Ann Oper Res (2007) 151:193–222

DOI 10.1007/s10479-006-0134-4

A portfolio-based evaluation of affine term structure

models

Andrea Beltratti · Paolo Colla

Published online: 1 December 2006

C Springer Science + Business Media, LLC 2007


Abstract We focus on affine term structure models as tools for active bond portfolio man-
agement. Our financial exercise comprises the following steps: 1) forecast the future values
of the state variables implied by several multi-factor models; 2) approximate the conditional
moments of the state vector to come up with discrete scenarios for the future state variables
3) compute bond returns for various maturities at future dates from the theoretical asset pric-
ing relations 4) solve the portfolio problem faced by an investor with a six month horizon
who takes into account the possibility to rebalance after one quarter. The sequence of opti-
mal portfolios is evaluated in terms of financial properties. We show that a financial based
evaluation of term structure models may yield results conflicting with those obtained from a
statistical evaluation.

Keywords Affine models . Dynamic optimization . Active asset allocation

Introduction

The term structure of interest rates has long been a fashionable topic among financial
economists, and the related literature is extremely wide even by historical standards. Duffie
and Kan (1996) have generalized previous works and shown that the completely affine class
(CA hereafter) nests celebrated models such as Vasicek (1977) and Cox, Ingersoll and Ross
(1985) as specific cases. Within the CA class the product between the volatility matrix and
the price of risk vector is linear in the state variables. Furthermore the variance of the log
state price deflator is linear in the state variables as well. Duffee (2002) has recently proposed
an extension to the essentially affine (EA hereafter) class. The key feature of EA models is
that the variance of the log state price deflator is not affine in the state variables. This opens
the way to model time variation in the price of risk not associated with time variation in the
volatility matrix. Comparing the three specifications Duffee (2002) describes as the best ones

A. Beltratti . P. Colla ()

Università Bocconi, Via Gobbi, 5, 20136 Milan, Italy
e-mail: [email protected]
Springer
194 Ann Oper Res (2007) 151:193–222

in his work, it emerges that the EA class performs better than the CA class in forecasting
bond yields. One important advantage of the EA structure is that it can produce expected
excess returns on 2-year and 10-year bonds which widely fluctuate between positive and
negative values. On the other hand, the proposed CA specification produces strictly positive
expected excess return to 2-year and 10-year bonds all over the sample period. Therefore EA
models are able to reconcile the small unconditional sample mean of bond excess returns
with their high standard deviation. For instance, one EA parametrization predicts expected
excess returns to the 10-year bond lower than −20% for some months during the 1970s and
1980s. While predictions arising from EA models seem to be compatible with the mentioned
empirical evidence, they might be difficult to justify from a general equilibrium perspective.
It is a matter of great concern that at some stage investors might choose to hold long term
bonds with such negative expected excess returns.
Dai and Singleton (2003) evaluate various multi-factor models with respect to their ability
in fitting the main stylized facts in bond prices’ dynamics. Among their main findings are:
(1) three-factor models do a better job than two-factor models in matching the persistence
in yields’ volatility and (2) a three-factor CA Gaussian model can replicate the negative
correlation between yield changes and the slope of the term structure, while multi-factor CIR
(1985) models1 are unable to do so. Taken together, these two findings imply that coming up
with a model able to simultaneously fit the dynamics of both the mean and the volatility of
yields can be a hard job.
We propose an evaluation of CA and EA term structure models as forecasting tools within
a multi-period optimal portfolio problem. Our paper is motivated by three concerns: (1) a
statistical evaluation of forecasting models can produce different results from those arising
from a financial metric; (2) a risk averse investor is interested in predicting not only the mean
of excess returns but also their risk, and (3) it is useful to understand what portfolio policy is
prescribed by a term structure model in order to evaluate its practical applicability.
The first point is simple. Many estimation techniques pin down parameters minimizing
the sum of squared errors, which results in high weights on extreme observations. This is
optimal only for specific loss functions used by the researcher, and produces well-known
results such that the model’s conditional mean coincides with the optimal forecast. However
it is easy to think of loss functions which would not select parameters by minimizing the
sum of squared errors. Aı̈t-Sahalia and Brandt (2001) selects the model specification based
on the first order conditions for utility maximization, thus by-passing the estimation phase
of the statistical forecasting model. In such a way variables and parameters are explicitly
relevant for portfolio choices. Another example is Christoffersen and Diebold (1997) that
solves the conditional forecasting problem for some specific loss functions of the decision-
maker. In several interesting cases the authors show how the conditional mean is not the
optimal forecast, which should instead stem from (a mixture of) the first two conditional
moments.
A corollary to this general argument is that a decision-maker may favor a forecasting
model which does not provide optimal performance from a particular statistical standpoint
like sum of squared errors minimization. There are several examples—for both the stock and
bond market– confirming that a low R 2 model may be useful for active asset allocation.2

1Duffee (2002) provides similar evidence within the EA class: the three-factor EA version of the Gaussian
model outperforms the multi-factor CIR.
2 Moreover, the forecasting specifications in the models mentioned in the text are not the best ones in their
respective categories. For example, an extended model which includes the term spread and the default spread
as supplement to the dividend yield achieves higher values of coefficient of determination.
Springer
Ann Oper Res (2007) 151:193–222 195

Breen, Glosten and Jagannathan (1989) consider a fairly simple forecasting model for stock
returns based solely on the interest rate level. The model is characterized by a R 2 between
2% and 8% —depending on the sample period—and may be employed in a stock/bond static
active asset allocation problem to produce a 2% return. Barberis (2000) documents that a
simple predictive equation for stock returns based on the dividend yield—with a R 2 as low
as 1%—, may deeply affect static and dynamic portfolio choices for long term investors.
Kandel and Stambaugh (1996) study a similar problem focusing on short term investors
within a Bayesian setup. Campbell and Viceira (2002) also concentrate on a dynamic asset
allocation problem based on a conditional forecasting model and resort to an approximate
solution, while Brennan, Schwartz and Lagnado (1997) finds the optimal dynamic portfolio
solving the representative investor’s first order conditions. The last two papers also make use
of predictive regressions for forecasting asset returns. The predictability problem is studied
with Bayesian methodologies by Avramov (2002) as well.
A recent example in the bond literature is Frauendorfer and Schürle (2001) which finds
conflicting results in an application to the Swiss bond market. The authors show that the one-
factor Vasicek (1977) model is preferred to both the one-factor CIR (1985) model and two-
factor models from a statistical standpoint (log likelihood maximization) as well as based on
its financial performance, even though two-factor models seem to better characterize yields’
empirical distribution. Applications of dynamic programming models to bonds have been
provided by Brennan, Schwartz and Lagnado (1997), Brennan and Xia (2000), Campbell
and Viceira (2001) and Walder (2002). In setting and solving the optimization problem
we rely on a scenario generation approach along the path set forth by Zenios and various
coauthors, among which Zenios (1993, 1995), Zenios and Kang (1993), Beltratti, Laurent and
Zenios (2004), Jobst and Zenios (2001a). However our model is explicitly forward-looking
because the investor solves a two-period problem—as described next in this introduction—,
whereas the cited papers consider static portfolio choices. We choose to work with the scenario
generation approach for several reasons, among which are: (1) it accounts for the introduction
of transaction costs and (2) it is flexible in allowing for a practically intuitive minimization
of the expected shortfall, and does not require a tractable utility function of the isoelastic
form—which is much harder to use in applications.
As to the second point above—the importance of predicting both mean excess returns
and risk—, the conflicting results reported above on the opportunity to simultaneously match
means and variances reinforce the idea that it may be interesting to evaluate term structure
models from a portfolio choice perspective. The portfolio optimization setting provides a
natural way to take into account the ability of a forecasting model to predict both expected
returns and a risk indicator. Furthermore it naturally embeds a metric to evaluate the efficiency
of the model, which can be tested for its contribution to the improvement in the objective
function to be optimized.
Finally, for practical applications it is important to characterize the dynamic portfolio
choices suggested by a given model. Investors prefer models which are relatively stable and
well balanced in terms of different asset classes’ holdings. Models requiring strong fluctu-
ations in asset shares are undesirable for various reasons. From a psychological standpoint,
investors may be unwilling to act very aggressively in the short run. Moreover in practice in-
vestors incur transactions costs for portfolio turnover, and there is evidence that in most cases
high turnover generates costs which severely damage portfolio performance. An important
part of our results will therefore be devoted to dynamically characterize portfolio choices
arising from the theoretical models in the presence of transaction costs.
Our empirical application hinges on the following steps. First we select the empirical
version of the term structure model, comparing three alternative specifications involving both
Springer
196 Ann Oper Res (2007) 151:193–222

EA and CA multi-factor models. We then use the selected models to generate scenarios for the
state vector. The scenarios are selected in such a way as to be compatible with various moments
of the distribution function associated with a given model. We determine scenarios such that
the means, variances and covariances at each node match the corresponding moments of
the joint density function for the state variables. At each node we then use the model to
evaluate the prices of discount bonds for any maturity. Given current discount bond prices,
we are therefore able to compute the rate of return on each available bond for any possible
future scenario under consideration. After scenarios have been generated, we write and solve
a two-period optimization problem. The basic time unit is the quarter, so that a two-period
problem corresponds to a six-month horizon. The investor knows that after three months there
will be a chance to modify the chosen asset allocation. We compute the wealth implied by
repeated application3 of our dynamic portfolio optimization problem to scenarios generated
by the different term structure models, taking into account transaction costs. We finally test
the resulting trading strategies to evaluate whether the results are statistically significant.
While we confirm Duffee’s (2002) result about the superiority of EA over CA models, we
show that the financial metric reverses the ranking of the models within the EA class. Duffee
(2002) has shown that a homoskedastic model has a better forecasting performance than a
model which allows for heteroskedasticity of bond returns. On the other hand we find that
the latter model allows for a better use of a dynamic portfolio strategy. The interpretation
is that allowing for time-varying higher moments may more than compensate for lower
forecasting power of the distribution’s mean. Performance tests do not show that the excess
return produced by the model is significantly positive, even though the superiority of the
heteroskedastic EA model is confirmed in terms of average performance. Moreover the
shortfalls produced by the two EA models are in general statistically significant. As for
the motivation to our exercise, on the basis of our results we claim that: (1) the statistical
evaluation of CA and EA models produces different results from those arising from our
financial metric, i.e. the performance of a dynamic bond allocation model which uses as
inputs scenarios generated from the statistical models; (2) the ability to predict not only
excess returns’ means but also their risk is crucial for a risk averse investor and explains the
difference between the statistical and the financial evaluations. Both results confirm that in
many cases, and the term structure of interest rates belongs to these cases, there is no such a
thing as “the best” model. The model to be used to solve a specific problem depends on the
objective of the decision-maker.
After this introduction the plan of the paper is as follows. Section 1 provides a general
introduction to the affine term structure class of models. Section 2 discusses the scenario
generation via lattice methods. Section 3 introduces the portfolio model and the dynamic
optimization while Section 4 presents the results. Section 5 concludes.

1 Term structure modelling within the affine class

In what follows we briefly review the building blocks in the dynamic term structure models
(DTSMs hereafter) literature. The interested reader may refer to Dai and Singleton (2000,
2003), Duffee (2002), Duffie and Kan (1996), and Fisher and Gilles (1996) for a thorough
discussion on both modelling and estimation techniques.

3 The second point of the planning horizon serves as a theoretical reference only, since the investor will never
actually implement the second stage decision.
Springer
Ann Oper Res (2007) 151:193–222 197

1.1 The affine framework

Consider an economy whose state is described by n variables (also called risk factors) Xt =
(X 1,t , . . . , X n,t ) following the diffusion:

dXt = μ (Xt ) dt + σ (Xt ) dWt (1)

where μ(Xt ) is n × 1 vector, σ (Xt ) is n × n matrix and the complete probability space
(, F, P) with the augmented filtration {Ft :t ≥ 0} is generated by n standard Brownian
motions Wt = (W1,t , . . . , Wn,t ) . Using no-arbitrage arguments one can write the risk factor
dynamics in (1) under the equivalent martingale measure Q as:

dXt = μ Q (Xt ) dt + σ (Xt ) dWtQ

Q 
where WtQ = (W1,t Q
, . . . , Wn,t ) is a vector of standard Brownian motions under the risk-
neutral measure Q. The n × 1 vector of market prices of risk Λ (Xt ) allows to move from
the physical probability measure to the risk-neutral one, given that the drift term under
Q is μ (Xt ) =  μ (Xt ) − σ (Xt ) Λ (Xt ). Therefore building a DTSM boils down to specify
Q

rt , μt , Λt , σt as functions of the state vector Xt , where rt denotes the riskless rate.

A DTSM is (exponential-) affine4 if and only if

r (Xt ) , μ (Xt ) , σ (Xt ) σ (Xt ) and σ (Xt ) Λ (Xt ) are affine in Xt (2)

Note that affinity in rt is equivalent to write the riskless rate as:

r (Xt ) = δ0 + δ  X t (3)

where δ0 is scalar and δ is n × 1. The key feature of the (exponential-) affine class is that
bond prices take the form:5

P (Xt , τ ) = exp(A (τ ) − B (τ ) Xt ) (4)

where A (τ ) is scalar, B (τ ) = (B1 (τ ) , . . . , Bn (τ )) , and τ is the bond’s residual time to

maturity. No-arbitrage condition on the price dynamics results in the coefficient A (τ ) and
the factor loadings B (τ ) obeying a PDE with boundary condition P (Xt , 0) = 1:

1
− Ȧ (τ ) + Ḃ (τ ) Xt − r (Xt ) − μ Q (Xt ) B (τ ) + B (τ ) σ (Xt ) σ (Xt ) B (τ ) = 0 (5a)
2
A (0) = 0 B (0) = 0 (5b)

4 Duffie and Kan (1996) do not carry out their analysis neither in terms of μ (Xt ) nor Λ (Xt ); on the other
hand they postulate an affine form for the short rate r (Xt ), the risk adjusted drift μ Q (Xt ) and the diffusion
σ (Xt ) σ (Xt ) . Thus with respect to Duffie and Kan’s (1996) requirement, condition (2) is stronger in that
imposes affinity in σ (Xt ) Λ (Xt ) as well. It follows however that under (2) the risk-adjusted drift μ Q (Xt ) is
affine as well.
5 Equivalently, bond yields yt,τ ≡ Y (Xt , τ ) are affine in the state vector:
log(P(Xt , τ )) 1
Y (Xt , τ ) ≡ − = (−A(τ ) + B(τ ) Xt )
τ τ
Springer
198 Ann Oper Res (2007) 151:193–222

Duffie and Kan (1996) show that within the affine class the diffusion σ (Xt ) has the following
structure:

σ (Xt ) =  S (Xt ) (6)

where  is n × n constant matrix and St is n × n diagonal with typical element:6


[S (Xt )]ii = αi + βi Xt (7)

where αi is scalar and βi is n × 1 vector, i.e. βi = (βi1 , . . . , βin ) . For notational purposes,
it is convenient to stack the αi ’s into the n × 1 vector α and the βi ’s vectors in the n × n
matrix β = (β 1 , . . . , β N ) such that:

S (Xt ) S (Xt ) = diag (α) + diag (βXt )

where diag(α) (respectively diag(βXt )) is a diagonal matrix obtained by inserting the vector
α (the vector βXt ) into the diagonal.
We now turn to specify the function Λ (Xt ), since the representation for the market price
of risk turns out to be the key difference between CA and EA models. In the CA class, the
market price of risk takes the form:

Λ (Xt ) = S (Xt ) λ1 (8)

where λ1 is n × 1 and S (Xt ) is given in equation (7). Therefore the instantaneous n vari-

ance
n of the market price of risk is affine in X t , since Λ (Xt ) Λ (X t ) = i=1 1i i +
λ 2
α

i=1 λ1 diag([β]i )λ1 X i,t .
On the other hand in the EA class the market price of risk is:

Λ (Xt ) = S (Xt ) λ1 + S − (Xt ) λ2 Xt (9)

where λ2 is n × n and S − (Xt ) is n × n diagonal with typical element7 :


− (αi + βi Xt )−1/2 if infXt (αi + βi Xt ) > 0
[S (Xt )]ii =
0 otherwise

Provided that EA models satisfy the affinity condition (2) bond prices are affine (see equation
(4)) and the PDE (5a) still holds together with boundary condition (5b). However the instan-
taneous variance of the market price of risk Λ (Xt ) Λ (Xt ) is not affine in Xt for λ2 = 0.

6 The requirement of a diagonal diffusion matrix is needed for identification of parameters in estimation. Some
affine models cannot be represented via a diagonal diffusion matrix, implying that the normalization in Dai
and Singleton (2000) cannot be imposed. In particular, affine models with more than three factors might not be
characterized by a diagonal diffusion matrix (see Cheridito, Filipović and Kimmel (2003) and the references
therein). In such cases, one should consider the diagonalization requirement (7) as a primitive condition on
the DTSM, rather than one of its properties.
7 The specification (9) is provided by Duffee (2002), and the expression for the S − matrix formalizes the

idea that the market price of risk goes to zero together with the volatility of the corresponding state variable.
However this restriction can be relaxed without impairing no-arbitrage conditions as in Cheridito, Filipović
and Kimmel (2003).
Springer
Ann Oper Res (2007) 151:193–222 199

This feature -arising from the richer specification for the market price of risk in equation (9)-
gives EA models higher flexibility in capturing the time variation in the price of risk.

1.2 Bond pricing within the CA and EA class

Starting from an affine form for the riskless rate as in equation (3) we specialize the affine
condition on the physical drift in (1) as:

μ (Xt ) = (K θ − K Xt ) (10)

where K is n × n and θ is n × 1. Without loss of generality we normalize throughout the

matrix  to be the identity matrix, such that equation (6) yields σ (Xt ) = S
(Xt ). It follows that
the instantaneous variance is σ (Xt )σ (Xt ) = S(Xt )S(Xt ) = diag(α) + i=1 n
diag([β]i )X i,t
and the PDE (5a) becomes:

− Ȧ (τ ) + Ḃ (τ ) Xt − r (Xt ) − (K θ − K Xt − S (Xt )
(Xt )) B (τ )

1 n
1  n

+ B (τ ) αi +
2
B (τ ) βi Xt = 0
2
(11)
2 i=1 i 2 i=1 i

Recall that (11) is valid for CA and EA models. Within the CA class the market price of risk
is given by equation (8). Thus:

S (Xt ) Λ (Xt ) = Ψ + Xt

where Ψ is n × 1 with Ψi = αi λi and  is n × n with typical row i = λi βi . The CA

specification (8) thus results in breaking down the PDE (11) into n + 1 ODEs:

1 n
Ȧ (τ ) = −δ0 − B (τ ) (K θ − Ψ) + B 2 (τ ) αi (12a)
2 i=1 i

 1  n
Ḃ (τ ) = δ − (K + ) B (τ ) − 2
B (τ ) βi (12b)
2 i=1 i

On the other hand in the EA class the market price of risk is defined in equation (9). In this
case:

S (Xt ) Λ (Xt ) = S 2 (Xt ) λ1 + I − λ2 Xt (13)

where I − ≡ S (Xt ) S − (Xt ) is n × n diagonal with typical element:

  
−
1 if infXt αi + βi Xt > 0
[I ]ii =
0 otherwise

Plugging (13) back into (11) gives the ODEs:

1 n
Ȧ (τ ) = −δ0 − B(τ ) (K θ − Ψ) + B 2 (τ )αi (14a)
2 i=1 i
Springer
200 Ann Oper Res (2007) 151:193–222


− 1 n
Ḃ (τ ) = δ − (K +  + I λ2 ) B (τ ) − 2
B (τ ) βi (14b)
2 i=1 i

Note that the EA class nests the CA class since for λ2 = 0 the specification for the market
price of risk (9) reduces to equation (8). It follows that the ODEs (14a–14b) for EA models
coincide with their CA counterpart (12a–12b).

1.3 Forecasting future yields: The first two conditional moments

In the previous section it has been clarified that bond prices at each time t are a function of
the current state vector. On the other hand the portfolio model requires forecasts of (scenarios
of) future prices. In order to predict future bond prices (or equivalently yields), one needs to
forecast the future state vector {XT }T >t conditional on the information available at time t.
Closed form formulas for the first two conditional moments are given in Duffee (2002). The
idea is to find the first two moments of a linear transformation of the state vector Xt . More
specifically, assume that the matrix K in equation (10) can be diagonalized:

K = N D N −1

where D is n × n diagonal with the i-th eigenvalue as the typical element and N is the n × n
matrix with the eigenvector associated with the i-th eigenvalue in its i-th column. Further
consider the transformation:

X∗t = N −1 Xt (15)

Then it can be shown (see Duffee (2002), pages 439–442) that for T > t:

E(X∗T |X∗t ) = θ ∗ + exp(−D(T − t))(X∗t − θ ∗ ) (16a)


n
var(X∗T |X∗t ) = b0 + ∗
bi X i,t (16b)
i=1

where θ ∗ = N −1 θ, and the ( j, k)-th element in the matrices b0 and bi , i = 1, . . . , n is defined

as follows:

1
[b0 ] jk = (1 − exp(−(T − t)([D] j j + [D]kk ))) [G 0 ] jk
[D] j j + [D]kk


n n
∗
+ θi [G i ] jk − [bi ] jk θi∗
i=1 i=1

[G i ] jk
[bi ] jk = (exp (− (T − t) [D]ii )
[D] j j + [D]kk − [D]ii
− exp(−(T − t)([D] j j + [D]kk )))

where G 0 =  ∗ diag(α)  ∗ , G i =  ∗ diag([β ∗ ]i ) ∗ ,  ∗ ≡ N −1  and β ∗ ≡ β N . The

first two conditional moments for the state vectors are obtained from equations (16a) and
Springer
Ann Oper Res (2007) 151:193–222 201

(16b) reversing the transformation (15):

E (XT |Xt ) = N E(X∗T |X∗t ) (17a)

var(XT |Xt ) = N var(X∗T |X∗t )N  (17b)

Once equipped with a DTSM parametrization for: (1) the riskfree rate (see equation (3));
(2) the physical drift (see equation (10)); (3) the volatility matrix (see (6) and (7)) and (4)
the market price of risk—either as in (8) or (9)—, one can use (17a,17b) together with the
appropriate ODEs—either (12a–12b) or (14a–14b)—to compute the first two conditional
moments of future bond prices.

2 Scenario generation via lattice method

For pricing purposes one needs to approximate the diffusion processes for the risk factors.
Lattice methods have been extensively used in the financial literature within derivative asset
pricing. In the aftermath of the binomial tree in Cox, Ross and Rubinstein (1979), several
authors devised lattices as useful tools in pricing claims. The basic strategy consists of
building a grid with sizes of the jumps and associated probabilities obtained equating the
first two moments of the underlying diffusion to those of the approximating distribution.
Nelson and Ramaswamy (1990) use binomial processes in approximating several univariate
distributions. In a multivariate setting one has to ensure the convergence of the approximating
distribution to the true distribution by matching the variables’ covariances as well. Boyle,
Evnine and Gibbs (1989) generalize the binomial lattice in Cox, Ross and Rubinstein (1979)
allowing each state variable to be proxied by a binomial process. Their method can be easily
extended to n variables, yielding processes with 2n jumps. Kamrad and Ritchken (1991)
build on the work in Boyle, Evnine and Gibbs (1989) allowing for a horizontal jump, i.e.
the possibility that the state vector does not move from its current value. In a n-dimensional
setting their model would result in 2n + 1 jumps.
Litterman and Scheinkman (1991) and more recently Chapman and Pearson (2001) have
shown that three factors are able to explain the majority of Treasury bond movements. Like
Ahn, Dittmar and Gallant (2002), Dai and Singleton (2000) and Duffee (2002) we therefore
set n = 3 in specifying the DTSMs for our dynamic portfolio problem. Given that the portfolio
model we implement (see Section 3 for further details) takes expected bond prices for two
future dates as inputs, we need a lattice for the state vector at time t + 1 and t + 2. Similarly
to Boyle, Evnine and Gibbs (1989) we choose a binomial tree in the three dimensions for
the state variables, thus resulting in 8 nodes for the state vector at time t + 1 and 64 nodes at
time t + 2.
Before discussing the three dimensional case, we describe our approximation method
with two state variables. The binomial lattice approximating each state variable {X i }i=1,2 is
represented in Figure 1. The starting point is the observation at time t of values for X 1,t and
X 2,t . At time t + 1 each state variable X i,t+1 can either move up to X i,t+1
u
with probability pi
or down to X i,t+1 with probability 1 − pi (see panel A in Figure 1). In general both the values
d

for the vector Xt+1 and the transition probabilities depend on the current time t, possibly via
the current state Xt . The quantities {X i,t+1
u
, X i,t+1
d
, pi }i=1,2 are chosen in order to match the
first two conditional moments for X 1 and X 2 . This amounts to solve the following system:

u
p1 X 1,t+1 + (1 − p1 ) X 1,t+1
d
= E (Xt+1 |Xt )1 (18a)
Springer
202 Ann Oper Res (2007) 151:193–222

Xiu,t+1 E (Xi,t+1 | Xt ) + var (Xi,t+1 | Xt )

pi 1/ 2

Xi,t Xi,t

1−pi 1 /2

Xid,t+1 E (Xi,t+1| Xt ) − var (Xi,t+1 | Xt )

Panel A Panel B

Fig. 1 Each state variable X i moves along a binomial tree over time. The general form of such a binomial
tree is depicted in Panel A. Panel B reports our specification with equal jump probabilities as described in
Section 2.

u
p2 X 2,t+1 + (1 − p2 ) X 2,t+1
d
= E (Xt+1 |Xt )2 (18b)
 u 2  d 2
p1 X 1,t+1 − E(Xt+1 |Xt )1 + (1 − p1 ) X 1,t+1 − E(Xt+1 |Xt )1 = var (Xt+1 |Xt )11 (18c)
 u 2  d 2
p2 X 2,t+1 − E(Xt+1 |Xt )2 + (1 − p2 ) X 2,t+1 − E(Xt+1 |Xt )2 = var (Xt+1 |Xt )22 (18d)

where E (Xt+1 |Xt )i and var (Xt+1 |Xt )ii are respectively the i-th element of the conditional
expectation vector and the (i, i) element of the conditional variance-covariance matrix. In
order to pin down the values for {X i,t+1
u
, X i,t+1
d
, pi }i=1,2 in the system (18a-18d) we impose
equal transition probabilities, i.e. p1 = p2 = 1/2, thus making the number of unknowns
equal to the number of equations.8 Let X iu be the deviation of the value X i,t+1 u
from the
conditional mean, i.e. X iu ≡ X i,t+1
u
− E (Xt+1 |Xt )i ( X id is defined similarly). Plugging
pi = 1/2 in the system (18a–18d) gives, for i = 1, 2:

X iu = − X id (19a)
 2  2
X iu + X id = 2var (Xt+1 |Xt )ii (19b)

Equation (19a) imposes symmetry around the conditional mean. Substituting  for X i =
X iu into eq. (19b) gives the value for the deviation at time t + 1 as X i = var (Xt+1 |Xt )ii
such that the future values for the state variables become:

u
X i,t+1 = E (Xt+1 |Xt )i + var (Xt+1 |Xt )ii (20a)

d
X i,t+1 = E (Xt+1 |Xt )i − var (Xt+1 |Xt )ii (20b)

for i = 1, 2 (see Figure 1 panel B).

We now turn to describe the tree for the state vector Xt . When both state variables are
approximated through a binomial lattice as in (20a, 20b), the state vector (and thus the price of
zero-coupon bonds) can be proxied by a 4-jump process. Let {X 1,t , X 2,t } denote the values for
the state variables at time t. Table 1 reports the values—together with associated probabilities

8 Our procedure differs from Boyle et al. (1989) since they assume values for Xt+1 that are symmetric
u
around the current value Xt , i.e. X i,t+1 = X i,t + X i and X i,t+1
d
= X i,t − X i , and then solve (18a–18d)
for { pi , X i }i=1,2 . On the other hand our method is closer to He (1990), where a trinomial tree with equal
probabilities for every state is employed in approximating each diffusion for the two-dimensional case.
Springer
Ann Oper Res (2007) 151:193–222 203

Table 1 State vector values and

probabilities (two-dimensional Value for Xt+1 Probability
case) 
u
X 1,t+1 , X 2,t+1
u
π1
Nodes for the state vector are
reported in the first column. Each 
state variable {X i }i=1,2 evolves
u
X 1,t+1 , X 2,t+1
d
π2
along a binomial tree. With two 
state variables the state vector
d
X 1,t+1 , X 2,t+1
u
π3
takes four values at time t + 1. 
The second column reports the d
X 1,t+1 , X 2,t+1
d
π4
associated probabilities.

X 1,u1
X1
var( X 1 | X 0 )11
E ( X 1 | X 0 )1

E (X1 | X 0 )
X 1,0 X 1,d1
X0 π4

t
π3
E (X 1 | X 0 )2
X 2,0
X2

Fig. 2 Each state variable {X i }i=1,2 moves along a binomial tree over time like in Figure 1. The whole state
vector’s dynamics is given by a two-dimensional lattice (values and probabilities are in Table 1). Figure 2
displays the main features of our lattice generation method in Section 2.

π1 , . . . , π4 – that the joint distribution for the state vector can take at time t + 1, while Figure
2 presents a graphical example of our procedure. Given the state vector values at time t + 1
as in (20a, 20b), the probability vector (π1 , . . . , π4 ) has to be chosen in order to replicate the
marginal distributions for the state variables {X 1 , X 2 } and the conditional covariance between
the two processes. Matching the marginal distributions and imposing that the probabilities
sum up to unity amounts 4 to set the system: π1 + π2 = 1/2, π3 + π4 = 1/2, π1 + π3 = 1/2,
π2 + π4 = 1/2 and i=1 πi = 1 which simplifies to:

π1 = π1 (21a)
π2 = 1/2 − π1 (21b)
π3 = 1/2 − π1 (21c)
π4 = π1 (21d)

Matching the conditional covariance requires further that:

π1 − π2 − π3 + π4 = ρ12 (Xt+1 |Xt ) (21e)

Springer
204 Ann Oper Res (2007) 151:193–222

Table 2 State vector values and

probabilities (three-dimensional Value for Xt+1 Probability
case) 
u
X 1,t+1 , X 2,t+1
u
, X 3,t+1
u
π1

u
X 1,t+1 , X 2,t+1
u
, X 3,t+1
d
π2

u
X 1,t+1 , X 2,t+1
d
, X 3,t+1
u
π3

u
X 1,t+1 , X 2,t+1
d
, X 3,t+1
d
π4

Nodes for the state vector are d
X 1,t+1 , X 2,t+1
u
, X 3,t+1
u
π5
reported in the first column. Each 
state variable {X i }i=1,2,3 evolves d
X 1,t+1 , X 2,t+1
u
, X 3,t+1
d
π6
along a binomial tree. With three 
state variables the state vector d
X 1,t+1 , X 2,t+1
d
, X 3,t+1
u
π7
takes eight values at time t + 1. 
The second column reports the d
X 1,t+1 , X 2,t+1
d
, X 3,t+1
d
π8
associated probabilities.

where ρ12 (Xt+1 |Xt ) denotes the conditional correlation between the two variables, i.e. for
i = j
 
ρi j (Xt+1 |Xt ) ≡ [var (Xt+1 |Xt )]i j / X i X j (22)

Thus the comovement between the two state variables allows to pin down one solution for
the probability vector (π1 , . . . , π4 ) in the system (21a–21e):

1
π1 = π4 = (1 + ρ12 (Xt+1 |Xt ))
4
1
π2 = π3 = (1 − ρ12 (Xt+1 |Xt ))
4

Notice that the probabilities {πi }i=14

are well defined, in that they all lie between zero and
unity.
With the two-dimensional case in mind, we now discuss our procedure for the three-factor
setup (the multi-factor DTSMs are described in Subsection 4.2). Each diffusion is proxied
by a binomial lattice as in Figure 1-panel B with equal transition probabilities and up/down
jumps as in equations (20a,20b). Since we are dealing with three state variables, the whole
state vector Xt+1 follows an 8-jump process as summarized in Table 2. Proceeding as before,
the probabilities {πi }i=1
8
are to be chosen in order to match the marginal distributions for the
state variables {X i,t }i=1 and the conditional covariances between the processes X i,t and X j,t ,
3

i = j.
Equivalently to the system (21a–21d) one gets:


8
πi = 1 (23a)
i=1

π1 + π2 + π3 + π4 = 1/2 (23b)
π1 + π2 + π5 + π6 = 1/2 (23c)
π1 + π3 + π5 + π7 = 1/2 (23d)
Springer
Ann Oper Res (2007) 151:193–222 205

Let ρ13 (·) and ρ23 (·) be defined as in eq. (22). Matching the conditional covariances results
in the three additional constraints:

(π1 + π2 − π3 − π4 − π5 − π6 + π7 + π8 ) = ρ12 (Xt+1 |Xt ) (23e)

(π1 − π2 + π3 − π4 − π5 + π6 − π7 + π8 ) = ρ13 (Xt+1 |Xt ) (23f)
(π1 − π2 − π3 + π4 + π5 − π6 − π7 + π8 ) = ρ23 (Xt+1 |Xt ) (23g)

The set of equations (23a–23g) describes a linear system in the probability vector π ≡
(π1 , . . . , π8 ) that admits infinitely many solutions. This indeterminacy arises in the litera-
ture when dealing with approximations for n > 2 variables (see Boyle, Evnine and Gibbs
(1989) and He (1990)). Following Boyle, Evnine and Gibbs (1989) we impose one addi-
tional constraint—suggested by the analysis for the two dimensional case—in order to get a
unique solution. Recall that with two state variables the probabilities of the extreme values
for both the state variables were the same, i.e. π1 = π4 . Thus we constrain the “all-up-node”
{X 1,t+1
u
, X 2,t+1
u
, X 3,t+1
u
} and the “all-down-node” {X 1,t+1
d
, X 2,t+1
d
, X 3,t+1
d
} to be equally likely:

π1 = π8 (23h)

Then the system (23a–23h) admits the following solution:

1
π1 = π8 = (1 + ρ12 + ρ13 + ρ23 )
8
1
π2 = π7 = (1 + ρ12 − ρ13 − ρ23 )
8
1
π3 = π6 = (1 − ρ12 + ρ13 − ρ23 )
8
1
π4 = π5 = (1 − ρ12 − ρ13 + ρ23 )
8

Since in our empirical study we model the state vector’s dynamics from time t up to t + 2,
we repeat the procedure outlined above9 in each of the 8 nodes resulting at time t + 1.

3 Portfolio model

We describe a multi-period portfolio model which takes as inputs the state vector forecasts
produced by the DTSMs introduced in Section 1 (see Subsection 4.2 for further details) via
the lattice procedure outlined in Section 2. Following the structure proposed by Beltratti,
Consiglio and Zenios (1998), investment decisions are measured in terms of dollars of face
value—rather than percentages invested in the various assets. The model describes the choices

9 Notice that even though the solution for the probability vector is now unique, nothing guarantees that the
probability vector lies between zero and one. A sufficient—albeit not necessary—condition for this to happen
is that the correlations among all state variables are all below 1/3 in absolute value. However the approximating
procedure we implement via the system (23a–23h) yields defined probabilities for all the DTSMs we consider
in our empirical application.
Springer
206 Ann Oper Res (2007) 151:193–222

of an investor with a T -period horizon and facing N risky assets and cash at the initial time
t = 0. We use the following notation:
St set of scenarios (nodes) anticipated at time t; st is a generic element of the set St ;
L t set of paths obtained combining scenarios from the sets S0 , S1 , . . . , St−1 (L t ⊆ S0 ×
S1 × · · · × St−1 ); lt is the information structure at time t (lt ∈ L t ) which describes the path
of nodes from time 0 until time t (t × 1 vector). At t = 0 no path is defined because all the
information is available for that period;
πlt probability of path lt ;
Pt (lt ) prices of assets at t, a function of the path between 0 and t (N × 1 vector);
ρt (lt ) one-period risk-free rate at t;
γ transaction costs on sales;
δ transaction costs on purchases;
c0 initial liquidity;
B0 initial portfolio (N × 1 vector);
1
= (1, . . . , 1) the unit vector of dimension (1 × N ).

At each time period the choice variable is denoted by the N × 1 vector Zt (lt ), giving the
quantity to be kept in the portfolio for each asset. Knowledge of Zt (lt ) together with the
initial holdings B0 gives the quantities that should be bought (Qt (lt ), N × 1 vector) or sold
(Yt (lt ), N × 1 vector), and in turn determines the amount of cash (vt (lt )). No short sales are
allowed. Notice that at t = 0 all the prices are known, and there is only one node so that the
control variable does not depend on the path.
At time t = 0 the value resulting from the original cash and from the sale of the assets in
the portfolio must be equal to the amount invested in liquidity and that invested for increasing
other assets (cashflow accounting), so that:

0 = (P0 + δ 1)Q
c0 + (P0 − γ 1)Y
0 + v0 (24)

Moreover each asset must satisfy an inventory balance constraint:

B0i + Q i0 = Y0i + Z 0i , ∀i ∈ N (25)

Decisions made at t = 1, . . . , T depend on the information structure lt and on previous

investment decisions. Similarly to what happens at the initial period t = 0 (see eq. (24)), the
increase in asset holdings must be equal to the income generated by the assets and the value
generated by sales. Therefore one has:

t (lt ) = [Pt (lt ) + δ 1]Q
ρt−1 (lt−1 )vt−1 (lt−1 ) + [Pt (lt ) − γ 1]Y
t (lt ) + vt (lt )
(∀lt ∈ L t )
(26)
Moreover for each path lt the amount of each asset sold or kept in the portfolio must be
equal to the amount bought or held from the previous period (see eq. (25)):

i
Z t−1 (lt−1 ) + Q it (lt ) = Yti (lt ) + Z ti (lt ) (∀i ∈ N , ∀lt ∈ L t ) (27)

We now describe the objective function. Let V Pt (lt ) be the value of the portfolio at time
t = 1, . . . , T computed for each path lt . V Pt (lt ) depends on the portfolio composition, the
Springer
Ann Oper Res (2007) 151:193–222 207

risky assets’ value, and the amount of liquidity at time t as follows:

V Pt (lt ) = vt (lt ) + Pt (lt )Zt (lt ) (∀lt ∈ L t )

Denoting with R P (l T ) the portfolio’s rate of return in the final period T computed for each
path lt , one has:

V PT (l T ) − V P0
R P (l T ) = (∀lt ∈ L t )
V P0

where V P0 ≡ c0 + P0 B0 is the initial portfolio value. Let us define for each t = 1, . . . , T and
for each path the one period shortfall associated with the portfolio held at time t:

eτt V Pt−1 (lt−1 ) − V Pt (lt ) if eτt V Pt−1 (lt−1 ) > V Pt (lt )
S Ft (lt ) = (∀lt ∈ L t )
0 otherwise
(28)
where τt is the minimum required return for the period (t − 1, t] . Consistently with eq. (28),
the opportunity cost to an investor corresponds to the case in which the value of the portfolio
at time t is lower than the value of the portfolio in the preceding period multiplied by the
factor eτt . The required return τt could be thought of as the increase in liabilities’ value or the
competitors’ expected rate of return. Alternative formulations can be easily implemented.
For example Chevalier and Ellison (1997) has shown that the penalty (in terms of assets under
management) fund managers bear when underperforming their competitors is not particularly
sensitive to the underperformance magnitude, while the reward increases strongly with the
overperformance. We could easily model this stylized fact by defining a fixed shortfall in case
of underperformance and a negative and increasing shortfall in case of overperformance. At
time T we define for each path l T the cumulative shortfall:

eτcum V P0 − V PT (l T ) if eτcum V P0 > V PT (l T )
S Fcum (l T ) = (∀lt ∈ L t ) (29)
0 otherwise

where τcum is the minimum return over the whole time horizon. Given cost measures for
shortfalls each period, {kt }t=1,..,T , and for the cumulative shortfall, kcum , the investor wants
to minimize a cost function based on the expected shortfall:
 
min TSF T = k1 πl1 S F1 (l1 ) + · · · + k T πlT S FT (l T )
l1 ∈L 1 l T ∈L T

+ kcum πlT S Fcum (l T ) (30)
l T ∈L T

under the constraints given by obtaining a return equal to the target R̄ at each final node:

R P (l T ) = R̄ (∀lt ∈ L t ) (31)

and by the cashflow accounting and inventory balance constraints (24–27). From the defi-
nition of the cost function it emerges that the higher the parameters τt and τcum , the higher
the shortfall associated with a given portfolio return. Minimizing the objective function (30)
given the constraint (31) amounts to risk minimization given a specific target return. This
Springer
208 Ann Oper Res (2007) 151:193–222

constrained minimization may be interpreted as an alternative to the standard efficient frontier

methodology in the presence of asymmetric risk. It has been used in a one-period optimization
problem by Zenios and Kang (1993), which shows that this approach is a linear program-
ming problem reformulation of the mean-absolute deviation (MAD) model originally due
to Konno and Yamazaki (1991). This objective function has recently been used within a
static optimization setup by Beltratti, Laurent and Zenios (2004) and by Jobst and Zenios
(2001b).
There is a growing interest in the use of the shortfall in portfolio optimization, given recent
attention to value-at-risk measures as well as portfolio optimization with non-normal assets.
Of course non-normal returns do not require the use of the shortfall, and other risk measures
like the standard deviation could be considered instead. However there are reasons to prefer
the expected shortfall. Indeed, the expected shortfall corresponds to the conditional loss and
to the conditional value-at-risk, see Embrechts, Klüppelberg and Mikosch (2003). Jobst and
Zenios (2001b) use the expected shortfall framework in an effort to embed credit risk into asset
management, and compare mean-absolute deviation portfolio optimization with expected
shortfall portfolio optimization. Bogentoft, Romeijn and Uryasev (2001) utilize the expected
shortfall for pension funds’ asset/liability management, while Topaloglou, Vladimirou and
Zenios (2002) employ it for an international portfolio choice problem claiming that “it can
be used to exercise some control on the lower tail of the return distribution and thus, it
is a suitable risk measure for skewed distributions” (Topaloglou, Vladimirou and Zenios
(2002), p. 6). Moreover, in practical applications, portfolio managers may be more sensitive
to risk measures approximating the expected amount of the loss, rather than to a measure
like the standard deviation which is independent of the portfolio value and the potential
loss.
Our analysis is therefore closer to the computation of (a sequence of) specific points on
a shortfall-expected return efficient frontier, rather than to the computation of (a sequence
of) optimal portfolios for a specific utility function. Of course we could introduce a param-
eter representing the relative weights of the target return and the expected shortfall in the
investor’s utility, and choose one optimal portfolio for each time period. Instead, we compute
a portfolio for each of the five target returns we consider in the empirical application (see
Subsection 4.3). Therefore our findings are not directly comparable to recent contributions
like Wachter (2003)—showing that increasing risk aversion forces the investor to buy the
bond with the same maturity as the consumption horizon—, because we are not solving an op-
timal consumption-investment problem. Also, our results may be more relevant to short-run
investors rather than long-run investors, as the latter are concerned with building portfolios
hedging the state variables and may deviate from the simple static efficient portfolios (see
Sangvinatsos and Wachter (2005) for a recent application to bond portfolios).
It is also important to compare EA and CA models in terms of their ability to form
efficient portfolios because in such a way a weak link with explicit utility maximization
is established. It is well known that different affine DTSMs are compatible with different
assumptions on the price of risk,10 and therefore with different representative investor’s utility
functions supporting the prices in general equilibrium. We are therefore minimizing the risk
of preferring one model to the other as the result of an unintended similarity between the
implicit utility function embedded in each DTSM and the explicit utility function evaluated
in our portfolio optimization model.

10 See Duffee (2002) for a comparison of the hypotheses maintained in CA and EA models.
Springer
Ann Oper Res (2007) 151:193–222 209

Table 3 Data set summary statistics

Maturity 3M 6M 1Y 2Y 5Y 10Y

Mean excess return, 1952–1998 0.6 0.8 0.9 0.9 2.9 −0.6
Mean excess return, 1952–1991 0.6 0.9 0.9 0.5 1.2 −1.9
Standard deviation, 1952–1998 0.5 0.9 1.7 3.1 6.3 9.8
Standard deviation, 1952–1991 0.4 0.8 1.6 2.9 6.2 9.6
Sharpe ratio, 1952–1998 0.36 0.29 0.16 0.08 0.13 0.02
Skewness, 1952–1998 1.7 1.9 0.7 −0.1 −0.5 0.3
Kurtosis, 1952–1998 11.5 16.0 12.0 9.1 5.8 2.8
Average yield, 1952–1998 5.46 5.69 5.90 6.14 6.46 6.67
Average yield, 1952–1991 5.64 5.88 6.08 6.27 6.53 6.68

The data set consists of month-end yields on zero-coupon U.S. Treasury bonds (available from
Duffee’s homepage at http://faculty.haas.berkeley.edu/duffee/affine.htm) over the sample
period January 1952-December 1998 for selected maturities (first row). For each maturity
we use continuously compounded excess returns over the overnight rate and report their
mean (second and third row), standard deviation (fourth and fifth row), Sharpe ratio (sixth
row), skewness (seventh row) and kurtosis (eight row). Average yields are reported in the
last two rows. All numbers are in percentage per year. Some of the statistics are computed
over the subsample January 1952-December 1991 for comparison with previous studies.

4 Data, implementation and results

4.1 Data

Our analysis is based on data from McCulloch and Kwon (1993) and Bliss (1997) available
from Duffee’s homepage at http://faculty.haas.berkeley.edu/duffee/affine.htm. The dataset
consists of month-end yields on zero-coupon US Treasury bonds with maturities of 3 and 6
months and 1, 2, 5, and 10 years for the sample period January 1952-December 1998. This
dataset is the standard for many studies of the bond market, see for example Duffee (2002).
We have obtained prices11 for other maturities by interpolating with the method proposed by
Nelson and Siegel (1987). Table 3 contains summary statistics about excess returns12 data
used in this study.
Between 1952 and 1998 the average monthly excess return increases with the bond matu-
rity, except for the 10 years bond.13 A similar result also holds for the subperiod 1952–1992
used in several studies by other researchers, who also report a similar finding (see for example
Campbell (2000)). Returns’ standard deviation is also increasing, but much more sharply.
As a consequence the Sharpe ratio is decreasing across maturities. Ilmanen (1996) reports
similar results over the period 1970–1994. Table 3 moreover shows that all excess returns are
characterized by non-normal distributions, especially at the short end of the term structure.
Non-normality justifies the use of the expected shortfall in our dynamic portfolio optimiza-
tion as discussed in Section 3. Finally Table 3 reports the average yield curve in the two
samples, displaying the standard concavity and upward slope. It is of course possible that

11 More specifically, we use the original data from McCulloch and Kwon (1993) over the sample period
January 1952-February 1992, which includes 29 maturities ranging from 1 month to 10 years. We interpolate
yields from Duffee’s homepage for the remaining period (March 1992–December 1998).
12 Excess returns are continuosly compounded with respect to the overnight rate.
13 The average rate of return on the ten year bond is positive, even though it is lower than the average riskless
rate.
Springer
210 Ann Oper Res (2007) 151:193–222

expected bond returns are larger than historical returns, particularly at the long end of the
term structure, due to unexpected inflation episodes during the 1970s and the 1980s—a point
made by Campbell (2000). However measurement error issues about expected bond returns
go beyond the scope of the paper, which is concerned with comparing alternative DTSMs
from a dynamic portfolio choice standpoint. It is instead crucial that the data set we use here
corresponds to the one employed for econometric estimation of these models. Morever, the
returns we report in the table are average values. What is relevant for our dynamic portfo-
lio choice model is the conditional expectation associated with the estimated econometric
models. The sample average is therefore not very relevant to our results.

4.2 Three-factor DTSMs specification

We follow Dai and Singleton (2000) and Duffee (2002), and define a C Am (n) model as a
completely affine model with n state variables of which m affect the instantaneous variance of
the state vector (similarly, E Am (n) defines an essentially affine model with n state variables
of which m affect the instantaneous variance of the state vector). Setting n = 3 the general
specification of three-factor models is given by:

rt = δ0 + δ1 X 1t + δ2 X 2,t + δ3 X 3,t
⎛ ⎞ ⎡⎛ ⎞ ⎛ ⎞⎛ ⎞⎤
X 1,t (K θ)1 k11 k12 k13 X 1,t
⎜ ⎟ ⎢⎜ ⎟ ⎜ ⎟⎜ ⎟⎥
d ⎝ X 2,t ⎠ = ⎣⎝ (K θ)2 ⎠ − ⎝ k21 k22 k23 ⎠ ⎝ X 2,t ⎠⎦ dt + S (Xt ) dWt
X 3,t (K θ)3 k31 k32 k33 X 3,t
 ⎛ ⎞

 X 1,t (32)
  ⎜ ⎟
[S (Xt )]ii = 
αi + βi1 βi2 βi3 ⎝ X 2,t ⎠
X 3,t
⎛ ⎞ ⎛ ⎞⎛ ⎞
λ1(1) λ2(11) λ2(12) λ2(13) X 1,t
⎜ ⎟ ⎜ ⎟⎜ ⎟
Λ (Xt ) = S (Xt ) ⎝ λ1(2) ⎠ + S − (Xt ) ⎝ λ2(21) λ2(22) λ2(23) ⎠ ⎝ X 2,t ⎠
λ1(3) λ2(31) λ2(32) λ2(33) X 3,t

When bringing the DTSM in (32) to the data, restrictions on the parameters are needed in
order to get an admissible specification, i.e. to yield positive conditional variances [S (Xt )]ii
(see Dai and Singleton (2000)). Further to these restrictions, Duffee (2002) reports preferred
specifications resulting from a new estimation of the model (32) after setting to zero all the pa-
rameters with absolute t-statistic below one. We consider the three preferred models estimated
in Duffee (2002) and use his estimated parameters. We refer the reader to Duffee (2002) for a
description of the estimation methodology and to his homepage for the parameters’ estimates.
The first model considered by Duffee (2002) is E A0 (3) and rules out heteroskedasticity
in the state variables.14 The second model is E A1 (3) and leaves some margin to capture

14For E A0 (3) the model restrictions (see Table 4) allow to write the first two conditional moments (see
(16a,16b)) as E(X∗T |X∗t ) = exp(−D(T − t))X∗t and var(X∗T |X∗t ) = b0 respectively, where the typical ele-
ment of the variance-covariance matrix is [b0 ] jk = [D] j j +[D]
1
kk
(1 − exp(−(T − t)([D] j j + [D]kk )))[G 0 ] jk
and G 0 =  ∗  ∗ . Therefore the conditional variance-covariance matrix does not depend on the current state
vector Xt under this specification.
Springer
Ann Oper Res (2007) 151:193–222 211

Table 4 Three-factor DTSMs parameterization

E A0 (3) E A1 (3) C A2 (3)

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
k11 0 0 k11 0 0 k11 k21 0
K ⎣ 0 k22 0 ⎦ ⎣k21 k22 k23 ⎦ ⎣k21 k22 0 ⎦
k31 0 k33 0 0 k33 k31 k32 k33
     
θ 0 0 0 θ1 0 0 θ1 θ2 0
     
α 1 1 1 0 1 1 0 0 1
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
0 0 0 1 0 0 1 0 0
β ⎣0 0 0⎦ ⎣β21 0 0⎦ ⎣0 1 0⎦
0 0 0 β31 0 0 0 1 0
      
λ1 λ1(1) λ1(2) λ1(3) λ1(1) λ1(2) 0 λ1(1) 0 λ1(3)
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
0 λ2(12) 0 0 0 0 0 0 0
λ2 ⎣ 0 λ2(22) 0 ⎦ ⎣λ2(21) 0 λ2(23) ⎦ ⎣0 0 0⎦
λ2(31) λ2(32) λ2(33) 0 0 λ2(33) 0 0 0
 
[S (Xt )]11 1 X 1,t X 1,t
 
[S (Xt )]22 1 1 + β21 X 1,t X 2,t
 
[S (Xt )]33 1 1 + β31 X 1,t 1 + X 1,t

Relevant parameters for admissible DTSMs are reported. The general specification for
three-factor DTSMs is given in (32). Restrictions on parameters are needed in order
to guarantee admissibility as in Dai and Singleton (2000); boldface numbers denote
parameters constrained in Duffee’s (2002) preferred specifications.

heteroskedasticity by means of one state variable. The third model, C A2 (3), is the richest
among those we consider in modeling the dynamics of volatility of interest rates. Table 4
reports the canonical representation as well as the restrictions imposed by Duffee (2002) for
each DTSM.

4.3 Dynamic portfolio model

In our application we consider an investor with a two-period horizon. Technically, our model
belongs to the class of two-stage stochastic programming problems, see Golub, Holmer
and McKendall (1995). Let st = (t, j) denote the scenario at time t and node j. Given
that scenarios are generated by binomial trees for the three state variables as described in
Section 2, we have S0 ≡ s0 = (0, 0), S1 = {(1, j)} j=1,...,8 , S2 = {(2, j)} j=1,...,64 . We think
of each period as one quarter. One quarter and two quarter rates are used to determine the
riskless rate respectively for the first and the second period. The risky portfolio is composed
of N = 39 assets (indexed by subscript i) corresponding to zero-coupon bonds with ma-
turity larger than three months, i.e. i = 1 denotes the six-month bond, i = 2 denotes the
nine-month bond and so on until the ten-year bond (i = 39). The asset prices are obtained
via the DTSMs in Subsection 4.2 solving the ODEs (either (12a–12b) or (14a–14b)) which
are themselves a function of the state variables obtained through the discrete approxima-
tion in Section 2.15 Therefore P0 = {P(0, 0, i)}i=1,...,39 , P1 (s0 , (1, 1)) = {P(1, 1, i)}i=1,...,39 ,
P1 (s0 , (1, 2)) = {P(1, 2, i)}i=1,...,39 and so on.

15It may be useful to notice that the probabilities determined in Section 2 are to be interpreted as true
probabilities, not as risk neutral probabilities. We do not need risk neutral probabilities because pricing is
Springer
212 Ann Oper Res (2007) 151:193–222

The other parameters characterizing our asset markets are chosen in the following way:
γ = δ = 5 basis points16 for transaction costs, and initial cash equal to 100. Within repeated
application of the dynamic portfolio problem to each quarter during our sample period,
we choose time-dependent parameters for the cost function (30) and the expected return
constraint (31). We set the target return in eq. (31) equal to (various multiples of) the average
yield on the current term structure (denoted by R̂t ). This choice describes an investor with
the following characteristics: (1) he expects yields to be constant, so that expected returns
corresponds to the yields currently obtained on the bonds and (2) he wants to achieve at
least—by using the quantitative model—a return related to the average yield conditions in
the bond market for each quarter. Given that the term structure is on average concave and
upward sloping (see Ilmanen (1996)) this is equivalent to setting a target return which is in
line with the short end of the term structure.
In the empirical analysis we repeat the computations for five different target return levels:
0.5 R̂t , 0.6 R̂t , 0.7 R̂t , 0.8 R̂t and R̂t . 17 This way we are able to compare portfolios for individ-
uals with different risk-aversion parameters. Of course when actually solving the portfolio
problem, we do not have any guarantee that the optimizer is able to find a solution under all
circumstances. In order to clarify this issue, assume that at a given time t the average yield,
i.e. across the term structure at the same time, gives R̂t = 5% and we choose the multiple
0.5. Thus the target return is equal to 2.5%. If the optimizer is not able to find any portfolio
achieving the 2.5% target, we decrease the target step by step until a solution is achieved. It
is worth noting that the occurrence of such target correction is quite rare for multiples up to
R̄t / R̂t = 0.8, while we have to adjust the target quite often when we use R̄t = R̂t .18
Finally, for each model we set τ1 = τ2 = ln( R̂t ), τcum = 2 ln( R̂t ) and k1 = k2 = kcum = 1.
The first choice reflects the hypothesis that the minimum required return corresponds to the
average yield of the term structure, while the second is a normalization aimed at describing
a case where all the shortfalls are equally important for the decision-maker.

4.4 Results

Duffee (2002) contains figures for the instantaneous expected excess returns on the 10-year
bond predicted by the three models described in Subsection 4.2. These pictures are useful
to motivate our study from an empirical point of view (see Duffee (2002), Figures 1–3).
They show very clearly how large is the difference in expected returns on holding long term

carried out by the theoretical DTSM. We can therefore use the model to properly study a portfolio problem
that has to be applied to the real world.
16 Over the period 1972–1997, Driessen, Melenberg and Nijman (2005) report a 1.6 basis point average
transaction cost (measured by the bid-ask spread midpoint) for Treasury bills with maturities from 1 month to
9 months. Transaction costs increase with the maturity from 0.6 b.p. for the 1 month T-bill up to approximately
3 b.p. for the 9-month maturity. Given that: (1) bonds are more volatile than short-maturity T-bills and (2) the
T-bill bid-ask spread increases with the time to maturity, it is reasonable to assume that transactions costs on
long-maturity bonds are higher than the average 1.6 b.p. T-bill bid-ask spread. Further, since our sample dates
back to 1952 and comprises U.S. Treasury bonds with maturity up to 10 years, it makes sense to assume 5 b.p
transaction costs.
17 The choice of these parameters is motivated by the fact that values lower than 0.5 resulted in portfolio almost
entirely invested in cash with very low asset turnover, thus making the asset allocation exercise uninteresting.
18 For example C A (3) does not reach the target return in six cases with the 0.8 multiple. We therefore
2
decrease the target to 0.75 R̂t (3 asset allocations), 0.7 R̂t (2 asset allocations) and 0.6 R̂t (1 asset allocation).
The other DTSMs are characterized by similar corrections. When using C A2 (3) and R̂t as target return, we
have to intervene on the target 24 times and decrease to values as low as 0.65 R̂t .
Springer
Ann Oper Res (2007) 151:193–222 213

bonds between the two DTSMs classes. The CA model generates positive excess returns
over the full sample, while the models belonging to the EA class—particularly E A0 (3)—
produce returns which fluctuate between positive and negative. Positive risk premia seem to
be more compatible with potential general equilibrium explanations of asset returns—arising
in equilibrium assuming risk aversion on the part of investor. However a model predicting
both positive and negative expected excess returns could be very useful as an input to active
asset management if the signals about future returns are the right ones. It is useful to point
out for completeness that similar figures for other maturities do not show such a clear cut
outcome, in that also C A2 (3) generates negative expected returns from time to time. This
implies that C A2 (3) may as well suggest time-varying dynamic asset allocation decisions
depending on the state variables’ values.
The three models also differ with respect to their implications for volatility: simply re-
call from Subsection 4.2 that E A0 (3) is homoskedastic and predicts constant conditional
volatility, while both E A1 (3) and C A2 (3) allow for time varying conditional volatility. This
evidence is therefore supportive towards our effort in this study: comparing models as inputs
for asset allocation is particularly interesting when the candidates are characterized by similar
statistical performance, but produce very different outputs in terms of financial choices. In
this case the relevant outputs are the conditional predictive density functions, characterized
for example in terms of the conditional mean and variance—which we have shown differ
across models. Under such conditions a financial decision-maker would be unsure about
which of the three DTSMs is best from a statistical point of view, and at the same time realize
that following one or another would produce different portfolio structures over time.
Table 5 reports descriptive statistics for the state variables19 for the three DTSMs in
Subsection 4.2. Note that E A0 (3) does not impose any restriction on the state vector’s sign
since the β’s are all set equal to zero. Thus it is not surprising that the state variables take
negative values. For the matrix S (Xt ) of conditional second moments to be well defined in
model E A1 (3) (respectively E A2 (3)) one needs X 1 (respectively X 1 and X 2 ) to be positive
everywhere. Table 5 (panels B and C respectively) shows that this is actually the case.
The interpretation of the three factors follows standard analyses, see for example Bühler
and Zimmermann (1996). For each DTSM we have run a linear regression of the yields (on
relevant maturities) on the state vector, and Figure 3 displays the resulting slope coefficients
for C A2 (3). The same qualitative behavior arises from the other two E A models and we do
not include the relevant figures for reasons of space. Figure 3 confirms the typical result that
the three state variables may be interpreted as a level, slope and concavity shock to the term
structure of interest rates.
Table 6 reports summary statistics for the total portfolio shortfall for the three models.20
The numbers in the table refer to the minimized value of the ex-ante total shortfall. This
means that the shortfall is computed based on the asset prices considered in the generated
scenarios rather than on the realized prices. Suppose for example that at time t the optimizer
selects—based on scenarios stemming from a given DTSM—portfolio Zt (lt ) for the path

19 For a given DTSM parametrization, we invert the relationship between yields and the state vector (derived
from Equation (4)). As in Duffee (2002) we consider bonds with maturities 6M, 2Y and 10Y as measured
without error.
20 In order to evaluate the robustness of our results to the structure of the lattice for the state variables we have
compared the shortfall obtained from our 8 scenario lattice to the shortfall obtained from a finer grid involving
512 scenarios, i.e. 8 values for the state vector every month, over 30 randomly chosen dates. The results show
that the mean shortfall obtained by the fine lattice is not statistically different from the mean shortfall obtained
from the coarse lattice.
Springer
214 Ann Oper Res (2007) 151:193–222

Table 5 State variables

descriptive statistics Panel A: E A0 (3)
State variable Mean Std. dev Min Max

X1 0.180 0.755 −1.945 3.106

X2 0.004 0.351 −1.388 1.402
X3 0.486 2.579 −3.592 7.687
Panel B: E A1 (3)
State variable Mean Std. dev Min Max

X1 22.694 12.550 2.881 56.364

X2 16.609 17.093 −14.296 88.144
X3 −0.663 1.074 −5.025 3.168
Panel C: C A2 (3)
The table reports mean (first
column), standard deviation State variable Mean Std. dev Min Max
(second column), minimum and
maximum (third and fourth X1 19.308 9.757 3.768 43.380
columns) for several three-factor X2 9.616 5.202 0.358 26.911
DTSMs. The models are X3 1.318 2.987 −4.056 14.825
described in Section 4.2.

lt and portfolios Zt (lt+1 ) for each lt+1 . We can compute at each time t + 1 and t + 2 the
portfolio’s expected shortfall considering whether the portfolio value—given respectively by
V Pt+1 (lt+1 ) and V Pt+2 (lt+2 )—lies below the minimum required return (see (28, 29)). The total
shortfall, i.e. the minimized value for TSF t+2 in eq. (30), may differ from the one we would
get for the actual ex-post shortfall because: (1) actual bond prices at time t, t + 1 and t + 2
may be different from the ones implied by the DTSM and (2) one can only compute one
quarterly ex-post shortfall—more precisely the ex-post shortfall associated with the asset
allocation Zt (lt ), but not the one arising from the portfolio choice Zt+1 (lt+1 )—since the asset
allocation at time t + 1 is not implemented. Differences across DTSMs are entirely due to
differences in predictive density functions because all the three models are used by the same
decision maker characterized by model-invariant parameters k1 , k2 , kcum , τ1 , τ2 and τcum . Of
course this comparison is preliminary to an analysis of the actual (ex-post) shortfall, which we
report in Table 8. The model with the most cautious approach would otherwise be selected by
the ex-ante shortfall approach, regardless of its true relevance for dynamic portfolio choices.
However we emphasize that the ex-ante analysis is useful to characterize the DTSMs in terms
of their relative attitude towards risk. Moreover, a comparison of the relative merits of the
three DTSMs in both the ex-ante and ex-post shortfalls may be informative to the extent that
such rankings do not change when moving from ex-ante to ex-post considerations. Analyzing
how the results change with modifications in the inputs, i.e. the prices, represents an important
form of robustness check.
Table 6 shows some interesting results. First, models E A1 (3) and C A2 (3) present a
positive relation between the average shortfall and the target return. This is consistent with
the efficient markets hypothesis view that greater expected return can be obtained only by
bearing more risk. The same does not strictly happen to E A0 (3). The most relevant result
in Table 6 from the point of view of our study is that—contrary to the forecasting results
in Duffee (2002)—E A0 (3) is not the best model in terms of shortfall minimization for each
target return. The best model is E A1 (3), since it always produces the lowest average total
shortfall but for the highest target return. The same ranking among the three models obtains
if one looks at the shortfall for the first period only (see Table 7). Results for the second
Springer
Ann Oper Res (2007) 151:193–222 215

1.5

0.5

0
0 2 4 6 8 10
-0.5

-1

slope concavity level

Fig. 3 For the model C A2 (3) the yields of the bonds with the 12 maturities included in the dataset have been
regressed on the estimated state variables and the resulting coefficients have been plotted. Each line in the
figure reports the 12 estimated coefficients of the yields of the 12 maturities (from one month to ten year) on
each of the three state variables. The lines have been defined as “slope,” “concavity” and “level” to interpret
the results. “Slope” is a negatively inclined function interpreted as a shock affecting yields of the bonds with
different maturities in a way negatively related to maturity. “Concavity” reports a function with a weak effect
on both short and long yields and a strong effects on yields of intermediate maturities. “Level” represents a
shock that affects in a similar way all yields.

Table 6 Ex ante shortfall

analysis (total shortfall) Panel A: E A0 (3)
R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
Mean 4.131 4.116 4.337 4.075 3.530
Std dev 2.085 2.298 2.862 3.054 4.654
Min 0.861 0.750 0.624 0.415 0.033
Max 9.477 9.341 9.658 8.982 20.639
Panel B: E A1 (3)
R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
Mean 3.107 3.363 3.665 3.762 5.029
Std dev 1.257 1.660 2.213 3.021 10.738
For each DTSM and
Min 0.859 0.818 0.621 0.413 0.030
risk-aversion parameter (given by
the ratio R̄t / R̂t in the first row) Max 6.462 7.167 7.750 15.672 66.609
the mean (second row), standard Panel C: C A2 (3)
deviation (third row), minimum
(fourth row) and maximum (fifth R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
row) of the total ex-ante shortfall Mean 3.308 3.526 3.721 4.285 13.249
is reported. The shortfall is Std dev 1.403 1.743 2.172 3.818 20.588
computed on the basis of the Min 0.860 0.830 0.621 0.413 0.026
theoretical bond prices produced Max 7.206 7.597 7.816 20.329 88.859
by the various models.

Springer
216 Ann Oper Res (2007) 151:193–222

Table 7 Ex ante shortfall

analysis (first quarter shortfall) Panel A: E A0 (3)
R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
Mean 1.929 1.885 1.919 1.710 1.201
Std dev 1.016 1.092 1.306 1.294 1.457
Min 0.299 0.269 0.251 0.168 0.0
Max 4.733 4.687 4.879 4.414 5.940
Panel B: E A1 (3)
R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
Mean 1.453 1.540 1.622 1.530 1.812
Std dev 0.636 0.807 1.016 1.116 3.674
For each DTSM and
Min 0.273 0.248 0.222 0.158 0.0
risk-aversion parameter (given by
the ratio R̄t / R̂t in the first row) Max 3.240 3.611 3.937 3.757 21.691
the mean (second row), standard Panel C: C A2 (3)
deviation (third row), minimum
(fourth row) and maximum (fifth R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
row) of the first quarter ex-ante Mean 1.539 1.602 1.627 1.747 4.834
shortfall is reported. The shortfall Std dev 0.707 0.848 0.980 1.419 7.039
is computed on the basis of the Min 0.258 0.300 0.261 0.156 0.0
theoretical bond prices produced Max 3.516 3.819 3.901 7.015 29.171
by the various models.

period shortfall and the cumulative shortfall—not shown for reasons of space—are similar.
This is important to evaluate the robustness of our results, since the ordering does not depend
only on one of the three shortfall measures included in our objective function. Tables 6 and
7 show the relevance in predicting accurately both the first and the higher order moments of
the state vector distribution for portfolio allocation purposes. Not only E A1 (3) produces the
lowest average shortfall, but it is also characterized by the least volatile shortfall in three out
of five cases. E A0 (3) is generally the worse model.21
Descriptive statistics for the ex-post, i.e. computed on the basis of actual market prices,
quarterly shortfall are reported in Table 8. These results are more relevant for practical
applications given that the investor relying to a DTSM has to evaluate its performance on the
basis of actual measures, i.e. on market prices and not on theoretical prices. Moreover for
dynamic optimization models this exercise is even more important than in other cases, since
the actual implementation of the portfolio strategy only regards the first of the two periods
included in the planning horizon. As usual, when coming to the second period, the model is
re-set and the new policy for the first of the two planning periods is implemented in place of
the second period policy. In other words the optimal policy is actually implemented only for
one quarter. Table 8 confirms the ex-ante analysis. E A0 (3)—the best model in terms of point
forecast of returns—is not the best one in terms of driving a dynamic asset allocation policy

21 We have performed a two-sided test on the difference in mean shortfalls (both total and first quarter) across
DTSMs. From the results—not reported for reasons of space but available from the authors upon request-, it
emerges that the difference in the ex-ante shortfall across models is significant in most cases, both for the total
and the first quarter shortfall. In particular the shortfall for E A0 (3) is always statistically different at usual
significance levels from the other two DTSMs in all cases but for the multiple 0.8. On the other hand, when
comparing E A1 (3) with C A2 (3) , the test for the equality of average ex-ante shortfalls does not reject the
null hypothesis in all cases but one (the unit multiple case). Taken together with Tables 6 and 7 this means that
the merits of E A1 (3) relative to E A0 (3) are significant for values of R̄t / R̂t up to (and including) 0.7, and
the performance of C A2 (3) is very close to E A1 (3) over the same range. On the other hand, E A0 (3) is the
best model whenever R̄t / R̂t = 1 while for R̄t / R̂t = 0.8 the ex-ante shortfall of the three DTSMs is almost
indistinguishable.
Springer
Ann Oper Res (2007) 151:193–222 217

Table 8 Ex post shortfall

analysis Panel A: E A0 (3)
R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
Mean 1.715 1.739 1.727 1.554 0.971
Std dev 1.265 1.349 1.511 1.589 2.285
Min 0.0 0.0 0.0 0.0 0.0
Max 9.838 9.518 9.536 9.207 21.229
Panel B: E A1 (3)
R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
Mean 1.529 1.627 1.653 1.554 1.406
Std dev 1.046 1.204 1.412 2.328 5.830
Min 0.0 0.0 0.0 0.0 0.0
For each DTSM and
risk-aversion parameter (given by Max 7.60 8.342 8.952 8.363 74.611
the ratio R̄t / R̂t in the first row) Panel C: C A2 (3)
the mean (second row), standard
deviation (third row), minimum R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
(fourth row) and maximum (fifth Mean 1.576 1.649 1.647 1.704 3.434
row) of the quarterly ex-post Std dev 1.164 1.298 1.428 2.478 14.836
shortfall is reported. The shortfall Min 0.0 0.0 0.0 0.0 0.0
is computed on the basis of the Max 9.235 9.681 9.355 28.260 180.068
actual bond prices.

which requires forecasting time-varying higher moments. In fact, this latter task cannot be
achieved by a homoskedastic model like E A0 (3). This illustrates in very simple terms the
main point of our paper.
Another relevant result here is that the ex-post shortfall is very close to its ex-ante coun-
terpart for all the models and this confirms the statistical validity of the three DTSMs at
the quarterly horizon. This is not surprising as the three specifications have been selected
as the best ones by Duffee (2002) from a wide set of potential candidates and are therefore
guaranteed to track the data well. An investor using the DTSMs we consider here can be
confident that ex-ante and ex-post measures generally coincide.22 Recall that the higher the
target return, the more unlikely is the optimizer to achieve the threshold R̄t (see footnote
18). As such we regard cases with a relatively low target return as the most interesting ones,
in that they guarantee a fair comparison across different DTSMs. In general E A1 (3) is still
the best model because it produces the lowest and least volatile shortfall in three cases, thus
confirming our ex-ante analysis.23 Perhaps it is not surprising to find that E A0 (3) performs
better for a high target return. In fact, large values for R̄t require more emphasis on the ability
to predict mean returns rather than risk, and this is exactly what model E A0 (3) achieves in
the statistical analyses of Duffee (2002). The possibility that the ranking of DTSMs depends
on the target return does not go against our general point, but actually reinforces it: a model
may be good or bad depending on the target return which the investor wants to achieve.

22 The test statistic for the equality of means between the ex-ante (first quarter) and ex-post shortfall does
not reject the null hypothesis at usual significance levels in all but one case (the average ex-post shortfall is
statistically different from the ex-ante shortfall at 10% statistical level for E A0 (3) and R̄t / R̂t = 0.5).
23 It is worth stressing that we are simulating the choice of an investor who is comparing the three DTSMs
from the point of view of their in-sample performance. In our view the investor would be happy to choose
the model producing the most efficient portfolio performance. We are not making the alternative exercise (see
Pastor (2000)) of assuming that one model is true and compute the utility cost for an investor using another
model. On the other hand we are simply claiming that an investor would choose the model, among the three
considered here, that achieves the target return with the lowest average shortfall.
Springer
218 Ann Oper Res (2007) 151:193–222

Table 9 Portfolio allocation

Panel A: E A0 (3)
R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
87.02 87.35 88.51 86.99 59.33
Cash
29.41 28.81 25.36 23.98 30.26
6.89 8.17 8.64 10.42 35.27
Short
20.70 23.23 22.95 23.69 28.22
3.22 2.57 1.31 1.14 3.42
Med
13.59 11.11 6.41 5.31 7.96
2.87 1.91 1.54 1.45 1.98
Long
9.13 6.78 4.96 3.80 3.85
Panel B: E A1 (3)
R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
82.27 82.41 81.99 77.73 52.32
Cash
34.53 33.90 33.33 34.15 29.82
8.19 11.16 13.60 17.48 25.84
Short
23.28 27.26 30.49 33.04 21.57
Descriptive statistics for the asset 7.06 4.48 2.61 3.28 16.80
Med
allocation implemented by the 19.94 12.92 8.08 9.86 18.29
DTSMs. For each DTSM and risk 2.47 1.94 1.79 1.51 5.04
Long
aversion parameter (given by the 7.16 5.69 4.55 3.52 16.81
ratio R̄t / R̂t in the first row)
portfolio weights are reported: Panel C: C A2 (3)
average weight (first row), and R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
standard deviation (second row).
81.13 80.58 76.78 69.27 24.15
The asset classes other than cash Cash
are defined as short-term 34.46 33.57 36.17 38.04 30.58
maturities (between 6 months and 4.44 7.88 13.18 20.10 31.88
Short
2 years), medium-term maturities 13.58 20.21 28.45 35.13 31.61
(between 27 months and 7 years) 12.10 8.78 4.90 2.06 1.04
Med
and long-term maturities 27.57 20.95 13.03 6.83 4.92
(between 87 months and 2.34 2.77 5.13 8.56 42.93
10 years). All numbers are in Long
6.24 8.19 16.39 21.35 36.20
percentage.

Table 9 reports summary statistics for average percentage portfolio weights, splitting the
risky assets in short-term maturities (between 6 months and 2 years), medium-term maturities
(between 27 months and 7 years) and long-term maturities (between 87 months and 10 years).
The goal here is to see whether it is possible to understand the dynamic portfolio policy in a
simple way. The results show that C A2 (3) has a tendency to decrease the share of cash for
higher target returns and to increase investment in short term bonds and long term bonds, with
a relative underweight of the medium term part of the term structure. E A0 (3) and E A1 (3)
decrease as well the share invested in cash as the target return increases, even though in a
less marked fashion. Moreover, they tend to substitute cash with short term bonds but do
not show any clear policy of increasing the share invested in medium and long term bonds.
This average policy is coherent with the ratio between expected return and risk historically
produced by bonds belonging to the various maturities described in Table 3.
Finally, Table 10 reports values for the performance test. The trading strategies associated
with the DTSMs can be evaluated relying to the literature on performance measurement based
on information about portfolio holdings. This literature tries to devise powerful performance
measures by exploiting the information contained in mutual funds’ portfolios. Usually these
measures are hard to compute in practice due to poor information about funds’ actual holdings.
Springer
Ann Oper Res (2007) 151:193–222 219

Table 10 Performance
measurement test Panel A: E A0 (3)
R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
ESM −1.213 −0.905 −0.603 −0.062 0.051
Descriptive statistics for
measuring the performance of the (0.415) (0.433) (0.393) (0.882) (0.788)
asset allocation strategies PCM −0.629 −0.537 −0.434 0.020 0.047
implemented by the DTSMs. For (0.305) (0.175) (0.150) (0.875) (0.532)
each DTSM and risk aversion
Panel B: E A1 (3)
parameter (given by the ratio
R̄t / R̂t in the first row) two tests R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
are reported: ESM is event study ESM 0.446 0.209 0.281 0.210 0.028
measure, looking at the ability of (0.736) (0.838) (0.684) (0.583) (0.934)
the model to produce time period
PCM 0.156 −0.003 0.021 0.116 0.033
returns larger than the
unconditional returns, PCM is (0.541) (0.988) (0.889) (0.273) (0.822)
portfolio change measure looking Panel C: C A2 (3)
at the ability of the model to
increase portfolio weights in R̄t / R̂t 0.5 0.6 0.7 0.8 1.0
those assets with the highest ESM 0.047 −0.008 0.075 0.252 0.094
return. All numbers are in (0.966) (0.992) (0.894) (0.624) (0.915)
percentage. The numbers in PCM −0.028 −0.107 −0.096 0.282 0.097
parentheses are the p-values of (0.911) (0.629) (0.625) (0.116) (0.694)
each statistic.

However we are comparing trading strategies which are known to us and we are able to
describe the portfolio chosen by each strategy every period. We refer to two of the simplest
(and earliest) measures proposed in the literature (see Grinblatt and Titman (1993)):
 T

N
t=1 rn,t
ESM t = wn,t rn,t − (33)
n=1
T

N
PCM t = rn,t (wn,t − wn,t−1 ) (34)
n=1

The former, the event study measure, computes the product between the portfolio weights
at the beginning of time t and the difference between time t returns and the average return—
which is taken as an estimate of the unconditional mean return. The second measure, the
portfolio change measure, is the product between the change in portfolio holdings and the
returns on the various assets. The idea is that a good strategy on average increases weights
on those assets producing the highest future return. Notice that PCM t can also be interpreted
as the return on a zero-cost portfolio. The two measures can be computed each period over a
given sample, and inference can be performed via a standard t-statistic.24 From Table 10 it
emerges that no model produces a performance test that is significantly different from zero.
An indication about the superiority of model E A1 (3) can however be retrieved from Table
10. As a matter of fact E A1 (3) produces a positive performance value in nine out of ten

24 Solnik (1993) generalizes the event study measure and allows for time-varying portfolio volatility (associated
with changing portfolio composition) by normalizing each measure with its the standard devation, i.e. wt V wt
where V is the unconditional variance covariance matrix. We have also computed the performance measure
with this heteroskedasticity correction. We have also computed a third version with an historical estimate of
the conditional variance-covariance matrix, wt Vt wt . However the results are not qualitatively different from
the simple version that we described in the text, so we have chosen not to report all the results for reasons of
space.
Springer
220 Ann Oper Res (2007) 151:193–222

cases, while the other two DTSMs either alternate between positive and negative signs or
generally produce negative performance measures.

5 Conclusions

We have used theoretical affine DTSMs in the context of active bond portfolio management.
We have selected three interesting DTSMs and employed them to produce forecasts for the
future values of the relevant state variables. Starting from the theoretical moments of the
state variables of the models, we have introduced binomial approximations to come up with
discrete scenarios for the future state variables. From the theoretical asset pricing relations
we have computed the bond prices for various maturities at the relevant future dates and the
associated returns. We have used these returns as inputs in a portfolio optimization problem
where an investor with a six month horizon takes into account the possibility to rebalance
after one quarter. The optimizer selects the optimal portfolio each quarter of our sample
period. As usual in the context of these problems, only the first stage of the optimal solution
is actually implemented. The sequence of optimal portfolios is then evaluated in terms of
financial properties.
The goals of the exercise were the following: (1) provide new evidence on the usefulness
of DTSMs for active asset allocation and (2) accumulate evidence on the relative merits of
various affine models from a financial metric standpoint, rather than the standard statistical
metric. Both goals are in our opinion important. There is a growing debate on the validity of
several DTSMs, especially in light of their low performance at replicating basic stylized facts
in the bond market, like the low average return and the high unconditional volatility. Moreover
the debate has so far been concentrated on a statistical evaluation of the models, while we
push towards a financial evaluation of the models. Also, there is an ongoing debate on the
possibility of implementing active asset allocation in efficient markets, and most analyses
have concentrated on the stock market. We try to shift the attention towards the bond market.
While still preliminary due to the specific sample we have used, we feel that our results are
useful to the overall debates about the validity of theoretical DTSMs and about the possibility
of using quantitative methods for active asset allocation.
We show that the superiority of EA over CA models claimed by Duffee (2002) also holds
from the point of view of dynamic portfolio optimization, even though our application does
not simply concentrate on expected returns but also takes risk into consideration. However
evaluating the models by means of a financial metric rather a purely statistical metric reverses
the order of the models within the EA class. The interpretation is that allowing for time-
varying higher moments may more than compensate for lower forecasting power of the
distribution’s mean. Performance tests do not allow to show that the excess return produced
by the model is significantly positive, even though the superiority of the heteroskedastic EA
model is confirmed in terms of average performance. Moreover, the shortfalls produced by
the two EA models are in general statistically significant.
At least three directions of future research seem promising to us. The first lies in a compari-
son of the two-period dynamic optimization model with the standard one-period optimization
model. The former is clearly superior to the latter in theory, even though in practice errors in
the production of scenarios may negatively affect the model’s operational performance. This
issue is interesting to analyze in the context of our current data set and theoretical models
for the term structure of interest rates. The second direction of future research lies in testing
the relevance of affine term structure models in the context of balanced portfolios involving
both stocks and bonds. Recent theoretical advances in theoretical pricing of bonds and stocks
Springer
Ann Oper Res (2007) 151:193–222 221

are very promising steps in this direction. Finally, our methodology can be used in order to
evaluate the quadratic DTSMs in Ahn, Dittmar and Gallant (2002), thus establishing a com-
parison between different DTSMs classes and an alternative perspective -based on financial
performance- to Brandt and Chapman’s (2002) recent work.

Acknowledgments We are indebted to Rita D’Ecclesia, Antonio Mele, Marco Pagano, Hyun Shin, Hercules
Vladimirou and five anonymous referees for detailed comments that have improved the article. We thank sem-
inar participants at the ‘Return predictability and portfolio allocation’ conference in Milan (2002, Università
Bocconi), the LSE/FMG meetings, and the CEMAPRE conference in Lisbon (2003, ISEG) for helpful com-
ments. Financing from the Centro di Economia Monetaria e Finanziaria “Paolo Baffi” at Università Bocconi
is gratefully acknowledged. Any errors are our own.

References

Aı̈t-Sahalia, Y. and M.W. Brandt. (2001). “Variable Selection for Portfolio Choice.” Journal of Finance, 56,
1297–1351.
Ahn, D.-H., R.F. Dittmar, and A.R. Gallant. (2002). “Quadratic Term Structure Models: Theory and Evidence.”
Review of Financial Studies, 15, 243–288.
Avramov, D. (2002). “Stock Return Predictability and Model Uncertainty.” Journal of Financial Economics,
64, 423–458.
Barberis, N. (2000). “Investing for the Long Run when Returns are Predictable.” Journal of Finance, 55,
225–264.
Beltratti, A., A. Consiglio, and S.A. Zenios. (1998). “Scenario Modeling for the Management of International
Bond Portfolios.” Annals of Operations Research, 85, 227–247.
Beltratti, A., A. Laurent, and S.A. Zenios. (2004). “Scenario Modeling of Selective Hedging Strategies.”
Journal of Economic Dynamics and Control, 28, 955–974.
Bliss, R.R. (1997). “Testing Term Structure Estimation Methods.” Advances in Futures and Options Research,
9, 197–231.
Bogentoft, E., H.E. Romeijn, and S. Uryasev. (2001). “Asset/Liability Management for Pension Funds Using
CVaR Constraints.” Journal of Risk Finance, 3, 57–71.
Boyle, P.P., J. Evnine, and S. Gibbs. (1989). “Numerical Evaluation of Multivariate Contingent Claims.”
Review of Financial Studies, 2, 241–250.
Brandt, M.W. and D.A. Chapman. (2002). “Comparing Multifactor Models of the Term Structure.” Mimeo,
The Wharton School.
Breen, W., L.R. Glosten, and R. Jagannathan. (1989). “Economic Significance of Predictable Variations in
Stock Index Returns.” Journal of Finance, 44, 1177–1189.
Brennan, M.J., E.S. Schwartz, and R. Lagnado. (1997). “Strategic Asset Allocation.” Journal of Economic
Dynamics and Control, 21, 1377–1402.
Brennan, M.J. and Y. Xia. (2000). “Stochastic Interest Rates and the Bond-Stock Mix.” European Finance
Review, 4, 197–210.
Bühler, A. and H. Zimmermann. (1996). “A Statistical Analysis of the Term Structure of Interest Rates in
Switzerland and Germany.” Journal of Fixed Income, December, 55–67.
Campbell, J.Y. (2000). “Asset Pricing at the Millennium.” Journal of Finance, 55, 1515–1567.
Campbell, J.Y. and L.M. Viceira. (2001). “Who Should Buy Long-Term Bonds?” American Economic Review,
91, 99–127.
Campbell, J.Y. and L.M. Viceira. (2002). Strategic Asset Allocation: Portfolio Choice for Long-Term Investors.
Oxford: Oxford University Press.
Chapman, D.A. and N.D. Pearson. (2001). “Recent Advances in Estimating Term-Structure Models.” Financial
Analysts Journal, 57, 77–95.
Cheridito, P., D. Filipović, and R.L. Kimmel. (2003). “Market Price of Risk Specification for Affine Models:
Theory and Evidence.” Mimeo, Princeton University.
Chevalier, J. and G. Ellison. (1997). “Risk Taking by Mutual Funds as a Response to Incentives.” Journal of
Political Economy, 105, 1167–1200.
Christoffersen, P.F. and F.X. Diebold. (1997). “Optimal Prediction under Asymmetric Loss.” Econometric
Theory, 13, 808–817.
Cox, J.C., J.E. Ingersoll, and S.A. Ross. (1985). “A Theory of the Term Structure of Interest Rates.” Econo-
metrica, 53, 385–407.
Springer
222 Ann Oper Res (2007) 151:193–222

Cox, J.C., S.A. Ross, and M. Rubinstein. (1979). “Option Pricing: a Simplified Approach.” Journal of Financial
Economics, 7, 229–263.
Dai, Q. and K.J. Singleton. (2000). “Specification Analysis of Affine Term Structure Models.” Journal of
Finance, 55, 1943–1978.
Dai, Q. and K.J. Singleton. (2003). “Term structure Modeling in Theory and Reality.” Review of Financial
Studies, 16, 631–678.
Duffee, G.R. (2002). “Term Premia and Interest Rate Forecasts in Affine Models.” Journal of Finance, 57,
405–443.
Duffie, D. and R.M. Kan. (1996). “A Yield-Factor Model of Interest Rates.” Mathematical Finance, 6, 379–406.
Driessen, J., B. Melenberg, and T. Nijman. (2005). “Testing Affine Term Structure Models in Case of Trans-
action Costs.” Journal of Econometrics, 126, 201–232.
Embrechts, P., C. Klüppelberg, and T. Mikosch. (2003). Modeling Extremal Events for Insurance and Finance,
vol. 33. Berlin, Germany: Springer-Verlag Applications of Mathematics.
Fisher, M. and C. Gilles. (1996). “Term Premia in Exponential-Affine Models of the Term Structure.” Mimeo,
Federal Reserve Board.
Frauendorfer, K. and M. Schürle. (2001). “Term Structure Models in Multistage Stochastic Programming:
Estimation and Approximation.” Annals of Operations Research, 100, 189–209.
Golub, B., M. Holmer, R. McKendall, L. Pohlman, and S.A. Zenios. (1995). “A Stochastic Programming
Model for Money Management.” European Journal of Operations Research, 85, 282–296.
Grinblatt, M. and S. Titman. (1993). “Performance Measurement without Benchmarks: an Examination of
Mutual Fund Returns.” Journal of Business, 66, 47–68.
He, H. (1990). “Convergence from Discrete to Continuous-Time Contingent Claim Prices.” Review of Financial
Studies, 3, 523–546.
Ilmanen, A. (1996). “Does Duration Extension Enhance Long-Term Expected Returns?” Journal of Fixed
Income, 6, 23–36.
Jobst, N.J. and S.A. Zenios. (2001a). “Extending Credit Risk (Pricing) Models for the Simulation of Portfolios
of Interest Rate and Credit Risk Sensitive Securities.” Working Paper 01-25, The Wharton School.
Jobst, N.J. and S.A. Zenios. (2001b). “The Tail that Wags the Dog: Integrating Credit Risk in Asset Portfolios.”
Journal of Risk Finance, 3, 31–43.
Kamrad, B. and P. Ritchken. (1991). “Multinomial Approximating Models for Options with k State Variables.”
Management Science, 37, 1640–1652.
Kandel, S. and R.F. Stambaugh. (1996). “On the Predictability of Stock Returns: an Asset-Allocation Per-
spective.” Journal of Finance, 51, 385–424.
Konno, H. and H. Yamazaki. (1991). “Mean-Absolute Deviation Portfolio Optimization Model and its Appli-
cations to Tokyo Stock Market.” Management Science, 37, 519–531.
Litterman, R. and J. Scheinkman. (1991). “Common Factors Affecting Bond Returns.” Journal of Fixed
Income, 1, 54–61.
McCulloch, J.H. and H.-C. Kwon. (1993). “U.S. Term Structure Data, 1947–1991.” Working Paper 93-6, Ohio
State University.
Nelson, C.R. and A.F. Siegel. (1987). “Parsimonious Modelling of Yield Curves.” Journal of Business, 60,
473–489.
Nelson, D.B. and K. Ramaswamy. (1990). “Simple Binomial Processes as Diffusion Approximations in
Financial Models.” Review of Financial Studies, 3, 393–430.
Pastor, L. (2000). “Portfolio Selection and Asset Pricing Models.” Journal of Finance, 55, 179–223.
Sangvinatsos, A. and J.A. Wachter. (2005). “Does the Failure of the Expectations Hypothesis Matter for
Long-Term Investors?” Journal of Finance, 60, 179–230.
Solnik, B. (1993). “The Performance of International Asset Allocation Strategies using Conditioning Infor-
mation.” Journal of Empirical Finance, 1, 33–55.
Topaloglou, N., H. Vladimirou, and S.A. Zenios. (2002). “CVaR Models with Selective Hedging for Interna-
tional Asset Allocation.” Journal of Banking and Finance, 26, 1535–1561.
Vasicek, O. (1977). “An Equilibrium Characterization of the Term Structure.” Journal of Financial Economics,
5, 177–188.
Wachter, J.A. (2003). “Risk Aversion and Allocation to Long Term Bonds.” Journal of Economic Theory, 112,
325–333.
Walder, R. (2002). “Dynamic Allocation of Treasury and Corporate Bond Portfolios.” Mimeo, Université de
Lausanne-HEC.
Zenios, S.A. and P. Kang. (1993). “Mean-Absolute Deviation Portfolio Optimization for Mortgage-Backed
Securities.” Annals of Operations Research, 45, 433–450.
Zenios, S.A. (ed.). (1993). Financial Optimization. Cambridge: Cambridge University Press.
Zenios, S.A. (1995). “Asset/Liability Management under Uncertainty for Fixed-Income Securities.” Annals
of Operations Research, 59, 77–97.
Springer
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.