An evaluation of survival model’s contribution to thrift
institution distress prediction”, Journal of Managerial Issue, Vol.13,No.4, pp. 401-417
[10] Chaduhuri, A, and De, K., (2011), “Fuzzy support vector machine for bankruptcy prediction”,
Applied Soft Computing, Vol. 11, No. 2, pp.2472-2486
[11] Chen, D., (2009), “Using neural networks and data mining techniques for financial distress prediction
model”. Expert Systems with Application, Vol. 36, No. 2, pp.4075-4086
[12] Donald, B. P. and Andrew, T. W. (2000), “Wavelet methods for time series analysis”, Cambridge:
Cambridge University press, 2010.
[13] Gepp, A. and Kumar, K. (2008), “The role of survival analysis in financial distress prediction”.
International Research Journal of Finance and Economic, Vol. 16, pp.13-34.
[14] Gibson, C. H. (2006), “Financial Reporting and analysis: using financial associating information”.
South-Western College Pub, 10 editions, April, 27.
[15] GIGS (2006): “Global Industry Classification Standard (GICS@)”, Standard & Poor’s, August, 2006
[16] Hong, L. and Guangjun, H. (2004), “Early failure warning of listed companies based on artificial
neural network method”, Accounting Research, Beijing ,pp. 42 – 46, February 2004.
[17] Lau, S. T., Cheng, B. W., and Hsieh, C. H., (2009), “Prediction model building with clustering-
launched classification and support vector machines in credit scoring”. Expert Systems with
Applications, Vol. 36, No. 4, pp. 7526-7556.
[18] Lee, C, and Lin, C., (2003), “A wavelet-based Neuron-fuzzy system and its applications”,
Proceedings of the International Joint Conference on Neural Networks, IEEE, pp. 1921-1926.
[19] Lin, L. and Piesse, J. (2004), “The identification of corporate distress in UK industrials: A conditional
probability analysis approach”, Applied Financial Economics, Vol. 14, pp. 73-82
[20] Lineesh, M. C. and C Jessy John, C. J. (2010), “Analysis of No stationary Time Series using Wavelet
Decomposition”, Nature and Science, Vol. 1., pp. 53-59.
[21] Lu, S., Ju, K. H. and Chon, K. H. (2001), “A new algorithm for linear and nonlinear ARMA model
parameter estimation using affine geometry”, IEEE Transaction Biomedical Engineering, Vol. 48, No,
10, pp. 1116-1124.

96
[22] Mcleay, S. and Omar, A. (2000), “The sensitivity of prediction models to the non-normality of bound
and unbounded financial ratios”, The British Accounting Review, Vol. 32, No. 2, pp. 213-230.
[23] Mossman, C. E., Bell, G. G. Swartz, L. M. and Turtle, H. (1998), “An empirical comparison of
bankruptcy models”, Financial Review, Vol. 33, No. 2, pp. 35-53.
[24] Ohlson, J. A. (1980), “Financial Ratios and the Probabilistic Prediction of Bankruptcy, Journal of
Accounting Research, (spring, 1980), Vol. 14, No. 1, pp. 109-131.
[25] Qin, J. and Zhang, B. (1997), “A goodness of fit test for logistic regression models based on case-
control data”. Biometrika, Vol. 84, pp. 609-618.
[26] Rommer, A. D. (2005), “A comparative analysis of the determinants of financial distress in French”,
Italian and Spanish firms, Working paper, Dan marks National Bank, Copenhagen. 18, May, 2005.
[27] Routledge, J. and Gadenne, D. (2000), “Financial distress, reorganization and corporate
performance”, Accounting and Financial, Vol. 40, No. 3, pp. 233-259.
[28] SAS, “The Box-Jenkins methodology for time series model”, SAS global Forum 2013,
support.sas.com/resources/papers/proceedings13/454-2013.pdf
[29] Suhartono and Subanar, (2009), “Development of Model Building Procedures in Wavelet Neural
Networks for Forecasting Non-Stationary Time Series”, European Journal of Scientific Research, Vol.
34, No. 3, pp. 416-427.
[30] Tan, C. N. W. and Dihardjo, H. (2001), “A study on using artificial neural networks to develop an
early warning predictor for credit union financial distress with comparison to the probit mode”,
Managerial Financial, Vol. 27, No. 4, pp. 56-77.
[31] Wang, J., Peng, H., Chen, Y. and Song, Z., (2006), “Adaptive Fuzzy Wavelet Control for a Class of
Uncertain Nonlinear Systems”. Proceedings of the Fifth International Conference on Machine
Learning and Cybernetics, Dalian.
[32] Wu, S., Lu X. (2001),”The Financial Distress Prediction Model of Listed Companies in China”,
Economic Research，Beijing, pp32-38
[33] Yang, Z., You, W. and Ji, G. (2011), “Using partial least squares and support vector machines for
bankruptcy prediction”, Expert Systems with Applications”. Vol.38, No.7, pp. 8336-8342.
[34] Yilmaz, S., and Oysal, Y., (2010), “Fuzzy wavelet neural network models for prediction and
identification of dynamical systems”, IEEE transactions on neural networks, pp. 1599-1609.
[35] Zhang, B. L., Coggins, K., Jabri, M. A., Dersch, D. and Flower, B. (2001), “Multi-resolution
Forecasting for Futures Trading Using Wavelet Decompositions”, IEEE Transactions on neural
networks, Vol. 12, No. 4,pp. 765-775.
[36] Zhang, J., Walter, G. G., Miao, Y., and Lee, W. N. W. (1995), “Wavelet neural network for function
learning”, IEEE Transactions on Signal Processing, Vol. 43, 1485-1497.
[37] Zhang, L. (2000), “Discriminant Analysis Model and Applications on Financial Distress”,
Forecasting, No. 6. pp 11-13, 2000
[38] Zhang, L., S. Chen, J. Yen, and E. Altman (2007), “Corporate Financial distress Diagnosis in China”.
In Working Paper Series: New York University.
[39] Zulkarmain, M. S., Mohamad, A. A. H., Annuar, M. N. and Zainal, A. M. (2001), “Forecasting
corporate failure in Malaysian industrial sector firms”, Asia Academy of Management Journal, Vol. 6,
No.1, pp. 15-30.

COMPARISON OF WAVELET NETWORK AND LOGISTIC REGRESSION IN PREDICTING ENTERPRISE FINANCIAL DISTRESS

More Related Content

What's hot (17)

Viewers also liked (15)

Similar to COMPARISON OF WAVELET NETWORK AND LOGISTIC REGRESSION IN PREDICTING ENTERPRISE FINANCIAL DISTRESS (20)

Recently uploaded (20)

COMPARISON OF WAVELET NETWORK AND LOGISTIC REGRESSION IN PREDICTING ENTERPRISE FINANCIAL DISTRESS