Information, Analysts, and the Cost of Debt Sattar A. Mansi *, William F. Maxwell **, and Darius P. Miller *** * Pamplin College of Business Virginia Tech Blacksburg, VA 24061 (703) 538-8406 smansi@vt.edu ** Eller College of Management University of Arizona Tucson, AZ 85721 (520) 621-1716 maxwell@eller.arizona.edu *** Kelley School of Business 1309 East Tenth Street Bloomington Indiana 47401-1701 812.855.3395 Fax: 812.855.5875 damiller@indiana.edu First Version: September 17, 2004 This Version: September 27, 2004 We would like to thank seminar participants at Wharton for comments and suggestions that significantly improved the paper. We thank Lehman Brothers Fixed Income group for providing debt data. All errors are the sole responsibility of the authors.
Information, Analysts, and the Cost of Debt Abstract We examine the relation between financial analyst forecast characteristics and the cost of debt financing. Using a sample of 5827 firm-year observations for the period 1982 through 2002, we find that higher analyst coverage, greater forecast accuracy, and lower forecast dispersion are associated with better credit ratings. We also find that forecast characteristics (accuracy with less dispersion and larger analyst following) are inversely related to the cost of debt financing beyond that incorporated in credit ratings. We further find that the economic impact of analysts forecasts is largest for firms where the composition of private information (i.e., more intangible assets) is highest. Overall, our results support the hypothesis that the information contained in analysts forecasts is related to firms cost of (debt) capital, and that the economic impact of analysts forecasts is related to the composition of firms private and public information.
1. Introduction Is the information environment of the firm related to its cost of capital? Also, does the composition of private and public information affect its economic impact? In this paper, we provide several empirical findings on these questions by investigating the role of analysts in the corporate bond market. First, we document that higher analyst coverage, greater forecast accuracy, and lower forecast dispersion are associated with better credit ratings. Second, we show that after controlling for firm and security specific information, analyst coverage, forecast accuracy, and forecast precision are negatively related to the required return on firms public corporate debt. Third, we show that the economic impact of analyst activity is largest for firms with relatively larger amounts of private information. Our results are robust to controls for default and market risk as well as corrections for the endogeneity between analyst coverage and performance. Overall, our results support the hypothesis that the information contained in analysts forecasts is related to firms cost of (debt) capital, and that the economic impact of analysts forecasts is related to the composition of firms private and public information. Easley and O Hara (2004) develop a theoretical model where investors know about all assets, but information is asymmetric. In their setting, private information represents a new form of systematic risk with investors demanding higher returns to hold stocks with greater private (and relatively less public) information. Their model predicts that information has a larger impact on the cost of capital of firms with relatively more private information. In conjunction with this work, a recent strand of research suggests that the composition of private and public information is related to analyst characteristics. Barron et al. (2002) find that earnings forecasts of high-intangible firms contain a higher proportion of private information. Barth et al. (2001) show that analyst coverage is higher, and that analyst expend greater effort in firms with more intangible assets. These empirical studies indicate that analyst activity is related to the level of private information in the firm. Taken together, these studies reinforce the relation between analysts and firms cost of capital and suggest that the composition of private and public information is important in understanding the economic impact of analysts forecasts. Note, however, that the association between the information contained in the properties of analysts forecasts and the cost of capital may not necessarily be negative. Diether, Malloy and Scherbian (2002) and Johnson (2004) find that lower dispersion of analysts forecasts is
associated with a higher equity returns. This implies that, for at least one attribute of analyst forecasts, reducing information uncertainty may actually increase the cost of capital. 1 Another strand of research considers the impact of estimation risk, or the uncertainty about the return distribution parameters of securities, on asset prices (see, e.g. Barry and Brown (1985) and Coles, Loewenstein, and Suay (1995)). In these models, an increase in information allows investors to better estimate asset returns, which in turn lowers the required rate of return. Another line of research examines the role of information that is incomplete, but not asymmetric (e.g., Merton (1987)). In this model, information can expand the awareness of a security, which broadens its shareholder base and reduces the required rate of return. Overall, these models predict that firms with better information environments enjoy a lower cost of capital. 2 Our study builds on prior empirical research which suggests that analysts forecasts play a role or reflect the information environment of the firm. Lang and Lundholm (1996), for example, find that company-provided information does not subsume the information contained in analysts forecasts. Healy, Hutton, and Palepu (1999) show that firms with more informative disclosures have higher analyst following, suggesting that firms disclosures and information production by analysts are complements. Barth and Hutton (2003) find that analysts earnings forecasts contain information about earnings persistence that facilitates more accurate pricing of accruals, a result that is consistent with the hypothesis that analysts are information intermediaries in the capital markets. Internationally, Bushman, Piotroski, and Smith (2004) and Lang, Lins, and Miller (2003) show a relation between increased disclosures and analyst following. This strand of literature suggests that analysts forecasts are at least indicative and in part contribute to the information environment of the firm. Our study contributes to several streams of literature. First, we extend the literature on how firms information structure is related to its cost of capital, documenting that the required return on corporate bonds is lower when firms have more analysts making more accurate forecasts with less dispersion. Empirical evidence on the direct impact of information contained in analysts forecasts on the cost of capital in equities is limited and to the best of our knowledge this is the first study that 1 Further, financial analysts earning forecasts, as a proxy for information quality, have been found to be related to firm risk and financial distress, and therefore may contribute to stock price volatility (Ciccone, 2003; Veronesi, 2000). Oyer (1998) also found that analyst forecast activities in the context of earning announcements are the biggest predictor of stock price volatility. Increase in equity volatility is associated with an increase in probability of financial distress (lower credit ratings) and in a wealth transfer from bondholders to stockholders. 2 It is, however, important to note the association between disclosure and the cost of capital can be negative. Diamond and Verrecchia (1991) show that disclosure changes the risks for market makers which influences the entry or exit of market makers. Therefore, disclosure can worsen or improve liquidity depending on the dealer decision. 2
examines this relation from the perspective of bondholders. For example, Botosan (1997) documents a negative association between corporate disclosures and the cost of equity capital in firms with low analyst following while Botosan and Plumlee (2002) show that the cost of equity is decreasing in annual report disclosures but increasing in the level of timely disclosures. 3 Sengupta (1998) provides evidence that disclosure quality is negatively related to the firm s at-issue cost of debt. Our findings on the economic impact of the information contained in analysts forecasts on credit spreads provides direct evidence on the link between information and the cost of capital for corporate bonds, a security that makes up a significant portion of many firms market capitalization and a market that is one of the world s largest. The bond market also provides a unique setting to examine the information content of analysts forecasts given the presence of other information intermediaries (i.e., credit rating agencies). 4 Further, bonds have shorter durations when compared to equity, and their valuations are well specified and less subject to the criticism that the results might be driven by misspecification of the equilibrium asset pricing model. Second, our results add to the literature on private information and intangibles in determining the cost of capital. While our results are consistent with hypothesis that better information environments lead to an overall lower cost of capital, we also find that the composition of private and public information is important. In this way, our finding that the relation between analysts forecasts and the cost of capital is larger for firms with relatively more private information (i.e., more intangible assets) provides empirical support for the model of Easley and O Hara (2004) which predicts the quantity of information, the dispersion of information, and the quality of public and private information are related to a firm s cost of capital. 5 Third, our findings add to recent research on private information and analyst activity that finds that analyst coverage is higher for firms with relatively more private information and that the information contained in these forecasts contains a higher proportion of that private information (Barth et al. (2001), and Barron et al. (2002)). Our results suggest the impact of this analyst coverage is economically larger in firms with relatively high level of private information. That is, while 3 More evidence exits using microstructure data. For example, Welker (1995) and Luez and Verrecchia (2000) document that increase disclosure results in lower bid-ask spreads and Easley, Hvidkjaer, and O Hara (2002) show that stocks with information based trading have higher rates of return. Gebhardt, Lee, and Swaminathan (2001) find that forecast accuracy is negatively related to a firm s implied cost of capital. 4 Ederington and Goh (1998) show that analysts forecasts are linked to credit ratings downgrades. Further, Millon and Thakor (1985) argue that credit rating agencies are formed to act as screening agents that certify the value of firms. There also exits a large literature on the role of credit ratings in affecting stock and bond prices (see, e.g., Hand, Holthausen, and Leftwich (1992)). 5 Easley and O Hara (2004) note on pp. 1573 that their model predicts that..companies benefit from having many analysts because analysts increase the precision of information and this lowers the companies cost of capital. 3
information production by analysts, on average, is associated with a lower cost of capital, the impact of analysts forecasts on the cost of capital is larger in firms with relatively more private information. In this way, our findings also add to our understanding of analysts role as information intermediaries. The remainder of the paper is organized as follows. Section 2 describes the data and methodology used to measure the cost of debt and forecast characteristics (forecast accuracy and dispersion, and analyst following). Section 3 presents the empirical results in a multivariate framework, discusses their implications, and provides alternative specifications for robustness. Section 4 concludes the paper. 2. Data Description 2.1. Data Sources In our analysis, we utilize financial, earning forecasts, corporate debt, insider and institutional holdings databases. Financial data are complied from Compustat over the 1984 to 2002 time period. Analyst information is gathered from the Institutional Broker Estimation System (I/B/E/S) annual consensus earnings forecast database. Lehman Brothers provided the data to calculate month-end credit spreads and bond rating information. 6 Compustat s Execucomp is used to collect insider holdings ownership and compensation information and is available beginning for 1992. Institutional holdings data is gathered from Thomson Financial Institutional Ownership database, which is available beginning in 1990. For a firm-year observation to be included in the analysis, data must be provided in the I/B/E/S based on consensus earnings per share; the Lehman Brothers database on the amount, yield to maturity, and duration of the firm's non-provisional public debt securities; and in the Compustat database for the control variables. After excluding financial and regulated industries and applying these requirements, we begin with a data set of 5,827 firm-year observations for 1,002 firms. 7 The subsample of firms with insider information and institutional holding information consists of 2,771 firm-year observations over the 1992 to 2000 time period. 6The Lehman Brothers data provides security specific information such as bid trading price, accrued interest, coupon, yield to maturity, duration, convexity, credit ratings from Moody s and S&P, and quote, issue, and maturity dates on non-provisional (options, warrants, and convertibles). It includes both investment and non-investment grade bonds. The database has been used extensively in both the accounting and finance literature. 7 To minimize survivorship bias, firms are allowed to exit and reenter the data set. 4
2.2. Measuring the Cost of Debt Financing For each debt security, the dependent variable, credit spread (Spread) is measured as the difference between the yield to maturity (Yield) on the firm s outstanding traded debt and its duration equivalent yield to maturity on a Treasury security. For firms with multiple debt securities, we estimate an overall credit spread for the firm based on a weighted average of the bonds credit spreads, with the weight being the amount outstanding for a bond divided by the total amount outstanding for all traded bonds in the firm. Hence, the weighted average firm credit spread (CreditSpread) for firm k is Credit Spread k = Credit J i= 1 Spreadi w i (1) where J is the number of bonds outstanding for firm k and w is the relative amount outstanding of bond i to the amount outstanding of all traded bonds for firm k. 8 2.3. Measuring Information Characteristics We use three variables from Lang and Lundholm (1996) as proxies for the information contained in analysts forecasts. ` Accuracy = The negative absolute value of the analyst forecast error (the actual EPS minus the median forecast deflated by the stock price. 9 Std. Deviation of Forecasts (Dispersion) Number of Analysts = The standard deviation of the inter-analyst forecast deflated by the stock price. = The natural log of the number of analysts following the stock. We make two adjustments. Given the well known problem of outliers, we winsorize Accuracy and Dispersion at the 99 percent level. In addition, we multiply these two measures by 100 to scale the variables relative to the dependent variable, credit spreads, which is reported in basis points. 8 For firms with multiple issues, we examine alternative weighting schemes (equal weighting and only using the most recently issued) and find no difference in our results. 9 Gu and Wu (2003) have found that using the median forecast consensus rather than the mean provides the most accurate forecast. 5
While each of the above information variables is unique, the three measures are highly correlated. In fact, when looking at the correlations in Table 3, we find that Accuracy has a correlation of -0.627 and 0.253 with the standard deviation of inter-analyst forecasts (Dispersion) and the number of analysts (NAnalysts), respectively. To separate out the unique impact of each variable, we orthogonalize Accuracy with respect to Dispersion and NAnalysts. That is, we estimate Accuracy = β 0 + β 1 (StdDev) + β 2 (Analysts) + ε (2) We use the residual term from the regression to create a measure of accuracy (AccuracyResidual), which represents the component of Accuracy that is orthogonal to Dispersion and NAnalysts. However, to ensure that the model is not misspecified from imposing a functional form or from measurement error, we also introduce Accuracy, Dispersion, and NAnalysts into our main models one at a time. In the spirit of Bushman, Piotroski, and Smith (2004), we also employ a factor analysis to determine the commonality in our three information variables. This allows us to present a measure of information that represents the overall analyst activity for the firm. An often-cited goal in constructing a multi-item index model is to identify attributes that have a high degree of intercorrelation and load on a single factor. We find that the three measure of analyst activity load on a single factor (InfoFactor) which has a correlation of 0.91, -0.88, and 0.36 with Accuracy, Disperion and NAnalysts, respectively. We align our variables credit spread and forecast characteristics for each firm-year observation by sampling the credit spread for the period following the actual earnings forecast. That is, once forecast accuracy is calculated, the credit spread is computed for the closest end of month period. For example, for a firm with a December 31, 1990, fiscal year end with actual earnings per share due in January 15, 1991, the credit spread is computed as of January 31, 1991. We calculate Dispersion and NAnalysts one month prior to the fiscal year-end, as O Brian and Bushan (1990) find that analyst activity levels off after the eleventh month of the fiscal year-end (FYE). However, we also reran our analysts using the month of FYE and found similar results. 6
2.4. Control Variables A significant determinant in bond yields is default risk, proxied by credit ratings. For firms with multiple bonds, there can be a variation in ratings within the same firm (due to priority claims). Rating agencies rate the most senior bonds of the firm and then notch bonds downwards with significantly lower levels of investor protections. The notching is typically one minor rating category for investment grade bonds and two minor rating categories for noninvestment grade debt. We report results using a firm s most senior bond rating, and though not reported, we find similar results to the primary specification using a weighted average of a firm s individual bond ratings. 10 Debt credit rating (Rating) is computed using the average of both Moody s and S&P bond ratings and represents the average credit rating at the date of the yield observation. 11 Bond ratings are computed using a conversion process in which AAA rated bonds are assigned a value of 22 and D rated bonds receive a value of 1. For example, a bond with an A1 rating from Moodys and an A+ from S&P would receive an average score of 18. Finally, to control for non-linear effects predominantly present in the noninvestment grade debt market (Mansi and Reeb (2002)), we include an indicator variable for bonds that are considered high yield (or noninvestment grade). Mansi, Maxwell, and Miller (2004) find that the third-party bonding and/or monitoring associated with auditor tenure is related to a firm s rating and spread. Hence, we include Tenure, which calculated as the natural log of the number of years the firm has had the same auditor. 12 We also include FirmAge to control for differences in the information environment of the firm (the log of the number of years from the original IPO). Ceteris paribus, a longer history could reduce information asymmetry between the firm and analysts, rating agencies, and investors. While bond rating is a good proxy for firm risk, it is not a perfect control for firm-risk since there is no expectation that firm risk is perfectly correlated with investors sentiments and a bond rating can lag investors assessment of firm risk (Hand, Holthausen, Leftwich (1992)). Thus, we also include a number of other firm specific measures of risk, which have been found to effect bond ratings and/or credit spreads. We include the size of a firm as it is related to the number of analysts 10 The results are available by request for any robustness tests reported but not tabulated. 11 Anderson, Mansi and Reeb (2003) suggest that using the average of both Moody s and S&P provides the most efficient measure of the default risk premium. We find similar results using either S&P or Moody s as the credit rating variable or indicator variables for each bond rating category. 12 The Tenure variable that takes a value of one in the year the first year financial information becomes available in COMPUSTAT or in 1974 as that is the first year this variable is available in COMPUSTAT. We keep a count for the number of years the firm continues to use the same auditor. There is a downward bias in this measure as the choice of auditing firm is unavailable before the firm enters COMPUSTAT, but this problem is somewhat mitigated by the backfilling of financial data when a firm is added to COMPUSTAT (see Banz and Breen [1986]). 7
following the firm and matters for both the probability of financial distress and the liquidity of a firm s bonds. Given the time period spanned in the study, we measure Size by first normalizing total assets to 1990 dollars and then taking the natural log. To control for capital structure effects, we include (i) the firm s leverage ratio (Leverage), computed as the total long-term debt plus notes payable divided by total assets, (ii) the firm s market-to-book ratio (MtB) as a measure of growth opportunities, computed as the market value of common equity divided by book value of total common equity, and (iii) the firm s working capital divided by total assets (Liquidity). In addition, we include firm profitability (Profitability) or the ratio of operating profit divided by total assets to control for performance and for the firm s information environment. Firm earning volatility (FirmRisk), which is likely related to both Accuracy and Dispersion, is computed as the standard deviation of the firm s return-on-equity over the last three years. We also include a variable to control for bond specific risk. Debt duration (Duration), or the discounted time weighted cash flow of the security divided by its price, represents the security s effective maturity. Duration measures the linearities in the price-yield relation and represents the securities' undiversifiable (or systematic) risk. Duration in this case is the weighted average duration of all public debt outstanding and represents the summation of the weighted durations of all bonds for each firm, with the weight being the amount of traded debt outstanding for an issue divided by the total amount of traded debt outstanding for the firm. In general, we expect duration to be negatively related to credit spread due to the negative relation in the price-yield relation. 13 For a subset of firms with available data, we also include information about the ownership structure of the firm. DeFusco, Johnson, and Zorn (1990) report that risk rises following the adoption of an executive stock option plan and that shareholder wealth rises around the announcement of a plan s adoption while bondholder wealth falls. This suggests management compensation practices can exacerbate the agency conflict between management/shareholders and bondholders (Core, Holthausen, Larcker (1999)). To control for this issue, we includes a measure of the ownership of the top management team in the firm (InsideOwn) and the percentage of their compensation that is related to long-term compensation (Compensation). Long-term compensation is defined as the equity based proportion of total compensation to the firm s top five executives, where equity based compensation is the sum of the total value of restricted stock granted, total value of 13 Although duration is negatively related to credit spreads, the sign of the relation could change due to the change in debt maturity, coupon, and the initial level of interest rates. As such, either dominant factor can change the sign of the relation. 8
stock options granted (using Black-Scholes), and long-term incentive payouts. This measure is then divided by total compensation (equity-based compensation, salary, bonus, other annual compensation, and all other compensation). We also include InstitOwn to represent the institutional ownership of the firm. Finally, we control for industry specific risk and time effects using indicator variables based on the Fama and French s 30 industry classification system and the years under study. 2.5. Descriptive and Univariate Statistics Table 1 provides sample descriptive statistics for the full sample of 1,002 firms. Included are the mean, median, standard deviation, 75th percentile, and 25th percentile values for the variables used in the analysis. The yield spread in the sample has mean and median values of about 237 and 165 basis points, with 25 th and 75 th percentiles of 90 and 331 basis points, respectively. Forecast Accuracy and Dispersion have a mean of about -1.368% and 0.663%, respectively. These measures are close to those reported by Lang and Lundholm (1996). The closer the accuracy variables is to zero the more accurate forecast. The mean and median rating of the firms in the sample equates to a S&P rating of BBB+. The debt has a mean duration of about 5.8 years and, in general, firms have an age of 23 years. In general, the firms in the sample are large with mean (median) total assets of approximately about $7.4 ($2.4) billion. The median leverage ratio is 32.4%, indicating that a large portion of the sample consists of firms that have significant long-term debt in their capital structure. This is expected due to the size of the firms in the sample. On average, there are 11 analysts following the stock, with 75 th and 25 th percentile values of 20 and 7 analysts, respectively. Finally, for the subsample of firms with ownership and compensation data, we find that institutions own a significant portion of the firm, an average of 54.1 percent. Insiders hold on average about 2.6 percent of the stock and on average about 44.6 percent of their compensation is equity related. Table 2 describes the industry distribution of the full sample of 5,827 firm-year observations using the Fama and French classification system based on 30 industries. While the sample covers all the non-regulated industries, the largest industry sectors are retail (9.0%), petroleum and natural gas (7.9%), chemicals (6.5%), business supplies (5.9%), healthcare (5.7%), business equipment (5.1%), and food products (4.8%). The lowest industry sectors are tobacco, beer and liquor, textiles, precious metals, and coal, each contribution less than one percent of the total sample. 9
Table 3 provides the correlation coefficients for the yield spreads, analyst forecast variables, and firm and security specific information including ownership variables (institutional and inside ownership). Using our full sample, we find that yield spreads are negatively related to forecast accuracy, analyst following, institutional ownership, compensation, tenure, age, size, market to book ratio, and profitability, but positively related to forecast dispersion, inside ownership, leverage, liquidity, and firm risk. Overall, the preliminary results from this correlation analysts suggest that firms with more accurate forecasts, less dispersion, and larger analyst following are more likely to have lower debt costs, which is consistent with the hypothesis that forecast characteristics contributes to the information environment, which in turn lowers the cost of capital. 3. Empirical Results 3.1 Bond Ratings and the Information Contained in Analysts Forecasts We first examine the relationship between the analyst activity and bond ratings. Credit ratings agencies are important information intermediaries in the corporate bond market, so it is possible that the information produced by ratings agencies may contain information about the properties of analysts forecasts. To test how the information contained in analysts forecasts is incorporated into credit ratings, we estimate the following model: Rating i,t = α I + β i (Information) + δ i (Firm Specific) + γ i (Industry) + λ i (Time) + ε I (3) where, β I is a vector of factor sensitivities to the information environment variables; δ I is a vector of factor sensitivities to firm specific risk variables; γ I is a vector of factor sensitivities to industry specific risk, and; λ I is a vector of factor sensitivities to time specific systematic risk. The results of the analysis are presented in Table 4. In the first three models, we examine the properties of analysts forecasts individually. Consistent with the hypothesis that the information contained in analysts forecasts influences credit ratings agencies, we find that all three of our information variables are significant and have the expected signs. Model 1 shows that after 10
controlling for firm and market risk, higher analyst following is associated with better credit ratings (coefficient =0.804, t-value=9.33). Model 2 shows that the median forecast accuracy is also positively related to credit ratings (coefficient=0.079, t-stat=6.89) while model 3 reports that forecast Dispersion is positively related to bond ratings (coefficient=-0.121, t-stat=6.00). In model 4, we include all three variables together (replacing Accuracy with AccuracyResidual) and again find all three variables of interest significant and correctly signed. Finally, in model 5 we examine the relation between our analyst information factor, InfoFactor, and credit ratings. Consistent with our previous findings and with the hypothesis that the information factor is capturing the information in analysts forecast, we find the information factor is significantly related to credit ratings (coefficient=0.468, t- stat=7.53). While imposing a linear factor structure on our variables does result in a loss of explanatory power, the magnitude is quite minimal (r-squared of 0.696 in model 4 versus 0.692 in model 5). Finally, we note that our previous OLS regressions assume uniform differences between the ordinal credit ratings. To verify that this assumption is not driving our results, we re-estimate our test using an ordered probit model. Model 6 of Table 4 reports that the coefficient on our test variable (InfoFactor) from the ordered probit estimation remains correctly signed and significant. Overall, the results from table 4 suggest that having more analysts making more accurate forecasts with less dispersion is associated with better credit ratings. Given that the information contained in our analyst variables appear to affect credit ratings, we create a rating variable which separates the unique aspect of a firm s bond rating from the information already incorporated in the other control variables (RatingResidual). To do this, we save the residuals for equation 3. This variable allows us to control for the unique information contained credit ratings when examining the cost of debt yet the variable is orthogonal to the analyst information variables as well as the other financial variables. This allows us to better measure the impact of the information contained in analysts forecasts while controlling for the information unique to credit ratings. We will, however, compare these results with those for the unorthogonalized ratings variable. 3.2 Credit Spreads and the Information Contained in Analysts Forecasts In this section, we examine the economic impact of the information contained in analysts forecasts by investigating the cross-sectional relation between our information proxies and the cost of debt financing, controlling for various risk measures. Our primary multivariate model is 11
Spread i,t = α i + β i (Information) + δ i (FirmRisk) + φ i (SecurityRisk) + γ i (Industry) + λ i (Time) + ε i (4) where, β i is a vector of factor sensitivities to information variables; δ i is a vector of factor sensitivities to firm specific risk variables; φ i is a vector of factor sensitivities to security specific risk variables; γ i is a vector of factor sensitivities to industry specific risk, and; λ i is a vector of factor sensitivities to time specific systematic risk. The results are presented in Table 5. All models include controls for risk as described in section 2. For the control variables, we find that the coefficient estimates have their expected sign and in general are significant. For debt related variables, the coefficients for rating residual, credit ratings, duration, age have their expected sign and all are significant. Duration is negatively related to the cost of debt. Credit rating is negative and significant, while age is positive and significant. For firm related variables, the coefficient estimate on firm size and firm leverage are both negative and significant. Firm risk is positively and significantly related to yield spread. Finally, the high yield dummy is positive and significant in the full sample indicating that risky debt is associated with higher debt costs. Overall, our models in Table 5 explain a significant portion in the variation of credit spread, as all models have adjusted r-squares of at least 61%. Model 1 of Table 5 shows that the coefficient on NAnalysts is negative and significant (-32.317, t-stat=6.29). Therefore, after controlling for known determinants of credit spreads, firms with higher analyst following enjoy a lower cost of debt. Model 2 reports that the coefficient on forecast Accuracy is negative and significant while Model 3 shows that the coefficient on forecast Dispersion is positive and significant. (-9.444, t-stat=18.67 and 8.626, t-stat=4.73, respectively). Taken together, these results suggest more analysts making more accurate forecasts with less dispersion are associated with a lower cost of debt capital. In model 4, we include the number of analysts, forecast dispersion and the residual of forecast accuracy together. We again find that all three are correctly signed and significant, which is consistent with each aspect of analyst activity being important in contributing an increased information environment of the firm, which in-turn results in a lower cost of (debt) capital. Model 5 of Table 5 replaces the individual proxies for the information contained in analysts forecasts with our information factor variable. We find that in increase in the information 12
factor lowers the cost of debt, all things equal. In order to measure the impact of using RatingResidual, that is, of purging the information on analyst activity from credit ratings, in model 6 we replace RatingResidual with raw credit ratings, Rating. We again find that analyst following, forecast accuracy and forecast dispersion are correctly signed and significant. As expected given the results of table 4, they are smaller in magnitude. For example, the coefficient on analyst following (NAnalysts) goes from -8.624 (t-stat=6.22) in model 4 to -8.169 (t-stat=5.90) in model 6. Therefore, while credit ratings contain some of the information contained in the properties of analysts forecasts, the properties of analyst s forecasts are important after controlling for the raw credit ratings as well. 14 Model 7 also employs the unadjusted credit ratings while examining the impact of the information factor. We again find that the coefficient on the information factor is correctly signed and significant. Overall, our results suggest that the information contained in analysts forecasts is an important factor in explaining the cost of debt, and that financial analysts, as information intermediaries, have an economic impact in the corporate bond market. 3.3 The Impact of Ownership Concentration and Compensation In this section, the test the sensitivity of our results to the inclusion of a number of additional control variables: (1) the stock ownership of the top management team in the firm; (2) the percentage of management s compensation that is equity based; and (3) the institutional ownership in the firm. We include management s stock ownership and equity based compensation as an additional control for agency conflicts. Therefore, we include a measure of the ownership of the top management team in the firm (InsideOwn) and the percentage of their compensation that is related to long-term compensation (Compensation). In addition to the ownership of the management team (insiders), we also include InstitOwn to represent the institutional ownership (outsiders) of the firm. Bhojraj and Sengupta (2003) find that concentrated ownership can exacerbate governance problems in the firm, which may be disproportionably borne by bondholders while Ashbaugh, Collins, and LaFond (2004) also that ownership concentration can affect firms credit ratings. Overall, the inclusion of these additional controls reduces the sample from 5827 firm-years to 2771 firm years. Table 6 presents the results. In model 1, we follow Table 4 and regress our variables of interest 14 As expected, the magnitude of our various financial control variables are also smaller when raw credit ratings are used as credit rating agencies also incorporate firm-level financial variables when assessing credit worthiness. 13
on bond ratings. We find the coefficients on the ownership by the management team and the ownership by institutions are negative and significant. This implies that blockholders have a detrimental effect on credit ratings, and is consistent with the findings of Bhojraj and Sengupta (2003) and Ashbaugh et al. (2004). Further, we find that higher the equity-based proportion of total compensation to the firm s top five executives is associated with lower credit ratings. This supports the findings of Core et al. (1999) who find that CEOs at firms with greater agency problems receive greater compensation. However, after controlling for these effects, we continue to find that our three analyst variables (NAnalysts, Dispersion, AccuracyResidual) are significant and correctly signed. Model 2 of Table 6 reexamines the impact of our analyst variables on credit spreads with the additional controls. While we find that these additional explanatory variables and significantly related to credit spreads as expected, we continue to find that our analyst variables are correctly signed and significant. Further, models 3 and 4 repeat the analysis using the information factor in place of the three analyst proxies with similar results. In sum, we find that both ownership concentration and management compensation are important to the cost of debt, but they do not subsume the information contained in analysts forecasts. 3.4. Endogeneity We examine several alternative specifications of our tests in an effort to mitigate any potential endogeneity concerns. To do this, we estimate firm fixed effect and two-stage least-squares models for the basic specification in the prior sections. The firm-fixed effect models helps us to understand how time-series variation in a firm s credit spreads is related to time-series variation in a firm s information and risk characteristics. The fixed effect models are reported Table 7 model 1, 3, 4, and 6. In all of the models, we find that Accuracy, Dispersion, NAnalysts, and InfoFactor all remain statistically significant. That is, variation of the analyst information proxies at the firm level is related to firm level changes in credit spreads, controlling for other factors. An added benefit of this methodology is that it provides a robustness test of our serial correlation corrected standard errors in our main pooled cross-sectional time-series models. We also employ a two stage least square instrumental variable (2SLS-IV) regression approach. In this test, we attempt to control for the possibility that analyst following is influenced by firm characteristics that are related to credit spreads that our previous models do not capture. To do this, 14
we first model the number of analysts following a firm. The first stage of the two-stage least squares estimates is: NAnalysts = α + β 1 Size + β 2 FirmRisk + β 3 Profitability + β 4 Correlation + β 5 Segments + ε (5) The variables are as described in section 3, but with the addition of two variables. Correlation, suggested by Lang and Lundholm (1996), represents the historical relation between stock returns and earnings, and Segments is the natural log of the number of operating segments the firm reports. Of significant concern is the identification of a variable related to accuracy and not related to credit spreads, we find the correlation to be unrelated to credit spreads but significantly related to the number of analysts. In the first stage, we find that all the variables are significantly related to NAnalysts. The second stage of these models are reported in Table 7 model 2 and 5. Using this approach, we find that NAnalysts, Dispersion, and Accuracy remain statistically significant in all models. 3.5. Additional Robustness Tests Though not reported we also perform a number of additional tests to examine the robustness of our results. To test the sensitivity of the results to outliers, we estimate the primary model using median regressions. As a second method to control for potential serial correlation in our data, we use the Fama-MacBeth (1973) procedure. To do this, we repeat the primary regression tests using a year by year regression and then obtain a t-statistics by averaging the coefficient for the 19 years and dividing by their standard deviation (adjusted for number of observation). Finally, we estimate the model using generalized least squares estimated with autocorrelation within panels and crosssectional correlation and/or heteroskedasticity across panels. We find similar results using any of these approaches. 3.6 Private Information and the Impact of Analyst Activity Our previous tests establish a link between the information contained analysts forecasts and the average cost of (debt) capital. In this section, we test if the relative roles of private and public information in the firm aid in explaining the economic impact of analysts forecasts. Recent research by Barth et al. (2001) and Barron et al. (2002) find evidence that the level of intangible assets is related to the amount of private information in the firm. We use this to test Easley and O Hara s (2004) conjecture that analyst information is more valuable in firms with a greater degree of private 15
information. We start off by focusing on two measures of the level of intangible information, intangible assets relative to total assets and research and development relative to operating expense. Since measures of intangible assets are highly correlated with industry classification, we focus on how intra-industry variation in these measures is related to the role of analyst information. To do this, we start by restricting the sample to firms within industries with at least 150 firm-year observations over the sample period. From this group, we form our two samples, Intangible and R&D samples. Next, we restrict the samples to industries with significant portion of intangible related information. For the Intangible sample, we require the mean intangible assets (intangible assets to total assets) of firms within the industry to exceed 10 percent. Panels A and B of Table 8 report the sample statistics. For this sample, we are left with 10 of the Fama and French 30 Industry classifications and 2183 firmyear observations. For the R&D sample, the average R&D to operating expense ratio for the industry had to exceed 1 percent. This restriction leaves a R&D sample of 8 industries and 2034 firm-year observations. We note that any cuts produce ad hoc classifications, but our goal is to set a significant requirement while leaving a sufficient sample to provide valid statistical inferences. When comparing the samples, we note that the choice of intangible measure has a significant impact in the sample as the only overlapping industries are Healthcare, Medical Equipment, Pharmaceutical; Fabricated Products and Machinery; and Everything Else. Given the vague nature of the Everything Else category, we also estimate the models excluding this category and find similar results. Panel C of Table 8 reports the regression results. Model 1 shows that the coefficients for both the information factor and the interaction between the information factor and the level of industryadjusted intangible assets are negative and significant. Model 2 repeats the analysts using firm-level fixed effects and find finds similar results. The findings from models 1 and 2 suggest that while the information contained in analysts forecast is associated on average with a lower cost of capital, the impact of this information is greater in firms with more intangible assets. In model 3, we partition the analysis on industry adjusted research and development expenses. We find that the coefficients on the information factor as well as the interaction of the information factor with research and development are negative and significant. Model 4 also confirms these results using firm-level fixed effects. Overall, the results from Table 8 support the hypothesis that the information contained in 16
analysts forecasts lowers the cost of (debt) capital and that the economic impact of this information is greater for firms with relatively larger amounts of private information. Our findings are consistent with the prediction of Easley and O Hara (2004) who argue that information production by analysts will decrease the cost of capital for all firms, but that this information should have a larger impact in firms where the composition of private versus public information is greatest. Further, in light of the findings of Barth et al. (2001) and Barron et al. (2002) that analysts in firms with more intangible assets are more numerous, expend greater effort, and produced forecasts that contain relatively more private information, our results suggest that the economic impact of these activities is translated into a lower cost of capital. 4. Conclusion Previous research leaves unanswered several questions regarding the role of analysts in the in pricing of financial assets, and in particular, corporate bonds. In this paper, we examine the relation between financial analyst forecast characteristics and the cost of debt financing. We find that higher analyst coverage, greater forecast accuracy, and lower forecast dispersion are associated with better credit ratings. We also show that after controlling for credit ratings, analyst coverage, forecast accuracy, and forecast precision are negatively related to the required return on firms public corporate debt. Finally, we document that the economic impact of analyst activity is largest for firms with relatively larger amounts of private information. Overall, our results support the hypothesis that the information contained in analysts forecasts is related to firms cost of (debt) capital, and that the economic impact of analysts forecasts is related to the composition of firms private and public information. While information production by analysts, on average, is associated with a lower cost of capital, the cost of capital impact of analyst forecasts is larger in firms with relatively more private information. In this way, our findings also add to our understanding of analysts role as information intermediaries. 17
References Anderson, R., S. Mansi, and D. Reeb, 2003, Founding family ownership and the agency cost of debt, Journal of Financial Economics, Forthcoming. Ashbuagh, H., D. Collins, and R. LaFond, 2004, Effects of corporate governance on firm credit ratings, Working Paper, University of Wisconsin-Madison, 1-53. Barth M. and A. Hutton, 2004, Analyst Earnings Forecast Revisions and the pricing of accruals, Working paper, Stanford University, Stanford, California, 1-59. Barth, M., R. Kasznik, and M. McNichols, 2001, Analyst coverage and intangible assets, Journal of Accounting Research 39, 1-334 Barron, O., D. Byard, O. Kim, 2002, Changes in analyst information around earnings announcements, Accounting Review 77, 821-846. Barry, and S. Brown, 1985, Differential information and security market equilibrium, Journal of Financial and Quantitative Analysis 20, 407-422. Bhojraj,S., and P. Sengupta, 2003, Effect of corporate governance on bond ratings and yields: The role of institutional investors and outside directors, Journal of Business 76, 455-475. Botosan, C., 1997, Disclosure level and the cost of equity capital, Accounting Review 72, 323-349. Botosan, C., and M. Plumlee, 2002, A reexamination of disclosure level and the expected cost of equity capital, Journal of Accounting Research, 40, 21-40. Bushman, R., Piotroski, J. and A. Smith, 2004, What determines Corporate Transparency?, Journal of Accounting Research, 42, 207-252. Ciccone, S., 2003, Analyst forecast properties, financial distress, and business risk, Working Paper, University of New Hampshire, Durham, 1-26. Coles, J., U. Lowenstein, and J. Suay, 1995, On equilibrium pricing under parameter uncertainty, Journal of Financial and Quantitative Analysis 30, 347-364. Core, J., R. Holthausen, and D. Larcker, 1999, Corporate governance, chief executive office, and firm performance, Journal of Financial Economics 51, 371-406. Diamond, D., and R. Verrecchia, 1991, Disclosure, liquidity, and the cost of capital, Journal of Finance 46, 1325-1359. Diether, K., C. Malloy, and A. Scherbina, 2002, Differences in opinion and the cross section of stock returns, Journal of Finance 57, 1130-1158. Easley, D. and M. O Hara, 2004, Information and the cost of capital, Journal of Finance 59, 1553-1583. 18