Page 43 - 臺大管理論叢第32卷第1期

NTU Management Review Vol. 32 No. 1 Apr. 2022

corresponding credit rating data from the Fixed Investment Securities Database (FISD).
The corporate bond yield data collected from TRACE covers 7,388 reference entities.
In order to match with the CDS data, we use Fuzzy string matching algorithm to merge
these two different data sets. We obtain the final data of 1,074 entities, which possess both

CDS market quotations and corporate bond yields simultaneously.
Regarding the hazard rates, Chen, Fabozzi, and Sverdlove (2010) have developed
a two-factor squared root process which can estimate the hazard rate and the liquidity

factor separately. They argue that their approach has several advantages. First, the model
has closed form solutions to the risk-free discount bonds and survival probabilities and
therefore the hazard rates can be computed efficiently. Secondly, the model fits credit
spreads reasonably well and is flexible enough to account for any form of dynamics in
credit spreads. However, this approach suffers a drawback that the hazard rate should be

re-estimated periodically.
Motivated by Chen et al. (2010), we estimate the hazad rate of the one-factor squared
root process designed for the credit risk model, with an uncented Kalman filter on two

different kinds of data: CDS market quotations combined with corporate bond yield
rates, and purely CDS. Since the uncented Kalman filter is composed of time update
and measurement update equations, we can have more dynamic and accurate estimates
of hazard rates. Due to the fact that CDS premium contain both the hazard rate and the
liquidity factor, we use the principal components analysis to isolate the hazard rate and

extract liquidity factors.
We start by estimating principal components of the CDS premiums across different
reference entities. This gives us six principal components. Then, the first principal

component, which accounts for 43% of total variations, is regressed on two kinds of hazard
rates estimated. Subsequently, we extract two liquidity factors by calculating the residuals
of each regression equation. One is extracted purely from CDS market quotations, and the
other is extracted from CDS market quotations combined with corporate bond yield rates.
Based on Anderson (2017), the residuals of each regression equation can be

representative of liquidity risk. Since this paper studies comovement between CDS
premium changes and explores the source of that increase, its finding reveals changes
in fundamentals (e. g. macroeconomic or market variables) account for only 23% of

the increase in covariance. The remaining increase is attributed to changes in liquidity

35