Bankruptcy Prediction Models and The Cost of Debt: Sattar A. Mansi, William F. Maxwell, and Andrew Zhang
Bankruptcy Prediction Models and The Cost of Debt: Sattar A. Mansi, William F. Maxwell, and Andrew Zhang
Bankruptcy Prediction Models and The Cost of Debt: Sattar A. Mansi, William F. Maxwell, and Andrew Zhang
Mansi is an Associate Professor of Finance at Virginia Tech, Maxwell is a Chaired Professor of Finance at Southern Methodist University, and Andrew (Jianzhong) Zhang is an Assistant Professor of Finance at the University of Nevada, Las Vegas. All remaining errors are ours.
1
Electronic copy available at: http://ssrn.com/abstract=1622407
1. Introduction The goal of financial distress models is to utilize information about the firm to construct a single variable allowing for the comparison of default risk across firms. The literature on financial distress risk provides four main models that forecast default including two accounting-based models (Altman, 1968; Ohlsons, 1980), one reduced form model (Campbell, Hilscher, and Szilagyi, 2010) and one structural model (Merton, 1974). These models lead to four proxy variables, namely, Altmans Z-Score, Ohlsons O-Score, CHS Score, and Mertons Distance to Default (Merton-DD). These measures are widely used in the literature to proxy for distress risk, especially in the areas of accounting and financial economics. According to Social Sciences Citation Index, as of December 2009, Altman (1968), Ohlson (1980), Merton (1974), and Campbell et al. (2010) have been cited 691, 295, 709, and 9 times respectively 1 Based solely on citations, the z-score and o-score remain highly popular distress risk measure in the academic literature, as they were 132 and 59 times in 2009 alone. Our results suggest that this is highly problematic. We find evidence suggesting that the Altman and Ohlson model are very weakly related to distress risk and in fact do no better than a naive distress measure, the leverage ratio. Given that empirical results differs based on the default measure utilized, the relative performance of these models in forecasting financial failures is of significant concern in the literature. 2 While the results are mixed, extant studies have never researched these four distress measures all together in a unified and comprehensive framework. Also ignored is the relative performance of these measures across different economic periods. Lastly, the relation between default measures and group of firms characterized by some important firm-specific variables like size, book-to-market ratio, and volatility has not been analyzed.
1
These variables
We conduct a short and likely incomplete survey by only considering papers recently published in top accounting and financing journals as of the year 2009. 2 For example, Agarwal and Taffler (2008) find the outperformance of Z-Score over Merton-DD. Hillegeist, Keating, Cram, and Lundstedt (2004) find that Merton-DD provides more information than either Z- or O-Score. Kealhofer and Kurbat (2001) argue that Merton-DD captures all of the information in bond ratings and well-known accounting variables. Beaver McNichols, and Rhie (2005) use a logit model with a longer time period finds that the ability of accounting ratios to predict bankruptcy remains. Duffie, Saita, and Wang (2007) show that Merton-DD has significant predictive power in a model of default probabilities over time. Shumway (2001) argues the advantage of a hazard model over the econometrical models underlying Altmans Z-Score and Ohlson O-Scores and the better performance of market-based variables over accounting-based variables. Bharath and Shumway (2008) find that one can easily construct a reduced form model that outperforms Merton-DD in out-of-sample forecasting ability, but hazard models that use the Merton-DD probability with other covariates have slightly better performance than models that omit the Merton-DD probability. Campbell et al (2010) find that adding Merton-DD to their reduced form model does not increase adjusted R-squared of their logit regression, and conclude that CHS model outperforms Merton-DD in bankruptcy forecasting.
2
Electronic copy available at: http://ssrn.com/abstract=1622407
significantly reflect the changes in corporate capital structure and the riskiness of corporate activity. According to Campbell et al. (2010), these changes can partially explain the changes in financial failures through time. As a result, to locate a distress measure that has robust superior forecasting power of bankruptcy is crucial to many empirical studies. Our results find a very significant difference across these models with accounting based measure faring poorly. In this paper, we attempt to reconcile the differences in extant studies by examining the relation between financial distress measures and bond spreads, an assessment of the markets' perception of risk, to determine the relative performance of these models. Given that bond investors require stable cash flows and demand default risk premium in yield spreads, many studies have related yield spreads to distress risk from various perspectives (see e.g., Almeida and Philippon, 2007; Anginer and Yildizhan, 2010; Bharath and Shumway, 2008; Davydendo and Strebulaev, 2007). We posit that distress risk is efficiently impounded in bond market prices; therefore, yield spreads are a significant reflection of the probability of financial failure. Hence, we are not modeling actual default risk as in prior research, but instead analyzing the ability of default models to capture market based distress risk. Our results indicate that the CHS Score is a superior forecasting model followed by Merton-DD, O-Score, and Z-Score. This is true in both early and late periods, and across normal and shock economic periods. The outperformance of CHS is more evident for firms with small size, low M/B ratio, and high equity volatility. To assess the relative performance of different models in forecasting financial failure, prior studies use prediction-oriented tests to distinguish between alternative models (see for example, Shumway, 2001; Bharath and Shumway, 2008). These tests usually rank sample firms by each measure into deciles (or quintiles) and assess prediction accuracy within a particular (typically one-year) time-horizon for each decile, either in-the-sample or out-of-sample. The model with the highest overall prediction accuracy is deemed the best. Hillegeist et al. (2004) argues that whether or not a single firm declares bankruptcy ex post does not provide evidence about whether the ex ante estimate is an accurate reflection of the true probability. They propose relative information content tests to compare the amount of distress-related information contained in each measure. Statistically, however, either test may have a relatively low test of power because the size of distressed sample is usually very small. For example, financial
failures were extremely rare until the late 1960s, and in the 3 years (1967 to 1969), there were no bankruptcies at all (Campbell et al., 2010). These studies also suffer from problems in defining financial failure and constructing distressed sample. Specifically, these studies conduct tests based on a sample of distressed firms, where a firm enters into the sample upon occurrence of one or more of the following events: default in payment of interest or principal, announcement of bankruptcy, or being delisted from exchanges for financial problems. Some important distress signals like dividend cut or omission have never been counted, however. These studies are usually different in their definition of financial distress and hence the sample construction of distressed firms. As a result, testing results are usually sample-specific and may be sensitive to sample size and sample period used in their model estimations. Even more importantly, firms often become financially distressed either much earlier or later than the actual default or bankruptcy filing date, and that lag or lead time can be significant. In practice, some financially distressed firms can manage its operating earnings, working capital, and cash flows to meet debt covenants and interest payment to delay default or bankruptcy. On the other hand, some firms will file for bankruptcy even when they are economically solvent. Strategic default brings firms into default early to break unfavorable contracts and obtain forced debt concessions. Additionally, a firm may have market assets greater than total liability but does not have enough cash and liquid assets to meet its current obligations and is forced to bankruptcy. A comparison between alternative bankruptcy forecasting models might be misleading when one measure of distress is timelier, and thus capture more early distress information than the others. Further, we know that many firms restructure their debt before entering bankruptcy. These firms would not show up in the sample of defaults, and even more problematic is that the characteristics related to firms choosing to restructure inside or outside of bankruptcy, such as firm size and M/B ratios, usually are also predictive variables in the distress measures (see Hotchkiss, John, Mooradian, and Thorburn, 2008). Additionally, previous studies when examining relative performance of distress models typically presume a dichotomous decision context; however, a decision-maker will more typically make a continuous decision choice (Hillegeist et al., 2004). For example, equity and debt holders have to assess the distress risk on a continuous basis to determine the required cost 4
of capital. In this study, we focus on the corporate bond spread since it is extracted based on publicly traded debt and represents not only a sensitive measure of distress risk, but also a timely and continuous one. Consequently, corporate bond spreads will provide a consistent and comprehensive benchmark for the assessment of the relative performance of different distress measures without the need of the sample construction of distressed firms. Using a large sample of bond data from the Lehman Brothers Fixed Income database covering the period from 1980 through 2006, we find that the four measures of financial distress are significantly related to bond spreads with CHS Score obtaining the highest t-statistic and adjusted R-squared followed by Merton-DD, O-Score and Z--Scores, before and after controlling for years to maturity and industry effect. Since it is well known that corporate spread is too high to be explained only by expected default, we use offering amount, bond age, and coupon rate to control for liquidity and tax effect, and find the relative outperformance of CHS remains intact. Extant studies suggest that the underperformance of accounting-based Z- and O-Scores might be due to the staleness of parameters and failure to accounting for R&D effects in their predictive variables. We use updated coefficients from Hillegeist et al. (2004) to construct updated Z- and O-Scores. We also follow Franzen, Rodgers, and Simin (2007) and re-construct these two scores by accounting for R&D effects. However, these two scores continue to have poor performance. The remainder of the paper is organized as follows. Section 2 describes data, distress measures, and control variables. Section 3 explains methodology and discusses the empirical results. Section 4 discusses and concludes.
nonconvertible bonds that are used in the Lehman Brothers Bond Indexes. The database is commonly used in the finance literature (e.g., Mansi, Maxwell, and Wald, 2009; Bharath and Shumway, 2008; Klock, Mansi, and Maxwell, 2005). For a firm-year observation to be included in the analysis, firms must have a debt issue that is present in the LBFI dataset and financial information must also be available in the Compustat and CRSP databases to compute all four distress measures (described in detail below). For firms with multiple issues, we only include the bond with the greatest term to maturity since it is more representative of the cost of debt. 3 We follow the convention in the literature to exclude regulated utilities with SIC codes 4949 to 4999 and financial firms with SIC codes 6000 to 6999, since some financial ratios used to construct distress measures are not applicable and have different meanings in these industries. Similar to Bharath and Shumway (2008) and Campbell and Taksler (2003) we also exclude all AAA rated bonds because the data for these bonds may not be reliable. We further eliminate all corporate floating rate debt, bonds with an odd frequency of coupon payments, inflation-indexed bonds, and bonds with less than one year to maturity due to their illiquidity. Merging the databases and applying these requirements yields a data set of 120,608 monthly observations on 1,752 firms covering the period from 1980 through 2006.
2.2 Measuring Bankruptcy Prediction Models We use four basic models along with derivations of the accounting models to measure distress risk. In addition, we also include in the analysis a naive benchmark, the market leverage ratio of the firm, defined as the sum of current long-term debt (annual item 34) and long-term debt (annual item 9) divided by market firm value (sum of current long-term debt (annual item 34), long-term debt (annual item 9), and market equity (product of annual item 199 and item 25). This serves as baseline to compare the effectiveness of the models. The first two models, Altman (1968) Z-Score and Ohlson (1980) O-Score, are based on accounting data. Since some constituents of O- and Z- scores are not available in the Compustat quarterly tape, we follow the convention in the literature and use Compustat annual tape to construct these two accounting-based measures and update them annually. We lag four months
3
Though not reported, we obtain similar results using the average spread across multiple issues for these firms.
to allow accounting information from annual filing to be available to bond investors when constructing these two measures. The primary specification for the z-score model is Z-Score=-1.2*wcta-1.4*reta-3.3*ebitta-0.60*mvliab-sata (1)
where wcta is working capital, or current assets (annual item 4) minus current liability (annual item 5) scaled by total assets (annual item 6), reta is retained earnings (annual item 36) scaled by total assets (annual item 6), ebitta is earnings before interest and tax (annual item 170 plus item 15) divided by total assets (annual item 6), mvliab is market value of equity (annual item 199 times item 25) divided by total liability (annual item 181), and sata is sales (annual item 12) divided by total assets (annual item 6). For comparison with other models, we reverse the signs of original coefficients so the Z-Score is increasing in the probability of bankruptcy. And the primary specification for the O score model is O-Score = -1.32 - 0.407*asset + 6.03*tlta - 1.43*wcta + 0.0757*clca - 2.37*nita - 1.83*ffotl + 0.285*intwo - 1.72*oeneg - 0.521*chin
(2)
where asset is the log of total assets (annual item 6), tlta is total liabilities (annual item 181) divided by total assets (annual item 6), wcta is working capital to total assets (as defined in equation (1) above), clca is current liability (annual item 5) divided by current assets (annual item 4), nita is net income (annual item 172) divided by total assets (annual item 6), ffotl is fund from operations defined as pretax income (annual item 170) plus depreciation (annual item 14) divided by total liability (annual item 181), intwo is a dummy variable that equals to 1 if the firm has negative book equity, chin is change in net income (annual item 172) from prior year divided by the sum of absolute values of current and prior year net income. Franzen, Rodgers, and Simin (2007) argue that since reporting of research and development (R&D) spending has increased over time, the effectiveness of these two accounting-based variables has deteriorated to the detriment of inferences related to distress risk. Therefore, we modify the definition of constituent variables to control for the effect of accounting for R&D. Adjusted net income is the sum of net income (annual item 172) plus R&D expense (annual item 46, 0 if missing) minus R&D amortization defined as 0.2 times the sum of the previous 5 years of R&D expense. If net income (annual item 172) is positive, we further control for the tax effect
by multiplying the adjusted net income by (1 - tax rate), where the tax rate is defined as the appropriate annual statutory tax rate: 46% in the period 1980-1986, 40% in the year 1987, 34% in 1988-1992, and 35% in 1993-2006. The variable adjust total assets is total assets plus R&D capital defined as RD+0.8*RDt-1+0.6*RDt-2+0.4*RDt-3+0.2*RDt-4. If net income (annual item 172) is positive, we adjust total liability as total liability (annual item 181) plus deferred tax liability defined as R&D capital times the tax rate. After adjusting net income, total assets, and total liability, we re-construct all constituent variables in Z- and O-Scores respectively and apply the original coefficients to these variables to construct a new set of Z- and O-Scores, labeled hereafter as adjusted Z-Score and Adjusted O-Score. Several studies have argued that the original coefficients in each model have substantially changed from their original values (e.g., Begley, Ming, and Watts, 1996; Hillegeist et al. 2004). Therefore, we use the coefficients based on the re-estimation by Hillegeist et al. and construct an updated Z-Score and O-Score measures. That is Updated Z-Score=-4.34-0.08 wcta+0.04*reta-0.1*ebitta-0.22mvliab+0.06sata Updated O-Score = -5.91 + 0.04*asset + 0.08*tlta + 0.01*wcta-0.01*clca + 1.20*nita + 0.18*ffotl + 0.01*intwo + 1.59*oeneg - 1.10*chin (4) (3)
where all constituent variables are defined in the same manner as in the original model without adjustment for the R&D effect. The third measure of distress risk is based on the structural model of Merton (1974). Recently, Hillegeist et al. (2004), Bharath and Shumway (2008), and Campbell et al. (2010) estimates the Merton (1974) model to obtain the distance to default (DD), a measure of the difference between the asset value of the firm (TA) and the face value of its debt (BD), scaled by the standard deviation of the firms asset value (SIGMA). We use an iterative procedure by solving a system of two nonlinear equations simultaneously to estimate asset value and asset volatility. In this system, equity is valued as a European call option on the value of the firms assets with time to maturity, T, equal to 1 year. We lag two months to allow the most recent accounting information from quarterly filings to be available to bond investors when we construct the DD measure. Specifically, we use short-term (quarterly item 45) plus one half long term (quarterly item 51) book debt to proxy for the face value of debt, or BD. This convention is
a simply way to take account of the fact that long-term debt may not mature until after the horizon of the distance to default calculation. If book debt (BD) is missing, we use BD=median (BD/TL)*TL, where the median is calculated for the entire dataset and TL is the total liabilities (quarterly item 54). This captures the fact that empirically, BD tends to be much smaller than TL. If BD=0, we use BD= median (BD/TL)*TL, where now we calculate the median only for small but nonzero values of BD (0 < BD <0.01). We use a 12-month rolling window (with at least 50 observations) standard deviation of daily stock return to measure equity volatility and use it as the starting value of asset volatility. Before calculating asset value and volatility, we adjust BD so that BD/ (ME+BD) is winsorized at the 0.5 and 99.5 percentiles of the crosssectional distribution, where ME is market equity. We compute Mertons distance to default, DD, as
DD =
1 SIGMA 2 2
(5)
where TA is total assets, Rbill is Treasure bill rate, SIGMA is asset volatility. We follow Campbell et al. (2010) and use 0.06 as an empirical proxy for the equity premium. Note that we put the negative sign before the traditional measure of distance to default so our DD is actually increasing in the probability of bankruptcy. Finally, we construct the CHS Score based on Campbell et al. (2010) reduced form model. We use Compustat quarterly tape and CRSP monthly and daily tapes to compute the fitted values using the coefficients in the 3rd column in Table 4 of Campbell et al. (2010). As before, we align each companys fiscal year appropriately with the calendar year, converting Compustat fiscal year quarterly data to a calendar basis, and then lag accounting data by 2 months. This adjustment ensures that the accounting data are available at the beginning of the month over which bankruptcy is measured. Our primary specification is CHS Score = -9.16 - 20.26*nimtaavg + 1.42*tlmta - 7.13*exretavg + 1.41*stdev - 0.045*rsize - 2.13*cashmta + 0.075*mtb - 0.058*price (6)
where nimtaavg and exretavg are the moving average of lagged four quarterly nimta and 12 monthly exret, respectively, with geometrically declining weights on lags. The variable nimta is computed as net income (quarterly item 69) divided by the sum of market equity (the product of number of shares outstanding and month end stock prices, both from CRSP) and total liability (quarterly item 54). The moving average nimtaavg suggests that a long history of loss is a better predictor of bankruptcy than one large quarterly loss in a single month. Exret is the monthly log excess return on each firms equity relative to the S&P 500 index. The moving average exretavg suggests that a sustained decline in stock market value is a better predictor of bankruptcy than a sudden drop in stock price in a single month. When lagged exert or nimta is missing, we replace it with the cross-sectional mean in order to avoid losing observations. Tlmta is the ratio of total liabilities (quarterly item 54) divided by the sum of market equity and total liabilities. Stdev is the annualized three-month rolling sample standard deviation. To eliminate cases in which few observations are available, we code stdev as missing if there are fewer than five nonzero observations over the three months used in the rolling window computation. When computing CHS Score, we replace missing stdev observations with the cross-sectional mean of stdev. Rsize is the relative size of each firm measured as the log ratio of its market equity to that of the S&P 500 index. Cashmta, a proxy for liquidity position of the firm, is the ratio of cash and short term investments (quarterly item 36) divided by the sum of market equity and total liabilities. Mtb is the ratio of market-to-book equity, where book equity is the sum of stockholders equity (quarterly item 60) and deferred tax credit (quarterly item 52) minus preferred stockholders equity (quarterly item 55). We adjust book equity by adding 10% of the difference between market and book equity. Price is the log price per share, truncated above at the $15. We winsorize all eight predictive variables at the 5th and 95th percentiles of their pooled distributions. In calculating summary statistics and conducting regression analysis, we winsorize all distress measures at the 1st and 99th percentiles to further limit the influence of outliers. As discussed before, all Z- and O-Scores are updated annually, while DD and CHS Scores are updated monthly. We expect the difference in updating frequency will affect the relative performance of these measures in forecasting bankruptcy and predicting bond yield spreads and conduct the robustness check based on annual data in Sections 3.2 and 3.3 below.
10
2.3 Measuring the Cost of Debt Financing We use the month-end bid price from the LBFI database to compute the cost of debt for each corporate bond. The cost of debt, or yield spread, is the difference between the yield to maturity on a corporate bond and the yield to maturity on its maturity equivalent Treasury security. The yield to maturity on a corporate bond is the discount rate that equates the present value of its future cash flows to its current price. The yield to maturity on a Treasury security is the yield on the constant maturity series obtained from the Federal Reserve Bank in its H15 release based on a par bond. In the cases where no corresponding Treasury yield is available for a given maturity, the yield spread is calculated using linear interpolation. For some of our analysis, we rely on raw yield spread. But due to the skewness of the yield data, we use the log of bond yield spread instead of raw yield spread in our regressions. When utilizing the log of the yield spread, we find a slightly higher r-squared in the regression results but our results are invariant when using raw yield spread.
2.4 Descriptive Statistics Panel A of Table 1 reports summary statistics for sample firms from 1980 to 2006. The mean, median, and standard deviation of the main variable in our analysis, yield spread, is 283, 189, and 288 basis points (bps), with upper and lower quartile values of 341 and 119, respectively. As discussed before, because the yield spread is highly skewed, we use the log transformed yield spread to approximate a normal distribution. Our nave measure and the four measures of financial distress risk market leverage, Z-score, O-Score, Merton-DD, and CHS Score have mean values of 0.399, -2.481, -1.062, -6.057, and -7.702, respectively. Compared to summary statistics reported in other studies, these numbers indicate that majority of our sample firms have low default probability. This is reasonable given that firms with bond issuances are much larger and financially healthier than firms in the CRSP-Compustat universal. [Insert Panel A of Table 1 about here] For the overall sample, the mean (median) market capitalization is $5.0 ($1.7) billion, a standard deviation of $10.7 billion, and 25th and 75th percentile values of $560 million and $4.75 billion, respectively. The median market-to-book ratio is 1.49, close to the mean of 1.87. Firms 11
in the sample have average annualized daily volatility of stock return of 37%. In terms of debt variables, the mean and median bond rating variable equates to S&P ratings of BB and BB-, respectively, with standard deviation ranging from B- to BBB+, which indicates a large portion of the sample has slightly below average quality debt. The mean and median traded debt has maturity of 15 and 13 years and, on average, has been outstanding for about 3.1 years with coupon rate of 9.368%. Panel B of Table 1 describes the industry distribution of the sample (in absolute number and in percentage) using the standard Security Industry Classification (SIC) codes. Industries include agriculture, forestry, fishing, mining, construction, manufacturing (food through petroleum and plastics through electronics), transportation and communications excluding utilities, wholesale and retail trade, real estate, services, and public administration. Based on the sample, a large portion is concentrated in manufacturing (about 47%), transportation and communication (22%), wholesale and retail trade (11%), services (10%), and mining and construction (9%). [Insert Panel B of Table 1 about here] Panel C of Table 1 provides the Pearson correlation between the variable used in the analysis and the four distress measures. We find that yield spread to be positively correlated with market leverage and each of the four measures of financial distress, with the highest correlation with CHS Score. The yield spread is also negatively correlated with bond ratings; evidence consistent with a general expectation that firms with higher distress measures and lower credit ratings have higher costs of debt financing. The pairwise correlation between any two measures of financial distress is also positive, with the highest correlation existing between Merton-DD and CHS Score (0.725). However, none of the correlation is close to a perfect correlation, indicating that these four measures may have captured different aspect of financial distress. Different measures also have different correlations with certain firm- and bond-specific characteristics. For example, while Z-Score and Merton-DD are negatively correlated with MTB, O-Score and CHS Score are positively correlated with MTB. This indicates that distress measures might have different significance for value and growth firms. For that reason, we also test the robustness of our results across firms that are different in size, MTB, volatility, and bond
12
liquidity. Years to maturity is negatively correlated to yield spreads, indicating that bonds with longer years to maturity have higher liquidity and lower yield spreads. Except for Z-Score, years to maturity are negatively correlated with distress measures, with correlation coefficients between -0.24 to -0.27. [Insert Panel C of Table 1 about here]
3. Empirical Analysis
3.1 Distress Measures and Bond Ratings Some studies rely on bond ratings to assess financial distress (see e.g., Avramov, Chordia, Jostova, and Philipov, 2008; Dhaliwal and Reynolds, 1994), and as first attempt at examining the veracity of the distress measures, we examine how the distress measures are related to bond rating. 4 While bond ratings can clearly be erroneous measures of default risk, they are known to be related to future default probability. A possible criticism of the rating analysis is that while the results do not reflect any problems with the distress risk measures, poor performance of some distress measures based on the rating analysis is simply due to problems with the bond ratings. For this reason, we move to an analysis of credit spreads in the next section. However, bond ratings serve as one measures of risk and provide a starting point for our analysis. For credit rating, we use the LBFI database to collect S&P bond credit ratings on a monthly basis and augment this dataset with Compustat data when available. We convert this information into major rating categories (AA, ..,B). The goal in this analysis is to see the relative spreads of the distress measure across bond ratings. To do this, we examine the median distress measure by bond rating. As a first condition, we should see a correct alignment of the default risk measures, a lower distress scores should equate to a higher rating. We report the median distress measure by rating category in the top row of each distress measure in Table 2, which is labeled All Periods. We find that all distress measures are correctly aligned. In this section, we
4 We believe that relying on bond ratings should be avoided given the significant potential for a selection bias as firms with bond ratings are significantly different than the average public firm. Most problematic is the fact that firms with bond ratings have a selection bias on characteristics known to influence asset returns, size, market-tobook, and volatility as firms with bond ratings are larger with lower than average market-to-book ratios and market volatility.
13
only focus on the originally constructed Z- and O-score, but our results are generally similar when using the adjusted and updated measures. One measure of the usefulness of a distress variable is its stability across time and industry since researchers often compare the risk of firms at different points in time and across different industries. Hence, we examine how stable the relation is between the distress measures and bond ratings. We examine whether the distress measures and bond rating correctly align over shorter-time frames, whether they are stable over time, and across industries. We begin by examining the stability of the distress measures over time with the idea that a stable distress measure allows for comparison of firms across time. As such, we compute the distress measures for the sample over three-year horizons and report the median levels within bond ratings. The results are presented in Table 2. Market leverage does not correctly align with bond ratings in two of the nine periods in the BBB thru B categories. The Z-score is highly problematic as the distress measure does not correctly align in six of the nine time periods (the relative rankings across the BBB, BB and B categories). All the other measures, correctly align with rankings in sub-time periods. As it relates to the stability of ranking over time, we find significant variation of rankings for market leverage, Z-score, O-score and Merton-DD. That is, the median value of BB in one time period might in fact be rated a BBB or B in another time period. We do find some variation over time for the CHS Score, but it is clearly the most stable. [Insert Table 2 about here] In Table 3, we next look at the comparability of distress measures across industries. Given our prior analysis, we hold time period constant in this analysis and only look at firms in the sample in June of 2006. Within industries, we find that the distress measures are for the most part correctly aligned except for the Business Services industry with the Z- and O-score, the Wholesale and Retail for the Z-score, and Communications for Market Leverage. Across industries, there is a huge variation as firms would be rated BBB or B with the same distress score depending on industry for the Z-score, O-score and Merton model (the Z-score has the greatest variation). The CHS Score varies across industries but the biggest discrepancy is less than the other distress measures. Overall, the distress measures mapping into bond ratings
14
suggests that it is difficult to have useful comparisons across time and industry. especially true for the Z-Score and less so for the CHS Score. [Insert Table 3 about here]
This is
3.2 Decile Sorts Noting the limitation of assuming that bond ratings are an accurate proxy for default risk, we move to examine the relation between distress measures and credit spreads. Credit spreads are assumed to reflect the markets perception of default risk plus a liquidity component. We begin with the assumption that for a distress measures to be an informative tool, it should differentiate the relative riskiness of firms. We begin our analysis by sorting firms into distress deciles based on the different distress measures. If the distress measure is properly capturing risk, we should see credit spreads increasing across deciles. Panel A of Table 4 provides analysis by examining the mean and median credit spread within distress measure deciles for a fixed time period (as of June 2006). The results suggest that, in general, all models correctly distinguish the firms in the most extreme deciles. Both the Merton DD and CHS distress models correctly align firms within deciles to credit spreads, spreads are increasing with the deciles. In contrast, the Z- and O-score models do not correctly distinguishing between the risks across the middle deciles. The Z and O-Score results are similar to the naive model, market leverage. [Insert Panel A of Table 4 about here] Next, we examine the likelihood that the default score sorts the firms into a similar decile based on measuring credit risk with spreads. That is, we do a double sort. We sort into deciles based on the default measure and also on credit spreads. We then examine what percentages of firms in the default measure decile are also in the same or the adjacent decile based on the credit spread sorting. This sort is performed in June of every year and the results are reported in Panel B of Table 4. We can classify a distress measure as being a more accurate depiction of market risk if there is a higher percentage in each decile. From the analysis, we find that market
15
leverage does a better job in every decile when compared to the Z- and that the O-Score seems to do slightly better than market leverage. Using the updated or R&D adjusted Z- and O-Score don't really change these conclusions. Both the Merton DD and CHS do better on average than market leverage. Across all the measures, the extremes are more easily differentiated than the middle deciles. [Insert Panel B of Table 4 about here] As an overall conclusion to the decile analysis, we find very little difference between the market leverage and the O-Score in performance. If anything the Z-score seems to perform worse than market leverage. The Merton DD and the CHS model seem to clearly outperform the other measures. On aggregate, the R&D adjusted Z- and O-score don't perform very differently than the unadjusted models, and the updated Z and O-score models seem to do a particularly poor job of assessing risk. While this is only one analysis, in the remainder of the paper we find similar results for the R&D adjusted and updated Z and O-score models. The R&D adjusted models in aggregate produce similar results to the original models, and the updated models perform worse than the original models. Hence for brevity, for the remainder of the paper, we focus solely on the original Z- and O-score models.
3.3 Distress Measures & Credit Spreads a Regression Analysis To assess the effect of distress measures on yield spread, we conduct cross-sectional regressions of spread relative to the distress measures. Our primary regression model is Spreadi,t+1 = 0 + 1(Distress Measure)i,t + 2(Time-to-Maturity) + i(Controls)t +e i,t (7)
where Spreadi,t+1 is the log yield spread for firm i, at month t+1, Distress is one of the four default measures used, Time-to-Maturity is the maturity of the firm's bond, and i, is a vector of factor sensitivities for control variables, which includes time to maturity and industry dummies defined at one-digit SIC codes. All of our tests are predictive in nature. We construct distress measures using publicly available stock and accounting information over one period of time, and evaluate whether these measures are associated with bond yield or change in bond yield in
16
a subsequent period. In this sense, our work also sheds light on out-of-sample predictability of financial failure of different distress measures. We note that all our distress measures are designed to predict financial failure in one-year horizon. We believe that these measures should also capture a significant part of distress risk in any short- and long horizons. To assess performance, we take a simple empirical approach. We regress each distress measure in the cross-section. To build a time-series, we follow the approach of Fama and MacBeth (1973) and report the time-series average of the coefficients, along with t-statistics based on Newey and West (1987) standard errors. The average adjusted R-squared is used to assess the relative power of these measures to explain the market's perception of risk, the credit spread. We also compare the results of the different distress measures to relying on market leverage and report, % Improvement, which reflects the additional explanatory power the distress measure provides beyond market leverage. The results of the regression tests are presented in Table 5. In Panel A, we regress credit spreads on the distress measure alone. In Panel B, we add time-to-maturity to ensure that differences in risk associated with different time-to-maturity is not driving our results. In Panel C, we then include industry control variables. In this framework, we find that the Z-Score underperforms market leverage in all three Panels. explaining credit spreads than market leverage. The O-Score is marginally better in The Merton DD and CHS Score are
dramatically better in explaining credit spreads than market leverage, Z-Score or O-Score with the CHS Score having slightly higher explanatory power than the Merton DD Score. In Panel D, we explore the sensitivity of yield spread to the change in distress risk. We regress change in yield spread from prior year on change in distress measure over the same time period in the spirit of Fama-MacBeth. Overall, CHS Score has the highest adjusted Rsquared followed by Merton-DD. As a further test of the explanatory power of these distress scores, we examine the ability of distress measures to explain credit spreads after any time constant explanatory variables are removed. In Table 6, we conduct further regression analysis. We begin by including firm and industry fixed effects in Panel A. We include the distress measures and time-to-maturity as the explanatory models. The industry fixed effect will capture industry specific risk. The Z-Score has the lowest R-squared across all the models, including market leverage. The O-Score
17
outperforms the market leverage model by 18%, but the Merton DD and CHS model have notably higher explanatory power. Next, we add year fixed effects to capture macro level risk changes in Panel B. While the explanatory power improves, the ranking of the different distress scores remains the same with the Z-Score having the lowest and the CHS score the highest. 5
3.4 Relative Performance across Industries A researcher may be concerned that distress measures could have significant variation across industries. First, distress measures use financial information in their construction, but firms in different industries have different accounting rules. Firms may be more or less sensitive to the risk factors included in distress measures, implying that the probability of bankruptcy can differ for firms in different industries with otherwise identical financial statements. Second, industries respond differently to macro economic shocks. Different industries also face different levels of competition. Despite the strong economic intuition
suggesting that industry effects should be an important component in bankruptcy prediction, not much attention has been paid to industry effects in extant literature, mostly likely due to the limited number of bankruptcies in each industry. The only exception is Chava and Jarrow (2004), which group sample firms into four industry categories and find that predictive variables have different power to forecast bankruptcy for different groupings. For a distress measure to be accurate, we would hope to see a high and similar R-Squared across industry. To assess if this is the case, in Table 7 Panel A, we separate sample firms by industry and examine the relative performance of each measure in each industry. We regress spread on each distress measure, controlling for time-to-maturity. As before, we report coefficient and t-statistic on each distress measure in upper panel and adjusted R-squared in bottom panel. Given that we do not have sufficient number of firms in agriculture industry (1digit SIC code = 0) and in other service (1-digit SIC code = 8) and public admin and other (1digit SIC code = 9) industries, we focus on 6 industries only. We find significant variation and low power for market leverage, Z-score and O-Score. We find appreciably higher explanatory
5 This is essentially an annual analysis since we only have one observation for each firm-year in the sample. This analysis thus alleviates a potential concern that the poor performance of Z- and O-Score is due to their relatively slowness in updating predictive constituents: these two scores are updated annually but Merton-DD and CHS Score are updated monthly. Though not tabulated, we also only use yield spread in December or yield spread in the fiscal ending month for each firm-year to conduct Fama-MacBeth regression analysis and obtain similar results.
18
power for the Merton DD and CHS Score as compared to the other 3 measures. This is universal across all industry types. We also find much less variation in the explanatory power across industries for these two measures. Overall, CHS Score is still the best measure of financial distress in all 6 industries. 3.5 Relative Performance over Time There is a broad variation in corporate failures over time, including the long-lasting increase in the 1980s and cyclical spikes in the early 1990s and early 2000s. A good bankruptcy forecasting models must quantify the time-series effects of the change in economic structure on financial distress. For this reason, in Table 7 Panel B, we conduct our regression analysis in three different time periods and in normal and shock economic periods, where shock period includes the year 1987, 1989, 1994, 1997, 2000, and 2001, and normal period includes the remaining years. It shows that CHS Score outperforms other measures in each time period and in both normal and shock period. The CHS Score outperformance is more evident during years of economic shocks. The CHS Score can explain 54% of cross-sectional variation of yield spread in shock periods. This number is much greater than the corresponding numbers for Merton-DD (45%), O-Score (36%), Z-Score (27%), and market leverage (37%).
3.6 Relative Performance over Size, Market-to-Book, and volatility quartiles So far we have shown that in both short and long horizons, the CHS Score outperforms the Merton-DD, which outperforms Z- and O-Scores with respect to their effects on yield spreads. As we know, distress risk is highly associated with equity size, market-to-book ratio, and volatility. Next we separate sample firms by size, market-to-book ratio, and volatility to examine further the relative performance of each distress measure. In Table 7 Panel C, we sort firms by market equity, market-to-book, and equity volatility respectively into four quartiles and conduct regression analysis in each of the two extreme quartiles and report results under lower and upper quartiles. We also conduct regression analysis on all firms in the two middle quartiles and report the results under middle quartiles. As before, we control for time to maturity and industry effect but only report coefficients on distress measures. We report Newey and West (1987) adjusted t-statistics and time-series average of adjusted R-squared for
19
each specification.
prominent in small firms. The Adjusted R-squared is around 25%-37% and decreases in size for Merton-DD and CHS Score and is around 22% for market leverage, O-Score, and Z-Score. In terms of relative performance, CHS outperforms other measures in smallest size quartiles but does not outperform other measures in larger size quartiles. Similarly, we sort sample firms into quartiles by market-to-book ratio and equity volatility. While CHS Score outperforms other three measures in each M/B quartile, it only outperforms other measures among firms with largest volatility. We conclude that CHS outperforms other measures among firms with highest distress risk. 3.7 Liquidity and Tax Effects The literature suggests that liquidity and tax factors explain a significant part of yield spreads. Because the LBFI does not provide data on bid-ask spreads, we proxy for crosssectional difference in corporate bond liquidity using time to maturity, log of bond age, and log of offering amount. We use time to maturity to control for liquidity risk in our main analysis. We also proxy for tax effect on the cross-sectional variation in bond spreads using bond coupon. We conduct two additional tests to examine the robustness of our results to liquidity and tax effects. First, we include log of offering amount (log of bond age) and coupon as regressors in the multivariate regression. We obtain similar results, which are not tabulated. Second, we sort firms into quartiles by bond age, offering amount and coupon rate respectively and conduct regression analysis in each quartile (results not reported). Our conclusion remains intact for each quartile, though CHS outperformance is relatively more evident among firms with high offering amount and coupon rate.
4. Conclusion The identification of an accurate measure of distress risk is of significant concern for financial research. A number of models are currently being used in the literature with outcomes that can be dependent on the measure utilized. In this paper, we compare the relative performance of the four most commonly used distress measures, namely, Altman (1968)s ZScore, Ohlson (1980)s O-Score, KMV-Merton distance to default (Merton-DD), and Campbell et
20
al. (CHS) measure. We utilize credit spreads to measure distress risk and compare the default models and our naive measure, market leverage, to see which one is the most accurate measure of distress risk. We define the best based on the distress measures ability to explain credit spreads in aggregate, over time, across industries, and relative to market priced risk factors (size, market-to-book, and volatility). We find that CHS Score has the greatest explanatory power over cross-sectional variation of cost of debt followed closely by Merton-DD. The two accounting-based measures, the Z-Score and O-Score, while having a significant impact on yield spread and thus containing distress information, perform poorly comparing to other two measures. The Z-Score, in fact, most often performs worse than what we consider the naive model, market leverage. We also find that adjusting predictive variables to control for R&D effect and using updated coefficients do not improve the relative performance of these two accounting measures. Our findings are consistent when relying on bond rating or credit spreads to assess risk. We find similar results when relying on either a regression or decile approach, which should mitigate a concern that we are not capturing some non-linearity in the regression analysis. The results suggest that relying on either the Z- or O-score for distress risk is highly problematic, and that there are easily calculated alternatives with notably greater power in capturing distress risk.
21
References
Agarwal, V. and R. Taffler, 2008. Comparing the performance of market-based and accountingbased bankruptcy prediction models, Journal of Banking and Finance 32, 1541-1551. Almeida, Heitor, and Thomas Philippon, 2007. The risk adjusted cost of financial distress, Journal of Finance 62, 2557-2586. Altman, E., 1968. Discriminant analysis and the prediction of corporate bankruptcy, Journal of Finance 23, 589-609. Anginer, Deniz, and Yildizhan, Celim, 2010. Is there a Distress Risk Anomaly? Corporate Bond Spread as a Proxy for Default Risk, Working Paper. Avramov, Doron, Tarun Chordia, Gergana Jostova, and Alexander Philipov, 2008. Dispersion in Analysts Earnings Forecasts and Credit Rating, Journal of Financial Economics 91, 83-101. Begley, J., J. Ming, and S. Watts, 1996. Bankruptcy classification errors in the 1980s: An empirical analysis of Altman's and Ohlson's models, Review of Accounting Studies 1-4, 267-284 Beim, D., 1992. Estimating bond liquidity, Working paper, Columbia University. Beaver, W., McNichols, M., and Rhie, J. 2005. Have financial statements become less informative? Evidence from the ability of financial ratios to predict bankruptcy, Review of Accounting Studies,10, 93122. Bharath, S. and T. Shumway, 2008. Forecasting default with the KMV-Merton model, Review of Financial Studies 21, 1339-1369. Campbell, John Y., Jens Hilscher, and Jan Szilagyi, 2010. In search of distress risk, Journal of Finance 63, 2899-2939. Campbell, J., and G. Taksler, 2003. Equity volatility and corporate bond yields, Journal of Finance 58, 2321-2349. Davydenko,S. and I. Strebulaev, 2007. Strategic actions and credit spreads: An empirical investigation, Journal of Finance 62, 2633. Dhaliwal, D., and S. Reynolds., 1994. The effect of the default risk of debt on the earnings response coefficient, The Accounting Review 69, 412-419. Duffie, D., L., Saita, and K. Wang, 2007, Multi-period corporate failure prediction with stochastic covariates, Journal of Financial Economics 83, 635-665. Fama, Eugene F., and James D. Macbeth, 1973. Risk, Return, and Equilibrium: Empirical Tests, Journal of Political Economy 71, 606-636.
22
Franzen, Laurel A., Kimberly J. Rodgers, and Timothy T. Simin, 2007. Measuring distress risk: The effect of R&D intensity, Journal of Finance 62, 2931-2967. Hillegeist, S., E. Keating, D. Cram, and K. Lundstedt, 2004. Assessing the probability of bankruptcy, Review of Accounting Studies 9, 5-34. Hotchkiss, Edith S., Kose John, Robetrt M. Mooradian, Karin S. Thorburn, 2008, Bankruptcy and the resolution of financial distress, working paper. Kealhofer, S., and M. Kurbat. 2001. The default prediction power of the Merton approach: relative to debt ratings and accounting variables (KMV LLC). Klock, M., S. Mansi, and W. Maxwell, 2005. Does corporate governance matter to bondholders? Journal of Financial and Quantitative Analysis 40, 693-719. Mansi, S., W. Maxwell, and J. Wald, 2009. Creditor Protection Laws and the Cost of Debt, Journal of Law and Economics 52, 701-726. Merton, R. 1974. On the Pricing of Corporate Debt: The Risk Structure of Interest Rates, Journal of Finance 29, 44970. Newey, Whitney K, and Kenneth D West, 1987. A Simple, Positive Semi-definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix, Econometrica 55, 703-708. Ohlson, J., 1980. Financial ratios and the probabilistic prediction of bankruptcy, Journal of Accounting Research 18, 109-131. Shumway, T., 2001. Forecasting bankruptcy more accurately: a simple Hazard model, Journal of Business 74, 101-124.
23
24
Market Leverage Z Score O Score KMV DD CHS Score R&D Adj Z Score R&D Adj O Score Up. Z Score Up. O Score MktCap Market-to-Book Volatility Maturity Rating
1.000 0.571 0.638 0.582 0.991 0.543 0.287 -0.363 0.247 0.406 -0.247 0.621
1.000 0.725 0.418 0.571 0.449 0.045 -0.340 -0.062 0.594 -0.270 0.686
1.000 0.416 0.634 0.391 0.168 -0.280 0.074 0.728 -0.236 0.611
Note: Our sample firms include all non-financial issuers in LBFI (198001-200612) with necessary CRSP and Compustat information to compute distress measures and controlling variables. For firms with multiple issues only the one with greatest maturity is selected so that each firm will only have one bond in the analysis. We remove AAA bonds and bonds with less than one-year to maturity. Z-Score is Altman (1968) and O-Score is Ohlson (1980) bankruptcy measures. R&D Adj. Z-Score and Adj. O-Score are constructed by following Franzen et al. (2007) to adjust the R&D effects for constituent variables. Up. Z-Score and Up. O-Score are constructed by using updated coefficients by Hillegeist et al. (2004). DD is distance to default based on Merton (1974) structural model. CHS Score is the bankruptcy score based on the reduced form model by Campbell et al. (2010). The signs of the coefficients have been changed to original, adjusted, and updated Z Scores and Merton-DD so they are increasing in the probability of bankruptcy. As a result, a greater value of all measures signifies higher probability of bankruptcy. Firm size is market equity cap. Market-to-book is the ratio of market equity to book equity. Volatility is the annualized 3-month rolling standard deviation of daily stock returns constructed on by following Campbell et al. (2010). Rating is S&P bond rating. Maturity is the number of years until bond maturity. All variables are winsorized at 1 and 99 percentiles, and the time series average of their cross-sectional distributions is reported in Panel A. Panel B reports the number of firm-month observations for each of the 8 industries. Panel C reports Pearson correlations between distress measures, yield spreads, and controlling variables.
25
Table 2 Bond Ratings and Distress Measures AA All Periods 1980-1982 1983-1985 1986-1988 1989-1991 1992-1994 1995-1997 1998-2000 2001-2003 2004-2006 All Periods 1980-1982 1983-1985 1986-1988 1989-1991 1992-1994 1995-1997 1998-2000 2001-2003 2004-2006 All Periods 1980-1982 1983-1985 1986-1988 1989-1991 1992-1994 1995-1997 1998-2000 2001-2003 2004-2006 All Periods 1980-1982 1983-1985 1986-1988 1989-1991 1992-1994 1995-1997 1998-2000 2001-2003 2004-2006 0.20 0.26 0.31 0.25 0.24 0.17 0.13 0.10 0.10 0.09 -3.24 -3.19 -2.78 -2.93 -2.82 -3.35 -3.72 -4.42 -4.79 -5.06 -2.37 -2.66 -2.42 -2.11 -2.01 -2.26 -2.44 -2.50 -2.51 -3.07 -9.55 -7.20 -8.63 -8.42 -10.31 -11.71 -12.55 -8.81 -8.75 -16.43 A 0.26 0.39 0.35 0.29 0.28 0.28 0.22 0.18 0.19 0.15 -2.92 -2.91 -2.77 -2.86 -2.79 -2.62 -2.83 -3.15 -3.19 -3.11 -1.83 -2.10 -1.91 -1.72 -1.71 -1.44 -1.80 -1.80 -1.91 -2.23 -7.68 -6.14 -7.05 -6.75 -8.42 -8.59 -9.62 -6.31 -6.33 -12.08 BBB Market Leverage 0.36 0.65 0.50 0.43 0.49 0.41 0.31 0.29 0.33 0.28 Z-Score -2.17 -2.06 -2.16 -2.13 -1.79 -1.78 -2.29 -2.31 -2.16 -2.48 O-Score -1.19 -1.30 -1.42 -1.35 -0.98 -0.81 -1.03 -1.15 -1.13 -1.68 Merton-DD -6.03 -4.91 -6.00 -5.33 -6.13 -6.68 -7.46 -4.84 -4.80 -9.24 CHS Score BB 0.44 0.50 0.53 0.47 0.48 0.52 0.44 0.42 0.45 0.35 -2.09 -2.57 -1.74 -1.90 -2.01 -1.82 -2.06 -2.19 -2.20 -2.14 -0.72 -0.55 -0.82 -0.66 -0.56 0.02 -0.52 -0.54 -0.87 -1.31 -4.21 -3.33 -3.97 -3.37 -4.63 -4.33 -4.95 -3.15 -3.48 -6.26 B 0.54 0.56 0.58 0.56 0.70 0.59 0.53 0.52 0.58 0.46 -1.75 -2.28 -1.91 -2.01 -1.71 -1.54 -1.70 -1.63 -1.42 -1.87 0.10 0.06 0.08 0.18 0.32 0.50 0.28 0.27 -0.02 -0.41 -2.83 -2.67 -3.07 -2.41 -2.10 -2.97 -3.32 -2.14 -2.30 -4.63 26
All Periods 1980-1982 1983-1985 1986-1988 1989-1991 1992-1994 1995-1997 1998-2000 2001-2003 2004-2006
-8.25 -8.21 -8.28 -8.24 -8.25 -8.30 -8.31 -8.03 -8.24 -8.44
-8.11 -8.13 -8.15 -8.11 -8.14 -8.08 -8.12 -7.85 -8.10 -8.33
-7.89 -7.91 -7.96 -7.89 -7.84 -7.87 -7.91 -7.62 -7.86 -8.14
-7.66 -7.61 -7.68 -7.64 -7.66 -7.51 -7.66 -7.26 -7.63 -7.98
-7.27 -7.33 -7.41 -7.29 -7.03 -7.25 -7.25 -6.92 -7.14 -7.67
Note: This table reports the median distress measures during the entire sample period and nine sub-periods by major rating categories.
27
Note: This table reports the median distress measures in June 2006 by industry and by major rating categories. We do not report results for AA bonds because many industries do not have bonds in this rating category.
28
Decile 1 2 3 4 5 6 7 8 9 10
DD 100 132 161 167 206 228 275 285 307 519
CHS 142 144 153 166 196 206 233 260 348 531
Panel A Average yield Spread Across Deciles 131 134 126 175 168 163 171 203 194 215 175 183 181 219 217 233 228 260 249 234 225 234 257 255 327 303 307 464 457 450 Panel B Double Sorting Percentage Analysis 51.7% 50.7% 51.4% 49.1% 53.5% 54.1% 42.6% 44.6% 45.0% 37.9% 44.8% 44.3% 38.5% 40.1% 42.5% 40.0% 42.4% 42.3% 37.3% 43.2% 42.5% 43.8% 47.4% 48.3% 57.9% 66.5% 66.6% 55.4% 70.9% 69.8%
1 2 3 4 5 6 7 8 9 10
50.9% 57.1% 43.5% 42.0% 40.6% 38.8% 39.7% 43.7% 53.8% 62.0%
45.6% 48.6% 38.7% 37.0% 34.6% 33.9% 33.9% 35.7% 37.4% 50.8%
45.2% 48.9% 38.9% 37.1% 34.8% 34.5% 34.0% 35.6% 37.1% 50.3%
11.0% 24.4% 28.5% 36.2% 37.5% 30.6% 23.6% 27.6% 44.9% 39.0%
52.9% 59.9% 51.2% 44.0% 44.4% 44.1% 50.0% 60.3% 72.3% 73.0%
45.5% 56.7% 46.2% 42.2% 40.5% 41.7% 42.9% 53.5% 68.8% 80.4%
Note: This table reports results based on single and double sorting. In Panel A, we sort firms into 10 deciles by each distress measure in June 2006 and compute the average yield spread for each decile. In Panel B, we sort firms into deciles based on each distress measure and also on credit spreads in June of each year. We then compute the time-series average of the percentage of firms in the distress measure decile that are also in the same or the adjacent deciles based on credit spread sorting.
29
Altman Z-Score 0.189 (29.60) 0.135 383 -58% 0.196 (29.82) -0.022 (-12.88) 0.213 383 -41% 0.235 (29.53) -0.017 (-8.29) Included 0.292 383 -29% 0.061 7.04 0.011 302 -34%
Ohlson O-Score 0.308 (33.23) 0.342 383 6% 0.294 (27.37) -0.008 (-3.28) 0.367 383 1% 0.294 (31.09) -0.004 (-1.71) Included 0.410 383 0% 0.029 5.77 0.005 302 -68%
Merton DD 0.154 (23.29) 0.435 383 35% 0.15 (23.36) -0.005 (-2.23) 0.452 383 25% 0.144 (21.55) -0.005 (-2.19) Included 0.474 383 16% 0.029 4.97 0.03 302 77%
CHS Score 0.731 (38.83) 0.467 383 45% 0.711 (37.34) -0.006 (-3.14) 0.485 383 34% 0.694 (34.02) -0.004 (-2.45) Included 0.514 383 25% 0.244 12.9 0.105 302 517%
2.042 (38.15) 0.323 383 n.a. 1.957 (37.67) -0.015 (-11.10) 0.362 383 n.a. 1.971 (39.81) -0.011 (-6.37) Included 0.410 383 n.a. 0.357 (6.16) 0.017 302 n.a.
Note: This table reports Fama-MacBeth regression analysis results. In Panels A, B, and C, the dependent variable is log yield spread. In Panel A, the independent variable is each distress measure. In Panel B, we control for time to maturity. In Panel C, we control for time to maturity and industry effects. In Panel D, the dependent variable is change in log yield spread from month t+1 to t+13, and the independent variable is yearly change in distress measures from t to t+12. We conduct cross-sectional regression each month and compute the time-series average of coefficients and adjust R-squared. All t-statistics (in parentheses) are adjusted using the Newey-West(1987) procedure with lag 12. Coefficients on intercept and industry dummies are not reported. % improvement is the percentage change in adjusted R-squared for each model compared to market leverage model.
30
Panel A Firm and Industry Fixed effect Distress Measures t-stat Maturity t-stat Adjusted R-Squared Number of Firms % Improvement 1.709 (29.12) -0.019 (-13.26) 0.312 10,306 n.a. 0.249 (22.6) -0.022 (-15.28) 0.282 10,306 -10% 0.286 (36.36) -0.011 (-7.68) 0.369 10,306 18% 0.131 (40.25) -0.011 (-9.58) 0.446 10,306 43% 0.683 (52.36) -0.009 (-6.9) 0.49 10,306 57%
Panel B Firm, Year, and Industry Fixed Coef t-stat Maturity t-stat Adjusted R-Squared Number of Firms % Improvement 1.811 (31.33) -0.014 (-10.06) 0.445 10,306 n.a. 0.246 (23.8) -0.02 (-13.53) 0.395 10,306 -11% 0.288 (37.94) -0.008 (-5.82) 0.483 10,306 9% 0.134 (38.2) -0.008 (-6.01) 0.524 10,306 18% 0.679 (52.25) -0.007 (-5.57) 0.575 10,306 29%
Note: This table reports results for regressions with fixed effects using sample firms in June of each year. Panel A reports regression of log yield spread with fixed firm and industry effects. Panel B reports regression of log yield spread with fixed firm, year, and industry effects. Coefficients on intercept and industry dummies are not reported.
31
Table 7 Regressions by Industry, Subperiods, Market Cap, Market-to-Book Ratio, and Volatility
Panel A. By Industry Mining and Construction 1.952 15.86 0.362 0.267 5.28 0.267 -26% 0.219 9.06 0.318 -12% 0.185 17.33 0.449 24% 0.629 15.12 0.438 21% Light Manufactured 2.564 42.81 0.461 0.236 24.23 0.233 -49% 0.329 23.29 0.383 -17% 0.141 19.62 0.441 -4% 0.827 35.03 0.478 4% Heavy Manufactured 2.034 33.40 0.381 0.267 22.28 0.237 -38% 0.312 34.66 0.381 0% 0.143 20.3 0.409 7% 0.704 30.65 0.448 18% Communication and Electronics 1.446 14.53 0.233 0.256 11.85 0.197 -15% 0.273 25.27 0.31 33% 0.139 12.77 0.399 71% 0.68 22.94 0.445 91% Whole and Retail Trade 2.002 20.96 0.401 0.195 16.16 0.253 -37% 0.297 16.71 0.384 -4% 0.164 22.58 0.489 22% 0.736 22.35 0.516 29% Business Service 1.357 -7.62 0.237 0.187 -6.95 0.197 -17% 0.191 -18.8 0.187 -21% 0.176 -10.03 0.435 84% 0.589 -10.06 0.438 85%
Market Leverage t-stat Adj. R-Squared Z Score t-stat Adj. R-Squared % Improvement O Score t-stat Adj. R-Squared % Improvement Merton DD t-stat Adj. R2 % Improvement CHS Score t-stat Adj. R-Squared % Improvement
32
33
Panel C. By Quartiles of Firm Size, Market-to-Book, and Volatility Bottom Quartile 1.101 (9.66) 0.225 0.177 (7.81) 0.191 -15% 0.167 (13.24) 0.227 1% 0.166 (11.75) 0.366 62% 0.443 (17.23) 0.446 98% 95 Market Cap Middle Quartiles 1.326 (22.32) 0.253 0.160 (22.43) 0.204 -19% 0.192 (27.12) 0.247 -2% 0.099 (17.24) 0.321 27% 0.557 (31.41) 0.318 26% 191 Upper Quartile 0.879 (11.31) 0.258 0.085 (6.78) 0.231 -11% 0.128 (9.09) 0.237 -8% 0.049 (10.01) 0.275 6% 0.383 (16.56) 0.257 0% 96 Bottom Quartile 1.691 (15.35) 0.388 0.190 (9.31) 0.283 -27% 0.291 (21.6) 0.409 5% 0.169 (12.72) 0.488 26% 0.666 (27.31) 0.571 47% 95 Market-to-Book Middle Quartiles 1.703 (23.17) 0.325 0.165 (18.7) 0.203 -38% 0.263 (24.49) 0.357 10% 0.113 (19.65) 0.361 11% 0.661 (30.9) 0.370 14% 191 Upper Quartile 2.318 (28.92) 0.496 0.274 (19.43) 0.416 -16% 0.305 (28.1) 0.517 4% 0.159 (26.15) 0.546 10% 0.686 (28.4) 0.550 11% 96 Bottom Quartile 1.608 (23.66) 0.296 0.164 (14.94) 0.211 -29% 0.189 (19.84) 0.241 -19% 0.070 (13.61) 0.265 -11% 0.606 (19.86) 0.252 -15% 96 Volatility Middle Quartiles 1.662 (32.75) 0.336 0.174 (28.03) 0.209 -38% 0.245 (22.33) 0.315 -6% 0.118 (19.25) 0.330 -2% 0.653 (33.29) 0.298 -11% 191 Upper Quartile 1.583 (19.43) 0.316 0.226 (13.27) 0.239 -25% 0.238 (18.47) 0.320 1% 0.247 (18.71) 0.471 49% 0.558 (22.9) 0.467 48% 96
Mkt. Leverage t-stat Adj. R-Squared Z-Score t-stat Adj. R-Squared % Improvement O-Score t-stat Adj. R-Squared % Improvement Merton-DD t-stat Adj. R-Squared % Improvement CHS Score t-stat Adj. R-Squared % Improvement Number of Firms
Note: In Panel A, we separate sample firms into 6 industries and conduct regression analysis within each industry. In Panel B, we separate sample firms into three time periods (1980-1990, 1991-2000, 2001-2006) and into normal and shock macroeconomic periods, where shocks years include 1987 (stock market crash), 1989 (real estate bubble), 1994 (high interest rates), 1997 (Asian financial crisis), 2000 (Internet bubble), and 2001 (9/11 incident), and normal years include all remaining years, and conduct regression in each period. In Panel C, we first sort firms into four quartiles every month from 198001-200612 by market equity cap, market-tobook ratio, and equity volatility respectively and then conduct regression within each quartile of Q1 and Q4 and two quartiles of Q2 and Q3 together. In all Panels we control for maturity. In Panels B and C we also control for industry effect. We run cross sectional regression on distress measures and report the timeseries average of coefficients and associated t-statistics (in the below parentheses) adjusted by Newey-West (1987) with lag 12. Also reported are the average adjusted R-squared and the number of firms in the cross-sectional regression. Coefficients and t-statistics for intercept, maturity, and industry dummies are
34