Determinants of Bond Credit Spread
An Empirical Analysis of Credit Spread Movements of Jumbo Covered
Bonds in The Netherlands for 2005 -2009
University of Groningen
Faculty of Economics and Business
Master Thesis
MSc Business Administration, Finance
Specialisation: Corporate Financial Management
Author: Robin Zonne
Student number: 1385771
Determinants of Bond Credit Spread
An Empirical Analysis of Credit Spread Movements of Jumbo Covered
Bonds in The Netherlands for 2005 -2009
Robin Zonne*
ABSTRACT
This study tries to explain the movements of credit spread of Jumbo covered bonds
in the Netherlands. Credit spread movements of Jumbo covered bonds are partially
explained by the slope of the yield curve -accounts for 3.4%- and changes in
liquidity of the bond -accounts for 2.5%- in the secondary market measured by the
proportional bid-ask spread. Remarkable result is that the economic climate, for
which the return and volatility of the stock market index function as a proxy, does
not influence changes in credit spread. The main implication of this study is that the
credit spread of covered bonds is not determined by probability of default or
changing recovery rates. This is in contrast with corporate bond credit spread
dynamics.
Keywords: determinants, credit spread, jumbo covered bonds, liquidity, financial
crisis
Table of Contents
1. INTRODUCTION ... 4
2. BACKGROUND AND LITERATURE REVIEW ... 9
2.1 Regulation in The Netherlands ... 9
2.1.1 The Issuer of Covered Bonds: Banks ... 10
2.1.2 The Cover Pool: The (value of) Residential Mortgages ... 11
2.1.3 Asset – Liability Liquidity Management ... 13
2.1.4 The Cover Pool Monitor and Banking Supervision... 14
2.2 Literature Review... 14
2.2.1 Economic Climate: Stock Market Return and Volatility ... 15
2.2.2 Interest Rate and Term Structure of Interest Rates ... 16
2.2.3 Leverage of the Issuer ... 17
2.2.4 Liquidity of the Covered Bond in the Secondary Market ... 17
2.2.5 Credit Spread Puzzle ... 18
3. DATA... 19
3.1 Dependent variable: Covered Bond Credit Spread ... 19
3.2 Explanatory Variables ... 22
3.2.1 The Economic Climate ... 23
3.2.2 The Interest Spot Rate and Slope of the Yield Curve ... 24
3.2.3 Leverage of the Issuer ... 25
3.2.4 Liquidity of the Covered Bond in the Secondary Market ... 25
3.2.5 US Treasury Rate... 26
3.2.6 Credit Spread Puzzle ... 26
4. METHODOLOGY ... 30
5.RESULTS ... 33
5.1 Correlation ... 33
5.2 Separate Bond Regressions ... 34
5.3 Separate Issuer and Average Credit Spread Bond Regressions ... 35
5.4 Regression Results from Periods with Structural Breaks ... 37
5.5 Robustness Check and Empirical Results ... 38
6.CONCLUSION ... 42
REFERENCES ... 45
1. INTRODUCTION
The financial crisis started mid-2007 and caused turmoil in the financial markets. Large asset write-downs were of big concern to banks. Credit became scarce, trust between banks and confidence of customers in banks reached an all time low since the globalization of the financial market. Banks relied on their central banks for credit as banks were tight with giving each other credit. This was triggered by the fall of Lehman Brothers in September 2008 and the need for liquidity became one of the main concerns for banks. This largest financial crisis since the 1930‟s was first called the „U.S. subprime-mortgage crisis‟. Mortgages in the United States from the subprime category started to default at in increasing rate. Not only banks that originated these mortgages suffered big losses, but also investors who invested money in these mortgages through securitizations lost lots of money.
One of the products that caused the financial crisis is mortgage-backed securities (MBS). Many homeowners in the US whose mortgage obligations exceeded the value of their homes chose to not pay their mortgages, to default and „walked away‟ (Weiner, 2009). This caused great uncertainty about the value of assets the banks bundled and sold to investors all over the world in these securitizations. Banks that kept some of these assets and investors who hold many of these securitized loans in their portfolios suffered big write-downs because of depressed asset market prices. This phenomenon has been aggravated by „mark-to-market‟ accounting, which requires that assets be marked down to the price for which there exist market consensus. When there are more sellers than buyers, the markdowns can be enormous, sometimes not reflecting the assets‟ eventually recovery value. Such events effectively demolished the market for securitized mortgages backed bonds with an 83% drop in securitized debt issuance in the U.S. (Levine, 2008).
residential mortgage-backed securities (RMBS) (European Mortgage Federation, Merrill Lynch European Structured Finance Monthly Report, September 2007). Covered bonds have become one of the largest asset classes in the European bond market in a relatively short time. Covered bonds are next to RMBS an important source of finance for mortgage lending. Covered bonds are different from (true sale) RMBS in several aspects. The most important difference is that with covered bonds the collateralized assets stay on the balance sheet of the issuer, whereas the collateral with RMBS transactions is taken off-balance and is put in a special purpose vehicle. Another main difference is that the interest and principal payments with covered bonds come from the general cash flow of the issuer, whereas with a RMBS transaction the interest and principal payments are made from the cash flows generated by the underlying assets. This creates a double recourse for covered bondholders. If the covered bond issuer defaults, the interest and principal payments are generated from the collateral assets which are ring fenced in the covered bond program contracts. With these contracts covered bondholders have a preferential claim over other claim holders. Both RMBS and covered bonds are backed by a pool of assets, but only for covered bonds this pool is actively managed. See appendix A for a complete overview of the differences between covered bonds and RMBS. This cover pool is usually put together so that the covered bond program is rated triple-A by the different rating agencies such as Moody‟s, Fitch and Standard & Poor‟s. This means that covered bonds are a serious alternative for Treasury bonds to bond investors interested in the highly rated securities (Packer et al. 2007).
that is risk-free. From a contingent-claim or no-arbitrage point of view, credit spreads exists for two reasons. The first reason is to compensate the bondholder for the risk of default. The second reason is to compensate the bondholder for the fact that in the event of default, the bondholder only receives a portion of the promised payments. According to Collin-Dufresne et al. (2001) it is of interest to examine how the credit spread respond to proxies for both changes in the probability of future default and for changes in the recovery rate.
To reflect the total credit risk of a bond, the yield of a bond is quoted against the yield on a risk-free benchmark. The credit spread of covered bonds used to be stable and low, i.e. covered bonds were perceived as almost as safe as government bonds. Government bonds are usually referred to as risk-free bonds, because the government can for example raise taxes to repay the bond at maturity. However, some counter examples do exist where a government has defaulted on its domestic currency debt, such as Russia in 1998 (the "ruble crisis"), though this is very rare. Also, the current situation in Spain and Greece shows that governments are not free of default risk. Although they have not (yet) defaulted, market participants respond to the measures the ECB and other European countries take. The yield required by investors to loan funds to governments reflects inflation expectations and the likelihood that the debt will be repaid. The yield on Greece‟s government debt rose during to 12.5% (d.d. 5-7-2010) from the average of 4.9% over the past 10 years (www.tradingeconomics.com, 6-8-2010). This is indicative for the fact that a government bond cannot always be seen as a risk-free benchmark.
Covered bonds are interesting for investors for several reasons. Covered bonds have a relatively high credit quality. The major credit rating agencies (Moody‟s, Fitch and Standard & Poor‟s) have slightly different approaches to rate covered bonds, but they all focus on the structure of the cover pool and the quality of the underlying mortgages. They factor in the issuer‟s rating to a lesser extent, although Standard & Poor‟s recently revised their rating methodology regarding covered bonds where the credit rating of covered bonds is directly linked to the credit rating of the issuer. Fitch and Moody‟s are changing their methodology and apply a tighter linkage between covered bond assessment and the issuer credit rating. Currently most covered bonds still carry ratings of double-A or triple-A. Covered bonds can offer investors higher yields than European government bonds without significantly altering the risk profiles of their conservative portfolios, and it serves purposes of diversification as well. Also, covered bonds give protection against event risk, which means that bondholders have a preferential claim on the assets in the cover pool if the issuer defaults.
mortgage loans create high credit ratings for covered bonds. The high credit ratings can result in a lower spread and thus lower interest payments to bondholders, which reduce the cost of funding for the originator. Covered bonds also enjoy a second cost advantage: the market for covered bonds is far more liquid than the market for products like asset-backed securities in Europe. Jumbo covered bonds fall into liquidity category II whereas asset-backed securities fall into category V (www.ecb.int, 2010). Issuers have more pricing power in liquid markets, as a function of supply and demand. It serves the purpose of diversification as well, because the issuer‟s covered bond usually receive higher credit ratings than its senior unsecured debt. Therefore, covered bonds attract a different group of investors, which helps to broaden the issuer‟s funding sources and creates a wider funding mix.
Covered bonds continue to be an important and open source of funding for banks. At the peak of the crisis, covered bonds were the only open market based source of refinancing available to banks outside of those subsidized by governments or central banks. This is underlined by a 30% rise in new issuance in 2008 to €650 billion, leading to a 13% overall increase in covered bonds outstanding to €2.38 trillion worldwide (European Covered Bond Fact book, 2009). In September 2005, ABN AMRO Bank N.V. was the first to issue a Dutch covered bond program, since then four other issuers have followed their example. The Dutch banks that have set up a covered bond program to date have issued „registered covered bonds‟. Structured covered bonds are based on a general law, such as a contract law. All Dutch covered bonds outstanding follow the same issuance structure: they use a special legal entity as holder of the legal title of covered bond collateral. On 1 July 2008 official covered bond legislation came into force in The Netherlands. To date, 29 countries have special covered bond legislation or arranged structured covered bonds on contractual basis in a general-law based framework. The regulation regarding covered bond programs set out in the next chapter, makes covered bonds safer debt investments than other mortgage backed securities such as RMBS. Safer investment means lower risk and thus a lower net yield for the covered bond investors. However, the market for covered bonds grows fast, in The Netherlands the aggregate initial covered bond program size is € 157 bn. Around € 30 bn. is actually issued in less than 5 years, which account for more than 5% of national mortgage debt. However, the academic literature about covered bonds and the determinants of credit spread of covered bonds is very scarce.
credit spread under structural models of corporate default (Merton, 1973) which are tested in academic literature. I also include a liquidity variable because some portion of the credit spread comes from the differences in liquidity between the two securities upon which the credit spread is measured (Chen, et al., 2007). Elton et al. (2007) found a tax-premium in their research on U.S. corporate bond credit spreads but this is not relevant for the Dutch market. The Netherlands treat income from government bonds and corporate bonds with the same rate for Dutch residents and does not have a distinction between federal and state tax. Therefore, a tax related variable is not included in this research. The credit spread of corporate bonds is determined by different factors. For every factor I construct a proxy to quantitatively measure the underlying factor. However, the variables tested in the literature that are hypothesized to affect the credit spread only explain about 25% of the movements of the credit spread (Amato and Remolano, 2003). Do the variables that explain part of the credit spread of standard corporate bonds also explain the credit spread of covered bonds? The strict regulation of covered bonds cause the yield to be higher than government bonds but lower than normal corporate bonds, but what causes these yields to change? First I review the existing literature on the determinants of credit spread. I then try to answer these questions based on the empirical part of this thesis. In light of the „credit spread puzzle‟ and to get more insight into credit spread movements, I also test for other variables that are hypothesized to influence the credit spread.
2. BACKGROUND AND LITERATURE REVIEW
Covered bonds have a short history in the Dutch market, but the German equivalent Pfandbrief exists since 1769 when it was introduced by Frederick the Great in Prussia (Skarabot, 2002). However, it was only when the Mortgage Bank Law was passed in 1899 that the Pfandbrief took its present form (Jobst, 2007). Germany has by far the biggest Pfandbrief market in the world but their market share is decreasing rapidly. Denmark, Spain and France also have developed markets in respectively Realkreditobligationer, Cédulas Hipotecarias and Obligations Foncières. As mentioned in the introduction, RMBS and covered bonds differ on some essential features. RMBS differ (among other aspects) from covered bonds in the strict regulations and laws around covered bonds which are not applicable to RMBS. To get a better understanding of the covered bond concept, I set out the aspects of the Dutch structure and regulations regarding covered bonds.
2.1 Regulation in The Netherlands
2.1.1 The Issuer of Covered Bonds: Banks
The regulation stipulates that the issuer needs to be a bank (that is a credit institution as meant in article 4(1) (a) Banking Consolidation Directive) that is licensed by the Dutch Central Bank (De
Nederlandsche Bank N.V.; DNB). The covered bonds are guaranteed by the CBC owning the
cover assets, which create a dual recourse for the covered bondholders. The CBC is a special purpose vehicle set up as a bankruptcy-remote entity. It is a private company with limited liability (besloten vennootschap met beperkte aansprakelijkheid) wholly owned by a foundation (stichting), without employees but with independent directors provided by a corporate service provider. The CBC has a limited corporate objects clause, so that any other party dealing with the CBC will be able to see that it is dealing with a special purpose vehicle.
Figure 1: General Dutch Covered Bond Structure
Interest rate Swap Providers
Total Return Swap Provider Structured Swap Providers Originators Assetmonitor Servicer / Administrator
Covered Bond Issuer (Credit Institution)
(Originator)
Covered Bond Investors Guarantor (Covered Bond Company)
Swap Agreements Servicing /
Administration Agreements
Asset Monitor Agreements
Guarantee Support Agreement
Covered Bonds Covered Bond Proceeds Covered Bond Collateral transfer
Guarantee Support Agreement
Security Rights in Covered Bond Collateral
Principal and Interest Payments Agent for Covered
Bond Investors
Asset Backed Guarantee Covered Bond Collateral transfer
Trustee
Source: Merril Lynch, The Covered Bond Book, 2008
2.1.2 The Cover Pool: The (value of) Residential Mortgages
Covered bonds are backed by cover assets in a cover pool which is actively (dynamic) managed. All Dutch covered bond programs (ABN AMRO, Achmea, ING, NIBC and SNS) are backed by residential mortgage loans and allow for inclusion of non-Dutch residential mortgage loans with some restrictions. All programs include substitution assets to be added in the cover pool. These assets can be euro denominated cash or other assets eligible under the Capital Requirements Directive1 (CRD) to collateralize covered bonds.
1
Table 1: Four types of covered bonds displaying structural complexity
Structured Covered Bonds (general-law-based)
issued outside the new regulation, not registered with DNB,
not UCITS 22(4) compliant, not CRD compliant
Covered Bonds (special-law-based)
issued along with the new regulation, not registered with DNB,
not UCITS 22(4) compliant, not CRD compliant
Registered Covered Bond (special-law-based)
launched along with the new regulation, registered with DNB, UCITS 22(4) compliant, but not CRD compliant.
Registered Covered Bond (special-law-based)
launched along with the new regulation, registered with DNB, UCITS 22(4) compliant, and CRD compliant.
Self constructed: information from Merril Lynch, The Covered Bond Book (2008)
In practice, all cover pools consist of Dutch residential mortgage loans and, in the program of NIBC, partly German residential mortgage loans. In other countries public debt or ship loans are also eligible as cover assets. The Dutch regulation only lists the general requirements of article 22(4) of the Undertakings for Collective Investment in Transferable Securities Directive2 (85/11/EC; UCITS). See table 1 for an overview of the different covered bond structures in the Netherlands. This means that Dutch covered bond regulation introduces CRD-compliance as an option, and not as a requirement. In covered bond regulation across Europe this is a new feature. This gives issuers of and investors in Dutch covered bonds the flexibility to choose whether they wish to issue or invest in covered bonds which are either only UCITS-compliant or both UCITS- and CRD-compliant. UCITS- and CRD-compliance of Dutch covered bonds can only be achieved if the covered bonds are registered by DNB. The DNB register indicates whether the relevant covered bonds are CRD-compliant, and the registered covered bonds are all UCITS-compliant. The ABN AMRO and ING covered bond programs are both UCITS- and CRD-compliant. The Achmea, NIBC and SNS covered bond programs are designed to be both UCITS- and CRD-compliant in all respects but one: they apply a 125% instead of an 80% Loan-To-Value (LTV) cut-off percentage. The CRD prescribes that covered bonds may be backed by residential mortgage loans up to the lesser of (a) the principal amount of the relevant mortgage and (b) 80% of the value of the underlying mortgaged property. However, Dutch residential mortgage loans may in practice have an LTV ratio of up to 125% (ECBC, 2009). All Dutch covered bond
2
programs take a two-step approach towards the handling of LTV-ratio‟s of Dutch residential mortgage loans, as follows. The first step is to check eligibility, which means the loan is only eligible as cover asset if the principal amount does not exceed 125% (the eligibility percentage) of the value of the underlying property at origination. Once a loan is part of the cover assets, the maximum value attributed to it in valuing the cover assets is a certain percentage (the LTV cut-off percentage) of the value of the underlying mortgaged property at that time3. The LTV cut-off percentage applied to Dutch residential mortgage loans is 80% for CRD-compliant cover assets (i.e. ABN AMRO, SNS and ING), 125% for non CRD-compliant cover assets (i.e. Achmea and NIBC) and 100% for residential mortgage loans that have a Dutch National Mortgage Guarantee (Nationale Hypotheek Garantie, NHG).
The CRD or the regulation does not prescribe which value of the underlying mortgaged property should be taken into account when calculating the LTV-ratio: the foreclosure value or the market value. Under the Dutch covered bond programs the eligibility percentage is applied to the foreclosure value at origination and the LTV cut-off percentage is applied to the market value of the mortgaged property. The market value is in turn calculated as the foreclosure value at origination divided by 85-90%, corrected for indexation4. This method is under criticism lately and other methods of calculating a fair market value are gaining ground. One of these methods is the Automated Valuation Model which does not rely on indices, but analyzes the location and dwelling characteristics. Based on this data it searches for the best reference sales and determines a fair market value for the relevant property.
2.1.3 Asset – Liability Liquidity Management
Asset liability management is the practice of managing risks that arise due to mismatches between the assets and liabilities of the bank. The cash inflows from the underlying mortgages do not exactly match with the cash outflows to the bondholders. This mismatch is called liquidity risk and covered bond issuers enter into different swap agreements to manage this liquidity risk. Also, all programs require the issuer to establish a reserve fund equal to 1 month‟s interest payments of all covered bond series if the issuer‟s short term rating falls below a certain rating („trigger rating‟). To cope with the liquidity risk on principal payments all covered bond programs use either a pre-maturity test which is taken every 6 (S&P) or 12 (Fitch and Moody‟s)
3
For example, if the relevant LTV cut-off percentage is 80% and a residential mortgage loan has a principal amount of 110 and is backed by mortgaged property with a value of 100, then this loan would be valued at 80 in the asset cover test determining the value of the cover assets. The 30 excess value of the loan would serve as extra credit enhancement.
4
months preceding the maturity of the relevant covered bond series, or a one-year maturity extension. This extension means that the original maturity date is extended by one year and is only available for any final redemption amount payable. Also, for all Dutch covered bond programs a minimum level of overcollateralization is required. Overcollateralization is the process of posting more collateral than is needed to obtain a certain level of security. Overcollateralization is measured by applying an asset cover test with asset percentages ranging from 70% to 94% (this varies per issuer and per time period and is determined with/by the rating agencies). This translates into a minimum overcollateralization of 14 % to 43%.
2.1.4 The Cover Pool Monitor and Banking Supervision
All Dutch covered bond issuers are required to regularly (monthly or quarterly) send out investor reports that contain detailed information about the cover assets and the outcome of the monthly asset cover test. The accuracy of the asset cover test calculation is required to be tested at least annually by an independent auditor. Each year the CBC is required to produce audited financial statements. Once a Dutch covered bond program has been registered by DNB as mentioned above, the issuer will have ongoing administration and reporting obligations towards DNB. If the covered bonds no longer meet the requirements set by the regulation or if the issuer no longer complies with its ongoing administration and reporting obligations towards DNB, direct contact between the two is established. If it comes to sanctions, it may be that an issuance-stop is imposed on the issuer, which will be disclosed by DNB in its register. DNB is entitled to ultimately stop and remove the registration of a covered bond. The latter never happened thus far, however, The Achmea covered bond company send out a notice 5 September 2008 that it had made administrative errors on their asset cover test. As a result thereof they in fact did not pass the asset cover test in earlier periods. KPMG performed the asset cover test after this, with no further direct consequences for the Achmea Covered Bond Company.
2.2 Literature Review
few authors that analyze the covered bond market and its credit spread. Packer et al. (2007) and Frino et al. (2007) explicitly center their article on covered bonds. The main sources of specialized information on covered bonds are the European Covered Bond Council (ECBC) and the rating agencies, in particular Moody‟s, Standard & Poor‟s and Fitch.
Bujalance and Ferreira (2004) analyze the covered bond market and the credit spread of covered bonds. They analyze the covered bond markets of Germany, France and Spain and the dynamic relationships between their average bond yields. They come up with a valuation model which includes determinants of the changes in these yields. Empirical researchers so far are only able to explain less than half (about 25% in most cases) of the variation in credit spreads. This low percentage of explaining the changes in credit spread is referred to as the „credit spread puzzle‟ (Amato and Remolano, 2003). Other academic literature analyzes determinants of credit spread with an approach based on theory, such as the structural model of default. Covered bonds are in effect different from standard corporate bonds, as explained in the earlier. However, financial institutions are corporations as well and the probability of default does exist for banks, something Lehmann Brothers showed in 2008. For the reason mentioned above, I review the existing academic literature on the determinants of credit spread changes of corporate bonds.
Appendix B gives an overview of the most recent or important research on bond credit spread. Some authors approach their research from a more theoretical point of view and build a model for explaining credit spreads. For example, Demchuck and Gibson (2006) build a structural two-factor model of default and compare the performance of their model with the model of Merton (1974), Longstaff and Schwartz (1995) and Collin-Dufresne et al. (2001). Other authors take a more empirical approach and test the variables found in past literature on their own dataset. The following factors are found in the academic literature on the subject.
2.2.1 Economic Climate: Stock Market Return and Volatility
influence of the economic climate on credit spreads. They use monthly observations of the credit spread of 688 different US corporate bonds to discover significant credit spread determinants. They find significant relations between credit spreads and determinants that capture the interest rates, stock market index and the overall state of the economy. They find significant evidence for the fact that when economic uncertainty is higher (e.g. during recessions) credit spreads widen. Demchuck and Gibson (2006) also mention the importance of the economic climate on credit spreads. They use the stock market performance to proxy for the economic climate. They find evidence that negative stock market returns cause the credit spread to widen. Longstaff and Schwartz (1995) notice in their study the importance of equity returns on fluctuations in corporate bond credit spreads. Collin-Dufresne et al. (2001) explain that according to the contingent-claim model when volatility increases, the probability of default increases and the credit spreads tends to widen. Van Landschoot (2008) compares determinants of bond yield dynamics between Euro and US dollar corporate bonds. She finds that Euro and US dollar bond yield spreads are significantly negatively related to the stock market return. She also finds a significant positive effect of the squared implied volatility on both bond yield spreads, but this effect is much stronger for the US dollar yield spreads than for the Euro yield spreads.
2.2.2 Interest Rate and Term Structure of Interest Rates
decreases the likelihood of mortgage refinancing and the existence of prepayment risk5 and leads to a decrease in credit spreads. This theory is consisted with the theory on interest rates from Longstaff and Schwartz (1995) and also with Collin-Dufresne et al. (2001). They state that theory predicts that an increase in the slope of the Treasury yield curve results in a decrease in credit spreads.
2.2.3 Leverage of the Issuer
Leverage is the third factor considered in this thesis to be a determinant of credit spread changes. Frino et al. (2007) conclude their study by saying that an important factor of default of a corporate bond is the leverage of the firm. Since default is one of the components of credit risk, they suggest that credit spreads can be explained by the leverage ratio of the issuing firms. They calculate the leverage ratio by dividing the firm‟s liabilities by the sum of the liabilities and the market capitalization. They find evidence for an effect of leverage on variation in credit spreads of corporate bonds. Collin-Dufresne et al. (2001) also find in their study that leverage and credit spreads are positively related, i.e. an increase in leverage of the issuing firm widens the credit spread. However, financial companies have a different capital structure than non-financial companies, which gives a completely different (much larger) leverage ratio and are therefore often excluded from the research sample.
2.2.4 Liquidity of the Covered Bond in the Secondary Market
The fourth factor discussed is liquidity of the covered bond in the secondary market. Holding a portfolio that‟s consists of illiquid assets increases risk for investors since illiquid assets cannot be converted without costs or delay. Houweling et al. (2005) analyze the effect of liquidity risk on bond spreads. They find that the liquidity premium to compensate for liquidity risk explains a significant portion of the observed credit spreads. They find evidence for the fact that this liquidity risk depends on the size of the bond issuance, the volatility of the bond yield, and the age of the bond. The empirical results of the paper of Liao and Tsai (2005) show that the corporate internal liquidity plays an essential role in explaining credit spreads of corporate bonds. Additionally, they find a systematic internal liquidity risk factor which can capture market-wide bond credit spread changes to a large extent. Amihud and Mendelson (1991) also find in their research on fixed income securities that higher yields belong to less liquid securities suggesting that liquidity and credit spread have a negative relation. Chen et al. (2007) find in their study on corporate yield spreads and bond liquidity evidence for the fact that wider credit spreads belong
5
to less liquid corporate bonds. In their paper on the „credit spread puzzle‟, Amato and Remolano (2003) state that part of the credit spread exists because of a liquidity premium. They state that it is easier to undertake transactions in equity and Treasury markets than in bond markets; investors must be compensated for this with a higher yield. Koziol and Sauerbier (2007) value illiquidity of bonds and conclude their paper by saying that liquidity is an important factor in explaining credit spreads. They mention that liquidity of a bond measures the ease with which a security can be converted into money or other assets.
2.2.5 Credit Spread Puzzle
3. DATA
This section will describe the data I use for the empirical part of this research. The objective is to investigate how well the variables identified in the previous chapter explain observed changes in credit spread. The data that I use in this thesis is described in more detail below. The variables considered in this analysis are supposed to capture the components of default risk and a liquidity premium. Next to this, I test for other variables that possibly explain changes in the credit spread of covered bonds. First, the dependent variable will be discussed in more detail and then the explanatory variables are described.
3.1 Dependent variable: Covered Bond Credit Spread
In contrast to non-Jumbo covered bonds, Jumbo covered bonds have a minimum issuance volume of €1bn, a fixed coupon payable once a year in arrears and bullet redemption. Additionally and more important, Jumbo covered bonds are supported by the commitment of at least five market makers to quote continuously two-way prices during trading hours as long as the instrument is sufficiently liquid. The issuers of covered bonds in The Netherlands are Dutch banks: ABN AMRO Bank N.V., SNS Bank N.V., Achmea Hypotheekbank N.V., ING Bank N.V., and NIB Capital. However, only three banks thus far have issued Jumbo covered bonds: ABN AMRO Bank N.V., ING Bank N.V., and Achmea Hypotheekbank N.V.
and offered rates for the swap. The choice of this curve as the benchmark curve is mainly driven by the fact that the swap curve is increasingly used as a risk free benchmark by financial institutions (Van Landschoot, 2008). Also, Feldhütter and Lando (2007)state that the riskless rate is better proxied by the swap rate than the Treasury rate for all maturities. Another advantage is that the credit spread is calculated automatically over the mid-swap with the same maturity and is retrieved from a Bloomberg terminal. See figure 1 for a visual representation of the credit spread of the separate Jumbo covered bonds and the average credit spread. This average credit spread is calculated by summing up the credit spreads of every bond i at time t and divide it by the number of bonds at time t. As one can see, spreads were pretty stable and low until the financial crisis became visible. From September 2007 a steady increase of spread can be seen and it exploded around the time that Lehman Brothers fell (which is around September 2008). In the methodology sector I elaborate on this further.
Figure 1: Credit spread of weekly observations of Dutch Jumbo covered bonds
Before I describe the explanatory variables I test for a unit root in the dependent variable. Regression of a non-stationary variable over another non-stationary variable can lead to a so-called spurious regression. This means that a regression „looks good‟ under standard measures (significant coefficient estimates and a high R2) but is actually worthless. For this, the augmented-Dickey-Fuller test (Dickey and Fuller, 1979, ADF) and the Phillips-Perron test (Phillips and Perron, 1988) can be applied. I analyze the dependent variable credit spread (simple average of the credit spreads of all bonds) and test for a unit root with the use of an ADF test. MacKinnon (1996) gives a critical value of -3.46 at the 1% significance level and the test statistic of the credit spread level gives -1.54. The test statistic is not larger in absolute value than the critical value so there is a unit root and the variable is non-stationary. A common move to overcome this is to use the first differences of the variable. The first differences of a time series are the series of changes from one period to the next. I test again for a unit root but now with the first differences. This gives a test statistic of -4.03 with the critical value of -3.46, which means that the null-hypothesis of a unit root can be rejected and the data is stationary. Unit root test statistics of the separate bonds display the same results and are stationary only if one takes the first differences (not reported).
Figure 2 shows the first differences, ΔCSi,t = CSi,t – CSi,t-1 of the Jumbo covered bonds. These are
Figure 2: First Differences of Credit Spread of Jumbo Covered Bonds
3.2 Explanatory Variables
In this section I describe the explanatory variables I use for the empirical part of this thesis. From the literature review in chapter 2.2 follow the variables that possibly explain the level of credit spread. The variables are based on articles among others from Collin-Dufresne et al. (2001), Van Landschoot (2008), Demchuck and Gibson (2006), Frino et al. (2007) and Longstaff and Schwartz (1995). Possible determinants of credit spread are: interest rates and the term structure of interest rates, the secondary market liquidity of the bond, the economic climate, the company‟s leverage ratio, the Dutch house price index, and the US treasury rate. I will discuss the proposed factors in more detail.
-100 -75 -50 -25 0 25 50 75 100 125 16 -9 -2005 3 -3 -2006 18 -8 -2006 2 -2 -2007 20 -7 -2007 4 -1 -2008 20 -6 -2008 5 -12 -2008 22 -5 -2009 6 -11 -2009 b asi s p o in ts in p erce n ta ge o ve r m id -s w ap ING 4,25% ING 5,25% ING 4,75% ING 3%
3.2.1 The Economic Climate
Collin-Dufresne et al. (2001), Demchuck and Gibson (2006) and Van Landschoot (2008) include in their research the return of a stock market index to proxy for the business climate. To capture the economic climate I use the stock market index and the proxy I use is the return on the Dutch AEX index. The data is retrieved from Datastream. I take the natural logarithmic return, which means that the rate of return is continuously compounded, and calculate this as follows: natural logarithmic return AEXt = ln (AEXt / AEXt-1)
Consistent with the articles mentioned in the literature review, the hypothesis is:
H10: there is a no relationship between stock market return and credit spread of Jumbo
covered bonds.
H1a: there is a negative relationship between stock market return and credit spread of
Jumbo covered bonds.
According to Collin-Dufresne et al. (2001) can, with a constant probability of default, changes in credit spread can be explained by changes in the expected recovery rate. So the main logic behind the negative relationship is that changes in credit spread can occur due to changes in expected recovery rate, which in turn is a function of the overall business climate.
Collin-Dufresne et al. (2001) explain that according to the contingent-claim approach when volatility increases, the probability of default increases and the credit spreads tends to widen. Demchuck and Gibson (2006) and Van Landschoot (2008) also use volatility in their research. The AEX volatility index (VAEX) reflects the implicit volatility of option prices and gives an indication of the economic climate. Accordingly, I use the VAEX as a second variable to proxy for the economic climate. Market data from the VAEX is available from September 3rd 2007 and for the previous period the values are simulated. This data is retrieved from Datastream. Consistent with the articles mentioned in the literature review, the hypothesis is:
H20: there is no relationship between stock market volatility and the credit spread of
Jumbo covered bonds.
H2a: there is a positive relationship between stock market volatility and the credit spread
3.2.2 The Interest Spot Rate and Slope of the Yield Curve
Following Bujalance and Ferreira (2004), Van Landschoot (2008), and Collin-Dufresne et al. (2001), I include the interest spot rate as an explanatory variable and use the three-month Euribor as proxy. Consistent with the articles mentioned in the literature review, the hypothesis is:
H30: there is a no relationship between the spot rate and credit spread of Jumbo covered
bonds.
H3a: there is a negative relationship between the spot rate and credit spread of Jumbo
covered bonds.
The rationale behind the expected negative relationship is that an increase in the interest rate implies an increase in the expected growth rate of the firm‟s asset value, which in turn reduces the probability of default and thus lowers the yield spread [Van Landschoot, (2008), Collin-Dufresne et al., (2001)].
The term structure of interest rates refers to the relationship between bond yields with different maturities. When you plot all the interest rates against their maturities you get the yield curve. The slope of the yield curve is of interest because it gives insight into the relation between short and long term interest rates. The yield curve can be upward or downward sloping and this depends on the expectations of market participants about economic future developments. Normally the yield curve is upward sloping which indicates normal economic conditions. To use this variable in empirical research a proxy for the slope of the yield curve is constructed. This yield curve proxy is usually calculated as the difference between the long-term Treasury bond yield with a maturity of 10 or 20 years and a short term rate, like the three-month government bond or the 2-year benchmark yield [see for example Collin-Dufresne et al. (2001) and Joutz et al. (2001)]. In this thesis I follow the method used by Van Landschoot (2008) and construct the proxy for the slope of the yield curve as follows: slope of the yield curve = Dutch 10 year Government bond yieldt – Dutch 3 month Government bond yieldt
Consistent with the articles mentioned in the literature review, the hypothesis is:
H40: : there is no relationship between the slope of the yield curve and credit spread of
Jumbo covered bonds.
H4a: there is a negative relationship between the slope of the yield curve and credit
spread of Jumbo covered bonds.
10 year yield and the 3 month yield is low, means that the economic outlook is weaker. With a weaker economy the credit spreads tends to be wider, hence the hypothesized sign.
3.2.3 Leverage of the Issuer
Collin-Dufresne et al. (2001) find in their study that leverage and credit spreads are positively related, i.e. an increase in leverage of the issuing firm increases the risk of default and this widens the credit spread. The leverage ratios Collin-Dufresne et al. (2001) use are reported as quarterly data. To use this variable in their research, they linearly interpolate leverage ratios from quarterly data. A straight line between the quarterly data points gives estimates of the leverage ratios for the months in between the quarterly data points. In this thesis I use weekly observations of credit spread and linearly interpolating quarterly data into weekly observations could possibly lead to wrong results. The leverage ratio of financial institutions differs a lot from non-financial companies. Dutch financial companies have a tier-1 capital of around 10%, which gives a leverage ratio around 10, which is much higher than non-financial companies normally have. The issuers of covered bonds in my sample are financial institutes. Therefore, I do not include this variable in the empirical part of this thesis.
3.2.4 Liquidity of the Covered Bond in the Secondary Market
Following among others Van Landschoot (2008) and Chen et al. (2007), I use liquidity as an explanatory variable to explain credit spread movements of Jumbo covered bonds. The proxy I construct that covers the liquidity variable is the proportional bid-ask spread of the bond. I subtract the bid price from the ask price and then divide it by the average bid-ask spread. I use quotes from the market makers and retrieve the data from Bloomberg. I use these market maker quotes instead of real transaction prices because these quotes are publicly available for all dates and all bonds. I calculate the proxy for liquidity as follows: proportional bid ask spreadi,t = (ask
price bondi,t – bid price bondi,t ) / average bid ask spread bondi
A lower proportional bid - ask spread means a more liquid market. A more liquid market carries less liquidity risk and thus lower credit spreads. Therefore, consistent with the articles mentioned in the literature review, the hypothesis is:
H50: there is no relationship between liquidity and credit spreads of Dutch Jumbo
covered bonds.
H5a: there is a positive relationship between liquidity and credit spreads of Dutch Jumbo
3.2.5 US Treasury Rate
Collin-Dufresne et al. (2001) and Van Landschoot (2008) include the US Treasury rate in their research on yield spread dynamics. Van Landschoot (2008) did not expect to find that US Treasury rates affect euro yield spreads and found it highly surprising that the results indicate that US interest rates still dominate the corporate bond market. They both test if changes in the level of US Treasury rates (they use the ten year and 3 month US Treasury rate respectively) are related to changes in the credit spread of bonds. As a proxy I use the middle rate in the secondary market of three month US Treasury bills. This data is retrieved from Datastream. The hypothesis is:
H60: there is no relationship between the US Treasury rate and credit spread of Jumbo
covered bonds
H6a: there is a negative relationship between the US Treasury rate and credit spread of
Jumbo covered bonds.
3.2.6 Credit Spread Puzzle
The hypothesis is:
H70: there is no relationship between the house price index and the credit spread of
Jumbo covered bonds.
H7a: there is a negative relationship between the house price index and the credit spread
of Jumbo covered bonds.
The rationale behind the expected negative relationship is that a higher house price increases the house owners‟ equity and therefore lowers the probability of default. The lower default rate means that there is a lower risk associated with the underlying collateral of covered bonds and thus the credit spread should be lower. Table 2 summarizes the variables that will be tested for and their hypothesized sign, i.e. if the relation with credit spread is expected to be positive or negative.
Table 2 . Overview from explanatory variables
Explanatory variable Hypothesized sign (Ha)
Return stock market index -
Volatility stock market index +
Interest spot rate (3 month Euribor) -
Slope of the yield curve (10y – 3m yield) -
Liquidity of the bond +
US Treasury rate (3 month Treasury rate) -
House price index -
I also test for a unit root in the explanatory variables. As mentioned before, regression of a non-stationary variable over another non-non-stationary variable can lead to a so-called spurious regression. If the independent variables are non-stationary, a commonly used intervention is to take the first differences so that the data is stationary and OLS can be applied.
See table 3 for the unit root statistics of the level and first differences of the variables.
period decrease of the index level. Therefore, a positive sign should be interpreted as a positive development for the house price index. A negative second difference means that the current period‟s increase is smaller than the previous period‟s increase of the index level, or that the decrease is larger than previous period decrease of the index level. Similarly, a negative sign should be interpreted as a negative development of the house price index. So, the hypothesized sign from hypothesis H7a as is explained above does not change.
Table 3. (In)dependent variable(s) and unit root test statistics from ADF test
variable t-Statistic level t-Statistic first difference (second difference)
average credit spread -1.544 -4.028
p-value 0.509 0.002
natural log return AEX index -14.853 -9.200
p-value 0.000 0.000
volatility index (VAEX) -2.297 -14.502
p-value 0.174 0.000 3month Euribor -0.408 -5.649 p-value 0.904 0.000 liquidity -2.126 -20.824 p-value 0.235 0.000 US Treasury rate 0.251 -13.597 p-value 0.975 0.000
house price index -2.012 -2.210 (-8.287)
p-value 0.281 0.203 (0.000)
slope yield curve -11.524 -13.025
p-value 0.000 0.000
lag length automatic based on Schwartz Info Criteria (maxlag = 14)
critical value (MacKinnon, 1996)= -3.46 one sided p-level at 1% confidence level
The results of the tests for a unit root in liquidity for each separate bond show the same results (not reported). Except for the liquidity of the Achmea 4.375% bond and the Achmea 4.75% bond which look stationary at level, all other bond liquidity proxies are stationary only when using the first difference.
of -4.45 basispoints. Recal that the credit spread is not measured over the risk-free government bond rate but over the European mid-swap.
Table 4. Descriptive statistics of (in)dependent variable(s) at level
(In)dependent variable obs. minimum maximum mean std.dev. skewness kurtosis Jarque-Bera prob.
average credit spread 223 -4.45 169.73 33.29 45.91 1.50 4.20 96.65 0.00
natural log return AEX index 219 -0.29 0.12 0.00 0.04 -1.93 17.02 1930.52 0.00
volatility index (VAEX) 220 10.78 81.22 24.72 13.26 1.87 6.91 269.03 0.00
3 month Euribor 230 0.66 5.38 3.17 1.43 -0.39 1.89 17.62 0.00
liquidity 222 0.00 0.12 0.04 0.02 0.77 2.62 23.17 0.00
US Treasury rate 220 0.02 5.05 2.78 1.95 -0.27 1.38 26.56 0.00
house price index 224 100.90 113.51 108.00 3.52 -0.35 2.10 12.08 0.00
4. METHODOLOGY
For the empirical part of this thesis I test for the influence of the determinants on the credit spread of Dutch Jumbo covered bonds. The dependent variable is the credit spread and the independent variables are the determinants which are described in more detail in chapter 3 Data. I want to measure the significance and influence of the independent variables found in the literature review on the credit spread level of Jumbo covered bonds in The Netherlands. Based on the variables found in the literature review and the data description in the previous chapters, this proposition gives the following formula as starting point:
CSi,t = α + β1Euribort + β2Houset + β3Liqi,t + β4AEXt + β5Slopet + β6USt + β7VAEXt + εt (1),
where CSi,t is the credit spread of bond i at time t, α is a constant, Euribort is the three month
Euribor at time t, Houset is the house price index at time t, Liqi,t is the proportional bid – ask
spread of bond i at time t, AEXt is the natural log-return of the AEX index at time t, Slopet is the
slope of the yield curve at time t, USt is the yield of the three month Treasury Bill in the
secondary market at time t, VAEXt is the level of the volatility AEX index at time t, and εt is an
error term.
The data is unbalanced panel data, which means the data is time series and cross sectional (panel) but not all data from all bonds from the same time period are available (unbalanced). The use of panel data increases the degrees of freedom and therefore increases the power of econometric estimates and also decreases problems with multicollinearity among determinants. The increased power of econometric estimates and decreased problems with multicollinearity is because the gap between the information requirements of the model and the information provided by the data decreases (Baltagi, 1995).
behavior in certain circumstances. For sample sizes like my dataset (220 observations), the violation of the normality assumption is practically inconsequential (Brooks, 2008). Keeping this in mind, results of the determinants of credit spread changes discussed in the next chapter can be used for valid inferences.
Recall from chapter 3 Data that all the explanatory variables (except the slope of the yield curve) as and the dependent variable are non-stationary at level. This means that regression of these variables can lead to a spurious regression as described in the previous chapter. A common intervention is to take the first differences of the variables, which are in effect the weekly changes from one week to the next. It is also possible to look for co-integrating relationships between the variables so that no observations are lost by taking the first difference. However, enough observations are available so that the method of taking the first differences is used. The consequence is that the economical meaning of the regression shifts from explaining the level of credit spread to explaining the movements in the level of credit spread with the use of the
movements in the level of the explanatory variables. This way it becomes a sort of volatility
analysis and the goal of volatility analysis must ultimately be to explain the causes of volatility (Engel, 2001). The causes of credit spread changes found by the model can be interpreted as determinants of credit spread movements.
Table 5. Explanatory variables and expected signs of the coefficients of the regression
ΔCSt = α + β1ΔEuribort + β22ndΔHouset + β3ΔLiqt + β4AEXt + β5Slopet + β6ΔUSt + β7ΔVAEXt + εt (2)
determinant description predicted sign
AEXt natural log return AEX index -
ΔVAEXt weekly change in volatility (AEX) index +
ΔEuribort
weekly change in 3m Euribor -
Slopet slope yield curve (10 year yield minus 3 month yield) -
ΔLiqt weekly change in 2nd market liquidity +
ΔUSt
weekly change in US Treasury rate -
2nd ΔHouset weekly 2nd change in house price index -
-where ΔCSt is the first difference of credit spread at time t, α is a constant, ΔEuribort is the first difference of the
three month Euribor at time t, 2ndΔHouset is the second difference of the house price index at time t, ΔLiqt is the first
difference of the proportional bid – ask spread of bond i at time t, AEXt is the natural log-return of the AEX index at
time t, Slopet is the slope of the yield curve at time t, ΔUSt is the first difference of the yield of the three month
Treasury Bill in the secondary market at time t, ΔVAEXt is the first difference of the volatility AEX index at time t, and
εt is an error term.
Following Collin-Dufresne et al. (2001), separate bonds with less than 25 observations are discarded for the individual regression estimates. For every individual bond i at time t I estimate the following regression:
ΔCS i,t= α0 + β1ΔEuribort + β2ΔHouset + β3ΔLiqi,t + β4AEXt + β5Slopet + β6ΔUSt + β7ΔVAEXt + εt (3),
5.RESULTS
This chapter gives the results that arise from the analysis as set out in the methodology chapter. I start with the correlation analysis. After that, the regression results of the separate bonds, separate issuers and different periods are presented and discussed. I end this chapter with a robustness check of the results by modifying the model specification (2) and (3).
5.1 Correlation
The correlation coefficients between the credit spread of the three different issuers (ABN AMRO, Achmea and ING Bank) are very high and range from 0.94 and 0.99 at the 1% significance level. This indicates that the observed credit spread movements are likely caused by the same underlying factors and that issuer specific characteristics are less important in explaining movements in credit spread.
rather high (-0.89) and significant, there is however no real theoretical justification for this high correlation.
Appendix C2 contains the correlation using Pearson„s r between the explanatory variables as used in the regressions. This gives rather different correlation statistics. Correlation between the natural log return of the AEX and slope of the yield curve and the correlation between Δ-volatility index and slope of the yield curve is significant at the 1% significance level. Respectively 4 and 2 correlations between explanatory variables are significant at the 5% significance level and 10% significance level. I omit some of the highly correlated variables and estimate the modified regression in paragraph 5 of this chapter.
5.2 Separate Bond Regressions
Appendix D1 shows the table with the estimated regression coefficients of model specification (3) for each separate bond over the entire sample size period. One Jumbo covered bond from ING and a Jumbo covered bond from ABN AMRO are discarded from the sample because they have too few observations, which leaves 10 separate bonds.
positive impact on the credit spread movements of the separate bonds. These results are in line with Collin-Dufresne et al. (2001), Houweling et al. (2005) and Longstaff et al. (1995) who conclude that (changes in) yield spreads are significantly affected by liquidity risk.
The findings of Van Landschoot (2008) of the influence of US Treasury rate on the Euro bond yield dynamics do not occur in the test results of this thesis. For none of the separate bond regressions I find a significant effect of US Treasury rate on Jumbo covered bond credit spread movements. This result emphasizes the fact that the result by Van Landschoot (2008) was indeed highly surprising and is probably not very persistent across other data samples. The proxy for the house price index is significant for 3 separate bonds with the coefficients being relatively large and they have the expected sign. The coefficients are relatively large because the house price index is not modified by taking the logarithm. The second differences used in the regression are therefore relatively large and the coefficients tend to be larger as well. As expected from the hypothesis, the slope of the yield curve has the expected sign for most Jumbo covered bonds and is significant for 6 bonds. Remarkable is that all 3 Achmea bonds are not significantly influenced by the slope of the yield curve. This result is in contrast with Collin-Dufresne et al. (2001) who find neither economically nor statistically significance of the slope of the yield curve in their test results. This might be explained by the difference in the construction of the slope of yield curve proxy. Collin-Dufresne et al. (2001) construct the slope of the yield by subtracting a 2-year yield from a 10 year yield, whereas I subtract a three month yield from the ten year yield. For one bond (ING 4.75%) I find no significant variables at all, which is probably caused by the low number of observations for the relevant bond (27 observations). The adjusted R2 ranges from 0.02 to 0.28, which means that the model specification explains the observed credit spread from 2% to 28% for the separate bonds. The Durbin-Watson statistics which measures autocorrelation ranges from 1.67 to 2.57, which is within a reasonable range to assume that autocorrelation does not influence the regression estimates. The F-statistics for the separate bonds are all pretty low, however they are mostly at the 1% significance level.
5.3 Separate Issuer and Average Credit Spread Bond Regressions
of the issuer Achmea, but for the issuer Achmea it is significant at a 5% significance level. A closer look into the separate bond regressions shows us that the p-value of the coefficient of the slope of the yield curve is just above 0.10 for all three regressions. This possibly explains the significance of this variable for the issuer Achmea, which is in line with the results of the non-Achmea separate bond regressions.
The regression results of issuer ING Δ-credit spread show only significant coefficients for the intercept and the slope of the yield curve. This result is expected since the separate regressions of the ING bonds did not contain many significant variables. The regression estimates for the issuer Achmea Δ-credit spread show that the credit spread movements are mainly caused by the slope of the yield curve proxy, the Δ-3month Euribor and the first differences of the liquidity proxy. These variables are all significant at a 1% or 5% level. The adjusted R2 is 0.12, which means that 11.9% of the credit spread movements are explained by this model specification. ABN AMRO was the first to issue a Jumbo covered bond in the Netherlands and shows from the regression estimates that the proxy for the house price index has a significant negative influence on credit spread movements. The Δ-3month Euribor and the slope of the yield curve are have the predicted sign and are significant at a 5% and 1% significance level respectively, which supports the alternative hypotheses H3a en H4a. The liquidity proxy is only significant in the regression of Achmea.
Achmea has the smallest amount issued in respect of Jumbo covered bonds compared to ABN AMRO and ING. This can be seen by the standard deviation of the liquidity proxy which is much higher for Achmea than for ING and ABN AMRO. The effect on credit spread is therefore different from the liquidity effect on ING and ABN AMRO. The adjusted R2 ranges from 0.08 to 0.12, which means that the model explains around 8 to 12% of the credit spread movements for the different issuers over the sample size. The Durbin-Watson statistic ranges from 1.62 to 2.35, which indicate to autocorrelation does not influence the regressions estimates much.
Figure 1 in chapter 3 shows that the level of credit spread rose exponentially during the financial crisis. Thus far the results are based on the entire sample size period. This can lead to values of the regression coefficients that are not very representative due to different market circumstances and events, such as the financial crisis and the fall of Lehmann Brothers. Therefore, I check for a structural break in the data with the use of Chow (1960) and Quandt-Andrews (1990). The Quandt-Andrews breakpoint test tests for one or more unknown structural breakpoints in the sample for a specified equation. The idea behind the Quandt-Andrews test is that a single Chow breakpoint test is performed at every observation between two observations. The k test statistics from those Chow tests are then summarized into one test statistic for a test against the null hypothesis of no breakpoints between 2 observations. The maximum lagrange F-statistic gives January 16th 2009 as most likely breakpoint. However, the test statistic is highly insignificant which means that the hypothesis of no structural break cannot be rejected. I also perform the Chow test which needs a specified date to test for a structural break. The break dates are based on historical events which can also be seen in figure 1 of chapter 3 Data. I use the start of the financial crisis and the fall of Lehmann Brothers as possible breakpoints. Based on these break dates, I run regressions of the different periods with Δ-average credit spread as dependent variable .
5.4 Regression Results from Periods with Structural Breaks
Appendix D3 gives an overview of the regressions run over different periods. Of interest are the differences in coefficient estimates and significances of the explanatory variables between the periods which are broken up from the original time period. The first 2 columns show the breakdown of the period based on the Quandt-Andrews test. The second 2 columns show the period broken up based on the start of the financial crisis and the third 2 columns in the table take the fall of the Lehman Brothers as structural breakpoint for the regressions.
The regressions based on Quandt-Andrews show the slope of the yield curve as a significant variable with the expected sign in both periods. The size of the coefficient differs from the first period, where it is -0.86 against -6.47 in the second period. The liquidity proxy is negative and significant for the first period, which is opposite according to the hypothesis.
period the coefficient is positive. This opposite sign can possibly explained by a vicious circle according to Sachs et al. (2009). They state that investors do not want to sell because they expect spreads to tighten further, dealers are no longer making their banks„ balance sheets available as a source of liquidity, and spreads are reducing because the market is getting squeezed. This means that most existing buy orders cannot be fulfilled, and isolated trades have the capability to trigger substantial spread movements. Swap spreads are reducing in this context precisely because liquidity is low.
The third set of 2 columns takes the fall of Lehmann Brothers as the point of structural break. Remarkable to see is that for the first period no single variable is significant, and the adjusted R2 is negative which has normally no statistical meaning. The volatility index, US Treasury rate and the house price index proxies are not significant for any of the different periods, which is in line with results from 5.2 and 5.3.
The US treasury rate proxy seems to have no impact at all based on the regressions so far. The surprising results by Van Landschoot (2008) are not borne out my data and I do not find much literature that confirms these results. This leads me to believe that the US Treasury rate has no influence on the credit spread movements of Dutch Jumbo covered bonds. For robustness I exclude the US Treasury rate and other insignificant variables. I run a regression with the dependent variable credit spread on one or a few variables and remove the abundant variables to improve the quality of the model specification.
5.5 Robustness Check and Empirical Results
I conduct robustness checks to see if the original model is sensitive to changes in the model specification. The insignificant variables from each separate issuer regression of model specification (3) are left out. The modified regression model (4) is slightly different per issuer because it comprises only the significant variables found from model specification (3).
ΔCSi,t = α + β1 ΔEuribort + β2 ΔLiqi,t + β3 2ndΔHouset + β4 Slopet+ εt (4)
