Bond rating changes from financial institutions:
What drives investors to react?
Aalze Hut
S2401932
26 June 2014
Supervisor: Prof.dr. L.J.R. Scholtens
Abstract
Using S&P credit rating changes for bonds issued by financial institutions, we use an event study to determine investors’ reactions to new bond rating information. We investigate several variables that might explain investors’ reactions by multiple regression on the Cumulative Average Adjusted Return (CAAR). Our results indicate that most market, firm and bond specific variables are helpful in explaining abnormal returns. We show that investors do react to bond rating changes, but that there are great differences between the results when we analyze at subgroup level consisting of: banks, insurance companies, financial companies, industrial companies and mutual & pension funds.
2
1. Introduction
The relationship between investors and credit ratings is under pressure since the reliability of credit ratings was not as reliable as was thought. Enron and WorldCom were able to maintain a triple-A rating shortly before default. After these scandals the Sarbanes-Oxley act was set in 2002 to improve corporate governance. Despite of this new law the next loss of reputation for Credit Rating Agencies (CRA’s) started in 2007. The US subprime mortgage crisis was partly due to the unreliable credit ratings issued by CRA’s who rated mortgage-backed securities with triple-A ratings. Two years later most of the mortgage-backed securities were downgraded to junk grade. Financial institutions still pay CRA’s to assess the quality of debt securities or structured financial instruments. Tirole (2011) even states that besides this conflict of interest, CRA’s advice on how to structure portfolios and pre-rating assessments all conspired to mislead the market. But if the CRA’s fail in their job to rate objective and accurate, questions raise about the impact of rating changes on investors’ behavior. What determinants make that investors decide to take action after a rating change?
First we wonder to what extent investors care about credit rating changes of bonds issued by financial institutions during our research period from 2005 to 2013. In part 1 we take a preparatory step by means of an event study where we compare investors’ reactions in different situations. We differentiate the direction of the rating changes (upgrades and downgrades), the ranking of the bonds (investment grade and junk grade), and the subgroup of the bonds (banks, insurance companies, financial companies, industrial companies and mutual & pension funds). To make sure that our results are robust we include a control event window and use multiple abnormal return measures. After we finished the whole event study in part 1 we continue in part 2 by developing meaningful multiple regression models that consists of market, firm and bond specific variables derived from existing literature. Because our regression models are based on the results of the event study we decided to divide this study in two parts. By again using the control event window and multiple Cumulative Average Adjusted Return (CAAR) methods we directly test whether our results are robust. We finally test if there is evidence for a changing investors’ reaction during our research period. Considering all results together we are able to answer our research question:
What drives investors to react to credit rating changes of bonds issued by financial institutions?
3 Hasan, and Tarazi (2013) recently did for Asian banks. But that does not clarify the extent of investors’ reactions. With our study we aim to identify which variables turned investors to react during 2005 to 2013.
This paper is organized as follows: In section two we present background information about bond event studies and multiple regression respectively. In third section we present part 1 of our paper. In part 1 we will introduce our event study methodology, data, hypothesis and results, respectively. In the fourth section we show part 2 that contains our multiple regression analysis consisting of a methodology, data, hypothesis and results section, respectively. In the last section we bring part 1 and 2 together and conclude.
2. Background
In this section we present an overview of the literature that is related to our research in a broad context. We shortly address some background and current issues with bond credit ratings to better understand the bond ratings’ right to exist. Thereafter we discuss the results of other bond event studies. Finally we present an overview of explanatory variables tested by other authors.
2.1 Bond Credit Ratings
4 the financial markets, the damaged reputation and the impact of rating changes attracting a lot of attention at the beginning of the twenty first century, what might have influence on investors’ reaction on bond rating changes.
2.2 Bond Event Studies
To explain what drives investors to react it is necessary that there is a price reaction that can be detected. Expected price reactions are easier to detect in event studies that analyze stocks instead of bonds due to the higher trading frequency of stock relative to bonds. Most papers on the effect of credit rating changes on stock prices conclude in line with common sense that updated ratings increase the stock prices and vice versa (see Hand, Holthausen, and Leftwich, 1992; Cantor and Packer, 1996; Hite and Warga, 1997; Kliger and Sarig, 2000). In other words, stock markets see rating information as new information, rely on this new information and adapt this ’efficiently’ in their prices (Fama, 1970). About the degree of efficiency in the bond market is more discussion.
5 Brister, Kennedy, and Lui (1994) were just attracted by this phenomenon and focused their research on junk bonds only. Now we have shortly discussed the study results about the direction of the rating, upgrading and downgrading, and the junk and investment grade, we refer back to the first sentence of this paragraph: Are there price reactions in bond event studies that can be detected? We think that the three points we discussed are inextricably linked to that question. Wansley and Clauretie (1985) found significant abnormal returns for downgraded bonds and stocks. The upgraded stocks were also significant, but the upgraded bonds were not. Katz (1974), and Ingram et al. (1983) showed significant abnormal bond returns after a rating change, while Weinstein (1977) did not. Hand et al. (1992) found in first instance also no significant bond returns, but when they excluded expected bond rating changes from their sample, the result were significant. The different outcomes indicate that in a bond rating event study the results of the research highly depend on the specifications of the input. Because most event studies take a unique set of observations it is difficult to compare different research results.
2.3 Bond regression studies
Important aspect of any type of multiple regression is the selection of the explanatory variables. In this paragraph we aim to give a brief overview of the independent variables that are most relevant to our research. The amount of existing literature about the drivers for investors to react to financial bond rating changes is limited. Hence we have to extend our scope to closely related researches too. For example papers that tried to explain bond yields or bond credit ratings. To keep the overview clear and structured we differentiate market, firm and bond specific variables.
In the first place there are macroeconomic market variables that can have influence on investors’ reactions to credit rating changes. The volatility of securities has a long tradition in finance (Cox and Ross, 1976). Bekeart and Wu (2000), and Collin-Dufresne, Goldstein, and Martin (2001) both carefully explain why the volatility should be relevant. However, the researches of Collin-Dufresne et al. (2001), and Corò, Dufour, and Varotto (2013) both show that the volatility is not significant. The government
bond yield variable in Corò et al. (2013) is doing much better since they find the variable to be highly
6 model. The results they find in the post-subprime period are significant for AAA and BBB rated bonds, but far from significant for AA and A rated. Grunert and Weber (2009) unsuccessfully include GDP
growth and unemployment rate in their model. An attempt of Collin-Dufresne et al. (2001) to include business climate as explanatory variable was also not successful.
In the second place we have the firms specific explanatory variables. It is Important to focus on independent variables that are related to financial institutions. We start with variables that have already proven themselves as explanatory variables. The leverage of the firm is such a variable (see Kliger and Sarig, 2000; Wringler and Watts, 1982; Pogue and Soldofsky, 1969; Corò et al., 2013; Pinches and Mingo, 1973; and Khieu, Mullineaux, and Yi, 2012). Despite of the overwhelming evidence are Collin-Dufresne et al (2001) showing that it is no golden rule to include a measure of leverage in the regression model. They thoroughly examine the effect of leverage on bond credit spreads with less convincing results. The total
assets are another frequently significant firm specific variable (see Kao and Wu, 1990; Khieu et al., 2012;
Belkaoui, 1980; Grunert and Weber, 2009; and Brister et al., 1994). Recently, Imbierowich and Rauch (2014) found no significance and Körs, Aktaș, and Doǧanay (2012) test besides total assets also some ratios created with total assets. They only find explanatory power in the long term liability to total assets
ratio. The debt ratio is according to Bhandari and Soldofsky (1979), and Koa and Wu (1990) worthy to
include. This time it is Belkaoui (1980) who disagrees with the explanatory power of the debt ratio. Other measures of solvability were successfully included (see Pinches and Mingo, 1973; Collin-Dufresne et al., 2001; and Körs et al. 2012).
Most models include a measure for earnings or cost-income ratio too. For example measures like
return on equity (Distinguin et al., 2013), earnings stability (Pinches and Mingo, 1973), or return on assets
7 change (see Distinguin et al., 2013; Corò et al., 2013; and Doumpos et al., 2014). Körs et al. (2012) include net sales without success in their model. Alp (2013) uses cash balances as independent variable for his data consisting of stocks. Although low cash balances do signal possible liquidity problems, are high cash balances a bit more complicated. If companies see no new good investment opportunities it can be a bad signal for investors to have much cash. Alp (2013) finds this variable to be significant at a ten percent level. Wringler and Watts (1982) earlier showed that the amount of cash of a firm was not able to explain why investors reacted to a bond rating change.
Besides the more general explanatory variables there are some papers who test variables that are especially interesting when analyzing financial institutions. For example the tier one capital ratio can have explanatory power for bank bonds, but is not available and relevant for industrial companies. The Tier one capital ratio represents the amount of high quality, non-repayable capital available to the bank (Higson, 1995). The Bank of International Settlements (BIS) changed the original Tier one capital requirement of Basel l to more restrictive requirements in Basel lll (Choudhry, 2012). The implementation of the new regulations for banks take time and the crisis is not contributing to the speed of this process. Despite of the changing definition of the Tier one capital is there evidence that investors care about this ratio. Distinguin et al. (2013) show in a univariate ordered logit regression that the Tier one capital ratio was significant at the one percent level for bonds issued by Asian banks. Distinguin et al. (2013) also tested with success the bank specific loan loss reserves to gross loans and Loan loss provision
to net interest revenue variables.
8 important for investors what is confirmed by Hand et al. (1992), Wansley et al. (1992), and Brister et al. (1994). Another bond specific variable is time to maturity. Just the time in years the bond is remaining active after the rating changes announcement. This measure is related to the probability of default by the simple rationale that a bond that is longer active has more time to default. Katz and Grier (1976) regressed the effects of bond rating changes for different maturity groups. Their conclusion was that the longer the maturity of a bond, the more precipitous will be its price reaction to adverse news. Kliger and Sarig (2000) use the duration instead of maturity. That also the size of the rating change can have impact on investors’ reaction to a bond rating changes is proven by Hand et al. (1992), and Wansley et al. (1992). Dummy variables for callability, call protection and secured by assets tested by Brister et al. (1994) did not result in significant results. Hand et al. (1992) were also unsuccessful with their credit
watch placement dummy.
There are many other possible explanatory variables that probably have explanatory power. We tried to keep this overview within bounds by skipping discussion of highly correlated variables. For a more comprehensive discussion of explanatory variables we refer to the financial bond study of Distinguin et al. (2013).
3. Event study
This section starts with a motivation of our event study methodology. Subsequently we will show the selected bonds and their descriptive statistics. After we have presented the event study hypotheses we developed to get clear what drives investors to react to bond rating changes, we will show the results.
3.1 Event study methodology
9 announcement of the credit rating change as event window. When the reclassified bonds are not traded on one of these days, they take the return on the first day the bond is traded again. Instead of largely choosing dates before the announcement date they took a range of 65 to 365 days after the event window as estimation window. They argue that an unusual post-rating change estimation window is better because there is evidence from previous studies that upgrades are preceded by positive average excess returns and vice versa. Combining the information it is clear that the ‘general optimal event period (event window and estimation window together)’ does not exist, or is not tractable due to all varieties in input variables. Although no authors write explicitly about a trial and error method to find their event and estimation window, most researchers will probably have done this some times. For determining the ‘optimal’ event window some short calculations of the adjusted returns of the days around the announcement day makes sense. Useful papers to get an indication of where to search are written by Goh and Ederington (1993), Hite and Warga (1997), and Katz and Grier (1976), who show their complete daily adjusted returns overview instead of arbitrarily presenting a selected event window. Based on the discussed event studies we choose to take an estimation window of 200 days. This estimation window is long enough to capture a good long term average return, but short enough to prevent too much bonds from falling out of the dataset due to unavailable prices. Generally the returns on event days in the event window are not included in the estimation window to prevent that the event influences the normal returns (MacKinlay, 1997). Because it is hard to derive a good fitting event window from the literature that exactly fits our research we choose to identify the price reactions to the rating change first. Therefore we need to determine the ‘abnormal returns’ first.
10 portfolio. The OLS Market method uses historical returns in a different way than the Mean method. An OLS regression of the historical bond returns on the bond market portfolio results in an estimated regression intercept (α) and slope coefficient (β) for each bond. The expected return for the bond is than determined by α plus β times the corresponding market portfolio return. Consequently, we determine the OLS Market method abnormal return by subtracting this expected return from the daily bond return. More complicated methods are introduced by for example Carhart (1997). He shows that the Fama- French three-factor model and his own variant, the Carhart four-factor model, are even better in detecting abnormal returns than the OLS Market method. Therefor he uses stock market capitalization information to calculate the small minus big (SMB) variable. Because we have no access to market capitalization information for our bonds we do not regard the methods described by Carhart (1997) as optional. The literature presents different reasons to prefer one of the methods described in Brown and Warner (1985) above another. Wansley et al. (1992) argue that the Mean method and OLS Market method require longer time series of returns to implement the model what would severely restrict the sample size when bonds are examined. Also the frequency of trading is often unknown what especially influences the sensitivity of the historical return using methods when the frequency is low. Brown and Warner (1985) themselves demonstrate that the power of the three models to reject the null hypothesis is almost the same in case of stock returns. Mackinlay (1995) adds to this discussion that the Mean method is probably the simplest method. The OLS Market method and the Market method both use a market portfolio what to some extent will lead to accounting for market-wide factors that are not included in the model. One weighty argument against the market portfolio using methods is that it requires a bond index that is closely related to the bonds in the sample (Hite and Warga, 1997). Finally, the OLS Market method is, as the name already implies, restricted to OLS assumptions that can lead to biased OLS estimates if the assumptions do not hold (Brown and Warner, 1985).
Considering the fact that we have not found articles making the comparison between three methods related to bond credit rating changes, and that all methods have their own strengths and weaknesses, we decide to measure the abnormal returns by all three methods proposed by Brown and Warner (1985). By doing this the risk of selecting a non-optimal method is eliminated and we directly have a robustness check for our results. We use the following generic definitions to calculate the abnormal returns for the Mean, Market, and OLS Market method, respectively:
11
(2)
( ) (3)
Where is the log return of bond i at day t during the event window, T is the number of days in the
estimation window, is the log return of bond i at day in the estimation window, and is the
return of the bond index that we use as market portfolio at day t. The and in equation 3 are the intercept and slope of bond i determined by an OLS regression of the bond return in time T on the index returns in time T. By cumulating all daily abnormal returns in the event window t, the Cumulative
Average Adjusted Return (CAARi) for bond i is determined:
∑ (4)
Where N stands for the number of summed event window days. The CAAR value depends on the method that is used to determine the abnormal return and is thus different for each abnormal return measurement method. We will use the CAAR of the bonds as dependent variables in our multiple regression analyses.
Having explained definitions of the AR and CAAR we can move on to the normal distribution test and subsequent event study methodology. Over the past decades many different statistical test have been developed to calculate the efficiency of the financial markets. Most tests give approximately the same outcomes, however, it is important to apply the test that fits best. The first step to find the correct test is to test whether the bond returns have a normal distribution, what we do with the test by Jarque and Bera (1980):
( ( ) ) (5)
Where n is the number of observations, S is the skewness, and K is the kurtosis of the distribution. The outcome of the Jarque-Bera test is asymptotically chi-squared distributed with two degrees of freedom. This means that the statistic can be used to test the null hypothesis of a normal distribution. H1 is in this case: There is no normal distribution of the log bond returns. After introducing
our bonds we will show that the bond returns are far from normally distributed. Hence, the null hypothesis can be rejected in favor of H1. Consequently, we have to use a non-parametric test to analyze
12 Corrado test mainly to identify what event window fits our data best. The first step of the Corrado test is to transform all bond time series into ranks. This implies that all downgraded abnormal bond returns are ranked from high to low, with the highest rank for the lowest abnormal return. And the other way around for all updated bonds that receive ranks from low to high, with the highest rank for the highest abnormal return.The ranking period includes the estimation and event window. We apply the Corrado (1989) test just as Hite and Warga (1997) to an event window of -12 to +12 surrounding the event. This 25 event window days together with our 200 days counting estimation window directly before the event window comes down to 225 (n) ranks with an average rank ( ̅ ) of 113. Let Rit stand for the rank of the
return on bond i on day t. Hence, we know that 225 ≥ Rit ≥ 1 and the excess rank of each bond is
determined by Rit – ̅ . Now the Corrado test statistic for event window day t can be calculated by:
∑ ( ̅) ( ) (6)
The standard deviation of the ranks (S(Ranks)) is calculated using the entire estimation and event window:
( ) √ ∑ ( ∑ ( ̅)) (7)
By ranking all excess bond returns the asymmetric original distribution is transformed into a uniform distribution across all excess bond returns (Corrado, 1989). One important restriction of the Corrado rank test is that it uses event windows of only one day. This is especially a great limitation when the event study uses bonds as observed security. Because of the low trading frequency of bonds, it probably takes longer to assimilate new information in prices. Nevertheless is the Corrado test a good fundament for solid event window determination and its results can be used as students t-statistics. The rank test of Cowan (1992) is a good example of a nonparametric test that solves the limitations of the Corrado test by enabling longer event windows. Following the same notations as the Corrado test the formula for the Cowan rank test is:
̅̅̅̅ ̅
[∑ ( ̅̅̅ ̅) ]
(8)
13 randomly determine an event window for the Cowan (1992) test, we will precede this event study method with a Corrado (1989) test.
When the event window is determined we will test the significance of the abnormal returns for three different subdivisions. We differentiate the direction of the rating changes (upgrades and downgrades), the ranking of the bonds (investment grade and junk grade), and the subgroup of the bonds (banks, insurance companies, financial companies, industrial companies and mutual & pension funds).
3.2 Event study data
14 In total 1763 different financial institutions issued 7304 bonds. All these straight bonds are issued in US dollars and are at least one year active between 2005 to 2013. For all bonds it is required that they already had a credit rating before the credit rating changed. The next requirement for the bonds is presence of a measure of creditworthiness. We obtained the Standard and Poors (S&P) historical bond ratings from Datastream for all 7304 bonds. In total 5347 bonds were rated by S&P, but more meaningful, only 2852 bonds had at least one rating change between the bond issue date and the end of 2013. Wansley et al. (1992) argues that in contrast with the stock markets there is no abnormal returns after a CreditWatch placement for bonds. They also argue that investors’ reactions to ‘real’ rating grade changes is independent from whether or not the bond received a CreditWatch placement. For that reason we removed within grade rating changes what means that there is no difference anymore between an upgrade from AA+ to AAA- and from AA to AAA.The choice for this generalization is supported by Hand et al. (1992) who argue that CreditWatch placements provide little evidence of abnormal bond returns associated with either indicated downgrades or indicated upgrades. Because most rating changes found place in the second half of our research period we decided to select for each bond only the oldest rating changes.
15
Table 1: Bond rating reclassification matrix.
For each bond i issued by a financial institution between January 2005 to December 2013 that met our data requirements and most importantly, had a S&P credit rating change, is presented in this matrix. The values below the x-diagonal represent the number of bonds that has been upgraded (329). All bonds above the x-diagonal are downgraded (699).
From/to AAA AA A BBB BB B CCC CC C D Total
AAA x 61 1 62 AA 3 x 102 19 124 A 33 x 269 302 BBB 1 136 x 123 6 266 BB 2 1 78 x 66 6 153 B 64 x 25 3 92 CCC 1 5 x 5 7 3 21 CC 4 x 3 7 C x 0 D 1 x 1 Total 4 96 240 366 188 78 35 8 7 6 1028
The most important results of table 1 is the difference between the number of upgraded bonds (329) and downgraded bonds (699). The behind this difference is probably the credit-crunch crisis that had a huge impact on the financial world. Further we see that the most rating changes are in the A and BBB rating class. We regard all bonds below the BBB grade as junk bonds just as Bongaerts et al. (2012) do.
Besides this bonds we also need a bond index to determine the abnormal return of the Market and OLS Market method. The bonds are from over the whole world, albeit all US Dollar denoted bonds, but a comprehensive bond index is not available. To get the best fitting index we run an OLS regression of four US dollar denoted bond indexes on the average log differenced return of all bonds included in this research. As extra check to assure we choose the best index we tested the correlation between the different indexes and the average bond returns. We show the results of this regression in table A.1 of appendix A. The conclusion of the R2 and the correlation is that, although the 3-5 year index is coming close, the 10+ year index from the Bank of America fits best. This is in line with our expectations since we use US long-term bond issued by financial institutions. Therefore we use the Bank of America Merrill Lynch US Financial Corp 10+ year index to calculate the abnormal returns and CAAR values for the Market and OLS Market method.
16 Corrado test results are presented in table B.1 in appendix B. Unfortunately are the Corrado test results even with this degree of detail not significant. Wansley et al. (1992) faced the same problem but solved it by cumulating the individual results. It is possible that bond prices barely react to new rating information or that the prices adapt the new rating information before or after the -12 to +12 days. However, it is also possible that the price reactions are only detectable if we cumulate the abnormal returns around the event announcement day like Wansley et al. (1992) show. Therefore we apply a more simplistic method to determine the event window. We calculate the CAAR values from equation 4 for the Mean, Market and OLS Market method. The first day of accumulation is again day -12 and the last day is still +12. If there is no price reaction we can expect for upgrades and downgrades an average CAAR around zero. In figure 1 we show the scatterplot of the CAAR’s with on the horizontal axis the event window days and on the vertical axis the percentage of price adjustment.
Figure 1: Scatterplot of the Mean, Market, and OLS Market CAAR.
This figure displays the scatterplot of the Mean, Market and OLS Market CAAR development with event day -12 as starting point. The horizontal axis shows the 24 event window days around the credit rating change announcement date. The vertical axis is the cumulative percentage of price adjustment.
Figure 1 shows little difference between the three CAAR measurement methods what is in line with the results of Brown and Warner (1985). Also the stronger impact of downgrades relative to upgrades corresponds to the results of Creigthon et. (2007), and Hite and Warga (1997). Most important at this stage is the horizontal axis that shows for downgraded bonds a downward sloping line from day -7 to day +1. There seems to be a weak upward trend from day -4 to day +1 for upgraded bonds. Taking the two trends together we select the event window of day -6 to +1 as our event study event window.
-3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12
17 Especially for the CAAR’s after the event day +1 we are uncertain whether including them is positive for our event window. To overcome this uncertainty we introduce a control event window from the literature. We select a control event window of day -5 to day +5 like Followill and Martell (1993), and Polonchek and Miller (1999) used.
Now we have determined estimation event window, a nonparametric multiday event study test can be applied if the data is not normally distributed. The for a normal distribution of the CAAR’s and their summary statistics are presented in table C.1, appendix C. The most important results from table C.1 are the high Jarque-Bera values that lead to normal distribution probabilities of zero for the abnormal returns. The consequence is that a non-parametric test should be used for the event study.
3.3 Event study hypotheses
We aim to assess what drives investors’ reactions to bond rating changes. As argued in the literature review is the analysis of bond price changes not always leading to similar results, due to slightly different input variables like event window and rating grade of the bonds. Although the event study is mainly a preparatory part to the multiple regression analysis, it can give some insight in price reactions. . In part 2 we will test possible drivers of investors’ reactions to bond rating changes. But the event study makes clear under what conditions investors react.
First of all does the literature suggest that investors are reacting stronger when the rating change concerns a downgrade instead of an upgrade (see Wansley and Clauretie, 1985; Creighton et al., 2007; and Dallocchio et al., 2007). In the second place there are Creighton et al. (2007), and Hite and Warga (1997) who show stronger investors’ reactions for junk bonds than investment grade bonds. And finally we can test the difference in investors’ reactions between the subgroups: banks, insurance companies, financial companies, industrial companies and mutual & pension funds. Hence, we formulate the following three null hypotheses:
Hypothesis 1: There is no significant difference in investors’ reaction between downgraded and upgraded bonds after a rating reclassifications.
Hypothesis 2: There is no significant difference in investors’ reaction between junk and investment grade bonds after a rating reclassifications.
18
3.4 Event study results
In table 2 we show the Cowan (1992) non-parametric event study results for our selected (-6, +1) and control event window(-5, +5).
Table 2: Cowan nonparametric results.
In this table we present the event study results calculated by the Cowan (1992) test. In panel A all downgraded and upgraded bonds are grouped together. In panel B we made a distinction between the bonds that are rated BBB or higher (investment grade), and all bonds below (Junk grade). Panel C shows the effect of separating the bond by their financial sector. All calculations are presented for the selected -6 to +1 and control -5 to +5 event windows. The percentages are the student-t statistic probabilities. The *,** and *** indicate statistical significance of the student’s t-statistic at the 10%, 5% and 1% level, respectively.
Panel A
(-6, +1) (-5, +5) Rating change direction Nr. of bonds significance significance
Downgrade 699 4.50% ** 3.97% ** Upgrade 329 29.39% 3.82% ** Panel B (-6, +1) (-5, +5) Rating change direction Rating class Nr. of bonds significance significance
Downgrade Investment grade 452 3.12% ** 2.94% **
Downgrade Junk grade 247 50.97% 47.26%
Upgrade Investment grade 254 85.76% 12.64%
Upgrade Junk grade 75 71.86% 31.48%
Panel C
(-6, +1) (-5, +5) Rating change direction Subgroup Nr. of bonds significance significance
Downgrade Bank 296 2.09% ** 17.27%
Downgrade Financial company 60 39.23% 23.70%
Downgrade Industrial company 212 51.46% 11.94%
Downgrade Insurance company 71 3.98% ** 5.01% *
Downgrade Mutual or Pension fund 50 28.78% 57.28%
Upgrade Bank 71 68.27% 60.38%
Upgrade Financial company 18 27.18% 3.31% **
Upgrade Industrial company 173 7.77% * 7.23% *
Upgrade Insurance company 44 71.10% 88.32%
Upgrade Mutual or Pension fund 17 3.58% ** 4.25% **
19 only significant group in panel B. Although Creighton et al. (2007), and Hite and Warga (1997) only used corporate bonds, our results are in contrast with their results. Based on the downgraded bonds in panel B we could reject the second hypothesis: ‘There is no significant difference in investors’ reaction between junk and investment grade bonds after a rating reclassifications’. However, the inconclusive results from the upgraded bonds prevent us to do that. In Panel C show the event study results of all subgroups. Three of the five calculation that are significant at a five percent level have a too small number of bonds to be reliable. The downgraded bank bonds are significant under the selected event window, but only the downgraded insurance bonds show significance robust over both event windows. The results in panel C show that there is sufficient evidence to reject the third hypothesis: ‘There is no significant difference in investors’ reaction between the subgroups after a rating reclassifications.’ For example the downgraded insurance company bonds are significant and the upgraded bank bonds are not. Further research of results in table 2 is interesting but unfortunately we must limit ourselves to the aim of this paper to search for drivers for investor reactions to financial bond rating changes. In part 2 of this paper we will only use the results from table 2 that a least have significant results at a ten percent confidence level for one of the two event windows. With this restrictive measure we aim to avoid making regressions that try to explaining price reactions that are not present.
4. Multiple regression
This section starts with the selection of the explanatory variables we derive from the theory. Subsequently we will show the availability and descriptive statistics of the explanatory variables. With the knowledge of the explanatory variable characteristics we continue with the multiple regression methodology. Then we can present our generic regression model and after the formulation of some hypotheses we finally move to the regression results.
4.1 Explanatory regression variables
20 correlation between the included and omitted variable. As we stressed in section two before, we are sometimes forced to substantiate the selection of our explanatory variables on closely related research instead of equal researches, due to lack of available literature. Taking this risks into account the search for the explanatory variables can start.
To explain abnormal bond returns we first take a look at the clusters of variables the existing literature distinguishes. Horrigan (1966) searched in a multivariate analysis for accounting data and financial ratios that could be used to assist in ‘long-term credit-administration decisions’. Four years later West (1970) came with an alternative approach to predict bond ratings. He criticized Horrigans’ methodology and lack of theoretical support for his hypothesis and variables selection. West (1970) referred to an article of Fisher (1959) to show that including non-financial ratios improves the predicting power of that model. Closer to the present the clustering is developed to more significant clusters than financial and non-financial ratios. Hwang (2013) includes twenty-four firm specific predictors and eleven macroeconomic predictors. Those firm specific predictors include four market variables, nineteen accounting variables and one composite industry variable. We choose to divide the possible explanatory variables in three clusters. Namely, market, firm and bond specific variables.
The first market specific variable is the volatility. In theory it is possible to subtract the future volatility of a firm by implying the volatility from the publicly traded options. Since not all bonds in our sample are issued by option trading financial institutions, we need to find a good substitute for this. Like Collin-Dufresne et al. (2001), the VIX index provided by the Chicago Board Options Exchange (CBOE VIX) is good alternative. This is because this time-series CBOE VIX is a key measure of market expectations of near-term volatility conveyed by S&P 500 stock index options prices. We subtracted the CBOE VIX from Datastream and use the notation (VIXi), where i denotes the VIX for bond i. A second market specific
variable is the 10-year US treasury rate. Collin-Dufresne et al. (2001), and Corò et al. (2013) both included successfully a ten year long measure of government interest rates. Since we only selected bonds issued in US dollars we use the daily time series of 10-year US treasury rates from Datastream. We denote this explanatory variable in the functional form (R10i).
21 variables from the literature but use papers with factor analyses such as Pinches and Mingo (1973) to avoid selection of variables with a high correlation.
The first firm specific explanatory variable we use is the logarithm of the total assets (TA logi) as
Pinches and Mingo (1973) transformed it for their factor analysis. Grunert and Weber (2009) successfully tested this TA logi at an one percent significance level. By taking the logarithm the skewed distribution of
the variables outliers are transformed to a more normal distribution. Our second firm variables is the logarithm of the turnover (T logi) as again Pinches and Mingo (1973), and Grunert and Weber (2009)
used it. A last firm specific variables for which we used a log transformation is the cash and cash equivalents variable (C logi) like Alp (2013) used. Although low cash balances do signal possible liquidity
problems, are high cash balances a bit more complicated. If companies see no new good investment opportunities it can be a bad signal for investors to have much cash. As measure for the solvency of the financial institutions we select the solvency ratio (Soli) like Collin-Dufresne et al. (2001), Körs et al.
(2012), Bhandari and Soldofsky (1979), and Koa and Wu (1990), who all included a measure of solvency successfully. In section two we discussed multiple ways of including an earnings measure in the regression model. We select the return on assets (ROAi) as earnings variable as for example Doumpos et
al. (2014) recently did.
The first five firms specific variables are available for almost all financial institution bond. Now we introduce four explanatory variables specifically for banks, and one specific for insurance companies. The first bank specific variable is the ratio between non-interest operating expenses and operating income. This type of cost-income ratio is often referred to as the cost to income ratio (C/Ii). Especially for
banks it is an important measure of the bank’s efficiency. The lower the cost to income ratio, the more efficient the firm is running. But this lower is better mentality can be dangerous if the reasons behind the ratio are not fully explored. This is probably an explanation why we cannot find empirical evidence for the cost to income ratio. However, the ratio is considered as a core income-ratio for banks (Choudhry, 2012). We think that the cost to income ratio is sufficient different from the ROA and that it at least should be tested if this variable is considered as credit rating change driver. The tier one core capital ratio (Tier 1i) is the second specific bank variable. Distinguin et al. (2013) showed that this variable was
highly significant for Asian banks. Especially for banks is net interest margin (NIMi) an interesting variable
to calculate to see whether they do well compared to competitors. We suggest that it is likely that investors do use this variable in their analysis too. In the same line of reasoning is the profit margin (PMi)
22 included by them based on the argument that it is an indicator for the bank’s asset quality. For all ten firm specific variables we used Orbis as data source.
Finally we selected three independent variables subtracted from the theory related specific to bonds. Of course we include the coupon rate (CRi) as an important bond risk measure like Perraudin and
Taylor (2004) did. All the bond specific coupon payments are fixed with constant frequencies and are reported as percentages. The next variable is a dummy variable and is used to represent a qualitative entity. Since we do not focus our research on investment and non-investment grade bonds, we try to capture this ‘boundary effect’ by a dummy variable. Investment grade (Invi) is a dummy variable that
takes the value of one if the bond rating moves out of the investment grade (downgrades) or into the investment grade (upgrades); and otherwise zero. This dummy variable is used by Hand et al. (1992), and Wansley et al. (1992). Remarkably, they both use exact the same definition of the dummy variable, but disagree over the expected sign of this variable for downgrades. We expect that Hand et al. (1992) were right and downgrades are expected to have negative influence on the prices of bonds and vice versa. The last variables is the maturity (Mati) tested by Katz and Grier (1976). The maturity is determined as time
the bond is remaining active after the rating change. For this three bond specific variables we used Datastream as data source. An overview of all explanatory variables we include in our research and their definitions is presented table D.1 in appendix D.
To give insight in the availability and the statistics of the independent variables for the 1028 bonds we show the descriptive statistics in table 3.
Table 3: Independent variable descriptive statistics.
The table presents summary statistics of the 1028 bonds that received rating change between 2005 and 2013. Independent variable Nr. of bonds available Mean Median Maximum Minimum Std. Dev.
VIX 1028 24.67 20.84 80.86 9.97 11.14
10-Year Treasury 1028 2.87 2.75 5.13 1.63 0.94
Coupon Rate 1028 6.59 6.40 13.00 2.00 1.73
Invest (dummy) 1028 0.20 0.00 1.00 0.00 0.40
Maturity 1028 11.72 7.67 89.22 0.30 11.25
Total assets log 766 7.70 7.69 9.48 2.08 0.98
Turnover log 773 6.74 6.85 8.19 1.15 0.90
Solvency 788 18.06 11.81 98.22 -94.92 17.65
Return On Assets 679 1.61 0.67 33.12 -54.91 7.30
Cash log 371 5.60 5.87 7.91 -0.08 1.19
LLR/GL 293 2.41 2.23 15.62 0.00 2.12
Net interest margin 334 1.84 1.70 66.93 -40.77 8.92
Profit margin 78 -6.99 6.45 26.03 -135.59 39.89
CI ratio 332 69.69 65.92 331.13 -199.07 38.65
23 In table 3 we see that for all 1028 bonds the market and bond specific variables are present. This means that we can use the first five explanatory variables for each regression model. The invest variable is a dummy and can therefore only take the values one or zero. One of the reason why the availability of the other variables is lower is the existence of bank and insurance specific variables. In table 4 we divided the bonds in subgroups and show the availability of the variables for downgraded and upgraded bonds in panel A and B, respectively. This insight in the variables is fundamental for the development of our regression models.
Table 4: Bond distribution to sector and independent variable availability.
Number of bonds that have the required variable information available separated by downgrades and upgrades. In the last column is the total amount of bonds that contain a certain variable presented. The maximal row shows the total amount of bonds that are labelled to a specific financial sector. We determine our minimum regression size to 40 and therefore are only the bold figures possible for regression analysis.
Panel A Downgrades Independent Variable Bank Insurance
Financial Company Industry Mutual & Pension Total VIX 296 71 60 212 50 689 10-Year Treasury 296 71 60 212 50 689 Coupon Rate 296 71 60 212 50 689 Invest 296 71 60 212 50 689 Maturity 296 71 60 212 50 689
Total assets log 269 44 44 123 43 523
Turnover log 266 44 38 144 35 527
Solvency 268 44 44 138 43 537
Return On Assets 270 0 41 118 42 471
Cash log 2 0 44 138 39 223
LLR/GL 228 0 0 0 0 228
Net interest margin 265 0 0 0 0 265
Profit margin 0 44 0 0 0 44
Cost to income ratio 263 0 0 0 0 263
Tier 1 capital ratio 197 0 0 0 0 197
Maximal 296 71 60 212 50 689
Panel B Upgrades Independent Variable Bank Insurance
Financial Company Industry Mutual & Pension Total VIX 71 44 18 173 17 323 10-Year Treasury 71 44 18 173 17 323 Coupon Rate 71 44 18 173 17 323 Invest 71 44 18 173 17 323 Maturity 71 44 18 173 17 323
24 Table 4: Continued.
Upgrades Independent Variable Bank Insurance
Financial Company Industry Mutual & Pension Total Turnover log 69 33 9 118 11 240 Solvency 69 33 12 116 14 244 Return On Assets 68 0 0 111 14 193 Cash log 0 0 12 116 14 142 LLR/GL 65 0 0 0 0 65
Net interest margin 69 0 0 0 0 69
Profit margin 0 33 0 0 0 33
Cost to income ratio 69 0 0 0 0 69
Tier 1 capital ratio 54 0 0 0 0 54
Maximal 71 44 18 173 17 323
For both panels in table 4 are the subgroups bank and industry the largest. Because we fix our minimum regression size to 40 bonds we can only regress at subgroup level if there are bold figures in the column. Independent of whether the event study results show reason to regress subgroups we can conclude from panel B that there are insufficient upgraded financial companies and mutual and pension funds.
Besides the six regression models we can estimate based on the event study results presented in table 2 we want to estimate some regressions with respect to the subprime crisis. Our bond sample and rating changes are from 2005 to 2013. According to Dick-Nielsen et al. (2012) is the subprime crisis having impact on the bond yield spread. We use the volatility VIX variable to estimate time periods with Quandt-Andrews’ breakpoint test. We shortly present the theory and methodology in the first part of appendix E. We determined two breakpoints: 20 July 2007 (breakpoint A) and 19 January 2012 (breakpoint B).
4.2 Multiple Regression methodology
25 In our short event study literature review in paragraph 2.2 we showed that the results of bond event studies are inconclusive in their outcomes. We think this is caused by slightly different inputs instead of inconsistencies. Since the aim of our paper is to identify the variables that can explain investors’ reactions to bond rating changes for the financial sector, especially the significant price reactions, the results of the event study are the fundament of the regression analysis. What type of regression analysis is most relevant seems to depend on characteristics like the aim and the variables of the research, but also on the period paper was written. Financial market research started long before computers with advanced software programs were invented. Where Pogue and Soldofsky (1969) really challenged themselves by including multiple dependent and independent variables in their research, are Körs, Aktaș, and Doǧanay (2012) including multiple discriminant analysis, ordered logit and ordered probit models in their paper with 27 independent variables. Körs et al. (2012) discuss in their literature different statistical models but state that artificial intelligence models are better able to predict S&P bond ratings than multivariate statistical models. Huang, Chen, Hsu, Chen, and Wu (2004) for example use artificial intelligence models like ‘support vector machines’ and ‘neural networks’ to analyze credit ratings. Although many statistical variants are used, we think that some form of multiple regression is what most bond price change explaining papers have in common. Breiman and Friedman (1997) discuss six statistical models related to predicting more than one dependent variable from the same set of explanatory variables. They argue that with high correlation between the dependent variables the regression results can exhibit superior prediction accuracy when the correct statistical method is used. However, one of their methods, the two block Partial Least Squares (PLS) method shows that separate multiple regression gives better results. Breiman and Friedman (1997) also test a curds and whey method. This suggests that when someone wants to predict a dependent variable for which highly correlated variables exist, the prediction accuracy of that dependent variable can be improved by using the correlated variables for multivariate regression. Another way to deal with more than one correlated dependent variables is by comparing the fits of different models. This means that multivariate multiple regression models are estimated as standard separated multiple regression models, and then afterwards, a comparison is made by a so called likelihood ratio test. Likelihood ratio tests express how many times more likely the data are under one model compared to the other. For this test the probability distribution developed by Wilks (1937), better known as the Wilks’ lambda distribution, can be used.
26 method in bond regression analysis (see Distinguin et al., 2013; Dick-Nielsen et al., 2012; and Collin-Dufresne et al., 2001). Multiple regression is based on standard OLS estimation techniques and it is important to assure that all underlying assumptions about the variables hold. Avoiding discussing all assumptions for each model again we choose to discuss the assumptions in general. When identified errors in the assumptions occur, the way the errors are solved is added. The first OLS assumption is that the expected value of the errors is equal to zero. This weak exogeneity assumption means that the independent variables are error-free, what is in our case an unrealistic assumption. Therefore we include a constant term in the models to satisfy this assumption. Secondly, the assumption of homoscedasticity must hold. The requirement is that the errors have a constant variance what makes them homoscedastic. Constant variance is for ratio variables more realistic than for accounting variables like total assets. To avoid heteroscedasticity in the regression models we transformed some variables a priori by taking the logarithm. Taking the logarithm of variables is an attempt to standardize extreme values and is only possible for positive values such as total assets. In Appendix A.1 we show what variables are transformed a priori. We test each regression for heteroscedasticity by the general test for heteroscedasticity from White (1980). If it appears that errors have no constant variance we will correct the regression result by using a heteroscedastic-consistent standard error estimates. This means that hypothesis testing becomes more conservative what makes it harder to get significant results. The third assumption is that the errors are uncorrelated with one another. The Durbin-Watson (1951) test is a test for first order autocorrelation. If this assumption fails to hold we have to select other variables. The forth assumption deals with the multicollinearity of the explanatory variables. High correlation between the independent variables or the errors creates inconsistent and biased parameter estimates. The simples way to solve this problem is the select other variables. The last assumption is about the normal distribution of the errors. From the possible solutions to this violation of normal distribution of the error terms we select the most passive one, namely, ignoring the violation. This rigorous measure is based on two considerations. In the first place is the violation of the normality assumption virtually inconsequential in case of sufficient large sample sizes. In the second place are the alternatives like taking the logarithms, include a dummy or using a whole other estimation technique not preferred.
27 the number of observations minus the number of estimated parameters. Further we show the R2 and the probability of the F-statistic for each model. The R-square (R2) is an important statistic in our research and measures the overall fit of the regression line, in the sense of measuring how close the points are to the estimated regression line in the scatterplot. R2 is computed as a fraction of the variance of the dependent variable explained by the regression. Since some of our regressions result in very low R2 we include the probabilities of the F-statistic too. These probabilities test the hypothesis that none of the variables actually explains anything.
Especially because it is useless to try to find explanatory variables that must explain far from significant abnormal returns, we will often refer to Cowan test results showed in table 2. To estimate robust regression estimates we will only use samples of more than 40 bonds. Some of the theoretically related independent variables are subgroup specific and therefore will the number of independent variables per regression model differ. Also the availability of the firm specific variables can influence the sample size drastically. It is of the utmost importance that our research results are reliable. Therefore we have to be critical and selective in creating models that can give insight in investors’ reaction to bond rating changes. To realize this in an clear way we add to each regression table a short motivation of the model. Here we only present the generic model including all independent variables:
⁄ ⁄ (9)
In equation 9 we show all fifteen possible explanatory variables together with a regression constant and error term i. The Mean, Market, and OLS Market CAAR’s are regressed on the explanatory variables individually for each model we test. As we show in table 3 are the market and bond specific variables available for all bonds. To simplify the model motivations in each regression table we will refer to these variables as the base model:
+ (10)
4.3 Multiple regression hypotheses
28 estimates. One of the reasons for us to use the Mean, Market , and OLS Market method was the argument of Brown and Warner (1985) that power to reject a null hypothesis was almost the same for all three methods. The lines in figure 1 suggests that their argument can be right. Because testing this in a hypothesis is too far from our research objective we will only appoint large differences between the methods.
Dick-Nielsen et al. (2012) show a difference in bond yield spreads before and during the subprime crisis. In appendix E we determined three time periods: before crisis (before period A: before 20 July 2007), early crisis (between period A and B: 20 July 2007- 19 January 2012) and late crisis (after period B: 19 January 2012). We test whether there is a difference over time a five percent significance level:
Hypothesis 4: There is no significant difference in investors’ reaction between period A and B.
The last and most important question is what explanatory variables drives investors to react to a bond rating change. Because we estimate six different models that all consist of three CAAR methods and two event window, a total of 36 regression models are estimated. Taking into account that we use in total 15 explanatory variables it makes no sense to assess each variable individually. Therefore we choose to present the percentage of significant regression for each explanatory variable in a table like Katz (1974) did. The implicit meaning of the results in that table is the answer to our research question: What drives investors to react to credit rating changes of bond issued by financial institutions?
4.4 Multiple regression results
With the results from the event study showed in table 2 we are able to develop six meaningful regression models that support our aim to assess what drives investors to react to bond rating changes. Then we will discuss the results of the regression estimates about the subprime crisis effects as a sensitivity analysis. Finally, we will present a table with an overview of the performance of the explanatory variables in the regression models.
29 ten percent level and having a sample size of at least 40 bonds. The first model we estimate concerns all downgraded bonds depicted in table 5 below.
Table 5: Multiple regression estimates for all downgraded bonds.
This table shows the regression estimates for all downgraded bonds that have available the variables TA logi, T logi and Soli besides the base model variables. Table 4, panel A shows that including other independent variables results in a sample that does not include all financial sectors. The event study results in table 2 for all downgraded bonds are 4.50% and 3.97% for the event windows (-6 +1) and (-5 +5), respectively. The model we estimate to explain is:
.
Mean CAAR Market CAAR OLS Market CAAR
(-6, +1) (-5, +5) (-6, +1) (-5, +5) (-6, +1) (-5, +5)
Coeff. Prob. Coeff. Prob. Coeff. Prob. Coeff. Prob. Coeff. Prob. Coeff. Prob. VIXi -0.140 0.006●●● -0.144 0.001●●● 0.053 0.002●●● 0.058 0.011●● 0.052 0.001●●● -0.008 0.034●●● Ri10 -0.859 0.326 0.055 0.923 -0.312 0.037●● -0.040 0.847 -0.195 0.251 -0.012 0.783 CRi 0.135 0.614 0.007 0.977 0.083 0.203 0.106 0.189 0.003 0.952 0.001 0.976 Invi 0.473 0.708 -1.105 0.314 0.650 0.057● 1.241 0.005●●● 0.875 0.029● -0.014 0.888 Mati -0.052 0.702 -0.010 0.915 -0.006 0.700 0.007 0.737 0.029 0.161 0.002 0.802 TA logi 1.401 0.210 1.564 0.169 -0.104 0.573 -0.538 0.012●● -0.269 0.084● 0.105 0.299 T logi -2.170 0.025●● -2.111 0.052● -0.086 0.589 0.207 0.293 -0.115 0.436 -0.183 0.064● Soli 0.108 0.111 0.092 0.180 0.005 0.605 -0.013 0.242 -0.003 0.706 0.007 0.233 Constant 5.032 0.460 2.610 0.655 0.198 0.903 0.414 0.837 1.621 0.283 0.365 0.454 R2 0.053 0.052 0.045 0.063 0.076 0.041 F-Prob. 0.001 0.001 0.004 0.000 0.000 0.008 N 499 499 499 499 499 499
The ●,●● and ●●● indicate statistical significance at the 10%, 5% and 1% level, respectively.
The results in table 5 show that the explanatory power of most variables is low. The volatility in the market measured by VIX is five times highly significant at the one percent level. However, the expected signs of the VIX coefficient are contradictory. Possibly are the constant or the other variables contaminating the regression estimate. For us this is the deplorable reason to value the expected sign only marginal. Probably as a result of the high VIX are the F-statistics showing that there is no reason to assume that the model has no explanatory power at all. The low average R2 of 5.5% suggests that the variables cannot explain investors’ reaction. Both event study result for downgraded bonds are significant at the five percent level. In the next model this is not the case. In appendix F, table F.1 we show the regression model for all upgraded bonds.
30 Although the difference is marginal, is the selected event window (-6 +1) showing higher R2’s than the control event window (-5 +5). We use the same variables as in table 5, but with an average R2 of 8.9% are the upgraded bond CAAR’s better explained.
Next, we analyze the regression estimates at subgroup level as a sensitivity analysis to the results in tables 5 and F.1. The first subgroup we analyze are the downgraded bank bonds presented in appendix F, table F.2. For bank we have thirteen explanatory variables available. The average R2 is with 23.1%
more promising than earlier regression results. Especially the 10-year US government treasury rate has a large share in high R2. In table F.2 is the Market CAAR model showing much higher R2 values than the other two methods. That more explanatory variables in the regression model is not necessarily better is proven by the downgraded insurance bonds.
In appendix F, table F.3 are the first eye-catching results the high R2 values. With an average R2 of 54.9% is this results higher than we expected. The F-statistic is for all six regressions very close to zero and especially the insurance specific profit margin variable is highly significant in three cases. But again is the sign of the coefficients contradictory. Further is the Mean CAAR outperforming the other two methods here. From the analysis of downgraded bonds we move to the only upgraded subgroup we can regress.
In appendix F, table F.4 we face for the first time a serious concern with two F-statistics. This means the there is a risk that the model has no explanatory power at all. The strange result is that the OLS Market CAAR is this time showing the highest R2. The OLS Market CAAR for the selected event window shows that eight out of the ten variables show at least significance at the ten percent level. For our last multiple regression model we leave the subgroups and move towards the investment grade bonds.
In appendix F, table F.5 we show the regression estimates of the downgraded investment grade bonds as a sensitivity analysis to the first two regression models. The solvability variables if for all six regression at least significant at the five percent level. This evidence shows that in this scenario investors are partially driven by the solvability of the firm. However, again is the sign of the solvability coefficient contradictory.
31 bonds to, presented in table 6. The breakpoint analysis determined in appendix E divides the bonds based on the announcement day of the rating change. Table 6 shows an overview of the key figures we derived from the regression models in appendix E.
Table 6: Breakpoint multiple regression estimates results for all bonds.
In this table we show the key figures of the OLS regressions we did on the Mean, Market, and OLS Market CAAR’s for both event windows. In the first block the upgraded bonds that are reclassified before and after period A (20 July 2007) are separated. In the same way show the second and third blocks the upgraded and downgraded bonds separated by the second breakpoint B (19 January 2012). Since we are very interested in the pre-crisis regression results we managed to keep enough upgraded bonds in the sample at the cost of dropping all company specific variables. Then the model becomes for all regressions:
. To compare the results we presented the R2 and F-statistic averages in the last column. In the appendix we show the complete regression results.
Panel A
Upgraded bonds Number Mean CAAR Market CAAR OLS Market CAAR
Period of Bonds Measure (-6, +1) (-5, +5) (-6, +1) (-5, +5) (-6, +1) (-5, +5) Average
<A 49 R2 0.143 0.250 0.403 0.433 0.291 0.086 0.268
<A 49 F-Prob. 0.233 0.025 0.000 0.000 0.009 0.549 0.136
>A 278 R2 0.078 0.080 0.129 0.034 0.106 0.105 0.089
>A 278 F-Prob. 0.001 0.000 0.000 0.094 0.000 0.000 0.016
Panel B
Upgraded bonds Number Mean CAAR Market CAAR OLS Market CAAR
Period of Bonds Measure (-6, +1) (-5, +5) (-6, +1) (-5, +5) (-6, +1) (-5, +5) Average
<B 200 R2 0.068 0.073 0.068 0.033 0.067 0.056 0.061
<B 200 F-Prob. 0.000 0.000 0.000 0.006 0.000 0.000 0.001
>B 122 R2 0.099 0.087 0.100 0.024 0.037 0.047 0.066
>B 122 F-Prob. 0.003 0.007 0.002 0.515 0.245 0.135 0.151
Panel C
Downgraded bonds Number Mean CAAR Market CAAR OLS Market CAAR
Period of Bonds Measure (-6, +1) (-5, +5) (-6, +1) (-5, +5) (-6, +1) (-5, +5) Average
<B 491 R2 0.034 0.045 0.092 0.038 0.101 0.114 0.071
<B 491 F-Prob. 0.234 0.108 0.002 0.179 0.001 0.000 0.087
>B 180 R2 0.054 0.059 0.174 0.050 0.053 0.054 0.074
>B 180 F-Prob. 0.255 0.209 0.000 0.298 0.269 0.259 0.215
32 of means calculation we find this difference to be significant with a probability of 0.6%. Applying this calculation to the regressions in panel B and C this results in 37.8% and 44.8% for upgraded and downgraded bonds, respectively. Our results suggest that it is likely that investors’ reaction to bond rating has changed since the crisis started like Dick-Nielsen et al. (2012). Therefore we reject our fourth hypothesis: ‘There is no significant difference in investors’ reaction between period A and B. Whether this results are robust over all financial bonds needs more focused research.
The event window and CAAR measurement method are worth a brief review. We spend a lot attention to the determination of the event window. With an average R2 of 21.0% for the selected event window (-6 +1) compared to an average R2 of 17.6% for the control event window (-5 +5), are the differences marginal but present. The average R2’s for the three different CAAR measures are 18.9% for the Mean, 20.2% for the Market, and 18.7% for the OLS Market method. This is according to what we observe in figure 1 and conclusion of Brown and Warner (1985).
Finally, in table 7 we show our implicit answer to the question what drives investors to react to bond rating changes for financial institutions.
Table 7: Significance of explanatory variables.
In this table we show the performance of the explanatory variables of the six event study based regression models. In the last column we sum the three columns in the middle and divide by the number of times regressed.
Explanatory variable Number of times regressed Times significant at the 10 % level Times significant at the 5% level Times significant at the 1% level Percentage of significant regressions VIX 36 3 4 9 44.4% 10-Year Treasury 36 4 5 4 36.1% Coupon Rate 36 2 3 2 19.4% Invest 30 5 3 3 36.7% Maturity 36 6 4 27.8%
Total assets log 36 4 5 2 30.6%
Turnover log 36 6 3 1 27.8%
Solvency 36 1 7 5 36.1%
Return On Assets 12 1 8.3%
Cash log 6 1 16.7%
LLR/GL 6 0.0%
Net interest margin 6 2 1 1 66.7%
Profit margin 6 1 1 3 83.3%
Cost to income ratio 6 1 1 33.3%
Tier 1 capital ratio 6 1 1 33.3%
Average 32.6%
