The Role of Credit Spreads and Structural Breaks in Forecasting the Term Structure of Korean Government Bond Yields

• Published date: 01 June 2015
• Date: 01 June 2015
• Author: Chang Hoon Lee,Kyu Ho Kang
• DOI: http://doi.org/10.1111/ajfs.12093

The Role of Credit Spreads and Structural

Breaks in Forecasting the Term Structure of

Korean Government Bond Yields*

Chang Hoon Lee

Department of Economics, Korea University

Kyu Ho Kang**

Department of Economics, Korea University

Received 31 December 2013; Accepted 23 June 2014

Abstract

We examine whether Korean credit spreads are informative enough to help improve the pre-

dictive accuracy of Korean government bond yields. To do this, we analyze a joint dynamic

Nelson-Siegel (DNS) model of Korean government bond yields and credit spreads. In the

model multiple change-points at unknown time points in the factor process are allowed in

order to capture the possibility of structural breaks in the yield and credit spread curve

dynamics. We find that the joint DNS model of the yield and credit spread curves outper-

forms the standard DNS model of the yield curve in terms of out-of-sample yield curve pre-

diction. Further, the predictive gains are maximized at the two change-points. The two

change-points seem to be closely associated with the beginning of the recent financial crisis

and the subsequent stabilization of Korean bond markets.

Keywords Dynamic Nelson-Siegel model; Out-of-sample forecasting; Posterior predictive cri-

terion; Bayesian MCMC simulation

JEL Classification: C11, E43, G12

1. Introduction

Default-free interest rates and corporate credit spreads are closely related through

uncertainty, financial crises, economic agents’ expectations, and business cycles.

According to the structural models for risky bond pricing proposed by Me rton (1974)

and Longstaff and Schwartz (1995), corporate credit spreads are determined by the

relation between the default-free interest rates and a firm’s asset value. If the economy

*We would like to thank the editor and two anonymous referees for their thoughtful and

helpful comments.

**Corresponding author: Kyu Ho Kang, Department of Economics, Korea University, Seoul

136-701, Korea. Tel: 82-2-3290-5132, Fax: 82-2-3290-2200, email: kyuho@korea.ac.kr.

Asia-Pacific Journal of Financial Studies (2015) 44, 353–386 doi:10.1111/ajfs.12093

©2015 Korean Securities Association 353

is in a recession, the default-free short-term interest rate decreases substantially and

the default probabilities of firms increase. Consequently, this economic dow nturn

causes a rise in the credit spreads that are the risk premium for taking the default risk.

Many studies (Duffee, 1996 1998; Collin-Dufresne et al., 2001) empirically confirm

this negative relation between risk-free interest rates and credit spreads.

1

Given the tight relation between government and corporate bond markets, infor-

mation in the credit spread curve may help improve the predictive accuracy of future

yield curves. Based on this idea, several recent studies (Christensen and Lopez, 2012;

Abdymomunov et al., 2014) suggest a joint model of the yield and credit spread

curves and examine whether credit spreads are informative in yield curve forecasting.

As noted in Christensen and Lopez (2012; henceforth, CL), the feedback effects

between the corporate bond markets and treasury bond markets are caused by the

business cycle. The business cycle affects the treasury bond markets through inves-

tors’ expectations concerning the response of central banks. It also has an influence

on the corporate bond markets because it changes the default probability of firms.

These effects are observed as changes in the Treasury yields and credit spreads.

Hence, credit markets and the interactions between the treasury bond yields and the

credit spreads of corporate bonds may help in improving the forecast accuracy of

the term structure of government bond yields since the financial crisis. The main

goal of CL is to examine whether information contained in the credit spreads can be

useful in forecasting the treasury yields. To do this, they employ a latent factor mod-

eling approach (i.e. the arbitrage-free dynamic Nelson-Siegel model, AFDNS) to cap-

ture the dynamics of the term structure of the Treasury yield curves.

2

On the other

hand, many studies have found that the credit spreads of corporate bonds tend to

be too high given the default probabilities This is the so called the credit spread puz-

zle, which holds across various rating and industry categories. CL take this as evi-

dence that systematic factors exist in credit spreads and that these factors are

common to credit spreads across business sectors and rating categories. Since the

systematic factors are unobserved they use latent factor models to extract them and

examine the feedback effects among the latent factors in the Treasury bond yields.

Specifically, they estimate the three-latent factor AFDNS model using Treasury yields

1

In particular, Duffee (1996) infers that the negative relation is mainly caused by the business

cycle.

2

The AFDNS models proposed by Christensen et al. (2011) belong to the class of affine arbi-

trage-free dynamic term structure (AFDTS) models proposed by Duffie and Kan (1996) that

impose the factorloading structure of the dynamic Nelson-Siegel (DNS) model proposed by

Diebold and Li (2006). The AFDTS and DNS models have complementary strengths and

weaknesses. The AFDTS models are theoretically rigorous (observance of arbitrage-free condi-

tion) but empirically disappointing. The DNS models are empirically successful but theoreti-

cally lacking (absence of the arbitrage-free condition). The AFDNS models take the strengths

(i.e. empirical tractability and the arbitrage-free condition) of the AFDNS and DNS models

by incorporating the DNS factor loading structure into the AFDTS models.

C. H. Lee and K. H. Kang

354 ©2015 Korean Securities Association

only via only the state-space model framework in the first step. In the second step,

they estimate the five-factor AFDNS (i.e. three yield factors and two credit spread

factors) model by combining Treasury yield factors with the credit spread factors.

Finally, they conduct the out-of-sample prediction using each model. As a result,

they find that there are two economy-wide systematic factors, interpreted as level

and slope factors, that affect credit spreads in four industry sectors. More impor-

tantly, adding information from the two systematic factors in the credit spreads

appears to improve the forecasting performance for the treasury bond yields.

Abdymomunov et al. (2014; henceforth, AKK) propose a five-factor DNS model

of the joint dynamics of the treasury yield and credit spread curves for the same

purpose with CL. In the AKK model framework, three latent factors, which are

interpreted as level, slope, and curvature factors, are assumed to explain the yield

curve dynamics. They also assume that the dynamics of the credit spread curves are

driven by two latent factors: level and slope. More importantly, they allow multiple

change-points at unknown time points in the factor process (which is assumed to

be a VAR(1) process) and impose the zero lower bound (ZLB) restriction caused

by the low short-term yield environment since the 2008 financial crisis to maximize

the yield curve predictability of their model.

3

For empirical analysis they employ a

Bayesian Markov Chain Monte Carlo (MCMC) simulation approach through the

Metropolis-Hastings algorithm. Their estimation results indicate that the joint DNS

model of the yield and credit spread curves outperforms the standard DNS and ran-

dom-walk models in forecasting the Treasury yield and credit spread curves. In

addition, the predictive gains are further enhanced when two structural breaks in

the factor process and a ZLB restriction are incorporated.

As mentioned earlier, while the AKK model and ours are extended versions of

the standard DNS model, the CL model is based on the AFDNS model, which has

theoretical advantages. Hence, the CL model provides rich structural interpretations.

However, CL do not consider the possibility of structural breaks or the ZLB. The

probability that a risk-free short rate in all models (AFDTS, AFDNS, DNS) takes a

negative value during the recent low interest rate period is substantial. In particular,

a ZLB restriction should be imposed on the density forecasting of government bond

yields in United States bond markets.

CL use the credit spread data for four rating and industry categories

4

to find

common risk factors while AKK consider the credit spread data for the aggregated

index (BAA-rated-index) to reduce the impact of industry-specific noise in the

predictions.

Since the 2008 global financial crisis the Korean corporate bond market has

matured substantially because of the increasing long-term capital needs of domestic

firms, banks’ risk management, and the widening gap between commercial bank’s

3

The treasury yields and credit spreads are assumed to follow a truncated normal distribu-

tion, TN[0,∞).

4

{Financials, Industrials, Banks, Utilities},{BBB, A, AA, AAA}.

Credit Spreads, Structural Breaks, and Yield Curves

©2015 Korean Securities Association 355