8 The determinants of credit ratings: Australian evidence.

Abstract:

We examine the impact that various financial and industry variables have on credit ratings issued for Australian firms by Standard and Poor's. Our ordered probit model indicates that interest coverage and leverage ratios have the most pronounced effect on credit ratings. Profitability variables and industry concentration measures are also important. Financial variables are helpful in discriminating between A- and BBB-rated firms, but are less precise in separating AA--and A-rated firms. We also document a consistent trend towards lower ratings--the standard required to achieve a particular rating is increasing over time.

Keywords:

CREDIT RATINGS; ORDERED PROBIT MODEL.

1. Introduction

A corporate credit rating is an independent evaluation of a firm's ability to make debt payments in a timely fashion. A credit rating may be assigned to a particular debt issue, or it may indicate the general ability of the firm to meet its obligations. When issuing a credit rating, the rating agencies incorporate quantitative and qualitative information obtained from public and private sources. The agency's analysis then produces a credit rating that represents the current opinion of the agency regarding the credit worthiness of an obligor.

Credit ratings are of great practical importance, as they impact the firm's cost of debt, its financing structure, and even its ability to continue trading. In a recent survey of chief financial officers, Graham and Harvey (2001) find that consideration of credit ratings is the second most important factor, after the maintenance of financial flexibility, in the decision to issue more debt. From the investment perspective, credit ratings are an independent source of credit analysis used at the operational level and for regulatory purposes and to limit agency costs. For example, credit ratings restrictions can be written into the mandate for the management of large public funds, restricting investment in bonds with lower credit ratings. Credit ratings are also used to manage counterparty default risk. For example, many Australian financial, investment, and energy companies commonly have credit policies that prevent the firm from entering contracts with counterparties that do not have an investment grade rating--at least without specific approval from the Board. Moreover, if a firm's credit rating is downgraded to sub-investment grade, this may trigger escape and penalty clauses under the terms of contracts with counterparties. Finally, the firm's cost of debt financing depends on its credit rating--debt becomes more expensive as credit ratings deteriorate.

For all of these reasons, firms carefully consider the impact on the credit rating every time they make a significant financial decision. For example, increasing leverage to fund an acquisition may have a negative impact on the firm's credit rating, which may in turn affect the firm's cash flows (as credit clauses are triggered and collateral or credit support is required). Clearly, the impact of any firm action on its credit rating is a primary consideration. Therefore, a framework for determining the likely effect of an increase in leverage (say) on the firm's credit rating is of considerable practical importance, as well as being of academic interest. The goal of this paper is to quantify the effect of the firm's decisions on its credit rating via financial and industry variables.

Credit ratings are assigned by commercial rating agencies, such as Standard and Poor's and Moody's, based on publicly available information and on confidential information provided to them by the firm. The factors that they consider in the process of determining a firm's rating, while not perfectly transparent, are generally taken to capture both financial information, such as financial accounting ratios, and non-quantifiable or 'subjective' factors, such as management quality, succession planning, industry characteristics, the competitive position of the firm, and so on.

While ratings agencies emphasise that both financial and non-financial factors matter, the academic literature has focused primarily on the ability of financial ratios to predict ratings. Particular attention has been paid to financial ratios that measure the firm's ability to service debt (interest coverage and cash flow measures, such as debt coverage), profitability (return on capital) and leverage (debt to assets). These financial ratios have also been identified by the ratings agencies themselves (e.g. Standard and Poor's 2003). Moreover, empirical research indicates that these ratios are able to explain a significant amount of the variation in corporate credit ratings.

In this paper, we examine the relationship between Australian credit ratings and a set of financial ratios and industry variables, such as systematic risk and industry concentration. We follow the existing literature (e.g. Blume, Lim & Mackinlay 1998) in using an ordered probit model to map our explanatory variables into credit ratings assigned by S&P to Australian firms between 1995 and 2002. We find that interest coverage and leverage ratios have the most pronounced effect on credit ratings, and that profitability variables and industry concentration measures are also important. We also document a consistent trend towards lower ratings--the standard required to achieve a particular rating is increasing over time. For instance, 20 out of the 22 firms that have a rating of A and above in 1995 would have earned a lower rating had the 2002 rating standard been applied. Over time, a stronger set of financial ratios is required to maintain the same credit rating.

This paper extends the analysis of Blume, Lim and Mackinlay (1998) in a number of directions. First, we apply the ordered probit model to Australian data. We are able to confirm that the tightening of ratings standards is an international phenomena and is not specific to U.S. capital markets. Second, we include a number of industry variables, as well as firm-specific financial variables. For example, our market concentration variable seeks to quantify the competitive position of the firm, which is often referred to as a qualitative factor that is examined by ratings agencies. Third, we examine the performance of the model, with tests of forecast accuracy and a series of out-of-sample tests. This is done to confirm that the strong statistical results are not simply the result of over-fitting the data in-sample. This leads us to identify that the model does a better job at discriminating between the lower rating categories--the classification of higher ratings appears to be driven by qualitative (non-financial) considerations.

The remainder of this paper is organised as follows. In section 2, we discuss the position of this paper within the related literature. Section 3 describes the data and the construction of our explanatory variables. Section 4 develops the ordered probit model that we use to map financial and industry variables to credit ratings. The results are discussed in section 5, and section 6 concludes.

2. Relevant Literature

The finance literature contains a number of papers that use financial accounting ratios and other publicly available information to predict credit ratings and corporate bankruptcies. Although ratings agencies argue that ratings are not simply surrogate estimates of corporate bankruptcy probabilities, ratings and corporate defaults are closely related in that bonds with lower ratings are more likely to default than bonds with higher ratings.

The bankruptcy prediction literature is centred around the seminal work of Altman (1968), which seeks to explain corporate bankruptcy status in the United States based on accounting and financial variables. Altman develops the 'Z score', using multiple discriminant analysis (MDA). This approach discriminates between two groups (bankrupt and non-bankrupt firms) by deriving a linear combination of certain pre-determined factors that best discriminates between the two groups. In particular, Altman demonstrates that a linear combination of financial ratios (including measures of leverage, sales efficiency, cash flow, and operating capital) performs well in discriminating between firms that went bankrupt in the subsequent year and those that did not.

The prediction of credit ratings is more difficult than the prediction of bankruptcy in two respects. First, there are more than two categories. Moody's and Standard and Poor's both use a scale with nine major grades. Although data limitations often prevent analysis of the bottom grades (e.g. there are very few bonds rated CC or C in any sample of credit ratings), the analysis will examine several grades of credit rating. This means that a simple binary discriminant analysis will be inappropriate and a more complex econometric technique will be required. Second, credit ratings are based not only on measurable quantitative data, but also on qualitative assessments of the firm's management team, corporate strategy, and industry position. This can make the mapping between financial data and credit ratings more difficult to quantify.

Nevertheless, there is a substantial literature that seeks to quantify the relationship between financial and industry data and credit ratings. This literature has progressed with the development of econometric techniques for analysing categorical dependent variables. In an early study, for example, Pogue and Soldofsky (1969) use a regression-based approach. They compare two of four rating categories at a time to determine the probability that a bond will have the higher of two ratings by assigning numerical values (1, 0) to credit rating categories. This probability is determined as a function of measures of leverage, profitability and size. They find that leverage and profitability have the greatest impact on bond ratings. In addition, the accuracy of their model in predicting ratings improves when the difference between the categories compared is greater (AAA vs. [BBB.sup.+] as opposed to [A.sup.+] vs. [BBB.sup.+]). While they avoid the problems associated with using ordinal data in a regression framework, by estimating the differences between categories separately, they do not make full use of all available information simultaneously.

Pinches and Mingo (1973) adopt a two-stage approach to assign ratings to particular bond issues. They begin by screening a set of potential independent variables using factor analysis and arrive at a set of six acceptable factors--subordination, years of consecutive dividends, issue size, and three financial ratios. The second stage of their analysis involves the use of multiple discriminant analysis (MDA) to classify bonds into rating categories by constructing linear functions that distinguish between categories by maximising the ratio of between-category variance to within-group variance. Their model correctly assigns ratings to approximately 69% of their sample, though the accuracy of their approach drops slightly when used on a 'holdout' sample. In general, while MDA considers each rating category as a different outcome, it does not capture the ordinal nature of credit ratings--the fact that a AAA-rated firm is more creditworthy than one rated AA.

More recent empirical studies have estimated the relationship between the explanatory variables and credit ratings using the ordered probit technique. This technique has been developed for settings in which the dependent variable is discrete, and takes on a finite number of values possessing a natural ordering. This makes it particularly appropriate for the study of credit ratings. Kaplan and Urwitz (1979) employ an ordered probit methodology on a sample of bonds with unchanged Moody's ratings and another of newly issued bonds. The independent variables include firm size, subordination, market beta and several financial ratios. The results indicate that subordination, total assets, leverage and market beta could predict two thirds of a hold-out sample of newly issued bonds. This approach is also used in the recent study by Blume, Lim and Mackinlay (1998) to classify rated bonds into rating categories as described by S&P. They modify the ordered probit model to cater for the non-linearity in the relationship between ratings and variables such as interest coverage. Furthermore, the use of panel data allows them to study any change in rating standards over time. In particular, they examine the popular view that credit quality has been progressively deteriorating. The basis of this view is the fact that recent years have seen persistently more credit rating downgrades than upgrades. This is consistent with a deterioration of credit quality, but also with an increase in the standards that are applied by ratings agencies. Blume, Lim and Mackinlay (1998) conclude in favour of the latter explanation. We examine this issue using Australian data and conclude that the increase in standards demanded by ratings agencies is an international phenomenon.

Only one study has examined credit ratings using Australian data, but it does not examine the relationship between credit ratings and financial and industry variables. Matolcsy and Lianto (1995) examine the incremental information content of bond rating revisions on stock prices, after controlling for accounting information, using a cross-sectional regression approach. The marginal effect on abnormal performance of a downgrade is a significant -14.6%, while upgrades have no incremental content. Their finding that only rating downgrades have information content is consistent with other studies.

3. Sample Selection

Our initial sample consists of Australian firms that have been rated by Standard and Poor's (S&P) between 1995 and 2002. Long-term domestic issuer credit ratings are obtained primarily through the S&P Ratings Direct website. Historical ratings for delisted firms not available from this website, are either obtained directly from S&P or from Bloomberg News Service. Concurrent and complete financial report information for the period 1993 to 2002 is collected from Bloomberg News Service. (1) Australian financial reports are commonly released two to three months after the balance date. Therefore, we define an annual financial report to be contemporaneous with the rating if it relates to the financial year-end that occurs three to fifteen months prior to the rating. This ensures that any changes based on information released in the annual report are captured in the corresponding rating. Also, years for which complete financial information was unavailable are excluded from the sample. In addition, banks and insurance firms are not included in the sample due to significant differences in accounting standards and the interpretation of several financial ratios (leverage in particular). The total number of ratings observations that meet the above criteria, and therefore form our initial sample, is 392.

The focus of this study is restricted to firms with ratings that are considered to be investment grade. This is because the vast majority (more than 90%) of Australian corporate bonds have investment grade ratings. We excluded 31 sub-investment grade ratings from our sample. These ratings varied from BB+ to CCC. Such a small number of observations spread across several rating categories would render estimation and interpretation speculative at best. Thus, our final sample consists of 361 observations of investment grade ratings. Table 1 provides summary statistics over time and by industry. Most observations are clustered in the A and BBB categories. Another interesting feature of our sample is the increase in the proportion of lower rated issuers over the sample period. The number of AA rated firms decline from a high of six in 1996 to one in 2002, while the number of BBB firms doubles over the same period. This may be partly due to the increased coverage over the period, but may more likely be a symptom of a decline in the average credit rating of the sample over time.

A summary by industry is presented in panel B of table 1. This indicates that the majority of the firms are clustered in the manufacturing industry, with more than double the number of the next most populous industry group, mining. The data are presented at the industry group level, which is rather broad. (2) Nevertheless, industry differences are noticeable at a very basic level through the clustering of observations within rating categories. For example, firms in transport and storage are clustered in the BBB category in contrast to communication services, which tend to cluster in the AA category.

3.1 Financial Ratios

Consistent with information provided by Standard and Poor's (2003) and with the approach used by Blume, Lira and Mackinlay (1998), we model the firm's credit rating as a function of its financial characteristics given by interest coverage, profitability and leverage. Credit ratings tend to be highly sensitive to the firm's interest coverage ratio--firms with higher coverage ratios are likely to have higher credit ratings. Cash flow or debt coverage ratios, such as free cash flows relative to total debt, are also crucial in credit analysis as this provides an indication of the firm's present ability to service its debt and meet its other financial obligations. A low cash-flow-to-total-debt ratio may be symptomatic of higher risk and a signal of weak prospects. High cash flow relative to total debt is associated with higher credit ratings. Another key factor is leverage, measured as debt to total assets. Other things equal, the greater the degree of financial leverage, the smaller the cushion the firm has with respect to any unanticipated changes to its fortunes. Higher leverage is usually associated with lower credit ratings. Profitability is another signal of the firm's ability to generate cash to meet its financial obligations--a high profitability ratio is more likely to be associated with a better credit rating. (3)

In table 2, we present descriptive statistics by rating category for a range of financial ratios that are likely to be relevant to the credit rating process. Some of these variables have been truncated or adjusted to remove the effects of extreme outliers, as explained in Section 3.5. The average interest coverage and profitability ratios are greater for the firms with higher credit ratings. As expected, AAA and AA rated firms have the highest debt coverage, but (curiously) BBB rated firms have better debt coverage than A-rated firms, on average. However, the average EBITDA interest coverage ratio (which measures a similar aspect of firm performance) for A-rated firms is nearly double that of BBB firms. This indicates that ratings agencies may place more weight on interest coverage ratios than on cash-flow-to-debt ratios. Of course, computing the relative importance of all of these financial ratios is one of the main goals of this paper.

Long-term debt leverage is generally higher for firms with lower ratings, while total debt leverage is similar for firms in the three ratings categories. This suggests that ratings agencies may place more weight on long-term debt relative to short-term debt. This is to be expected, particularly if short-term interest coverage ratios are solid.

In results not presented in this paper, we find that for some financial ratios the mean and median differ considerably. This indicates the presence of skewness and the potential for outliers. In particular, our sample contains a small number of extremely high interest coverage and debt coverage ratios. For example, the maximum EBITDA interest coverage ratio is 156.07. This, of course, occurs when a firm that has very little debt is highly profitable. Since the average AAA/AA-rated company in the sample has an EBITDA interest coverage ratio of 15.52, whether a particular company has a ratio of 100 or 156 probably makes little difference to a rating agency. We therefore truncate this variable in our modelling, as described in Section 3.5.

3.2 Systematic Risk

We follow past research by including the equity beta as a measure of systematic risk. Blume, Lim and Mackinlay (1998), for example, note that 'The hypothesis is that a firm will be less able to service its debt for given accounting ratios as its equity risk increases' (p. 1395). Other things equal, a firm with a higher equity beta is expected to have a lower credit rating. Of course, this effect may be difficult to identify in the data because other things are not equal. In particular, firms in industries that have high systematic risk may make more conservative financing choices and have stronger financial ratios.

Firm level betas are obtained from the Centre for Research in Finance database at the Australian Graduate School of Management. The data is used to calculate annual portfolio betas, constructed using the MSCI--S&P Global Industry Classification Standard (GICS) at the industry group level. To calculate industry betas, we include all the listed Australian firms that have a market capitalisation in excess of $50 million as very small firms are more prone to thin trading which is likely to result in estimation error. Industry betas are computed by weighting firm level betas by market capitalisation as at June 30, for each sample year. This is used as an additional explanatory variable for the credit rating of each sample firm. Table 2 indicates that lower rated firms do tend to come from industries with higher betas, although the differences are economically small. In particular, the distributions of industry betas for A- and BBB-rated firms are almost indistinguishable.

3.3 Industry Competition

Ratings agencies suggest that credit ratings should also depend, in part, on the firm's business environment. Numerous industry characteristics including competitiveness, barriers to entry, exposure to technological change, regulatory environment and vulnerability to economic cycles can have a significant influence on the level of business risk a firm faces. For example, a firm in a monopolistic industry faces less uncertainty and can afford a greater degree of leverage, other things equal, than a firm in a highly competitive technology-based industry. Ultimately, the inclusion of industry differences into a model of credit ratings as a measure of industry effects may improve the explanatory power. For instance, Iskander and Emery (1994) show that industry factors (using broad industry dummy variables) play a significant role in credit ratings, even after controlling for financial characteristics.

In this study, we capture the effects of industry factors using a measure of competition in the industry. The degree of competition a firm faces within an industry affects the level of operational uncertainty and the amount of excess profits it can generate. Firms facing less competition are less sensitive to economic cycles and are more able to access capital at a lower cost. From a credit perspective, greater stability of cash flows translates into a lower level of business risk and hence lower credit risk. Industry competitiveness can be observed through micro-economic modelling of the market share of participants, homogeneity of products, and so on.

Demsetz (1973) notes that the effects of industry concentration benefit the large firms within an industry rather than small firms. Porter (1979) explains this to be a result of 'mobility barriers' that are a function of investment in R&D, advertising, and economies of scale that protect large firms. As a result, high concentration in an industry acts to the benefit of large firms and to the detriment of small firms and new entrants. Issuer credit ratings in Australia are made where there is a commercial demand, and this leads to coverage of predominantly larger firms. It is therefore likely that the effect of industry concentration will be positive for the available sample of rated entities--our sample consists of large firms that benefit from being in a concentrated industry. Thus, we expect that within our sample (of large firms) those firms in more concentrated industries will enjoy higher credit ratings, other things equal.

We measure industry competitiveness by the four-firm concentration ratio, the percentage of market production supplied by the four largest firms in the industry, published by the Australian Bureau of Statistics (ABS) in 'Industry Concentration Statistics'. The ABS uses the Australia and New Zealand System of Industry Classification (ANZSIC) standard rather than the GICS standard. The industry concentration data is available at the ANZSIC sub-industry division comprising 42 sub-industries. This ratio measures the contribution to total industry sales (or service) revenue by the four leading firms in that industry. The statistics are only available for financial years ending 1999 through to 2001. We assume that the concentration statistics have remained relatively stable since 2001 as the firms in our sample are predominantly the largest firms in the Australian market.

Table 2 indicates that lower rated firms do tend to come from industries with lower industry concentration ratios, although the differences are economically small. In particular, the distributions of industry concentration ratios for A- and BBB-rated firms are almost indistinguishable.

3.4 Transformation of Financial Ratios

In assigning credit ratings, the agencies adopt a longer-term perspective by using a process known as 'rating through the cycle' (e.g. Standard and Poor's, 2003). This is usually implemented by considering three-year averages of relevant financial ratios rather than just the most recent observations. We follow this process by using three-;fear averages of the financial ratios in our model using data from 1993-2002. (4)

The ordered probit approach (that we describe in detail in the next section) models credit ratings as depending on a linear function of the independent variables. In results not presented in this paper, we find that some of the explanatory variables are highly skewed. In particular, there are a number of extreme outliers in the interest and debt coverage variables caused primarily by companies with very small amounts of debt--a standard low-denominator problem. As a result, the assumption of a linear relationship between the financial variable and the credit rating may be violated for these variables. In addition, the interest coverage ratio can be negative when earnings are negative. The magnitude of the interest coverage ratio may not be meaningful for observations with a small interest expense and large negative earnings. To highlight the problem at hand, according to S&P's criteria, an increase in EBIT interest coverage from 2 to 6 would be considered significant, and potentially lead to an upgrade from BB to A, other things equal. In contrast, an increase from 96 to 100 would carry little significance and is unlikely to have any effect on a firm already rated AAA.

The interest coverage and debt coverage variables are therefore modified to overcome these limitations. We use a procedure whereby ratios greater than that of the median AAA rated firm and lower than the median BBB rated firm are truncated. First, any annual observation of interest coverage less than negative one is set to negative one. Any annual observation of EBIT interest coverage greater than 25 is set to 25, on the basis that increases in value beyond 25 convey no additional information. Similarly, EBITDA interest coverage is capped at 30. The truncation values of 25 and 30 for interest coverage are also supported by empirical research by Blume, Lim and Mackinlay (1998). They find that increments in interest coverage from 0 to 5 tend to be most informative, and increases in interest coverage beyond 20 provide no statistically significant incremental information in the modelling of credit ratings. That is, our transformations are based on empirical observation and the guidance offered by S&P.

Debt coverage variables are also plagued by extreme observations. A debt coverage ratio of one indicates coverage of one hundred percent, which is indicative of a AAA-rated issuer. Increments beyond one indicate the firm's cash flows are more than sufficient to cover the entire amount of debt on issue. Extreme observations, positive and negative, are less likely to be informative. Our sample contains a number of extreme debt coverage observations which skew the sample. For example, Incitec in 2001 and 2002 had very high operating cash flow coverage of debt of 19 and 13 times, respectively. The high ratio was a result of a low level of debt financing relative to previous years combined with a stronger than historical cash flow. For this reason, observations of the cash flow and operating funds debt coverage ratios greater than five are set equal to five and those less than negative one are set to negative one. Finally, negative leverage ratios (due to negative reported net debt) are set to zero.

Table 2 presents a full set of descriptive statistics for our transformed interest coverage, debt coverage, and leverage ratios and the other (untransformed) variables that are used in the remainder of the paper.

4. Methodology

In this paper, we seek to map financial and industry variables to credit ratings. While creditworthiness is a continuous variable that is economically more meaningful, it cannot be observed. Therefore we focus on credit ratings, which should relate closely to creditworthiness. The structure of credit ratings, however, presents several econometric issues. First, the ratings are discrete rather than continuous. This means that standard least squares techniques are inappropriate--some form of limited dependent variable technique is preferred. Second, there is a natural ordering to the ratings--AA is a higher rating than A, which is a higher rating than BBB. This makes multiple discriminant analysis inappropriate. Third, the ratings categories are not necessarily evenly spaced--the BBB rating category, for example, may traverse a wider range of financial and industry variables than the other categories. Kaplan and Urwitz (1979) and Blume, Lim and Mackinlay (1998) discuss many of these issues.

The approach we adopt in this paper is the ordered probit model as proposed in Hausman, Lo and Mackinlay (1992) and implemented in a credit ratings setting by Blume, Lim and Mackinlay (1998). This statistical technique is designed for empirical studies in which the dependent variable is discrete and takes on a finite number of values that possess a natural ordering. We examine three ordered rating classes: AAA/AA, A and BBB. We combine the AAA and AA rating classes as there are few Australian corporations with AAA ratings and the characteristics of AAA- and AA-rated companies are similar. The ordered probit approach recognises: (i) that A is higher than BBB, not merely different (a problem with multiple discriminant analysis); and, (ii) that the three classes are not necessarily evenly spaced (a problem with regression analysis--that is, the difference between the AAA/AA class and the A class may be more or less than the difference between the A class and the BBB class.)

In our setting, the dependent variable, the credit rating of company i in year t, [Y.sub.it], takes one of three values--1 if the company is rated AAA or AA, 2 if the company is rated A and 3 if it is rated BBB:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

The values of [Y.sub.it] are censored data in that they can take only one of three possible values. To relate these credit ratings to our explanatory variables, we define

[Y.sup.*.sub.it] = [X.sup.'.sub.it] [beta] + [[epsilon].sub.it] (2)

where [X.sub.it], is a vector of explanatory variables, [beta] is a vector of coefficients to be estimated, and [[epsilon].sub.it], is a standard normal residual. Here, credit ratings depend on certain quantifiable factors given by [X.sub.it] and unobservable factors given by [[epsilon].sub.it]. The ordered probit model relates the unobserved variable [Y.sup.*.sub.it] to the observed credit rating [Y.sub.it] as follows:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (3)

Thus, the probability of a particular set of explanatory variables being associated with a particular credit rating is given by

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (4)

where [[epsilon].sub.it] ~ N(0,1).

The parameters [[alpha].sub.1] and [[alpha].sub.2] are chosen to reflect the proportion of observations in the sample that fall within each rating category. A higher value of [[alpha].sub.1] will increase the number of observations that are classified as AAA/AA. Symmetrically, a higher value of [[alpha].sub.2] will decrease the number of observations that are classified as BBB. These parameters, therefore, depend on the proportion of observations in the sample that fall into each of the three rating categories. Also, it is clear from equation (4) that higher values of the linear combination of explanatory variables, [X.sub.it] [beta] imply that a lower credit rating is more likely. This implies that an increase in an explanatory variable with a positive (negative) coefficient reduces (enhances) credit quality. We use standard maximum likelihood techniques to estimate the coefficients ([beta]) and the values of [[alpha].sub.1] and [[alpha].sub.2].

We assess the significance of individual parameters using standard statistical tests. The goodness of fit of the estimated probit model can be evaluated with reference to the percentage of sample outcomes it predicts accurately. To calculate this measure, the expected outcomes from the model are compared with the actual sample. A matrix of model predictions versus actual observations allows a more detailed evaluation of the model's predictions. A serious error is defined by a prediction that is two categories different to the observed rating--for example, a forecast of a rating AA when the actual rating is BBB.

To further aid in interpretation of the coefficients, the product of the standard deviations of the explanatory variables and model coefficient is also included. This product is a measure of economic significance. It represents the change in the conditional expectation of the credit rating, in response to a one standard deviation change in the underlying explanatory variable. A comparison of this product to the size of the partition (e.g. [[alpha].sub.1] - [[alpha].sub.2]) provides an indication of the economic significance of the respective coefficients. Another measure of economic significance that we present is the number of credit ratings that would change categories if the financial ratios change, in either direction, by one standard deviation. This provides us with some insight into the relative importance of these ratios in the determination of a firm's credit rating.

5. Discussion of Results

We begin our analysis by evaluating the determinants of Australian credit ratings in a series of univariate ordered probit models. These univariate tests of the key financial ratios and industry variables provide an evaluation of the hypothesised relationships and an indication of relative economic significance. The results are presented in table 3. A positive coefficient indicates a lower credit rating (recall that our category 1 contains AA-rated firms and category 3 contains BBB-rated firms). The cash flow variables and total debt leverage are insignificant and industry beta is only marginally significant. All other coefficients are statistically significant and in the hypothesised direction.

The parameters [[alpha].sub.1] and [[alpha].sub.2] play the role of partitioning a standard normal distribution into three regions--one for each credit rating class. Consider, for example, their interpretation in relation to EBIT interest coverage. The ordered probit model suggests that any firm for which the product of the [beta] coefficient and the measured EBIT interest coverage is greater than -0.2829 is likely rated in the BBB class. This corresponds to an EBIT interest coverage ratio of 4.6, so firms with EBIT interest coverage of about five or more tend to have credit ratings at A or above. Similarly, firms for which the product of the [beta] coefficient and the measured EBIT interest coverage is less than -1.6863 is likely rated in the AAA/AA class. This corresponds to an EBIT interest coverage ratio of 27. However, we have capped EBIT interest coverage at 25 in our analysis (see Section 3.5). Therefore, the univariate ordered probit model does not predict that any firms are in the AAA/AA class, based solely on the EBIT interest coverage ratio. This is because AAA/AA firms cannot be discriminated from A-rated firms on the basis of EBIT interest coverage. Nearly 20% of the A-rated firms in our sample have EBIT interest coverage ratios in excess of the median EBIT interest coverage ratio of the AAA/AA-rated firms. This, together with the fact that our sample contains 152 A-rated firms and only 34 with AAA/AA ratings, is why the model does not predict any firms in the top category. To do so, would require an increase in the estimate of [[alpha].sub.1]. However, this would cause more problems than it would solve, in that a relatively larger number of A-rated firms would then be mis-classified as belonging to the higher rating class.

Economic significance can be assessed with reference to the final two columns in table 3. The second from last column reports the product of the estimated coefficient and the standard deviation (across all observations) of the independent variable. This provides an indication of how much the latent continuous variable [Y.sup.*.sub.it] would move if the independent variable were one standard deviation greater. This should be interpreted with reference to the estimates of [[alpha].sub.1] and [[alpha].sub.2]. For example, the median EBIT interest coverage ratio in our sample is about 4. Therefore, at this median value, [X.sup.'.sub.it][beta] is -0.248 (4x-0.062), which is greater than the estimate of [[alpha].sub.2]. This implies that the model would assign this median firm to the BBB category. An increase in the EBIT interest coverage ratio of one standard deviation would reduce the estimate of [X.sup.'.sub.it][beta] by 0.260 to -0.508. Since this is less than the estimate of [[alpha].sub.2], the firm would be assigned to the A category. In this case, the EBIT interest coverage ratio is economically significant.

The final column of table 3 reports the proportion of firm credit ratings that are correctly assigned by the univariate ordered probit models. In our setting, there are three ratings categories so a strategy of randomly assigning firms to a ratings category has an expected success rate of 33%. However, 175 of our 361 observations (48.48%) are from the BBB category. Therefore, we test the success rates of our models against a naive strategy that assigns all firms to the BBB category.

Interest coverage, profitability (operating margin), and long-term debt leverage display the highest prediction success rates. In contrast, both debt coverage ratios and total debt leverage show a prediction success rate of 42%, which is superior to the strategy of random prediction but inferior to the naive BBB prediction. Further examination of these explanatory variables suggests the univariate model incorporating these variables predicted an A rating for the entire sample, so that the resultant prediction success rate reflects the proportion of A-rated firms in our sample. The second to last column confirms that these variables, in isolation, are economically insignificant.

In aggregate, the results in table 3 are consistent with the U.S. evidence presented in Blume, Lim and Mackinlay (1998) who report that interest coverage, profitability and long-term debt leverage provide strong explanatory power in the credit rating process.

Both industry variables have limited predictive ability. The coefficient relating to industry beta is statistically insignificant, its predictive power is inferior to a naive BBB prediction, and it is also economically insignificant. The coefficient relating to industry concentration is marginally significant, however the predictive performance and economic significance are poor. This is likely due to the fact that industry variables are factored into the credit rating decision only after the strength of the financial ratios are considered.

For this reason, we turn next to an expanded model in which a range of financial and industry variables are analysed simultaneously. We begin by reporting the correlation between these variables in table 4. The eight financial ratios can be grouped into four pairs that address interest coverage, debt coverage, profitability and leverage. Table 4 indicates that the correlation between the interest coverage, debt coverage, and leverage variables (shown in the shaded cells) is economically large and statistically significant. The two measures of profitability (Return on Capital and Operating Margin) are not highly correlated, indicative of the different aspects of profitability that each measures. In order to minimise the problems associated with multicollinearity, we estimate a model with only four ratios, one from each of the four pairs. We are guided by existing evidence, such as Blume, Lim and Mackinlay (1998), in our choice of variables. In particular, we adopt EBIT interest coverage, Operating funds to total debt, Operating margin and Long-term debt leverage as our explanatory financial variables. (5) In addition to these financial variables, our full model also includes the two industry variables; industry beta and industry concentration. Table 4 indicates that these two variables are significantly positively correlated. There is no a prori reason shy firms in more concentrated industries should have higher systematic risk. This positive correlation is more an artifact of market structures in Australia. For example the average beta and concentration ratios are 1.60 and 32% for media and entertainment firms, compared with 0.49 and 14% for food and beverage firms. For various structural reasons, there was more competition among low-beta industries in our sample.

Finally, Blume, Lim and Mackinlay (1998) also examine the potential effect of increasing standards being applied by credit rating agencies. Over the last 15 years, there has been a documented decline in the average credit rating issued for U.S. firms. This could be due either to: (i) a decline in the average credit quality of U.S. firms; or, (ii) an increase in the credit rating standards being applied. The U.S. evidence supports the latter. We test for this relationship in the Australian context by including a time variable. This variable takes the value of zero in the first year of our sample (1995), one in the second year of our sample, through to seven in the final year of our sample (2002). If there has been an increase in agency standards, the coefficient on this time variable would be positive. Thus, the same set of financial and industry variables would support a lower credit rating over time (recall that AAA/AA is category 1 and BBB is category 3 in our analysis).

Our full model, therefore, includes financial and industry variables and a time trend. The results of this expanded ordered probit model are presented in table 5, panel A. The significantly positive coefficient on the time variable supports the findings of Blume, Lira and Mackinlay (1998). That is, the rating standards used by agencies are increasing over time. This implies that any decline in average credit ratings is consistent with increasingly stringent standards being used by agencies when assigning ratings--conditional on the variables identified in this study. A firm would need to improve its financial and industry ratios over the course of our sample just to maintain its credit rating. Indeed, we find that of the 22 firms that have a rating of A or above in 1995, 20 would have been rated at least one category lower by the standards that were applies in 2002, had they maintained an identical set of financial and industry variables.

Table 5, panel A, also suggests that all of the financial and industry variables have a significant impact on credit ratings, and all but debt coverage are in the predicted direction. The results suggest that a higher debt coverage ratio (Operating funds/Total debt) is associated with a lower credit rating. The most likely explanation for the anomalous finding is the high degree of collinearity between debt coverage and interest coverage (see table 4). The coefficient on the interest coverage ratio has more than doubled relative to the univariate model in table 3. This increase in the effect of interest coverage is likely compensated for by the anomalous negative effect of debt coverage.

This is further evident in table 5, panel B, which documents the number of firms that would have their credit rating upgraded or downgraded if each financial variable were increased by one standard deviation. The first row, for example, shows that the model forecasts 357 ratings at A or BBB. Of these, the model would have assigned 331 to a higher rating had the interest coverage ratio been one standard deviation higher. (6) Similarly, the model assigns 187 firms to the AAA/AA and A categories. Of these, the model would have assigned 186 to a lower rating had the interest coverage ratio been one standard deviation lower. Table 5, panel B indicates that the EBIT interest coverage ratio has the most dramatic effect on credit ratings. However, multicollinearity between the various financial variables remains a concern, so we focus more on the predictive success of the model and a series of out-of-sample forecast tests.

Our full model correctly assigns credit ratings to 61.5% of our sample firms. Table 5, panel C compares our model forecasts with actual credit ratings. The model correctly predicts more than two-thirds of the A and BBB ratings. However, the model only assigns four firms to the AAA/AA category. This is due to the fact that (i) the financial and industry variables are quite similar for firms in the AAA/AA and A categories, and (ii) there are almost five times as many firms in the A category than in the AAA/AA category. Thus, a firm with strong financial and industry variables is (within our sample) statistically more likely to be an A-rated firm. Therefore prediction success (and our likelihood function) is likely to be maximised by assigning very few observations to the top category. This is also consistent with the findings of Blume, Lim and Mackinlay (1998) who conjecture that an omitted variable may explain which of the firms with strong financial ratios are assigned to the top category. For example, an assessment of management quality may differentiate between AA and A- rated firms. However, Blume, Lim and Mackinlay (1998) note that it is difficult to address this issue, 'without defining the omitted variable itself.' Nevertheless, we can conclude that financial and industry variables are able to discriminate between credit ratings for all but the relatively few firms in the highest rating category.

Another approach to evaluate the predictive ability of our model is to examine its out-of-sample forecast performance. This is done by estimating the ordered probit model (excluding the time trend variable) using data from seven years of our sample and using the estimates to predict the ratings of the remaining year. For example, we predict the ratings for 1995 using the parameters estimated from 1996-2002. We predict ratings for 1996 using parameters estimated from 1995 and 1997-2002, and so on. The results of these out-of-sample tests are presented in table 5, panel D. In all cases, the predictive ability exceeds 50%, and in four of the eight cases the forecast success rate is significantly higher than what would be achieved from a naive prediction that all firms have BBB ratings. Therefore, we conclude that the ordered probit model that includes firm-specific financial variables and industry betas and concentration ratios performs well in describing how Australian credit ratings are determined.

6. Conclusion

We examine the impact that various financial and industry variables have on credit ratings issued for Australian firms by Standard and Poor's. Our ordered probit model indicates that interest coverage and leverage ratios have the most pronounced effect on credit ratings. Profitability variables and industry concentration measures are also important. We also document a consistent trend towards lower ratings--the standard required to achieve a particular rating is increasing over time. These results serve to corroborate and extend similar evidence from U.S. markets presented by Blume, Lim and Mackinlay (1998).

While the ordered probit model performs well in terms of prediction success rates and out-of-sample forecast tests, its weakness lies in the ability of the model to discriminate between firms in the highest ratings categories. In particular, all firms rated A and above tend to appear similar in relation to the financial ratios and industry variables that we examine. Further work in this area should be directed at identifying variables that are able to distinguish between AA- and A-rated firms. Alternatively, if no such identifying variables can be uncovered, an investigation of the qualitative procedures employed in this regard by ratings agencies would be informative.

The authors wish to thank Diane Del Guercio, Tom Smith, an anonymous referee and seminar participants at the Australian National University for their comments. The Australian Research Council provided funds for this project under Grant DP03-42953.

(Date of receipt of final transcript: November 29, 2006. Accepted by Garry Twite and David Gallagher, Area Editor.)

References

ABS, 'Industry Concentration Statistics', cat. No. 8140.0.55.001.

Altman, E.I., 1968, 'Financial ratios, discriminant analysis and prediction of corporate bankruptcy', Journal of Finance, vol. 23, pp. 589-609.

Blume, M.E., Lim, F. & Mackinlay, C. 1998, 'The declining credit quality of U.S. corporate debt: Myth or reality?', The Journal of Finance, vol. 53, pp. 1389-413.

Demsetz, H. 1973, 'Industry structure, market rivalry, and public policy', The Journal of Law and Economics, vol. 16, pp. 1-9.

Graham, J.R. & Harvey, C. 2001, 'The theory and practice of corporate finance: Evidence from the field', Journal of Financial Economics, vol. 60, pp. 187-243.

Hausman, J., Lo, A.W. & MacKinlay, A.C. 1992, 'An ordered probit analysis of transaction stock prices', Journal of Financial Economics, vol. 31, pp. 319-79.

Iskander, M.E. & Emery, D.R. 1994, 'An empirical investigation of the role of indenture provisions in determining bond ratings', Journal of Banking and Finance, vol. 18, vol. 93-111.

Kaplan, R. & G. Urwitz, 1979, 'Statistical models on bond ratings: A methodological inquiry', Journal of Business, vol. 52, pp. 231-61.

Matolcsy, Z.P. & Lianto, T. 1995, 'The incremental information content of bond rating revisions: The Australian evidence', Journal of Banking and Finance, vol. 19, pp. 891-902.

Pinches, G.E. & Mingo, K.A. 1973, 'A multivariate analysis of industrial bond ratings' Journal of Finance, vol. 28, pp. 1-18.

Pogue, T.F. & Soldofsky, R.M. 1969, 'What's in a bond rating?,' Journal of Financial and Quantitative Analysis, vol. 4, pp. 201-28.

Porter, M.E. 1979, 'The structure within industries and companies' performance', The Review of Economics and Statistics, vol. 61, pp. 214-27.

Standard and Poor's, 2003, 'Corporate Ratings Criteria,' Standard and Poor's Corporation, New York, www2.standardandpoors.com/spf/pdf/fixedincome/CorpCrit2003r-jun.pdf.

(1.) The sample period of the financial data is greater to allow for three-year averaging.

(2.) Under S&P's GICS structure, there are 10 broad sectors which are sub-divided into 24 industry groups. These groups are further partitioned into 62 industries that are then divided into 132 sub-industries.

(3.) The specific accounting ratios that we examine are defined as follows: EBIT (EBITDA) interest coverage is the ratio of EBIT (EBITDA) to Total Interest Expense. Leverage ratios are given by Long-term (Total) debt to Capitalisation (long-term debt + shareholder's equity + minority interests). Profitability is captured by return on capital (EBIT to [Average debt + equity capital]) and Operating income to Sales (EBITDA to Sales). Debt coverage is measured by two ratios--Cash flow from Operations to Total debt and Funds from operations [EBITDA--interest expense--income tax expense] to total debt.

(4.) In cases where three years of data was not available (i.e. newly listed companies), only one year of data is used in the firm's first year in sample. In the following year, a two-year average is taken and three-year averages are used thereafter.

(5.) Recall that our two profitability measures are not highly correlated. For this reason, we examined a model that included both, but the performance of that model was insignificantly different from one that omitted return on capital. This is likely to be due to the correlation between return on capital and other explanatory variables.

(5.) Note that we ignore observations that the model has already assigned to the top rating category as these cannot be further upgraded.

Stephen Gray [dagger]

Alexsander Mirkovic [dagger]

Vanitha Ragunathan [dagger]

[dagger] UQ Business School, University of Queensland, Brisbane, QLD 4072. Email: [e-mail]

Table 1
Distribution of Sample Observations Over Time, Rating Class and
Industry

This table describes the sample of Standard and Poor's credit ratings
of Australian firms used in this study. The time period covered is
1995 to 2002. The data are sourced directly from S&P.

                                  AAA         AA        A      BBB

Panel A. Corporate Credit Ratings by Year

2002                                           1        23      27
2001                                           2        21      26
2000                                           4        17      28
1999                                           5        20      23
1998                                           5        19      23
1997                                           5        17      19
1996                               1           6        18      15
1995                                           5        17      14

Total                              1          33       152     175

Panel B. Corporate Credit Ratings by Industry

Mining                                         7        25      34
Manufacturing                                  4        44      86
Electricity, Gas and Water                              13       9
Supply
Construction                                             4       9
Wholesale Trade                                2         6       1
Retail Trade                                            14       2
Transport and Storage                                    5       9
Communication Services             1           9         8       1
Property and Business                         11        28       6
Services
Private Community Services                                       2
Personal and Other Services                              5      16
Total                              1          33       152     175

                               Investment    Below
                                 Grade        BBB     Total

Panel A. Corporate Credit Ratings by Year

2002                               51          7        58
2001                               49          7        56
2000                               49          7        56
1999                               48          5        53
1998                               47          3        50
1997                               41          1        42
1996                               40          0        40
1995                               36          1        37

Total                             361         31       392

Panel B. Corporate Credit Ratings by Industry

Mining                             66          8        74
Manufacturing                     134         16       150
Electricity, Gas and Water         22         --        22
Supply
Construction                       13         --        13
Wholesale Trade                     9          1        10
Retail Trade                       16         --        16
Transport and Storage              14         --        14
Communication Services             19         --        19
Property and Business              45         --        45
Services
Private Community Services          2         --         2
Personal and Other Services        21          6        27
Total                             361         31       392

Table 2
Descriptive Statistics of Transformed Financial Ratios and Industry
Variables

This table presents summary statistics for three transformed
financial ratios * (interest coverage, debt coverage and
leverage), profitability ratios and industry variables for
all firms in our sample, according to rating categories. The
sample period used is 1995 through to 2002 using a panel of
361 observations. Ratings data are sourced directly from S&P
and financial data from Bloomberg News Service.

                                                           Standard
                                       Mean      Median    Deviation

AAA/AA
  EBIT interest coverage               8.39       7.42        5.17
  EBITDA interest coverage            10.44      10.49        5.84
  Operating funds/Total debt           0.71       0.54        0.95
  Operating cash flows/Total debt      0.67       0.53        0.91
  Return on Capital                    0.17       0.13        0.09
  Operating Margin                     0.37       0.33        0.19
  LT debt leverage                     0.21       0.18        0.16
  Total debt leverage                  0.42       0.42        0.33
  Industry beta                        0.86       0.80        0.25
  Industry concentration (%)          42.08      32.91       32.74

A
  EBIT interest coverage               5.30       4.32        3.35
  EBITDA interest coverage             7.50       6.44        4.31
  Operating funds/Total debt           0.40       0.31        0.42
  Operating cash flows/Total debt      0.45       0.34        0.52
  Return on Capital                    0.12       0.10        0.06
  Operating Margin                     0.28       0.18        0.22
  LT debt leverage                     0.26       0.30        0.16
  Total debt leverage                  0.40       0.38        0.19
  Industry beta                        0.95       1.01        0.32
  Industry concentration (%)          34.98      34.30       19.83

BBB
  EBIT interest coverage               4.60       3.32        4.36
  EBITDA interest coverage             7.09       5.28        5.84
  Operating funds/Total debt           0.50       0.26        0.75
  Operating cash flows/Total debt      0.54       0.28        0.77
  Return on Capital                    0.10       0.10        0.06
  Operating Margin                     0.21       0.14        0.18
  LT debt leverage                     0.34       0.36        0.16
  Total debt leverage                  0.41       0.40        0.17
  Industry beta                        0.96       0.95        0.31
  Industry concentration (%)          33.29      30.60       15.77

                                       Max.       Min.

AAA/AA
  EBIT interest coverage              25.00       1.93
  EBITDA interest coverage            30.00       1.99
  Operating funds/Total debt           5.00       0.06
  Operating cash flows/Total debt      4.50      -0.52
  Return on Capital                    0.37       0.07
  Operating Margin                     0.69       0.12
  LT debt leverage                     0.68       0.00
  Total debt leverage                  1.52       0.03
  Industry beta                        1.39       0.36
  Industry concentration (%)          87.70       9.00

A
  EBIT interest coverage              25.00       0.38
  EBITDA interest coverage            30.00       2.31
  Operating funds/Total debt           4.22       0.07
  Operating cash flows/Total debt      5.00      -0.25
  Return on Capital                    0.34      -0.003
  Operating Margin                     0.80       0.04
  LT debt leverage                     0.72       0.00
  Total debt leverage                  1.64       0.06
  Industry beta                        1.60       0.12
  Industry concentration (%)          85.07       8.60

BBB
  EBIT interest coverage              22.61      -0.15
  EBITDA interest coverage            30.00       0.32
  Operating funds/Total debt           4.26      -0.29
  Operating cash flows/Total debt      4.61       0.03
  Return on Capital                    0.48      -0.01
  Operating Margin                     0.80       0.04
  LT debt leverage                     0.95       0.00
  Total debt leverage                  0.98       0.08
  Industry beta                        1.60       0.13
  Industry concentration (%)          87.70       7.50

                                     Skewness   Kurtosis

AAA/AA
  EBIT interest coverage               1.76       6.66
  EBITDA interest coverage             1.91       7.85
  Operating funds/Total debt           3.48      14.67
  Operating cash flows/Total debt      3.03      12.44
  Return on Capital                    0.76       2.15
  Operating Margin                     0.35       1.68
  LT debt leverage                     0.46       2.27
  Total debt leverage                  1.88       6.87
  Industry beta                        0.70       2.89
  Industry concentration (%)           0.41       1.43

A
  EBIT interest coverage               2.45      11.55
  EBITDA interest coverage             1.95       8.16
  Operating funds/Total debt           5.69      47.12
  Operating cash flows/Total debt      5.30      41.06
  Return on Capital                    1.70       6.18
  Operating Margin                     1.10       2.80
  LT debt leverage                     0.12       3.19
  Total debt leverage                  2.56      14.06
  Industry beta                       -0.48       2.26
  Industry concentration (%)           0.70       2.90

BBB
  EBIT interest coverage               2.15       7.27
  EBITDA interest coverage             2.14       7.55
  Operating funds/Total debt           2.99      11.81
  Operating cash flows/Total debt      2.94      11.73
  Return on Capital                    2.09      11.43
  Operating Margin                     2.14       6.46
  LT debt leverage                     0.79       6.50
  Total debt leverage                  1.05       5.10
  Industry beta                       -0.31       2.36
  Industry concentration (%)           0.52       2.81

Note: * Extreme values of interest coverage and debt coverage ratios
are truncated. EBIT (EBITDA) interest coverage ratios are capped at
25 (30). Interest coverage ratios less than -1 are set to--1. Debt
coverage ratios are capped at a maximum of 5 and a minimum of -1.
Negative leverage ratios are set to zero.

Table 3
Univariate Ordered Probit Model Estimates of Independent
Financial Variables

The ordered probit model is estimated separately for each of the
eight financial variables and two industry variables. The model
is estimated for the whole sample of 361 observations over the
sample period of 1995 to 2002. Ratings data are sourced directly
from S&P and financial data is from Bloomberg News Service.
P-values for the percentage of correct predictions are based on
binomial distribution where p is 0.4848, consistent with the
success rate of a naive predition that all firms are rated BBB.
This nave prediction model would yield 175 successes in our
sample.

                                   [[alpha].sub.1]    [[alpha].sub.2]

EBIT interest coverage                 -1.6863            -0.2829

EBITDA interest coverage               -1.5731            -0.1982

Operating funds/Total debt             -1.3333             0.0213

Operating cash flows/Total debt        -1.3158             0.0378

Return on capital                      -1.837             -0.4244

Operating margin                       -1.7213            -0.306

LT debt leverage                       -0.7551             0.6849

Total debt leverage                    -1.3211             0.0326

Industry beta                          -1.0753             0.2839

Industry concentration                 -1.5402            -0.1746

                                          B            Coefficient x
                                    (Std. Errors)        Std. Dev.

EBIT interest coverage               -0.062               -0.260
                                     (0.018) (a)

EBITDA interest coverage             -0.032               -0.168
                                     (0.013) (a)

Operating funds/Total debt           -0.036               -0.024
                                     (0.113)

Operating cash flows/Total debt      -0.001                0.000
                                     (0.102)

Return on capital                    -4.043               -0.276
                                     (1.059) (a)

Operating margin                     -1.393               -0.284
                                     (0.295) (a)

LT debt leverage                      2.187                0.368
                                     (0.407) (a)

Total debt leverage                  -0.014               -0.003
                                     (0.347)

Industry beta                         0.259                0.080
                                     (0.184)

Industry concentration               -0.616               -0.1216
                                     (0.349) (a)

                                      % Correct
                                     Predictions
                                      (p-value)

EBIT interest coverage                55.4
                                      (0.0049)

EBITDA interest coverage              52.1
                                      (0.0776)

Operating funds/Total debt            42.1
                                      (0.9934)

Operating cash flows/Total debt       42.1
                                      (0.9934)

Return on capital                     47.1
                                      (0.6820)

Operating margin                      54.3
                                      (0.0118)

LT debt leverage                      56.2
                                      (0.0019)

Total debt leverage                   42.1
                                      (0.9934)

Industry beta                         42.9
                                      (0.9487)

Industry concentration                41.8
                                      (0.9952)

Note: (a,b,c) indicates statistical significance at the 1%, 5% and
10% levels respectively.

Table 4
Correlation Matrix of Transformed Variables and Industry
Characteristics

This table presents the correlation between pairs of variables used
in our ordered probit models. The sample period used is 1995 through
to 2002 using a panel of 361 observations. Financial data is from
Bloomberg News Service, industry betas are from CRIF and concentration
ratios are from ABS.

                        EBIT          EBITDA       Operating
                      Interest       Interest        Funds/
                      Coverage       Coverage      Total Debt

EBIT interest
coverage

EBITDA interest       0.950 (a)
coverage

Operating             0.836 (a)      0.870 (a)
funds/Total debt

Operating cash        0.806 (a)      0.857 (a)      0.977 (a)
flows/Total debt

Return on             0.658 (a)      0.660 (a)      0.595 (a)
Capital

Operating             0.131 (a)     -0.007         -0.096 (c)
Margin

LT debt leverage     -0.498 (a)     -0.434 (a)     -0.399 (a)

Total debt           -0.472 (a)     -0.472 (a)     -0.432 (a)
leverage

Industry beta        -0.081         -0.016          0.016

Industry             -0.094 (c)      0.041          0.041
concentration
(%)

                     Operating
                        CF/         Return on      Operating
                     Total Debt      Capital         Margin

EBIT interest
coverage

EBITDA interest
coverage

Operating
funds/Total debt

Operating cash
flows/Total debt

Return on             0.575 (a)
Capital

Operating            -0.101 (c)      0.055
Margin

LT debt leverage     -0.384 (a)     -0.042         -0.147 (a)

Total debt           -0.455 (a)     -0.040         -0.035
leverage

Industry beta         0.028         -0.043         -0.207 (a)

Industry             -0.037          0.296 (a)     -0.007
concentration
(%)

                      LT Debt       Total Debt      Industry
                      Leverage       Leverage         Beta

EBIT interest
coverage

EBITDA interest
coverage

Operating
funds/Total debt

Operating cash
flows/Total debt

Return on
Capital

Operating
Margin

LT debt leverage

Total debt            0.576 (a)
leverage

Industry beta         0.044         -0.103 (b)

Industry              0.355 (a)      0.181 (a)      0.163 (a)
concentration
(%)

Note: (a,b,c) indicates statistical significance at the 1%, 5% and
10% levels respectively.

Table 5
The Full Ordered Probit Model

This table presents results for the ordered probit model that
incorporates four financial variables, two industry variables and
a time trend. The model is estimated for the whole sample of 361
observations over the sample period of 1995 to 2002. Ratings data
are sourced directly from S&P and financial data is from Bloomberg
News Service.

Panel A presents parameter estimates from the full ordered probit
model. The overall success rate of the model is 61.50%. The estimates
of the sample partition parameters are [[alpha].sub.1] = -1.0452 and
[[alpha].sub.2] = 0.6770.

Panel B reports the proportion of model predictions that would change
if the financial variable were increased by one standard deviation
while all other data is unchanged. A total of 357 (187) observations
can potentially be upgraded (downgraded). Observations already in the
top (bottom) category cannot be further upgraded (downgraded).

Panel A: Parameter Estimates, for the Full Ordered Probit Model

                                                Standard
                               Coefficient       Error        p-value

Year dummy                        0.132          0.031        0.0000
EBIT interest coverage           -0.150          0.038        0.0001
Operating funds/Total debt        1.016          0.309        0.0010
Operating margin                 -0.759          0.406        0.0616
LT debt leverage                  2.866          0.614        0.0000
Industry beta                     0.418          0.207        0.0435
Industry concentration           -0.018          0.004        0.0000

Panel B: The Impact of Financial Variables on Credit Ratings

                                   Improvement of One
                                   Standard Deviation

                                Upgraded       Downgraded

EBIT interest coverage          331/357           0
                                (92.72%)

Operating finds/Total debt       51/357           0
                                (14.29%)

Operating margin                  5/357           0
                                 (1.4%)

LT debt leverage                  0              28/357
                                                (14.97%)

                                   Improvement of One
                                   Standard Deviation

                                Upgraded       Downgraded

EBIT interest coverage            0             186/187
                                                (99.47%)

Operating finds/Total debt        0              58/187
                                                (31.02%)

Operating margin                  0               9/187
                                                 (4.81%)

LT debt leverage                 23/187           0
                                 (6.44%)

The Full Ordered Probit Model

Panel C compares the predictions of the full ordered probit model
with actual S&P ratings.

Panel D presents the predictive accuracy of our final model (without
the time trend variable). In each case, the model is estimated with
data from all other years and the prediction accuracy is reported for
the holdout year. P-values for the percentage of correct predictions
are based on binomial distribution where p is 0.4848, consistent with
the success rate of a naive predition that all firms are rated BBB.

Panel C: Prediction Success Matrix

                                 Actual
                   AAA,AA           A         BBB

Forecast
  AAA,AA              2              2          0          4
  A                  28            100         55        183
  BBB                 4             50        120        174
                     34            152        175

Panel D: Predictive Ability of the Model on a Holdout Sample

                 Percentage
                 of Correct
Holdout Year     Predictions     p-value

1995                69.44        0.0089
1996                67.50        0.0048
1997                56.10        0.1286
1998                51.06        0.4168
1999                54.16        0.2596
2000                57.14        0.1421
2001                71.43        0.0003
2002                66.67        0.0029