DTS (Duration Times Spread)

SM

T
A
M
A new measure of spread exposure in credit portfolios.

R
FO
Arik Ben Dor, Lev Dynkin, Jay Hyman, Patrick Houweling,

Y
Erik van Leeuwen, and Olaf Penninga

N
A
IN
ARIK BEN DOR
is an analyst, Quantitative

LE
Portfolio Strategy, at sset allocation in a portfolio or a benchmark is
Lehman Brothers in

A
C
New York City. typically expressed in terms of percentage of

TI
[email protected] market value. It is widely recognized that this
is not sufficient for fixed-income portfolios,

R
LEV DYNKIN
is a managing director, A
where differences in duration can cause two portfolios
with the same allocation of market weights to have
IS
head of Quantitative
Portfolio Strategy at extremely different exposures to macro-level risks.
TH

Lehman Brothers in As a result many fixed-income portfolio managers

New York City.
[email protected]
have become accustomed to expressing their allocations
E

in terms of contributions to duration—the product of the

C

JAY HYMAN percentage of portfolio market value represented by a

U

is a senior analyst, given market cell and the average duration of securities
D

Quantitative Portfolio
in that cell. This represents the sensitivity of the portfolio
O

Strategy at Lehman
Brothers. to a parallel shift in yields across all securities within this
R

[email protected] market cell. For credit portfolios in particular, the corre-

EP

sponding measure would be contributions to spread dura-

PATRICK HOUWELING tion, measuring the sensitivity to a parallel shift in spreads.1
R

is a senior quantitative
researcher at Robeco Asset Determining the set of active spread duration con-
TO

Management in Rotterdam, tributions to market cells or issuers is one of the primary

The Netherlands. decisions required of credit portfolio managers.Yet all
[email protected]
L

spread durations are not created equal. Just as one could

A

ERIK VAN LEEUWEN create a portfolio that matches the benchmark exactly by
G

is a senior portfolio market weights, but clearly takes more credit risk (e.g., by
LE

manager at Robeco Asset investing in the longest-duration credits within each cell),
Management in Rotterdam,
IL

The Netherlands. one could match the benchmark exactly by spread dura-
[email protected] tion contributions and still take more credit risk—by
IS

choosing the securities with the widest spreads within

OLAF PENNINGA each cell.
IT

is a senior portfolio
manager at Robeco Asset These bonds presumably trade wider than their peer
Management in Rotterdam, groups for a reason—that is, the market consensus has
The Netherlands. determined that they are more risky—and are often
[email protected]
referred to as high-beta because their spreads tend to react
DTS is a registered service mark of Lehman Brothers. more strongly than the rest of the market to a systematic

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 77

Copyright © 2007
shock. Portfolio managers are well aware of this effect, period September 1989-January 2005, we demonstrate
but many tend to treat it as a secondary effect rather than that the volatility of spread changes (both systematic and
as an intrinsic part of the allocation process. idiosyncratic) is indeed linearly proportional to spread
To reflect the view that higher-spread credits level.This relation holds for both investment-grade and
represent greater exposures to sector-specific risks, we high-yield credit, irrespective of the sector, duration, or
propose a simple risk sensitivity measure that uses spreads time period. Additional analysis (to be published later)
as a fundamental element in the credit portfolio manage- indicates that these results are not confined to the realm
ment process. We represent sector exposures by contri- of U.S. corporate bonds but also apply to European cor-
butions to Duration Times Spread (DTS), computed as the porate bonds and credit default swaps (CDX and iTraxx),
product of market weight, spread duration, and spread. regardless of the spread reference curve (e.g.,Treasury or
For example, an overweight of 5% to a market cell imple- LIBOR).This explains why relative spread volatilities of
mented by purchasing bonds with a spread of 80 basis spread asset classes are much more stable than absolute
points and spread duration of three years is equivalent to spread volatilities, both across different sectors and credit
an overweight of 3% using bonds with an average spread quality tiers and also over time.
of 50 b.p. and spread duration of eight years. The paradigm shift we advocate has many implica-
To understand the intuition behind the new measure tions for portfolio managers, both in terms of the way
we propose, consider the return due strictly to change in they manage exposures to industry and quality factors
spread—Rspread. Let D denote the spread duration of a bond (systematic risk) and in terms of their approach to issuer
and s its spread; the spread change return is then given by:2 exposures (non-systematic risk). Throughout this article,
we present evidence that the relative spread change
Rspread = − D∆s (1) approach offers increased insight into both of these sources
of risk.
It is quite easy to see that this is equivalent to

∆s ANALYSIS OF SPREAD BEHAVIOR

Rspread = − Ds (2) OF CORPORATE BONDS
s
That is, just as spread duration is the sensitivity to How should the risk associated with a particular
an absolute change in spread (e.g., spreads widen by 5 market sector be measured? Typically, for lack of any better
b.p.), DTS (Ds) is the sensitivity to a relative change in estimate, the historical volatility of a particular sector over
spread. Note that this notion of relative spread change some previous time period is used to forecast its volatility
provides for a formal expression of the idea mentioned ear- for the coming period.3
lier—that credits with wider spreads are riskier since they For this approach to be reliable, we would like to find
tend to experience greater spread changes. that these volatilities are fairly stable. Unfortunately, this
In the absolute spread change approach in Equation is not always the case.
(1), the volatility of excess returns is approximated by: As an example, Exhibit 1 shows the 36-month
trailing volatility of spread changes for various credit
σ return ≅ Dσ spread
absolute
(3)
ratings in the Lehman Brothers Credit Index between
September 1989 and January 2005. It is clear that spread
while in the relative spread change approach of Equation volatility changed dramatically over the period, declining
(2), excess return volatility follows: steadily until 1998 and then increasing significantly
through 2005. The dramatic rise in spread volatility since
σ return ≅ Dsσ spread
relative
(4)
1998 was only partially a response to the Russian crisis
and the Long-Term Capital Management debacle, as
Given that the two representations are equivalent, volatility has not reverted to its pre-1998 level.
why should one of them be preferable over the other? If the investment-grade corporate universe is instead
We provide ample evidence that the second approach, partitioned by spread levels, we find considerably more
based on relative spread changes, has an advantage due to stable spread volatilities, as seen in Exhibit 2. After an ini-
the stability of the associated volatility estimates. Using a tial shock in 1998, the volatilities within each spread bucket
large sample with over 560,000 observations over the revert almost exactly to their pre-1998 level (beginning

78 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
EXHIBIT 1
Spread Change Volatility by Credit Rating (trailing 36 months: 9/89-1/05)

Source: Lehman Brothers.

EXHIBIT 2
Spread Change Volatility by Spread Range (trailing 36 months: 9/89-1/05)

Source: Lehman Brothers.

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 79

Copyright © 2007
in August 2001, exactly 36 months after the Russian crisis illustrates that while a modest stability advantage is gained
occurred). In this respect, one could relate the results of by measuring volatility of relative spread changes, the
Exhibit 1 to an increase in spreads—both across the market improvement is not as great as we might have hoped, and
and within each quality group. the plot seems to show that even relative spread changes
As Equation (4) suggests, a potential remedy to the are quite unstable. This apparent instability, however, is
volatility instability problem is to approximate the absolute due only to the dramatic events that took place in the
spread volatility (basis points per month) by multiplying second half of 1998.
the historically observed relative spread volatility (% per When we recompute the two time series excluding
month) by the current spread level (b.p.). This can help the four observations representing the period 8/98-11/98,
stabilize the process if relative spread volatility is more the difference between the modified time series is striking.
stable than absolute spread volatility.The results in Exhibit From a low of 3 b.p.per month in mid-1997,absolute spread
2 point in this direction, and indicate a relationship volatility increases steadily through a high of 16 b.p. per
between spread level and volatility. month in 2002-2003, growing by a factor of five. Relative
Exhibit 3 plots side by side the volatility of absolute spread volatility, however, increases more modestly over the
and relative spread changes of all bonds in the Lehman same period, from 3% per month to 7% per month.
Brothers Credit index rated Baa (relative spread changes Another demonstration of the enhanced stability of
are calculated simply as the ratio of spread change to relative spreads is seen when comparing of the volatilities
the beginning-of-month spread level). The comparison of various market segments over distinct time periods.We

EXHIBIT 3
Absolute and Relative Spread Change Volatility of Baa Credit (trailing 36 months)

Actual data replaced with missing values for months that are excluded.
Source: Lehman Brothers.

80 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
have already identified 1998 as a critical turning point for Second, relative spread volatilities of various sectors
the credit markets, due to the combined effect of the are quite tightly clustered, ranging from 5% to a bit over
Russian default and the Long-Term Capital Management 10%, while the range of absolute spread volatilities is much
crisis. To what extent is volatility information prior to wider, ranging from 5 b.p. per month to more than 20
1998 relevant in the post-1998 period? b.p. per month.
In Exhibit 4, we plot pre-1998 volatility on the x Our results so far clearly indicate that absolute spread
axis, and post-1998 volatility on the y axis.We do this for volatility is highly unstable and tends to rise with increasing
two different measures of volatility:absolute spread volatility, spread. Computing volatilities based on relative spread
and relative spread volatility.4 Each point on the graph rep- change however, generates a more stable time series.These
resents a particular sector-quality cell of the Lehman findings have important implications for the appropriate
Brothers Credit Index, which is divided into eight industry way to measure excess return volatility and demonstrate
groups by three quality cells. Points along the diagonal line the need to better understand the behavior of spread
reflect identical volatilities in both time periods.5 changes.
Two clear phenomena can be observed here. First, To analyze the behavior of spread changes, we first
most of the observations representing absolute spread examine the dynamics of month-to-month changes in
volatilities are located quite far above the line, pointing to spreads of individual bonds.When spreads widen or tighten
an increase in volatility in the second period of the sample, across a sector, do they tend to shift in parallel or shift
even though the events of 1998 are not reflected in the proportionally? This key issue should determine how we
data. Relative spread volatilities, however, are quite stable, measure exposures to systematic spread changes.
with almost all observations located on the 45-degree line If spreads change in a relative fashion, then the
or very close to it.This is because the pickup in volatility volatility of systematic spread changes across a given sector
in the second period was accompanied by a similar increase of the market should be proportional to the average spread
in spreads. of that sector. This is true when comparing the risk of

EXHIBIT 4
Absolute and Relative Spread Change Volatility Before and After 1998

Source: Lehman Brothers.

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 81

Copyright © 2007
different sectors at a given time, or when examining the Higher-spread securities are riskier in that they are affected
volatility of a given sector at different times. more by a widening or tightening of spreads relative to
We also examine issuer-specific (or idiosyncratic) lower-spread securities with similar characteristics.
spread volatility. The dispersion of spread changes among To analyze the behavior of spread changes across
the various issuers within a given market cell, or the degree different periods and market segments, we estimate the
to which the spread changes of individual issuers can parameters of two models based on Equations (5) and (6).
diverge from those of the rest of the sector, also tends to The first model corresponds to the parallel shift approach
be proportional to spread. shown in Equation (5):
We investigate each of these issues using monthly
spread data from the Lehman Brothers Credit Index his- ∆si ,t = α J ,t + ε i ,t (7)
torical database.The dataset includes more than 15 years,from
The second model reflects the notion of a propor-
September 1989 through January 2005, and provides
tional shift in spreads as in Equation (6):
monthly spreads,spread changes,durations,and excess returns
for all bonds in the Credit Index. For the sections of our ∆si ,t = β J ,t ⋅ si ,t + ε i ,t
study that include high-yield bonds as well as investment- (8)
grade,we augment the dataset with historical data from the Comparing Equation (8) to Equation (6) reveals that
Lehman Brothers HighYield Index. A more detailed descrip- the slope coefficient we estimate, bJ,t corresponds to the
tion of the dataset can be found in the appendix. proportional systematic spread change, DsJ,t/sJ,t.These two
models are nested in a more general model that allows
Dynamics of Spread Change for both proportional and parallel spread changes to take
place simultaneously:
To understand why absolute spread volatility is so
unstable, we first need to examine at a more fundamental ∆si ,t = α J ,t + β J ,t ⋅ si ,t + ε i ,t (9)
level how spreads of individual securities change in a given
Exhibit 5 shows changes in spreads experienced by
month. One basic formulation of the change in spread of
the large issuers that constitute the Communications sector
some bond i at time t is that the overall change is simply
of the Lehman Brothers Corporate Index against their
the sum of two parts, systematic and idiosyncratic:
beginning-of-month spreads in January 2001.6 It is clear
that this sectorwide rally was not characterized by a purely
∆si ,t = ∆s J ,t + ∆siidiosyncratic
,t for i∈J (5) parallel shift; rather, issuers with wider spreads tightened
by more.
where J denotes a peer group of bonds with similar risk
Exhibit 6 reports the regression results when we fit
characteristics (such as Financials rated Baa with duration
the three general models of spread change to the data in
of up to five years). This formulation is equivalent to
this example. The results verify that spreads in the
assuming that spreads change in a parallel fashion across
Communications sector in January 2001 changed in a
all securities in a given market cell J (captured by Dsj,t).
proportional fashion.The slope estimate is highly signi-
Alternatively,if changes in spreads are proportional to
ficant, and the high R2 (97.1%) indicates that the model
spread level,we have (omitting the subscript t for simplicity):
fits the data well.7
The combined model that allows for a simultaneous
∆si ∆s J ∆siidio
= + (6) parallel shift achieves only a slightly better fit (97.7%) and
si sJ si yields a somewhat unintuitive result; it shows that the
sector widens by a parallel shift of 16 b.p. and simultane-
 ∆s J  ously tightens by a relative spread change of -28%. We
∆si = si   + ∆si
idio
or: therefore estimate a fourth model, which is essentially a
 J 
s
variant of the combined model:
Equation (6) reflects the idea that systematic ∆si ,t = α J ,t + β J ,t ( si ,t − s J ,t ) + ε i ,t (10)
spread changes are proportional to the current (system-
atic) spread level and that the sensitivity of each security Normalizing spreads by subtracting the average
to a systematic spread change depends on its level of spread. spread level in Equation (10) yields slope coefficients and

82 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
EXHIBIT 5
Average Spreads and Spread Changes for Large Issuers in the Communications Sector (January 2001)

Source: Lehman Brothers.

EXHIBIT 6
Regression Estimates of Various Models of Spread Change—Communications

R2 values reported in the last column are based on 1,480 individual regressions (185 months 3 8 sectors).
Source: Lehman Brothers.

R2 that are identical to those generated by the combined We conduct an analysis similar to that in Exhibit 6
model, but now the intercept α J ,t represents the average using individual bond data in all eight sectors and 185
spread change in the sample. This model expresses the months included in the sample. Our hypothesis that the
month’s events as a parallel tightening of 245 b.p. cou- relative model provides in general an accurate descrip-
pled with an additional relative shift, with a slope of 228%, tion of the dynamic of spread changes has several testable
that represents how much more spreads move for issuers implications. First, the aggregate R2 for the relative model
with above-average spreads, and how much less they move should be significantly better than that of the parallel
for issuers with below-average spreads. model, and almost as good as that of the combined model.

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 83

Copyright © 2007
Second, we would like to find that the slope factor is sta- and tighten by more in a rally. There are essentially no
tistically significant (as indicated by the t-statistic) in most examples of large parallel spread movements when the
months and sectors. Third, the realizations of the slope slope factor moves in the opposite direction. This clear
and the parallel shift factor in the combined model with linear relationship between the shift and slope factors
normalized spread should be in the same direction, espe- serves as an additional validation of the relative model.
cially whenever the market experiences a large move. That
is, in all significant spread changes, issues with wider spreads Systematic Spread Volatility
experience larger moves in the same direction.
We find support for all three implications.The last The security-level analysis has established that sys-
column of Exhibit 6 reports the aggregate R2 for these tematic changes in spreads are proportional to the sys-
regressions across all sectors and months. The relative tematic level of spread consistent with Equation (6).We
model explains twice as much variation in spreads (33%) now go one step further, and examine the relation between
as the parallel shift model (16.9%) and almost as much as systematic spread volatility and the level of spreads.To do
the less restrictive combined model. Only about a third this, we would like to partition our dataset by spread level,
of spread movements are explained because in many separately measure the volatility of each spread bucket,
months there is little systematic change in spreads, and and examine the relation between spread level and spread
spread changes are largely idiosyncratic. Still, the slope volatility.
factor is statistically significant 73% of the time. The nature of the dataset presents several challenges,
Exhibit 7 shows that large spread changes are accom- however. First, it is far from homogeneous—it includes
panied by slope changes in the same direction (the bonds from different industries, credit qualities, and matu-
correlation between the two is 80%).That is, bonds that rities. Second, the spreads of corporate bonds exhibited
trade at wider spreads will widen by more in a widening large variation over the course of the period studied, so

EXHIBIT 7
Regression Coefficients for Shift and Slope Factors

Source: Lehman Brothers.

84 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
that the populations of any fixed spread buckets vary sub- of the partition and sample population can be found in
stantially from one time period to another. Our goal was the appendix).
to design a partition fine enough so that the bonds in The systematic spread change in cell J in month t can
each cell share similar risk characteristics, yet coarse enough be represented simply as the average spread change across
so that our cells are sufficiently well populated over the all bonds in that bucket in month t.Therefore, for each cell
course of the period to give statistically meaningful results. in the partition,we compute every month the median spread,
The credit index is first partitioned rather coarsely the average spread change, and the cross-sectional standard
by sector (Financials, Industrials, and Utilities) and then deviation of spread change.This procedure produces 51 dis-
further subdivided by duration (short, medium, and long). tinct time series datasets; each consists of a fairly homoge-
To ensure that every sector-duration cell is well-populated neous set of bonds for which we have monthly spreads and
each month, we do not use prespecified duration levels spread changes.We then calculate the time series volatility
but rather divide each sector into three equally populated of these systematic spread changes.9 Similarly, the spread
duration groups.8 level for bucket J is calculated as the time series average of
In the last step, bonds in each sector-duration cell the monthly median spread (rather than the average spread).
are assigned to one of several spread-level buckets.To allow The relation between the volatility of systematic spread
a detailed partitioning of the entire spread range while changes and spread level is plotted in Exhibit 8,where each
minimizing the number of months that a bucket is sparsely observation represents one of the 51 buckets in the parti-
populated, the spread break points differ from sector to tion. Exhibit 8 illustrates a clear relation between spread
sector. In addition, the Financial and Industrial sectors are volatility and spread level. Higher spreads are accompanied
divided into six spread buckets, while the Utilities sector by higher volatilities for all sector-duration cells. Relatively
has only five spread buckets (a more detailed description minor differences can be seen between Industrials and the

EXHIBIT 8
Time Series Volatility of Systematic Spread Changes versus Spread Level (9/89-1/05)

Source: Lehman Brothers.

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 85

Copyright © 2007
other two broad sectors. Similarly, duration does not seem The 250-300 b.p. spread region represents the boundary
to have any significant systematic effect on the results.10 between investment-grade and high-yield bonds. For a
Nonetheless, the results shown in Exhibit 8 do not good part of the time period of our study, these spread cells
perfectly corroborate our hypothesis of proportional spread are very lightly populated by our investment-grade bond
volatility, which would predict that all our observations (or sample. Because we exclude any cell with fewer than 20
at least all observations within a given sector) should lie bonds, the summary results for these cells may be less
along a diagonal line that passes through the origin, of robust than desired.
the form: To examine the relation between systematic spread
change volatility and spread level beyond the 250 b.p.
σ spread
absolute
(s) ≅ θs (11) level, we repeat the analysis including all bonds rated Ba
and B.This increases the sample size by roughly 35%, to
While the points at the left side of Exhibit 8 seem 565,602 observations.We employ the same sector x dura-
to fit this description, the points to the right, representing tion x spread partition, with the addition of several spread
higher spread levels, do not seem to continue along this buckets to accommodate the widening of the spread range
line. Rather, volatility seems to flatten out beyond the 200 (the number of cells increases to 66).
to 250 b.p. range. Is it possible that spread volatility does Exhibit 9 plots the relation between systematic spread
not continue to grow linearly when spreads increase volatility and spread level using both investment-grade
beyond a certain point? and high-yield data.The linear relation between the two
Before we reject our hypothesis, we should question now extends out through spreads of 400 b.p. As before, the
the significance of these few highest-spread observations. three observations that represent the highest-spread bucket

EXHIBIT 9
Systematic Spread Volatility versus Spread Level (investment-grade 1 high-yield bonds)

Monthly observations for all bonds rated Aaa-B (9/1989-1/2005).

Source: Lehman Brothers.

86 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
in the Industrials sector (circled) have somewhat lower- The idiosyncratic spread change of bond i in market
than-expected spread volatility. Once again, we suspect cell J at time t is defined as the difference between its
the statistical relevance of these most extreme data points. spread change and the average spread change for the cell
The simple linear model of Equation (11) provides in that month:
an excellent fit to the data, with u equal to 9.1% if we use
all observations or 9.4% if we exclude the three circled ∆siidio
,t = ∆si ,t − ∆s J ,t (12)
outliers. Hence, the results suggest that the historical
volatility of systematic spread movements can be expressed The volatility of idiosyncratic spread changes is then
quite compactly, with only minor dependence on sector exactly equal to the cross-sectional standard deviation of
or maturity, in terms of a relative spread change volatility total spread changes.11
of about 9% per month. That is, spread volatility for a Exhibit 10 shows a scatterplot of the cross-sectional
market segment trading at 50 b.p. should be about 4.5 volatility for all months and spread buckets including high-
b.p. per month, while that of a market segment at 200 b.p. yield bonds. This plot clearly shows the general pattern
should be about 18 b.p. per month. of volatilities increasing with spread, as well as the rela-
tive paucity of data at the higher-spread levels.
Idiosyncratic Spread Volatility To obtain a single measure of idiosyncratic spread
volatility for each bucket, we pool all observations of idio-
To study the spread dependence of idiosyncratic spread syncratic risk in a given market cell J over all bonds and
volatility, we use the same partition as for our study of sys- all months, and compute the standard deviation. This
tematic spread volatility.Instead of the average spread change pooled measure of idiosyncratic spread volatility per market
experienced within a given cell in a given month, we cell is plotted in Exhibit 11 against the median spread of
examine the dispersion of spread changes within each cell. the cell.

EXHIBIT 10
Volatility of Idiosyncratic Spread Change versus Spread Level

Monthly calculations (9/1989-1/2005) for all bonds rated Aaa-B, computed separately by sector, duration, and spread bucket (N 5 7,250).
Source: Lehman Brothers.

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 87

Copyright © 2007
EXHIBIT 11
Pooled Idiosyncratic Spread Volatility versus Spread Level

Each observation represents the standard deviation of idiosyncratic spread changes aggregated across all sample months separately by sector, duration, and spread
bucket for all bonds rated Aaa-B (9/1989-1/2005).
Source: Lehman Brothers.

The linear relation between spread and spread an intercept and a spread slope factor. We follow the same
volatility is strikingly clear. Observations that represent approach for idiosyncratic spread volatility, except that we
buckets populated almost exclusively by high-yield bonds use the monthly cross-sectional volatility estimates.
exhibit more variation than those representing investment- Panels A and B of Exhibit 12 present the yearly
grade bonds,but follow the exact same pattern. The regres- spread slope estimates and the corresponding adjusted R2.
sion results indicate a zero intercept, but the estimated The results are plotted separately for systematic and idio-
slope coefficient (the relative volatility of idiosyncratic yield syncratic volatility. The estimated coefficients are all highly
change) is somewhat higher than estimated previously, significant, with t-statistics ranging between 15 and 30
11.5% versus 9.4%. for both systematic and idiosyncratic spread volatility.
Not surprisingly, Exhibit 12 reveals that including
Stability of Spread Behavior high-yield data generally increases the spread estimate for
both systematic and idiosyncratic volatility. The spike in
We have established that spread volatility is linearly volatility caused by the 1998 Russian crisis is evident in
proportional to the level of spread. What is the extent of the high estimate of spread slope in 1998 (except for idio-
time variation in the spread slope or the change in spread syncratic volatility with high-yield). Excluding 1998, the
volatility as spreads vary? spread slope estimates are remarkably stable despite the
For each bucket we compute the yearly systematic limited number of observations in the estimation.
spread volatility and corresponding average spread level Panel B of Exhibit 12 reveals that the regressions
(using 12 months of average spread change).12 We then have better and more stable explanatory power when high-
regress these estimates of systematic spread volatility against yield securities are included.When we analyze investment-

88 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
EXHIBIT 12
Yearly Regression of Spread Volatility Against Spread Level (1/1990-12/2004)

Source: Lehman Brothers.

grade data only, the R2 of our regressions goes as low as DTS, Spread Duration, and Excess Returns
40% for systematic volatility and 30% for idiosyncratic
volatility. When we include high-yield data, the regres- If the volatility of both systematic and idiosyncratic
sion results are much better, achieving R2 values of con- spread changes is proportional to the level of spread, the
sistently over 70% for systematic volatility and 60% for volatility of excess returns should be linearly related to
idiosyncratic volatility. DTS, with the proportionality factor equal to the volatility
Overall, this pattern confirms that relative spread of relative spread changes over the corresponding period
changes characterize both investment-grade and high- [see Equation (4)].
yield credit. To examine this prediction, each month bonds are
assigned to quintiles according to their DTS value. Each
quintile is further subdivided into six buckets based on
A NEW MEASURE OF
spread. Every month the average excess returns and median
EXCESS RETURN VOLATILITY
DTS are calculated, and then the time series volatility of
What are the implications of spread proportionality? excess returns and average DTS are calculated separately
Which measure—duration times spread, or spread dura- for each bucket. This formulation yields two empirical
tion—is more appropriate for representing the risk of predictions:
credit securities? We show here that excess return volatility 1. Excess return volatility should increase linearly with
increases linearly with DTS, consistent with the formu- DTS, where the ratio of the two (or slope) represents
lation in Equation (4). Furthermore, portfolios with very the volatility of relative spread changes previously
different spreads and spread durations but with similar estimated.
DTS exhibit the same excess return volatility. 2. The level of excess return volatility should be approx-
For example, a portfolio with a weighted spread of imately equal across spread buckets with a similar
200 b.p. and spread duration of two years is as risky as a DTS value.
portfolio with a spread of 100 b.p. and spread duration of
four years.We also show that DTS generates better esti- The results of the analysis, presented in Exhibit 13,
mates of future excess return volatility than those calcu- strongly support both empirical predictions, even though
lated by spread duration. we do not control for industry, quality, maturity, or any
other effect.

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 89

Copyright © 2007
EXHIBIT 13
Excess Return Volatility versus DTS (9/1989-1/2005: investment-grade credit)

Bonds are first divided into DTS quintiles and then further subdivided into six buckets by spread level.
Source: Lehman Brothers.

First, it is clear that excess return volatility increases bottom spread buckets in the second DTS quintile (shown
with the level of DTS and that a straight line through the in boldface) exhibit very close DTS values of 299 and 320.
origin provides an excellent fit. This is confirmed by a Yet they have very different spread and spread duration
regression of the excess return volatility on average DTS, characteristics; bonds in the top bucket have average spread
which finds a fit of 98% and an insignificant intercept. duration of 5.48 and trade at a spread of 54 b.p., while
The slope estimate is 8.8%, which is in line with the esti- bonds in the bottom cell have spread duration of 2.53 and
mated slope from the analysis of systematic spread volatility. a spread of 127 b.p. Hence,a portfolio of high-spread bonds
Second, consistent with prediction (2), observations with short duration can be as risky as a portfolio of low-
representing the same DTS quintile but with differing spread bonds with high duration, as long as they both have
spread levels exhibit very similar excess return volatilities. roughly the same duration times spread.13
The one exception to this is in the highest DTS quintile,
where the subdivision by spread causes wide variations in A Comparison of
DTS as well.As a result, the points no longer form a tight Excess Return Volatility Forecasts
cluster, although they do continue to exhibit the same
general relation between DTS and volatility. A natural step to extend the analysis is to examine
To demonstrate the significance of the second result, which approach provides a better forecast of the excess
Exhibit 14 reports the average spread and spread duration return volatility of a portfolio:
for all 30 buckets. It illustrates the extent of the variation 1. Spread duration 3 historical volatility of absolute
in spreads and corresponding spread duration across buckets spread change.
with almost identical DTS. For example, the top and 2. DTS 3 historical volatility of relative spread change.
90 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
EXHIBIT 14
Summary Statistics by DTS and Spread Buckets

Source: Lehman Brothers.

To directly compare the forecasting accuracy of the volatility forecasts generated by the two measures differ.
two measures, for every month we compute the realized If we explicitly write the expression for the ratio of the
excess return of each of the 24 buckets in the Lehman two measures at month t for some bucket J:
Brothers Credit Index (8 sectors 3 3 credit ratings). The
∑D × s
∆s J
carry component (spread/12) is stripped from the realized σt( sJ
)× i ,t i ,t
i∈J
excess return, and the random part is then divided by one Vol ratio J,t =
of the two forecasts of excess return volatility.14 σ t ( ∆s J )× ∑D i ,t
i∈J
If the projected excess return volatility is an unbiased
estimate of the “true”volatility, then the time series volatility θ × ∑ Di ,t × si ,t
of these standardized excess return realizations should be i∈J
≅
very close to 1.0. θ × s J ,t × D J ,t
Our premise is that relative spread change volatility
is a more timely measure than absolute spread change ∑D
i∈J
i ,t × ( s J ,t + siidio
,t )

volatility, as it can react almost instantaneously to a change ≅

s J ,t × D J ,t
in market conditions. Hence we expect the sample time
series standard deviation of excess returns to be closer to s J ,t × D J ,t s J ,t
≅ =
1.0 under approach 2 than under approach 1. A volatility s J ,t × D J ,t s J ,t (13)
measure that adjusts more quickly for changing market
conditions should also generate less extreme realizations The volatility measure based on relative spread changes
(realizations that fall above or below two or three standard reflects the current spread level of bucket J, while the
deviations) than a measure that is slower to react. volatility measure based on absolute spread changes reflects
Before we examine the results of the forecasting, it the time-weighted average spread the bucket has exhib-
may be helpful to look at the conditions under which the ited over the volatility estimation period (denoted s J ,t ).

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 91

Copyright © 2007
 If, for example, the systematic spread level of bucket A comparison of the three sets of observations reveals
J over the estimation period is unchanged, the ratio would that using absolute spread changes produces downward-
be equal to 1.0. Otherwise, it would be above or below (upward-) biased estimates of volatility for the entire
1.0, depending on whether the current spread is above or available history (the previous 36 months). As a result, the
below the historical average. Using a shorter period for average standard deviation of normalized excess returns
estimating spread change volatility will not necessarily using the entire history is above 1.0 (1.14) and using the
reduce the difference between the two measures, if the partial history is below 1.0 (0.92). The observations
long-term historical spread is a better reflection of the generated using relative spread changes are evenly spread
current spread environment than the recent past. around 1.0, and the average standard deviation of stan-
Exhibit 15 displays the mean and standard deviation dardized excess returns is 1.01. A close examination of the
of the time series of normalized residuals separately for results does not suggest any relation between the devia-
each volatility measure (each observation represents one tion from 1.0 and the sector-quality bucket.
of the 24 buckets).The normalized residuals are generated These findings support our empirical prediction and
from the entire history of returns available at the begin- are also consistent with the analysis of the ratio of the two
ning of each month. In addition, Figure 15 shows the mean volatility measures. Excess return volatility estimates based
and standard deviation of the normalized residuals when on absolute spread changes are very sensitive to the length
the absolute spread change volatility is calculated over the of the estimation period; they may overreact when using
previous 36 months only. This corresponds to the approach too few data points and can be slow to adjust when using
many investors take in periods of exceptionally low or a long history. What is the optimal estimation period is
high volatility, namely, to rely only on recent data. not clear ex-ante when using absolute spread changes.

EXHIBIT 15
Mean and Standard Deviation of Normalized Excess Return Realizations

Conditional volatility estimates are computed monthly by sector and credit quality based on the entire available history or previous 36 months, using monthly
spread changes observations (9/89-1/05).
Source: Lehman Brothers.

92 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
 A longer estimation period is always desirable for face value, one might say there is no lower bound for
proportional spread changes, however, because it improves volatility and that spread volatility should decline to almost
the accuracy of the proportionality factor, while at the zero for very low-spread securities. Spread volatility, how-
same time the volatility estimate adjusts instantaneously ever, is not driven solely by changes in credit risk but also
because of multiplication by the current spread level.15 by non-credit risk-based factors. Non-credit risk-based
The second empirical prediction states that there spread changes can occur because of noise (i.e., pricing
should be a lower percentage of extreme realizations errors);demand/supply imbalance (for example,when secu-
(positive or negative) in the case of relative rather than rities enter or exit the Lehman Brothers Corporate Index);
absolute spread change volatility. Exhibit 16 plots a his- and other factors.
togram of the standardized excess return realizations for Spread volatility (systematic or idiosyncratic) can
all sector-quality cells for the two volatility measures. The therefore be represented as the sum of two terms: a con-
standard normal distribution is also displayed. stant term that reflects non-credit risk-based spread
Not surprisingly, the histogram reveals that both volatility, and a second term that represents spread volatility
volatility estimators generate distributions that are nega- due to changes in credit risk (which may be approximated
tively skewed (-2.67 and -1.35 for the relative and absolute by a linear function of spread) as follows:
spread change-based volatility measures).With respect to
the percentage of outliers, 7.06% of the observations in σ ( ∆s ) = σ non
2
credit risk + θ 2s2 (14)
the distribution based on absolute spread changes are
located beyond two standard deviations from the mean. Equation (14) makes it clear that for sufficiently high
In the case of the distribution based on relative spread spreads, the second term dominates the first, and spread
changes, the same figure is almost half, at 4.03%. volatility can be approximated well by a linear function of
spread, as we find for corporate bonds. As spreads tighten
THE SCOPE OF DTS and approach zero, the first term dominates, and spread
volatility should converge to some minimum structural level.
What do our findings imply for the level of spread Agency debentures provide a natural framework to
volatility as spreads approach zero? Taking our results at examine the behavior of spread volatility for very low spreads.

EXHIBIT 16
Distribution of Standardized Excess Returns

Based on observations (9/1992—1/2005) grouped across all sectors and credit ratings.
Source: Lehman Brothers.

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 93

Copyright © 2007
Because of market perception that securities issued by the shape and fits nicely with that of long-financials. A regres-
three major agencies are backed by the U.S. government, sion of spread volatility against spread level reveals a flatter
these securities typically trade at very low spreads.Between slope than we estimated for corporates (5.7% versus 9%),
September 1989 and April 2005,the median agency spread consistent with Equation (14).17
stayed between 20 and 50 b.p. except for a few distinct An analysis of idiosyncratic volatility indicates in a
months. We follow the same approach as for corporates. similar fashion that volatility increases moderately as spreads
Each month,bonds are partitioned according to beginning- increase from 20 b.p. to 80 b.p. and indicates a structural
of-month spread level.Average spread change and median volatility level of 4.0-4.5 b.p. per month. The fact that
spread level are computed separately for each bucket.We idiosyncratic structural volatility is higher than the corre-
then examine the relation between the time series volatility sponding systematic level is to be expected, as pricing noise
and average (median) spread level of each bucket. should be more pronounced for individual securities.
The sample spans roughly the same time period as To complete the analysis, we partition the sample
the corporate dataset (September 1989-April 2005) and into 12 DTS buckets and plot the excess return volatility
includes all Aaa-rated, non-callable debentures in the of each bucket against its DTS (Exhibit 18). Similar to
Lehman Brothers Agency Index.16 As before, we discard corporate bonds, excess return volatility increases linearly
extreme observations (in either the top or bottom per- with DTS (the estimated slope from the regression is 9.8%
centile of the spread distribution). Since the total number versus 8.8% for corporates). As the DTS approaches zero,
of observations (73,000) is about 17% of the corporate however, there is a clear flattening of the relation, and
sample size, we use only eight spread buckets. volatility does not decline further. Indeed, the regression
The results are presented in Exhibit 17. For com- yields a significant intercept of 3 b.p., which is consistent
parison, we also show the spread volatility of long- with our previous estimate for the structural level of sys-
duration financials which share many of the characteristics temic volatility.
as agencies.
The plot in Exhibit 17 illustrates that spread volatility DTS Across Seniority Classes
is roughly constant for spreads below 20 b.p., and the level
of structural systematic volatility is about 2.5-3.0 b.p. per Probably some of the most convincing evidence in
month. Above 20 b.p., the relation takes the usual linearsupport of the DTS concept is that portfolios that are
remarkably different in terms of their spread and spread
duration, but with a similar product of the
two (DTS), exhibit the same excess return
volatility.Underpinning this result is the issue
EXHIBIT 17 of whether credit risk is fully captured by
Systematic Spread Volatility Versus Spread Level of Agencies spreads. If spreads incorporate on average all
(9/1989-4/2005) publicly available information related to credit
risk, then all portfolios with similar DTS
should have the same level of excess return
volatility.
We re-examine this issue in the context
of debt seniority by looking at portfolios of
bonds from different seniority classes (e.g.,
senior notes, debentures) but with a very
similar DTS. If spreads already incorporate
the likelihood of default and the recovery
value in such a case, all portfolios should
exhibit the same excess return volatility. Such
a result would provide further support for
our earlier findings.
While credit ratings naturally lend
Source: Lehman Brothers. themselves to cross-sectional comparisons,

94 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
EXHIBIT 18 grouping bonds into portfolios according to
Excess Return Volatility versus DTS (9/1989-4/2005): agencies seniority class is inappropriate.18
To address these issues, we perform a
more detailed analysis at the issuer level (iden-
tified by ticker). Each month, we construct
two portfolios for each issuer, SENIOR and
SUBORD, which include all the securities
(often just a single one) defined as senior and
subordinated. Months when only one of the
portfolios is populated are eliminated.
We first compute the market-weighted
duration times spread and excess return for
each portfolio and the DTS ratio of the
SENIOR portfolio to the SUBORD port-
folio. We then match the DTS of the
SENIOR portfolio to that of the SUBORD
Source: Lehman Brothers.
portfolio (i.e., the DTS is scaled up or down)
and adjust the excess return accordingly.
Hence for every issuer, we have a time series
of excess returns for two portfolios with the
EXHIBIT 19 same DTS each month.19
Summary Statistics for Senior and Subordinated Portfolios Using this approach for portfolio
construction has clear advantages over the
cross-sectional technique. First, it controls
for any issuer-specific effect. Second, it accu-
rately captures the relative seniority of dif-
ferent claims. Third, the fact that by
construction the two portfolios have the
same DTS has testable implications: The
ratio of excess return volatility of the two
portfolios should be 1.0 on average. In addi-
tion, any difference in excess return should
be relatively small and reflect only idiosyn-
cratic risk (for example, one portfolio may
include bonds that on average are smaller
Source: Lehman Brothers. and older than bonds in the second port-
folio and are therefore less liquid).
Exhibit 19 presents the 25th percentile,
50th percentile, and 75th percentile of the
constructing portfolios based on debt seniority is more of ratio of excess return volatility for the SENIOR and
a challenge. The classification of a bond as senior or sub- SUBORD portfolios as well as the difference in average
ordinated depends on its payment priority in case of a excess returns.As the ordering among seniority classes is not
default. The recovery value of any bond is affected by always clear, the table presents these statistics for different
other claims on the same issuer that are more or less senior compositions of the SENIOR and SUBORD portfolios.
to that bond. Across issuers, however, the same seniority For example, the second row reports senior notes
class does not necessarily imply similar recovery value in under SENIOR and notes under SUBORD. There were
case of a default. Furthermore, even for a given issuer it 353 different issuers with these portfolios populated over
is not always clear if a certain claim is senior to another some time period.The median ratio of excess returns volatil-
claim (e.g., a debenture versus a senior note). Thus simply ities is 0.94, indicating no significant difference between

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 95

Copyright © 2007
the two portfolios. One-quarter of the issuers exhibited We obtain similar results for other compositions of the
ratios below 0.79,and one-quarter of the issuers ratios above two portfolios that are reported in Exhibit 19.
1.08, with the remaining half falling between these values.
The typical performance of the two portfolios is also very SUMMARY AND IMPLICATIONS
similar and the median difference is 1 basis point per month FOR PORTFOLIO MANAGERS
(i.e., the SUBORD portfolio underperforms).
The results reported for other portfolio compositions We have presented a detailed analysis of the behavior
of spread changes. Using our extensive corporate bond
are similar (in particular the last row, which represents the
most inclusive case), and do not indicate the two port- database, which spans 15 years and includes well over
folios exhibit different risk characteristics. 560,000 observations, we demonstrate that spread changes
To examine the relation between duration times are proportional to the level of spread. Systematic changes
in spread across a sector tend to follow a pattern of relative
spread and excess return volatility across various seniority
classes, each SENIOR and SUBORD portfolio (con- spread change; that is, bonds trading at wider spreads expe-
rience greater spread changes.The systematic spread volatility
structed for each issuer) is assigned to one of the DTS quin-
tiles each month. We then calculate the weighted excess of a given sector (if viewed in terms of absolute spread
return and DTS for each quintile (separately by senioritychanges) is proportional to the median spread in the sector;
class).The two aggregate portfolios in each quintile havethe non-systematic spread volatility of a particular bond or
issuer is proportional to its spread as well. These findings
the exact same DTS since at the issuer level the DTS of the
SENIOR and SUBORD portfolios is equal by construc- hold irrespective of sector, duration, or time period.
tion. As before, we compute the time series volatility of In a sense, these results are not altogether surprising.
excess returns and the average DTS of the ten portfolios.The lognormal models typically used to represent changes
in interest rates assume that changes in yield are propor-
Exhibit 20 presents the results of the analysis com-
tional to current yield levels. Models for pricing credit
paring senior notes to notes and senior debentures to sub-
derivatives have used a similar lognormal model to describe
ordinated debentures.The scatterplot shows that the linear
relation between excess return volatility and DTS is pre-changes in credit spreads (see Schönbucher [1999]).
served, and that the slope does not depend on the seniority An assumption of lognormal spread changes would
imply two things: 1) That spread changes are proportional
level. In both cases, there is an almost exact match between
the volatilities of the SENIOR and SUBORD portfolios. to spreads, and 2) that the relative spread changes are nor-
mally distributed. Our results can be seen as
providing empirical evidence to support the
first of these assumptions, but not necessarily
EXHIBIT 20 the second.
Excess Return Volatility versus DTS Across Seniority Classes There are several implications for a
portfolio manager who wishes to act on
these results. First, the best measure of expo-
sure to a systematic change in spread within
a given sector or industry is not the contri-
bution to spread duration, but the contri-
bution to duration times spread, or DTS.At
many asset management firms, the targeted
active exposures for a portfolio relative to
its benchmark are expressed as contribution
to duration overweights and underweights
along a sector by quality grid—and reports
on the actual portfolio follow the same
format. In the relative spread change approach
managers would express their targeted
overweights and underweights in terms of
Source: Lehman Brothers. contributions to DTS instead.

96 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
 Second, our finding that the volatility of non- A typical approach might be to match the dollar
systematic return is proportional to DTS offers a simple durations of the two bonds, or to go long and short credit
mechanism for defining an issuer limit policy that enforces default swaps of the same maturities with the same notional
smaller positions in more risky credits. Many investors amounts. If issuer A trades at a wider spread than issuer
specify some ad hoc weight cap by credit quality to con- B, however, our results would indicate that a better hedge
trol issuer-specific risk.20 Alternatively, we can set a limit against marketwide spread changes could be obtained by
on the overall contribution to DTS for any single issuer. using more of issuer B, so as to match the contributions
For example, say the product of market weight 3 to DTS on the two sides of the trade.
spread 3 duration must be 5.0 or less.Then, a position in Should portfolio management tools such as risk and
issuer A, with a spread of 100 basis points and a duration performance attribution models be modified to view
of five years, could be up to 1.00% of portfolio market sector exposures in terms of DTS contributions and sector
value, while a position in issuer B, with a spread of 150 spread changes in relative terms? The answer in our
and an average duration of ten years, would be limited to opinion is yes, in both cases.
0.33%. A risk model for any asset class is essentially a set of
Establishing issuer limits based on spreads has advan- factors that characterize the main risks to which securi-
tages and disadvantages compared to a ratings-based ties in that asset class are exposed. The risk of an indi-
approach. One advantage, as described above, is that it is vidual security or portfolio depends on its risk loadings
simple to specify a single uniform limit that requires or sensitivity to the set of factors and their past realiza-
increasing diversification with increasing risk.The key dif- tions. For credit-risky securities, the traditional risk fac-
ference between the two approaches, though, concerns tors typically measure absolute spread changes in terms of
how often issuer limits are adjusted. a sector-by-quality partition that spans the universe of
In a ratings-based framework, bond positions that bonds. Specifying the risk factors in terms of relative spread
are within policy on the date of purchase will tend to changes instead has two important benefits.
remain within policy unless they are downgraded. A First, such factors would exhibit more stability over
spread-based constraint, by contrast, however, is by its very time and allow better forward-looking risk forecasts.
nature constantly adjusted as spreads change. One pos- Second, the partition by quality would no longer be nec-
sible result is that as spreads widen, a position that was essary to control risk, and each sector can be represented
within policy when purchased can drift over the allow- by a single risk factor. This would allow managers to
able DTS limit. Strict enforcement of this policy, requiring express more focused views, essentially trading off the
forced sales to keep all issuer exposures to stay within the elimination of the quality-based factors with a more finely
limit, could become very distracting to managers, and grained partition by industry.
incur excessive transaction costs as spreads trade up and Similarly, a key goal for attribution models is to
down. match the allocation process as closely as possible. If and
One possible solution would be to specify one when a manager starts to state allocation decisions in
threshold for new purchases and a higher one at which terms of DTS exposures, performance attribution should
forced sales would be triggered. This could provide a mech- follow suit.
anism that adapts to market events more quickly than the One practical difficulty that may arise in the imple-
rating agencies without introducing undue instability. mentation of DTS-based models is an increased vulner-
Another possible disadvantage of the DTS-based ability to pricing noise. For the most part, models of
issuer cap is that it allows for large positions in low-spread portfolio risk and reporting of active portfolio weights
issuers and exposes the portfolio to credit torpedoes. This rely largely on structural information. Small discrepancies
too would argue for using the DTS-based approach in in asset pricing give rise to small discrepancies in market
conjunction with caps on market weights. values, but potentially larger variations in spreads.
Third, there could be hedging implications. Say a Managers who rely heavily on contribution-to-DTS
hedge fund manager has a view on the relative perform- exposures will need to implement strict quality controls
ance of two issuers within the same industry, and would on pricing.
like to capitalize on this view by going long issuer A and Our investigation of the relation between duration
short issuer B in a market-neutral manner. How do we times spread and excess return volatility has focused almost
define market neutrality? entirely on investment-grade credit in the U.S. We have

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 97

Copyright © 2007
also included, however, some results from U.S. high-yield We believe the duration times spread approach
credit that show that the idea of proportional spread accurately represents the impact of spread changes on
changes carries through to high-yield as well as agency excess returns, and that acceptance of this result could
debentures. Additional analysis we have conducted estab- have wide-ranging effects on portfolio management prac-
lishes that all our results are valid for European corporate tice throughout the industry. We anticipate continued
bonds as well as for credit default swaps. Furthermore, the research in this area on several fronts, including extension
findings are robust to the return horizon (monthly or to other asset classes, and implementation of DTS-based
weekly) and to the reference curve used to calculate the features into portfolio analytics offerings.
spreads (Treasury or LIBOR).
Indeed, we believe that perhaps one of the most APPENDIX: DATA DESCRIPTION
useful applications of DTS will be in the management of
core-plus portfolios that combine both investment-grade The dataset used in the empirical analysis covers
and high-yield assets. Traditionally, investment-grade credit September 1989 through January 2005 (a total of 185
portfolios are managed according to contributions to dura- months).The sample includes all the bonds in the Lehman
tion, while high-yield portfolios are managed according Brothers Credit Index excluding 1) zero-coupon bonds,
to market value weights; using contributions to DTS across 2) callable bonds, and 3) bonds with non-positive spreads.
both markets could help unify this process. The final dataset provides a total of 416,783 observations.
Skeptics may point out that in high-yield markets, Exhibit A-1 shows the sample by sector and year.
especially when moving toward the distressed segment, We also extend the analysis to include high-yield
neither durations nor spreads are particularly meaningful, bonds rated Ba and B (trading at a price above 80 to mit-
and the market tends to trade on price, based on an esti- igate potential default effects), which increases the number
mated recovery value. A useful property of DTS in that of observations by roughly 35% (from 416,783 to 565,602).
context is that in the case of distressed issuers, where Spread figures are model-driven and can exhibit
shorter-duration securities tend to have artificially high extreme values (especially since the methodology for com-
spreads, DTS is fairly constant across the maturity spec- puting option-adjusted-spreads changed during the sample
trum, so that managing issuer contributions to DTS period). To mitigate the effect of outliers, we exclude
becomes roughly equivalent to managing issuer market observations where changes in spread fall above the 99th
weights. percentile or below the 1st percentile. As a result, monthly
spread changes included in the analysis range from 260
b.p. to 178 b.p.
Figure A2 outlines the exact break-
down into spread buckets by industry and
EXHIBIT A-1 maturity we employ in analyzing the rela-
Sample Population by Sector and Year tion between spread volatility and spread
level. Note that because spread has a ten-
dency to widen with maturity, the popula-
tion of the short-maturity bucket is
concentrated in the lowest-spread bucket
(denoted by 1) while the opposite holds for
the long-maturity bucket.
We also report for each bucket the per-
centage of months during the sample period
when the bond population exceeds 20. This
statistic is of interest, because months with
fewer than 20 observations are filtered out of
any volatility calculation.The percentage of
months with sufficient number of observa-
Investment-grade bonds only; number of bonds is as of December of each year. tions varies between 30% and 50% for Utili-
Source: Lehman Brothers. ties and 50%-80% for Financial and Industrials.

98 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007
EXHIBIT A-2
Sample Partition by Sector, Duration, and Spread

Sample includes IG bonds only, between 9/89 and 1/05. Spread breakpoints, cell population, and percentage of months a bucket is populated by more than
20 bonds.
Source: Lehman Brothers.

ENDNOTES of a volatility change and the time it is first reflected in our

estimates.
1
Spread is the constant (absolute) shift to the zero-coupon 4
To enable the two to be shown on the same set of axes,
discount curve in all scenarios that is required to ensure that the both absolute and relative spread volatility are expressed in units
model value of the bond (average value over all scenarios) equals of similar magnitudes.The interpretation is different, however.
the observed market price. An absolute spread change of 10 represents a 10 b.p. parallel
2
Spread change return is closely related to excess return, shift across a sector, while a relative spread change of 10 means
the return a corporate bond earns over a duration-matched that all spreads in the sector move by 10% of their current values
Treasury bond. Excess return can be approximated by the sum (e.g., from 50 to 55, from 200 to 220).
of the spread change return and an additional component due 5
The eight sectors are: Banking, Finance, Basic Industry,
to spread carry. Consumer Cyclical, Consumer Non-Cyclical, Communica-
3
This practice leads to perennial questions about how tions, Energy, and Utility. Bonds are assigned to one of three
much history should be used in such estimation.A longer time quality cells:Aaa/Aa,A, and Baa.
period leads to more stable estimates of volatility; a shorter time 6
Large issuers are defined as issuers that have outstanding
period (or a weighting scheme that gives more weight to recent issues with market value in excess of 1% of the sector aggre-
observations) makes the estimate less stable, but better able to gate market value. In this case, 17 issuers represent 216 out-
adapt to fundamental changes in the marketplace. In either case, standing issues.
the large swings in volatility that the market can experience 7
Since we compare models with and without an inter-
mean that we are always trying to catch up to market events, cept, Exhibit 6 reports uncentered R2 values calculated using
and there will always be some amount of lag between the time

WINTER 2007 THE JOURNAL OF PORTFOLIO MANAGEMENT 99

Copyright © 2007
the total sum of squares (without subtracting the average spread 14
Although the carry component is time-varying, we ana-
change) rather than a centered R2. lyze each month’s excess return conditioned on the beginning-
8
We find that the distribution of spread duration varies of-month spread.We can therefore treat the carry component
significantly across time and therefore does not allow for a par- as deterministic.
tition based on constant spread duration values. 15
A longer estimation period is always desirable as long
9
Despite our efforts to ensure uniform cell populations, as the proportionality factor is stable across periods, which we
some cells are very sparsely populated (or even empty) in some find to be the case.
months. Months when a cell is populated by fewer than 20 16
Including publicly issued debt of U.S. government agen-
bonds are not used in the analysis. As a robustness check, we cies, quasi-federal corporations, and corporate or foreign debt
repeat the analysis using the entire available time series of sys- guaranteed by the U.S. government (such as USAID securities).
tematic spread changes and a weighted volatility estimate (where 17
The results are unchanged when we exclude issues with
the weighting factor is the number of observations in each a market value below $300 million or non-U.S. agencies.
month).The results are essentially unchanged. 18
When a bank is owned by a holding company, for
10
Instead of a single spread coefficient, we also estimate example, owners of subordinated claims issued by the bank have
an unconstrained model that allows the spread slope coefficient priority in case of a default over owners of a senior claim issued
to vary by sector and duration as follows: by the holding company.
19
σ ( ∆s )i ,d,s = si ,d,s ( β + βFin ⋅ I Fin + βInd ⋅ I Ind + β Med ⋅ I Med + βLong ⋅ I Long ) + ε i ,d,s Excess returns can be adjusted by the same scaling factor
(the ratio of DTS of the two portfolios) because they are lin-
where i, d, and s denote the sector-duration-spread combina-
early related to DTS.To implement this in practice, we would
tion of each observation. IFin and IInd are dummy variables equal
need to take into account financing costs. As we do not form
to 1 if i 5 Financials or Industrials, respectively, and zero other-
a trading strategy, however, but rather examine whether similar
wise. Similarly, IMed and ILong equal 1 if d 5 medium or long,
DTS portfolios exhibit similar excess return volatilities, we can
respectively, and zero otherwise.The results confirm that rela-
ignore borrowing costs.
tive spread volatility is not restricted to a single sector or matu- 20
For example, an investment policy may specify that no
rity, and that there is roughly a 9 b.p. per month pickup in
more than 1% of the portfolio market value can be invested in
volatility for every 100 b.p. increase in spread.
11 securities of any single Baa-rated issuer, no more than 2% in any
To be consistent with Equation (6), idiosyncratic spread
A-rated issuer, and no more than 4% in any Aa-rated issuer.
change should be defined as ∆siidio ,t = ∆si ,t − ( si ,t s J ,t ) ∆s J ,t . As we are

carrying out this test over relatively narrow spread buckets,

however, there is very little difference in practice between the REFERENCE
two definitions.
12
Depending on the sample composition and population, Schönbucher, Philipp. “A Libor Market Model with Default
this produces between 38 and 66 observations yearly for sys- Risk.” Working paper, University of Bonn, 1999.
tematic volatility and 300 to 500 observations for idiosyncratic
volatility. As before, only observations that represent buckets To order reprints of this article, please contact Dewey Palmieri at
populated by at least 20 bonds during the entire year are included [email protected] or 212-224-3675.
in the analysis.
13
Our findings do not change when we repeat the analysis
using other partitions.

100 DTS (DURATION TIMES SPREAD) WINTER 2007

Copyright © 2007