Treasury Yields and Credit Spread Dynamics: A

Regime-Switching Approach
Dimitris A. Georgoutsos* and Thomas I. Kounitis

Department of Accounting and Finance,

Athens University of Economics and Business,
76 Patission Street, Athens GR-10434, Greece
E-mail: [email protected] , E-mail: [email protected]

March, 2014

Abstract

The purpose of this paper is to shed new light on the conflicting empirical
evidence on the relationship between credit spreads and Treasury rates. Following a
general-to-specific modelling approach, we were unable to accept the presence of a
long-run relationship between Baa credit spreads and long-term Treasury rates. At the
same time, and in support of the structural models on credit risk modelling, a negative
short-run relationship was obtained by means of impulse response functions.
Subsequently, by employing a regime-switching estimation technique, we were able
to establish the importance of the Treasury yield curve slope for the Baa credit spread
determination in periods characterized by low interest rate volatility. Finally, we were
able to provide evidence of an asymmetric response of the Baa credit spread to term
spread changes according to the source of these changes, i.e. short or long term
Treasury rates.

JEL classification: C32; E43; E44; G12

Keywords: Credit spreads; Term spreads; Regime shifts.

(*) correspondent author

Electroniccopy
Electronic copy available
available at:
at:https://ssrn.com/abstract=2569083
http://ssrn.com/abstract=2569083
1. Introduction

The enquiry into the determinants of corporate credit spreads has stimulated a

large body of academic research over recent years. The existing literature on credit

risk modeling revolves around two fundamental theoretical approaches: the structural

and reduced form models. Structural form models treat corporate debt as being

equivalent to a synthetic asset consisting of a long position on a risk free bond and a

short position on a put option written on the value of the firm and struck at the face

value of the debt. Within this class of models, default is basically determined by the

firm’s asset value relative to some default threshold. In contrast, reduced form models

employ market information instead of the company’s financial fundamentals and take

as a premise that bonds, when grouped by ratings, are homogenous with respect to

risk.

The present study departs from the main body of empirical studies on credit

spread determinants and focuses on a segment of the relevant literature that deals

solely with the relationship between risk free interest rates and the corporate –

Treasury yield spread (e.g., Longstaff and Schwartz, 1995; Duffee, 1998; Neal et al.,

2000).1 In particular, the main issue we address in this paper is whether the slope of

the Treasury yield curve constitutes an explanatory factor of corporate credit spreads.

The underlying rationale is that the slope of the yield curve acts as an indicator of

future economic activity and is therefore closely related to credit spreads which

quantify, among others, default risks and liquidity squeezes. For example, an increase

in the yield curve slope is associated, under the expectations hypothesis and

neglecting term premiums, with improving economic conditions, higher expected

future short rates, a higher drift of the company’s value process and therefore a

smaller probability of default. Equivalently, a negative slope might be the outcome of

Electroniccopy
Electronic copy available
available at:
at:https://ssrn.com/abstract=2569083
http://ssrn.com/abstract=2569083
a tighter monetary policy, often intended to reduce inflationary pressures, which

makes a slowing-down over subsequent months more likely and therefore increases

the probability of default (e.g. Estrella and Trubin, 2006). Despite the fact that

structural models do not include the yield curve slope as a possible explanatory

variable, it is nevertheless present in many empirical specifications of the stochastic

behavior of credit spreads; still, the empirical evidence is rather conflicting.

Methodologically, this paper is most closely associated with the empirical work

of Neal et al. (2000, 2012). However, our analysis is novel in a number of ways. First,

we employ a vector error-correction methodology, in accordance with Johansen

(1995), using monthly bond yields of the 1- and 10-year U.S. Treasuries as well as

Moody’s Baa bond index over the period 1960 to 2010. This specification permits the

incorporation of the market’s perspective on future growth, as this is reflected in the

slope of the yield curve, in the empirical investigation of credit spreads. The approach

adopted in this paper departs from a large part of the existing literature that initially

tests the credit spread series (i.e. the difference between corporate and government

bond yields) for unit roots and then applies, accordingly, either cointegration or

“conventional” econometric techniques (Davies, 2004). Instead, our approach is

integrated in the sense that we identify equilibrium relationships among the variables

under study, within the context of cointegration. We provide evidence for the

existence of two cointegrating vectors of the form (1, -1), between Baa and 10-year

Treasury rates, and between 10- and 1-year Treasury rates. The first cointegrating

vector rejects the hypothesis that there exists a long run relationship between credit

spreads and Treasury rates. However, through an impulse response analysis, we show

that in the short run the Baa credit spread declines due to a rise in the 10-year

Treasury rate and that it requires more than 40 months to return to its initial level. Our

Electronic copy available at: https://ssrn.com/abstract=2569083

results are consistent with the evidence of, among others, Longstaff and Schwartz

(1995) and Duffee (1998), who reported a negative short run relationship between

changes in credit spreads and Treasuries. They differ however from the findings of

Neal et al. (2000, 2012) and Davies (2008), who reported a widening Baa credit

spread in the long run due to positive Treasury rate changes and a negative short run

reaction only for the first periods in their impulse response exercise.

The second contribution of this paper involves the examination of the temporal

stability of the aforementioned long-run relationships, using a sequence of tests

suggested by Hansen and Johansen (1993). The results indicate that two cointegrating

vectors are established after 1990, and that the single cointegrating vector prior to that

date is identified with the yield curve slope and not the credit spread. In the early

1980’s, the tests carried out for the constancy of the cointegration space and the

stability of the estimated eigenvectors indicate a structural break.

Third, we further extend the short-run dynamic analysis by allowing for the

presence of different regimes, which are found to be related to the inflation rate and

industrial production. There is currently a large volume of empirical literature that

documents nonlinearities in interest rates as well as the suitability of regime-switching

models for capturing this feature. For instance, Ang and Bekaert (2002) found

overwhelming evidence for multiple regimes in the data generating process of the 3-

month short rates for the U.S., Germany, and Great Britain. Clarida et al. (2006)

estimated a non-linear VECM that is theoretically based on the expectations model of

the term structure. They estimated a MS-VECM, and showed that it provides good in-

and out-of-sample fits and that the regimes are related to the state of the business

cycle and the inflation rate. Our paper employs a Markov-switching vector

equilibrium correction model (MS-VECM), using the estimation techniques

Electronic copy available at: https://ssrn.com/abstract=2569083

developed by Krolzig (1997). The analysis indicates that the yield curve slope

embodies useful information for the credit spread determination during low

conditional volatility periods only. Through an impulse response analysis, we explore

the behavior of credit spreads with respect to the source of term spread changes, i.e.

the 1- and 10-year Treasury rates. The obtained evidence suggests that credit spreads

react asymmetrically to shocks in short- and long-term rates when either of these

shocks results in identical initial term spread changes.

The remainder of the paper is organized as follows. Section 2 briefly reviews the

empirical evidence on the relation between credit risk, risk-free interest rates and the

yield curve slope. Section 3 presents the dataset and describes the econometric

methodology, with an emphasis on the application of Markov-switching techniques to

cointegrated vector autoregressions. Section 4 reports and interprets the empirical

results from the cointegration analysis, the MS-VECM and the impulse responses.

Section 5 concludes.

2. Related literature

Structural models follow the framework set out by Merton (1974), who

incorporated option pricing theory to the pricing of corporate debt. Among the

assumptions embedded in Merton’s model are that the value of the company assets is

the only random variable, and that interest rates are constant. Higher spot interest

rates increase the risk-neutral drift of the company value process and this reduces the

probability that a default threshold will be hit; assuming that the initial company value

remains unchanged. Longstaff and Schwartz (1995) allowed for stochastic interest

rates, and their model still implies a negative relation between credit spreads and the

level of risk-free interest rates. The magnitude of this relation depends on the value of

the instantaneous relation between the asset value and the interest rate processes.

Electronic copy available at: https://ssrn.com/abstract=2569083

Their model was tested on monthly data for eleven Moody’s corporate bond yield

averages for the period 1977-1992 and the results appear to be consistent with theory.

Duffee (1998) argued that the negative relation between corporate bond yield spreads

and Treasury yields is much stronger for callable bonds, and this was validated

empirically when tested on monthly investment-grade corporate bond yields for the

period 1985-1995. Collin-Dufresne et al. (2001) investigated how well the

explanatory variables of the structural framework explain observed credit spread

changes. Their regression analysis can explain only 25 percent of the observed spread

changes, but again an increase in the risk-free rate lowers the credit spread that is

calculated from quotes on non-callable debt of industrial firms over the period 1988-

1997. They failed to identify a statistically significant presence of the slope of the

term structure among the determinants of the credit spreads.

Neal et al. (2000) questioned the empirical findings of previous studies on the

grounds that their models were not specified correctly. They argued that previous

empirical specifications did not incorporate equilibrium relationships between the

variables. This was deemed important because the theoretical models offer long-run

predictions. Furthermore, the estimates from these studies would have been biased if

corporate and Treasury rates had been cointegrated. Cointegration analysis is

necessitated from the evidence documented in a large part of the empirical literature

that either assumes, or finds, that interest rates are non-stationary processes. Even if

interest rate series have a root very close to one, they are better approximated by

integrated of order one, I(1), processes, and cointegration techniques are then more

appropriate (Stock, 1997). Neal et al. (2000) employed Johansen’s cointegration

technique and established the existence of cointegration between the yields of

Moody’s Aaa and Baa indices and the 10-year Treasury bonds. They tested the

Electronic copy available at: https://ssrn.com/abstract=2569083

hypothesis that the cointegrating vector between the yields of corporate and

government bond rates is (1, -1), and failed to reject it only for the Aaa series. As

concerns the Baa series, they found that higher Treasury rates eventually increase

credit spreads and this result was repeated when the Baa series was replaced by the

Lehman Brothers Corporate Index on BAA rated bonds (Neal et al., 2012).

Davies (2004) used cointegration and regime-switching techniques in an attempt

to identify empirically the short- and long-run determinants of credit spreads for the

period 1986-2003. He found that in the long run, credit spreads are inversely related

to the risk-free rate, whilst over the short-run horizon, the impact of the risk-free rate

depends on the volatility of the regime. He further identified a negative long-run

relationship between the credit spread series and the yield curve slope, while within

his dynamic analysis this result was confirmed only for the high volatility period.

Both the credit spread and the slope of the yield curve were treated as I(1) series.

Davies (2008) examined the question of the credit spread determinants over a much

longer time period that spanned 83 years from 1921 to 2004. He found that credit

spreads are positively related to the risk-free interest rate both in the short- and long-

run, which contradicts the findings of Davies (2004). In this specification, the yield

curve slope was treated as I(0), and thus was included as an exogenous variable in his

system. Applying a Markov-switching approach, he reported that the slope is a

significant determinant of the Baa credit spread only under a high volatility regime.2

The existence, or lack of, credit spread stationarity comprises a controversial

issue in the empirical finance literature. On the one hand, there are studies by, among

others, Duffie and Singleton (1999) and Pringent et al. (2001), which have established

empirically, the mean reversion property of credit spreads.3 On the other hand,

Pedrosa and Roll (1998) examined the daily time series properties of credit spreads

Electronic copy available at: https://ssrn.com/abstract=2569083

from 1995 to 1997, and were not able to reject the null hypothesis that credit spreads

are non-stationary. However, the mean reversion property of credit spreads is present

in many valuation models of defaultable debt, since the opposite case of non-

stationarity has two rather implausible features: an explosive volatility structure over

time and the possible realization of negative values. Bhanot (2005) has persuasively

argued that the existence of mean reversion in corporate bond index spreads, similar

to the ones used in this paper, is a result of the ratings based classification of bonds

included in the index and the related effects of survival.

Bierens et al. (2003) proposed an econometric model of the credit spreads that

incorporates, among others, portfolio rebalancing, ARCH conditional

heteroscedasticity and lagged market factors, and found that a steepening Treasury

yield curve predicts falling credit spreads for investment-grade bonds. However, this

finding was insignificant in the prediction of high-yield credit spreads. Alexander and

Kaeck (2007) concentrated on Credit Default Swap (CDS) spreads, as more direct

measures of credit risk, and indicated that they are negatively related to interest rates

and unrelated to the yield curve slope. Schaefer and Strebulaev (2008) and Ericsson,

Jacobs and Oviedo (2009) showed that credit spreads and credit default swaps

respectively cannot be explained by some of the determinants of theoretical models

such as interest rates. Wu and Zhang (2008), through a dynamic factor analysis,

identified three macroeconomic risk factors, the inflation rate, the real output growth

and the volatility of financial markets and then they showed that positive inflation

shocks increase Treasury yields and credit spreads across all maturities and credit

rating classes. Thus, they established a positive relationship between Treasury rates

and credit spreads.

Electronic copy available at: https://ssrn.com/abstract=2569083

3. Data and methodology

This analysis employs monthly data for the 1- and 10-year constant maturity U.S.

Treasury bonds and the Moody’s Baa seasoned bond index. We consider the 1-year

Treasury rate to be the short-term interest rate, and this is a compromise between the

2-year Treasury rate and the 3-month bill yield employed in previous studies.4 We

selected the 10-year constant maturity Treasury bond rate as the long-term interest

rate, basically due to its relatively long history. The Moody’s index considered in the

analysis is created from an equally weighted sample of yields on 75-100 long maturity

(more than 20 years) US Baa-rated bonds issued by large non-financial corporations.

As Davies (2004) notes, the average duration of the corporate bonds is expected to be

close to 10 years; this is justified on the grounds that due to the default risk embedded

in corporate bonds, their duration is actually shorter compared to Treasuries of similar

maturity. For the purposes of this study we choose to concentrate only on the spread

of Baa bonds since this is at the bottom of the investment-grade category, and the

inclusion of the Aaa spread would probably blur its stochastic properties. The dataset

spans the period from January 1960 to September 2010, and was obtained from the

Federal Reserve Board (Federal Reserve Statistical Release H.15-Selected Interest

Rates).

The methodological framework followed in this paper consists of two stages. In

the first stage, we apply cointegration analysis, using the multivariate estimation

technique developed by Johansen (1995). Specifically, we assume that the data

generating process of the I(1) stochastic variables is a Gaussian vector autoregressive

model of finite order k, VAR(k), which can be expressed in a vector error-correction

model (VECM) form as:

Electronic copy available at: https://ssrn.com/abstract=2569083

 k −1
∆z t = µ + ∑i =1 Γι ∆z t −i +Π z t −1 + ε t , t = 1,....T . (1)

Where: zt is a (nx1) vector of endogenous variables (i.e. the yields on the Baa bond

index, the 1- and 10-year Treasuries), µ is the deterministic term, and εt is a (nx1)

multivariate normal random error with mean zero and a time independent variance-

covariance matrix. The tests for cointegration involve the estimation of the rank, r, of

Π since this is equal to the number of cointegrating vectors. If 0<r<n, then there are r

stationary linear combinations of the elements of z, and n-r non-stationary common

stochastic trends. In this case, there exist (nxr) matrices α and β, such that:

Π = αβ ' (2)
where a is the adjustment coefficients matrix, and β is the matrix of the cointegration

vectors.

Hansen and Johansen (1993) have suggested methods to examine the parameter

constancy in cointegrated VAR models. Specifically, three tests have been

constructed under the two VAR representations; in the “Z-representation”, all

parameters of the cointegrated VAR system are re-estimated during the recursions,

while under the “R-representation”, only the long-run parameters are re-estimated.

The first test, called the rank test, gives a sequence of trace statistics obtained from

the recursive estimation of the model, scaled by the appropriate critical value (5% in

our case). A second test deals with the null hypothesis of constancy of the

cointegration space for a given cointegration rank. Hansen and Johansen (1993)

proposed a likelihood ratio test that is constructed by comparing the likelihood

function from each recursive sub-sample with the likelihood function from the full

sample. The third test examines the constancy of the individual elements of the

cointegrating vectors.

10

Electronic copy available at: https://ssrn.com/abstract=2569083

 In the second stage of the analysis, we shift our attention to the short-run

dynamics of our model and explore them by employing the MS-VECM introduced by

Krolzig (1997). In the MS-VECM developed by Krolzig (1997), the regime

generating process is assumed to be an ergodic Markov chain with a finite number of

states, st є {1,…,M}, defined by the transition probabilities:

M
pij = Pr( st +1 = j / st = i ), ∑p ij =1 ∀i, j ∈ {1,..., M }. (3)
j =1

The cointegrated MS-VAR can be expressed in a vector error-correction model

(VECM) form as:

k −1
∆zt = µ ( st ) + ∑ i =1 Γi ( st )∆zt −i +α ( st ) β ′zt −1 + ε t , t = 1,....T (4)

where εt ~ NIID (0, Σ(st)), and α and β are defined in Eq. (2). In this paper, we

estimate a MS-VECM with 2 regimes and 3 lags, allowing for regime shifts in the

intercept, the autoregressive terms, Γ, and the variance-covariance matrix. The so-

called MSIAH(2)-VECM(3) can be written as:

 µ1 + Γ11∆zt −1 + Γ12 ∆zt − 2 + Γ13∆zt −3 +

 +α ECT1 + α ECT2 + u , if s = 1
 11 t −1 t −1
∆zt = 
12 1t t
(5)
µ
 2 + Γ 21∆zt −1 + Γ 22 ∆zt −2 + Γ 23 ∆z t −3 +
+α 21ECT1t −1 + α 22 ECT2t −1 + u2t , if st = 2

where the two error correction terms, ECT1 and ECT2, represent the long-run

equilibrium relationships identified by the Johansen cointegration analysis.

Saikkonen and Luukkonen (1997) stated that the Johansen cointegration technique is

consistent even if the data generating process is a non-Gaussian VAR. The remaining

parameters of the MS-VECM are estimated using an Expectation-Maximization (EM)

algorithm (Krolzig, 1997).

11

Electronic copy available at: https://ssrn.com/abstract=2569083

4. Empirical results

4.1 Cointegration analysis

The analysis starts with the computation of a battery of unit root tests, including

the Augmented Dickey-Fuller (ADF) and Phillips-Perron (PP) tests, the DF-GLS test

developed by Elliott, Rothenberg, and Stock (1996), as well as the KPSS test by

Kwiatkowski, Phillips, Schmidt, and Shin (1992). The test results, presented in Table

1, indicate that we are unable to reject the unit root hypothesis for the levels of the

series at conventional levels of significance. In contrast, the unit root hypothesis is

rejected for the differenced series.

[Insert Table 1 here]

Next, we apply the Johansen (1995) maximum likelihood procedure in a VAR for

zt=[Baa, Tcm10y, Tcm1y]t. Starting with a maximum lag length of six, we ended up

with a model with four lags and a constant restricted in the cointegration space, on the

basis of conventional information criteria and test statistics.5 Table 2 presents the

trace and maximum eigenvalue test statistics along with the critical values at the 95%

significance level. The evidence supports the existence of two cointegrating vectors,

at the 95% significance level, and therefore of one common stochastic trend.6 We

proceeded with the identification of the system by imposing two restrictions on each

cointegrating vector, so that the vectors are expressed by (1, -1, 0) and (0, 1, -1),

among the variables Baat, Tcm10yt, and Tcm1yt. As depicted in Table 2, the null

hypothesis could not be rejected at conventional levels of significance (p-value of

0.351). Thus, we provide evidence on the existence of cointegration between the Baa

corporate and long-term Treasury rates, and between long- and short-term Treasury

rates, with corresponding cointegrating vectors of (1, -1).7 The restricted constant

12

Electronic copy available at: https://ssrn.com/abstract=2569083

term in the first cointegrating vector is interpreted as a constant risk premium

coefficient, and its estimated value of 1.93 percent almost equals the sample average

of the credit spread. The second cointegrating vector implies stationarity for the

spread between the 10- and 1-year Treasury bond yields, with an estimated term

premium of 0.96 percent, as compared to the sample term premium average of 0.86

percent. This finding is consistent with the long-run implication of the Expectations

Hypothesis of the term structure and it has been replicated by many other studies,

particularly for the short end of the yield curve.

[Insert Table 2 here]

The short-run dynamics of the estimated cointegrated model, presented in Table

2, indicate that the first cointegration relation (i.e. the credit spread), is statistically

significant only in the equation for ∆Tcm1y. The estimated adjustment coefficient

implies that an increase in the credit spread is negatively associated with the 1-year

yield. This result might be interpreted as a reaction of monetary policy to signs of

rising defaults and/or the presence of liquidity squeezes. The second cointegration

relation (i.e. the term spread), is statistically significant in the equation for ∆Baa and

marginally so for the ∆Tcm10y, with the point estimates of the adjustment coefficients

indicating that a steeper yield curve is associated with a declining credit spread.

However, since the two estimates are not statistically different from each other, it is

safer to conclude that the credit spread remains intact.

In order to examine the reaction of the credit spread to interest rate shocks, we

compute impulse response functions from the estimated VECM(3) by imposing the

cointegrating vectors identified before (see Mills, 1999). The chosen

13

Electronic copy available at: https://ssrn.com/abstract=2569083

orthogonalization of the variance-covariance matrix implies that a shock in the 1-year

Treasury yield has a contemporaneous impact on both the corporate rate and the 10-

year Treasury yield, while on the other end a shock on corporate yields has no

contemporaneous impact on government yields. Figure 1 displays the cumulative

response of the credit spread series to a one hundred basis points (b.p.) shock in the

10-year Treasury yield. It is shown that, initially, the Baa variable rises to a smaller

extent than the Treasury yield, which implies that the Baa spread declines. Gradually,

the initial decline becomes smaller, but even after 40 months the spread is still 20 b.p.

lower than its pre-shock level.8

[Insert Figure 1 here]

The cointegration analysis ends with the presentation of the Hansen and Johansen

(1993) stability tests of the cointegration results. Figure 2a illustrates that the rank of

the cointegration space is dependent on the sample used for its estimation.

Specifically, it provides clear evidence of two cointegrating vectors only after 1993.

In light of this finding, we re-estimated our model for the period up to December

1990, and failed to reject the null hypothesis for the existence of a single cointegrating

vector, which was identified as the term spread. These empirical results support the

considerations that the U.S. corporate bond market developed substantially over the

last 20 years, which might explain the lack of cointegration between corporate and

government bond yields on samples extending up to early 1990s. This is further

supported by evidence displayed in Figure 2b suggesting that the constancy of the

cointegration space is established only after 2000. Figure 2c shows that the first

eigenvalue, associated with the yield curve slope, underwent a structural break in the

early 1980s.9 Similar results, in another context, are reported in a number of studies

14

Electronic copy available at: https://ssrn.com/abstract=2569083

on the term structure of the U.S. government bond market.10 The second eigenvalue

exhibits a more stable behavior, particularly after 1982. Overall, the stability tests

indicate that the second cointegrating vector appears in the 1990s, and that the

cointegration space experiences a structural break in the early 1980s.

[Insert Figure 2 here]

4.2 Markov-switching analysis

On the basis of the bottom-up specification strategy of Krolzig (1997), the null

hypotheses stating that the intercept, the variance-covariance matrix and the

autoregressive parameters are not regime dependent were all rejected at the 5% level

(see Table 3).

[Insert Table 3 here]

According to the estimation results presented in Table 4, the log-likelihood value

obtained from the MSIAH(2)-VECM(3) is higher than that from the linear system,

whereas the null hypothesis of linearity is strongly rejected according to the non-

standard likelihood ratio test of Davies (1987).11 In addition, the AIC, HQ, and SC

information criteria are all in favor of the MSIAH(2) - VECM(3). The standard

deviations of the equations in regime 1 are substantially larger than those of regime 2;

thus, we identify regime 1 as the high volatility regime, that lasts approximately 5.15

months, and regime 2 as the low volatility regime whose duration is roughly 17.69

months. The transition probabilities p11 and p22, which express the possibility of

regime clustering, indicate that there is an 80 percent probability that a high volatility

15

Electronic copy available at: https://ssrn.com/abstract=2569083

regime will be followed by a similar one, while the corresponding probability is 94

percent for the low volatility period.

[Insert Table 4 here]

Figure 3 provides regime classification information as this is expressed by the

smoothed probabilities of being in the high and low volatility regimes. Regime 1,

defined as the high volatility regime, covers several economic events such as the

Federal Reserve’s “new operating procedures” from 1979 to 1982 and the subprime

mortgage crisis of 2007-1010, as well as the periods of 1973-1975 and 2001-2002 that

have been defined as US recessions by the NBER. In addition, regime 1 is shown to

have prevailed during the 1970s and early 1980s, which correspond to times of high

and persistent inflation.

[Insert Figure 3 here]

Ang and Bekaert (2002) proposed the computation of a statistic, known as the

regime classification measure (RCM), so as to evaluate the quality of the regime

classification produced by the regime-switching model. The RCM statistic is defined

as RCM ( M ) = 100 M M (1 / T )∑Tt=1 (∏ Mj=1 p j ,t ) , with M denoting the number of regimes,

T the number of observations, and pj,t the smoothed probabilities of being in regime

j=1,…M. They argued that if the classifying ability of the model is almost perfect,

then the smoothed probabilities of observing one of the regimes ought to be close to

unity and the RCM close to zero. In our case, the value of the RCM statistic was found

to be 16.66, which highlights the good classifying ability of our model.

16

Electronic copy available at: https://ssrn.com/abstract=2569083

 As a final verification of the relative performance of our non-linear model in

comparison to the linear one, we perform forecasting exercises for the out-of-sample

period from January 2009 to September 2010.12 Table 5 reports the Mean Absolute

Prediction Error (MAPE) and the Root Mean Squared Prediction Error (RMSPE) for

one-step prediction errors, whereby it is shown that the MSIAH(2)-VECM(3)

outperforms the VECM(3), in terms of its forecasting ability, for one of the three

series and has almost the same predictive power for the other two.

[Insert Table 5 here]

In accordance with Clarida et al. (2006), we provide evidence on the existence of

a linkage between the regime switching mechanism and a number of macroeconomic

variables. Specifically, we estimate a logit model which involves the construction of a

binary dependent variable that takes the value of one, or zero, when the probability of

being in the high volatility period is higher, or smaller, than 0.5. The explanatory

variables used are the output and inflation gaps, as measured by the deviation of the

industrial output and the inflation rate from their respective Hodrick-Prescott trends.

The results of the logit estimation are reported in Table 6, which depicts that de-

trended inflation is the most important factor in explaining the dynamics of the

switching mechanism. The sign of the estimated parameter is positive, thus indicating

that the probability of being in the high volatility regime is higher when inflation is

above its estimated trend.13 The output gap is also found to be statistically significant,

being related to the probability of being in the high volatility regime in a negative

way. This is not surprising, since the two major time intervals in the estimated high

volatility regime coincide with the recessions of 1973-1975 and 1981-1982.

17

Electronic copy available at: https://ssrn.com/abstract=2569083

 [Insert Table 6 here]

The estimated adjustment coefficients of the MSIAH(2)-VECM(3) are reported in

Table 4. Under both regimes, the first error-correction term (i.e. the credit spread), is

not found to be statistically significant. The second error-correction term (i.e. the

yield curve slope), is statistically significant in the high volatility regime, in the

equations for ∆Baa and ∆Tcm10y, with estimated coefficients of -0.067 and -0.106,

respectively. In the low volatility regime, the term spread is significant only in the

equation for ∆Baa, with a substantially smaller, in absolute values, adjustment

coefficient of -0.021. These results confirm previous empirical evidence documented

by, among others, Ang and Bekaert (2002), which shows a larger correlation, in

absolute values, between short rates and term spreads during high volatility periods.

The point estimates of the term spread variable, under regime 1, in the equations for

∆Baa and ∆Tcm10y indicate that an increase of the term spread is associated with

wider credit spreads. However, since the estimates are not significantly different from

each other it is safer to conclude that the credit spread remains unchanged. During the

low volatility period, the yield curve slope does not affect the dynamic behavior of the

1- or the 10-year Treasury bond yields, at the 95% level. In contrast, this variable is

inversely related to Baa rate changes, so that a steeper yield curve reduces the credit

spread. This finding is in accordance with the Expectation Hypothesis, which

associates an increasing slope to improving economic conditions.

The regime switching analysis is concluded by studying, by means of an impulse

response analysis, the reaction of credit spread to term spread changes caused by

shocks in the 1- and the 10-year Treasury rates. For each one of the two regimes

18

Electronic copy available at: https://ssrn.com/abstract=2569083

estimated in eq. (5) we apply the methodology of Ehrmann et al. (2003) for the

production of regime dependent impulse responses whereas a Choleski decomposition

of the variance-covariance matrix, similar to the one in section 4.1, ensures its exact

identification. Figure 4 illustrates the cumulative response of the Baa credit spread to

a one hundred basis points interest rate shocks. We observe a negative reaction of the

credit spread to short-term Treasury rate shocks, under both regimes (Fig. 4A, 4B).

This implies that a flattening of the yield curve, caused by an increase in short rates,

leads to a reduction of the credit spread which contradicts the long-standing view that

the slope of the yield curve is a leading indicator of future economic conditions.

According to this argument, and neglecting liquidity premiums, we would expect the

slope of the yield curve and the credit spread to be inversely related. In a similar vein,

tightening of the liquidity conditions, exemplified by the increase of the short run

interest rates, should be accompanied by an increase in credit spreads. In the case

when the flattening of the yield curve is being caused by a negative shock in the 10-

year Treasury rate, the results are consistent with the Expectations Hypothesis under

which a flattening of the yield curve signifies a deteriorating economic environment

(Fig. 4C, 4D).

[Insert Figure 4 here]

These results are in agreement to the evidence from the long run cointegration

analysis of the estimated VECM in section 4.1. However the lack of statistical

significance of the impulse responses makes it unsafe to draw a firm conclusion about

the short run relationship between Treasury rates and credit spreads.

5. Conclusion

19

Electronic copy available at: https://ssrn.com/abstract=2569083

 This paper attempts to re-examine the relation between corporate and government

bond rates which stands at the heart of some of the most well-known models for

corporate debt pricing. We depart from the existing empirical literature, since we

focus solely on the relationship between corporate and long/short Treasury rates over

both a long- and short-run horizon, as well as within a Markov-switching

environment.

Our analysis yielded the following results. First, the estimated VECM identified a

one-to-one long-run relationship between Baa-rated corporate bond and the 10-year

Treasury yields. However, the short-run analysis, by means of impulse responses,

provided weak support to structural models since a negative reaction of credit spreads

to long term Treasury rate changes was found. We also failed to establish the

Treasury yield curve slope as an additional determinant of the credit spread. Overall,

our results are in broad agreement with the existing literature and highlight the

importance of the chosen sample period and the term of the risk-free interest rates for

the interpretation of the differences observed in the empirical research.

Second, we pursued a Markov-switching approach in order to study the dynamics

of our model allowing for the presence of different regimes. The obtained evidence

suggested that, in the high volatility period, the yield curve slope does not affect the

credit spread. However, this was reversed in the low volatility period where a

relationship consistent with the Expectation Hypothesis of the term structure was

found to exist. We also analyzed, by means of an impulse response analysis, the

reaction of the credit spread to shocks in both short- and long-term Treasury rates. We

validated that the credit spread variable reacts asymmetrically to Treasury yield curve

slope changes according to the source of the shock that has caused the slope change.

20

Electronic copy available at: https://ssrn.com/abstract=2569083

 Overall, our empirical findings stress the importance of treating the interest rate -

credit spread relationship as a dynamic one that varies both over time and according

to the prevailing business cycle fluctuations. To the extent that theoretical models of

credit spread dynamics capture patterns of behavior observed in the data, our evidence

compromises a number of conflicting results found in the relevant literature.

21

Electronic copy available at: https://ssrn.com/abstract=2569083

Endnotes
1
Throughout this study we refer interchangeably to credit spread and corporate –
Treasury yield spread, although a number of papers refute their close association and
assign to credit risk a minor importance as an explanatory factor of the latter variable
(see Elton et al., 2001; Collin-Dufresne et al., 2001, Longstaff et al., 2005; Dionne et
al., 2004).
2
It should be mentioned that Davies (2004) and (2008) measured the reaction of the
credit spread series in relation to the 2-year Treasury rate in the first study and the 3-
month Treasury bill rate in the second one. Our results are more comparable to those
of Neal et al. (2000) who employed a sample from 1960 to 1997 and measured the
reaction of the credit spread in relation to the 10-year constant maturity Treasury
series. The overall evidence stresses the importance of the chosen sample period and
risk free interest rate for the interpretation of the divergent results in the literature.
3
It has also been found that higher-rated spreads tend to revert much faster to their
long-term mean than lower-rated spreads (Prigent et al., 2001).
4
The 2-year Treasury bond yield is used by, among others, Collin-Dufresne et al.
(2001) and Davies (2004), the 3-month bill yield by Duffee (1998) and Davies (2008)
and the Merrill Lynch 1-3 years Treasury index yield by Bierens et al. (2003).
5
The Lagrange Multiplier test for autocorrelation of order four (one) has a marginal
significance level equal to 7% (20%).
6
Conditional on the existence of two cointegrating vectors, we tested for the
exclusion of each one of the three variables from the cointegration space. In each
case, we easily rejected, at conventional levels of significance, the null hypothesis that
the respective variable is excluded. We also tested the variables for weakly
exogeneity, and failed to reject the null hypothesis of exogeneity for the 1-year
Treasury yield only.
7
Bevan and Garzarelli (2000) used a similar data basis to ours, except for the 1-year
Treasury yield, for the period from January 1960 to December 1999, within an
autoregressive distributed lags (ARDL) approach. They showed that, in the long-run,
a 100 basis point increase in the level of the Treasury yield increases the Baa credit
spread by 4 basis points.
8
We apply a one hundred basis points shock and present cumulative reactions in
order to preserve comparability with the results on impulse responses in the next
section and have a clear connection with the findings from the long run cointegration
analysis respectively. Standard errors are not reported since they indicate, as it is
common with this methodology, that the responses are not statistically different from
zero. Our focus therefore is more on the type of the reactions, positive or negative,
and not on their quantitative measurement.
9
This is based on the evidence that the lower bounds of the estimated eigenvalues
before 1982 are larger than the upper bounds of their values after that date.

22

Electronic copy available at: https://ssrn.com/abstract=2569083

10
For example, Hansen (2003) found a structural break in the 1979-1982 period, and
after taking account of this break in the dynamics of the VECM he finds constant
cointegrating vectors that satisfy the Expectations Hypothesis.
11
The distribution of the LR is likely to be different from the adjusted χ2distribution
proposed by Davies (1987). This is due to the violation of the regularity conditions
under which this test is valid.
12
Forecasts for period t are calculated as follows:
M
Et −1 ( zt ) = zt −1 + Et −1 (∆zt ) = zt −1 + ∑ pij ∆zt , where pij and ∆zti are given in Eqs. (3)
j =1

and (4).
13
Other researchers have found similar results. For instance, David (2008) established
the usefulness of the inflation rate, and the short nominal rate, as predictors of
business cycles. Davies (2008) also highlights the usefulness of the inflation rate as a
regime differentiating variable in a Self-Extracting Threshold model. Wu and Zhang
(2008), in a different setting, have identified inflation rate and real output growth as
the main macroeconomic risk factors.

23

Electronic copy available at: https://ssrn.com/abstract=2569083

References

Alexander, C., Kaeck, A., 2007. Regime dependent determinants of credit default
swap spreads. Journal of Banking and Finance 32, 1008-1021.
Ang, A., Bekaert, G., 2002. Regime switches in interest rates. Journal of Business &
Economic Statistics 20, 163-182.
Bevan, A., Garzarelli, F., 2000. Corporate bond spreads and the business cycle:
Introducing GS-Spread. The Journal of Fixed Income 9, 8-18.
Bhanot, K., 2005. What causes mean reversion in corporate bond index spreads? The
impact of survival. Journal of Banking & Finance 29, 1385–1403.
Bierens, H., Huang, J.-Z., Kong, W., 2003. An econometric model of credit spreads
with rebalancing, ARCH and jump effects. Working paper, Penn State University.
Clarida, R., Sarno, L., Taylor, M., Valente, G., 2006. The role of asymmetries and
regime shifts in the term structure of interest rates. Journal of Business 79, 1193-
1224.
Collin-Dufresne, P., Goldstein, R., Martin, S., 2001. The determinants of credit spread
changes. Journal of Finance 56, 2177-2207.
David, A., 2008. Inflation Uncertainty, Asset Valuations, and the Credit Spread
Puzzle. The Review of Financial Studies 21(6), 2487-2534.
Davies, A., 2008. Credit spread determinants: An 85 year perspective. Journal of
Financial Markets 11, 180–197
Davies, A., 2004. Credit spread modeling with regime-switching techniques. The
Journal of Fixed Income 14, 36-48
Davies, R., 1987. Hypothesis testing when a nuisance parameter is present only under
the alternative. Biometrika 74, 33–43
Dionne, G., Gauthier, G., Hammami, K., Maurice, M., Simonato, J.-G., 2004. Default
risk on corporate yield spreads. Working paper, HEC Montreal.
Duffee, G., 1998. The relation between Treasury yields and corporate bond yield
spreads. Journal of Finance 53, 2225-2241.
Duffie, D., Singleton, K., 1999. Modeling term structures of defaultable bonds. The
Review of Financial Studies 12, 687-720.
Ehrmann, M., Ellison, M. and Valla, N., 2003. Regime Dependent Impulse Response
Functions in a Markov Switching Vector Autoregression Model. Economics
Letters 78(3), 295-299.
Elliott, G., Rothenberg, T., Stock, J., 1996. Efficient tests for an autoregressive unit
root. Econometrica 64, 813-836.
Elton, E., Gruber, M., Agrawal, D., Mann, C., 2001. Explaining the rate spread on
corporate bonds. Journal of Finance 56, 247-277.
Ericsson, J., Jacobs, K. and Oviedo, R.. 2009. The Determinants of Credit Default
Swap Premia. Journal of Financial and Quantitative Analysis. 44(1), 109-132.
Estrella, A., 1998. A new measure of fit for equations with dichotomous dependent
variables. Journal of Business and Economic Statistics 16, 198-205.
________ and Trubin, M., 2006. The Yield Curve as a Leading Indicator: Some
Practical Issues. Current Issues in Economics and Finance, Federal Reserve Bank
of New York, Vol. 12, No. 5, July/August 2006.
Hansen, H., Johansen, S., 1993. Recursive estimation in cointegrated VAR-models.
Working paper, University of Copenhagen, Institute of Mathematical Statistics.
Hansen, P.R., 2003. Structural changes in the cointegrated vector autoregressive
model. Journal of Econometrics 114, 261-295.
Johansen, S., 1995. Likelihood-Based Inference in Cointegrated Vector

24

Electronic copy available at: https://ssrn.com/abstract=2569083

 Autoregressive Models. Oxford University Press, Oxford.
Krolzig, H.-M., 1997. Markov-Switching Vector Autoregressions. Modelling,
Statistical Inference and Application to Business Cycle Analysis. Springer,
Berlin.
Kwiatkowski, D., P. Phillips, P. Schmidt, Shin, Y., 1992. Testing the null hypothesis
of stationarity against the alternative of a unit root. Journal of Econometrics 54,
159-178.
Longstaff, F., Mithal, S., Neis, E., 2005. Corporate yield spreads: Default risk or
liquidity? New evidence from the Credit-Default Swap market. Journal of
Finance 60, 2213-2253.
Longstaff, F., Schwartz, E., 1995. A simple approach to valuing risky fixed and
floating rate debt. Journal of Finance 50, 789-819.
Merton, R., 1974. On the pricing of corporate debt: The risk structure of interest rates.
Journal of Finance 29, 449-470.
Mills, T., 1999. The Econometric Modelling of Financial Time Series. Cambridge
University Press.
Neal, R., Rolph, D., Morris, C., 2000. Interest rates and credit spread dynamics.
Working paper, Federal Reserve Bank of Kansas City.
______, Rolph, D., Dupoyet, B., Jiang, X., 2012. Interest rate and Credit Spread
Dynamics. Working paper, Indiana University.
Osterwald-Lenum, M., 1992. A note with quantiles of the asymptotic distribution of
the maximum likelihood cointegration rank test statistics. Oxford Bulletin of
Economics and Statistics 54, 461-472.
Pedrosa, M., Roll, R., 1998. Systematic risk in corporate bond credit spreads. Journal
of Fixed Income 8, 7-26.
Prigent, J.-L., Renault, O., Scaillet, O., 2001. An empirical investigation in credit
spread indices. Journal of Risk 3, 27-55.
Saikkonen, P., Luukkonen, R., 1997. Testing cointegration in infinite order vector
autoregressive processes. Journal of Econometrics 81, 93-126.
Schaefer, S. Strebulaev, I., 2008. Structural models of Credit Risk are useful:
Evidence from Hedge Ratios on Corporate Bonds. Journal of Financial
Economics, 90, 1-19.
Stock, J., 1997. Cointegration, Long-Run Comovements, and Long-Horizon
Forecasting. In: Kreps D & Wallis KF (eds.) Advances in Econometrics:
Proceedings of the Seventh World Congress of the Econometric Society, vol.3.
Cambridge University Press, Cambridge, pp. 34-60.
Wu, L., Zhang, F. X., 2008. A No-arbitrage Analysis of Macroeconomic
Determinants of the Credit Spread Term Structure. Management Science, 54,
1160-1175.

25

Electronic copy available at: https://ssrn.com/abstract=2569083

Appendix

Table 1
Unit root tests for the levels and first differences of the series. 1960:1-2010:9.
ADF test PP test DF-GLS test KPSS test
Baa -1.44 -1.30 -0.84 3.35*
Tcm10y -1.59 -1.51 -1.22 2.88*
Tcm1y -1.68 -1.93 -1.59 1.18*

∆Baa -11.38* -14.39* -10.60* 0.44

∆Tcm10y -12.76* -16.64* -12.61* 0.21
∆Tcm1y -10.24* -15.39* -8.88* 0.09
Note: ADF stands for the Augmented Dickey-Fuller t-test, and PP for the Phillips-Perron test. DF-GLS
denotes the modified Dickey-Fuller test developed by Elliott, Rothenberg, and Stock (1996). The ADF,
PP, and DF-GLS tests consider the null hypothesis of a unit root. KPSS denotes the test suggested by
Kwiatkowski et al. (1992), which considers the null hypothesis of level stationarity. Lag length is
determined on the basis of the Schwarz Criterion. (*) denotes statistical significance at the 5% level.

Table 2
Johansen maximum likelihood cointegration results. 1960:1-2010:9
H0:r p-r Trace Trace*
0 3 64.815* 63.575*
1 2 24.209* 23.026*
2 1 2.672 2.560

Long-run cointegration relationships

Baa Tcm10y Tcm1y Constant
General 1.000 -3.467 2.301 1.383
-0.825 1.000 -0.141 1.226
Restricted 1.000 -1.000 0.000 -1.929
0.000 1.000 -1.000 -0.959
Q(2) = 2.094 (p-value = 0.351)

Adjustment coefficients
ECT1 ECT2
∆Baa -0.014 (-1.190) -0.033 (-3.729)
∆Tcm10y 0.004 (0.246) -0.027 (-2.226)
∆Tcm1y -0.057 (-2.341) 0.026 (1.419)
Trace* stands for the Bartlett-corrected trace test. * denotes statistical significance at the 5% level, on
the basis of the Osterwald-Lenum (1992) critical values. Q denotes a likelihood ratio test for the over-
identifying restrictions appearing in the restricted model. It is distributed as a χ2 with the corresponding
degrees of freedom given in parentheses. ECT1 (ECT2) denotes the cointegrating relationship related to
the credit spread (the term spread). T-values are reported in parentheses.

26

Electronic copy available at: https://ssrn.com/abstract=2569083

Table 3
Bottom-up specification strategy
Restricted model Unrestricted model LR1
340.632 380.655 64.64*

LR2
340.655 676.837 672.364*

LR3
676.837 707.850 62.026*
LR1 is a test statistic of the null hypothesis of no regime dependent intercept. LR2 is a test statistic of
the null hypothesis of no regime dependent variance-covariance matrix. LR3 is a test statistic of the null
hypothesis of no regime dependent autoregressive parameters. The test statistics are calculated as
2(lnL-lnLr), where L and Lr denote the unrestricted and restricted maximum likelihood respectively. (*)
denotes statistical significance at the 5% level.

Table 4
Estimation results for the MSIAH(2)-VECM(3), 1960:1-2010:9
MSIAH(2)-VECM(3) Linear VECM(3)
Log-likelihood 707.850 340.632
AIC criterion -2.042 -0.981
HQ criterion -1.799 -0.862
SC criterion -1.419 -0.676
LR linearity test = 734.435 [0.00]*

Regime 1 Regime 2
Duration (months) 5.15 17.69
Unconditional probability 0.23 0.77
p11 = 0.806
p22 = 0.943

Regime-dependent equilibrium-correction coefficient estimates

Regime 1 ∆Baat ∆Tcm10yt ∆Tcm1yt
Constant 0.101 (1.171) -0.104 (-0.899) 0.149 (0.7441)
ECT1t-1 -0.019 (-0.492) 0.061 (1.186) -0.080 (-0.8833)
ECT2t-1 -0.067 (-2.166) -0.106 (-2.563) 0.008 (0.1124)
Std Error 0.306 0.405 0.713

Regime 2 ∆Baat ∆Tcm10yt ∆Tcm1yt

Constant 0.041 (2.173) -0.010 (-0.363) 0.030 (1.019)
ECT1t-1 -0.014 (-1.113) 0.021 (1.189) -0.022 (-1.150)
ECT2t-1 -0.021 (-2.445) -0.017 (-1.383) 0.021 (1.597)
Std Error 0.135 0.194 0.207
*Probability values in squared brackets (Davies, 1987). p11 (p22) stands for the transition probability of
remaining in regime 1(2). To conserve space, only the estimates of the adjustment coefficients of the
two error correction terms are reported. ECT1 (ECT2) denotes the cointegrating relationship of the
credit spread (term spread). T-values in parentheses.

27

Electronic copy available at: https://ssrn.com/abstract=2569083

Table 5
Forecast analysis for the period 2009:1 to 2010:9
Baa Tcm10y Tcm1y
MAPE
MSIAH(2)-VECM(3) 0.1624 0.1642 0.0635
Linear VECM(3) 0.1497 0.1603 0.1296

RMSPE
MSIAH(2)-VECM(3) 0.1897 0.1880 0.0763
Linear VECM(3) 0.1752 0.2098 0.1769
MAPE stands for the Mean Absolute Prediction Error. RMSPE stands for the Root Mean Squared
Prediction Error.

Table 6
Logit estimation results. 1960:1-2010:9
Variables Coefficients T-Statistics
Constant -1.446 -13.220
x−x -0.307 -3.520
π −π 0.777 6.719
Pseudo-R2 = 0.089
The binary dependent variable is equal to one when the probability of being in the high volatility
period is greater than 0.5, and equal to zero when the corresponding probability is smaller than 0.5.
The explanatory variables are a constant, the output gap x − x , and the inflation gap π − π . Pseudo-R2
denotes Estrella’s (1998) measure of goodness of fit for logit models.

28

Electronic copy available at: https://ssrn.com/abstract=2569083

 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39

-0.1

-0.2

-0.4

-0.5

-0.6

Baa credit spread

Fig. 1. Cumulative monthly impulse response of the credit spread to a one hundred
basis points shock in the 10-year Treasury bond yield.

29

Electronic copy available at: https://ssrn.com/abstract=2569083

 (a)
a)

(b)
a)

(c)
a)

Fig. 2. Cointegration stability tests. (a) The trace test. (b) The constancy of the
cointegration space test. (c) The eigenvalue test.

30

Electronic copy available at: https://ssrn.com/abstract=2569083

 Probabilities of Regime 1
1.00
filtered
smoothed
0.75

0.50

0.25

1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
Probabilities of Regime 2
1.00
filtered
smoothed
0.75

0.50

0.25

1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010

Fig. 3. Regime probabilities of the MSIAH(2)-VECM(3). The first figure refers to the
high volatility regime, and the second one to the low volatility regime.

31

Electronic copy available at: https://ssrn.com/abstract=2569083

(A) One unit positive shock in the 1-year Treasury rate - high volatility regime

0,1
0
-0,1 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
-0,2
-0,3
-0,4
-0,5
-0,6
-0,7
-0,8
-0,9

Baa_spread term-spread

(B) One unit positive shock in the 1-year Treasury rate - low volatility regime

0
-0,1 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
-0,2
-0,3
-0,4
-0,5
-0,6
-0,7
-0,8

Baa_spread term_spread

( C) One unit negative shock in the 10-year Treasury rate - high volatility regime

0,5

0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
-0,5

-1

-1,5

Baa_spread term_spread

(D) One unit negative shock in the 10-year Treasury rate - low volatility regime

0,6
0,4
0,2
0
-0,2 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
-0,4
-0,6
-0,8
-1
-1,2
-1,4

Baa_spread term_spread

Fig. 4. Cumulative impulse response of the credit and term-spreads

32

Electronic copy available at: https://ssrn.com/abstract=2569083