Temporary Alienation or Lasting Separation ? Covered and Government Bond Spread Movements during and after the Crisis of 2007-2009

This study examines the interrelationship of asset swap spreads on government and mortgage covered bonds in Germany, France, Italy, and Spain between 2007 and mid-2014. Using a local least squares estimator with time varying parameters, we find that in all of the four countries under investigation, the pattern of spread movements for these two bond classes underwent significant changes over time. In Germany, where the confidence of market participants in the solidity of public finances appears to be largely unshaken, spreads were driven apart due to “flight to safety” effects in times of turmoil, and drew closer again when the situation steadied. Yet in France, Italy, and Spain, the (partial) erosion of confidence in the sustainability of government debts led to a protracted weakening of the linkage between the spread movements of government and mortgage covered bonds.


Introduction
Covered bonds are interest-bearing, dual recourse bonds which are backed by a pool of underlying loans.The cover pools can either consists of commercial or residential mortgage loans, loans to public sector entities, or shipping and aircraft loans that serve as collateral for investors.Covered bondholders benefit from a dual protection in the form of a preferential claim over the cover assets in addition to a claim against the creditworthiness of the issuer itself.
Due to these additional safety layers, covered bonds have, over most of their history, been regarded as close substitutes for government bonds from the respective issuer's country of residence.In a "well-behaved" market, yields on covered bonds and yields on domestic government bonds of the comparable maturity can hence reasonably be expected to move roughly in parallel as time passes, with a small yield advantage for covered bonds that reflects the residual risk of a joint default by the issuer and the cover pool.
Yet the global financial crisis of 2007-2009, which originated in the U.S. subprime crunch, culminated with the collapse of Lehman Brothers in September 2008, and subsequently transformed into a sovereign debt crisis for some of the countries affected, has left deep marks on global financial markets, abruptly challenging patterns in asset pricing relationships that had been previously taken for granted, as Eichert and Rudolf (2013, p. 71) put it.There are three main reasons to suspect that this diagnosis also holds true for the price relationship between government bonds and covered bonds in the Euro currency area: • Firstly, in countries with a strong reputation for stable government finances, the yield gap between covered and government-issued bonds can be expected to widen substantially in periods of severe market distress because anxious investors flee riskier assets in favour of (alleged) safe haven investments.
• Secondly, in countries where a protracted turmoil within the financial sector raises the spectre of a possible government default, the formerly imagined safety advantage of government bonds presumably wears out the more obvious the severity of the crisis becomes.In principle, it could even be possible for lenders to remain solvent while their domestic government defaults, particularly if the former are considered "systemically relevant" and thus eligible for capital and/or liquidity support by some supranational entity.On the other hand, it is hard to imagine a sovereign debt crisis not being accompanied by a wave of defaults in the market for domestic cover assets.Such a situation might instead induce governments to change the legal framework in favour of defaulted borrowers, barring lenders from taking possession of the cover assets and thus passing on any resulting losses to either covered bond issuers or investors.In any case, it is plausible to assume that the resulting uncertainty can be reasonably expected to induce a prolonged decoupling of the yields on otherwise comparable covered and government-issued bonds.
• Thirdly, selective central bank bond purchases aimed at stabilising targeted market segments and resolving banks' refinancing problems in times of crisis almost inevitably impact (or, as critics would have it, "distort") the price relationships prevailing on the market.For the issue addressed here, three distinct asset purchase programmes by the European Central Bank (ECB) deserve particular attention.These are: o The first Covered Bond Purchase Programme (CBPP1), lasting from July 2009 through June 2010 and involving a nominal value of € 60bn (source: ECB, 2011).
o The Securities Market Programme (SMP, May 2010 to February, 2012), leading to the purchase of around € 214bn of government bonds from Greece, Portugal, Spain, Italy, and Ireland (sources: Eser et al., 2012, andECB, 2013), and, o The second Covered Bond Purchase Programme (CBPP2, November 2011 through October 2012), once again amounting to a nominal value of € 60bn (source: ECB, 2012a).
The fact that the ECB has announced a third Covered Bond purchase programme in October 2014 (see ECB, 2014) with a view to stimulating credit supply and averting deflation underscores that (at least) the last-mentioned point will remain very present in the near future.
Against this background, the purpose of this paper is to trace the impact which the recent crisis and the policy measures aimed at its resolution have made on the market's perception of the default risk of covered vis-à-vis government bonds in four European economies (Germany, France, Italy, and Spain).The investigation concentrates on the mortgage-covered segment of the covered bond market because this currently is by far the largest one in terms of outstanding value (see, e.g., European Banking Authority, 2014).More specifically, this paper seeks to answer the following questions: • Did the long-standing tendency of covered and government bond prices to move in an equidirectional manner over time unravel in the course of the crisis?And if so, was the observed decoupling of these two market segments a short-lived or long-lasting phenomenon?
• Are there any marked differences in the development of pricing patterns for both asset subclasses between the individual countries under investigation?
• How did the unconventional monetary policy measures taken by the European Central Bank in response to the crisis affect risk premia on both sub-markets in the countries involved?
The remainder of this paper is organized as follows: The following section 2 provides a brief review of the relevant literature.The data in use are described and some related descriptive statistics are then presented in Section 3. Section 4 contains a brief description of the statistical model and estimation methods used.The empirical results are presented and commented upon in Section 5.The paper ends with a short summary and some conclusions (section 6).

Literature Review
The levels of credit spreads in the bond market, and their movements over time, have been analysed in a considerable number of earlier studies, most of which have focused on the U.S. Examples include Chen (2010), Longstaff, Mithal, and Neis (2005), Elton et al. (2001), Ng and Phelps (2011), as well as Churm and Panigirtzoglou (2005), among others.Most of these studies, however, relate to unsecured bonds issued by corporates or (less frequently) by financial institutions.
Given the size of the covered bond market and its importance for financial market stability, it might come as a surprise that existing empirical investigations of the behaviour of credit spreads in this market are comparatively small in number.Most of the related research focuses on the German Pfandbrief market and often centres around the yield gap between Pfandbriefe and German government bonds; see, e.g., Bühler andHies (1998), andJobst (2006).A predictive model relating the 10 year Pfandbrief spread to a number of macroeconomic factors has been developed by Rees (2001), whereas Koziol and Sauerbier (2007), Siewert and Vonhof (2011) as well as Kempf et al. (2012) examine the linkage between Pfandbrief spreads and market liquidity measures.Moreover, the effects that both credit risk and market liquidity have on the related markets have been examined by Breger and Stovel (2004) as well as Sünderhauf (2006).
So far, few empirical investigations of the behaviour covered bond spreads on a cross-national basis have been carried out.The contributions by Packer et al. (2007), and Volk and Hillenbrand (2006) give evidence of the significant impact that the issuer's country of residence has on covered bond yields, which can, at least in part, be ascribed to the lack of a common regulation of national covered bond markets in Europe.Another very revealing study in this context is the one by Prokopczuk and Vonhoff (2012) who, on the basis of a large panel dataset, find that cross country differences in asset swap spreads on covered bonds are a lot more pronounced in times of economic turmoil than under stress-free economic conditions.
The current paper builds on this body of research, while shifting attention more towards the various footprints the subprime mortgage crisis of 2007-2009, the subsequent sovereign debt crisis in the peripheral countries of the Euro currency area, and the ECB's crisis management operations have left on the markets for both covered and government bonds.

Underlying Data and Descriptive Statistics
The purpose of this empirical investigation is to identify and examine structural changes in the market-implied credit risk of mortgage covered vs. sovereign bonds.To this end, average market credit spreads pertaining to these two asset subclasses are used as a measurement criterion.More specifically, the country-specific iBoxx Mortgage Covered Bond and Sovereign Bond indexes (both being provided by Markit Ltd., a global, financial information company) are used for measuring the performance of the selected covered bond markets.Markit indices are widely used as benchmarks by investors and asset managers.They benefit from multiple-contributor pricing from selected leading financial institutions, which also provide support to the index family in research and trading.
To isolate the credit risk premium from the remaining economic drivers behind the observable index levels, the daily asset swap spread of each index segment (covered and sovereign) is gathered for each of the four countries under investigation.In line with Choudry (2008), it can be characterized as follows: An asset swap is a package that combines an interest rate swap with a cash bond, thus transforming the interest rate basis of the bond.Typically, a fixed rate bond will be combined with an interest rate swap in which the bondholder pays fixed and receives floating coupon, the latter of which is a short-term interbank rates such as Euribor/Libor and increased by a spread, which is referred to as the asset swap spread.Its level at any given time reflects the credit risk of the underlying bond relative to the inter-bank credit risk.
For the purpose this investigation, asset swap spreads pertaining to the different iBoxx indexes in use are calculated, in line with Markit (2010, p. 16), by weighting the asset swap spread of each bond included in the index with its corresponding market capitalization and duration: For the sake of completeness, it must be mentioned that asset swap spreads based on bond market indexes like iBoxx do not constitute a perfect representation of the market's perception of the underlying default risks.This is due to the following reasons: The composition of the indexes may change over time, the constituents of two different members of the index family may differ in terms of duration, and government bonds usually tend to be more actively traded than covered bonds (see Dick-Nielsen et al., 2012).Although all these factors add some "noise" to the data, the above measure is used for this investigation because there simply are no comparable bond market indexes that do not suffer from these ailments.
The database used consists of asset swap spread data for both government bonds and mortgage covered bonds from Germany, France, Italy, and Spain.The sampling period ranges January 1, 2007 to May 28, 2014.In the case of Germany, France, and Spain, this corresponds to 1,914 total observations, each referring to a particular trading day.For Italy, covered bond spread data are only available from January 1, 2009, onwards, which means that in this case, only the period starting at that date could be investigated.
Here and in the following, the level of the mortgage covered bond spread at any trading day t are denoted by C., and its change between two subsequent trading days t and t-1 by ΔC t .Likewise, S t and ΔS t , respectively, stand for the level and the day-to-day change in the sovereign bond spread.Country-specific descriptive statistics for these quantities during the sampling period are given below:

The Regression Equation
In order to capture a possible, temporary or permanent divergence of the risk premia associated with governments bonds and mortgage-covered bonds from the same country, a linear regression model with time varying coefficients is used.In the following, ΔC t denotes the change in the asset swap spreads on covered bonds between two subsequent trading days t and t-1, and ΔS t stands for the change in the sovereign bond spread in the related country during the same period.Then, the regression equation by which the co-movement, or the permanent or temporary drifting apart of these two quantities is being modeled, reads: where the error terms ε t are mutually independent with . This specification, which may appear somewhat peculiar at first sight, is chosen because at least some of the possible parameter constellations it permits lend themselves to a straightforward interpretation: • In periods where neither α(t) nor β(t) differs significantly from zero, covered bond spreads and sovereign spreads tend to move in an essentially parallel manner, disturbed only by the realizations of the error term ε t This is the state one would expect to prevail if the risk content of both bond classes and the risk appetite (or risk aversion) of investors remained unchanged over time.
• Values of β(t) that are significantly below zero but greater than (-1) would imply that, at the particular point t in time, an increase in the sovereign spread tends to be accompanied by an equidirectional, but less pronounced change in mortgage-covered bond spreads.Economically, such an observation implies that as the market-implied default risk of the sovereign increases, the perceived "safe haven" property of government bonds gradually tends to wear out.
• Values of β(t) that do not significantly differ from (-1) indicate that at the time of measurement, mortgage-covered bond spreads have completely decoupled (because in this case, equation ( 2) would just boil down to ΔC t = α(t) + ε t ).
• Values of β(t) that lie significantly below (-1) suggest countervailing movements of sovereign and mortgage-covered bond spreads • A value of α(t) that significantly exceeds zero implies that the market-implied default risk of mortgage-covered bonds, relative to sovereign bonds with the same country of origin, has increased by a larger amount than the concurrent move in the sovereign bond spread would have led one to expect.
• Likewise, a value of α(t) that is significantly below zero indicates that the market-implied default risk of sovereign bonds, as compared to mortgage-covered bonds from the same country of origin, has decreased further than suggested by the simultaneous change in the covered bond spread.
By examining the developments of the regression coefficients in (2) over time, it is possible to discern different "régimes" as to the relationship between mortgage-covered and sovereign bond spreads, and, perhaps, to trace them back to evolving or already completed changes in the underlying economic fundamentals.The following subsection deals with the estimation method applied for this purpose.

Estimation Method: Locally Weighted Least Squares
For the functions α(t) and β(t), which describe the development of the regression coefficients over time, no particular form has been specified.Rather, it will only be assumed in this paper that these functions are smooth, so that, for every particularly value t 0 of t, they can be approximated by a constant in a reasonably chosen neighbourhood of t 0 , In the following, the parameter h, referred to as the bandwidth, stands for the size of this neighbourhood, and the so-called kernel function K( .) denotes a non-negative weight function.(Here and in the remainder of this subsection, the description closely follows Fan et al. (2003).The local regression technique described, inter alia, by Fan and Gijbels (1996), consists of finding estimates α(t 0 ) and β(t 0 ) for each of the possible values t 0 of t by minimizing the locally weighted least-squares criterion function ( ) and β(t 0 ).In the particular context, we choose K( .) to have one-sided support ranging from minus infinity to zero, to reflect the fact that between two observation times only past data can be used for predictive purposes.More specifically, we set K( .) to ( ) Here, ( ) ⋅ φ denotes the density function of the Standard Normal distribution.This choice is only one of several possible forms for the kernel function (see Cleveland & Devlin, 1988, for an overview), but experience suggests that the particular form of the kernel function chosen usually does not significantly affect the estimation process.
Then, the locally weighted least squares estimators of α(t 0 ) and β(t 0 ), denoted by , are the minimizers of the weighted least-squares criterion (3).

Bandwidth Selection
The bandwidth h equation (3) determines how quickly the weights of the past observations decrease as the distance from the reference date t 0 grows.In cases where h is "too small", the resulting parameter estimates tend to "fit the noise", i.e. to be too sensitive to the specific realizations of the random influences present in the data, to possess excessive variance, and to be poorly generalizable.On the other hand, choosing h to be "too large" will cause important features in the unknown, true functions α(t) and β(t) to go unnoticed.In the context of this investigation, the proposed solution to this dilemma is to follow Härdle (1990, section 5.1.1.)in choosing the "optimal" value h * of h in such a way that it minimizes the "leave-one-out criterion" ( ) The intuition behind this decision criterion is that a reasonable choice of h should be the one that best predicts the dependent variable (ΔC t0 -ΔS t0 ) by using only observations from the past (t < t 0 ).

Standard Errors and Estimated Confidence Intervals
Assuming that the sample size is large enough for the coefficient estimates to be approximately normally distributed, the estimated standard errors α(t) and β(t) can be used to calculate asymptotic confidence intervals for these coefficient estimates, and to assess whether or not they are sufficiently distant from certain reference values (e.g.zero or minus one in the cases discussed above) to support (or reject) the hypotheses associated with such an observation.Since the regression (3) is merely a special case of the weighed least squares estimation technique (see, e.g.Gouriéroux and Monfort, section 8.3.), the related asymptotic properties apply.The formulas for standard errors can thus be adopted from p. 115 of Fan and Gijbels (1996).Throughout the following discussion of the estimation results, a parameter estimate is considered significantly different from a reference value v if v lies outside the surrounding two-sided 95% confidence interval.

Germany
The time paths of the parameter estimates for Germany, along with the related 95% confidence intervals (CI), are   is significantly below zero most of the time (frequency: 95.85%), but the percentage of observations where this parameter is significantly above (-1) is recognizably lower (69.3%)than in the case of its eastern neighbour.The development of the parameter estimates over the sampling period is displayed in Figures 3 and 4.
stays close to zero until, in autumn 2008, the events surrounding the "Lehman shock" drive it to new highs, reflecting, once again, the "flight to safety" effect sketched above.Subsequently, the combined impact of the fiscal and monetary policy measures, last but not least CBPP1, reverse this impulse, first driving ) ( ˆt α sharply into the negative and eventually leading to a temporary restoration of the relative calm prevailing before September 2008. The rather close resemblance between the examined spread movement patterns came to an abrupt end as the fears about the possibility of sovereign defaults in one or more of the "peripheral" member states of the Eurozone (Greece, Ireland, Italy, Portugal, and Spain, or GIIPS) grew in July, 2011.The sharp upward jump in ) ( ˆt α at the beginning of this sub-period presumably reflect the large extent to which French banks used to be exposed to the default risk of the GIIPS governments and financial institutions -particularly from Greece -at that time (see, e.g., Curley, 2012;Daneshkhu, 2012).The second, massive wave of government bond purchases conducted by the ECB (roughly between August, 2011 and January, 2012; see Trebesch & Zettelmeyer, 2013, p. 40) in the course of its SMP programme appears to have brought a rather short-lived and incomplete relief from these tensions, as the sharp and very erratic movements in the ) ( ˆt α in that period of time suggest.Apparently, it is not until the effects of CBPP2 begin to materialise near the end of 2011 that ) ( ˆt α first declines and then becomes negative for a period of about ten weeks following January 27, 2012.Yet because of French banks' considerable exposure to neighbouring Spain, ) ( ˆt α briefly shot back up into the positive terrain in June 2012, when that country faced difficulty in accessing bond markets (Traynor & Watt, 2012).hovers closely to (-1), indicating a protracted weakening, or even a near-complete break-up of the previously observed linkage between the spread levels for mortgage-covered and government bonds.A plausible explanation for this phenomenon is that in France, the huge transfer of private sector risk to the public sector that occurred in the course of the state-sponsored rescue packages has led to an erosion of the perceived safety advantage of government debt over covered bonds.As a consequence, the solvency of both the government and the financial sector have nowadays become significantly dependent on the continuing readiness (and ability) of the European Central Bank "to do whatever it takes to preserve the euro", as ECB President Mario Draghi put it at the Global Investment Conference in London on July 26, 2012 (see ECB, 2012b), shortly before announcing the Outright Monetary Transactions (OMT) programme allowing for (theoretically) unlimited purchases of Eurozone sovereign bonds in the secondary market if needed.) ( ˆt β is han (-1) in -half years (-1) in the remaining Spain differs from the three other economies examined here in that the financial crisis that the country underwent in the years after 2007 can, at least to a considerable extent, be attributed to the bursting of a domestic property bubble in the first half of 2008.However, in much 2008 and early 2009, it appeared that Spanish banks would emerge from the financial crisis relatively unscathed because of their strong capital and provisioning buffers (see International Monetary Fund, 2009, p. 45).The fact that the flight-to-quality effect diagnosed for Germany and France during the period from (roughly) mid-September 2008 to January 2009 is hardly detectable in the case of Spain can, at least in part, be attributed to this phenomenon.

The behaviour of
The first cracks in the apparently solid foundation of the Spanish banking sector occurred later during the first quarter of 2009, when the medium-sized the savings bank Caja Castilla La Mancha (CCM), which had been heavily exposed to the property sector, suffered a capital shortfall and was subsequently taken over by the government (see Reuters, 2012).In the time series of our parameter estimates, this development is reflected by a first, sharp upward move in ) ( ˆt α during that period of time.However, the announcement and subsequent execution of CBPP1 in mid-2009 appear to have impacted the Spanish mortgage-covered bond market strongly, as the temporary yet pronounced move of ) ( ˆt α into negative territory in the second half of the year suggests.The fact that, in June 2009, the Spanish government established the banking bailout and reconstruction programme "Fondo de Reestructuración Ordenada Bancaria" (FROB; see Hugh, 2010), probably also played a role in this development.
The period of relative calm that followed came to an abrupt end when in May, 2010, FROB was forced to take control of Cajasur, another savings bank with troubled property loans (see Reuters, 2012) during the sampling period can be linked to specific events within the Spanish banking sector: • In the second quarter of 2011, Caja de Ahorros del Mediterráneo (CAM; Mediterranean Savings Bank) slid into financial difficulties and was taken over by the FROB in July.
• Between mid-October and early December 2011, a large number of Spanish banks were downgraded by rating agencies S&P and Fitch (see McDermott, 2011), raising fresh concerns about the inability of further savings banks to deal with their bad real estate loans or raise additional capital from the market.This prompted the government to partially nationalize some of the weakest cajas in early 2012.This appears to have brought some relief to the mortgage-covered bond market, as the move of ) ( ˆt α into negative territory by the end of January, 2011, indicates. • In May, 2012, the credit ratings of several Spanish banks were downgraded, some to "junk" status.Bankia, the country's largest mortgage lender (which had resulted from the merger in December of 2010 of Caja Madrid, Bancaja and five other smaller savings banks), was nationalised on 9 May, and on 25 May it announced that it would require a bailout of €23.5 billion to cover losses from failed mortgages (sources: Minder, 2012;Santos, 2014).
In June, 2012, Spain abandoned the attempt to re-develop its distressed bank unilaterally and joined Greece, Ireland, and Portugal in requesting a rescue package of up to €100bn from the European Financial Stability Facility (ESM; source : Tremlett, 2012).In December, 2012, the Spanish government raised a specific request for aid amounting to € 39.5 bn, which the ESM granted the FROB shortly thereafter.Another tranche totaling € 1.5 bn was released and subsequently paid out in January 2013 (source: Rose, 2013).In the above estimates, the relief this step (and the effects of CBPP2) brought to the Spanish banking sector is reflected in the fact that in August, 2012, ) ( ˆt α moves into negative territory and stays there over most of the remaining sampling period. The tendency of mortgage covered and government bond spreads to move equidirectionally, as measured by the parameter ) ( ˆt β , appears to be weaker in Spain than in France and (particularly) in Germany, as the lower median value of this parameter estimate and its somewhat erratic fluctuations, particularly before July 2010, indicate.During most of the time that follows, ) ( ˆt β essentially moves up and down between (-0.7) and (-1), indicating an almost complete dissolution of any observed linkage between the spread levels for both bond types.Like in the case of France after November 2011, a plausible explanation for this having occurred is the evanescence of the perceived solvency advantage of the government over covered bond issuers due to the assumption of private sector default risk by the state.In Spain, too, the financial health of domestic banks and the public sector have become mutually dependent on each other, and jointly dependent on the continued willingness, and ability, of the ECB to quench any anxiety on the part of investors by large-scale secondary market bond purchases when required.

Summary and Conclusions
This paper examined the interrelationship of government and mortgage covered bond spreads during a sampling period which stretches from the incubation time of the most recent financial crisis in early 2007 through to the year 2014, using a linear regression model with time varying parameters based on a locally weighted least squares estimation procedure.
In all four countries under investigation, the accustomed tendency of government and mortgage covered bond spreads to move equidirectionally has at least been interrupted during the sampling period.In Germany, where the confidence of market participants in the solidity of public finances appears to be largely unshaken, credit spreads on government bonds and mortgage covered bonds were temporarily driven apart due to "flight to safety" effects when the financial crisis of 2007-2009 and the subsequent Eurozone sovereign debt crisis peaked, and drew closer again as the impacts of the crisis-management efforts by governments and central banks became effective.In all other countries examined-France, Italy, and Spain-the (partial) erosion of confidence in the sustainability of government debts has led to a protracted weakening of the linkage between the spread movements of government and mortgage covered bonds which, in the cases of Italy and Spain, has come close to a complete uncoupling.The policy reactions of the ECB, most notably the Covered Bond Purchase Programmes, have also had a temporary yet momentous impact on the relationship between these two variables.Although the stabilization measures taken by the ECB have succeeded in bringing down the borrowing costs of the highly indebted member states, the relationship between the spread movements of government and mortgage-covered bonds in these countries has so far not returned to its pre-crisis normal.A possible interpretation of this finding is that in these countries, the financial health of domestic banks and the public sector have become mutually dependent on each other, and jointly dependent on the continued willingness, and ability, of the ECB to act as a lender of last resort if required.
following August, 2011, is (at least) just as telling.During most of the time span between November 2011 and the end of the sampling period,

Table 1 .
Descriptive statistics for Germany (all figures shown in basis points)

Table 2 .
Descriptive statistics for France (all figures shown in basis points) t -ΔS t Sampling period Jan 2007-May 2014 Jan 2007-May 2014 Jan 2007-May 2014 Jan 2007-May 2014 Jan 2007-May 2014

Table 3 .
Descriptive statistics for Italy (all figures shown in basis points) t -ΔS t Sampling period Jan 2007-May 2014 Jan 2007-May 2014 Jan 2009-May 2014 Jan 2009-May 2014 Jan 2009-May 2014

Table 4 .
Descriptive statistics for Spain (all figures shown in basis points) Sampling period Jan 2007-May 2014 Jan 2007-May 2014 Jan 2007-May 2014 Jan 2007-May 2014 Jan 2007-May 2014 55%, hence exceeding the corresponding value for Germany, yet without being conspicuously high in comparison.Like in Germany,

Table 6 .
Descriptive statistics of parameter estimates: France In a first phase of the sampling period, lasting from January 2007 through August 2011, the movement patterns of the coefficient estimates in France broadly resemble the ones observed in Germany: