PREDICTING DEFAULT BY INDIAN FIRMS USING STATISTICAL METHODS

Rajeev Kumar Upadhyay1

ABSTRACT

This study assesses the classification accuracies of two statistical methods namely, multiple discriminant analysis and logistic
regression approach in default prediction. It uses a small sample of 32 India firms listed on Bombay Stock Exchange for the
sample period of six years over 2010-11 to 2015-16. Two models have been built using the two statistical methods. Results of the
study clearly indicate that there are no significant differences in the classification accuracies because of change in statistical
methods. Different statistical measures clearly suggest that the two developed models have comparative classification accuracies
in default prediction. However, the specification and robustness of logit model is found to be on lower side than that of the
discriminant model contrary to large numbers of studies. The weaker model specification of logistic model may be result of
small sample size.

Keywords: Default, Bankruptcy, Insolvency, Multiple Discriminant Analysis, Logistic Regression

INTRODUCTION
Any form of external liability is risk to any firm. It is true for
every size of firm. It is quite possible that a firm may fail to
meet its obligations in time. It can happen even with financially
healthy firms. In today’s times when large numbers of the
borrowers are highly leveraged, even the slightest negative
fluctuations in the rate of interests may prove very costly for
the firms (Bandyopadhyay A., 2006). In response to a number
of factors (external and internal) as well as Basel norms for
banking, RBI has been issuing guidelines to the banking and
financial sectors to deal with increasing non-performing assets
but these guidelines have not been able to prevent events of
default by corporate borrowers.
Even after concentrated efforts to improve the situation on
NPAs, the problem of NPA is alarming. The numbers of
defaults as well as wilful defaults are increasing. As a result
the banking and financial system is forced to take deep debt
haircut to recover a portion of its NPA. As per RBI financial
stability report 2016, the NPA of scheduled banks has
increased significantly (RBI, 2016). In such circumstances,
prior information about probable defaults by corporate
borrowers could be of very help for the whole banking and
financial system for corrective measures, both at operative as
well as regulatory levels.
As per available literature, there are few studies on different
methods and models for default prediction in Indian context.
Moreover these studies cannot be termed to be enough. The
present study assesses the predictive ability of two statistical
methods namely multiple discriminant analysis and logistic
regression approach in default prediction. It uses a small
sample of 32 India firms listed on Bombay Stock Exchange
for the sample period of 2010-11 to 2015-16. For the purpose
of study two models have been built using the two statistical
methods.
LITERATURE REVIEW
For more than a century there have been efforts to predict
probable defaults before those actually occur. However, prior
to the breaktrhough study by Beaver (1966), most of the
studies had been using univariate methods that could give
multiple signals. Beaver exhibited the importance of diffrernt
financial ratios in prediction of failure of the firms as well as
role of credit extension to corporate by banking and financial
sector (Beaver, 1966). Altman (1968) extended the work of
Beaver (1966) by developing a single and directional credit
scoring model. This model used multivariate method to a clear
signal about possible default. This study used multiple
discriminant analysis to build a Z-score model using different
financial and market ratios to arrive at a single score. The
developed score(Z-score) could provide clear information
about the possible default by the firm in near future. The
developed model has classification accuracy of 94 percent
two years prior to default (Altman E. I., 1968). This model
revolutionsed the whole arena of defualt prediction forever.
Even today, banking and financial sector across world is using
this method for developing internal credit scoring models for
evaluting loan applications along with other methods.
Over time different approaches for default prediction have
been developed. In theory when the value of assets is lower
than that of the liabilities or there is lack of liquidity, firms
tend to default (Wilcox, 1971). Ohlson (1980) applied
conditional Logit model on a data set with 105 bankrupt and
2058 non-bankrupt firms from the US over the period of 1970-
1976 using 7 financial ratios and 2 categorical variables. The
classification accuracy was lower than that of the studies based
on MDA (Ohlson, 1980). Scott (1981) found that a few

1
Assistant Professor, Sri Aurobindo College (Evening), University of Delhi, E-mail : rajeevupadhyay@live.in

J-GIBS Volume 11, Number 1, January-December 2019 89

Electronic copy available at: https://ssrn.com/abstract=3630799

financial and strucutral ratios can be helpful to understand
the default process (Scott, 1981). A comparison of ZETA
model and Z-score model shows that ZETA model provides
improved accuracy. However, there is no certain explanation
to it but use of more relavant and larger data of industrial
firms in the analysis for ZETA analysis may provide
explanation (Altman E. I., 2000). Liang (2003) conducted a
comparative study on Chinese firms between MDA and
logistic regression model exhibiting that logistic regression
provides relatively higher predictive ability than that of the
MDA (Liang, 2003).
In Indian context Bandyopadhyay (2006) developed two
models based on MDA and logit approach. The Z-score model
exhibited high predictive power in detecting defaulting firms
and outformed the original Altman’s model. The logit model
clearly indicates that inclusion of financial and non-financial
paprameters improves the predictive ability of the model. The
default probability calculted using Logit can also be helpful
in deciding credit risk capital as well pricing of debt
(Bandyopadhyay A. , 2006). The information from structural
models along with firm specific information can imporve the
accuracy of prediction in credit scoring models like logit
model (Bandyopadhyay A. , 2007). Market value of the assets
and volatility in the value of assets are important factors in
default prediction and have inverse relationship with default
probabilities (Sharma, Singh, & Upadhyay, 2014).
Singla and Singh (2017) tried to establish a relationship
between the size of the company and probability of default
for the firms from Indian steel sector using logit and Altman’s
Z score model. The results of the study clearly point that the
size of the firm is inversely related with the probability of
default (Singla & Singh, 2017). This is an interesting result
but it requires further validations with large sample over larger
sample period. Upadhyay (2018) uses four methods, namely
MDA, Logit, Poisson regression based reduced form model
and Merton distant default model over sample period from
2004 to 2016 on 2450 cases. The findings of the study suggest
that logit model yields the highest classification accuracy of
97.7 percent and reduced form model yields the lowest
classification accuracy of 82.3 percent and classification
accuracies of developed models are found to be higher than
that of many other studies (Upadhyay, 2018).
RESEARCH GAP AND OBJECTIVES OF THE
STUDY
There is a need for more studies in the area of default
prediction which could explore different aspects of default
prediction models and practices such as impact of size, sector
and change in statistical methods and models on classification
accuracy.The present study aims to assess classification
accuracies of developed model with the change in statistical
method. For predicting default of bank loans by Indian
corporates, this study uses two statistical methods namely
90
Electronic copy available at: https://ssrn.com/abstract=3630799
multiple discriminant analysis and binary logistic regression
for dichotomous response over a sample period of six years
from 2010-11 to 2015-16 for a relatively small sample.The
results of the study have been compared to arrive at a
conclusion as which method offers better results. This study
has intentionally used relatively a small sample of 32 firms,
half defaulting and half non-defaulting firms listed in Bombay
Stock Exchange with the purpose to assess the change in
predictive ability of default prediction models for small sample
size.
Data and Methods
For the purpose of study, total 192 cases have been used from
32 firms listed on Bombay Stock Exchange. These 32 firms
have been selected randomly of which 16 have defaulted at
least once during the sample period of six years and 16 firms
have never defaulted over the sample period. These firms are
non-banking and non-financial firms from different industries
(manufacturing and services). However there is risk of
sampling bias as well as selection biases (Upadhyay, 2018).
To develop default prediction model (DPM) using the multiple
discriminant analysis, original variables used by Altman
(1968) have been used in the study. Similarly for the logit
model, original variables used by Ohlson (1980) have been
used. The financial data has been collected from the respective
financial statements of these firms. The information on default
has been collected from auditors’ report from the respective
annual reports published by the firms. The market information
has been collected from Bombay Stock Exchange.
Multiple Discriminant Method
For the multiple discriminant analysis, the study has used the
same variables used by Altman (1968) model to arrive upon
Z-scores for the every case (Altman E. I., 1968).
Z-Score = ± – ²1 NWC/TA + ²2RE/TA + ²3EBIT/TA + ²4MVE/
BVD + ²5Sales/TA
Where
NWC/TA = Net working capital to total assets ratio
RE/TA = Retained earnings to total assets ratio
EBIT/TA = EBIT to total assets ratio
MVE/BVD = Market value of equity to book value of debt
ratio
Sales/TA = Sales to total assets ratio
The collected data for the ratios after checking for outliers
has been processed in SPSS using multiple discriminant
analysis and as a result, the analysis yields the following
model:
Z = -1.099 – 0.488NWC/TA – 0.391RE/TA + 10.229EBIT/
TA + 0MVE/BVD + 0.334Sales/TA
J-GIBS Volume 11, Number 1, January-December 2019
Logistic Regression Method DATA ANALYSIS AND INTERPRETATIONS

For the logistic regression the study has used the same
variables used by Ohlson (1980) model to build O-score model
for the sample on the lines of following model (Ohlson, 1980).
O-Score = ± – ²1 WC/TA + ²2TL/TA + ²3CL/CA + ²4NI/TA +
²5FFO/TL + ²6CHIN + ²7Size + ²8X + ²9Y
Where
WC/TA = Working capital to total assets ratio
TL/TA = Total liabilities to total assets ratio
CL/CA = Current liabilities to current assets ratio
NI/TA = Net income to total assets ratio
FFO/TL = Funds from operations to total liabilities
CHIN = (NIt - NIti)/(INItI + INIt-il), where NIt is net income
for the most recent period. The denominator acts as a level
indicator. The variable is thus intended to measure change in
net income
Size = Log (TA/GNP price-level index)
X = 1 if TL>TA, 0 otherwise
Y = 1 if a net loss for the last two years, 0 otherwise
The collected data for the ratios after checking for outliers
has been processed in SPSS using logistic regression analysis
and as a result, the analysis yields the following model:
O-Score = -1.289-0.974WC/TA + 2.48TL/TA + 0.21CL/CA
- 4.307NI/TA + 0.614FFO/TL + 0.251CHIN - 0.302Log (TA/
GNP) + 17.54X + 2.617Y
Part One: Multiple Discriminant Analysis
Table 1.1 provides information about mean and standard
deviation of the ratios used in multiple discriminant analysis
to predict default. From the table, it is clear that there is huge
difference among the ratios between the two groups. The mean
of ratios of non-defaulting group are found to be more robust
than that of the defaulting group. Similarly standard deviation
of ratios of the defaulting group is unfavourably high in
comparison to non-defaulting group. On comparing the mean
and standard deviation of each ratio for both the groups, it
signifies that these two groups are clearly different and may
be making two separate clusters. Though, this requires further
investigations. However, this pattern is more evident in case
of market value of equity to book value of debt ratio as the
mean of this ratio has deteriorated sharply from 101.63 to
0.53. The findings of the study are in confirmation with the
findings of Wilcox (1971) which finds that when a firm is
about to default its key ratios relating to liquidity and structure
start deteriorating over time (Wilcox, 1971).
From the table 1.2, it is clear that there are weak correlations
among most of the variables besides negative correlations
among a few. However, there are strong positive correlations
between NWC/TA and RE/TA and RE/TA EBIT/TA. Strong
positive correlation is undesirable in the analysis as weak and
negative correlation is favourable condition for predicting
default (Cochran, 1964).

Table 1.1: Group Statistics

Mean Std. Deviation

No Default Default Total No Default Default Total
NWC/TA 0.2160 0.1664 0.2067 0.1695 0.3852 0.2253
RE/TA 0.3879 -0.0610 0.3040 0.1698 0.9272 0.4593
EBIT/TA 0.1370 -0.0445 0.1031 0.0767 0.1935 0.129
MVE/BVD 101.639 0.5349 82.7517 292.705 0.6538 266.73
Sales/TA 0.8762 0.5457 0.8145 0.6577 0.3746 0.6274
Table 1.2: Correlation

NWC/TA RE/TA EBIT/TA MVE/BVD Sales/TA

NWC/TA 1.000 0.537 0.319 -0.128 -0.150
RE/TA 1.000 0.750 0.028 -0.010
EBIT/TA 1.000 0.180 0.077
MVE/BVD 1.000 0.158
Sales/TA 1.000

J-GIBS Volume 11, Number 1, January-December 2019 91

Electronic copy available at: https://ssrn.com/abstract=3630799

From the table 1.3, it is clear that the Eigen value for the
analysis is found to be 0.475 which not very high. High eigen
value indicates towards robust specification of the model and
is favourable but for the developed model this value is on
lower side. This signifies lower level of robustness of the
specified model. However the canonical correlation for the
specified model is slightly on higher side at 0.567, thus
improving model’s robustness.
Table 1.3: Eigenvalues
& Pal (2003), Agarwal & Taffler (2008), Mishra & Singh
(2016), Gupta (2014) and Upadhyay (2018). The classification
accuracy for the non-defaulting firms is 90.5 percent with
Type I error of 9.5 percent. The classification for the defaulting
firms is 82.4 percent with Type II error of 17.6 percent.
Considering the small sample size as well as the
misclassifications, the accuracy of the specified model can
be termed satisfactory.
Table 1.6: Classification Results (In-Sample)

Observed Predicted Group Total

Eigenvalue percent of Cumulative Canonical
Membership
Variance percent Correlation
No Default Default
0.475 100.0 100.0 0.567
No Default 143 15 158
Table 1.4 provides information about the Wilks’ Lambda test. Count
Default 6 28 34
In multiple discriminant analysis Wilks’ Lambda tells about
the strength of discriminatory power of the specified model. No Default 90.5 9.5 100.0
percent
Lower is the value, higher is the discriminatory power of the Default 17.6 82.4 100.0
specified model. The value of Wilks’ Lambda for the specified Overall Accuracy 89.1 percent
model is 0.678. This clearly signifies that the specified model
has lower discriminatory power. Part Two: Logistic Regression
Table 1.4: Wilks’ Lambda Table 2.1 gives information about the goodness of fit of the
developed logit model in predicting default by Indian firms
Wilks' Lambda Chi-square Sig. on the banking borrowings. The observed Nagelkerke R
square is found to be 0.518. This signifies that only 51.8
0.678 68.931 0.000
percent of variations can be predicted with the help of the
From the table 1.5, it is clear that in the process of building developed logit model. This value is relatively weak and may
the model, market value of equity to book value of debt would be the cause of errors in default prediction.
be excluded from the analysis as the coefficient for the factor Table 2.1: Model Summary
is found to be zero. This means that this variable has no
discriminatory power as far as this analysis is concerned. -2 Log likelihood Cox & Snell R Nagelkerke R
However, this observation is contrary to the conclusion that Square Square
the mean values and standard deviation provide. This may be
confirmed by further investigations with larger sample and 107.017 0.314 0.518
wider sample period. However the importance of EBIT to From table 2.2 it is clear that the Wald test gives very poor
total assets ratio is very high as coefficient value is as high as values for almost every variable in the model except Y (Y = 1
10.229. if a net loss for the last two years, 0 otherwise). This clearly
Table 1.5: Canonical Discriminant Function Coefficients indicates the relatively weaker discriminatory strength of the
model.
Variables ?
Table 2.2 Variables in the Equation
NWC/TA -0.488
Variables ? W ald Sig.
RE/TA -0.391
W C/TA -0.974 0.409 0.522
EBIT/TA 10.229
TL/TA 2.480 1.018 0.313
MVE/BVD 0.000 CL/CA 0.021 0.084 0.772
Sales/TA 0.334 NI/TA -4.307 0.410 0.522
(Constant) -1.099 FFO/TL 0.614 0.050 0.823
CHIN 0.251 0.796 0.372
From the classification result from table 1.6, it is clear that
Log(TA/G NP) -0.302 1.311 0.252
overall sample classification accuracy of the specified model
is 89.1 percent. The classification accuracyof the developed X(1) 17.540 0.000 0.999
model is comparable to findings of many studies such as Y(1) 2.617 7.925 0.005
Altman (1968), Bandyopadhyay (2006), Acharya, Chatterjee, Constant -1.289 0.188 0.665

92 J-GIBS Volume 11, Number 1, January-December 2019

Electronic copy available at: https://ssrn.com/abstract=3630799

 Table 2.3: Classification Results (In-Sample)
Observed
No Default
No Default
Count
Default
No Default
percent
Default
Overall Percentage
Table 3.1: Comparative Results(In-Sample)
Multiple Discriminant Analysis
Developed Model
Overall Accuracy 89.5
Type I Error 9.5
Type II Error 17.6
The overall sample classification accuracy of the derived
model is found to be 90.2 percent. This level of classification
accuracy is comparable to results of many studies such Ohlson
(1980), Bandyopadhyay (2006) Acharya, Chatterjee, & Pal
(2003), Agarwal & Taffler (2008), Mishra & Singh (2016),
Gupta (2014), Sharma, Singh, & Upadhyay (2014) and
Upadhyay (2018). The classification accuracy for the non-
defaulting firms is 96.9 percent with Type I error of 3.1
percent. The classification accuracy for the defaulting firms
is merely 58.8 percent with Type II error of 41.2 percent.
However the overall misclassification is 9.2 percent only.
Considering the small sample size as well as the overall
misclassifications, the accuracy of the specified model can
be termed satisfactory but the high level of type II error makes
the specification of model weaker.
Part Three: Comparison of Results
From the table 3.1, on comparing the two developed models,
it is observed that there are no significant differences between
the two models on account of overall accuracy and Type I
error but on the basis of Type II errors, the relative robustness
and specification of the developed discriminant model is found
to be stronger than that of the logit model. So it can be
concluded that for a relatively small sample size and small
sample period, there are no significant differences in
classification accuracies of default prediction models.
However the robustness and specifications of the developed
logit models seems to be affected either by small sample size
or small sample period or both. No tool has been used to
directly assess the impact of sample period in this study.
Though, the fact remains intact that there must be some impact
of sample period as the changing values of ratios must be
reflecting on the result. So vaguely it can be argued that the
weak model specification and robustness might have been
J-GIBS Volume 11, Number 1, January-December 2019
Electronic copy available at: https://ssrn.com/abstract=3630799
Predicted Group Membership Total
Default
154 5 96.9
14 20 58.8
96.9 3.1 100.0
41.2 58.8 100.0
90.2
Logistic Regression
Developed Model
90.2
3.1
41.2
caused by the small sample size than the small sample period.
Moreover, it requires further investigations to confirm this
result.
CONCLUSION
The overall sample classification accuracies are found to be
89.1 percent with misclassification of 10.9 percent for
discriminant model and 90.2 percent with misclassification
of 41.2 percent for logit model. So on the basis of accuracies,
it can be concluded that there are no significant differences
in the classification accuracies of developed default prediction
models (DPM) because of the change in statistical methods.
Also the classification accuracies of the developed models
are comparable to many studies like Altman (1968), Ohlson
(1980), Bandyopadhyay (2006) Acharya, Chatterjee, & Pal
(2003), Agarwal & Taffler (2008), Mishra & Singh (2016),
Gupta (2014), Sharma, Singh, & Upadhyay (2014) and
Upadhyay (2018). So it can be concluded that small sample
size seems to have no negative impact on the classification
accuracies of default prediction models.
However the relative robustness as well as the specification
of the developed discriminant model is found to be higher
than that of the developed logit model. So it can be said that
the robustness and specifications of the developed models is
sensitive to change in statistical method.The weak model
robustness and specification may be the cause of higher type
II error in the developed logit model which may have been
caused by small sample size. Though, it requires further
investigation to confirm this result.
For the defaulting firms, the key ratios relating to liquidity
and structure start deteriorating than that of the non-defaulting
firms. From the mean and standard deviation of each ratio for
both the groups, it seems that these two groups are making
93
two separate clusters. Though, it again requires further
investigations. Also in case of discriminant model, market
value of equity to book value of debt is found to be
insignificant signifying that market variables are not important
for discriminant model. However all these conclusions are
based on the original variables used in Altman (1968) and
Ohlson (1980). A study with more variables may provide
better results.
REFERENCES
1. Altman, E. I. (1968). Fianncial Ratios, Discriminant
Analysis and Prediction of Corporate Bankruptcy . The
Journal of FInance, 589-609.
2. Altman, E. I. (2000). Predicting financial distress of
companies: revisiting the Z-score and ZETA models. New
York: Stern School of Business, New York University.
3. Bandyopadhyay, A. (2006). Predicting probability of
default of Indian corporate bonds: logistic and Z-score
model approaches. The Journal of Risk Finance, 255-
272.
4. Bandyopadhyay, A. (2006). Predicting probability of
default of Indian corporate bonds: logistic and Z-score
model approaches. The Journal of Risk Finance, 255-
272.
5. Bandyopadhyay, A. (2007). Mapping corporate drift
towards default Part 2: a hybrid credit-scoring model.
The Journal of Risk Finance, 8(1), 46-55.
6. Beaver, W. (1966). Financial Ratios as Predictors of
Failure, Empirical Research in Accounting: Selected
Studies. Journal of Accounting Research, 5, 71-111.
7. Cochran, W. G. (1964). On the preformance of the linear
discriminant analysis. Technometric, 179-190.
94
Electronic copy available at: https://ssrn.com/abstract=3630799
8. Liang, Q. (2003). CORPORATE FINANCIAL
DISTRESS DIAGNOSIS IN CHINA: EMPIRICAL
ANALYSIS USING CREDIT SCORING MODELS.
Hitotsubashi Journal of Commerce and Management,
38, 13-28.
9. Ohlson, J. A. (1980). Financial Ratios and Probabilistic
Prediction of Bankruptcy. Journal of Accounting
Research, 18(1), 109-131.
10. Ohlson, J. A. (1980). Financial Ratios and the
Probabilistic Prediction of Bankruptcy. Journal of
Accounting Research, 18(1), 109-131.
11. RBI. (2016). Financial Stability Report December 2016.
RBI.
12. Scott, J. (1981). The probability of bankruptcy: a
comparison of empirical predictions and theoretical
models. Journal of Banking and Finance, 5, 317-44.
13. Sharma, C. S., Singh, R. K., & Upadhyay, R. K. (2014).
Predicting Probability of Default. Primax International
Journal of Commerce and Management Research, 2(3),
5-13.
14. Singla, R., & Singh, G. (2017). Assessing the Probability
of Failure by Using Altman’s Model and Exploring its
Relationship with Company Size: An Evidence from
Indian Steel Sector. Journal of Technology Management
for Growing Economies, 2, 167-180.
15. Upadhyay, R. K. (2018). Predicting Probability of Debt
Default: A Study of Corporate Debt Market in India and
other Countries (Doctoral Thesis). University of Delhi.
16. Wilcox, J. (1971). A simple theory of financial ratios as
predictors of failure. Journal of Accounting Research,
9(2), 389-95.
J-GIBS Volume 11, Number 1, January-December 2019
Electronic copy available at: https://ssrn.com/abstract=3630799