MODELING CREDIT RISK IN INDIA: ẐIndia

The Z-Score Model developed for Indian Companies

SAANJALI KOTAHWALA

An honors thesis submitted in partial fulfillment

of the requirements for the degree of

Bachelor of Science

Undergraduate College

Leonard N. Stern School of Business

New York University

May 2014

PROFESSOR MARTI G. SUBRAHMANYAM PROFESSOR EDWARD I. ALTMAN

Faculty Adviser Thesis Adviser
ABSTRACT

Recognizing the success of adapting the Altman Z-Score model (1968) to different

subsets of companies such as SMEs, Emerging Market companies, private companies, and

companies subject to extraordinary administration, we develop a distress prediction model

specifically for Indian companies. A data set of publically traded companies in India is collected

and various financial ratios are analyzed. The most predictive of these ratios are selected by

running multiple logistic regressions. Validation of the model is conducted by running the ratios

from the model on the entire data set leaving one company each time, a method provided by

Lachenbruch (1967). As per validation, the prediction power of the model has 89.09% accuracy.
ACKNOWLEDGEMENTS

Professor Ed Altman, for his guidance and expertise without which there was no hope of finding
meaning in the data. More importantly, for his commitment, patience, and warmth, which made
the work fulfilling.

My mother, Indira Kotahwala, for teaching me to take responsibility.

My father, Alok Kotahwala, for motivating the topic of my research.

The team at Royal India, for helping me in the collection of data.

Professor Subrahmanyam, for coordinating the program and giving me the opportunity to do this
research.

Professor Marco Avellaneda and Professor Patrick Perry, for explaining the nuts and bolts of
different regression methods.
 I. INTRODUCTION

Credit risk models have a wide range of applicability. From the company’s perspective,

the more accurate the assessment of its risk, the more accurately its risk will be priced in terms of

interest rates and size of loans and advances. Bank capital requirements can also be affected by

different risk models. A previous study showed that building a model specifically for the SMEs

was more effective than a generic model and therefore lowered Basel II capital requirements for

SMEs (Altman & Sabato 2007). From the bank’s perspective, a model that is accurate and can be

applied with relative ease helps to take quick yet informed decisions when dealing with a large

number of clients. It helps them quantify and manage risk across different products and

geographies.

Considering the success in adapting the generic models in many previous cases, we aim

at developing a specific model for India so that we yield better prediction accuracy than the

generic model, which for our purposes is the Z’’-Score model. Our goal is to find a set of ratios

that has the most predictive power of a company’s credit worthiness. We therefore analyze 15

financial ratios of 55 publicly traded Indian companies and try to narrow down to a few ratios.

While our output gives a probability of default, the use of the model can be seen more as

identifying whether a company seems more similar to one that defaulted a year later or one that

remained healthy.

The Z-Score gained popularity due to its accuracy and ease of applicability. Seeking

these two goals, we furthered the model for companies in India to see if accounting for country-

specific characteristics by choice of data, yields a more accurate model.

 Lehmann (2003) has shown that using qualitative variables, i.e. subjective judgments of

credit analysts, improves prediction quality. Our model does not account for any qualitative

variables and therefore can be further improved by incorporating such input.

II. REVIEW OF RELEVANT LITERATURE

II. a. Generic Models & India-Specific Model

One of the most well-known distress prediction models, the Altman Z-Score (1968), uses

four financial statement ratios and a stock market variable. It was developed with 66 American

manufacturing companies, with an equal number of defaulted and non-defaulted firms. The Z’-

Score was a later adaptation of the original model to private companies (1983). Extending the

model for non-US, non-manufacturers and emerging markets, the Z’’-Score was introduced in

1995, by analyzing a sample of Mexican companies. The ratios of Z’’ were the same as the Z’,

excluding the Asset Turnover ratio because of its sensitivity to industry and country.

Tables 1,2,3 outline how the score was calculated in each model.

Z=
+1.2 Working Capital / Total Assets
+1.4 Retained Earnings / Total Assets
+3.3 EBIT / Total Assets
+0.6 Market Value Equity / Book Value of Total Debt
+0.999 Sales / Total Assets

TABLE 1 (Source: Altman, 1968)

Z’ =
+0.717 Working Capital / Total Assets
+ 0.847 Retained Earnings / Total Assets
+3.107 EBIT / Total Assets
 +0.420 Book Value Equity / Total Liabilities
+0.998 Sales / Total Assets

TABLE 2 (Source: Altman, 1983)

Z’’ =
+6.56 Working Capital / Total Assets
+ 3.26 Retained Earnings / Total Assets
+6.72 EBIT / Total Assets
+1.05 Book Value Equity / Total Liabilities

TABLE 3 (Source: Altman, Hartzell & Peck, 1995)

An India-specific model was developed by Bhatia (1988) for identifying ‘sick’

companies, referring to those companies that continue to operate despite incurring losses for 2

years, or has four successive defaults on its debt service obligations, or taxes in arrears for 1-2

years. A sample of 18 sick and 18 healthy companies in the period 1976-95 was used, and seven

ratios were shortlisted. The Type I accuracy was 87.1% and Type II error was 86.6%. Validation

on a hold out sample of 20 healthy and 28 sick companies was performed and the results verified

the efficacy of the model. Table 4 lists the coefficients of the discriminant analysis by Bhatia.

Y=
+6.56 Current Ratio
+ 3.26 Stock of Finished Goods / Sales
+6.72 Profit After Tax / Net Worth
+1.05 Interest / Value of Output
+6.56 Cash Flow / Total Debt
+ 3.26 Working Capital Management Ratio
+6.72 Sales / Total Assets

TABLE 4 (Source: Bhatia, U. 1988)

II. b. Choice of Regression

The seminal works in the field of default prediction studies were those by Beaver (1967)

and Altman (1968). Altman had used the Multiple Discriminant Analysis (MDA) technique for

creating the Z-Score, and for long, MDA was the general tool used in default prediction studies.

After many scholars pointed out two drawbacks of the method- 1) MDA assumes that the

independent variables are multivariate normally distributed 2) Variance-Covariance matrices are

equal across defaulted and non-defaulted firms (McLeay and Omar 2000), Ohlson (1980) for the

first time used a logit regression for default prediction. While his model had a lower

classification accuracy than Altman’s Z and Z’’, the reasons for using a logit model in default

prediction were powerful.

III. MODEL DEVELOPMENT

III. a. Data Set

Our analysis uses financial data from 55 companies of which 21 are defaulted companies

and 34 are non-defaulted. While the early models used equal number of defaulted and non-

defaulted firms, in later studies, the number of defaulted companies in the set was chosen so that

the prior probability input was the same as the expected average default rate (Altman & Sabato

2006). For our set, the prior probability is 38%, which is considerably higher than 5.3%, the

overall default rate for CRISIL-rated firms (includes approximately 6400 Indian firms).

Moreover, since we could match at most 34 non-defaulted companies to the 21 defaulted

companies, we kept 55 companies in our set.

 The companies were identified using two resources, Fitch’s Update of Indian FCCB

Redemption for FY2013 and cases registered with the Board for Industrial & Financial

Reconstruction (BIFR India). The financial data was collected from company filings with the

Bombay Stock Exchange, India. The defaults in the data set occurred between 2009 and 2012.

The set contains companies from the following industries- Pharmaceuticals, Construction &

Contracting, Telecom Equipment, Telecom Services, Coke Manufacturing, Computer Software,

Computer Hardware, Textiles, Edible Oils and Solvents, Ceramics, Sponge Iron, Mining &

Minerals, and Sugar.

The non-defaults were matched with the defaulted companies with respect to year of

default, industry, and either size of sales or size of total assets in order to establish comparability.

The size of sales for the companies in the set falls in the range of 15 million USD to 1.5 billion

USD. All financial ratios were collected from a year prior to default, so the model developed is a

1-year default prediction model. The y variable was taken to be 0 for non-defaults and 1 for

defaulted companies.

III. b. Selection of Variables

While there are a large number of ratios to choose from, we collected 15 financial ratios.

These ratios were collected across five categories – Leverage, Liquidity, Profitability, Activity

and Coverage. The different categories were selected in order to capture different measures of a

company’s operations as explained in Altman’s paper. Within each category, some of the ratios

are the ones developed by Altman for his original Z-Score model, and others are common ratios

used in the general discipline of Accounting.

A list of the ratios used in the analysis is presented in Table 5.

 Ratio Category Variables Used

1. Long Term Debt / Book Value Equity

2. Debt / EBITDA
Leverage
3. Short Term Debt / Book Value Equity

1. Current Assets / Current Liabilities

Liquidity 2. Cash / Total Assets

3. Working Capital / Total Assets

1. Gross Profit / Sales

Profitability 2. EBITDA / Total Assets

3. Net Income / Total Assets

4. Retained Earnings / Total Assets

1. Sales / Total Assets

Activity 2. Accounts Receivable / Sales * 365

3. Accounts Payable / Cost of Goods Sold * 365

1. EBITDA / Interest Expenses

Coverage 2. EBIT / Interest Expenses

TABLE 5
III. c. Logistic Regression & Results

After running multiple combinations of different number of variables and using forward

and backward stepwise logistic regression, we developed the model shown in Table 6. The signs

of the coefficients are consistent with our expectations; we expect higher Short Term Debt /

Equity, lower EBITDA / Total Assets and lower Reserves / Total Assets to predict a higher

chance of default.

Log ( pd / (1-pd) ) =
+2.805
+ 0.293 Short Term Debt / Book Value Equity
-19.869 EBITDA / Total Assets
-5.473 Retained Earnings / Total Assets

TABLE 6

The p-values of the coefficients are lower than .035 indicating that there is strong

statistical evidence of a relation between the variables and the default event. Table 7 gives the p-

values.

VARIABLE COEFFICIENT STD ERROR OF COEFF Z VALUE P VALUE

Constant 2.80505 1.25367 2.24 0.025
Short Term Debt / BV Equity 0.29306 0.12726 2.30 0.021
EBITDA / Total Assets -19.8693 9.06304 -2.19 0.028
Retained Earnings / Total Assets -5.47297 2.57488 -2.13 0.034

TABLE 7

The Deviance test, an equivalent of the sum of squares of residuals in Ordinary Least

Squares for logistic regression, is also statistically significant with a p-value of .98. A low p-

value for the Deviance test indicates that the predicted probabilities deviate from the observed
probabilities in a manner that the binomial does not predict. The Log-Likelihood test, an

equivalent of the F test, has a p-value of 0 up to three significant digits, providing evidence that

there exists a significantly strong relation between the selected variables and the default event.

With regard to misclassification rates, for a cutoff of 0.5, the Type I Error, cases when

the model predicts a non-default when the firm defaulted is 9.52% and the Type II Error, cases

when the model predicts a default when the firm in fact did not default, is 8.82%.

III. d. Validation Results

Given the size of the sample we found it appropriate to use Lachenbruch’s method of

leaving-one-out validation. As per validation results, we found that the model has an accuracy of

89.09%. The 10.91% error comes from 3 Type I and 3 Type II errors in the sample of 55

companies, where we used the same cutoff score of 0.5.

III. e. Running Z’’ on the Sample

After running the ratios from the Z’’ model, we get a good model that works well on the

data set and gives statistically significant results. Table 8 shows the coefficients along with the p

values (all < 0.05) obtained by regressing the Z’’ ratios on the India sample.

VARIABLE COEFFICIENT STD ERROR OF COEFF Z VALUE P VALUE

Constant 4.79226 2.11856 2.26 0.024
Working Capital / Total Assets 5.99765 3.03577 1.98 0.048
Retained Earnings / Total Assets -7.39158 2.81348 -2.63 0.009
EBIT / Total Assets -30.2255 12.8587 -2.35 0.019
Book Value Equity / Total Liabilities -16.1025 7.83849 -2.05 0.040

TABLE 8
 The Deviance Test for this model is significant at a p-value of .97. The sign of the

Working Capital / Total Assets is of concern since it is contrary to expectation (i.e. a higher

WC/TA ratio should give lower probability of default); the signs of other variables are

consistent. For calculating the misclassification rates, we chose a cutoff of 0.09 which gave a

Type I error of 19% and a Type II error of 8.82%.

IV. CONCLUSION

Developing the Z-Score for Indian companies gives a statistically significant model. The

validation of the model suggests 89.09% accuracy. The model can be further improved with a

different set of potential variables and also by the inclusion of qualitative variables.
REFERENCES & SOURCES

Altman, E. I. ‘Financial Ratios, Discriminant Analysis and the Prediction of Corporate

Bankruptcy’, Journal of Finance, Vol. 23, No. 4, 1968

Altman, E. I., J. Hartzell, and M. Peck, ‘A Scoring System for Emerging Market Corporate
Debt’, Salomon Brothers, 1995

Altman, E. I., and G. Sabato, ‘Modeling Credit Risk for SMEs: Evidence from the US Market’,
2006

Beaver, W., ‘Financial Ratios as Predictors of Failure’, Journal of Accounting Research, Vol. 4,
(Supplement), 1967

Bhatia, U., ‘Predicting Corporate Sickness in India’, Studies in Banking & Finance 7: 91-103,
1988

Joshi, Mukherjee, Valecha, ‘Update of Indian FCCB Redemption, FY 13’, India Ratings &
Research, 2012

Lachenbruch, M. Ray Mickey, ‘Estimation of Error Rates in Discriminant Analysis’,

Technometrics, 1968

Lehmann, B., ‘Is it Worth the While? The Relevance of Qualitative Information in Credit rating’,
Helsinki, 2003

McLeay, S., and A. Omar, ‘The Sensitivity of Prediction Models to the Non-Normality of
Bounded and Unbounded Financial Ratios’, British Accounting Review, 2000

Ohlson, J., ‘Financial Ratios and the Probabilistic Prediction of Bankruptcy’, Journal of
Accounting Research, Vol. 18, 1980

Panicker, Vemuri, Vasu, Kacker, ‘CRISIL Default Study’, CRISIL Annual Default & Ratings
Transition Study, 2012