Professional Documents
Culture Documents
Financial Performance and Predicting The Risk of Bankruptcy - A Case of Selected Cement Companies in India
Financial Performance and Predicting The Risk of Bankruptcy - A Case of Selected Cement Companies in India
net/publication/266493500
CITATIONS READS
9 77
1 author:
Ozgur Turetken
Ryerson University
45 PUBLICATIONS 489 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Ozgur Turetken on 18 February 2015.
Comparative Study
Ozgur Turetken
E-mail: turetken@temple.edu
Predicting Financial Performance Of Publicly Traded Turkish Firms: A
Comparative Study
Abstract
This study aims to predict the financial performance of publicly traded Turkish firms
using their available financial data. Financial performance is measured by a firm’s inclusion in
an index of top performers, where the inclusion of firms is based on the value and sales volume
of their stocks. Two alternative techniques, multiple discriminant analysis and neural networks,
are used for this prediction problem, and their prediction accuracy is compared. The results
suggest that the chosen financial ratios lead to useful predictions with both techniques, where
neural networks perform better than discriminant analysis when trained with a larger sample. It
was also found that both techniques provide better predictions for average and poor performers
than they do for top performers without statistically significant differences between the
Prediction of the financial performance of a firm is important for a number of groups that
have interest in the firm. Creditors, auditors, stockholders, and senior management all need a
methodology or model based on some available information that will be helpful in the
and as a decision making tool for auditors [16,17,21,23,33,35]. Many financial performance
predictions involve the assignment of a firm to one of a number of pre-determined classes. Such
classifications concerning the performance and viability of a firm is a complicated process where
systems [20], have been used to tackle the classification problem. Among these, statistical
methods have been the traditional way of building such predictive models [31]. The most
frequently used statistical methods for a prediction that involves assignment of an observation to
either one of two prespecified classes (a 2-way classification problem) are multiple discriminant
analysis (MDA), binary logistic regression (logit) analysis, binary probit analysis, and ordinary
More recently, Neural Network (NN) models have gained a lot of popularity as an
finance applications of neural networks shows that while most of the research in the area
suggests that NN models are superior to statistical models in terms of prediction accuracy, the
results are mixed [34]. Further, it has been suggested that the superiority of one class of
the amount [20], the accuracy [2,22], and other characteristics [14,20] of the data used in the
model. Even studies that suggest a combined statistical and artificial intelligence based
technique stress the importance of determining the context in which each technique is more
successful [24].
To improve our understanding of the mentioned techniques, this paper presents a study of
the predictive capabilities of neural networks and multiple discriminant analysis in a new
domain. The aim of the study is to predict financial performance of publicly traded Turkish
firms using neural networks and discriminant analysis, and to compare their performance. Since
the list of publicly traded Turkish firms is a new data set for such a comparison, the results will
shed light on the generalizability of earlier findings of studies that were done in different
countries. This is important, because possible differences between the findings of this study and
those of others may point to the existence of certain macro variables depending on national
economies or the period of time a study is conducted. Meanwhile, the lack of such differences
would increase our confidence in the relative strength and weakness of each technique.
In addition, we study the effect of training sample size in building the models for its
previously proposed effect on prediction accuracy [20]. Finally, the dependent variable in our
study is financial success measured by a firm’s inclusion in the Istanbul Stock Exchange 100
(ISE100) index rather than failure or bankruptcy, which has been the focus of most previous
The organization of the rest of the paper is as follows. In the next section, relevant
literature on the topic is reviewed. The third section describes the study by displaying the data
4
set, and detailing the research questions. Also, the methods used to build the models, and the
predictive models per se are covered in this section. The results are given in the fourth section,
Corporate performance prediction has been of interest to scholars for more than three
decades. Much earlier work on financial performance prediction has focused on bankruptcy
cases for their importance to a number of stakeholders such as creditors, auditors, stockholders,
and senior management. Beaver [3] is one of the pioneering researchers to study whether
financial ratios provide useful information in bankruptcy prediction. Altman [1] expanded this
idea, and proposed the use of multiple discriminant analysis (MDA) built on a set of five
financial ratios in bankruptcy prediction. This study found that the MDA model could
successfully predict bankruptcy in 94% of the cases in the sample. Since Altman, MDA models
have been used extensively for the bankruptcy prediction problem [5,19], while logit has
emerged as another popular technique [4]. The major evolution in these studies has been the
identification of different financial and economic ratios that improve predictive performance
[33].
MDA and logit assume certain characteristics of the underlying data such as normality
therefore they are parametric techniques. A neural network model, on the other hand, is non-
parametric since it does not assume the data to have any specific characteristics. This non-
parametric nature of neural network based prediction models has been attractive to researchers
for various kinds of problems that require predictions based on financial variables. Among such
problems are loan application approval decisions [10], prediction of corporate bond rankings and
5
profitability [7], prediction of initial stock prices [16] and stock returns [18], prediction of mutual
fund performance [15], credit card fraud prevention [27], signature forgery detection [9,26] as
have been made on various dimensions, and the effect of different model-building decisions on
the relative success of a technique has been studied. For example, Tam [30] and Tam and Kiang
[31] designed a detailed study to compare several models built for prediction of bankruptcies in
the U.S. banking sector in the late eighties. These comparisons were made on the robustness,
predictive accuracy, adaptability, and explanatory capability dimensions. The results suggested
that neural networks offer better predictive accuracy than discriminant analysis, factor-logistic
analysis, k nearest neighbors, and ID3 especially when the data set is multimodal. Neural
networks were also more robust due to their non-parametric nature, and showed better
adaptability in model adjustment. On the other hand, the studies observed that neural networks
are harder to configure, computationally inefficient, and lacked the explanatory capability of
In another study, a neural network was trained to predict thrift failures and the resultant
model was evaluated using a logit model as a performance benchmark for correct classification
ratios [28]. This study found that the neural network, which uses the same financial ratios as the
corresponding logit model, achieves a higher degree of prediction accuracy, and is more robust.
Fletcher and Goss [8] used company failure data to build a neural network and a logit model, and
found that the neural network technique more accurately predicts bankruptcy than the logit
model.
6
Wilson and Sharda [33] designed a comprehensive study to compare the prediction
accuracy of neural networks and multiple discriminant analysis using bankruptcy data. Using
simulation, the authors created 180 different training and test samples with one of three differing
ratios of bankrupt and non-bankrupt firms in each subsample. Their major findings were that the
neural networks outperform discriminant analysis in almost all cases, and the composition of the
training set has a significant effect where the best training was achieved when the ratio of the
Tsukuda and Baba [32] used 48 different neural network models to predict Japanese
corporate bankruptcy, and compared the results to those of multiple discriminant analysis. The
comparison suggested partial support for neural networks such that the neural networks were
superior in the classification of non-bankrupt firms, whereas both techniques were equivalent in
Boritz and Kennedy [6] found that the performance of neural networks in comparison to
that of discriminant analysis, logit, and probit is effected by the choice of variables and sampling
error in the learning sample. It was also noted that the performance of a back propagation neural
A study by Luther [25] developed a prediction model using neural networks for the
outcome of bankruptcy, based on the firm’s financial ratios at the time of filing for Chapter 11.
The performance of the neural network models varying on parameters such as the cut-off points
was compared to that of a logit model tested on the same data set. The conclusion was that the
neural network approach is more robust than logit in predicting the outcome.
Although most of the literature seems to agree that neural networks are better predictors
of financial performance, the agreement is not unanimous. For example, contrary results were
7
found in [35] where the discriminant analysis outperformed different kinds of neural network
This brief review of the relevant work shows that for a number of reasons that may or
may not be controlled by the model builder, the relative success of neural networks over
statistical methods is case specific. Although there is general agreement that neural networks are
more robust and statistical techniques such as discriminant analysis are easier to compute and
interpret, the predictive accuracy of each technique depends on the data. For these reasons, this
study compares the relative performance of neural networks and multiple discriminant analysis
in a different context. What makes the context different is the data set, which is different from
what has been previously studied, and the choice of independent and dependent variables. The
next section discusses the characteristics of the data set and variables.
The data set is taken from the Istanbul Stock Exchange (ISE) bulletins and includes the
firms whose stocks were traded on ISE in 1998. The total sample size is 283. As a surrogate for
financial performance, the ISE 100 index listings are used. The ISE 100 list is formed by
ranking the firms by their market values, which is the product of the number of stocks and the
price of each stock. The top firms in this ranking are included in the index. Normally, the
number of firms included is 100. However, for various reasons that are not relevant to this study,
the exchange may decide that not all 100 firms will be included, and the list is made shorter. For
example, the data set used in our study has 94 firms in the index. The ISE 100 index is updated
quarterly, and the inclusion of a firm is a decision based on its activities from the previous year.
8
Therefore, whether a firm is in the index for the first quarter of the year 1999 is considered to be
assigned a value of 1 if a firm is listed on the ISE 100 index in the first quarter of 1999, and a
value of 0 otherwise. We believe this approach to represent financial performance is useful since
the ability to predict top performers in the stock market would facilitate a quick and easy first
Financial (accounting) ratios of a firm are arguably better measures of a firm’s current
performance than the individual items on the financial statement (Brigham et al. 1999).
Accordingly, the independent (predictor) variables in the study are financial ratios. It has been
established in the literature (for example, [23]) that the ratios that have predictive power for
financial performance can be grouped under three main categories: 1. profitability, 2. debt, 3.
liquidity and activity. Based on the availability of data, two profitability (return on total assets,
basic earning power), two debt (debt ratio, paid in capital by total assets), and one activity (book
value by total assets) variables are used in this study. The following is a brief description of
these variables:
- NP/TA (Return on Total Assets): This is the ratio of the net profit to total assets of a firm and
- EBIT/TA (Basic Earning Power): This ratio is calculated by dividing earnings before interest
and taxes by total assets. This ratio shows the raw earning power of the firm’s asset before
the influence of taxes and leverage, and is useful for comparing firms with different tax
9
- FD/TA (Debt ratio): This ratio is calculated by dividing the sum of current liabilities and
long-term debt by total assets. Debt ratio measures the percentage of funds provided by
creditors.
- PIC/TA (Paid in Capital/Total Assets): This is similar to debt ratio; it measures the
- BV/TA (Book Value/Total Assets): Book value of a firm’s stock is the ratio of the common
maximum degree of separation between two groups (i.e. the two levels of the dependent
variable). This discriminant function is either a linear combination of the attributes (i.e.
The linear discriminant function for multiple discriminant analysis (MDA) is of the following
where
10
Observations are classified as 1 if y is less than a cut-off value or as 2 otherwise. The
cut-off value is determined by averaging the y values for all observations. This procedure
generates a discriminant function based on linear combinations of the predictor variables, which
provide the best discrimination between the groups. The function is generated from a sample of
observations for which group membership is known. The function can then be applied to new
As a general rule, the discriminant function is constructed using a portion of the sample
data while the rest of the sample is used to test the performance of the function in classifying
new data. These two samples are called analysis and test (hold-out) samples respectively.
Neural Network (NN) models are among the most popular machine learning techniques
used for prediction. Neural Networks are biologically inspired statistical tools [33]. In a neural
network model, there are two or more layers in which there are individual processing elements,
called neurons. Each neuron has an input and an output connection point. Neurons are joined to
one another in the same or different layers. There is a weight associated with each connection
and these weights are used to calculate the weighted combination of the “signals” that a neuron
receives and processes (i.e. applies a nonlinear transformation). As a result of this neural
processing, an output signal is produced and sent to other neurons that the output point is
connected to. This signal is used by the receiving neurons as input. These neurons in different
layers and the weighted connections between them constitute the neural network.
In a typical feed-forward neural network, information flows from the input neurons (in
the input layer) to the output layer of the network through hidden layer(s). The output (at the
11
output layer) thus produced is compared to the actual output and the difference between the two
is used as feedback to the input and hidden layers for adjusting the connection weights. This
approach is one of the most popular neural network training methods known as
“backpropagation”. The backpropagation network is presented with many such input/output data
pairs before it starts mapping the relationships between dependent and independent variables in
numeric data. This process is the actual training of the network, and like the human brain, the
memory or “knowledge” of the resulting network is not stored in a particular location, but in the
overall pattern of connections that determine the network’s architecture. The result is a
computational system capable of “learning from experience” and therefore increasing the
After training (lowering the output error to an acceptable level), the network model is
“tested” by taking a set of new input data that were not used in training, and by feeding it into the
network to observe the output it will produce. The resulting output “guess” is then compared to
the known answer and an accuracy measure is calculated. This helps in determining how well
the neural network has “learned” to place the correct dependencies and weights to the various
In this study we try to answer a number of research questions. First, since it has been
proposed that the choice of predictor variables affects the performance of a prediction technique
[6], we would like to see whether the set of financial ratios we have chosen is actually a good
12
Next, we would like to compare the predictive accuracy of multiple discriminant analysis
(MDA) and neural networks (NN) on the specific data set of this study. It has been proposed
that the size of the training sample affects the relative performance of different prediction
techniques [20] where neural networks perform better with a larger sample size. Therefore we
would like to study the effect of training sample size in comparing neural networks and
discriminant analysis. The suggested lower limit for the analysis (training) sample size for
discriminant analysis is 20 times the number of independent variables [12]. Based on that, the
analysis subsample size in this study should be at least 100. On the other hand, when the training
sample size is too large, neural networks tend to “overfit” the model to data, which is
occasionally referred to as “memorizing” the data. Memorization degrades the neural network
performance due to the failure of the models to generalize. Therefore, we wanted to avoid
memorization. For this reason, we imposed an upper limit on the size of the training sample size,
and set that limit to half of the overall sample size. This way, we obtained two training sample
sizes: the “small” sample includes 100 cases (firms), and the “large” one includes 141 cases.
Our final research question concerns the observation that the cost of misclassification in
financial predictions is not the same for good performers and poor performers [32,33]. As it is
more costly to decide that a “to go bankrupt” firm will survive than to decide that a healthy firm
will go bankrupt in bankruptcy studies, we argue that it is more costly to classify an average or a
poor performer as a top performer than classifying a top performer as a poor or average one.
Therefore we also compare neural networks and discriminant analysis for the correct
classification of both groups of firms (i.e. ISE 100 firms and others).
13
3.4 Models
The multiple discriminant analysis method operates on the assumption that the data for
the independent variables are normally distributed. To test for normality, normal p-p plots have
been drawn for each of the five independent variables as depicted in the appendix. For these
plots and the subsequent statistical analyses of the study, Statistical Package for Social Sciences,
SPSS is used. The p-p plots suggest that the return total assets, the basic earning power, and the
debt ratio can be considered normally distributed for practical purposes. However, the deviation
from normality for the “book value to total assets ratio” (BV/TA) and “paid in capital to total
assets ratio” (PIC/TA) variables is considerable. As a remedy, these variables are put through a
logarithmic transformation, and the distributions of the resulting variables are found to be
normally distributed for all practical purposes. These two “transformed” variables and the three
As discussed before, two MDA models, one with an analysis (training) sample size of
100 cases (firms), and another with an analysis sample size of 141 cases are built. The model
with 100 cases in the analysis sample is arbitrarily called MDA1 while the other one is called
MDA2. The total sample size in the study is 283, therefore the test sample size in MDA1 is 183,
The analyses are performed using the stepwise model. To avoid a possible bias towards
the larger group, equal prior probabilities for the two groups are set. This way, the best chance
14
3.4.1.1 MDA1
MDA1 divides the total sample into an analysis sample of 100 cases, and a test sample of
183 cases. An important assumption in MDA is the equality of the variance-covariance matrices
across the different groups. This equality means that the cases from different groups come from
similar data generation processes, and this is particularly important for a model to have
explanatory capability. Box’s M tests are used in MDA to test whether this equality is valid or
not. Figure 1 depicts the Box’s M test results, and it is seen that the variance-covariance
matrices are significantly different between the two groups (p = 0.034). This means that this
model only has predictive power but no explanatory power. In other words, the subsequent
discriminant function cannot be used to examine the effect of the individual variables on the
dependent variable. This is not a particular problem since the purpose of the study is the
general, are not used for explanatory purposes hence the lack of explanatory capability of the
MDA models does not hurt the comparisons. The correct classification rates of MDA1 are
displayed in Table 1.
3.4.1.2 MDA2
This is the model that divides the total sample into equal-sized (i.e. 141 vs. 142) analysis
15
The Box’s M test results in Figure 2 suggest that the variance-covariance matrices in the
sample of this analysis are significantly different between the two groups (p=0.027). Similar to
the previous model, the discriminant function here is good for prediction purposes only. The
The NN model with 5 input (one for each predictor), 8 hidden, and 2 output nodes is built
using ThinksPro – Neural Networks for Trial Version. The network is built on a multi-layer
normal feed-forward architecture. Each node in the output layer is a dummy variable and takes a
value of 1 if a case belongs to the class that the node represents and 0 otherwise. The sigmoid
function (t(f) = 1/(1+exp(-f))) is used for transforming the incoming signal to output at each
node. The input data to the network are preprocessed by performing a z-transformation, i.e., for
each case and every independent variable in the data set, a z-score is calculated by taking the
difference between the independent variable value of a specific case and the average of that
independent variable over the entire data set, and dividing this difference by the standard
deviation of the variable. The sigmoid function has an S-shaped curve, and its value is virtually
constant unless the input to the function is within a certain interval. When inputs from outside
this interval are used in the network, this phenomenon of having a constant output value
materializes, and this is known as the saturation of the network. The z-transformation is required
to avoid saturation especially when the sigmoid function is used at individual nodes.
The network is trained using the backpropagation (BPN) algorithm, and is considered
“trained” when the (mean standard) classification error reaches a local minimum. This decision
16
is the result of a judgment call. The classification error is defined as the root mean square of the
As was the case in the MDA models, two neural networks are trained. The model with
100 cases (firms) in the training (analysis) sample is called NN1, and the one with 141 cases in
3.4.2.1 NN1
The classification error as described above on the network trained on 100 samples
reaches a local minimum after 900 iterations where the mean standard error is approximately
3.4.2.2 NN2
The classification error on the network trained on 141 samples reaches a local minimum
after 800 iterations where the mean standard error is approximately 0.27. The correct
4. Results
Three predictive performance measures for the predictive models are defined:
1. group 1 (ISE100 firms) correct classification rate, 2. group 2 correct classification rate, 3. total
correct classification rate. These measures belonging to the predictive models developed in the
17
Our first research question is regarding whether the set of our chosen financial ratios can
successfully predict financial performance. To assess the value of the predictor set, we use
Press’s Q statistic. Press’s Q statistic is a test to determine whether the model results are
significantly better than those that would be obtained by chance only [12]. Press’s Q statistic is
calculated as follows:
where
k = number of groups.
The critical value of Q (Chi-square value for 1 degree of freedom at the desired
confidence level) at a significance level of 0.05 is approximately 3.84. Based on this, all the
results of our models are better than chance values. Accordingly, we suggest that the predictor
variable set produces acceptable predictions both for MDA and NN.
18
The second research question we wanted to tackle was regarding the relative performance
results of Table 1 shows that the comparison of MDA and NN cannot be made without
considering the effect of training sample size such that NNs outperform MDA (68.3% vs. 63.4%)
when the training sample is large while MDA outperforms NN (60.1% vs. 57.4%) when the
Our last research question concerns the relative performance of neural networks and
discriminant analysis for the correct classification of top financial performers (i.e. ISE 100 firms)
versus others. As seen in Table 1, both neural networks and discriminant analysis are better at
correctly classifying non-ISE 100 firms than ISE 100 firms. To test whether this difference is
significant, the non-parametric Wilcoxon signed-rank test is used. The Wilcoxon signed-rank
test is a procedure used with two related variables to test the hypothesis that the two variables
have the same distribution. This test takes into account information about the magnitude of
differences within pairs and gives more weight to pairs that show large differences than those
that show small differences. The test statistic is based on the ranks of the absolute values of the
The Wilcoxon signed-rank test results displayed in the top portion of Figure 4 suggest
that the difference between the correct classification rate of ISE 100 firms and non-ISE 100 firms
(non-ISE 100 > ISE 100) is borderline significant (p = 0.068). To test whether the difference
between groupwise classification rates are technique dependent, we examine the difference in
19
groupwise classification rates for MDA and NN separately (Table 2). As seen in the bottom
portion of Figure 4 and Table 2, neural networks and discriminant analysis are not significantly
different (p = 0.180) in terms of differences between groupwise classification rates. Rather, the
difference in correct classification rate between groups depends on the training sample size. In
the next section, we discuss the implications of these results, and draw conclusions.
This paper aimed to explore whether the financial performance of the firms whose stocks
are traded in the Istanbul Stock Exchange (ISE) could be predicted using a set of available
financial ratios. Our test results showed that both the multiple discriminant analysis and the
neural network models built using this set of ratios of the 283 publicly traded firms yielded
predictions that are better than chance values, and are therefore usable for predictive purposes.
Our analyses were performed for a data set that was not analyzed in this context before, therefore
the findings confirm the validity of predicting financial performance by firm-level variables
We observe that the correct classification rates for the test samples were between 57.4%
and 68.4%, which are lower than the typical figures reported for bankruptcy prediction models,
which are typically above 90%. However, it should be noted that unlike bankruptcy prediction
models, this study uses a market value-related variable to represent financial performance.
Although market value of a firm depends on its performance, it is also affected by other factors
20
such as the macro-level economical and political state of the country, and stock market
conditions. Therefore being able to predict stock performance solely by micro-level variables is
nontrivial, and would be a good tool for financial analysts and individual investors. On the other
hand, despite the good results of our models, we believe their value will be best utilized if
complemented by macro level predictions and other widely used techniques such as time series
forecasting. Details of how such integrative models should be built were beyond the scope of
The previous literature reviewed in section 2 suggested that the neural network models
would outperform discriminant analysis in classification accuracy. The results of this study
suggest that this is only true if the neural network model is trained with a large enough data set.
Our conclusion is that neural networks would be particularly useful for macroeconomic studies
where the data set is typically very large whereas for data samples that have sizes around the
recommended minimum for discriminant analysis, neural networks would not perform to their
best capacity. Our recommendation for such cases is to prefer traditional statistical models
especially given the complexity of neural networks, their level of sensitivity to model building
decisions, and the time it takes to build them. Earlier work with simulated data [20] also
suggests that neural networks perform significantly better when the training sample size is large,
however the results do not provide a minimum training sample size recommendation. Although
we did not test the comparative performance of the techniques over a continuously changing
range of the training sample size, we were able to see a difference for the larger training sample
size, which is roughly 28 times the number of independent variables. Here, we were limited by a
data set of certain size, but future research should study whether a certain multiple of the number
of independent variables (for example, a conservative suggestion would be 30) could be a good
21
cut off point for neural networks to outperform discriminant analysis. Likewise, the size of our
sample also limited us from observing a good upper limit for neural network training sample
sizes beyond which over-fitting (i.e. memorization) would occur, and that remains to be an open
Our results also showed that using the predictor variable set, both the discriminant
analysis and neural network models had a propensity towards the average or poor performing
firms hence the classification rates for these firms are significantly better than those in the top-
performing group. It can be argued that the “cost” of classifying a poor or average performer as
a top performer is higher than classifying a top performer as an average or poor performer. For
that reason, the predictions of our models may be better than just a simple average of the correct
classification rates of each group since they perform better when the misclassification cost is
higher. On the other hand, there was no significant difference between discriminant analysis and
neural networks in terms of the relative performance in classifying each group. Therefore, our
results fail to provide a recommendation as to what technique to use for problems where the
22
Appendix
.75
.50
Expected Cum Prob
.25
0.00
0.00 .25 .50 .75 1.00
.75
.50
Expected Cum Prob
.25
0.00
0.00 .25 .50 .75 1.00
.75
.50
Expected Cum Prob
.25
0.00
0.00 .25 .50 .75 1.00
23
Normal P-P Plot of NETP/TA
1.00
.75
.50
Expected Cum Prob
.25
0.00
0.00 .25 .50 .75 1.00
.75
.50
Expected Cum Prob
.25
0.00
0.00 .25 .50 .75 1.00
24
- Normal P-P Plots Of The Transformed Variables
.75
Expected Cum Prob
.50
.25
0.00
0.00 .25 .50 .75 1.00
.75
Expected Cum Prob
.50
.25
0.00
0.00 .25 .50 .75 1.00
25
References
Regression and Neural Networks as Data Quality Varies: A Business Value Approach. Journal of
4. Bell, T. B., Ribar, G. S. and Verchio, J. R. Neural Nets vs. Logistic Regression: A comparison
of each model's ability to predict commercial banks' failures. The Deloitte and Touch/University
12(1): 1-25.
8. Fletcher, D. and Goss, E. Forecasting with neural networks, an application using bankruptcy
10. Gallant, S. Connectionist Expert Systems. Communications of the ACM 1988; (February):
152-169.
11. Gessner, G., Malhotra N. K., Kamakura W. A., Zmijewski, M. E. Estimating Models with
Binary Dependent Variables: Some Theoretical and Empirical Observations. Journal of Business
12. Hair, J. F., Anderson, R. E., Tatham, R. L. and Black, W. C. Multivariate Data Analysis.
13. Hansen, J. V., McDonald, J. B. and Stice, J. D. Artificial intelligence and generalized
14. Hill, T., O'Connor and M., R., W. Neural Network Models for Time Series Forecasts.
15. Indro, D. C., Jiang, C. X., Patuwo, B. E. and Zhang, G. P. Predicting Mutual Fund
16. Jain, B. A. and Nag, B. N. Artificial Neural Network Models for Pricing Initial Public
17. Jo, H. and Han, I. Integration of Case-Based Forecasting, Neural Network, and Discriminant
Analysis for Bankruptcy Prediction. Expert Systems with Applications 1996; 11(4): 415-422.
18. Kanas, A. and Yannopoulos, A. Comparing Linear and Nonlinear Forecasts for Stock
27
21. Kim, K. and Han, I. Genetic Algorithms Approach To Feature Discretization in Artificial
Neural Networks for The Prediction of Stock Price Index. Expert Systems with Applications
22. Klein, B. D. and Rossin, D. F. Data Quality in Neural Network Models: Effect of Error Rate
23. Kryzanowski, L., Galler, M. and Wright, D. W. Using Artificial Neural Networks to Pick
24. Lee, K. C., Han, I. and Youngsig, K. Hybrid Neural Networks for bankruptcy prediction.
25. Luther, R. K. An artificial neural network approach to predicting the outcome of Chapter 11
bankruptcy. The Journal of Business and Economic Studies 1998; 4(1): 57-73.
26. Mighell, D. Back-Propagation and its Application to Handwritten Signature Verification. In:
D. S. Teuraetsky editor. Advances in Neural Information Processing Systems. San Mateo, CA:
27. Rochester, J. E. New Business Users for Neurocomputing. IS Analyzer 1990; (February): 1-
17.
28. Salchenberger, L. M., Cinar, E. M. and Lash, N. A. Neural networks: a new tool for
29. Swales, G. S. J. and Yoon, Y. Applying artificial neural networks to investment analysis.
30. Tam, K. Y. Neural network models and the prediction of bank bankruptcy. OMEGA 1991;
19(5): 429-455.
28
31. Tam, K. Y. and Kiang, M. Y. Managerial Applications of Neural Networks: The Case of
32. Tsukuda J. and I., B. S. Predicting Japanese Corporate Bankruptcy in Terms of Financial
Data Using Neural Network. In: 16th Annual Conference on Computers and Industrial
Engineering, 1994.
33. Wilson, R. L. and Sharda, R. Bankruptcy prediction using neural networks. Decision Support
34. Wong, B. K. and Selvi, Y. Neural Network Applications in Finance: A Review and Analysis
35. Yang, Z. R., Platt, M. B. and Platt, H. D. Probabilistic Neural Networks in Bankruptcy
36. Zopounidis, C., Doumpos, M. and Matsatsinis, N. F. (1997). On the Use of Knowledge-based
Decision Support Systems in Financial Management: A Survey. Decision Support Systems 1997;
20(1997): 259-277.
29
Figures
Figure 1
Box’s M Test Results for MDA1
Box's M 4.563
F Approx. 4.497
df1 1
df2 13219.585
Sig. .034
Figure 2
Box’s M Test Results for MDA2
Box's M 4.927
F Approx. 4.874
df1 1
df224536.062
Sig. .027
Figure 3
Correct Classification Rate versus Sample Size
70.00%
Correct clasification rate
68.00%
66.00%
64.00%
62.00%
MDA
60.00%
58.00% NN
56.00%
54.00%
52.00%
50.00%
100 140
Sample size
Figure 4
Comparison of Groupwise Correct Classification Rates Wilcoxon Signed-Rank Test
Results
NONISE100 -
ISE100
Z -1.826
Asymp. Sig. (2-tailed) .068
NN - MDA
Z -1.342
Asymp. Sig. (2-tailed) .180
31
Tables
Table 1
Correct Classification Rates
Table 2
Correct Classification Rates by Groups