Bond Portfolio Management Using the Dynamic

Nelson-Siegel Model
João F. Caldeira1 , Guilherme V. Moura∗2 , and André A. P. Santos2

1 Department of Economics, Universidade Federal do Rio Grande do Sul, Porto Alegre/RS - Brazil.
2 Department of Economics, Universidade Federal de Santa Catarina, Florianópolis/SC - Brazil.

February 11, 2014

Abstract

Factor models for the yield curve, such as the dynamic version of the Nelson-Siegel model
proposed by Diebold and Li (2006) and its arbitrage-free counterpart proposed by Christensen,
Diebold and Rudebusch (2011), have been extensively applied to forecast bond yields. In this
paper, we propose a novel utilization of these models in bond portfolio management. More
specifically, we derive closed form expressions for the vector of expected bond returns and
for its covariance matrix based on a general class of dynamic factor models, and use these
estimates to obtain optimal mean-variance and duration-constrained mean-variance bond
portfolios. An empirical application involving a large data set of US Treasuries shows that
the proposed portfolio policy outperform a broad set of traditional yield curve strategies used
in bond desks. Moreover, we find that the investor with a quadratic utility function is willing
to pay a performance fee to adopt the proposed mean-variance bond portfolios.
Keywords: Bond indexing, Kalman filter, maximum likelihood estimation, out-of-sample eval-
uation, portfolio optimization

∗
Corresponding author. Department of Economics, Federal University of Santa Catarina, 88049970 Florianópo-
lis/SC - Brazil. Tel. +55 48 3721-6652; fax +55 48 3721-9585. E-mail: [email protected]

1
1 Introduction
Dynamic factor models for the term structure of interest rates have been extensively and

successfully applied in forecasting bond yields. The dynamic version of the Nelson and Siegel

(1987) model proposed by Diebold and Li (2006) and its arbitrage-free version proposed by

Christensen et al. (2011) have gained popularity among financial market practitioners and

central banks (BIS, 2005; Gürkaynak et al., 2007). Existing evidence suggests that these

specifications are remarkably well suited both to fit the term structure, as well as to forecast

its movements (see, for example, Diebold and Rudebusch, 2011, and the references therein).

This evidence makes one tempted to investigate the extent to which these specifications can

be applied to other problems. It would be interesting, for instance, to use this approach to

support a portfolio policy that exploits the risk-return tradeoff in bond returns. In this sense,

the main question of this paper is the following: can we go beyond forecasting and use these

specifications to perform bond portfolio selection in a mean-variance context? Even though

the mean-variance criterion is well established in building quantitative portfolio strategies, a

number of questions need to be addressed in order to motivate its application in fixed income.

Is this approach relevant to the fixed income world? A satisfactory answer to these questions

depends on an adequate examination of the comparative performance of the resulting mean-

variance bond portfolio vis-a-vis traditional portfolio strategies employed in bond desks. Is

the comparative performance analysis favorable to the proposed portfolio policy? Moreover,

is the investor willing to pay a performance fee in order to switch from a traditional bond

portfolio policy to the mean-variance policy? These are fundamental questions that must

be answered in order to validate the use of the mean-variance paradigm in the fixed income

context.

We address each of these questions by providing a realistic empirical applications involving

a data set based on a large panel of monthly time series of U.S. zero-coupon bonds with ma-

turities up to 10 years over the period from January 1970 to December 2009. We initially im-

plement the dynamic version of the Nelson and Siegel (1987) model proposed by Diebold and

Li (2006) (hereafter DNS) and its arbitrage-free version proposed by Christensen et al. (2011)

2
(hereafter AFDNS) specifications along with efficient procedures for estimation of factors and

parameters as proposed by Jungbacker and Koopman (2008) in order to generate forecasts of

bond returns based on an expanding window. These are then used to solve alternative versions

of the mean-variance optimization problem. First, we solve the shortselling-constrained mean-

variance problem for alternative values for the risk aversion coefficient. Second, we extend

the standard mean-variance formulation and propose the duration-constrained mean-variance

problem, thus obtaining an optimal mean-variance portfolio that matches the duration of a

given benchmark. The results are benchmarked against alternative bond portfolio strategies

widely employed in bond desks, such as spread, barbell, bullet, and ladder strategies.

The empirical evidence shows that the proposed mean-variance bond portfolios seems to

be very reasonable alternatives to the portfolios built upon traditional yield curve strategies.

Our results show that the shortselling-constrained mean-variance portfolios of US bonds with

monthly re-balancing achieved an annualized average gross return ranging from 5.9% to 11.6%

and an annualized standard deviation ranging from 1.3% to 12.1%. The risk-adjusted per-

formance measured by the annualized Sharpe ratio ranges from 0.35 to 0.67. Investors with

higher (lower) risk-aversion tend to invest in portfolios with lower (higher) duration. We also

employ a test for the Sharpe ratio based on the bootstrap procedure of Politis and Romano

(1994), which allows us to formally compare portfolio policies in terms of their risk-adjusted

returns. The results reveal that in most cases the mean-variance portfolios outperform all

benchmark policies in terms of Sharpe ratio. Our results also indicate that an investor with

quadratic utility is willing to pay a fee to adopt the mean-variance portfolio policy in all cases,

and that this performance fee increases with the investor’s risk aversion coefficient. On aver-

age, the investor with risk aversion coefficient equal to 1 is willing to pay a performance fee of

2 to 15 basis points per year to switch from traditional bond portfolio strategies to a mean-

variance portfolio strategy. We also assess the impacts of transaction costs by implementing

the mean-variance portfolios under quarterly re-balancing frequency. We find that lowering

the portfolio re-balancing frequency leads to a substantial decrease in portfolio turnover, al-

though accompanied by decreases in the risk-adjusted performance measured by the Sharpe

ratio. Nevertheless, for some values of the risk aversion coefficient the mean-variance portfo-

3
lios under quarterly re-balancing still outperform all benchmark strategies considered in the

paper.

We also move forward the literature by considering a novel duration-constrained mean-

variance optimization. In this formulation, we solve the shortselling-constrained mean-variance

problem by adding a target duration constraint for the bond portfolio. This approach is par-

ticularly appealing, since in practice fixed-income portfolios tend to be selected in order to

approximate the duration of a benchmark or to replicate the performance of this benchmark

in terms of return and volatility. Our results reveal that the risk-adjusted performance of the

duration-constrained mean-variance bond portfolios are superior to that of the benchmark

strategies with equal duration levels. Moreover, the duration-constrained mean-variance pol-

icy yields portfolios with lower risk (standard deviation) in comparison to other benchmark

portfolios with similar duration levels.

Our approach to the bond portfolio allocation problem is based on the the mean-variance

framework introduced by Markowitz (1952), which is one of the milestones of modern finance

theory. In this framework, individuals choose their allocations in risky assets based on the

trade-off between expected return and risk. A common criticism of this approach is that

it is a myopic single-period portfolio policy, with no concerns involving longer, multi-period

investment horizons. However, the the errors coming from solving a complicated dynamic

optimization problem can outweigh the expected utility gain from investing optimally as

opposed to myopically; see Brandt (2009) for a discussion. Consequently, the application of

the mean-variance approach is still widespread among practitioners and academics.

In order to implement the mean-variance optimization in practice, it is common to ob-

tain estimates of the vector of expected returns and its covariance matrix and plug these

estimators in an analytical or a numerical solution to the mean-variance problem. We show

that factor models for the yield curve can substantially simplify the process of mean-variance

bond portfolio selection, since they allow the computation of expected bond returns and their

covariance matrix in closed form. Specifically, we show how to obtain these expressions based

on a general class of dynamic affine factor models that includes the DNS and AFDNS models

as special cases, and use them to obtain optimal bond portfolio allocations based on an active

4
portfolio policy grounded on the mean-variance framework. Two appealing features of our

approach are that i) all ingredients necessary to calculating the closed form estimators are

easily retrieved from the one-step Kalman filter estimation of the DNS and AFDNS models,

and ii) since it is based on factor specifications, our approach is reasonably parsimonious and

suitable for high-dimensional applications in which a large number of fixed income securities

is involved.

Existing literature shows few references suggesting the use of mean-variance approach

to bond portfolio selection. Korn and Koziol (2006) pioneered this literature by employing

the Vasicek (1977) model to perform mean-variance bond portfolio selection. Our approach,

on the other hand, allows the use of a broad family of affine term structure models to per-

form bond portfolio selection. In particular, our approach nests the one considered in Korn

and Koziol (2006), which is based on the Vasicek (1977) model and extends to other affine

specifications with enhanced forecasting power, such as the DNS and AFDNS models.

At least two reasons can be pointed out in order to justify the lack of mean-variance

optimization applied to fixed-income portfolios. The first reason is the relative stability

and low historical volatility of this asset class, which discouraged the use of sophisticated

methods to exploit the risk-return trade-off in fixed-income assets. However, this situation

has been changing rapidly in recent years, even in markets where these assets have low default

probability (Korn and Koziol, 2006). The recurrence of turbulent episodes in global markets

usually brings high volatility to bond prices, which increases the importance of adopting

portfolio selection approaches that take into account the risk-return trade-off in bond returns.

The second reason refers to the difficulties in modeling bond returns and the covariance

matrix of bond returns (Korn and Koziol, 2006; Puhle, 2008). Fabozzi and Fong (1994) argue

that if a covariance matrix of bond returns is available, the process of portfolio optimization

using fixed-income securities is similar to that of equity portfolios. However, one should

bear in mind that fixed-income securities have finite maturities and promise to pay face

value at maturity. In this sense, the end of the year price of a bond with two years to

maturity is indeed a random variable. However, the price of that same bond in two years is a

deterministic quantity (disregarding the risk of default) given by its face value. This implies

5
that the statistical properties of price and return of a fixed-income security depends on their

maturity. Thus, both bond price and bond return are non-ergodic processes, and therefore

traditional statistical techniques cannot be used to directly model the expected return and

volatility of these assets; see Meucci (2009) for a discussion. One possibility to circumvent

this difficulty is to employ factor models for the term structure of interest rates. As argued by

Korn and Koziol (2006), the great advantage of using this approach is the possibility to model

constant (or fixed) maturity yields. We show in the paper that this allows the estimation of

the conditional distribution of bond yields without relying on bond maturities.

The paper is organized as follows. Section 2 describes the factor models used for modeling

the term structure and shows how to convert yield forecasts into bond return forecasts.

Section 3 discusses the estimation procedure for expected bond returns and for the conditional

covariance matrix of bond returns. Section 4 discusses the empirical applications to both

portfolio optimization. Finally, Section 5 concludes.

2 Bond returns and dynamic factor models

In this section we show how a general class of dynamic factor models for the yield curve can

be used to compute the joint distribution of the log-return of N bonds. Factor models for

the term structure of interest rates allow us to obtain closed form expressions for the vector

of expected yields and its covariance matrix. From these moments, we show how to compute

the distribution of bond prices and bond returns, which are key ingredients to bond portfolio

optimization according to the mean-variance approach proposed by Markowitz (1952).

2.1 Dynamic factor models for the yield curve

Dynamic factor models play a major role in econometrics, allowing the explanation of

a large set of time series in terms of a small number of unobserved common factors (see,

among others, Fama and French, 1993; Stock and Watson, 2002; Jungbacker et al., 2013).

Many specifications for the yield curve can be viewed as dynamic factor models with a set of

restrictions imposed on the factor loadings.

6
 We consider a set of time series of bond yields with N different maturities τ = [τ1 , . . . , τN ]0 .

The yield at time t of a security with maturity τi is denoted by yi,t for t = 1, . . . , T and

i = 1, . . . , N . The N × 1 vector of all yields at time t is given by

yt = [y1,t , y2,t , . . . , yN,t ]0 , t = 1, . . . , T.

The general specification of the dynamic factor model is given by

yt = Γ + Λft + εt , εt ∼ N ID (0, Σ) , t = 1, . . . , T, (1)

where Γ is a N × 1 vector of intercepts, Λ is a N × K matrix of factor loadings, ft is a

K−dimensional stochastic process, εt is the N × 1 vector of disturbances with covariance

matrix given by Σ. As usual in the yield curve literature, we restrict the covariance matrix Σ

to be diagonal (see, for example, Diebold and Li, 2006; Diebold et al., 2006). The dynamic

factors ft are modeled by the following stochastic process:

ft = µ + Υft−1 + ηt , ηt ∼ N ID (0, Ω) , t = 1, . . . , T, (2)

where µ is a K × 1 vector of constants, Υ is the K × K state-transition matrix, and Ω is the

covariance matrix of the disturbance vector ηt , which is independent of the vector of residuals

εt ∀t.

Equations (1) and (2) characterize a linear and Gaussian state space model, thus the

Kalman filter can be used to obtain estimates of the unobserved factors, as well as to construct

the log-likelihood function. The specification for ft is fairly general, however, in modeling

yield curves the usual specifications for ft are either a first-order autoregressive process,

thus yielding a independent-factor model, or a first order vector autoregressive process, thus

yielding a correlated-factor model (see, among others, Diebold et al., 2006).

The yield curve specifications considered in this paper are the main variants of the original

formulation of the Nelson and Siegel (1987) factor model, namely the dynamic Nelson-Siegel

model proposed by Diebold and Li (2006), and its arbitrage-free version, proposed by Chris-

7
tensen et al. (2011). The alternative specifications considered can be captured in the general

dynamic factor model formulation in (1) and (2) with different restrictions imposed on the

vector of intercepts Γ, on the loading matrix Λ and on the state-transition equation (2). More

specifically, denoting the ith line of Λ by Λ(τi ), both the dynamic Nelson-Siegel model and

its arbitrage-free counterpart imply that

1 − e−λτi 1 − e−λτi
    
λτi
Λ(τi ) = 1, , −e , (3)
λτi λτi

where the decaying factor λ is assumed to be constant over time. This is in accordance with

Diebold and Li (2006) and the common finding that time variations in λ have only a negligible

impact on the model’s fit and prediction power (see, for example, Haustsch and Ou, 2012).

The restrictions in (3) allows us to interpret the three factors as the level of the yield curve,

its slope and curvature (Diebold and Li, 2006).

However, the models differ with respect to restrictions imposed on Γ, Υ, and Ω. While

the dynamic Nelson-Siegel model of Diebold et al. (2006) sets every element of the vector Γ

to zero and does not impose any additional restrictions on Υ or Ω, Christensen et al. (2011)

have shown that it is necessary to restrict Γ, Υ, and Ω to ensure that the DNS model is

arbitrage-free. We detail in Appendix 1 the derivation of the restrictions in Γ, Υ, and Ω

required to ensure an arbitrage-free DNS model.

2.2 The distribution of log-returns

Bond portfolio management requires estimates of the expected return of each bond, as

well as estimates of their covariance matrix. However, factor models for the term structure

of interest rates are designed to model only bond yields. Nevertheless, it is possible to obtain

expressions for the expected bond return and for the conditional covariance matrix of bond

returns based on the distribution of the expected yields. We provide these expression below.

Given the system of equations in (1) and (2), the distribution of expected yields yt|t−1 is

8
N (µyt|t−1 , Σyt|t−1 ), where

µyt|t−1 = Λft|t−1 , (4)

and

Σyt|t−1 = Λ Ωt|t−1 + ΥQt−1 Υ0 Λ0 + Σt|t−1 ,

(5)

where ft|t−1 = Et−1 [ft ], and Qt−1 = Vart−1 [ft−1 ] take into account the uncertainty in the

estimates of the unobserved factors (see Harvey et al., 1992).

(τ )
Taking into account that the price of a bond at time t, Pt , is the present value at time
n o
(τ ) (τ )
t of $1 receivable τ periods ahead, we compute the bond price as Pt = exp −τ · yt . To
(τ )
compute the unexpected return, rt , of holding that bond from t − 1 to t while its maturity

decreases from τ to τ − 1, we use the log-return expression,

(τ −1)
!
(τ ) Pt (τ −1) (τ ) (τ ) (τ −1)
rt = log (τ )
= log Pt − log Pt−1 = τ · yt−1 − (τ − 1) · yt . (6)
Pt−1

(τ −1)
Letting yt|t−1 denote the one-step-ahead forecast of a continuously compounded zero-

coupon nominal yield-to-maturity, together with equations (4)-(6), it is easy to see that the
 
(τ )
distribution of the vector of expected log-returns, rt|t−1 , is N µr(τ ) , Σr(τ ) , where
t|t−1 t|t−1

(τ )
µr(τ ) = −(τ − 1) ⊗ µy(τ −1) + τ ⊗ yt−1 , (7)
t|t−1 t|t−1

 
 
0 0

0 
Σrt|t−1 = (τ − 1)(τ − 1) ⊗ Λ Ωt|t−1 + ΥQt−1 Υ Λ + Σt|t−1 
 , (8)
| {z }
Σ (τ −1)
y
t|t−1

and ⊗ is the Hadamard (elementwise) multiplication.

The results (7) and (8) show that it is possible to obtain closed form expressions for the

expected log-returns of bonds and their covariance matrix based on dynamic factor models

for the yield curve. These estimates are key ingredients to the problem of bond portfolio

9
management, as discussed in section 4. It is worth noting that all ingredients necessary to

calculating the closed form estimators are easily retrieved from the Kalman filter estimation

discussed in Section 3. In particular, the values of ft|t−1 and Qt−1 in (4) and (5), respectively,

are direct products of the Kalman filter recursions and are readily available. Furthermore,

the closed form estimators are based on a general formulation for a dynamic factor model for

the yield curve, which implies that it is also applicable to other affine term structure models.

3 Estimation procedure
Given the state space formulation of the dynamic factor model presented in (1) and (2),

the Kalman filter can be used to obtain the likelihood function via the prediction error decom-

position, as well as filtered estimates of the states and of their covariance matrices. However,

the computational burden associate with the Kalman filter recursions depends crucially on

the dimension of both the state and observation vectors. Moreover, in yield curve models

the dimension of the observation vector (N × 1) is often much larger than that of the state

vector (K × 1). In these circumstances, Jungbacker and Koopman (2008) have shown that

significant computational gains can be achieved by a simple transformation. First, define the

N × N and the K × N matrices:

 
AL
A= , AL = CΛ(λ)0 Σ−1 ,
 
AH

respectively, where C can be any K × K invertible matrix, and AH is chosen to guarantee

−1
that A is full rank. Selecting C = Λ(λ)0 Σ−1 Λ(λ) implies:

          
AL yt   ft   t  AL ε t  AL ε
  C 0 
Ayt =  = + ,   ∼ N 0,   . (9)

AH yt 0 A H εt AH εt 0 ΣH

The law of motion of the factors in (2) is not affected by the transformation. Note that

AH yt is neither dependent on ft , nor correlated with AL yt , and therefore does not need to be

10
considered for the estimation of the factors. This implies that the Kalman filter only needs

to be applied to the low dimensional subvector AL yt for signal extraction, generating large

computational gains when N >> K (see Table 1 of Jungbacker and Koopman, 2008).

Denote l(y) the log-likelihood function of the untransformed model in (1) and (2), where

y = (y10 , . . . , yT0 )0 . Evaluation of l(y) can also take advantage of the transformations presented

above. Jungbacker and Koopman (2008) show that the log-likelihood of the untransformed

model can be represented as

T
T |Σ| 1 X 0 −1
l(y) = c + l(y L ) − log − et Σ et , (10)
2 |C| 2
t=1

where c is a constant independent of both y and the parameters, y L = (AL y10 , . . . , AL yT0 )0 ,

l y L is the log-likelihood function of the reduced system, and et = yt − Λ(λ)ft . Note that

computation of matrix AH is not required at any point, as proved in Lemma 2 of Jungbacker

and Koopman (2008).

4 Application to bond portfolio management

In this Section we describe the empirical application carried out in the paper. We describe

the data set, the mean-variance approach to bond portfolio optimization, the benchmark

strategies, and the methodology to assess portfolio characteristics such as performance and

composition. Finally, we discuss the results.

4.1 Data

The data set consists of end-of-month continuously compounded yields on U.S. zero-coupon

bonds. This panel of monthly time series of yields was constructed from the CRSP un-

smoothed Fama and Bliss (1987) forward rates by Jungbacker et al. (2013) and is publicly

available on the Journal of Applied Econometrics Data Archive, as part of their supplementary

material. The data set consists of 17 maturities over the period from January 1970 to Decem-

ber 2009. The maturities analyzed are τ = 3, 6, 9, 12, 15, 18, 21, 24, 30, 36, 48, 60, 72, 84, 96, 108

11
and 120 months. This choice provides us with a panel of 480 monthly observations on 16

different maturities. The 3-month rate is not included in the optimization process and is

taken as the risk free rate.

Table 1 provides descriptive statistics for this data set. For each maturity, we report mean,

standard deviation, minimum, maximum and one-month, one-year and two-year sample au-

tocorrelation and two-month partial autocorrelation. The summary statistics confirm some

stylized facts common to yield curve data: the sample average curve is upward sloping and

concave, volatility is decreasing with maturity, and autocorrelations are very high. The esti-

mate of the partial autocorrelation function suggest that autoregressive processes of limited

lag order will fit the data well since only the first coefficient is significant for most maturities,

while the second lag coefficients are relatively small. Figure 1 depicts the data set and the

evolution of the yield in the period analyzed.

[Table 1 about here.]

[Figure 1 about here.]

4.2 Bond portfolio optimization according to the mean-variance

framework

To illustrate the applicability of the proposed estimators of expected bond returns and condi-

tional covariance matrix of bond returns defined in Section 2.2, we consider the mean-variance

optimization problem of fixed-income portfolios. The formulation of mean-variance portfolio

is given by
min wt Σrt|t−1 wt − 1δ wt0 µrt|t−1
wt

subject to
(11)
wt0 ι = 1

wt ≥ 0,

where µrt|t−1 is defined in (7), Σrt|t−1 is given in (8), wt is the vector of optimal weights,

ι is a vector of ones with dimension N × 1, and δ is the risk aversion coefficient. In our

12
empirical implementation, we solve the mean-variance optimization problem considering four

alternative values for the risk aversion coefficient δ, in particular, we evaluate (11) for δ =

{1×10−4 , 1×10−3 , 1×10−2 , 1×10−1 , 0.5, 1.0}. Finally, we consider the case in which shortsales

are restricted by adding to (11) a constraint to avoid negative weights, i.e. wt ≥ 0. Previous

works show that adding such a restriction can substantially improve performance, especially

reducing the turnover of the portfolio, see Jagannathan and Ma (2003), among others.

Duration-constrained mean-variance optimization

It is also worth noting that fixed-income portfolio managers commonly employ a strategy

known as bond indexing, which consisting in building a portfolio that replicates risk factors

of a given benchmark index or portfolio (Fabozzi and Fong, 1994). One of the most common

risk factors considered in this strategy is the duration, which is a standard measure of the bond

portfolio sensitivity to changes in yields. The bond indexing strategy can be incorporated in

the mean-variance context by means of a duration-constrained mean-variance optimization

problem. In this setting, we solve the mean-variance optimization problem in (11) with an

additional constraint on the duration of the optimal portfolio, thus obtaining an optimal

portfolio that matches the duration of a given benchmark. The restriction on the duration of

the optimal portfolio is defined as τp = w0 τ , where τp is the portfolio duration and τ is the

vector of individual bond durations. The choice of the target portfolio duration τp can be made

in order to match the duration of a given benchmark or strategy (e.g. a fixed income index,

other fixed income portfolio etc). In our empirical application we obtain duration-constrained

mean-variance bond portfolios with a target portfolio duration of τp = {1, 3, 5, 7, 9} years for

an investor with a risk aversion coefficient δ = 1.

4.3 Methodology for evaluating portfolio performance and im-

plementation details

The performance of optimal mean-variance portfolios is evaluated in terms of average return

(Rp ), average excess return relative to the risk-free rate (Rp,ex ), standard deviation (volatility)

13
of returns (b
σ ), Sharpe Ratio (SR), and turnover. These statistics are calculated as follows:

T −1
1 X 0
Rp = wt Rt+1
T −1
t=1
T −1
1 X
Rp,ex = (w0 Rt+1 − Rft+1 )
T − 1 t=τ t
v
u 1 TX
u −1
σ̂ = t (wt0 Rt+1 − µ̂)2
T −1
t=1
Rp,ex
SR =
σ̂
T −1 N
1 XX
Turnover = (|wj,t+1 − wj,t |),
T −1
t=1 j=1

where wj,t is the weight of the asset j in the portfolio in period t before the re-balancing,

while wj,t+1 is the desired weight of the asset j in period t + 1. As pointed out by DeMiguel

et al. (2009), the turnover as defined above, can be interpreted as the average fraction of

wealth traded in each period. Rf denotes the risk-free rate. We consider the risk free rate to

be the 3-month Treasury.

In order to evaluate the composition of the optimal bond portfolios, we also report the

average duration (in years) of the mean-variance portfolios. The average bond portfolio

duration is calculated as

T −1
1 X 0
Average portfolio duration = wt τ,
T −1
t=1

where τ is the vector of individual bond durations.

We also follow Fleming et al. (2001) and Fleming and Kirby (2003) and use a utility-based

approach to measure the value of the performance gains associated with employing a given

bond portfolio strategy. We assume the investor has a quadratic utility function given by

 
γ 2
U (Rp,t ) = W0 1 + Rp,t − (1 + Rp,t ) ,
2 (1 + γ)

where Rp,t = wt0 Rt+1 is the portfolio return, γ is the investor’s relative risk aversion and W0

14
 is the initial wealth. In order to compare two alternative bond portfolio strategies (Rp1 and

Rp2 ), we determine the maximum performance fee a risk-averse investor would be willing to

pay to switch from using one portfolio policy to another. That is, we determine the value of

∆ such that
T
X −1 T
X −1
U (Rp1,t ) = U (Rp1,t − ∆).
t=1 t=1

This constant represents the maximum return the investor would be willing to sacrifice each

month in order to capture the performance gains associated with switching to the second

portfolio policy. We report the value of ∆ as an annualized basis point fee for each value of

the risk aversion coefficient.

In order to implement the models, we employ a recursive estimation strategy based on

an expanding estimation window. Departing from the first t observations, all models are

estimated and their corresponding one-step-ahead estimate of the vector of expected bond

returns and its covariance matrix are computed using the results in (7) and (8) for each of

the specifications considered in the paper. Next, we use these expressions to obtain optimal

mean-variance portfolios. Finally, we add one observation to the sample, and re-estimate

all models to obtain another one-step-ahead estimate of the vector of expected bond returns

and its conditional covariance matrix. This process is repeated until the end of the data set

is reached. Therefore, we end up with a sample of T − t out-of-sample observations, where

T is the length of the data set. We use as initial estimation window a sample of t = 120

monthly observations, which yields 360 out-of-sample monthly observations. We adopt a

similar procedure to implement the benchmark portfolio strategies discussed in Section 4.4

4.4 Description of the benchmark strategies

To assess the relative performance of the proposed bond portfolio policy based on the mean-

variance criterion, we consider a spectrum of popular yield curve strategies employed in

the majority of bond desks worldwide.1 These strategies involve positioning a portfolio to
1
Jones (1991), Fabozzi and Fong (1994) and Fabozzi (2002) provide a detailed explanation of traditional yield
curve strategies.

15
capitalize on expected changes in the shape of the Treasury yield curve. We consider four

alternative yield curve strategies: spread (or slope) strategy, bullet strategy, barbell strategy,

and ladder strategy. In the spread strategy, we assume the investor shortsells the 1-year bond

and buys the 10-year bond. In the bullet strategy, the portfolio is constructed so that the

maturity of the bonds in the portfolio are highly concentrated at one point on the yield curve.

We implement the bullet strategy by considering six alternative portfolios invested in the 1-

year, 3-year, 5-year, 7-year, 9-year, and 10-year bond. In the barbell strategy, the maturity of

the bonds included in the portfolio is concentrated at two extreme maturities. We implement

the barbell strategy by considering a portfolio equally-weighted in the 1-year and 10-year

bonds. Finally, in a ladder strategy the portfolio is constructed to have approximately equal

amounts of each maturity. We implement this strategy by considering a portfolio equally-

weighted in the 16 maturities of the data set described in Section 4.1. It is worth noting that

each of these strategies will result in different performance when the yield curve shifts. The

actual performance will depend on both the type of shift and the magnitude of the shift.

Table 2 reports the performance of the benchmark strategies. The statistics of returns,

standard deviation and Sharpe ratio are annualized. We observe that the highest average

gross and excess returns are achieved the bullet portfolio invested in the 9-year bond (9.5%

and 3.7%, respectively). The strategy with lower portfolio risk in terms of standard deviation

is 1-year bond portfolio (2%). In terms of risk-adjusted returns, we find that the 3-year bond

portfolio outperforms all other strategies, since it achieves the highest Sharpe ratio (0.447).

[Table 2 about here.]

In order to assess the relative performance of the approach proposed in the paper, we con-

sider as main benchmark strategy the one with highest Sharpe ratio, which is the 3-year bond

portfolio. The stationary bootstrap of Politis and Romano (1994) with B=1.000 resamples

and block size b = 5 was used to test the statistical significance of differences between Sharpe

ratios of optimal portfolios relative to the benchmark portfolio. The methodology suggested

in Ledoit and Wolf (2008, Note 3.2) was used to obtain p-values.

16
 4.5 Results

Mean-variance bond portfolios

In this section, we present the out-of-sample results of optimal mean-variance portfolios for

the data discussed in Section 4.1 and for different levels of risk aversion δ. The specifications

used to model expected bond returns are the dynamic version of the 3-factor Nelson-Siegel

model (DNS) discussed in Section 2.2 and its arbitrage-free counterpart (AFDNS). To assess

the robustness of the results, we consider two alternative specifications to model the factor

dynamics (AR(1) and VAR(1)). Optimal portfolio compositions are re-balanced on a monthly

basis. Table 3 reports the performance results of mean-variance bond portfolios. The statistics

of returns, standard deviation and Sharpe ratio are annualized. We use an asterisk to indicate

the instances in which the Sharpe ratio of the mean-variance portfolios is statistically higher

than that obtained by the benchmark portfolio strategy with highest Sharpe ratio (3-year

bond) at a significance level of 10%.2 In order to have an idea about portfolio composition,

we also report the average duration (in years) of the mean-variance portfolios. All figures are

based on out-of-sample observations.

The results in Table 3 show that the optimal mean-variance portfolios of US bonds have

risk-return profiles that vary substantially across the levels of risk tolerance considered. The

Table shows that the optimal mean-variance portfolio of US bonds achieved an annualized

average gross return ranging from 5.9% to 11.6% and an annualized standard deviation rang-

ing from 1.3% to 12.1%. The risk-adjusted performance measured by the annualized Sharpe

ratio ranges from 0.35 to 0.67. As expected, we observe that an increase in the risk aversion

coefficient δ leads to decreases in portfolio risk (measured by the standard deviation) as well

as to decreases in portfolio average return and portfolio turnover. Moreover, we find that

an increase in the risk aversion coefficient leads to optimal portfolio composition with lower

duration (i.e. invested in short term maturities). This result is also mostly expected, since

investors with higher risk aversion can lower portfolio risk by investing in shorter maturities.

For instance, the average portfolio duration across specifications for an investor with risk
2
Table 2 shows the performance of all benchmark bond portfolio strategies.

17
aversion coefficient δ = 1 is 0.19 years whereas the same figure for an investor with δ = 0.01

is 7.27 years.

A comparison among alternative specifications for mean-variance portfolios suggest that

both DNS and AFDNS models deliver good results in comparison to the benchmark strategy.

In this sense, it is difficult to identify which one is the best. A similar conclusion is achieved

when comparing the performance among alternative specifications for the factor dynamics

(AR and VAR).

The most important message in 3 is that the mean-variance bond portfolios seems to be

very reasonable alternatives to the portfolios built upon the traditional yield curve strate-

gies discussed in Section 4.4. We observe that several specifications for the mean-variance

portfolios achieved higher Sharpe ratios in comparison to all benchmark policies considered.

Overall, the best performance in terms of Sharpe ratio (0.675) is achieved by the mean-

variance portfolio obtained with the AFDNS/VAR specification with δ = 0.1. This figure is

statistically and substantially higher than the one achieved by the benchmark bullet portfolio

invested in the 3-year bond (0.447).

It is also worth noting that the differences in the portfolios’ risk-return profiles indicate

that investors with both low and high risk appetite can benefit from adopting the mean-

variance policy. For instance, a mean-variance investor with risk aversion δ = 1 × 10−4 and

employing the DNS/AR specification would select a portfolio invested mostly in long term

maturities (average portfolio duration of 6.7 years) thus achieving a Sharpe ratio of 0.574,

which is higher that the Sharpe ratios of all benchmark policies. Similarly, a mean-variance

investor with risk aversion δ = 0.5 and employing the AFDNS/AR specification would select

a portfolio invested mostly in short term maturities (average portfolio duration of 0.7 years)

thus achieving a Sharpe ratio of 0.539, which is also higher that the Sharpe ratios of all

benchmark policies.

[Table 3 about here.]

We further illustrate the performance of the optimal US mean-variance portfolios by plot-

ting in Figure 2 the cumulative portfolio returns over the out-of-sample period obtained with

18
the AFDNS/AR specification when δ = 0.01. The Figure also plots the cumulative returns

of three benchmark bond portfolios strategies (barbell portfolio, 3-year bond, and 10-year

bond). We observe a striking difference in favor of the mean-variance portfolio throughout

the full the sample period. Moreover, this specification for the mean-variance portfolios yields

not only higher cumulative returns, but also higher average returns (both gross and excess)

and higher Sharpe ratio in comparison to the benchmark strategies, as indicated in Table 3.

[Figure 2 about here.]

The results reported in Table 3 and illustrated in Figure 2 suggest that the mean-variance

portfolios obtained with the proposed estimators for the vector of expected bond returns

and its covariance matrix deliver better risk-adjusted performance with respect to traditional

bond portfolio strategies. However, it is still unclear how much an investor is willing to pay

to switch from a traditional bond portfolio strategy to a mean-variance portfolio strategy. In

order to address this question, we report in Table 4 the annualized performance fee an investor

with quadratic utility is willing to pay to change from a each of the traditional bond portfolio

strategy considered in Section 4.4 to a mean-variance portfolio strategy. The results in Table

4 are unambiguous and corroborate those in Table 3. We find that the investor is willing to

pay a fee to adopt the mean-variance portfolio policy in all cases, and that this performance

fee increases with the investor’s risk aversion coefficient. For instance, an investor with risk

aversion coefficient δ = 1 is willing to pay an annualized fee of 34 basis points to change

from the spread bond portfolio policy to the mean-variance portfolio policy. On average, the

investor is willing to pay a performance fee of 2 to 15 basis points to switch from traditional

bond portfolio strategies to a mean-variance portfolio strategy.

[Table 4 about here.]

Duration-constrained mean-variance bond portfolios

The results of the duration-constrained mean-variance bond portfolios are shown in Table 5.

We find that an increase in the target portfolio duration leads to optimal portfolios with higher

19
average returns, higher average excess returns and higher standard deviation. This result is

expected, since a higher target duration will lead to portfolio compositions invested in longer

maturities and, therefore, with higher returns and risk. The most important result of this

analysis, however, is the relative performance of the duration-constrained mean-variance bond

portfolios with respect to the performance of traditional bond portfolio strategies. For each

value of the target portfolio duration considered in Table 5, we compare the risk-adjusted

performance of the duration-constrained mean-variance portfolios with that obtained by a

portfolio invested in a bond with equal duration. For instance, in the case of a target portfolio

duration of 1 year, the performance of the mean-variance portfolios are compared to that

obtained by a portfolio invested in the 1-year bond. Finally, we assume an investor with risk

aversion coefficient δ = 1. We have also conducted this analysis for alternative values for the

risk aversion coefficient δ. The results are very similar to those reported here.

The risk-adjusted performance of the duration-constrained mean-variance bond portfolios

are superior to that of the benchmark strategies with equal duration levels in several cases.

For instance, when considering a target portfolio duration of 5 years, the duration-constrained

mean-variance portfolio obtained with the AFDNS/AR specification obtained a Sharpe ratio

of 0.429, whereas the same value for the portfolio invested in the 5-year bond is 0.387. More-

over, we find that an investor with a quadratic utility function is willing to pay a fee to adopt

the duration-constrained mean-variance portfolio policy in all cases.

Another important result is that the volatility (standard deviation) of the duration-

constrained portfolios is always lower than that of the benchmark portfolio with equal dura-

tion. For instance, the annualized standard deviation of the duration-constrained mean-

variance portfolio with a 3-year target duration is 4.07, whereas the same value for the

benchmark portfolio invested in the 3-year bond is 4.73. Similar results are obtained for

other duration levels. This suggests that diversification plays an important role here, since

it is possible to achieve a desired duration level while reducing portfolio risk by investing in

bonds with different maturities.

[Table 5 about here.]

20
 Summarizing the results reported in Tables 3 to 5, we find that the optimal mean-variance

portfolios based on the closed-form estimates for the vector of expected bond returns and the

covariance matrix of the bond returns presented in Section 2.2 have superior out-of-sample

performance with respect to the benchmark policies from at least two standpoints. First, the

risk-adjusted performance of mean-variance portfolios based on dynamic factor models are

significantly better than that of the benchmark policies. Second, we find that an investor

is willing to pay a performance fee to switch from traditional yield curve strategies to the

mean-variance and duration-constrained mean-variance portfolios.

Assessing the impact of portfolio re-balancing frequency

The results discussed in Table 3 assume a monthly re-balancing frequency. The transaction

costs involved in this re-balancing frequency might degrade the performance of the portfolios

and hinder its implementation in practice. Thus, the performance of optimized portfolios is

also evaluated in the case of quarterly re-balancing frequency. A potentially negative effect

of adopting a lower re-balancing frequency is that the optimal compositions may become

outdated.

Table 6 brings the results of the optimal mean-variance portfolios under quarterly re-

balancing frequency. As expected, we find that lowering the portfolio re-balancing frequency

leads to a substantial decrease in portfolio turnover. For instance, the average turnover of the

mean-variance portfolio under monthly re-balancing across all specifications is 0.719 whereas

the same figure for the quarterly re-balancing is 0.233. This result suggests that lowering

the portfolio re-balancing frequency can lead to substantial decreases in transaction costs.

We also find, however, that, on average, lowering the portfolio re-balancing frequency leads

to decreases in the risk-adjusted performance measured by the Sharpe ratio. Nevertheless,

for some values of the risk aversion coefficient the mean-variance portfolios obtained with

the DNS/VAR specification outperform the benchmark strategy (3-year bond portfolio). For

instance, when δ = 1 × 10−1 , the mean-variance portfolios obtained with the DNS/VAR

specification achieved a Sharpe ratio of 0.539, whereas the same figure for the benchmark

strategy is 0.447. This result suggests that the mean-variance portfolios can outperform

21
traditional yield curve strategies even when optimal portfolio compositions are re-balanced

less frequently.

[Table 6 about here.]

5 Concluding remarks
The mean-variance approach introduced by Markowitz (1952) to obtain optimal portfolios has

been widely used by market participants and largely documented in the academic literature.

However, the use of this methodology for the optimization of portfolios composed of fixed-

income securities has received little attention in the literature. In order to address this

shortcoming, this paper adopts the mean-variance approach to bond portfolio optimization

based on dynamic factor models for the term structure of interest rate, such as the dynamic

version of the Nelson-Siegel model proposed by Diebold and Li (2006) and its arbitrage-free

version proposed by Christensen et al. (2011). These are standard specifications to model

interest rates widely used by market participants and academics. In this paper, we extend

their utilization to a broader context of bond portfolio selection.

We show that factor models for the yield curve simplify the process of bond portfolio

management, since it allows the computation of the vector expected bond returns and its

covariance matrix in closed form. We show how to obtain closed-form expressions for these

two moments based on a general class of dynamic factor models, and use them to obtain opti-

mal mean-variance and duration-constrained mean-variance portfolios. Our evidence indicate

that the mean-variance portfolios obtained with the proposed estimators for the vector of ex-

pected bond returns and its covariance matrix deliver better risk-adjusted performance with

respect to traditional bond portfolio strategies. Moreover, an investor with quadratic utility

is willing to pay a performance fee to change from a each of the traditional bond portfolio

strategies to a mean-variance portfolio strategy.

22
Appendix: The arbitrage-free Nelson-Siegel model
(AFDNS)
The theoretical weakness of the standard DNS model is that is not defined under an arbitrage free
setting. Christensen et al. (2011) have extended this model to overcome this by adhering to the
standard continuous time affine diffusion processes developed in Duffie and Kan (1996). The resulting
class of AFDNS models involves considering the filtered probability space (Ω, F, (Ft ), Q) with filtration
(Ft ) = {Ft : t ≤ 0} satisfying the usual conditions of Williams (1997). Here Q denotes the risk neutral
measure and we will denote the real world probability measure by P . If the resulting risk neutral
dynamic factors, generically denoted by state vector Xt , for the DNS term structure model are defined
on such a probability space and follow a Markov process defined on set M ⊆ Rn that is a solution to
the stochastic differential equation given by
 
dXt = K(t)P Θ(t)P − Xt dt + Σ(t)D(Xt , t)dWtP , (12)

with WQ a standard Brownian Motion in Rn defined according to the filtration (Ft ). Furthermore, as
in Duffie and Kan (1996) one assumes the drifts and dynamics are bounded continuous functions such
that ΘQ : [0, T ] → Rn , KQ : [0, T ] → Rn×n and volatility Σ : [0, T ] → Rn×n . Finally, they assume the
diagonal mapping D : M × [0, T ] → Rn×n with diagonal elements given by
q
[D]ii = γ i (t) + δ i (t)Xt1 + . . . + δ n (t)Xtn ∀i ∈ {1, . . . , n},

where each γ i : [0, T ] → Rn and δ i : [0, T ] → Rn×n are bounded continuous functions. In addition
Duffie and Kan (1996) assume the instantaneous risk neutral rate is an affine function of the state
variables given by rt = r0 (t) + r1 (t)0 Xt with bounded continuous functions r0 : [0, T ] → R and
r1 : [0, T ] → Rn . Under this affine formulation of the continuous latent factor dynamics Duffie and
Kan (1996) proved that a closed form analytic expression for the zero-coupon bond prices is attained
as a linear function of the latent dynamic factors. This gives the dynamic zero-coupon yield at time t
for a bond with maturity T as
" Z #!
T
1 1
y(t, T ) = − log E −Q
ru du =− [A(t, T ) + B(t, T )0 Xt ] (13)
T −t t T −t

with A(t, T ) and B(t, T ) obtained as solutions to the system of ordinary differential equations ode?s
with boundary conditions A(0) = 0 and B(0) = 0 given by
n
dA(t, T )
0 1X 0
0
= r1 + KQ B(t, T ) − [Σ B(t, T )B(t, T )0 Σ]ii δ i (14)
dt 2 i=1
n
dB(t, T )
0 1X 0
0
= r0 + B(t, T ) KQ ΘQ − [Σ B(t, T )B(t, T )0 Σ]ii γ i . (15)
dt 2 i=1

Since we would like to work under a model that is the arbitrage-free equivalent model for the
DNS specification, we see that by considering the generic affine latent factor dynamic s.d.e. structure
proposed in Equation (13), it is clear that to recover the equivalent AFDNS model one requires
solutions to this system of odes given by

B 1 (t, T ) = − (T − t)
1 − e−λ(T −t)
B 2 (t, T ) = −
λ (T − t)
1 − e−λ(T −t)
B 3 (t, T ) = − + (T − t) e−λ(T −t)
λ

23
which results in a solution for y(t, T ) given by,

1 − e−λ(T −t) 1 − e−λ(T −t)

   
A(t, T )
yt (τ ) = X1t + X2t + X3t − e−λ(T −t) − , (16)
λ (T − t) λ (T − t) T −t

where the additional term − A(t,T )

T −t is the correction to the DNS model to ensure it satisfies the
theoretical properties of being arbitrage free.
Unlike the DNS-type models, the AFDNS-type models of Christensen et al. (2011) impose partic-
ular structure on both the risk neutral and real world latent factor dynamic processes. The ordinary
differential equations for the coefficients of A(t, T ) and B(t, T ) are only functions of r1 and KQ . How-
ever, under the assumptions of Christensen et al. (2011) in which the mean state variable levels under
the risk neutral measure at zero, the additional terms for the drift ΘQ and volatility matrix Σ appear
in the solution to the yield adjustment term A(t,T )
T −t which is found as a solution of the form,

3 Z T
A(t, T ) 1 1 X
− =− [Σ0 B(s, T )B(s, T )0 Σ]ii ds. (17)
T −t 2 T − t i=1 t

Obtaining a solution to the system of equations that satisfy these solutions can be achieved for
a large family of possible AFDNS models, such as those discussed in Christensen et al. (2011) in
Proposition 1 which presents one such model in which the instantaneous risk neutral rate is given by
rt = X1t + X2t where the latent state variables Xt = (X1t , X2t , X3t ) are described by the following
system of equations under the risk neutral measure Q given by
     Q     Q 
dX1t 0 0 0 θ1 dX1t dW1t
 dX2t  =  0 λ −λ   θQ  −  dX2t  dt + Σ  dW Q  , λ > 0. (18)
2 2t
dX3t 0 0 λ θ3Q dX3t Q
dW3t

In addition, it has been shown that the solution to the equation specifying the yield adjustment
term A(t, T ) results in a restriction to the volatility matrix of the form
 
σ11 0 0
Σ =  σ21 σ22 0 . (19)
σ31 σ32 σ33

This completes the specification of the AFDNS model under the risk neutral measure.

Independent and correlated AFDNS models

Opposed to the DNS case, the AFDNS models are formulated in continuous time. To estimate models
in this environment a change of probability measure is needed, as the theoretical part above focused
on Q-dynamics. By setting up an affine risk specification as in Duffee (2002) the equation in (12)
can be formulated to remain affine under P-dynamics (see Christensen et al. (2011) for full details).
Because of this, any vector Θ and matrix K under the P-measure preserve the risk neutral dynamics
discussed above. This flexibility guarantees that identical (P-measure) models can be estimated, so
our focus is on an independent and correlated model.
AFDNS are an affine model, meaning analytical formula exist for yields. Using the solution already
presented in Equation (16) yields most have the following relationship with the state variables

yt = A + BXt + εt (20)

Note that the matrix B is identical in the DNS and AFDNS models (compare with Equation (??) and
Equation (16)). The only difference is the addition of the vector A containing the yield-adjustment
terms in the AFDNS models. Because the AFNS is a continuous model the “time dimension” is

24
modeled in terms of dynamics instead of a time-series model such as an AR or VAR. However, it does
not mean similar models cannot be estimated to ensure comparison, as shown below.
For the independent-factor AFDNS model the dynamics of the state variables (working under the
P-measure now)
   P   P      P

dX1t κ11 0 0 θ1 X1t σ11 0 0 dW1t
 dX2t  =  0 κP22 0   θ2P  −  X2t  dt +  0 σ22 0   dW2t P 
. (21)
P P P
dX3t 0 0 κ33 θ3 X3t 0 0 σ33 dW3t

In the correlated-factor AFDNS model, the three shocks may be correlated, and there may be full
interaction among the factors as they adjust to the steady state
   P   P      
dX1t κ11 κP12 κP13 θ1 X1t σ11 0 0 P
dW1t
 dX2t  =  κP21 κP22 κP23   θ2P  −  X2t  dt +  σ21 σ22 P 
0   dW2t . (22)
dX3t κP31 κP32 κP33 θ3P X3t σ31 σ32 σ33 P
dW3t

From this specification the resemblance to the DNS independent and correlated model is striking. This
is the most flexible version of the AFDNS models where all parameters are identified.

An exact expression for the covariance of the continuous-time

AFDNS model
In this paper, the models are estimated using maximum likelihood estimation method based on the
Kalman filter. The AFDNS model can be formulated in a state space form as follow. For the
continuous-time AFDNS models, the conditional mean vector and the conditional covariance matrix
are
h P
i P
E [XT |Ft ] = I − e−K ∆t ΘP + e−K ∆t Xt (23)
Z ∆t
P P 0
V [XT |Ft ] = e−K s ΣΣ0 e−(K ) s ds (24)
0

To estimate AFNS models in the state-space Kalman-filter maximum-likelihood estimation frame-

work proposed by Christensen et al. (2011) we must compute the conditional covariance matrix
Z ∆t T
e−K s ΣΣT e−(K ) s ds
P P
P
V [ Xt | Ft−1 ] = (25)
0

of discrete observations. Since the estimation is an intense computational process, we need to provide
fast intermediate calculations. One approach is to approximate the integral and the matrix exponential.
Another approach uses the diagonalization of KP to calculate the integral exactly. Due to the reduced
size of the matrix KP , the latter is significantly faster than the former, at least when compared to
naı̈ve numerical integration techniques.
The AFNS state transition equation is




Xt = I − exp −KP ∆t ΘP + exp −KP ∆t Xt−1 + ηt ,

where ∆t is the time between the observations at t and t − 1, with measurement equation

yt = A + BXt + εt ,

and error structure      

ηt 0 Q 0
∼N , ,
εt 0 0 H
where H is diagonal, Q = V P [ Xt | Ft−1 ] and the transition and measurement errors are assumed

25
orthogonal to the initial state.
The computation of Q is straightforward if we use the diagonalization of KP

KP = V ΛV −1 , (26)

where V contains the eigenvectors of KP , and Λ is a diagonal matrix containing the eigenvalues (λi )
of KP . We refer the reader to ? and ? for details on linear algebra theory and on the computational
aspects of linear algebra, respectively.
Substituting (26) in (25), using

exp(−KP s) = V exp(−Λs)V −1 ,

and similarly

T
T
exp(− KP s) = V −1 exp(−Λs)V T ,
we obtain !
Z ∆t
Q=V exp(−Λs)Ω exp(−Λs)ds V T ,
0

T
where Ω = (ωij )n×n = V −1 ΣΣT V −1 . Since the exponential of a diagonal matrix with entries −λi s
is a diagonal matrix with entries e−λi s , each term of the matrix under the integral is (ωij e−(λi +λj )s )n×n .
Integration yields an expression which only involves matrix multiplications
 
ωij  −(λi +λj )∆t
Q=V 1−e V T. (27)
λi + λj n×n

Stationarity of the system under the P-measure is ensured if the real component of all the eigen-
values of KP is positive, and this condition is imposed in all estimations. For this reason, we can start
the Kalman filter at the unconditional mean, X0 = ΘP , and covariance matrix, Σ0 . In particular,
R∞ P T
the unconditional variance Σ0 = 0 e−K s ΣΣT e−(K ) s ds used in the initialization of the filter is
P

easily obtained from the above expression. Assuming KP has eigenvalues with positive real parts, the
integral is convergent to  
ωij
Σ0 = V V T. (28)
λi + λj n×n

26
References
BIS, 2005. Zero-coupon yield curves: technical documentation. Technical Report. Bank for Interna-
tional Settlements.

Brandt, M., 2009. Portfolio choice problems. Handbook of Financial Econometrics 1, 269–336.

Christensen, J.H., Diebold, F.X., Rudebusch, G.D., 2011. The affine arbitrage-free class of nelson-siegel
term structure models. Journal of Econometrics 164, 4–20.

DeMiguel, V., Garlappi, L., Uppal, R., 2009. Optimal versus naive diversification: how inefficient is
the 1/N portfolio strategy? Review of Financial Studies 22, 1915–1953.

Diebold, F., Li, C., 2006. Forecasting the term structure of government bond yields. Journal of
Econometrics 130, 337–364.

Diebold, F.X., Rudebusch, G.D., 2011. The Dynamic Nelson-Siegel Approach to Yield Curve Modeling
and Forecasting. mimeo.

Diebold, F.X., Rudebusch, G.D., Aruoba, S.B., 2006. The macroeconomy and the yield curve: a
dynamic latent factor approach. Journal of Econometrics 131, 309–338.

Duffee, G.R., 2002. Term premia and interest rate forecasts in affine models. Journal of Finance 57,
405–443.

Duffie, D., Kan, R., 1996. A yield-factor model of interest rates. Mathematical Finance 6, 379–406.

Fabozzi, F., 2002. Fixed Income Securities. John Wiley & Sons, Hoboken, NJ.

Fabozzi, F., Fong, G., 1994. Advanced fixed income portfolio management: the state of the art.
Probus.

Fama, E., French, K., 1993. Common risk factors in the returns on stocks and bonds. Journal of
Financial Economics 33, 3–56.

Fama, E.F., Bliss, R.R., 1987. The information in long-maturity forward rates. The American Eco-
nomic Review , 680–692.

Fleming, J., Kirby, C., 2003. A Closer Look at the Relation between GARCH and Stochastic Autore-
gressive Volatility. Journal of Financial Econometrics 1, 365–419.

Fleming, J., Kirby, C., Ostdiek, B., 2001. The economic value of volatility timing. The Journal of
Finance 56, 329–352.

Gürkaynak, R., Sack, B., Wright, J., 2007. The US treasury yield curve: 1961 to the present. Journal
of Monetary Economics 54, 2291–2304.

Harvey, A., Ruiz, E., Sentana, E., 1992. Unobserved component time series models with ARCH
disturbances. Journal of Econometrics 52, 129–157.

Haustsch, N., Ou, Y., 2012. Analyzing interest rate risk: Stochastic volatility in the term structure of
government bond yields. Journal of Banking & Finance 36, 2988–3007.

Jagannathan, R., Ma, T., 2003. Risk reduction in large portfolios: Why imposing the wrong constraints
helps. The Journal of Finance 58, 1651–1684.

Jones, F.J., 1991. Yield curve strategies. The Journal of Fixed Income 1, 43–48.

27
Jungbacker, B., Koopman, S., 2008. Likelihood-based analysis for dynamic factor models. Tinbergen
Institute Discussion Paper Found on http://www.tinbergen.nl.

Jungbacker, B., Koopman, S.J., van der Wel, M., 2013. Smooth dynamic factor analysis with an
application to the us term structure of interest rates. Journal of Applied Econometrics forthcoming.

Korn, O., Koziol, C., 2006. Bond portfolio optimization. The Journal of Fixed Income 15, 48–60.

Ledoit, O., Wolf, M., 2008. Robust performance hypothesis testing with the Sharpe ratio. Journal of
Empirical Finance 15, 850–859.

Markowitz, H., 1952. Portfolio selection. Journal of Finance 7, 77–91.

Meucci, A., 2009. Risk and Asset Allocation. Springer Finance, Springer.

Nelson, C.R.N., Siegel, A.F., 1987. Parsimonious modeling of yield curves. The Journal of Business
60, 473–489.

Politis, D., Romano, J., 1994. The stationary bootstrap. Journal of the American Statistical Associ-
ation 89, 1303–1313.

Puhle, M., 2008. Bond portfolio optimization. Lecture notes in economics and mathematical systems,
Springer.

Stock, J., Watson, M., 2002. Forecasting using principal components from a large number of predictors.
Journal of the American Statistical Association 97, 1167–1179.

Vasicek, O., 1977. An equilibrium characterization of the term structure. Journal of Financial Eco-
nomics 5, 177–188.

Williams, D., 1997. Probability with Martingales. Cambridge University Press.

28
Figures

Figure 1: US term structure dynamics over time

Note: This figure details the evolution of the US term structure of interest rates over 1970:01-2009:12. We use the following
maturities: 3, 6, 9, 12, 15, 18, 21, 24, 30, 36, 48, 60, 72, 84, 96, 108 and 120 months.

18

16

14

12
Yield (%)

10

0
120

96

72
Ma
tur 48
ity 2010
(M 2005
on 24 1995
2000
ths 1985
1990
) 0 1975
1980
1970
Time

29
Figure 2: Cumulative returns for the US mean-variance portfolio

30
Mean−variance − AFDNS/AR
Barbell portfolio
25
3−year bond
10−year bond
20

15

10

−5
50 100 150 200 250 300 350

30
Tables

Table 1: Descriptive statistics for the U.S. Treasuries data set

Summary statistics for US-Treasury data set consisting of monthly yield data from January 1970 to December 2009. We
show for each maturity mean, standard deviation, minimum, maximum, skewness, kurtosis, three (1 month, 12 month
and 24 month) autocorrelations (Acf, ρb(1), ρb(12) and ρb(24))) and two-month partial autocorrelation.

Acf Pacf

τ Mean Std Dev Min Max Skew Kurt ρb(1) ρb(12) ρb(24) α
b(2)

6 5.97 3.10 0.15 16.48 0.67 3.82 0.980 0.763 0.516 -0.126
9 6.08 3.09 0.19 16.39 0.63 3.71 0.981 0.771 0.538 -0.145
12 6.17 3.05 0.25 16.10 0.57 3.59 0.981 0.777 0.552 -0.154
15 6.25 3.03 0.38 16.06 0.52 3.49 0.982 0.785 0.571 -0.149
18 6.32 3.01 0.44 16.22 0.52 3.46 0.983 0.792 0.585 -0.153
21 6.39 2.99 0.53 16.17 0.53 3.46 0.983 0.797 0.598 -0.143
24 6.42 2.94 0.53 15.81 0.52 3.40 0.983 0.799 0.610 -0.163
30 6.51 2.88 0.82 15.43 0.50 3.32 0.983 0.801 0.627 -0.135
36 6.60 2.83 0.98 15.54 0.53 3.35 0.984 0.814 0.642 -0.132
48 6.76 2.75 1.02 15.60 0.57 3.33 0.985 0.823 0.664 -0.112
60 6.85 2.67 1.56 15.13 0.61 3.28 0.986 0.832 0.685 -0.103
72 6.96 2.64 1.53 15.11 0.64 3.26 0.987 0.842 0.702 -0.106
84 7.03 2.57 2.18 15.02 0.71 3.30 0.987 0.841 0.709 -0.124
96 7.07 2.54 2.11 15.05 0.75 3.29 0.988 0.850 0.721 -0.121
108 7.10 2.52 2.15 15.11 0.80 3.33 0.988 0.853 0.724 -0.141
120 7.07 2.46 2.68 15.19 0.86 3.41 0.988 0.843 0.717 -0.117

31
 Table 2: Performance of traditional yield curve strategies
The Table reports performance statistics for the traditional yield curve strategies. The statistics of returns, standard
deviation and Sharpe ratio are annualized and the average portfolio duration is measured in years. The excess return is
calculated using the 3-month rate as a risk-free asset.

Yield curve Mean Mean excess Standard Sharpe

strategy return (%) return (%) deviation ratio
Spread portfolio 2.253 2.253 11.544 0.195
Barbell portfolio 7.814 2.022 7.184 0.282
Laddered portfolio 7.784 1.992 5.214 0.382
1-year bond 6.688 0.896 2.021 0.443
3-year bond 7.906 2.113 4.728 0.447
5-year bond 8.614 2.822 7.298 0.387
7-year bond 9.232 3.440 9.603 0.358
9-year bond 9.491 3.698 11.738 0.315
10-year bond 8.941 3.149 12.875 0.245

32
 Table 3: Performance of optimal US-Treasuries mean-variance portfolios
The Table reports performance statistics for mean-variance portfolios using US zero-coupon yields with maturities equal
to 6, 9, 12, 15, 18, 21, 24, 30, 36, 48, 60, 72, 84, 96, 108 and 120 months. The yield curve model used is the dynamic
version of the Nelson-Siegel 3-factor model (DNS) and its arbitrage-free version (AFDNS). The optimal portfolios are
re-balanced on a monthly basis. The statistics of returns, standard deviation and Sharpe ratio are annualized and the
average portfolio duration is measured in years. The excess return is calculated using the 3-month rate as a risk-free asset.
δ denotes the value of the risk aversion coefficient. Asterisks indicate that the coefficient is statistically higher than that
obtained benchmark bond portfolio (3-year bullet portfolio) at a significance level of 10%.

Yield curve Factor Mean Mean excess Standard Sharpe Turnover Average
model dynamics return (%) return (%) deviation ratio duration
δ = 1 × 10−4
DNS AR 11.656 5.864 10.208 0.574∗ 0.933 6.735
DNS VAR 11.547 5.755 10.701 0.538∗ 0.922 7.459
AFDNS AR 10.101 4.309 11.181 0.385 0.743 7.638
AFDNS VAR 11.181 5.389 11.095 0.486∗ 0.866 7.441
δ = 1 × 10−3
DNS AR 11.656 5.864 10.191 0.575∗ 0.950 6.690
DNS VAR 11.513 5.721 10.685 0.535∗ 0.931 7.422
AFDNS AR 10.125 4.333 11.140 0.389 0.752 7.585
AFDNS VAR 11.149 5.357 11.013 0.486∗ 0.866 7.386
δ = 1 × 10−2
DNS AR 11.216 5.424 10.021 0.541∗ 0.935 6.208
DNS VAR 11.423 5.631 10.408 0.541∗ 0.961 7.004
AFDNS AR 10.144 4.352 10.740 0.405 0.817 7.040
AFDNS VAR 10.890 5.098 10.591 0.481∗ 0.924 6.846
δ = 1 × 10−1
DNS AR 9.057 3.264 5.248 0.622∗ 0.959 2.179
DNS VAR 9.516 3.724 5.517 0.675∗ 1.074 2.829
AFDNS AR 9.182 3.390 5.771 0.587∗ 0.931 2.439
AFDNS VAR 9.477 3.685 6.178 0.596∗ 1.021 2.654
δ = 0.5
DNS AR 6.734 0.942 1.902 0.495∗ 0.507 0.685
DNS VAR 6.695 0.903 1.767 0.511∗ 0.479 0.724
AFDNS AR 6.931 1.139 2.114 0.539∗ 0.544 0.748
AFDNS VAR 7.181 1.388 2.777 0.500∗ 0.545 0.786
δ = 1.0
DNS AR 6.334 0.542 1.492 0.363 0.133 0.526
DNS VAR 6.297 0.505 1.424 0.354 0.104 0.525
AFDNS AR 6.386 0.594 1.505 0.395 0.160 0.539
AFDNS VAR 6.539 0.747 1.909 0.391 0.201 0.567

33
Table 4: Performance fee to switch from traditional yield curve strategies to the
mean-variance strategy
The Table reports the annualized performance fee (∆) in basis points that the investor is willing to pay to change from a
traditional bond portfolio strategy to a mean-variance portfolio strategy. δ denotes the value of the risk aversion coefficient.

Yield curve Factor

model dynamics ∆spread ∆barbell ∆1Y ∆3Y ∆5Y ∆7Y ∆9Y ∆10Y ∆ladder
δ = 1 × 10−4
DNS AR 9.403 3.839 4.963 3.746 3.039 2.424 2.167 2.718 3.868
DNS VAR 9.293 3.729 4.853 3.636 2.929 2.314 2.057 2.608 3.758
AFDNS AR 7.848 2.283 3.407 2.190 1.484 0.868 0.611 1.162 2.312
AFDNS VAR 8.927 3.363 4.487 3.270 2.563 1.948 1.691 2.242 3.392
δ = 1 × 10−3
DNS AR 9.403 3.809 4.910 3.703 3.012 2.414 2.179 2.741 3.828
DNS VAR 9.255 3.661 4.762 3.555 2.864 2.266 2.031 2.593 3.680
AFDNS AR 7.864 2.270 3.372 2.165 1.473 0.876 0.640 1.203 2.289
AFDNS VAR 8.888 3.294 4.395 3.189 2.497 1.900 1.664 2.226 3.313
δ = 1 × 10−2
DNS AR 8.987 3.100 3.974 2.866 2.322 1.907 1.878 2.556 3.010
DNS VAR 9.150 3.264 4.138 3.030 2.486 2.071 2.042 2.720 3.174
AFDNS AR 7.861 1.975 2.849 1.741 1.197 0.782 0.753 1.431 1.884
AFDNS VAR 8.609 2.722 3.596 2.488 1.944 1.530 1.500 2.178 2.632
δ = 1 × 10−1
DNS AR 10.686 2.140 0.949 0.737 1.538 2.784 4.633 6.362 1.060
DNS VAR 10.940 2.395 1.203 0.992 1.793 3.039 4.887 6.617 1.314
AFDNS AR 10.530 1.985 0.793 0.582 1.383 2.629 4.478 6.207 0.904
AFDNS VAR 10.557 2.012 0.820 0.608 1.410 2.656 4.504 6.233 0.931
δ = 0.5
DNS AR 23.977 7.474 0.100 2.572 7.398 13.614 21.083 25.958 3.430
DNS VAR 24.040 7.537 0.163 2.635 7.462 13.678 21.147 26.021 3.493
AFDNS AR 23.929 7.426 0.052 2.524 7.351 13.567 21.036 25.910 3.382
AFDNS VAR 23.508 7.005 -0.369 2.103 6.930 13.146 20.615 25.489 2.961
δ=1
DNS AR 33.981 12.007 0.383 4.700 12.294 21.927 33.259 40.296 5.926
DNS VAR 34.021 12.047 0.423 4.740 12.334 21.967 33.299 40.336 5.966
AFDNS AR 33.983 12.009 0.385 4.702 12.296 21.929 33.262 40.298 5.928
AFDNS VAR 33.674 11.700 0.076 4.393 11.987 21.620 32.952 39.989 5.619

Average 15.805 5.127 2.279 2.786 4.666 7.244 10.599 13.171 3.252

34
Table 5: Performance of optimal US-Treasuries duration-constrained mean-variance
portfolios
The Table reports performance statistics for duration-constrained mean-variance portfolios using US zero-coupon yields
with maturities equal to 6, 9, 12, 15, 18, 21, 24, 30, 36, 48, 60, 72, 84, 96, 108 and 120 months. The yield curve model used
is the dynamic version of the Nelson-Siegel 3-factor model (DNS) and its arbitrage-free version (AFDNS). The optimal
portfolios are re-balanced on a monthly basis. The statistics of returns, standard deviation and Sharpe ratio are annualized.
The excess return is calculated using the 3-month rate as a risk-free asset. ∆ denotes the annualized performance fee
(in basis points) that the investor is willing to pay to change from a bond portfolio strategy to a mean-variance portfolio
with equal duration. Asterisks indicate that the coefficient is statistically different from that obtained by the bullet bond
portfolio strategy with equal duration at a significance level of 10%. We assume an investor with risk aversion coefficient
δ = 1.

Yield curve Factor Mean Mean excess Standard Sharpe Turnover ∆ (b.p.)
model dynamics return (%) return (%) deviation ratio
Target portfolio duration=1-year
DNS AR 6.595 0.803 1.765 0.455∗ 0.337 0.221
DNS VAR 6.571 0.779 1.753 0.444 0.267 0.226
AFDNS AR 6.626 0.834 1.853 0.450 0.326 0.149
AFDNS VAR 6.589 0.797 1.874 0.425 0.310 0.121
Target portfolio duration=3-year
DNS AR 7.611 1.818 4.071 0.447 0.214 1.390
DNS VAR 7.596 1.804 4.071 0.443 0.195 1.388
AFDNS AR 7.645 1.853 4.102 0.452∗ 0.241 1.333
AFDNS VAR 7.720 1.928 4.398 0.438 0.261 0.720
Target portfolio duration=5-year
DNS AR 8.515 2.723 6.603 0.412∗ 0.163 2.395
DNS VAR 8.514 2.722 6.605 0.412∗ 0.166 2.389
∗
AFDNS AR 8.629 2.837 6.607 0.429 0.203 2.398
AFDNS VAR 8.666 2.873 6.886 0.417∗ 0.252 1.464
Target portfolio duration=7-year
DNS AR 9.409 3.617 9.186 0.394∗ 0.166 1.972
DNS VAR 9.407 3.614 9.187 0.393∗ 0.166 1.969
AFDNS AR 9.576 3.784 9.162 0.413∗ 0.207 2.101
∗
AFDNS VAR 9.628 3.836 9.339 0.411 0.232 1.288
Target portfolio duration=9-year
DNS AR 9.619 3.827 11.653 0.328 0.067 0.511
DNS VAR 9.619 3.827 11.653 0.328 0.067 0.510
AFDNS AR 9.993 4.201 11.669 0.360∗ 0.126 0.450
AFDNS VAR 9.927 4.134 11.691 0.354∗ 0.130 0.319

35
Table 6: Performance of optimal US-Treasuries mean-variance portfolios under quar-
terly re-balancing frequency
The Table reports performance statistics for mean-variance portfolios using US zero-coupon yields with maturities equal
to 6, 9, 12, 15, 18, 21, 24, 30, 36, 48, 60, 72, 84, 96, 108 and 120 months. The yield curve model used is the dynamic
version of the Nelson-Siegel 3-factor model (DNS) and its arbitrage-free version (AFDNS). The optimal portfolios are
re-balanced on a quarterly basis. The statistics of returns, standard deviation and Sharpe ratio are annualized and the
average portfolio duration is measured in years. The excess return is calculated using the 3-month rate as a risk-free asset.
δ denotes the value of the risk aversion coefficient. Asterisks indicate that the coefficient is statistically higher than that
obtained benchmark bond portfolio at a significance level of 10%.

Yield curve Factor Mean Mean excess Standard Sharpe Turnover Average
model dynamics return (%) return (%) deviation ratio duration
δ = 1 × 10−4
DNS AR 10.568 4.776 9.851 0.485∗ 0.313 6.774
DNS VAR 10.608 4.815 10.203 0.472∗ 0.302 7.448
AFDNS AR 9.676 3.884 10.615 0.366 0.251 7.306
AFDNS VAR 9.776 3.984 10.210 0.390 0.296 7.061
−3
δ = 1 × 10
DNS AR 10.566 4.774 9.841 0.485∗ 0.322 6.729
DNS VAR 10.609 4.816 10.175 0.473∗ 0.298 7.393
AFDNS AR 9.675 3.883 10.613 0.366 0.251 7.306
AFDNS VAR 9.731 3.939 10.177 0.387 0.294 7.028
δ = 1 × 10−2
DNS AR 10.109 4.317 9.540 0.453∗ 0.309 6.209
DNS VAR 10.655 4.863 9.990 0.487∗ 0.290 7.015
AFDNS AR 9.581 3.789 10.222 0.371 0.252 6.874
AFDNS VAR 9.646 3.854 9.919 0.389 0.294 6.734
δ = 1 × 10−1
DNS AR 8.122 2.329 4.323 0.539∗ 0.307 2.218
DNS VAR 8.276 2.483 4.733 0.525∗ 0.327 2.845
AFDNS AR 7.740 1.947 4.894 0.398 0.292 2.486
AFDNS VAR 7.346 1.553 5.524 0.281 0.287 2.808
δ = 0.5
DNS AR 6.560 0.768 1.701 0.452 0.173 0.697
DNS VAR 6.563 0.770 1.690 0.456∗ 0.157 0.736
AFDNS AR 6.665 0.873 1.949 0.448 0.184 0.775
AFDNS VAR 6.620 0.828 2.151 0.385 0.177 0.793
δ=1
DNS AR 6.252 0.460 1.408 0.327 0.048 0.530
DNS VAR 6.240 0.448 1.409 0.318 0.041 0.529
AFDNS AR 6.301 0.509 1.458 0.349 0.052 0.549
AFDNS VAR 6.308 0.516 1.599 0.322 0.072 0.571

36