Fixed Income Performance Attribution

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/228317513
Fixed Income Performance Attribution
Article in SSRN Electronic Journal · May 2010

DOI: 10.2139/ssrn.1619185
CITATIONS READS
10 22,653
3 authors, including:
Stéphane Daul Nicholas Sharp

Pictet Asset Management 7 PUBLICATIONS 59 CITATIONS
22 PUBLICATIONS 704 CITATIONS
SEE PROFILE
SEE PROFILE
All content following this page was uploaded by Nicholas Sharp on 06 November 2017.
The user has requested enhancement of the downloaded file.

Fixed Income Performance Attribution
Authors
Stéphane Daul
RiskMetrics Group, Inc.
[email protected]
Nicholas Sharp
[email protected]
Lars Qvigstad Sørensen

[email protected]
Contents
1 Introduction 3
2 Total Return 5
2.1 Payments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Transactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Currency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3 Factor Model 9
3.1 Factor duration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Currency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2.1 Forward premium and currency surprise . . . . . . . . . . . . . . . . . . . . . 13
3.3 Carry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4 Yield curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.1 Effective duration and convexity . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4.2 Key-rate durations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.5 Credit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6 Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.7 Inflation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1
2
4 Relative Performance Attribution 33
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Pure asset-grouping approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.3 Pure factor model approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.4 Hybrid attribution model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.5 Multi-period attribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.6 Impact of pricing difference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5 Exposures 45
5.1 Partial and full exposure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.2 Performance legs and notional cash . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.3 Portfolio weight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.4 Other derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6 Benchmarks 51
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2 Benchmark rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.2.1 Selection criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2.2 Rebalancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2.3 Total return . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.2.4 Index calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.4 Replicating returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.5 Validation of benchmark data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Chapter 1
Introduction
Performance attribution is a set of techniques used by portfolio managers for the purpose of measuring
and explaining portfolio performance relative to a benchmark. The objective is to quantify the impact
that active management decisions have on the relative performance (i.e. active return) over a given
interval of time.
Each decision taken by the portfolio manager is associated with an attribution effect that quanti-
tatively measures the amount by which the decision affects the relative performance. The attribution
effects aggregate together to fully account for the active return. Because attribution effects describe the
impact of active management decisions it is important that the attribution model reflects the decision-
making process. A long-only government bond portfolio is certainly managed differently from an
emerging-market portfolio, or a credit derivatives portfolio.
The standard attribution methodology, introduced by Brinson and Fachler (1985), attributes the
active return to asset allocation and security selection decisions. The methodology boils down to
grouping positions in a reference portfolio in such a way to make the attribution effects apparent, and
hence its current name, the asset-grouping approach. This approach has been widely extended and is
particularly well suited for equity portfolios. 1
However for fixed-income portfolios this approach is not as successful. Fixed-income managers
need specialized attribution models that for example incorporate all the effects of yield-curve move-
ments. For instance it is difficult to find any reference portfolio that captures only a butterfly effect of
the yield curve. Another route is the factor based approach, where the performance of all securities is
first decomposed using systematic factors and then aggregated. We consider a configurable, extendable
1 This is actually the approach that RiskMetrics has taken for equity attribution, see Soor (2010).
4 CHAPTER 1. INTRODUCTION
hybrid approach to fixed-income performance attribution where attribution is decomposed into many
fixed-income factors as well as simultaneously carrying out an asset-grouping approach to performance
attribution where necessary.
To present the content of this review paper, we show what an attribution report could look like
in Table 1.1. This fictious example presents a portfolio of fixed income instruments along with their
relative attribution to the benchmark. We start by calculating the total return of each securities as
shown in Chapter 2, and report the active return (I), i.e. the security return relative to the benchmark.
The active return is then decomposed (II) using the fixed income factor model developed in Chapter
3. The factors may not capture all investment decisions, and we apply an asset-grouping approach on
the residual (III) as detailed in Chapter 4. Finally we look in more details at how exposures need to be
calculated in Chapter 5 and we analyze benchmarks in Chapter 6.
Table 1.1
Illustrative attribution report
(I) (II) (III)

Active Return Currency Carry Yield Curve Credit Volatility Residual Allocation Selection
Security A 0.170% 0.075% 0.002% 0.133% 0.003% 0.007% -0.050% -0.037% -0.012%
Security B -0.041% -0.075% -0.001% -0.111% -0.002% 0.001% 0.147% 0.111% 0.037%
Security C 0.447% 0.000% 0.059% 0.441% 0.002% -0.008% -0.047% -0.035% -0.012%
Security D 0.220% 0.050% 0.024% 0.150% 0.007% 0.010% -0.020% -0.015% -0.005%
Security E -0.111% -0.030% -0.002% -0.281% -0.017% 0.025% 0.195% 0.146% 0.049%
Security F 0.364% 0.075% 0.052% 0.631% -0.027% 0.062% -0.429% -0.322% -0.107%
Security G 0.052% -0.075% 0.000% 0.080% 0.001% 0.008% 0.038% 0.028% 0.009%
Security H -0.044% 0.000% 0.000% -0.162% -0.004% 0.001% 0.121% 0.091% 0.030%
Security I 0.213% 0.050% 0.003% 0.032% 0.002% -0.004% 0.129% 0.097% 0.032%
Security J 0.120% -0.030% 0.007% 0.327% 0.003% 0.008% -0.194% -0.145% -0.048%
Security K 0.120% 0.075% 0.011% 0.168% 0.002% 0.006% -0.143% -0.107% -0.036%
Security L -0.462% -0.075% 0.002% -1.110% -0.041% -0.090% 0.852% 0.639% 0.213%
Security M -0.131% 0.000% 0.000% -0.121% -0.007% -0.020% 0.016% 0.012% 0.004%
Total 0.916% 0.041% 0.154% 0.177% -0.078% 0.005% 0.617% 0.463% 0.154%
Chapter 2
Total Return
The total return is the percentage change in the value of a security resulting from changes in the market
value of the security, interest income accrued or received, and reinvestment income over the attribution
period. The attribution period is defined as the time that elapses between the start date t1 and end date
t2 .
The total return rtotal of a fixed-income security (in the security’s local currency) over the attribution
period [t1 ,t2 ] is defined as
V (t2 ) −V (t1 ) +C(t1 ,t2 )
rtotal = , (2.1)
V (t1 )
where V (t1 ) is the market value (this is the dirty/invoice price in the case of bonds) of the security at
the start of the attribution period, and similarly V (t2 ) is the market value of the security at the end of
the attribution period. Also, C(t1 ,t2 ) is the sum of payments received during the attribution period.
2.1 Payments
There are two sources of cash that accumulate in a portfolio, these are coupon payments and principal
payments, the latter includes both scheduled and unscheduled prepayments.
Coupon payments
Interest income in the form of coupon payments is the most common form of cash that accumulates in
the portfolio. This form of payment represents the interest income the holder is entitled to based on
6 CHAPTER 2. TOTAL RETURN
the total par amount outstanding. The structure and frequency of coupon payments will vary between
securities. Examples include U.S. Treasury securities which pay semi-annual coupons and MBS fixed-
rate passthrough securities which pay monthly coupons. These are two examples of fixed-coupon
securities, there are other coupon types such as floating coupons, step coupons, or inflation-linked
coupons where the coupon rate must be multiplied by the appropriate index ratio.
Principal payments
The three types of principals payments that must also be accounted for when calculating the total return
are:
1. Sinking funds – Bonds with sinking fund schedules make partial principal payments prior to the
bond’s maturity date.
2. Called securities – Bonds with a call provision allow the issuer to buy (call) back the bond at a
specified price at specified points in the future. Bonds may be fully or partially called. Being
called corresponds to cash at the call price entering the portfolio.
3. Principal prepayments – Principal can be paid early in the form of unscheduled prepayments for
securitized products (MBS, ABS and CMBS). The timing of prepayments and the time of cash
actually entering the portfolio will vary from security to security.
Reinvestment of coupon payments
Payments received are typically reinvested at the reinvestment rate re (tc ) as
C(t1 ,t2 ) = ∑ Ctc · [1 + re (tc )](t2 −tc ) , (2.2)

tc ∈[t1 ,t2 ]
where payments received, Ctc , are coupon payments or principal payments paid during the attribution
period [t1 ,t2 ], and tc is the time of a payment. It is possible that the payments received, Ctc , are rein-
vested at the reinvestment rate re (tc ), e.g. the 1-month LIBID rate (with daily compounding) as of the
prior month-end. In this case the number of days between the receipt of the cash flow and the end of
the attribution period is t2 − tc .
It is also permissible that a policy of not reinvesting payments received is followed. On one hand,
investors may be able to outperform a benchmark if they reinvest cash but on the other hand, they may
also underperform by having to pay transaction costs where a benchmark does not.
Transactions 7
2.2 Transactions
So far we have assumed that performance measurement has been carried out on a static portfolio basis,
i.e. the return is calculated based on holdings snapshots at t1 and t2 . A more relevant approach is to
take account of all transactions that affect the portfolio. This would allow the contribution of trading
activity to the total return to be determined.
To perform transaction-based attribution, we can choose the attribution period to be daily so that
the effect of all transactions are applied at the correct time. For a transaction-based approach the total
return (2.1) of a security is such that
V (t + 1) −V (t) +C(t,t + 1) − P(t) + S(t)

rtotal = , (2.3)
V (t)
where P(t) and S(t) are purchases made and sales made on day t, respectively.1 For information on the
implementation of RiskMetrics transactions-based approach see Soor and Costigan (2010).
It is well known that return calculations based on the buy-and-hold approach are very approximate
(see Vann, 1999, for a study supporting this claim) when compared with calculations that take intra-
attribution period changes in portfolio holdings into account. Buy-and-hold return calculation is still
very important though as the difference between this and the transaction-based return gives a measure
of the value added from transactions, and as such is an integral aspect of attribution analysis.
2.3 Currency
If a bond is denominated in a currency different from the portfolio’s base currency, then part of the
return will be due to foreign exchange rate movements. This is treated as a separate component of the
return and will be part of our factor model. The total return (in base currency) of a security is defined
local as
as the currency return rcurrency , plus the local total return rtotal
base local
rtotal = rcurrency + rtotal . (2.4)
The currency return comprises the currency appreciation and its cross product with the local total
return. Further details are provided in Section 3.2.
1 This formulation assumes that external cash flows occur at the end of the day. Assumptions regarding intra-day cash
flows are also permissible.
8 CHAPTER 2. TOTAL RETURN
Chapter 3
Factor Model
In order to decompose total returns, we need a model that describes how changes in specific factors
affect returns. Models provide approximations to a complex real world. Each model has pros and
cons, and there are always tradeoffs involved when choosing one particular model over another. In
addition, some models may perform well under certain market conditions but poorly under other market
conditions. Therefore, a solution that offers the user flexibility is proposed. We want to attribute returns
between times t1 and t2 to various decisions taken by the portfolio manager. A common factor model is
based on the premise that there exist common factors which affect prices on large classes of securities.
A security’s price change as a response to a factor is determined by the security’s sensitivity to the
factor. At a given time, a factor model with K factors for a security i takes the form
K
ri = ∑ βi,k fk + εi, (3.1)
k=1
where ri is the total return on security i, fk is the return of factor k, βi,k represents the sensitivity of
security i to a movement in factor k, and εi is the residual return that cannot be explained by factors. We
can decompose returns into a systematic and an idiosyncratic component. The systematic component is
the part of the return that is attributable to exposure to factors. The residual is the part of the return that
remains after the factor return decomposition. It contains the idiosyncratic component and whatever is
non-attributable because of data problems. Note that the factor return, fk , is common across all assets
and buckets.
10 CHAPTER 3. FACTOR MODEL
3.1 Factor duration
There are several ways to measure factor sensitivities and factor realizations. One class of models esti-
mates factor returns by cross-sectional linear regressions, similar to the procedure in Fama and MacBeth
(1973). Another class of models constructs factor returns from observable variables and then estimates
factor sensitivities by linear regressions. The advantage of this method is that it does not require so-
phisticated pricing functions. The drawback is that the estimated factors or their sensitivities do not
necessarily correspond to concepts portfolio managers can relate to.
Our approach differs from the two above. Our factors are the actual observed changes in the risk
factors. We have pricing functions for all fixed income instruments. The sensitivity to a change in
a factor is found by repricing the security after applying shifts to the factor. This provides a more
intuitive way to assess the sources of returns as the actual factors are used instead of latent variables.
The cost, however, is that our approach is more computationally demanding.
To set the stage, assume that the market value of a fixed income instrument is a function of N risk
factors
V = g(x1 , x2 , . . . , xN ). (3.2)
The effect on market value of small changes in the factors can be approximated with a second order
Taylor series expansion
N
∂g 1 N N ∂ 2g
∆V = ∑ ∆xk + ∑ ∑ ∆xk ∆xℓ . (3.3)
k=1 ∂ xk 2 k=1 ℓ=1 ∂ xℓ ∂ xk
Typically, cross terms are ignored so that the expression simplifies to
N
∂g 1 N
∂ 2g
∆V = ∑ ∂ xk ∆xk + 2 ∑ ∂ x2 ∆xk2. (3.4)
k=1 k=1 k
The first order terms are referred to as durations with respect to a specific factor and the second order
terms are referred to as convexities. Convexities capture the fact that the relation between a bond’s
price and the factor change is non-linear. We define the factor duration as the negative of the percent
sensitivity of the security’s price with respect to a small shift in the factor. The factor duration with
respect to factor i is thus
1 ∂V
Dk = − , (3.5)
V ∂ xk
where ∂ V /∂ xk represents the derivative of the security price with respect to the factor, i.e. the factor
sensitivity. The return contribution from a specific factor k is thus given by
rfactor k = −Dk × fk , (3.6)

Currency 11
where fk = ∆xk is the ex-post realized return on factor k.
The decomposition of returns can be thought of as a process where returns are attributed to factors
successively. For each security in the portfolio, we decompose the local total return into the following
components:
1. Currency: Each portfolio has a base currency in which market values are computed. If a bond
is denominated in a currency different from the base currency, then part of the return in base
currency will be due to foreign exchange rate movements. These are separated and treated as a
separate factor. Further details are provided in Section 3.2.
2. Carry: The carry captures the part of a bond’s return that is due simply to the passage of time.
It typically consists of two components: the income that accrues to the holder of the bond over
time and the change in clean price, provided that the yield curve remains the same. Both these
effects are detailed in Section 3.3.
3. Yield curve: The yield curve return is the return on the bond that is due to changes in the
government yield curve. For regular bonds, this is typically one of the most important factors. It
is discussed further in Section 3.4
4. Credit: Investors buying bonds that are not backed by the full faith and credit of a government
not subject to default risk, will command a premium to take on the risk of not getting back all
promised coupons and principal. Widening or tightening of credit spreads will impact the returns
on such bonds and is treated in Section 3.5.
5. Volatility: The volatility return captures the change in a fixed income security’s price because
of changes in volatility. This component can be significant for options and instruments that have
embedded options. It is discussed more in Section 3.6
6. Inflation: Changes in the nominal yield curve can be due to changes in the real interest rate
or changes concerning expectations about inflation and associated risk premia. Section 3.7 de-
tails how we can decompose the yield curve return into a real rate component and a break-even
inflation component.
3.2 Currency
In the case of multi-currency portfolios the local currency of a security will often be different from the
base currency of the portfolio and there will be a contribution to the total return from movements in the
foreign exchange rate during the attribution period. Hence, we separate out the component of return
due to currency movements, rcurrency , as shown in Eq. (2.4).
The total return of a security expressed in terms of the local currency and the exchange rate from
local to base currency can be found as follows. First let us consider an investor who buys foreign
currency at time t1 and sells the position back in base currency at time t2 , the currency appreciation rFX ,
or the realized return from this investment in foreign currency, is defined as:
χ (t2 ) − χ (t1 )
rFX = , (3.7)
χ (t1 )
where χ (t1 ) and χ (t2 ) are the exchange rates from local currency to base currency at the beginning and
the end of the holding period, respectively.
For an investor who buys a security denominated in a foreign currency whose market value in the
local currency is V (t) and where the exchange rate from local currency to base currency is χ (t), the
market value in the base currency Vbase (t) is given by
Vbase (t) = V (t)χ (t). (3.8)
Let us assume there is an intermediate cash flow, e.g. a coupon payment, during the holding period,
where Ctc is the value of a payment made at time tc in the local currency during the attribution period
[t1 ,t2 ] . In the presence of intermediate cash flows an assumption about reinvestment must be made to
compute the total return. To keep the exposition as simple as possible, we assume that the intra-month
payment earns zero interest. Furthermore, we assume that the intermediate payment is kept in local
currency until the end of the attribution period, at which point it is converted back into base currency.
Relaxing these assumptions is not hard; it is just a matter of bookkeeping. With these assumptions, the
local total return can be expressed as
V (t2 ) +Ctc
rlocal = − 1. (3.9)
V (t1 )
By substituting (3.8) into (2.1) and with these assumptions we can express the total return in base
currency as follows
[V (t2 ) +Ctc ] χ (t2 )

rbase = · −1
V (t1 ) χ (t1 )
= (rlocal + 1)(rFX + 1) − 1
= rlocal + rFX + rlocal rFX . (3.10)
Currency 13
Finally, by rearranging (2.4) and by making the substitution (3.10) for rbase we have an expression for
the currency return
rcurrency = rbase − rlocal

= rFX (1 + rlocal ). (3.11)
Notice that the currency return contains an interaction component (rFX · rlocal ) that causes a difference
between the return on a currency investment in the cash market and a security’s currency return.
Example
Assume an investor whose base currency is USD. Assume further that the investor buys a UK gilt
denominated in GBP with a dirty price of GBP 95 at an FX rate of 1.5 USD/GBP (here V (t1 ) = 95
and χ (t1 ) = 1.5). His cash outflow at the beginning of the attribution period is thus 142.5 USD.
During the month, the investor receives a coupon payment of 5, i.e. C(t1 ,t2 ) = 5, this coupon, by
assumption, earns no interest until the end of the attribution period. At the end of the attribution
period, the investor sells the bond at a dirty price of GBP 93 (so V (t2 ) = 93). The FX rate at the
end of the period is 1.6 USD/GBP (χ (t2 ) = 1.6), at which time the bond is sold and the proceeds
from the sale and the received coupon are exchanged back into base currency. The cash inflow is thus
(93 + 5) × 1.6 USD/GBP = 156.8 USD. We can now easily compute the local total return, the total
return in base currency, and the currency return. From (3.9) the return in local currency is
93 + 5
rlocal = − 1 = 3.16%. (3.12)
95
From (3.10) the total return in base currency is
1.6 × (93 + 5)
rbase = − 1 = 10.04%. (3.13)
1.5 × 95
Finally, from (3.11) the currency return rcurrency is 10.04% − 3.16% = 6.88%.
3.2.1 Forward premium and currency surprise
It is possible to decompose currency appreciation (3.7) into two components (see Ankrim and Hensel,
1994):
1. Forward premium – The expected interest rate differential between two countries.
2. Currency surprise – The unexpected movement of the base currency relative to its forward ex-
change rate or market predicted rate.
By defining the forward exchange rate of the base currency at time t1 for conversion through a forward
contract at time t2 as Fχ (t1 ,t2 ), we can rewrite the currency appreciation as follows
[χ (t2 ) − Fχ (t1 ,t2 )] + [Fχ (t1 ,t2 ) − χ (t1 )]

rFX = . (3.14)
χ (t1 )
The term in the first set of brackets in the numerator is not known until t2 , whereas the term in the
second set of brackets is known at t1 . Hence, the first term is the currency surprise,
χ (t2 ) − Fχ (t1 ,t2 )

rsurprise = , (3.15)
χ (t1 )
and the second term is the forward premium,
Fχ (t1 ,t2 ) − χ (t1 )

rforward = . (3.16)
χ (t1 )
One of the benefits of splitting the currency return between forward premium and currency surprise
is that a portfolio manager can only hedge away the currency exposure at forward rates, he can therefore
only expect to neutralize the forward premium element on a fully hedged portfolio.
Example
If we take the same market scenario and UK Gilt from our previous example, and we suppose that the
1-month forward exchange rate is Fχ (t1 ,t2 ) = 1.515. Then from Eq. (3.7), the currency appreciation is
rFX = (1.6 − 1.5)/1.5 = 6.67%, (3.17)
the currency surprise, according to (3.15), is
rsurprise = (1.6 − 1.515)/1.5 = 5.67%, (3.18)
and the forward premium, according to (3.16), is
rforward = (1.515 − 1.5)/1.5 = 1%, (3.19)
In this example it is evident that the majority of the currency appreciation is due to the unexpected
movement of the base currency (USD) relative to its forward exchange rate.
Carry 15
3.3 Carry
The carry return on a bond is deterministic at the outset of the attribution period. It is the holding
period return on the bond if bond market conditions do not change, hence the factor is the time. Unlike
for equities, even if market conditions stay the same, the return on a coupon bond is not zero. First,
the coupon accrues deterministically. If we disregard reinvestment of any intra-period cash flows, the
coupon return on a bond between times t1 and t2 is
AI(t2 ) − AI(t1 ) +C(t1 ,t2 )
rcoupon = , (3.20)
P(t1 ) + AI(t1 )
where AI is the accrued interest, P is the clean price, and C(t1 ,t2 ) denotes any payments between times
t1 and t2 . C(t1 ,t2 ) typically refers to coupon payments but other deterministic payments between t1 and
t2 must also be taken into account (see Section 2.1 for details).
The roll-down return is the return that is due to the change in the clean price of the bond by rolling
time forward while keeping the yield curve that prevailed at t1 .
P(t2 |yield(t1 )) − P(t1 )
rroll = . (3.21)
P(t1 ) + AI(t1 )
The carry return is computed as the sum of Eqs. (3.20) and (3.21)
rcarry = rcoupon + rroll . (3.22)
Example
We illustrate the carry computation with an example. The example bond is the U.S. Treasury bond
with coupon rate 9 18 and maturity date May 15, 2018. We compute the carry on this bond between
October 30 and November 30, 2009. The dirty price on October 30 was V (t1 ) = 148.38. The accrued
interest was AI(t1 ) = 4.17 and hence the clean price was P(t1 ) = 144.22.1 This bond paid a coupon of
0.5 × 9.125 per 100 face value on October 15. The accrued interest AI(t2 ) = 0.38 is therefore less than
AI(t1 ). The coupon return is thus given by
0.38 − 4.17 + 0.5 × 9.125
rcoupon = = 0.522%. (3.23)
148.38
The computation of the roll-down return is a little more involved than computing the coupon return,
and requires repricing the bond. The first step is to calibrate our model to match the observed market
1 Small numeric differences in this section are due to rounding as the actual computations have been carried out with
higher precision than two decimals.
price. This is achieved by adding an option-adjusted spread (OAS) of 20.84 bps to the government
spot curve. Second, we need to compute a new ’time to cash flow’ vector. This vector is found by
setting the settlement date to t2 . The third step is to roll along the yield curve in time such that each
spot rate y(τi ) is replaced with y(τi − ∆t), where ∆t = t2 − t1 . We then reprice the bond using the OAS
at t1 and find that the hypothetical clean price at t2 is P(t2 ) = 143.98. The roll-down return can then be
computed as the hypothetical change in clean price divided by the dirty price at t1
143.98 − 144.22
rroll = = −0.156% (3.24)
148.38
In our example, the carry return on this bond is thus
rcarry = 0.522% − 0.156% = 0.366%. (3.25)
The roll-down return therefore includes two components: 1) a bond converges to its par value as it
approaches maturity. This is the pull-to-par effect. 2) Different rates are used to discount cash flows as
we are ’rolling down the yield curve’. A more mathematical treatment of the carry computations and
a proof of consistency with the Taylor series expansion framework introduced in Section 3.1 can be
found in Sorensen (2010a).
3.4 Yield curve
The term yield curve refers to a plot of yield to maturity on either par bonds or zero coupon bonds
against time to maturity in a certain market segment, such as Euro government bonds or U.S. corporate
AAA bonds. The yield curve reflects market participants’ expectations about the future, and as such,
summarizes borrowing costs for different maturities and issuers. RiskMetrics Group constructs zero
curves for specific maturities for many fixed income markets. Most often, it is the case that there exists
market prices for coupon bonds, but zero coupon prices are hard to get. In such cases, we estimate zero
coupon bond prices from coupon bond prices using a technique called bootstrapping (see, for example,
Malz, 2002 and Shi, 2009 for a description of the construction of curves).
Changes in the yield curve affect a number of market participants. First, movements in the yield
curve are one of the main drivers of returns on fixed income securities, and consequently affect bond in-
vestors. Second, the short rate is one of the main tools for macroeconomic policy. Third, the yield curve
in large part determines the costs of mortgage borrowing. The yield curve is therefore followed both
by investors, policymakers, and the general public. As shown by Litterman and Scheinkman (1991),
movements in the yield curve can in large part be summarized by the first two principal components.
These roughly correspond to a shift in the level of interest rates and a change in the slope of the term
Yield curve 17
structure. A model for returns due to a parallel shift of the yield curve is presented in Section 3.4.1. A
model that also incorporates returns due to slope changes is presented in Section 3.4.2.
3.4.1 Effective duration and convexity
The yield curve return is modeled as the bond’s return sensitivity to changes in the yield curve. The
traditional measure for capturing a bullet bond’s exposure to changes in interest rates was the Macaulay
duration or modified duration. In the case of plain vanilla fixed rate bullet bonds, there exist analytical
formulas to compute Macaulay and modified duration. However, for more complex instruments, such
as mortgage backed securities or bonds with embedded options, analytical formulas for duration are
not available. Instead, we compute the effective duration by taking the central finite difference
1 V (+δ y) −V (−δ y)
D=− , (3.26)
V 2δ y
where D is the effective duration, V is the current price and V(+δ y) and V (−δ y) are the prices obtained
by shifting the whole yield curve up and down by δ y basis points respectively. For complex fixed
income instruments, duration often loses its interpretation as the weighted average maturity of the
instrument. For instance, Interest Only (IO) strips typically have negative duration. In such instances,
it is more fruitful to consider effective duration and think of it as a measure of a security’s interest
rate sensitivity. The benefit of using effective duration is that by applying shifts and repeatedly calling
the pricing function, the chain rule is automatically taken into account. For example, if we consider a
callable bond, the effective duration will automatically capture the interest rate sensitivity of both the
bullet bond and the embedded option. That is, the second term of the effective duration in Eq. (3.26)
will be approximately equal to
∂ Vcb ∂ Vcb ∂ Vbb ∂ Vcb ∂ Vo
= + (3.27)
∂y ∂ Vbb ∂ y ∂ Vo ∂ y
∂ Vbb ∂ Vo
= − . (3.28)
∂y ∂y
where Vcb denotes the market value of the callable bond, Vbb the market value of the embedded bullet
bond, and Vo the market value of the embedded option.
Duration represents the first-order term from a Taylor series expansion of bond prices with respect
to yields. As such, it represents a linear approximation of the sensitivity of prices to yield changes.
However, the relation between bond prices and interest rates is not linear. Second-order terms are
captured by convexity. Effective convexity is computed as
1 V (+δ y) − 2V +V (−δ y)
C= , (3.29)
V (δ y)2
While the convexity contribution is often small, it can be substantial. Some portfolio strategies, such
as the bullet versus barbell, are explicit convexity strategies. We therefore include the convexity effect
of a parallel shift in yields in our attribution model. The return contribution due to a parallel shift in
yields is thus
1
rshift = rduration + rconvexity = −D∆y + C(∆y)2 . (3.30)
2
The effective duration D and the convexity C in Eq. (3.30) represent sensitivities to the average change
in the yield curve. RiskMetrics Group provides spot rates for a large number of maturity nodes. The
average change in the yield curve depends on which nodes we choose to compute the average. For
consistency with our approach for key rate durations in Section 3.4.2, we compute the average change
in yields over a set of K key rates
1 K
∆y = ∑ ∆yk , (3.31)
K k=1
where ∆yk is shorthand notation for the change in the spot rate for constant maturity τk between the
start date (t1 ) and end date (t2 ) of the attribution period
∆yk = y(t2 , τk ) − y(t1 , τk ). (3.32)
While effective duration is a much used and important tool in portfolio analysis and risk manage-
ment, it is important to understand that it assumes that yield curve changes occur in a parallel way.
However, in practice, yield curve changes are often far from parallel. Figure 3.1a shows the yield
curves as of month-end October and month-end November, 2009. Visual inspection reveals that the
shift was largest in absolute value in the medium term. Figure 3.1b plots the changes in yields for each
time to maturity and shows that yields in general decreased between October and November and that
the maximum decrease was around the five year point. The figure makes it clear that this was not a
parallel shift.
3.4.2 Key-rate durations
To remedy the shortcomings of the traditional duration measures, Ho (1992) introduced the key rate
durations (KRD). The key rate durations are computed by considering a factor model, where the factors
are the chosen key rates for the currency in question. Key rate duration is the negative of the percentage
price change due to a small shift in that key rate k and is defined as
1 V (+δ yk ) −V (−δ yk )
KRDk = − , (3.33)
V 2δ yk
Yield curve 19
Figure 3.1
Yield Curves and Yield Curve Changes
(a) Yield curves (b) Change in yields
5 10
5
4 0
−5
∆ Yield (bps)
3
Yield [%]
30−Oct−2009 −10
30−Nov−2009 −15
2
−20
1 −25
−30
0 −35
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Maturity [years] Maturity (years)
where δ yk is the shift applied to the k-th key rate. See, for example, Sorensen (2010b) for a detailed
description of the implementation of key rate durations. The yield curve component of a bond’s return
is the part of the return that can be explained by movements in the bond’s benchmark curve. Figure 3.2
shows the effective duration and key rate durations for a U.S. Treasury benchmark as of October 30,
2009.
The key rates depends on the market and for example, we use the following key rates for the U.S.
Treasury,:
Key rates = {6m, 2y, 5y, 10y, 20y, 30y}, (3.34)
which have been found representative of the entire yield curve.
We use the key rate durations to capture the contributions to returns that come from a reshaping
of the yield curve. It is therefore the effect over and beyond the parallel shift of the yield curve. The
additional contribution to returns from key rate durations is labeled the shape return. For each key rate,
we compute the difference between that key rate duration and the effective duration divided by the
number of key rates. This difference is multiplied with the change in the corresponding key rate and
summed over all key rates such that
K
rshape = − ∑ KRDk ∆yk + D∆y. (3.35)
k=1
Figure 3.2
Effective duration and key rate durations
Effective Duration
4
Key Rate Durations
Duration [Years]
0
0 5 10 15 20 25 30
Key Rates [years]
Combining Eqs. (3.30) and (3.35), we get the return contribution from government yield curve changes
ryield curve =rshift + rshape

K
1
= − ∑ KRDk ∆yk + C(∆y)2 . (3.36)
k=1 2
An appealing feature of key rate duration models is that, using six key rates as for US Treasury, we
can capture more changes than with the usual twist and butterfly effects. And further unlike for princi-
pal components models, we do not require a stationary covariance structure of interest rate changes
Example
We continue with our example bond from Section 3.3 and consider the return from month-end October
to month-end November, 2009 on the 9 18 U.S. Treasury bond with maturity date May 15, 2018. This
U.S. Treasury bond has no embedded option or risk of default. Hence, the two main drivers of returns
will be the carry and the yield curve return. The total return on the bond between October and Novem-
ber, 2009 was 1.83%. As we showed in Section 3.3, the carry return was 0.37%. Table 3.1 shows the
returns due to key rate duration exposures. The header row shows the term of each key rate. Since
this is a U.S. bond, the key rate nodes are those in Eq. (3.34). The first and second rows show the zero
Yield curve 21
Table 3.1
Key rate yields and durations
Time to Maturity [Yrs] 0.5 2 5 10 20 30

Key Rates Start [%] 0.16 0.90 2.35 3.56 4.30 4.55
Key Rates End [%] 0.16 0.67 2.04 3.35 4.27 4.57
∆ Key Rates [%] 0.00 -0.23 -0.31 -0.21 -0.03 0.02
KRD 0.05 0.33 2.28 3.76 0.00 0.00
KRD Return [%] 0.00 0.08 0.71 0.79 0.00 0.00
coupon yield for each key rate node at the beginning and end of the attribution period. The third row
displays the change in key rate yields and shows that the medium term rates decreased significantly
over the period. The fourth row shows the key rate durations. The last row shows the returns due to
exposures to the key rate durations and is computed by multiplying the negative of each of the key rate
durations with the realized shift in the corresponding key rate. The sum across key rates of the key rate
duration returns in Table 3.1 is 1.58%.
Table 3.2 shows the return decomposition when taking the carry, shift, and shape return into ac-
count. The second column shows the decomposition of our sample bond. The average change in the
key rates of the U.S. government curve was ∆y = −0.127%, which yields −D∆y = 0.82%. In this
example, the positive return due to convexity is only 0.004% and is indistinguishable from zero in the
table given our level of precision. The additional return due to a non-parallel shift of the yield curve is
thus rshape = 1.58% − 0.82% = 0.76%. The shift and shape returns are of similar size, and illustrate the
importance of taking non-parallel movements in they yield curve into account. The decomposition of
returns is done security by security. We can aggregate the security level results and decompose the re-
turns of a portfolio. The last column of the table provides the results for a Treasury index. The signs of
the return components are the same as for the example bond, but the shift and shape returns are smaller
in magnitude. This is consistent with the lower duration of the index. Both return decompositions
leave a small but not negligible residual.
Table 3.2
Treasury return decomposition
1
Return US T 9 8 US T Index
Total 1.83% 1.40%
Carry 0.37% 0.28%
Shift (Duration & Convexity) 0.82% 0.67%
Shape 0.76% 0.36%
Residual -0.12% 0.09%
Duration 6.44 5.22
3.5 Credit
Every bond that is not backed by the full faith and credit of the government is subject to credit risk.
Investors command a spread over government yields to accept the possibility of default when investing
in corporate bonds. Movements in the spread versus Treasuries therefore influence bond prices and
returns. It is generally agreed that there is no risk of default by investing in U.S. Treasuries. However,
bond prices for other governments may reflect a probability of default.
RiskMetrics Group provides a large number of curves. The level of detail depends on the market.
Figure 3.3 shows the yield curves for US Corp All Industry A, and USD Government on December 31,
2009. As expected, the government zero yields are below the corporate curve for all maturities.
The spread return captures the component of a bond’s return that is due to changes in credit markets.
Each bond is specified both with a curve for discounting cash flows and a reference government curve.
For each bond, we start by computing the total change in spread versus the government curve. Part of
this spread will be due to the change in spreads between the sector curve, for example US Corp All
Industry A, and the government curve. The part of the total spread change that is not explained by this
curve spread change is an issue-specific idiosyncratic spread change. The sensitivity to spread changes
is assessed by the spread duration and spread convexity. Spread duration is computed as
1 V (+δ s) −V (−δ s)
Dspread = − , (3.37)
V 2δ s
where Dspread is the spread duration and δ s is the shift applied to the spread to compute the central
Credit 23
Figure 3.3
Yield curve for US Corp All Industry A and USD Government on 31-Dec-2009
5
Yield [%]
3 USD Govt
US Corp All Industry A
2
0
0 5 10 15 20
Maturity [Years]
difference. The spread convexity is computed as
1 V (+δ s) − 2V +V (−δ s)
Cspread = . (3.38)
V (δ s)2
For each bond, we compute the option-adjusted spread, s. This is the constant spread that we must
add to the spot government curve in order to match the model price of the bond with the traded price.
This spread will in general change from s(t1 ) to s(t2 ) between dates t1 and t2 . The spread change,
∆s = s(t2 ) − s(t1 ), can be decomposed into the change in the swap spread, the sector spread versus the
swap curve, and an idiosyncratic spread
∆s = ∆sswap + ∆ssector + ∆sidiosyncratic , (3.39)
where ∆sswap = ∆(yswap −ygovernment ) is the average change in spread between the swap and government
curve. The average is computed over the set of key rates. Similarly, ∆ssector = ∆(ysector − yswap ) is the
average change in spread between the sector and swap curve. The sector curves are constructed using
bootstrapping methods on bonds in specific rating and industry segments; US Corp Financial Aa and
US Corp Utility Baa are two examples of what we define as sector curves. Shi (2009) gives a detailed
description of how these curves are constructed.
The return contribution from widening or tightening sector spreads can be modeled as common
factor returns. The change in the idiosyncratic spread, which can be due to company-specific news or
liquidity changes for a specific issue, contributes to the residual return on the bond. This is the part of
the bond’s total return that common factor changes cannot account for.
The contribution from a change in spreads can be approximated by −Dspread ∆s + 12 Cspread (∆s)2 .2
Using the spread decomposition in Eq. (3.39), we define the spread return, rspread , as the return due to
changes in the swap and sector spread
1 1
rspread = −Dspread ∆sswap + Cspread (∆sswap )2 −Dspread ∆ssector + Cspread (∆ssector )2 . (3.40)
| {z2 }| {z2 }
swap spread return sector spread return
Example
In this example, we decompose returns from month-end November to month-end December, 2009 on a
corporate bond issued by Coca Cola Enterprises. The bond has a coupon rate of 7% and a final maturity
date on October 1, 2026. It is mapped to the discount curve US Corp All Industry A. We find that the
spread duration of this bond was Dspread = 10.47 and the spread convexity was Cspread = 146. The
spread movements are reported in Panel A of Table 3.3. At the beginning of the attribution period, the
bond spread versus the government curve was 130.50 bps. By year-end, the bond spread had decreased
to 102.50 bps, a decrease of 28 bps. Next, we consider the spread change in the swap versus the
government curve. Using the available key rates, we find that this spread increased by ∆sswap = 1.68
bps, yielding a negative contribution to the return on the bond by rswap = −18 bps. Finally, we consider
the spread change in the US Corp All Industry A versus the swap curve, ∆ssector = −18.23 bps. The
contribution from this spread change is rsector = 193 bps. The sum of these two contributions gives the
common factor spread return, rspread = rswap + rsector = 176 bps.
Panel B of Table 3.3 shows the expanded return decomposition for this bond. The total return was
negative, −2.47%, despite the positive return contribution from spread returns. This was due to an
increase in the Treasury yield curve which is captured by the shift and shape returns, as described in
Section 3.4. A visual representation of the return decomposition is given in Fig. 3.4, which clearly
shows that the positive contribution from the spread return is not large enough to outweigh the negative
contribution from the shift and shape returns. Even after accounting for carry, shift, shape, and spread
returns, the remaining residual is not small in this case. This is because Coca Cola spreads decreased
more than spreads on the US Corp All Industry A curve between November and December.
2 Recently, Dor et al. (2007) have suggested that duration times spread be used instead. They show that portfolios with
different spreads and spread durations, but similar duration times spread (DTS) exhibit the same return volatility.
Credit 25
Table 3.3
Spread return and corporate bond return decomposition
Statistic Coca Cola 7% 1-Oct-2026

Panel A: Spread Changes
Dspread 10.47
Cspread 146
∆s -28.00 bps
∆sswap 1.68 bps
∆ssector -18.23 bps
Panel B: Return Decomposition
Total -2.47%
Carry 0.46%
Shift -4.71%
Shape -1.25%
Spread 1.76%
Residual 1.27%
Figure 3.4
Corporate bond return decomposition
2 Spread
Residual
Carry
Total Shift Shape
Return [%]
−2
−4
−6
3.6 Volatility
Several fixed income instruments include options. For such instruments, the market value depends on
interest rate volatility. A prominent example is callable bonds, which are standard fixed coupon bonds,
but where the issuer retains the right to buy back the bond at par during a specified period before
maturity. An investor who is long a callable bond is thus long a regular bond and short a call option
on the bond. Hence, part of the return on the callable bond will be driven by the option’s price change.
The value of the option increases with an increase in interest rate volatility. This implies that the value
of the callable bond decreases with an increase in volatility.
Figure 3.5
Cap volatilities on two dates: 29-May-2009 (blue) and 5-Jun-2009 (red)
0.8
Cap volatility
0.6
0.4
0.2
0
0 2 4 6 8 10
Maturity [years]
To value embedded options, we resort to models for the interest-rate dynamics, for example the
Black-Derman-Toy model (Black et al., 1990). These models typically need a market-implied volatility
curve for calibration. Hence any movement in such curve will imply a price change. In Fig. 3.5 we
show the volatility curves obtained from cap prices at two different dates.
Our volatility factor is defined as the average shift of the market-implied volatility curve over our
K key nodes,
1 K
∆σ = ∑ σk (t2 ) − σk (t1 ) (3.41)
K k=1
with the key nodes being
k ∈ {1y, 2y, 3y, 5y, 10y}. (3.42)
Volatility 27
The exposure to the volatility factor is given by the bond’s vega. To compute vega, the volatility curve
is shifted up and down by a small amount δ σ (e.g. 1bp) and obtained as
V (+δ σ ) −V (−δ σ )
v= , (3.43)
2δ σ
where V (+δ σ ) is the instrument value obtained by increasing all the volatility curve by +δ σ and
V (−δ σ ) is the instrument value obtained by decreasing all the volatility curve by −δ σ , the other
factors being constant.
The volatility return is then obtained as

1
rvolatility = v∆σ . (3.44)
V
Example
Table 3.4
Characteristics of the two bonds
Callable bond Regular bond

Coupon 5% 5%
Maturity date 3-Sep-2020 3-Sep-2020
Country US US
Strike date 3-Oct-2015
Strike price 100
To illustrate the volatility factor we study two similar bonds. The first is a simple US Treasury
callable bond and the second is a similar bond, but with no embedded option. Table 3.4 shows the
characteristics of the bonds. These bonds have been priced using RiskMetrics models on two consec-
utive dates, June 1st and 2nd 2009, and we report the various factor sensitivities in Table 3.5.
We first notice that the effective duration of the callable bond is significantly lower than that of
the regular bond. Indeed lowering the interest rate will increase the chance the bond is called and
hence reduce the value of the bond. Further we see a zero vega as expected for the regular bond and, a
negative vega for the callable bond because the higher the volatility the higher the chance to meet the
strike price, hence the lower the price.
During the two analysis dates, the cap volatility surface has decreased by
∆σ = −1.149% (3.45)
Table 3.5
Duration, convexity and vega of the two bonds
Callable Bond Regular Bond

Effective duration 6.95 8.27
Effective convexity 0.731 0.838
Vega -0.163 0.000
Table 3.6
Return decomposition for the two bonds
Return Callable Bond Regular Bond

Total 0.538% 0.548%
Carry 0.017% 0.015%
Duration 0.264% 0.315%
Convexity 0.004% 0.006%
Volatility 0.187% 0.000%
Residual 0.066% 0.212%
Inflation 29
and the US Treasury curve has decreased by
∆y = −0.038%, (3.46)
yielding the results in Table 3.6. We see that the duration effect is not as important for the callable bond
as for the regular bond, but this is balanced by the volatility effect. We acknowledge that the residuals
are still large and a full key-rate duration decomposition should be applied.
3.7 Inflation
In Section 3.4, we assessed the impact of nominal yield curve changes on bond returns. In this section,
we decompose returns due to changes in the government curve into those caused by changes in real
rates and those caused by changes in break-even inflation.
The consumer price index (CPI) is a measure estimating the average price of consumer goods and
services purchased by households. This index can be used to adjust for the effect of inflation on the real
value of money, salaries, pensions, and regulated or contracted prices. Many governments have issued
inflation-linked bonds. The relevant quantity for pricing these bonds is the break-even inflation, as
shown in Couderc (2008). The break-even inflation is not only the expected inflation, it also contains
a compensation that investors require to accept inflation risk. Nominal interest rates yt (τ ) can be
expressed as
1 + yt (τ ) = [1 + rt (τ )] [1 + it (τ )] , (3.47)
where rt (τ ) is the real yield curve and it (τ ) is the break-even inflation curve. RiskMetrics computes
the break-even inflation curve from nominal and inflation-indexed securities as shown for illustration in
Fig. 3.6. Here we plot the Euro nominal curve and the French break-even inflation curve on a particular
date.
We neglect the cross-product term and rewrite Equation (3.47) as
yt (τ ) = rt (τ ) + it (τ ). (3.48)
The real yield curve is then obtained by subtracting the break-even inflation from the nominal yield
curve.
When the security has no dependency on inflation, its price will change only when the real curve
moves. To capture such effects we need to introduce dual durations, namely the real duration
1 ∂V
Dr = − (3.49)
V ∂r
Figure 3.6
EUR nominal yield curve and France break-even inflation curve on 11-May-2006
Nominal
3 Inflation
Yield [%]
0
0 5 10 15 20 25 30
Maturity [year]
and the inflation duration

1 ∂V
Di = − . (3.50)
V ∂i
As illustration consider a simple zero-coupon inflation-linked bond paying a face value F scaled by
inflation at maturity T . Consider further a flat real yield curve r(τ ) = r and a flat break-even inflation
curve i(τ ) = i. The value of the bond is then calculated as
F(1 + i)T
V= . (3.51)
(1 + i)T (1 + r)T
For such bond, we see that Di = 0 and Dr = T /(1 + r).
As for the Treasury curve effect we treat parallel shifts in real and break-even inflation rates as
factors. Consistent with the computation of the parallel shift for the nominal yield curve in Eq. (3.31),
the parallel shift in the real interest rate curve is
1 K
∆r = ∑ ∆rk ,
K k=1
(3.52)
where ∆rk = rt2 (τk ) − rt1 (τk ). The same set of K key nodes is used to compute both the nominal and
the real shift. The parallel shift in the break-even inflation curve is similarly
1 K
∆i = ∑ ∆ik ,
K k=1
(3.53)
Inflation 31
where ∆ik = it2 (τk ) − it1 (τk ) is the change in the break-even inflation curve between two dates for the
constant maturity τk .
The effective real duration of the security is calculated as
1 V (+δ r) −V (−δ r)
Dreal = − , (3.54)
V 2δ r
where V (+δ r) (V (−δ r)) is the value of the security calculated using the real yield curve shifted up
(down) with δ r. The inflation duration is similarly computed as
1 V (+δ i) −V (−δ i)
Dinflation = − , (3.55)
V 2δ i
where V (+δ i) (V (−δ i)) is the value of the security calculated using the inflation curve shifted up
(down) with δ i.
The returns due to real yield curve and inflation effects are then
rreal = −Dreal ∆r, (3.56)
and
rinflation = −Dinflation ∆i. (3.57)
The sum of the contribution from Eqs. (3.56) and (3.57) is equal to rshift in Eq. (3.30) less the contri-
bution from convexity for non-inflation linked bonds.
Example
Table 3.7
Characteristics of the two bonds
Inflation-linked Bond Regular Bond

Coupon 2.5% 2.5%
Maturity date 17-Jul-2020 17-Jul-2020
Country US US
CPI USCPI –
To illustrate the inflation factor we study two similar bonds. The first is an inflation-linked bond
and the second is a similar bond, but with no inflation protection. Table 3.7 shows the characteristics
Table 3.8
Dual durations of the two bonds
Duration Inflation-linked Bond Regular Bond

Nominal – 6.36
Inflation 0.0018 6.36
Real 6.41 6.36
of the bonds. These bonds have been priced using RiskMetrics models on two consecutive dates, June
1st and 2nd 2009, and we report the various factor sensitivities in Table 3.8.
We first notice that all three durations of the regular bond are equal and that the inflation duration
of the inflation-linked bond is almost zero (up to non-linear pricing effects), as expected. During the
two analysis dates, the factor realizations are
∆i = −0.038% (3.58)
and
∆r = −0.010%. (3.59)
This yields the results in Table 3.9, where we notice that the inflation-linked bond has zero return
attributed to inflation, as expected.
Table 3.9
Return decomposition for the two bonds
Return Inflation-linked Bond Regular Bond

Total 0.098% 0.531%
Carry 0.013% 0.013%
Inflation 0.000% 0.238%
Real 0.066% 0.065%
Residual 0.019% 0.214%
Chapter 4
Relative Performance Attribution
4.1 Introduction
After having measured the total return of each fixed income security in the portfolio, and further decom-
posed it into various effects using factors, in this chapter we look at the attribution of the performance
of the portfolio relative to the benchmark. We will first describe the method for a single period, which
in most cases is daily and then show how to aggregate these daily results to the reporting period, e.g.
monthly.
For a single period, the return of the portfolio is obtained as
RP = ∑ wPi riP (4.1)

i
and the return of the benchmark as

RB = ∑ wBi riB , (4.2)
i
where wPi and wBi are the weights (exposures) of position i in the portfolio and benchmark, while riP
and riB their return over the period. For a given security, these returns can be different for example due
to trading, see Section 2.2.
As stated in Chapter 1, the portfolio manager is interested in his active return, i.e. his return relative
to the benchmark:
RP − RB = ∑ wPi riP − ∑ wBi riB . (4.3)
i i
The portfolio manager has access to all constituents and characteristics of the benchmark and he
34 CHAPTER 4. RELATIVE PERFORMANCE ATTRIBUTION
Figure 4.1
Hybrid performance attribution
takes various bets to try outperform it. Performance attribution is an ex-post analysis that attributes this
active return to the various bets the portfolio manager took.
The first method for performance attribution was introduced by Brinson and Fachler (1985) and
captures the investment decision by considering reference portfolios or semi-notional funds, which
boils down to grouping the assets differently. The asset-grouping approach consists essentially of two
main attribution effects, namely asset (or sector) allocation and security selection, and is very efficient
for equity portfolios. However this is not not fully satisfactory for fixed income portfolios, where many
other effects need to be captured, such as carry or yield curve movements.
Another approach is to resort to a factor model. One first decomposes the return of all securities into
systematic factors, and then aggregates all sensitivities as the attribution effect. However the drawback
is that some effects cannot be captured by a systematic factor.
We have developed a hybrid attribution method that take the best of both while leaving their dis-
advantages apart. The hybrid attribution method is schematically shown in Fig. 4.1. We first consider
as many factors as needed, which will leave an unexplained part (or residual), on which we apply a
standard asset-grouping method. The hybrid method has two interesting limiting cases. The first is
when no factors are used, this is then a standard asset-grouping approach. An emerging-market bond
manager might like to see the results from this attribution method, if the market is not liquid enough
to produce reasonable yield curves. While the other limit is when we use enough factors to produce
negligible residuals. This approach could be the choice of a Treasury fund manager.
Pure asset-grouping approach 35
In the following we first describe the two limiting cases, and then introduce the hybrid approach.
We also show to link the single attribution period to a longer reporting period and finally go through
the treatment of mispricing.
4.2 Pure asset-grouping approach
In addition to the notation introduced in Section 4.1 we suppose the universe of securities is partitioned
in sectors (labeled by J) and introduce the following notation:
wBJ is the weight of sector J in the benchmark.
wPJ is the weight of sector J in the portfolio.
rJB is the return of sector J in the benchmark.
rJP is the return of sector J in the portfolio.
The sectors typically depend on the investment process the portfolio manager follows, and can be
industrial sectors, ratings, or asset classes. Further there are no restrictions to consider also sub-sectors,
or more complex hierarchies.
The weights of sector J in the portfolio and benchmark are given by
wPJ = ∑ wPi wBJ = ∑ wBi (4.4)

i∈J i∈J
and the scaled returns of sector J are

1 1
rJP = P ∑
wJ i∈J
wPi riP rJB = B ∑
wJ i∈J
wBi riB (4.5)
Using Equation (4.1) and Equation (4.2), the active return can be written as
RP − RB = ∑ wPJ rJP − ∑ wBJ rJB , (4.6)

J J
where the summation is now over all sectors.
A standard way to decompose this active return is to group the terms differently, such as to make
apparent a sector allocation attribution term and a security selection term. The allocation and selection
attribution term at the sector level are shown in Table 4.1. For each sector J, the allocation term is the
relative weight of the sector (wPJ − wBJ ) times the over/underperformance of the sector relative to the
total benchmark return (rJB − RB ). The selection term for sector J is the portfolio weight of the sector
wPJ times the over/underperformance of the sector return (rJP − rJB ). The sum over all sectors of the
allocation and selection returns is equal to the total excess return over the benchmark.
Table 4.1
Asset-grouping performance attribution
Sector Sector Allocation Security Selection

.. .. ..
. . .
J (wPJ − wBJ )(rJB − RB ) wPJ (rJP − rJB )
.. .. ..
. . .
Total ∑J (wPJ − wBJ )rJB ∑J wPJ (rJP − rJB )
In this formulation we have chosen a top-down security selection, including also the interaction
term. There are various variations of such method as described in Soor (2010).
4.3 Pure factor model approach
When the factor model introduced in Chapter 3 yields a negligible residual we can use an attribution
based on factors only. Assuming a zero residual in Eq. (3.1) we decompose the return of all securities
in the portfolio and benchmark as
riP = ∑ βi,k
P
fk riB = ∑ βi,k
B
fk . (4.7)
k k
The beta-sensitivities to factor k of the portfolio and benchmark are
βkP = ∑ wPi βi,k

P
βkB = ∑ wBi βi,k
B
. (4.8)
i i
The factor-based attribution terms are shown in Table 4.2. For each factor k, the allocation term
is the relative beta (βkP − βkB ) times the realization of the factor fk . The sum over all factor returns is
equal to the total excess return over the benchmark.
In case a factor is only uniquely defined for a sector, for example credit spreads for a rating sector,
we extend the attribution as follows. We label the sector-specific factor as fJ,k and introduce the sector-
Hybrid attribution model 37
Table 4.2
Factor-based performance attribution
Factor Attribution
.. ..
. .
k (βkP − βkB ) fk
.. ..
. .
Total ∑k (βkP − βkB ) fk
specific sensitivity for the portfolio and benchmark as
βJ,k
P
= ∑ wPi βi,k
P
βJ,k
B
= ∑ wBi βi,k
B
. (4.9)
i∈J i∈J
Notice that here the sector-specific sensitivities are not rescaled by the sector weight as for the sector
returns (4.5). The attribution to factor k in sector J is then
( P )
βJ,k − βJ,k
B
fJ,k (4.10)
where the factor realization fJ,k is the same for all securities in the sector.
Finally, in case the residuals are not negligible we can aggregate them as an additive term:
∑(εiP − εiB). (4.11)

i
4.4 Hybrid attribution model
In some cases it is difficult to construct a consistent factor for a given effect, or the investment process
does not consider a common factor for all securities. We then use a mixed approach. In that approach,
first the return is decomposed using factors, with an idiosyncratic or residual term remaining. Second,
this idiosyncratic term is explained using the asset-grouping approach. For example in the case of
corporate bonds, the factors could be carry (Section 3.3) and yield curve (Section 3.4), and the residual
would include the return due to credit spread movements and all other specifics.
We use the notation introduced in Sections 4.2 and 4.3 to write the return of security i in the
portfolio and benchmark as
riP = ∑ βi,k
P
fk + εiP riB = ∑ βi,k
B
fk + εiB (4.12)
k k
where we introduce the residual return as εiP for the portfolio and εiB for the benchmark.
Using the portfolio and benchmark beta defined in Eq. (4.8), the active return is then
( )
RP − RB = ∑(βkP − βkB ) fk + ∑ wPi εiP − wBi εiB . (4.13)
k i
We introduce the sector residuals
εJP = ∑ εiP εJB = ∑ εiB , (4.14)

i∈J i∈J
to obtain the hybrid attribution terms by treating factors as in Section 4.3 and residuals as in Section 4.2.
Attribution results are shown in Table 4.3.
Table 4.3
Hybrid performance attribution
Factor Sector Factor Attribution Sector Allocation Security Selection

.. .. .. .. ..
. . . . .
k J (βJ,k
P −βB )f
J,k J,k – –
.. .. .. .. ..
. . . . .
Residuals J – (wPJ − wBJ )εJB wPJ (εJP − εJB )
Total ∑k,J (βJ,k − βJ,k
P B )f
J,k ∑J (wPJ − wBJ )εJB ∑J wPJ (εJP − εJB )
Example
To illustrate the hybrid attribution method we analyze a small corporate bond portfolio. As shown in
Fig. 4.2 we consider two factor effects, carry and duration, and apply the asset-grouping approach to
the residual, which will contain the return due to the sector credit spread movements.
The portfolio and benchmark positions are listed in Table 4.4. The holding period is one week,
from 29 May 2009 to 5 June 2009 and we show the resulting sector allocation and duration on 29 May
2009 in Table 4.5. The active return is 29.32 bps.
The first step consists of applying our factor model yielding the decomposition shown in Table 4.6.
The shift of the yield curve was + 0.26%, resulting in negative returns for all bonds, larger than the
carry return in absolute terms. However, not all bonds have negative total returns, due to different
Hybrid attribution model 39
Figure 4.2
Corporate bond attribution
Table 4.4
Universe of corporate bonds
Security Sector Portfolio Weight Benchmark Weight

CAM 6.375 07/15/2018 Energy 10.1% 2.8%
CSGN 6.125 11/15/2011 Financials 0.0% 23.5%
CSGN 6.5 01/15/2012 Financials 71.3% 19.8%
CSGN 6 02/15/2018 Financials 0.0% 12.5%
NRC 4.375 10/01/2010 Financials 0.0% 3.4%
NRC 7.25 03/01/2012 Financials 0.0% 11.0%
VZ 4.35 02/15/2013 Telecom 18.6% 5.2%
VZ 5.5 02/15/2018 Telecom 0.0% 10.2%
VZ 6.4 02/15/2038 Telecom 0.0% 11.6%
Table 4.5
Sector allocation and duration
Portfolio Benchmark
Sector weights Energy 10.1% 2.8%
Financials 71.3% 70.2%
Telecom 18.6% 27.0%
Duration 3.0 4.5
Table 4.6
Return decomposition of all bonds
Security Sector Total Return Carry Return Duration Return Residual

CAM 6.375 07/15/2018 Energy 0.21% 0.16% -1.69% 1.74%
CSGN 6.125 11/15/2011 Financials -0.83% 0.14% -0.60% -0.37%
CSGN 6.5 01/15/2012 Financials -0.82% 0.05% -0.62% -0.25%
CSGN 6 02/15/2018 Financials -0.88% 0.08% -1.68% 0.72%
NRC 4.375 10/01/2010 Financials -0.48% 0.08% -0.34% -0.23%
NRC 7.25 03/01/2012 Financials -1.34% 0.14% -0.64% -0.84%
VZ 4.35 02/15/2013 Telecom -1.01% 0.08% -0.88% -0.21%
VZ 5.5 02/15/2018 Telecom -1.26% 0.11% -1.76% 0.39%
VZ 6.4 02/15/2038 Telecom -2.08% 0.13% -3.07% 0.86%
Table 4.7
Hybrid performance attribution [in bps]
Sector Active Return Carry Duration Sector Allocation Security Selection

Energy 1.51 1.13 -12.26 12.64 0.00
Telecom 23.27 -1.57 41.57 -3.99 -12.75
Financials 4.54 -3.41 11.38 -0.22 -3.20
Total 29.32 -3.85 40.69 8.43 -15.95
Multi-period attribution 41
credit spread shifts included in the residuals. The next step is to apply the asset-grouping approach on
the residuals yielding the final attribution of the active return in Table 4.7.
These results are in agreement with the insights we have analyzing the characteristics of the port-
folio in Table 4.5. There is a large negative duration bet and since the Treasury curve increased during
the analysis period, we get a large positive attribution. We also notice that the sector allocation bets are
not large in comparison to stock picking bets, understandable as the portfolio only has one single bond
in each sector. This example also shows the importance of using factor for fixed income performance
attribution, as highlighted by the fact that duration attribution is much more important than the other
effects.
4.5 Multi-period attribution
So far we have only considered a single attribution period, which is thought to be daily. This is the
basic building block for longer attribution period, such as monthly or quarterly. As we get holdings of
the portfolio including transactions and the constituents of the benchmarks on a daily frequency. To
report the performance attribution over a longer period, we need to aggregate the daily calculations.
For example, to get the performance attribution of an European manager in February 2010, we need to
aggregate the 20 daily total returns and attribution.
The returns that we considered so far were discrete relative returns from t to t + 1 day, defined as
Vi (t + 1)
ri,t,t+1 = − 1, (4.15)
Vi (t)
where Vi (t) is the market value of the instrument i at time t (measured in days). This formulation is
very useful because the return of a portfolio, from time t to t + 1, can be calculated as the weighted
sum of these discrete relative returns,
Rt,t+1 = ∑ wi ri,t,t+1 (4.16)

i
where wi is the portfolio weight of instrument i at time t. This weight should represent the exposure of
the position and is not always simply its market value (see Chapter 5). Further this is consistent with
the arithmetic attribution that we have introduced in Section 4.4. In other words the active return can
k
be decomposed in a sum of attribution terms At,t+1 , as
P
Rt,t+1 − Rt,t+1
B
= ∑ At,t+1
k
(4.17)
k
We consider now a reporting period, from t1 = 0 to t2 = T , consisting of T daily steps. The return
of security i over that reporting period is
Vi (T )
ri,0,T = −1 (4.18)
Vi (0)
and the portfolio total return over that period is
R0,T = ∑ wi ri,0,T (4.19)

i
under a static portfolio assumption. The problem is to account for daily transactions and that factor
sensitivities evolve during the reporting period, yielding the question of what value to take for the
attribution, i.e. start value or average?
The solution is to perform daily attribution, accounting for any transactions and then aggregate
the attributions to the reporting period. The problem is that discrete relative return and arithmetic
attribution do not add up over time. A number of methodologies (known as smoothing algorithms)
have been developed to achieve this (see Bacon (2008)) and a comparison of these methodologies can
be found in Menchero (2004). Among them is a simple yet effective methodology proposed by Carino
(1999) for the asset-grouping approach.
The idea is to resort to logarithmic returns defined as

Vi (t + 1)
r̃i,t,t+1 = log (4.20)
Vi (t)
with the property that the full period return is obtained by summing up the daily returns,
Vi (T ) T −1
r̃i,0,T = log = ∑ r̃i,t,t+1 . (4.21)
Vi (0) t=0
In the following ( ˜ ) stands for log-returns.
The daily arithmetic attributions are transformed in daily logarithmic attributions, summed up to
get the attributions over the reporting period and then transformed back to arithmetic attributions. To
that purpose we define the daily smoothing factor
P
R̃t,t+1 − R̃t,t+1
B
κt,t+1 = . (4.22)
P
Rt,t+1 − Rt,t+1
B
P
In case, Rt,t+1 B
= Rt,t+1 we set it to κt,t+1 = 1. While the full period smoothing factor is defined as
R̃P0,T − R̃B0,T
κ0,T = . (4.23)
RP0,T − RB0,T
Impact of pricing difference 43
Writing the daily attribution as Eq. (4.17) we get the daily logarithmic active return decomposed as
P
R̃t,t+1 − R̃t,t+1
B
= ∑ κt,t+1 At,t+1
k
. (4.24)
k
These attribution terms are summed up over the reporting period and transformed back to relative
returns. Using 4.23, the aggregated attribution effects are then
κt,t+1 k
Ak0,T = ∑ A (4.25)
t κ0,T t,t+1
providing the following decomposition of the active return of the reporting period:
RP0,T − RB0,T = ∑ Ak0,T . (4.26)

k
4.6 Impact of pricing difference
It has already been mentioned that security returns in the portfolio may be different from corresponding
security returns in the benchmark due to trading, see Section 2.2. Another possible difference in returns
can occur due to differences in the benchmark’s and portfolio’s market prices from which returns are
calculated.
It is highly likey that a portfolio manager may take market prices from a different source than the
source which is used by a benchmark vendor, which may lead to a different price in the portfolio and
benchmark for a given security. Also, a vendor’s price may be calculated from an arithmetic mean
of two price quotes, which would also lead to a potential difference with the portfolio price (taken
from a single quote).1 For example Iboxx follows this consolidated price or mid-price approach. Mid-
prices are also used in FTSE indices, again with prices calculated as the arithmetic mean between two
quotes.2
It is easy to see that ambiguity may arise as prices are available from multiple sources as well as
the use of mid-prices for a given security. As accurate attribution relies on an accurate performance
calculation, if there is a difference in prices between the benchmark and the portfolio we must remove
this before calculating the active return. Otherwise, the impact of pricing differences could compromise
the integrity of the attribution analysis.
1 Government bonds traditionally have far more price quotes than corporate bonds which will increase the likelihood of
pricing differences.
2 FTSE apply a quality control check to ensure each single price in the index is unbiased and representative of the market
so they generally do not include a bond issue if there is only one price available from a single market maker.
To remove the impact of pricing differences on portfolio returns it is necessary to adjust portfolio
prices. Let πi be the difference (only due to different price sources) between the portfolio price and the
benchmark price
πi (t) = ViB (t) −ViP (t), (4.27)
where ViP (t) and ViB (t) are the market prices of instrument i in the portfolio and benchmark, respec-
tively, at the end of a given day t. Then ṼiP (t) is the adjusted portfolio price for security i at time
t
ṼiP (t) = ViP (t) + πi (t). (4.28)
It is then possible to calculate the adjusted return, r̃iP , over the holding period by substitution of (4.28)
into the equation for the local total return of an instrument (2.1). Then the adjusted portfolio return,
R̃P , due to adjusted portfolio prices is equal to
R̃P = ∑ wPi r̃iP . (4.29)

i
The impact (or return), RI , due to the pricing difference, is the difference between the adjusted bench-
mark return and the benchmark return,
RI = RP − R̃P . (4.30)
Finally the active return (4.3) less the impact of any pricing difference RI is
R̃P − RB = ∑ wPi r̃iP − ∑ wBi riB . (4.31)

i i
Chapter 5
Exposures
For attribution security weights according to market value is straightforward, this is simply the secu-
rity market value divided by the total portfolio market value. However, when the portfolio contains
derivatives calculation of weights becomes more involved. Rather than weight being calculated simply
according to market value instead it is more appropriate to calculate weights from the full exposure
value of each derivative. The problem with the simple market value approach is that the exposure of
the portfolio is vastly understated. We show how full exposure can be used to express derivative secu-
rities in terms of their underlying synthetic assets so that their returns contribute towards the portfolio’s
total return.
5.1 Partial and full exposure
The partial exposure of a security is its market value (or present value). The full exposure of a derivative
security is captured by decomposing the security according to its underlying synthetic assets. Only then
can the appropriate underlying exposure and thus the source of return of the derivative, due to a change
in value of its underlying asset(s), be captured.
For a basic instrument partial exposure is equivalent to full exposure, where basic implies the asset
is not a derivative instrument, examples include cash holdings and bonds. For a derivative asset the
partial exposure represents the value the asset could be sold for, whereas the full exposure quantifies
the economic exposure of the derivative due to movements in the underlying asset and as such, weights
calculated using the derivatives full exposure is more appropriate.
46 CHAPTER 5. EXPOSURES
By restating the portfolio in terms of the full exposure of each security the effective weight of
each position can be defined and used for performance measurement and attribution. All derivative
positions (futures, options, etc) are replaced by their underlying synthetic assets and notional cash for
which performance can be calculated exactly.
5.2 Performance legs and notional cash
As full exposure requires derivatives to be broken up into their underlying synthetic assets it is ap-
propriate to introduce the concept of performance legs. A performance leg is either the underlying
exposure value or notional cash. Note, all assets expressed in terms of their partial exposure have a
single performance leg which is always equal to the market value of the asset. Notional cash is the
amount of cash that is necessary to balance the exposure of an underlying asset such that the addition
of all performance legs is equal to the market value of the derivative. This adjustment to offset the un-
derlying exposure leg prevents derivative returns from being overstated, the notional cash is a synthetic
performance leg.
5.3 Portfolio weight
Let us define the weight of each position in a portfolio considered in terms of its partial exposure. For
a position i, where i = 1, . . . , N and where Vi is the market value or partial exposure of position i, the
weight wi of the position i in the portfolio is
Vi
wi = . (5.1)
∑i Vi
For a portfolio considered in terms of its full exposure, weights are calculated using performance
legs, where Ṽi is the value of each performance leg for a security i. For derivatives, Ṽi is the underlying
exposure value or the complementary notional cash and for basic assets, Ṽi is simply the market value.
Hence we define the effective weight w̃i as
Ṽi
w̃i = . (5.2)
∑i Ṽi
Portfolio weight 47
Examples
To illustrate the need for full exposure we move to several specific examples. We begin with the trivial
case of a bond to introduce the approach.
Bonds
For bonds full exposure is equivalent to partial exposure as they have a single performance leg, which
has value equal to the position market value.
Position market value = Quantity × Market price × FX rate. (5.3)
A note on currency rates is appropriate. Since the market price of the bond is in the security’s local
currency the FX rate scales the market price so that exposure value is in the base currency of the
portfolio. Quantity is the number of bonds held (which is > 0 for long positions and < 0 for short
positions) multiplied by the notional of the bond (or amount outstanding), i.e. Quantity = Amount ×
Notional.
Consider a long position in a single US Treasury bond that trades at 59 cents on a dollar. The
notional amount is USD 1,000,000. The portfolio’s base currency is EUR and the USD/EUR FX rate
is 1.4131. This bond has a single performance leg and its exposure value in the base currency is equal
to the position market value, from (5.3) this is as follows
Exposure value = 1, 000, 000 × 0.59 × 1/1.4131

= 417, 521.761. (5.4)
Bond futures
Futures are the most basic form of derivative. The futures price is set so that no payment is made when
the contract is written, such that, at initiation, the futures contract has zero market value. Through
marking to market each day the futures contract is effectively rewritten each day at the new futures
price, which is determined by the underlying bond price. Hence the value of the futures contract after
the daily settlement (of payments towards the eventual purchase of the underlying asset) will always be
zero. Purchasing a bond future is equivalent to holding a long bond position and a short cash position.
Bond futures have two legs of exposure, one leg is the underlying bond and the other leg is notional
cash.
For bond futures, Market Value = 0. The first performance leg is the exposure value of the under-
lying asset:
Underlying asset exposure value = Quantity × Quoted bond price × FX rate. (5.5)
Since the bond future’s market value is not equal to the underlying bond’s exposure value an amount of
notional cash is required to offset the difference between the first performance leg and the bond futures
market value. The notional cash amount is the second performance leg:
Notional cash = Position market value − Underlying asset exposure value. (5.6)
Let us illustrate this with a numerical example. Consider a short bond future contract of size 500 on
a notional amount of 200,000. The current quoted bond price is 101.4535 and the security currency and
base currency are both USD. We first calculate the position market value, which is zero after marking
to market. Thus from (5.5) the first performance leg is
Underlying asset exposure value = −500 × 200, 000 × 101.4535/100 × 1

= −101, 453, 500, (5.7)
and finally from (5.6) the second performance leg is
Notional cash = 0 − (−101, 453, 500)

= 101, 453, 500. (5.8)
From this example it is evident that the position market value/partial exposure of zero is misleading as
it does not at all give an intuitive sense of the exposure of the position. This clearly shows that full
exposure is needed to decompose the futures contribution to the portfolio’s total return.
Effect of settlement
If a position has a settlement date which is forward of the trading date, i.e. the position enters the port-
folio prior to settlement, the positions market value is not equal to its partial exposure. This produces
an extra performance leg in the case of basic assets as an amount of notional cash is required to offset
the difference between the position market value and the exposure of the asset.
This is best understood through a simple worked example. If we take the US Treasury bond from
the earlier example and say that the bond will settle in three days, and we also assume that the given
Other derivatives 49
settlement price has a present value of 0.56, then the positions market value is as follows
Position market value = Quantity × (Market price − Settlement price present value) × FX rate
= 1, 000, 000 × (0.59 − 0.56) × 1/1.4131
= 21, 229.92. (5.9)
This is very different to the exposure value in (5.4) of 417,521.7607, the resulting notional cash is as
follows
Notional cash = 21, 229.92 − (417, 521.761) = −396, 291.841. (5.10)
As before the notional cash offsets the difference between the positions market value and the exposure
value of the asset.
5.4 Other derivatives
The performance analysis of increasingly complex derivatives using full exposure is possible as long
as the derivatives can be split into basic (non-derivative) instruments.
For a pay fixed, receive float swap, the first performance leg is the exposure value of being long
the underlying floating rate note, and the second performance leg is the exposure value of being short
the corresponding fixed-coupon bond. For a swap there is no notional cash as the swap market value is
equal to the addition of the two underlying bond exposures, there is no need to offset anything.
Option contracts like any other asset generate economic exposures. The economic exposure of
option contracts is not linear like futures contracts, and will change depending on whether the option
is a call or a put and by how much above or below the underlying price is compared to the exercise
price. To attribute the performance of an option we express the change in value of the option in terms
of the change in value of the underlying asset. To measure the full exposure of an option contract it
is appropriate to scale the underlying asset value by the option delta. The delta of an option is the
rate of change of the option value with respect to a change in value of the underlying asset. The first
performance leg is the delta-adjusted exposure value of the underlying asset. Since the bond option’s
market value will not be equal to the underlying bond’s exposure value an amount of notional cash is
required to offset the difference between the first performance leg and the market value of the bond
option position. The second performance leg is the notional cash.
Finally consider a credit default swap (CDS), here a protection buyer makes a series of premium
payments to the seller in exchange for payment of the notional in the event the underlying bond defaults.
The first performance leg is the exposure value of being long a risky bond. The second performance
leg is the exposure value of being long a fixed payer interest-rate swap. The third performance leg
is the exposure value of a default free loan. This methodology is also applicable to a CDS index
derivative. The approach discussed earlier for options can be used to readily capture the full exposure
of swaptions (options on a swap), caps and floors (which are options on bonds). With all derivatives
the partial exposure vastly understates the true (full) exposure of the position.
Chapter 6
Benchmarks
6.1 Introduction
A benchmark is a portfolio that is used as a reference, it consists of a predetermined set of securities

against which managers can measure the performance of their portfolio. The set of securities could be
a published benchmark or a combination of a set of published benchmarks.
Benchmarks should have the following key characteristics:
• Representative: in terms of span and weight of appropriate markets, instruments and issues that
reflect opportunities available to investors;
• Investible and replicable: by including only securities in which an investor can trade at short
notice and for which market prices exist;
• Accurate and reliable: so that index return calculations reflect the actual changes in the value
of a portfolio consisting of the same securities; daily and timely, so managers can measure
performance immediately and make adjustments to their investment strategy.
A common distinction between benchmarks is by classification of the constituents, for example

a dissection of the universe of securities into Treasury, government related (agency or sovereign),
corporate, and securitized. A more granular distinction can be made by grouping securities by their
style (macro, long-short, relative value, etc), or by their geography (global, Americas, Europe, Asia-
Pacific, Africa).
52 CHAPTER 6. BENCHMARKS
Benchmarks attempt to provide an accurate, comprehensive depiction of the performance and fun-
damental characteristics of a given market. They are also geared towards a manager’s specific invest-
ment strategy. For instance J.P. Morgan have two versions of their Emerging Markets Bond Index
(EMBI), one is the standard EMBI Global and the other is the EMBI Global Diversified. The EMBI
Global uses a traditional market-capitalization approach to determine the weight of each individual
bond issue, whereas the EMBI Global Diversified limits the weights of index countries with larger
debt stocks by only including a specified portion of these countries’ eligible current face amounts of
debt outstanding. The former benchmark is geared towards active managers of large portfolios or any
portfolio that, regardless of size, faces daily fluctuations in its balance of investable funds. The latter
is geared towards managers who face limitations on the amount of portfolio exposure they can take to
individual issuers.
This chapter begins by discussing the rules by which benchmark level and composition is deter-
mined, we then give examples of the data that is necessary for fixed-income performance attribution.
Replication of vendor returns is discussed as this is needed so that validation checks on daily data
received can be performed; only then can we ensure the integrity of the data used in performance
attribution analysis.
6.2 Benchmark rules
The composition and level of a published benchmark is governed by a set of pre-defined rules, these
rules are both vendor and benchmark/index specific and can be summarized as follows:
• Selection criteria – How bonds enter and exit the benchmark, for example, by bond type, issuer,
time to maturity, or outstanding amount.
• Rebalancing – The details of how the weights of securities are adjusted and at what time interval.
• Total return – The return methodology and how cash flows which arise from the constituent
securities are handled in terms of reinvestment.
• Index calculation – How the constituent market prices aggregate to the index price.
Benchmark rules 53
6.2.1 Selection criteria
It is possible that constituents are removed from the benchmark due to securities that become ineligible
for inclusion during the month. This could be due to ratings changes, called bonds, securities falling
below a given time-to-maturity. Also, bonds can enter the benchmark due to bond issues that are newly
eligible because of ratings changes or newly issued bonds. A very important attribute of any security
belonging to a benchmark is that it should be easy to invest in it, as it is essential that the benchmark can
be reproduced by a manager. An example of a very different selection criteria exists for J.P. Morgan’s
EMBI, where a security must belong to a country that has been classified as having a low or middle
per capita income by the World Bank for at least two consecutive years.
6.2.2 Rebalancing
Rebalancing is performed to ensure that the benchmark accurately reflects the available market supply
of investable bonds.
Merrill Lynch and FTSE indices are rebalanced on the last calender day of the month based on
information available in the marketplace. For the Iboxx Top 30 Index, rebalancing is performed every
quarter, on the last calender day of February, May, August and November. Barclays Capital (formerly
Lehman Brothers) benchmark indices belong to either a returns (monthly rebalanced) or statistics uni-
verse (no rebalancing) and constituents eligibility for index inclusion is evaluated on a monthly basis
to determine index composition, where each index consists of two universes of securities:
1. Returns universe – The securities within an index of this type are determined at the beginning
of each month and held fixed until the beginning of the next month, i.e. securities do not move
either into or out of the index. Daily and monthly returns reflect the performance of the returns
universe. This monthly rebalancing means the returns universe holds the market cap of the
constituents constant throughout the month which means that a fund manager avoids having to
hit a moving target.
2. Statistics universe – A dynamic set of bonds that changes daily to reflect the latest composi-
tion of the market. This universe serves as a projection of the next month’s returns universe.
Statistics such as market values, sector weightings, and various averages (e.g. coupon, dura-
tion, maturity, yield, price) are updated and reported daily. At the end of each month, the latest
statistics universe becomes the returns universe for the coming month. The statistics universe
allows a manager to monitor changes in the index throughout a month. Active managers can
modify their portfolios as the index changes, while passive managers can be prepared to execute
all rebalancing transactions at the end of the month to match the upcoming returns universe.
6.2.3 Total return
The standard methodology to calculate total return is based on price changes, income received, and,
where applicable, currency value changes. Beyond this each vendor has its own specific rules regarding
how cash flows affect returns calculations. For all Lehman Brothers indices intra-month cash flows
contribute to monthly returns but the cash flows are not reinvested during the month and do not earn
a reinvestment return. For Merrill Lynch intra-month cash flows are reinvested at the beginning-of-
month 1-month LIBID (London Interbank Bank Bid Rate), with daily compounding. At the end of
each month, cash and reinvestment income are removed from the index and weights are recalculated.
For IBoxx indices, cash from received coupons is reinvested in the money market. The interest rate
used is the one-month LIBID. LIBID is defined as the US-Dollar one month LIBOR less 1/8 (12.5
basis points) as of the coupon payment date.
6.2.4 Index calculation
The calculation of the daily index value requires the following to be known:
1. The list of instruments to be included and their amounts outstanding.
2. The daily total return of each instrument.
3. The weight of each instrument as of the prior business day’s close.
An index’s total return is the weighted average of the total return of all constituents. Typically the
weighting factor is the market value at the start of the period. Where the market value accounts for
the market price of an index-eligible security and its accrued interest. One contrasting case is for the
EMBI Global Diversified benchmark which applies different proportions of each issuer’s current face
amount outstanding from countries with larger debt stocks, so that countries with large current face
amounts outstanding of index-eligible debt will have instrument allocations and, thus, index capital-
ization weights reduced.
Data 55
6.3 Data
Daily published data from vendors such as returns, pricing and bond characteristic data should be used
where available at both index level and constituent level.
Some typical daily index level data needed for performance attribution is given in Table 6.1. Various
statistics such as portfolio duration and the index coupon rate among others are very useful for a
portfolio manager to see. Market values and index returns are also required in other currencies if
supplied by the vendor.
Table 6.1
Daily index level data
Field name Example

1. Market value in base currency. $8384475.78
2. Number of securities in the index. 158
3. Index level in base currency. 6227.12
4. Return frequency, e.g. daily, monthly, quarterly, yearly. DAILY
5. Hedging – stating if the index returns are hedged or not. Y
6. Price return in base currency. 0.570383%
7. Total return in base currency. 0.646883%
8. Weighted average coupon rate of the index. 10.743664
9. Weighted average remaining maturity of the index. 11.269746
10. Weighted average yield to maturity of the index. 2.22%
11. Weighted average yield to worst of the index. 1.82%
12. Weighted average effective duration of the index. 5.27
13. Weighted average modified duration of the index. 5.21
14. Weighted average effective convexity of the index. 0.55
15. Weighted average modified convexity of the index. 0.51
16. Weighted average issuer OAS of the index. 54bps
Some typical daily constituent level data required for performance attribution is given in Table 6.2.
The data needed per security includes security market values so relative weights can be calculated. For
returns, prices and cash flows are needed, which in turn requires information on coupon details and
conventions, and also includes any non-standard features such as non-uniform first coupons. Again,
market values and security returns in other currencies are required if supplied by the vendor.
Table 6.2
Daily constituent level data
Field name Example

1. Market value in base currency. $125700
2. Outstanding or notional in base currency. $83800
3. Clean quoted price. 143.473765
4. Accrued interest. 2.445652
5. Dirty price. 145.919417
6. Payment made in base currency. 4137.625
7. Coupon type, e.g. fixed, floating, zero, step, hybrid. FIXED
8. Annual coupon rate. 9.875
9. Coupon payment frequency per year. 2
10. Day count convention followed e.g. actual/actual. ACT/ACT
11. Maturity date. 20-Sep-2026
12. Price return in base currency. 0.542782727%
13. Total return in base currency. 0.634256839%
14. Yield to maturity. 2.22%
15. Yield to worst. 1.82%
16. Effective duration. 5.23
17. Modified duration. 4.53
18. Effective convexity. 0.55
19. Modified convexity. 0.51
20. Issuer OAS. 54bps
21. Index ratio, for inflation linked bonds. 1.606312
22. Real price of an inflation linked bonds. 99.779999
23. Real accrued interest of an inflation linked bond. 1.24448
Replicating returns 57
6.4 Replicating returns
It is vital to be able to reconcile with published vendor returns. This is necessary to ensure that vendor
data used in performance attribution analysis is sound. Replication of vendor returns is possible only
if each vendor’s returns methodology can be replicated.
The main difficulty in replication of fixed-income benchmark returns is the correct calculation of
coupon amounts and the timing of the payments. For fixed-coupon bonds, once the inception date,
first and last coupon date, maturity date, annual coupon payment and coupon frequency and ex-day
convention for each bond is known the coupon my be determined exactly. For bonds paying floating
coupons and bonds making unscheduled principal payments (see Section 2.1 for details on all payment
types), unless benchmark providers publish their daily security-level rates of returns or at least the
payments made, determination of exact payment amounts and timings for returns calculations becomes
increasingly difficult.
A further issue with benchmark replication is the treatment of cash. For coupons or principal paid,
the vendor may either decide to place the cash received into a holding account or to reinvest the cash
in the benchmark’s securities.
Example
To illustrate some of the subtleties involved in replicating benchmark returns we examine the 1-day
local return of a bond for a sample of different vendors. The example bond is the Republic of Austria
government bond which matures on 01/15/2018, the full terms and conditions are given in Table 6.3.
The characteristics of interest needed for replicating returns are the coupon details; the bond pays an
annual coupon of 4.65% on January 15.
In Table 6.4 actual pricing data and 1-day returns are given for the 3 sample vendors. The vendors
under examination are as follows: Citigroup (shown as Citi), iBoxx Markit (shown as iBoxx), and J.P.
Morgan (shown as JPM). The published 1-day local total return (see the column headed Published)
for 01/15/2009 is compared to the 1-day local total return rtotal , calculated according to Eq. (2.1) using
the clean price P(t), accrued interest AI(t), and coupon payment amount Ctc (reported by each vendor)
for 2 consecutive days, the residual is the difference between the published vendor return and the
calculated return.
There are some very interesting points to notice about the results.
Table 6.3
Terms and conditions of Republic of Austria Government Bond, 4.65% 15-Jan-2018
ISIN AT0000385745
Currency EUR
Issue Date 22-Jan-03
Maturity Date 15-Jan-18
First Coupon Date 15-Jan-04
Coupon type Fixed Annual
Coupon 4.65%
Day count Act/Act
Outstanding 10321379000
Par 1000
Table 6.4
Vendor pricing data and 1-day returns for Republic of Austria Government Bond
Date P(t) AI(t) Ctc Published rtotal Residual

Citi
14-Jan-2009 107.15362 4.6373 -
15-Jan-2009 107.40437 0 4.65 0.235675 0.235663 0.000012
iBoxx
14-Jan-2009 107.11179 4.6372951 -
15-Jan-2009 107.40461 0 4.65 0.273403 0.273403 0
JPM
14-Jan-2009 107.3843 0.051 -
15-Jan-2009 107.6324 0.0637 - 0.242808 0.242751 0.000057
All values are in percent.
Validation of benchmark data 59
(i) The coupon was deemed to be paid on January 15 for Citi and iBoxx, evident from the accrued
interest being zero for these vendors on this day but in the case of JPM, the coupon was deemed
to have already been paid by January 15. This leads to a difference in returns between vendors.
(ii) For the same bond the clean price reported by each vendor is different on corresponding days,
and thus, so is the 1-day return, which is consistent with Section 4.6, where the impact of pricing
differences on returns due to vendors using different data sources is discussed.
(iii) The published vendor return is replicated by rtotal to an extremely high degree of accuracy, in the
case of iBoxx, the residual is actually zero at this level of precision.
6.5 Validation of benchmark data
Daily validation of vendor benchmark data is necessary to ensure the integrity of the data used in all
performance attribution analysis. This involves checking data received from a vendor against calculated
results and setting allowed tolerances for the difference between the two. If checks exceed specified
tolerances then the data feed from the vendor needs to be examined closely for errors. Some examples
of data checks are: constituent total return should equal the published constituent return; aggregated
constituent returns should equal published index total return; calculated statistics like yield-to-maturity,
effective duration, modified duration, etc, should equal published statistics.
Bibliography
Ankrim, E. and Hensel, C. (1994). Multicurrency performance attribution, Financial Analysts Journal
50(2): 29–35.
Bacon, C. (2008). Practical Portfolio Performance Measurement and Attribution, John Wiley & Sons.
Black, F., Derman, E. and Toy, W. (1990). A one-factor model of interest rates and its applications to
treasury bond options, Financial Analysts Journal pp. 33–39.
Brinson, G. and Fachler, N. (1985). Measuring non-U.S. equity portfolio performance, The Journal of
Portfolio Management 11(3): 73–76.
Carino, D. (1999). Combining attribution effects over time, Journal of Performance Measurement
Summer: 5–14.
Couderc, F. (2008). Inflation across the board, RiskMetrics Journal 8(1): 61–87.
Dor, A., Dynkin, L., Hyman, J., Houweling, P., Van Leeuwen, E. and Penninga, O. (2007). DTS
(Duration Times Spread), Journal of Portfolio Management 33(2): 77–100.
Fama, E. and MacBeth, J. (1973). Risk, return, and equilibrium: Empirical tests, Journal of Political
Economy 81(3).
Ho, T. S. Y. (1992). Key Rate Durations, The Journal of Fixed Income 2(2): 29–44.
Litterman, R. and Scheinkman, J. (1991). Common factors affecting bond returns, The Journal of
Fixed Income 1(1): 54–61.
Malz, A. (2002). Estimation of zero-coupon curves in datametrics, RiskMetrics Journal 1.
Menchero, J. (2004). Multiperiod arithmetic attribution, Financial Analysts Journal 60: 76–91.
Shi, D. (2009). Issuer specific curves—new construction model, Research technical note, RiskMetrics
Group.
61
Soor, R. (2010). Performance manager attribution methodology guide – phase 1, Research technical
note, RiskMetrics Group.
Soor, R. and Costigan, B. (2010). Performance manager return calculation methodology guide, Re-
search technical note, RiskMetrics Group.
Sorensen, L. (2010a). Computing carry for performance attribution, Research technical note, Risk-
Metrics Group.
Sorensen, L. (2010b). Key rate durations, Research technical note, RiskMetrics Group.
Vann, P. (1999). An analysis of a source of errors in performance measurement, MSCI Barra research
insights, MSCI Barra.
View publication stats

Fixed Income Performance Attribution

Uploaded by

Copyright:

Available Formats

Fixed Income Performance Attribution

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fixed Income Performance Attribution

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Fixed Income Performance Attribution

Article in SSRN Electronic Journal · May 2010

Stéphane Daul Nicholas Sharp

The user has requested enhancement of the downloaded file.

Lars Qvigstad Sørensen

3.1 Factor duration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3.2.1 Forward premium and currency surprise . . . . . . . . . . . . . . . . . . . . . 13

3.4 Yield curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.4.1 Effective duration and convexity . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.4.2 Key-rate durations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4 Relative Performance Attribution 33

4.2 Pure asset-grouping approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.3 Pure factor model approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.4 Hybrid attribution model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.5 Multi-period attribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.6 Impact of pricing difference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

5.1 Partial and full exposure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

5.2 Performance legs and notional cash . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5.3 Portfolio weight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5.4 Other derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6.2 Benchmark rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

6.2.1 Selection criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

6.2.3 Total return . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

6.2.4 Index calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

6.4 Replicating returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

6.5 Validation of benchmark data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

(I) (II) (III)

Reinvestment of coupon payments

Payments received are typically reinvested at the reinvestment rate re (tc ) as

C(t1 ,t2 ) = ∑ Ctc · [1 + re (tc )](t2 −tc ) , (2.2)

V (t + 1) −V (t) +C(t,t + 1) − P(t) + S(t)

3.1 Factor duration

rfactor k = −Dk × fk , (3.6)

where fk = ∆xk is the ex-post realized return on factor k.

Vbase (t) = V (t)χ (t). (3.8)

[V (t2 ) +Ctc ] χ (t2 )

rcurrency = rbase − rlocal

3.2.1 Forward premium and currency surprise

[χ (t2 ) − Fχ (t1 ,t2 )] + [Fχ (t1 ,t2 ) − χ (t1 )]

χ (t2 ) − Fχ (t1 ,t2 )

Fχ (t1 ,t2 ) − χ (t1 )

rFX = (1.6 − 1.5)/1.5 = 6.67%, (3.17)

the currency surprise, according to (3.15), is

rsurprise = (1.6 − 1.515)/1.5 = 5.67%, (3.18)

and the forward premium, according to (3.16), is

rforward = (1.515 − 1.5)/1.5 = 1%, (3.19)

rcarry = rcoupon + rroll . (3.22)

rcarry = 0.522% − 0.156% = 0.366%. (3.25)

3.4 Yield curve

3.4.1 Effective duration and convexity

∆yk = y(t2 , τk ) − y(t1 , τk ). (3.32)

3.4.2 Key-rate durations

(a) Yield curves (b) Change in yields

which have been found representative of the entire yield curve.

ryield curve =rshift + rshape

Time to Maturity [Yrs] 0.5 2 5 10 20 30

difference. The spread convexity is computed as

∆s = ∆sswap + ∆ssector + ∆sidiosyncratic , (3.39)

Statistic Coca Cola 7% 1-Oct-2026