Full text of DOI:10.1111/0022-1082.00278

THE JOURNAL OF FINANCE • VOL. LV, NO. 5 • OCT. 2000
Specification Analysis of Affine
Term Structure Models
QIANG DAI and KENNETH J. SINGLETON*
ABSTRACT
This paper explores the structural differences and relative goodness-of-fits of af-
fine term structure models ~ATSMs!. Within the family of ATSMs there is a trade-
off between flexibility in modeling the conditional correlations and volatilities of
the risk factors. This trade-off is formalized by our classification of N-factor affine
family into N ϩ 1 non-nested subfamilies of models. Specializing to three-factor
ATSMs, our analysis suggests, based on theoretical considerations and empirical
evidence, that some subfamilies of ATSMs are better suited than others to explain-
ing historical interest rate behavior.
IN SPECIFYING
A
DYNAMIC TERM STRUCTURE MODEL—one that describes the co-
movement over time of short- and long-term bond yields—researchers are
inevitably confronted with trade-offs between the richness of econometric
representations of the state variables and the computational burdens of pric-
ing and estimation. It is perhaps not surprising then that virtually all of the
empirical implementations of multifactor term structure models that use
time series data on long- and short-term bond yields simultaneously have
focused on special cases of “affine” term structure models ~ATSMs!. An ATSM
accommodates time-varying means and volatilities of the state variables
through affine specifications of the risk-neutral drift and volatility coeffi-
cients. At the same time, ATSMs yield essentially closed-form expressions
for zero-coupon-bond prices ~Duffie and Kan ~1996!!, which greatly facili-
tates pricing and econometric implementation.
The focus on ATSMs extends back at least to the pathbreaking studies by
Vasicek ~1977! and Cox, Ingersoll, and Ross ~1985!, who presumed that the
instantaneous short rate r~t! was an affine function of an N-dimensional
state vector Y~t!, r~t! ϭ d0 ϩ dy' Y~t!, and that Y~t! followed Gaussian and
square-root diffusions, respectively. More recently, researchers have ex-
plored formulations of ATSMs that extend the one-factor Markov represen-
* Dai is from New York University. Singleton is from Stanford University. We thank Darrell
Duffie for extensive discussions, Ron Gallant and Jun Liu for helpful comments, the editor and
the referees for valuable inputs, and the Financial Research Initiative of the Graduate School
of Business at Stanford University for financial support. We are also grateful for comments
from seminar participants at the NBER Asset Pricing Group meeting, the Western Finance
Association Annual Meetings, Duke University, London Business School, New York University,
Northwestern University, University of Chicago, University of Michigan, University of Wash-
ington at St. Louis, Columbia University, Carnegie Mellon University, and CIRANO.
1943

1944
The Journal of Finance
!
tation of the short-rate, dr~t! ϭ ~u Ϫ r~t!! dt ϩ v dB~t!, by introducing a
stochastic long-run mean u~t! and a volatility v~t! of r~t! that are affine
functions of ~r~t!, u~t!,v~t!! ~e.g., Chen ~1996!, Balduzzi et al. ~1966!!.1 These
and related ATSMs underpin extensive literatures on the pricing of bonds
and interest-rate derivatives and also underlie many of the pricing systems
used by the financial industry. Yet, in spite of their central importance in
the term structure literature, the structural differences and relative empir-
ical goodness-of-fits of ATSMs remain largely unexplored.
This paper characterizes, both formally and intuitively, the differences and
similarities among affine specifications of term structure models and as-
sesses their strengths and weaknesses as empirical models of interest rate
behavior. We begin our specification analysis by developing a comprehensive
classification of ATSMs with the following convenient features: ~i! whether a
specification of an affine model leads to well-defined bond prices—a prop-
erty that we will refer to as admissibility ~Section I!—is easily verified; ~ii!
all admissible N-factor ATSMs are uniquely classified into N ϩ 1 non-nested
subfamilies; and ~iii! for each of the N ϩ 1 subfamilies, there exists a max-
imal model that nests econometrically all other models within this subfam-
ily. With this classification scheme in place, we answer the following questions.
@Q1# Given an ATSM ~e.g., one of the popular specifications in the litera-
ture!, is it “maximally flexible,” and, if not, what are the overidentifying
restrictions that it imposes on yield curve dynamics?
@Q2# Are extant models, or their maximal counterparts, sufficiently flex-
ible to describe simultaneously the historical movements in short- and
long-term bond yields?
Ideally, a specification analysis of ATSMs could begin with the specifica-
tion of an all-encompassing ATSM, and then all other ATSMs could be stud-
ied as special cases. However, for a specification to be admissible, constraints
must be imposed on the dynamic interactions among the state variables, and
these constraints turn out to preclude the existence of such an all-encompassing
model. Therefore, as a preliminary step in our specification analysis, we
characterize the family of admissible ATSMs.2 This is accomplished by clas-
sifying ATSMs into N ϩ 1 subfamilies according to the number “m” of the Ys
~more precisely, the number of independent linear combinations of Ys! that
determine the conditional variance matrix of Y, and then using this classi-
fication to provide a formal and intuitive characterization of ~minimal known!
sufficient conditions for admissibility. Each of the N ϩ 1 subfamilies of ad-
missible models is shown to have a maximal element, a feature we exploit in
answering Q1 and Q2.
1Nonlinear models, such as those studied by Chan et al. ~1992!, can be extended in similar
fashion to multi-factor models, such as those of Andersen and Lund ~1998!, that fall outside the
affine family.
2
The problem of admissibility was not previously an issue in empirical implementations of
ATSMs, because the special structure of Gaussian and CIR-style models made verification that
bond prices are well defined relatively straightforward.

Affine Term Structure Models
1945
The usefulness of this classification scheme is illustrated by specializing to
the case of N ϭ 3 and describing in detail the nature of the four maximal mod-
els for the three-factor family of ATSMs. This discussion highlights an impor-
tant trade-off within the family of ATSMs between the dependence of the
conditional variance of each Yi~t! on Y~t! and the admissible structure of the
correlation matrix for Y. Gaussian models offer complete flexibility with re-
gard to the signs and magnitudes of conditional and unconditional correla-
tions among the Ys but at the “cost” of the apparently counterfactual assumption
of constant conditional variances ~m ϭ 0!. At the other end of the spectrum of
volatility specifications lies ~what we refer to as! the correlated square-root dif-
fusion ~CSR! model that has all three state variables driving conditional vol-
atilities ~m ϭ 3!. However, admissibility of models in this subfamily requires
that the conditional correlations of the state variables be zero and that their
unconditional correlations be non-negative. In between the Gaussian and CS-
Rmodels lie two subfamilies of ATSMs with time-varying conditional volatil-
ities of the state variables and unconstrained signs of ~some of! their correlations.
Specializing further, we show that the Vasicek ~Gaussian!, BDFS, Chen,
and CIR models are classified into distinct subfamilies. Moreover, compar-
ing these models to the maximal models in their respective subfamilies, we
find that, in every case except the Gaussian models, these models impose
potentially strong overidentifying restrictions relative to the maximal model.
Thus, we answer Q1 by showing that there exist identified, admissible ATSMs
that allow much richer interdependencies among the factors than have here-
tofore been studied.
One notable illustration of this point is our finding that the standard as-
sumption of independent risk factors in CIR-style models ~see, e.g., Chen
and Scott ~1993!, Pearson and Sun ~1994!, and Duffie and Singleton ~1997!!
is not necessary either for admissibility of these models or for zero-coupon-
bond prices to be known ~essentially! in closed form.3 At the same time we
show that, when the correlations in these CSR models are nonzero, they
must be positive for the model to be admissible. The data on U.S. interest
rates seems to call for negative correlations among the risk factors ~see Sec-
tion II.C!. Because CSR models are theoretically incapable of generating
negative correlations, we conclude that they are not consistent with the his-
torical behavior of U.S. interest rates.
Given the absence of time-varying volatility in Gaussian ~m ϭ 0! models
and the impossibility of negatively correlated risk factors in CSR ~m ϭ 3!
models, in answering Q2, we focus on the two subfamilies of N ϭ 3 models
in which the stochastic volatilities of the Ys are controlled by one ~m ϭ 1!
and two ~m ϭ 2! state variables. The maximal model for the subfamily
m ϭ 1 nests the BDFS model, whereas the maximal model for the m ϭ 2
subfamily nests the Chen model.
We compute simulated method of moments ~SMM! estimates ~Duffie and
Singleton ~1993!, Gallant and Tauchen ~1996!! of our maximal ATSMs. Whereas
most of the empirical studies of term structure models have focused on U.S.
3
We are grateful to Jun Liu for pointing out this implication of our admissibility conditions.

1946
The Journal of Finance
Treasury yield data, we follow Duffie and Singleton ~1997! and study LIBOR-
based yields from the ordinary, fixed-for-variable rate swap market. The
primary motivation for choosing swap instead of Treasury yields is that the
former are relatively unencumbered by institutional factors that are not fully
accounted for in standard arbitrage-free term structure models, including
the ATSMs studied in this paper ~see Section II.A!. The models pass several
formal goodness-of-fit tests. Moreover, the restrictions implicit in the Chen
and BDFS models are strongly rejected. The substantial improvements in
goodness-of-fit for the newly introduced, maximal models are traced directly
to their more flexibly parameterized correlations among the state variables.
Negatively correlated diffusions are central to the models’ abilities to match
the volatility structure and non-normality of changes in bond yields.
The remainder of the paper is organized as follows. Section I presents
general results pertaining to the classification, admissibility, and identifi-
cation of the family of N-factor affine term structure models. Section I.B
specializes the classification results to the family of three-factor affine term
structure models and characterizes explicitly the nature of the overidenti-
fying restrictions in extant models relative to our more flexible, maximal
models. Section II.A explains our estimation strategy and data. Section II.B
discusses the econometric identification of risk premiums in affine models.
Section II.C presents our empirical results. Finally, Section III concludes.
I. A Characterization of Admissible ATSMs
Absent arbitrage opportunities, the time-t price of a zero-coupon bond that
matures at time T, P~t, t!, is given by
t
r ds
P~t, t! ϭ EtQ eϪ
,
~1!
͵
s
t
ͫ ͬ
where EQ denotes expectation under the risk-neutral measure. An N-factor
affine term structure model is obtained under the assumptions that the in-
stantaneous short rate r~t! is an affine function of a vector of unobserved
state variables Y~t! ϭ ~Y1~t!,Y2~t!,...,YN~t!!,
N
r~t! ϭ d0 ϩ
di Yi~t! [ d0 ϩ dy' Y~t!,
~2!
(
iϭ1
and that Y~t! follows an “affine diffusion,”
G
!
dY~t! ϭ KE uD Ϫ Y~t! dt ϩ ⌺ S~t!dW~t!.
~3!
~
!
G
W~t! is an N–dimensional independent standard Brownian motion under Q,
KE and ⌺ are N ϫ N matrices, which may be nondiagonal and asymmetric,
and S~t! is a diagonal matrix with the ith diagonal element given by
@S~t!#ii ϭ ai ϩ bi' Y~t!.
~4!

Affine Term Structure Models
1947
Both the drifts in equation ~3! and the conditional variances in equation ~4!
of the state variables are affine in Y~t!.
Provided a parameterization is admissible, we know from Duffie and Kan
~1996! that
'
P~t, t! ϭ eA~t!ϪB~t! Y~t!
,
~5!
where A~t! and B~t! satisfy the ordinary differential equations ~ODEs!
N
dA~t!
dt
1
2
'
ϭ ϪuD 'KE B~t! ϩ
@⌺' B~t!#2i ai Ϫ d0,
~6!
~7!
(
iϭ1
N
dB~t!
dt
1
'
ϭ ϪKE B~t! Ϫ
@⌺' B~t!#2i bi ϩ dy.
(
2
iϭ1
These ODEs, which can be solved easily through numerical integration start-
ing from the initial conditions A~0! ϭ 0 and B~0! ϭ 0Nϫ1, are completely
determined by the specification of the risk-neutral dynamics of r~t!, in equa-
tions ~2! through ~4!.
To use the ~essentially! closed-form expression of equation ~5! in empirical
studies of ATSMs we also need to know the distributions of Y~t! and P~t, t!
under the actual probability measure P. We assume that the market prices
of risk, ⌳~t!, are given by
!
⌳~t! ϭ S~t!l,
~8!
where l is an N ϫ 1 vector of constants. Under this assumption, the process
for Y~t! under P also has the affine form,4
!
dY~t! ϭ K ⌰ Ϫ Y~t! dt ϩ ⌺ S~t!dW~t!,
~9!
~
!
where W~t! is an N-dimensional vector of independent standard Brownian
motions under P, K ϭ KE Ϫ ⌺⌽, ⌰ ϭ KϪ1~KE uD ϩ ⌺c!, the ith row of ⌽ is given
by li bi', and c is an N–vector whose ith element is given by li ai.
4
Our formulation can be easily generalized to a nonaffine diffusion for Y under P, while
preserving the tractable pricing relations in equation ~5!, as long as the specification of ⌳~t!
preserves the affine structure of Y under Q. See Duffie ~1998! for a generalization of our frame-
work along this line. We do not pursue this issue here, partly because our primary focus is on
the correlation structure of Y~t! and partly because we find it difficult to accurately estimate
the market prices of risk, even for our simple parameterization of ⌳~t! ~see Section II.B!.

1948
The Journal of Finance
A. A Canonical Representation of Admissible ATSMs
The general specification in equation ~9! does not lend itself to a specifi-
cation analysis because, for an arbitrary choice of the parameter vector c [
~K, ⌰, ⌺,B, a!, where B [ ~b1,..., bN! denotes the matrix of coefficients on Y
in the @S~t!#ii, the conditional variances @S~t!#ii may not be positive over the
range of Y. We will refer to a specification of c as admissible if the resulting
@S~t!#ii are strictly positive, for all i. There is no admissibility problem if
bi ϭ 0. However, outside of this special case, to assure admissibility we find
it necessary to constrain the drift parameters ~K and ⌰! and diffusion coef-
ficients ~⌺ and B!. Moreover, our requirements for admissibility become in-
creasingly stringent as the number of state variables determining @S~t!#ii
increases.
To formalize what we mean by the family of admissible N-factor ATSMs,
let m ϭ rank~B! index the degree of dependence of the conditional variances
on the number of state variables. Using this index, we classify each ATSM
uniquely into one of N ϩ 1 subfamilies based on its value of m. Not all
N-factor ATSMs with index value m are admissible, so, for each m, we let
Am~N ! denote those that are admissible. Next, we define the canonical rep-
resentation of Am~N ! as follows.
Definition 1 @Canonical Representation of Am~N !# : For each m, we parti-
'
'
tion Y~t! as Y ' ϭ ~Y B ,Y D !, where Y B is m ϫ 1 and Y D is ~N Ϫ m! ϫ 1, and
define the canonical representation of Am~N ! as the special case of equa-
tion ~9! with
BB
Kmϫm
0mϫ~NϪm!
K ϭ
,
~10!
ͫ
ͬ
DB
DD
K~NϪm!ϫm K~NϪm!ϫ~NϪm!
for m Ͼ 0, and K is either upper or lower triangular for m ϭ 0,
B
mϫ1
⌰
⌰ ϭ
,
~11!
~12!
~13!
ͩ ͪ
0~NϪm!ϫ1
⌺ ϭ I,
0mϫ1
a ϭ
,
ͩ ͪ
1~NϪm!ϫ1
BD
Imϫm
B
mϫ~NϪm!
B ϭ
;
~14!
ͫ
ͬ
0~NϪm!ϫm 0~NϪm!ϫ~NϪm!

Affine Term Structure Models
1949
with the following parametric restrictions imposed:
di Ն 0,
m ϩ 1 Յ i Յ N,
~15!
~16!
m
Ki ⌰ [
Kij ⌰j Ͼ 0,
1 Յ i Յ m,
j Þ i,
(
jϭ1
Kij Յ 0,
1 Յ j Յ m,
1 Յ i Յ m,
1 Յ i Յ m, m ϩ 1 Յ j Յ N.
~17!
~18!
~19!
⌰i Ն 0,
Bij Ն 0,
Finally, Am~N ! is formally defined as the set of all ATSMs that are nested
special cases of our canonical model or of any equivalent model obtained by
an invariant transformation of the canonical model. Invariant transforma-
tions, which are formally defined in Appendix A, preserve admissibility and
identification and leave the short rate ~and hence bond prices! unchanged.
The assumed structure of B assures that rank~B! ϭ m for the mth canon-
ical representation. To verify that it resides in Am~N !, note that the instan-
taneous conditional correlations among the Y B~t! are zero, whereas the
instantaneous correlations among the Y D~t! are governed by the parameters
Bij, because ⌺ ϭ I.5 Because the conditional covariance matrix of Y depends
only on Y B and equation ~19! holds, admissibility is established if Y B~t! is
strictly positive. The positivity of Y B is assured by zero restrictions in the
upper right m ϫ ~N Ϫ m! block of K and the constraints in equations ~17!
and ~18!. In Appendix A we show formally that the zero restrictions in equa-
tions ~10! through ~14! and sign restrictions in equations ~15! through ~19!
are sufficient conditions for admissibility. To assure that the state process is
stationary, we need to impose the additional constraint that all of the eigen-
values of K are strictly positive. We impose this constraint in our empirical
estimation.
Not only is the canonical representation admissible, but it is also “maxi-
mal” in Am~N ! in the sense that, given m, we have imposed the minimal
known sufficient condition for admissibility and then imposed minimal nor-
malizations for econometric identification ~see Appendix A!.6 However, our
canonical representation is not the unique maximal model in Am~N !. Rather,
there is an equivalence class AMm~N ! of maximal models obtained by in-
variant transformations of the canonical representation. The representation
5
All of the state variables may be mutually correlated over any finite sampling interval due
to feedback through the drift matrix K.
6
Because the conditions for admissibility are sufficient, but are not known to be necessary,
we cannot rule out the possibility that there are admissible, econometrically identified ATSMs
that nest our canonical models as special cases. Importantly, all of the extant ATSMs in the
literature reside within Am~N !, for some m and N.

1950
The Journal of Finance
in equations ~10! through ~19! was chosen as our canonical representation
among equivalent maximal models in AMm~N !, because of the relative ease
with which admissibility and identification can be verified and the paramet-
ric restrictions in equations ~15! through ~19! can be imposed in econometric
implementations.
As will be illustrated subsequently, the canonical representation of Am~N !
is often not as convenient as other members of AMm~N ! for interpreting the
state variables within a particular ATSM. In particular, the literature has
often chosen to parameterize ATSMs with the riskless rate r being one of the
state variables. Any such Ar ~“affine in r”! representation is typically in
Am~N !, for some m and N, and therefore has an equivalent representation in
which r~t! ϭ d0 ϩ dy' Y~t!, with Y~t! treated as an unobserved state vector ~an
AY or “affine in Y” representation!. Similarly, any AY representation ~e.g., a
CIR model! typically has an equivalent Ar representation. In Section I.B we
present the equivalent Ar and AY representations of several extant ATSMs
and also their maximally flexible counterparts.
An implication of our classification scheme is that an exhaustive specifi-
cation analysis of the family of admissible N-factor ATSMs requires the ex-
amination of N ϩ 1 non-nested, maximal models.
B. Three-Factor ATSMs
In this section we explore in considerably more depth the implications of
our classification scheme for the specification of ATSMs. Particular atten-
tion is given to interpreting the term structure dynamics associated with
our canonical models and to the nature of the overidentifying restrictions
imposed in several ATSMs in the literature. To better link up with the em-
pirical term structure literature, we fix N ϭ 3 and examine the four associ-
ated subfamilies of admissible ATSMs.
B.1. A0~3!
If m ϭ 0, then none of the Ys affect the volatility of Y~t!, so the state
variables are homoskedastic and Y~t! follows a three-dimensional Gaussian
diffusion. The elements of c for the canonical representation of AM0~3! are
given by
k11
0
0
0
1
0
0
1
0
0
k21 k22
K ϭ
,
⌺ ϭ
,
΄
΅ ΄ ΅
0
0
1
k31 k32 k33
0
1
1
0
0
0
0
0
0
⌰ ϭ
0
,
a ϭ
,
B ϭ
,
΄ ΅
΂ ΃ ΂ ΃
0
1
0
0
0
where k11 Ͼ 0, k22 Ͼ 0, and k33 Ͼ 0.

Affine Term Structure Models
1951
Gaussian ATSMs were studied theoretically by Vasicek ~1977! and Lange-
tieg ~1980!, among many others. A recent empirical implementation of a two-
factor Gaussian model is found in Jegadeesh and Pennacchi ~1996!.
B.2. A1~3!
The family A1~3! is characterized by the assumption that one of the Ys
determines the conditional volatility of all three state variables. One mem-
ber of A1~3! is the BDFS model:
!
du~t! ϭ m~uT Ϫ u~t!!dt ϩ h[ u~t!dBv~t!,
du~t! ϭ n~uN Ϫ u~t!!dt ϩ zdBu~t!,
~20!
Z
!
dr~t! ϭ k~u~t! Ϫ r~t!!dt ϩ u~t!dBr ~t!,
Z
with the only nonzero diffusion correlation being cov~dBv~t!, dBr~t!! ϭ rrv dt.
Rewriting the short rate equation in equation ~20! as
dr~t! ϭ k~u~t! Ϫ r~t!!dt ϩ 1 Ϫ r2 u~t!dBr ~t! ϩ rrv u~t!dBv~t!,
~21!
!
!
!
rv
where Br ~t! and Bv~t! are independent, and replacing u~t! by v~t! ϭ
~1 Ϫ r2 !u~t!, uT by vS ϭ ~1 Ϫ r2 !uT , and h[ by h ϭ 1 Ϫ r2 h[ , we obtain the
!
rv
rv
rv
BDFS model expressed in our notational convention for ATSMs,
!
dv~t! ϭ m~vS Ϫ v~t!!dt ϩ h v~t!dBv~t!,
du~t! ϭ n~uN Ϫ u~t!!dt ϩ zdBu~t!,
~22!
!
!
dr~t! ϭ k~u~t! Ϫ r~t!!dt ϩ v~t!dBr ~t! ϩ srv h v~t!dBv~t!,
where srv ϭ r 0h 1 Ϫ r2 .
!
rv
rv
The first state variable, v~t!, is a volatility factor, because it affects the
short rate process only through the conditional volatility of r. The second
state variable, u~t!, is the “central tendency” of r. The short rate mean re-
verts to its central tendency u~t! at rate k.
For interpreting the restrictions in the BDFS and related models, it is
convenient to work with the following maximal model in AM1~3!, presented
in its Ar representation:
!
dv~t! ϭ m~vS Ϫ v~t!!dt ϩ h v~t!dBv~t!,
du~t! ϭ n~uN Ϫ u~t!!dt ϩ z2 ϩ b v~t!dB ~t!
!
u
u
!
ϩ suv h v~t!dBv~t! ϩ sur a ϩ v~t!dB ~t!,
~23!
!
r
r
dr~t! ϭ krv ~vS Ϫ v~t!!dt ϩ k~u~t! Ϫ r~t!!dt ϩ
a ϩ v~t!dBr ~t!
r
!
ϩ srv h v~t!dBv~t! ϩ sru z2 ϩ b v~t!dB ~t!.
!
!
u
u

1952
The Journal of Finance
The BDFS model is the special case of the expressions in equation ~23! in
which the parameters in square boxes are set to zero. Thus, relative to this
maximal model, the BDFS model constrains the conditional correlation be-
tween r and u to zero ~sru ϭ sur ϭ 0!. Additionally, it precludes the volatility
shock v from affecting the volatility of the central tendency factor u ~suv ϭ 0!.
Finally, the BDFS model constrains krv ϭ 0 so that v cannot affect the drift
of r. Freeing up these restrictions gives us a more flexible ATSM in which v
is still naturally interpreted as the volatility shock but in which u is perhaps
not as naturally interpreted as the central tendency of r. The overidentifying
restrictions imposed in the BDFS model are examined empirically in
Section II.C.
Though the model specified by the expressions in equation ~23! is conve-
nient for interpreting the popular A1~3! models, verifying that the model is
maximal and, indeed, that it is admissible, is not straightforward. To check
admissibility, it is much more convenient to work directly with the following
equivalent AY representation ~see Appendix E, Section E.1!:
r~t! ϭ d0 ϩ d1 Y1~t! ϩ Y2~t! ϩ Y3~t!,
~24!
~25!
Y1~t!
Y1~t!
u1
k11
0
0
k22
0
0
0
—
—
—
0
d
ϭ
ϫ
Ϫ
dt
Y2~t!
Y2~t!
΄
΄
΅
΄
΅
΂ ΃
΂ ΃ ΂ ΃
0
k33
Y3~t!
0
Y3~t!
S11~ t !
0
0
0
1
0
0
s21
1
s23
0
0
S22~t!
dB~t!,
΅ ΄
Ί
΅
s31 s32
1
0
S33~t!
where
S11~ t ! ϭ Y1~ t !,
S22~t! ϭ a2 ϩ @b2#1 Y1~t!,
S33~t! ϭ a3 ϩ @b3#1 Y1~t!.
~26!
That this representation is in AM1~3! follows from its equivalence to our
canonical representation of AM1~3!. This can be shown easily by diagonaliz-
ing ⌺ and by normalizing the scale of S22 and S33, while freeing up d2 and d3
in the expression r~t! ϭ d0 ϩ d1Y1~t! ϩ d2Y2~t! ϩ d3Y3~t!. All three diffusions

Affine Term Structure Models
1953
may be conditionally correlated, and all three conditional variances may de-
pend on Y1. However, admissibility requires that s12 ϭ 0 and s13 ϭ 0, in
which case Y1 follows a univariate square-root process that is strictly positive.
The equivalent AY representation of the BDFS model is obtained from
equation ~24! by setting all of the parameters in square boxes to zero, except
for s32 which is set to Ϫ1. An immediate implication of this observation is
that the BDFS model unnecessarily constrains the instantaneous short rate
to be an affine function of only two of the three state variables ~d1 ϭ 0!. This
is an implication of the assumption that the volatility factor v~t! enters r
only through its volatility and, therefore, it affects r only indirectly through
its effects on the distribution of ~Y2~t!,Y3~t!!. This constraint on dy is a fea-
ture of many of the extant models in the literature, including the model of
Anderson and Lund ~1998!.
B.3. A2~3!
The family A2~3! is characterized by the assumption that the volatilities of
Y~t! are determined by affine functions of two of the three Y’s. A member of
this subfamily is the model proposed by Chen ~1996!:
!
dv~t! ϭ m~vS Ϫ v~t!!dt ϩ h v~t!dW1~t!,
!
du~t! ϭ n~uN Ϫ u~t!!dt ϩ z u~t!dW2~t!,
~27!
!
dr~t! ϭ k~u~t! Ϫ r~t!!dt ϩ v~t!dW3~t!,
with the Brownian motions assumed to be mutually independent. As in the
BDFS model, v and u are interpreted as the stochastic volatility and central
tendency, respectively, of r. A primary difference between the Chen and BDFS
models, and the one that explains their classifications into different subfam-
ilies, is that u in the former follows a square-root diffusion, whereas it is
Gaussian in the latter.
A convenient maximal model for interpreting the overidentifying restric-
tions in the Chen model is
!
dv~t! ϭ m~vS Ϫ v~t!!dt ϩ kvu ~uN Ϫ u~t!!dt ϩ h v~t! dW1~t!,
!
du~t! ϭ n~uN Ϫ u~t!!dt ϩ kuv ~vS Ϫ v~t!!dt ϩ z u~t! dW2~t!,
dr~t! ϭ krv ~vS Ϫ v~t!!dt
~28!
ϩ kru ~uN Ϫ u~t!!dt Ϫ k~uN Ϫ u~t!!dt ϩ k rS Ϫ r~t! dt
~
!
!
!
ϩ srv h v~t! dW1~t! ϩ sru z u~t! dW2~t!
ϩ
a
r
ϩ bu u~t! ϩ v~t! dW3~t!.
!

1954
The Journal of Finance
The Chen model is obtained as a special case with the parameters in square
boxes set to zero except that rS ϭ uN .
Clearly, within this maximal model, u and v are no longer naturally inter-
preted as the central tendency and volatility factors for r. There may be
feedback between u and v through their drifts, and both of these variables
may enter the drift of r. Moreover, the volatility of r may depend on both u
and v. Also, the Chen model unnecessarily constrains the correlations be-
tween u~t! and r~t! and between v~t! and r~t! to 0. All of these restrictions
are examined empirically in Section II.C.
Again, we turn to the equivalent AY representation of the model specified
in equations ~28! to verify that it is admissible and maximal ~see Appen-
dix E, Section E.2!:
r~t! ϭ d0 ϩ d1 Y1~t! ϩ Y2~t! ϩ Y3~t!,
~29!
Y1~t!
Y2~t!
Y1~t!
Y2~t!
—
u1
u2
—
k11 k12
k21 k22
0
0
d
ϭ
Ϫ
dt
—
΄
΅
΄
΅
΂ ΃
΂ ΃ ΂ ΃
0
0
k33
Y3~t!
Y3~t!
0
S11~ t !
0
0
0
1
0
0
1
0
0
1
0
0
S22~t!
ϩ
dB~t!,
΄
΅ ΄
Ί
΅
s31 s32
0
S33~t!
~30!
~31!
where
S11~ t ! ϭ @b1#1 Y1~t!,
S22~t! ϭ @b2#2 Y2~t!,
S33~t! ϭ a3 ϩ Y1~t! ϩ @b3#2 Y2~t!.
With the first two state variables driving volatility, k12 and k21 must be less
than or equal to zero in order to assure that Y1 and Y2 remain strictly pos-
itive. That is, ~Y1~t!,Y2~t!! is a bivariate, correlated square-root diffusion.
Additionally, admissibility requires that Y1 and Y2 be conditionally uncorre-
lated and Y3 not enter the drift of these variables. Y3 can be conditionally
correlated with ~Y1,Y2!, and its variance may be an affine function of ~Y1,Y2!.

Affine Term Structure Models
1955
The corresponding restrictions on the AY representation in equations ~29!
through ~31! implied by the Chen model are obtained by setting the square
boxes in equation ~30! to zero except that s32 ϭ Ϫ1 and d0 ϭ Ϫd1u1 Ϫ u2 ϩ
qu2 ϭ Ϫu2k220k33. As in the BDFS model, in the Chen model r is constrained
to be an affine function of only two of the three state variables.
B.4. A3~3!
The final subfamily of models has m ϭ 3 so that all three Ys determine the
volatility structure. The canonical representation of AM3~3! has parameters
k11 k12 k13
1
0
0
1
0
0
k21 k22 k23
K ϭ
,
⌺ ϭ
,
΄
΅ ΄ ΅
0
0
1
k31 k32 k33
⌰ ϭ ~u1, u2, u3!', a ϭ 0, and B ϭ I3, where kii Ͼ 0 for 1 Յ i Յ 3, kij Յ 0 for
1 Յ i Þ j Յ 3, ui Ͼ 0 for 1 Յ i Յ 3.
With both ⌺ and B equal to identity matrices, the diffusion term of this
model is identical to that in the N-factor model based on independent square-
root diffusions ~often referred to as the CIR model!. With B diagonal, the
requirements of admissibility preclude relaxation of the assumption that ⌺
is diagonal. However, admissibility does not require that K be diagonal, as in
the classical CIR model, but rather only that the off-diagonal elements of K
be less than or equal to zero ~see equation ~17!!. Thus, the canonical repre-
sentation is a correlated, square-root ~CSR! diffusion model. It follows that
the empirical implementations of multifactor CIR-style models with inde-
pendent state variables by Chen and Scott ~1993!, Pearson and Sun ~1994!,
and Duffie and Singleton ~1997!, among others, have imposed overidentify-
ing restrictions by forcing K to be diagonal. In this three-factor model, a
diagonal K implies six overidentifying restrictions.
C. Comparative Properties of Three-Factor ATSMs
In concluding this section, we highlight some of the similarities and dif-
ferences among ATSMs and motivate the subsequent empirical investigation
of three-factor models.
Positivity of the instantaneous short rate r:
As a general rule, for three-factor ATSMs, 3 Ϫ m of the state variables in
Am~3! models may take on negative values. That is, the Gaussian ~m ϭ 0!
model allows all three state variables to become negative, the A1~3! model
allows two of the three state variables to become negative, and so on. There-
fore, only in the case of models in A3~3! are we assured that r~t! Ͼ 0, pro-
vided that we constrain d0 and all elements of dy are non-negative. Emphasis
on having r~t! positive has, at least partially, led some researchers to con-

1956
The Journal of Finance
sider models outside the affine class, such as the discrete-time “level-
GARCH” model of Brenner, Harjes, and Kroner ~1996! and Koedijk et al.
~1997! or the diffusion specifications of Andersen and Lund ~1997! and Gal-
lant and Tauchen ~1998!. These authors assure positivity by having the short
rate volatility be proportional to the product of powers of the short rate and
another non-negative process.
Conditional second moments of zero-coupon bond yields:
For ATSMs in Am~N !, the conditional variances of zero-coupon-bond yields
are determined by m common factors. However, the additional flexibility in
specifying conditional volatilities that comes with increasing m is typically
accompanied by less flexibility in specifying conditional correlations. The
nature of the conditional correlations accommodated within AMm~3! can be
DB
DD
seen most easily by normalizing K~NϪm!ϫm to zero and K~NϪm!ϫ~NϪm! to a
diagonal matrix and concurrently freeing up ⌺DB and the off-diagonal ele-
ments of ⌺DD. This gives an equivalent model to our canonical representa-
tion of Am~N ! ~see Appendix C!. It follows that the admissibility constraints
accommodate nonzero conditional correlations of unconstrained signs be-
tween each element of Y D and the entire state vector Y~t!. For instance,
with N ϭ 3 and m ϭ 3, the state variables are conditionally uncorrelated. On
the other hand, with m ϭ 2, only two state variables determine the volatility
of Y~t!, but Y1~t! may be conditionally correlated with both Y2~t! and Y3~t!.
Unconditional correlations among the state variables:
In the Gaussian model ~m ϭ 0!, the signs of the nonzero elements of K are
unconstrained, and, hence, unconditional correlations among the state vari-
ables may be positive or negative. On the other hand, for the CSR model
with m ϭ 3, the unconditional correlations among the state variables must
be non-negative. This is an implication of the zero conditional correlations
and the sign restrictions on the off-diagonal elements of K required by the
admissibility conditions. The case of m ϭ 1 is similar to the Gaussian model
in that ⌺ may induce positive or negative conditional or unconditional cor-
relations among the Ys. Finally, in the case of m ϭ 2, the first state variable
may be negatively correlated with the other two, but correlation between
Y2~t! and Y3~t! must be non-negative.
Notice that a limitation of the affine family of term structure models is
that one cannot simultaneously allow for negative correlations among the
state variables and require that r~t! be strictly positive.
These observations motivate the focus of our subsequent empirical analy-
sis of three-factor ATSMs on the two branches A1~3! and A2~3!. Models with
m ϭ 1 or m ϭ 2 have the potential to explain the widely documented condi-
tional heteroskedasticity and excess kurtosis in zero-coupon-bond yields, while
being flexible with regard to both the magnitudes and signs of the admis-

Affine Term Structure Models
1957
sible correlations among the state variables. In contrast, models in the branch
A0~3! have the counterfactual implication that zero yields are conditionally
normal with constant conditional second moments.
Though models in A3~3! allow for time-varying volatilities and correlated
factors, the requirement that conditional correlations be zero and uncondi-
tional factor correlations be positive turns out to be inconsistent with his-
torical U.S. data. Evidence suggestive of the importance of negative factor
correlations can be gleaned from previous empirical studies of multifactor
CIR models, which presume that the factors are statistically independent.
For instance, we computed the sample correlation between the implied state
variables from the two-factor ~N ϭ 2! CIR model studied in Duffie and Single-
ton ~1997! and found a correlation of approximately Ϫ0.5 instead of zero.7
Additional evidence on the importance of negative correlations within A1~3!
and A2~3! is presented in Section II.C.
II. Empirical Analysis of ATSMs
A. Data and Estimation Method
For our empirical analysis we study yields on ordinary, fixed-for-variable
rate U.S. dollar swap contracts. A primary motivation for this choice is that
swap markets provide true “constant maturity” yield data, whereas in the
Treasury market the maturities of “constant maturity” yields are only ap-
proximately constant or the data represent interpolated series. Additionally,
the on-the-run treasuries that are often used in empirical studies are typi-
cally on “special” in the repo market, so, strictly speaking, the Treasury data
should be adjusted for repo specials prior to an empirical analysis. Unfor-
tunately, the requisite data for making these adjustments are not readily
available, and, consequently, such adjustments are rarely made.
Though the institutional structures of dollar swap and U.S. Treasury mar-
kets are different, some of the basic distributional characteristics of the as-
sociated yields are similar. For instance, principal component analyses yield
very similar looking “level,” “slope,” and “curvature” factors for both mar-
kets.8 Additionally, the shapes of the “term structures” of yield volatilities
examined in Section II.C are qualitatively similar for both data sets. Nev-
ertheless, empirical studies of ATSMs with swap and Treasury yield data
may lead to different conclusions for at least two reasons. One is the differ-
ence in institutional features outlined above. Another is that the histories of
yield data available in swap markets are in general shorter than those for
7
Duffie and Singleton ~1997! assume that the two- and 10-year swap yields are priced per-
fectly by their two-factor model. Thus, using their pricing model evaluated at the maximum
likelihood estimates of the parameters, implied state variables are computed as functions of
these two swap yields.
8
See, for example, Litterman and Scheinkman ~1997! for a principal components analysis of
the U.S. Treasury data.

1958
The Journal of Finance
Figure 1. Time Series of Swap Yields. Our sample covers the period from April 3, 1987, to
August 23, 1996, weekly. The yields plotted in this graph include, from the lowest to the highest
line ~with occasional cross-overs when the yield curves are inverted!, six-month LIBOR, two-,
three-, four-, five-, seven-, and 10-year swap rates. The maturities used for estimation are the
six-month LIBOR, two-, and 10-year swap rates.
Treasury markets. In particular, available histories of swap yields do not
include the periods of the oil price shocks of the early 1970s, the monetary
experiment in the early 1980s, and so on.
The observed data y were chosen to be the yields on six-month LIBOR and
two-year and 10-year fixed-for-variable rate swaps sampled weekly from April
3, 1987, to August 23, 1996 ~see Figure 1 for a time series plot of the LIBOR
and swap yields!. The length of the sample period was determined in part by
the unavailability of reliable swap data for years prior to 1987. The yields
are ordered in y according to increasing maturity ~i.e., y1 is the six-month
LIBOR rate, etc.!.
The conditional likelihood function of the state vector Y~t! is not known
for general affine models. Therefore, we pursue the method of simulated
moments ~SMM! proposed by Duffie and Singleton ~1993! and Gallant and
Tauchen ~1996!. A key issue for the SMM estimation strategy is the selection
of moments. Following Gallant and Tauchen ~1996!, we use the scores of the
likelihood function from an auxiliary model that describes the time series

Affine Term Structure Models
1959
properties of bond yields as the moment conditions for the SMM estimator.
More precisely, let yt denote a vector of yields on bonds with different ma-
turities, xt' ϭ ~yt', yt'Ϫ1,..., yt'Ϫᐉ!, and f ~yt 6xtϪ1, g! denote the conditional den-
sity of y associated with the auxiliary description of the yield data. We searched
for the best f for our data set along numerous model expansion paths, guided
by a model selection criterion, as outlined in Gallant and Tauchen ~1996!.
The end result of this search was the auxiliary model
f ~yt 6xtϪ1, g! ϭ c~xtϪ1! e ϩ @h~z 6xtϪ1!# 2 n~z !,
~32!
@
#
0
t
t
where n~{! is the density function of the standard normal distribution, e0 is
a small positive number,9 h~z6x! is a Hermite polynomial in z, c~xtϪ1! is a
normalization constant, and xtϪ1 is the conditioning set. We let zt be the
normalized version of yt, defined by
zt ϭ RxϪ,1tϪ1~yt Ϫ mx, tϪ1!.
~33!
In the terminology of Gallant and Tauchen ~1996!, the auxiliary model may
be described as “Non-Gaussian, VAR~1!, ARCH~2!, Homogeneous-Innovation.”
“VAR~1!” refers to the fact that the shift vector mx, tϪ1 is linear with elements
that are functions of Lm ϭ 1 lags of y, in that
c1 ϩ c4 y1, tϪ1 ϩ c7 y2, tϪ1 ϩ c10 y3, tϪ1
c2 ϩ c5 y1, tϪ1 ϩ c8 y2, tϪ1 ϩ c11 y3, tϪ1
c3 ϩ c6 y1, tϪ1 ϩ c9 y2, tϪ1 ϩ c12 y3, tϪ1
m
x, tϪ1 ϭ
.
~34!
΂
΃
“ARCH~2!” refers to the fact that the scale transformation Rx, tϪ1 is taken to
be of the ARCH~Lr!-form, with Lr ϭ 2,
t1 ϩ t76e1, tϪ1
6
t2
t4
ϩ t256e1, tϪ2
6
0
t3 ϩ t156e2, tϪ16
t5
R
x, tϪ1 ϭ
,
~35!
ϩ t336e2, tϪ2
6
΄
΅
0
0
t6 ϩ t246e3, tϪ16
ϩ t426e3, tϪ2
6
9
Our implementation of SMM with an auxiliary model differs from many previous imple-
mentations by our inclusion of the constant e0 in the SNP density function. Though e0 is iden-
tified if the scale of h~z6x! is fixed, Gallant and Long ~1997! encountered numerical instability
in estimating SNP models with e0 treated as a free parameter. Therefore, we chose to fix both
e0 and the constant term of h~z6x! at nonzero constants. We set e0 ϭ 0.01 and verified that the
estimated auxiliary model was essentially unchanged by this setting instead of zero.

1960
The Journal of Finance
where et ϭ yt Ϫ mx, tϪ1. Thus, the starting point for our SNP conditional
density for y is a first-order vector autoregression ~VAR!, with innovations
that are conditionally normal and follow an ARCH process of order two:
n~y6mx, ⌺x!, where ⌺x, tϪ1 ϭ Rx, tϪ1 Rx' , tϪ1
.
“Non-Gaussian” refers to the fact that the conditional density is obtained
by scaling the normal density n~zt! ~the “Gaussian, VAR~1!, ARCH~2!” part!
by the square of the Hermite polynomial h~zt 6xtϪ1!, where h is a polynomial
of order Kz ϭ 4 in zt, that is,
4
3
h~zt 6xtϪ1! ϭ A1 ϩ
3~lϪ1!ϩ1ϩi zil, t
.
~36!
( ( A
lϭ1 iϭ1
Finally, “Homogeneous-Innovation” refers to the fact that the coefficients in
the Hermite polynomial h~zt 6xtϪ1! are constants, independent of the condi-
tioning information.10
Having selected the moment conditions, we proceed using the standard
“optimal” GMM criterion function ~Hansen ~1982!, Duffie and Singleton ~1993!!,
a quadratic form in the sample moments
T
1
?
log f ~ytf6xtfϪ1, gT !.
~37!
(
T
?g
tϭ1
Under regularity, this SMM estimator is consistent for f, even if the auxil-
iary model does not describe the true joint distribution of yt. Efficiency con-
siderations, on the other hand, motivated our extensive search for among
semi-nonparametric specifications of f ~yt 6xtϪ1, g!. The analysis in Gallant
and Long ~1997! implies that, for our term structure model and selection
strategy for an auxiliary density f ~yt 6xtϪ1, g!, our SMM estimator is asymp-
totically efficient.11
The auxiliary model selected shows that the time series of LIBOR and
swap rates over the sample period we examined are remarkably “unevent-
ful”: a low order ARCH-like specification was able to capture the time vari-
ation in the conditional second moments. This contrasts with the findings in
the Treasury market by Andersen and Lund ~1997!, for example, who found,
for a different sample period, a GARCH-like or high order ARCH-like spec-
ification for the conditional variance.
10
Note that the term “Homogeneous-Innovation” does not imply that the conditional vari-
ances of the yields are constants ~because of the “ARCH~2!” terms!.
11
More precisely if, for a given order of the polynomial terms in the approximation to the
density f, sample size is increased to infinity, and then the order of the polynomial is increased,
the resulting SMM estimator approaches the efficiency of the maximum likelihood estimator. It
follows that our SMM estimator is more efficient ~asymptotically! than the quasi-maximum
likelihood estimator proposed recently by Fisher and Gilles ~1996!.

Affine Term Structure Models
1961
With A1 normalized to 1, the free parameters of the SNP model are
g ϭ ~Aj : 2 Յ j Յ 13; cj : 1 Յ j Յ 12;
~38!
t : j ϭ 1,2,...,7,15,24,25,33,42!.
j
B. Identification of the Market Prices of Risk
In Gaussian and square-root diffusion models of Y~t!, the parameters l
governing the term premiums enter the A~t! and B~t! in equation ~5! sym-
metrically with other parameters, and this leads naturally to the question of
under what circumstances l is identified in ATSMs. This section argues that
l is generally identified, except for certain Gaussian models.
The “identification condition” in GMM estimation is the assumption that
the expected value of the derivative of the moment equations with respect to
the model parameters have full rank ~Hansen ~1982!, Assumption 3.4!. To
simplify notation in our setting, we let zt' [ ~yt , xt'Ϫ1! and f ~zt, g! denote the
auxiliary SNP conditional density function used to construct moment condi-
tions. Using this notation, the rank condition for our SMM estimation prob-
lem is that the matrix
?2 log f
?g?f'
?2 log f
?g?ztf
?ztf
?f'
0
~ztf , g0! ϭ E
~ztf , g0!
~39!
0
0
D0 [ E
'
ͫ
ͬ ͫ
ͬ
has full rank. The rank of D0 is at most min~dim~g!, dim~f!!, so clearly a
necessary condition for identification is that dim~g! Ն dim~f!, where dim
denotes the dimension of the vector.
The market price of risk l ~or any other parameter! is not identified if
there exists another parameter, say d0 ʦ f, such that ?ztf0?l and ?ztf0?d0 are
collinear, in which case D0 clearly does not have full rank. For an example,
it is easy to check that in the case of the one-factor Gaussian model esti-
mated with a zero yield, the market price of risk is not identified because
the partial derivatives of the zero yield with respect to both l and d0 are
constant and therefore proportional to each other.
In a one-factor setting, there are two sources of identification of l. One is
the use of coupon bond yields instead of zero yields to estimate the model.
The nonlinear mapping between the coupon yield and the underlying ~Gauss-
ian or otherwise! state variable implies that both ?ztf0?l and ?ztf0?u are
state dependent and are not collinear. Another source of identification is the
assumption that the state variable follows a non-Gaussian process. In the
case of a square-root ~CIR! model, for example, it is easy to check that, even
though ?ztf0?l is a constant, none of the partial derivatives of ztf with re-
spect to the other structural parameters is a constant. Consequently, l is
identified.

1962
The Journal of Finance
This intuition for one-factor models can easily be generalized to the case of
N-factor affine models for N Ͼ 1. The identification results for the general
case are as follows:12 When zero yields are used to estimate a Gaussian
model ~in the A0~N ! branch!, one out of N market prices of risk is not iden-
tified.13 When zero yields are used to estimate non-Gaussian models ~in the
Am~N ! branch with 1 Յ m Յ N !, one out of N market prices of risk is not
identified unless, at least for one k with m ϩ 1 Յ k Յ N, bk is not identically
zero. When coupon yields are used to estimate affine models, all of the mar-
ket prices of risk are identified.
Intuitively, in cases where all market prices are identified, the common
source of identification is the stochastic or time-varying nature of risk pre-
mia, which can be induced either by a non-Gaussian model or a nonlinear
mapping between the observed yields and the underlying state variables.
When the risk premia are not time varying, they may not be separately
identified from the level of the unobserved short rate.14
C. Empirical Analysis of Swap Yield Curves
We estimated six ATSMs in A1~3! and A2~3! and report the overall goodness-
of-fit, chi-square tests for these models in Table I. Both the Chen and BDFS
models, denoted by A1~3!BDFS and A2~3!Chen, respectively, have large chi-
square statistics relative to their degrees of freedom. In contrast, the corre-
sponding maximal models, denoted by Am~3!Max, for m ϭ 1,2, are not rejected
at conventional significance levels. However, the improved fits of A1~3!Max
~compared to A1~3!BDFS! and A2~3!Max ~compared to A2~3!Chen! were achieved
with six and eight additional degrees of freedom, so we were concerned about
overfitting. This concern was reinforced by the relatively large standard er-
rors for most of the estimated parameters in the Max models, displayed in
the second columns of Tables II and III. Therefore, we also present the re-
sults for two intermediate models, A1~3!DS and A2~3!DS ~the DS indicating
that these are our preferred models!, that constrain some of the parameters
in the Max models. Relative to the more constrained BDFS and Chen mod-
els, the A1~3!DS and A2~3!DS models allow nonzero values of the parameters
~sru, sur! and ~kuv, krv, srv!, respectively. The DS models are not rejected at
conventional significance levels and have fewer parameters than the Max
models, and most of the estimated parameters are statistically significant at
conventional levels. Therefore, we will focus primarily on the DS models in
subsequent discussion.
12
Details are available from the authors upon request and from the Journal of Finance
website.
13
More precisely, d0 and the N prices of risk can not be separately identified. The fact that
there is only one underidentified market price of risk in an N-factor model with N Ϫ m Gaussian-
like factors may appear surprising at first. This is explained by the fact that we have con-
strained the long run means of the Gaussian-like factors to zero so that d0 is the only free
parameter that determines the overall level of the instantaneous short rate.
14
This intuition suggests that if the instantaneous short rate is observed, or if the model is
forced to exactly match the unconditional mean of an additional yield, an otherwise unidenti-
fied market price of risk can become identified.

Affine Term Structure Models
1963
Table I
Overall Goodness-of-Fit
This table lists the test statistics ~third column! for overidentifying restrictions obtained by
setting the simulated sample moments in equation ~37! to zero. The degree of freedom ~fourth
column! is the difference between the dimension of gT and that of f. The fifth column gives the
p-values for a chi-square distribution with the specified degrees of freedom. A1~3!Max, specified
in equation ~23!, is the “maximal” model in branch A1~3!. A1~3!BDFS is obtained by constraining
all parameters in the square boxes in equation ~23! to zero. A1~3!DS is our preferred model in
this branch, obtained by relaxing sur and sru from A1~3!BDFS. A2~3!Max, specified in equation
~28!, is the “maximal” model in branch A2~3!. A2~3!Chen is obtained by constraining all param-
eters in the square boxes in equation ~28! to zero, except that rS ϭ uN . A2~3!DS is our preferred
model in this branch, obtained by relaxing kuv and srv from A2~3!Chen. krv is not a free param-
eter but is nonzero. It has a deterministic relationship to kuv through their dependence on a
common degree of freedom, that is, k21 ~see equation ~A9!.
Branch A1~3!
BDFS
BDFS ϩ sur ϩ sru
Maximal
x2
d.f.
p-value
A1~3!BDFS
A1~3!DS
A1~3!Max
84.212
28.911
28.901
25
23
19
0.000%
18.328%
6.756%
Branch A2~3!
x2
d.f.
p-value
A2~3!Chen
A2~3!DS
A2~3!Max
Chen
129.887
22.931
16.398
26
24
18
0.000%
52.387%
56.479%
Chen ϩ kuv ϩ ~krv! ϩ srv
Maximal
The reason that the DS models do a better job “explaining” the swap rates,
as measured by the x2 statistics, than the A1~3!BDFS and A2~3!Chen models is
that the former allow a more flexible correlation structure of the state vari-
ables. In the A1~3! branch, the A1~3!BDFS model only allows a nonzero condi-
tional correlation between the short rate and its stochastic volatility ~srv Þ 0!.
The A1~3!DS model also allows the short rate and its stochastic central ten-
dency to be conditionally correlated ~sru Þ 0 and sur Þ 0!. ~Recall, from equa-
tion ~23!, that relaxing these constraints affects both the u~t! and r~t! processes.!
In the A2~3! branch, the A2~3!Chen model assumes that r~t!, u~t!, and v~t!
are all pairwise, conditionally uncorrelated. In contrast, the model A2~3!DS al-
lows the short rate to be conditionally correlated with its stochastic volatility
~srv Þ 0! and allows the stochastic volatility to influence the conditional mean
of the short rate ~krv Þ 0! and its stochastic central tendency ~kuv Þ 0!.15
Moreover, in both branches, it is the introduction of negative conditional
correlations among the state variables that seems to be important ~see the
third columns of Tables II and III!. Such negative correlations are ruled out
a priori in the CSR models ~family A3~3! in Section I!. Hence, these findings
support our focus on the branches A1~3! and A2~3! in attempting to describe
the conditional distribution of swap yields.
15
Though we relax three constraints, this amounts to two additional degrees of freedom,
because kuv and krv are controlled by the single parameter k21 in the AY representation, given
the constraint d1 ϭ 0. See equation ~A9!.

1964
The Journal of Finance
Table II
EMM Estimators for Branch A1(3)
The EMM estimators reported here and in Table III are obtained by minimizing the EMM
criterion function based on simulated sample moments in equation ~37!. The parameters in the
first column pertain to the Ar representation ~equation ~23!! of models in the AYM1~3! branch.
Parameters indicated by “fixed” are restricted to zero. t-ratios are in parentheses.
Parameter
Estimate ~t-ratio!
A1~3!DS
Ar
A1~3!BDFS
A1~3!Max
m
n
krv
k
vS
0.602 ~4.246!
0.0523 ~7.138!
0 ~fixed!
2.05 ~9.185!
0.000156 ~6.630!
0.14 ~8.391!
0 ~fixed!
491 ~3.224!
0 ~fixed!
0 ~fixed!
0.000113 ~9.493!
0 ~fixed!
0.365 ~6.981!
0.226 ~14.784!
0 ~fixed!
17.4 ~3.405!
0.015 ~1.544!
0.0827 ~16.044!
0 ~fixed!
4.27 ~2.645!
Ϫ3.42 ~Ϫ1.754!
Ϫ0.0943 ~Ϫ2.708!
0.0002 ~4.069!
0 ~fixed!
0.00782 ~1.382!
0 ~fixed!
Ϫ0.344 ~Ϫ0.064!
31.7 ~2.410!
9.32 ~2.296!
0.366 ~5.579!
0.228 ~9.720!
0.0348 ~0.001!
18 ~3.547!
0.0158 ~1.323!
0.0827 ~13.530!
0.0212 ~0.012!
4.2 ~1.936!
uN
suv
srv
sru
sur
z2
ar
h2
bu
lv
lu
lr
Ϫ3.77 ~Ϫ1.668!
Ϫ0.0886 ~Ϫ2.470!
0.000208 ~2.742!
3.26eϪ14 ~0.000!
0.00839 ~1.222!
7.9eϪ10 ~0.000!
Ϫ0.27 ~Ϫ0.045!
30.2 ~1.289!
5.18eϪ05 ~1.934!
0 ~fixed!
6.79eϩ04 ~1.886!
31 ~3.392!
1.66eϩ03 ~3.312!
9.39 ~1.153!
Figure 2 displays the t-ratios for testing whether the fitted scores of the
auxiliary model, computed from the models A1~3!BDFS, A1~3!DS, A2~3!Chen
,
and A2~3!DS, are zero. For a correctly specified model ~and assuming that
asymptotic approximations to distributions are reliable!, the sample scores
should be small relative to their standard errors. The graphs for models
A1~3!BDFS and A2~3!Chen show that about half of the fitted scores have large-
sample t-ratios larger than 2. In contrast, only two of the t-ratios are larger
than 2 for model A1~3!DS, and none are larger than 2 for model A2~3!DS
.
The first 12 scores of the auxiliary model, marked by “A” near the hori-
zontal axis, are associated with parameters that govern the non-normality of
the conditional distribution of the swap yields. The second 12 scores, marked
by “⌿,” are related to parameters describing the conditional first moments,
and the last 12 scores, marked by “t,” are related to the parameters of the
conditional covariance matrix. A notable feature of the t-ratios for the indi-
vidual scores is that they are often large for models A1~3!BDFS and A2~3!Chen
for all three groups A, ⌿, and t. Thus, the additional nonzero conditional
correlations in the DS models help explain not only the conditional second
moments of swap yields but their persistence and non-normality also.
The estimated values of the parameters for the Ar representations of the
models are displayed in Tables II and III. Though perhaps not immediately
evident from the Ar representations, the DS models maintain the constraint

Affine Term Structure Models
1965
Table III
EMM Estimators for Branch A2(3)
The parameters pertain to the Ar representation ~equation~28!! of models in the AYM2~3! branch.
Parameters indicated by “fixed” are restricted to zero except that rS is constrained to be equal to
uN , whenever it is “fixed.” t-ratios are in parentheses.
Parameter
Estimate ~t-ratio!
A1~3!DS
Ar
A1~3!Chen
A1~3!Max
m
1.24 ~4.107!
0 ~fixed!
0.636 ~4.383!
Ϫ33.9 ~Ϫ2.377!
Ϫ35.3 ~Ϫ2.367!
0 ~fixed!
0.103 ~2.078!
2.7 ~7.432!
0.291 ~1.648!
Ϫ12.4 ~Ϫ1.132!
Ϫ274 ~Ϫ1.077!
Ϫ0.0021 ~Ϫ0.411!
0.0871 ~1.027!
3.54 ~4.457!
0.000315 ~2.051!
0.0136 ~0.994!
0.053 ~2.784!
Ϫ133 ~Ϫ2.438!
Ϫ0.0953 ~Ϫ0.167!
1.12eϪ09 ~0.000!
7.04eϪ05 ~1.679!
0.00237 ~0.002!
1.92eϪ05 ~0.002!
7.58eϩ03 ~1.064!
Ϫ174 ~Ϫ1.050!
Ϫ349 ~Ϫ1.540!
kuv
krv
kvu
n
k
vS
0 ~fixed!
0 ~fixed!
0.0757 ~4.287!
2.19 ~8.618!
0.000206 ~7.456!
0.0416 ~7.909!
0.0416 ~fixed!
0 ~fixed!
0.000239 ~5.792!
0.0259 ~4.006!
0.0259 ~fixed!
Ϫ182 ~Ϫ3.620!
0 ~fixed!
uN
rS
srv
sru
ar
h2
z2
bu
lv
lu
lr
0 ~fixed!
0 ~fixed!
0 ~fixed!
0.000393 ~2.873!
0.00253 ~7.507!
0 ~fixed!
Ϫ1.9eϩ03 ~Ϫ3.946!
Ϫ35.2 ~Ϫ5.080!
Ϫ121 ~Ϫ4.973!
0.000119 ~2.083!
0.00312 ~3.129!
0 ~fixed!
1.3eϩ04 ~1.761!
Ϫ152 ~Ϫ2.923!
Ϫ692 ~Ϫ4.066!
from the A1~3!BDFS and A2~3!Chen models that d1 ϭ 0. That is, in the AY
representation, the instantaneous riskless rate is an affine function of only
the second and third state variables. The test statistics in Table I suggest
that this constraint is not inconsistent with the data.
The estimates of the mean reversion parameters ~m, n, k! of the state vari-
ables ~v~t!, u~t!,r~t!! are ~0.37,0.23,17.4! and ~0.64,0.10,2.7! for models A1~3!DS
and A2~3!DS, respectively. As in previous empirical studies ~e.g., Balduzzi
et al. ~1996! and Andersen and Lunk ~1998!, the “central tendency” factor
u~t! shows much slower mean reversion ~smaller n! than the rate k at which
gaps between u and r are closed in the short rate equation. Put differently,
in model A1~3!DS, r~t! reverts relatively quickly to a process u~t! that is
itself reverting slowly to a constant long-run mean uN .16 In both DS models,
the “central tendency” factor has the smallest mean reversion.
An important cautionary note at this juncture is that comparisons across
models of mean reversion coefficients ~or, more generally, coefficients of the
drifts! may not be meaningful even if the models are nested. The reason is
that changing the correlations among the state variables can be thought of
as a “rotation” of the unobserved states Y~t!. Therefore, the meaning of la-
bels such as “central tendency” or “volatility” in terms of yield curve move-
16
This interpretation does not hold exactly in model A2~3!DS, because krv is nonzero.

1966
The Journal of Finance
Figure 2. Fitted SNP Scores. Each subpanel plots the fitted SNP scores for one of the four
models, normalized by their standard deviations. The fitted scores are the simulated sample
moments given in equation ~37!, evaluated at the EMM estimators. Each normalized score has
an asymptotic standard normal distribution. The horizontal axes are the SNP parameters g.
The first group of 12 is designated by the letter “A,” the middle group of 12 is designated by the
letter “⌿,” and the last group of 12 is designated by the letter “t.” The meanings of the param-
eters and the interpretations of the t-ratios are discussed in Section II.C.
ments may not be the same across models. To illustrate this point, consider
the models in A1~3!. In the model A1~3!BDFS, the correlation between changes
in u~t! and changes in the 10-year swap rate is 0.98. The close association
between the long-term swap rate and central tendency is intuitive, because
r~t! mean reverts to u~t!. Nevertheless, this interpretation is not invariant
to relaxation of the constraints sur ϭ 0 and sru ϭ 0, which gives model A1~3!DS
.
In the latter model, changes in u~t! are most highly correlated with changes
in the two-year swap rate ~correlation ϭ 0.95!. Because the two- and 10-year
swap rates differ in their persistence, this explains the larger value of n
17
~faster mean reversion of u~t!! in model A1~3!DS than in model A1~3!BDFS
.
17
Similar observations apply to the volatility factor v~t!. In both models, v~t! is well proxied
by a butterfly position that is long 10-year swap and LIBOR contracts and short two-year
contracts. However, the weights in these butterflies turn out to be quite different across the
models.

Affine Term Structure Models
1967
Table IV
Moments of Pricing Errors (in basis points)
“Mean” ~second column! is the sample mean of the pricing errors for the swap yields and models
indicated in the first column. “Std” ~third column! is the sample standard deviation and “r”
~fourth column! is the first-order autocorrelation of the pricing errors. The columns labeled
Q-Invert and Q-Steep display the sample means of the pricing errors for the days on which the
slope of the yield curve was in the lowest ~inverted! and highest ~steep! quartile of its distri-
bution.
Model0Swap
A1~3!DS
Mean
Std.
r
Q-Invert
Q-Steep
3 yr
5 yr
7 yr
Ϫ11.3
16.9
Ϫ12.7
9.6
16.5
10.1
0.95
0.97
0.94
Ϫ8.1
Ϫ12.0
Ϫ9.1
Ϫ16.6
Ϫ26.6
Ϫ17.6
3 yr
5 yr
7 yr
Ϫ43.1
Ϫ63.3
Ϫ47.5
11.6
12.1
8.3
0.97
0.96
0.94
Ϫ55.4
Ϫ75.6
Ϫ54.1
Ϫ27.7
Ϫ49.1
Ϫ38.1
A2~3!DS
Does the evidence recommend one of the intermediate models, A1~3!DS or
A2~3!DS, over the other? Ultimately, the answer to this question must de-
pend on how the models will be used ~e.g., risk management, pricing options,
etc.!. Even within the term structure context, these models are not nested,
so formal assessments of relative fit are nontrivial. However, we offer sev-
eral observations that suggest that, focusing on term structure dynamics
within the affine family, model A1~3!DS provides a somewhat better fit. Con-
sider first the properties of the time series of pricing errors. Table IV presents
the within-sample means, standard deviations, and first-order autocorrela-
tions of the pricing error for the yields on swaps with the three intermediate
maturities three, five, and seven years, none of which were used in estimat-
ing the parameters.18 Model A1~3!DS has notably smaller average pricing
errors than model A2~3!DS, though both models have a tendency to imply
higher yields than what we observed.
Second, the feedback effect in the drift due to krv Þ 0 and kuv Þ 0 in model
A2~3!DS is also accommodated by model A1~3!DS. However, the results for
model A1~3!DS suggest that nonzero values of these ks are not essential for
fitting the moments of swap yields used in estimation, once sur and sru are
allowed to be nonzero. Within model A2~3!DS, the admissibility conditions
preclude relaxation of the constraint sur ϭ 0, because of its richer formula-
tion of conditional volatility. Admissibility also requires that kuv ~and there-
fore krv! be negative. Consequently, the stochastic central tendency and the
stochastic volatility must have a positive unconditional correlation. Taken
together, these findings suggest that the negative correlations among the
state variables called for by the data are not easily accommodated within
the A2~3! branch.
18
These pricing errors were computed by inverting the models for the implied values of the
state variables, using the six-month and two- and 10-year swap yields, and then computing the
differences between the actual and model-implied swap rates for the intermediate maturities,
with the latter evaluated at the implied state variables.

1968
The Journal of Finance
Figure 3. Term Structure of Yield Volatility. The volatility here is defined as the sample
standard deviation of weekly changes of LIBOR and swap yields ~see Figure 1!. Except for the
observed term structure of volatility ~line!, which is computed from the observed yields, the
other four curves are computed from yields simulated from the four models, using EMM esti-
mators reported in Table II and Table III.
Third, we also examined the shapes of the implied term structures of ~un-
conditional! swap yield volatilities. The solid, uppermost line in Figure 3
displays the historical sample standard deviations of differences in yields.
The other lines display the sample variances computed using long, simu-
lated time series of swap yields from the models evaluated at their esti-
mated parameter values.19 Notably, the term structure of historical sample
volatilities is hump-shaped, with a peak around two years. Hump-shaped
volatility curves can be induced in ATSMs either through negative correla-
tion among the state variables or by hump-shaped loadings B~t! on Y~t! in
equation ~5!. The intuition for this lies in the interplay between the negative
correlations among the shocks to the risk factors and different speeds of
mean reversion of the state variables. For expository ease, consider the case
19
We stress that the model-implied volatilities were computed by simulation and not from
yields computed with the implied state variables. Thus, Figure 3 displays the population vol-
atilities implied by the models, conditional on the estimated parameter values. We have found
that using implied swap yields to compute sample moments often leads to substantially biased
estimates of the population values.

Affine Term Structure Models
1969
of two factors where the first factor has a faster rate of mean reversion
~larger k! than the second. In affine models, k plays a critical role in the rate
at which the factor weights ~the B~t! in equation ~7!! tend to zero as matu-
rity t is increased. At short maturities, the volatilities of both factors will
typically affect overall yield volatility. As t increases, the influence of the
first factor will die out at a faster rate than that of the second factor. Thus,
for long maturities, yield volatility will be driven primarily by the second
factor and volatility will decline with maturity. A hump can occur, because
the negative correlation contributes to a lower yield volatility at the shorter
maturities. As maturity increases, the negative contribution of correlation to
yield volatility declines as the importance of the first factor declines. That
models with independent, mean-reverting state variables cannot induce a
hump can be seen from inspection of the loadings implied by the CIR model.
Models in A1~3! and A2~3! can exploit both of these mechanisms to match
historical volatilities ~whereas models in A3~3! only have the latter mecha-
nism!.20 All of the model-implied, volatility term structures in Figure 3 have
a hump. However, model A1~3!DS appears to fit the volatility of swap yields
much better than model A2~3!DS
.
Finally, when we computed the implied yield curves from model A2~3!DS
,
we found that there were often pronounced “kinks” at the short end of the
yield curve, whereas those implied by model A1~3!DS were generally smooth.
In light of the small goodness-of-fit statistics for the A2~3!DS model, we were
puzzled by the frequency of kinks in yield curves, the large average pricing er-
rors, and the underestimation of yield volatilities. The preceding discussion of
the constraints on the conditional correlations implied by the admissibility con-
ditions, together with inspection of the form of the risk-neutral drifts, leads us
to the following conjecture: the market prices of risk were set, in part, to rep-
licate the effects of a nonzero sur ~which cannot be done directly! at the ex-
pense of sensible shapes of implied yield curves and smaller pricing errors. To
explore the validity of this conjecture, we simply reduced the market prices of
risk by 20 percent in absolute value in model A2~3!DS and found that the im-
plied yield curves were essentially free of kinks and, equally importantly, seemed
to line up well with the historical yield curves.21
There is also evidence that all of the models examined fail to capture some
aspects of swap yield distributions. In particular, in Table IV, columns 5 and
6, we report the average pricing errors for dates when the slope of the swap
curve was in the lowest ~“Q-Invert”! and highest ~“Q-Steep”! quartiles of the
historically observed slopes.22 In the case of model A1~3!DS, the average pric-
ing errors are larger when the swap curve is steeply upward sloping than
when it is inverted. The reverse is true for model A2~3!DS. This suggests
20
These observations provide further motivation for our interest in the branches A1~3! and
A2~3!.
21
The criterion function used in estimation does not impose a penalty for kinks in spot
curves or choppy forward-rate curves. Such penalties could, of course, be introduced in practice.
Nor does the criterion function force the means of the swap rates observed historically and
simulated from the models to be the same.
22
Slope is the difference between the 10- and two-year swap yields.

1970
The Journal of Finance
that there may be some omitted nonlinearity in these affine models.23 Also,
though the standard deviations of the pricing errors are small relative to
those of the swap yields themselves, the errors are highly persistent ~see
column 4 of Table IV!. Such persistence points to some misspecification of
the model for intermediate maturities.
III. Conclusion
In this paper we present a complete characterization of the admissible and
identified affine term structure models, according to the most general known
sufficient conditions for admissibility. For N-factor models, there are N ϩ 1
non-nested classes of admissible models. For each class, we characterize the
“maximally flexible” canonical model and the nature of the admissible factor
correlations and conditional volatilities that these canonical models can
accommodate.
Our empirical analysis of the family of three-factor affine models leads us to
the following conclusions about their empirical properties. First, across a wide
variety of parameterizations of ATSMs, the data consistently called for nega-
tive conditional correlations among the state variables. Such correlations are
precluded in multifactor CIR models and, therefore, this finding would not have
been directly evident from previous empirical studies of these models.24 Non-
zero conditional correlations are also precluded in the affine version of the Chen
model, and only limited nonzero correlations were permitted in the BDFS model.
The empirical results from the A1~3! and A2~3! branches suggests that the lim-
ited correlation accommodated by these models largely explains the associated
large chi-square statistics. The importance of negative correlation may not have
been more apparent from previous empirical studies, because many of these
studies used data on the short rate alone to estimate multifactor models. We
find that a key role of the factor correlations is in explaining the shape of the
term structure of volatility of bond yields, and this is revealed most clearly
through the simultaneous analysis of long- and short-term bond yields.
Second, within the affine family, we demonstrated an important trade-off be-
tween the structure of factor volatilities on the one hand and admissible non-
zero conditional correlations of the factors on the other. The “maximally flexible”
models in A1~3! give the most flexibility in specifying conditional correlations,
while still accommodating some time-varying volatility ~in this case driven by
one factor, m ϭ 1!. Models in A2~3! offer more flexibility in specifying time-
varying volatility ~as it may depend on two factors, m ϭ 2!, but admissibility
requires a relatively more restrictive correlation structure. For our data set on
dollar swap yields and our sample period, we find that flexibility with regard
23
In a one-factor setting, Ait-Sahalia ~1996! finds evidence for nonlinearity in the drifts of
short rates, although some recent work such as Chapman and Pearson ~1999! suggests that
such evidence needs to be interpreted with caution. Boudoukh et al. ~1998! also provide evi-
dence for a nonlinear relationship between slope and level in a two-factor setting.
24
As noted in Section I, our finding was implicitly present in studies of CIR models, because
the correlations among the implied state variables are strongly negative, in contrast to the
implications of the models.

Affine Term Structure Models
1971
to the specification of correlations is more important than flexibility in spec-
ifying volatilities. Overall, the preferred A1~3!DS model in A1~3! seems to fit
better than the preferred model A2~3!DS in A2~3!. We conjecture that, for other
markets or different sample periods, where conditional volatility is much more
pronounced in the data, the relative goodness-of-fit of models in the branches
A1~3! and A2~3! may change. Indeed, our analysis suggests that the affine fam-
ily of term structure models will have the most difficulty describing interest
rate behavior in settings where conditional volatility is pronounced and the fac-
tors are strongly negatively correlated.
Though extant ATSMs have evidently failed to capture key features of
historical dollar interest rate behavior, many basic features of these models
are supported by our empirical analysis. Specifically, both the A1~3!DS and
A2~3!DS models build on the literature that posits a short-rate process with
a stochastic central tendency and volatility. The ~implicit! restriction in these
three-factor models that the short rate can be expressed as an affine func-
tion of only two of the three factors is supported by the empirical evidence,
and, hence, this restriction is imposed in the DS models. Furthermore, con-
sistent with the analysis of a central tendency factor in Andersen and Lund
~1998! and Balduzzi et al. ~1996!, we find that the short rate tends to mean-
revert relatively quickly to a factor that itself has a relatively slow rate of
mean reversion to its own constant long-run mean. However, in our pre-
ferred model A1~3!DS, neither of the two factors determining the short rate,
besides itself, is literally interpretable as the central tendency of r. Also,
even though factors may “look” like central tendency or volatility, the mean-
ing of these constructs in terms of their induced changes in the shapes of
yield curves varies substantially across specifications of the factor correlations.
Finally, all of the models examined in this paper presume that the market
prices of risk are proportional to the volatilities of the state variables. Two
important, and potentially restrictive, implications of this formulation are
that the state variables follow affine diffusions under both the actual and
risk-neutral probabilities and that the signs of the market prices of risk do
not change over time. The evidence suggests that our formulation of risk
premiums may underlie the difficulty we found in matching the sample mo-
ments of swap yields, particularly with model A2~3!DS. Nonlinear formula-
tions of the risk premiums ~including formulations with time-dependent signs!
can be accommodated directly within the affine framework, as long as the
state variables follow affine diffusions under the risk-neutral distribution.
The empirical significance of such extensions of the affine framework ex-
plored here is an interesting topic for future research.
Appendix A. Invariant Transformations
In defining the class of admissible ATSMs, it will be necessary to undertake
various transformations and rescalings of the state and parameter vectors in
ways that leave the instantaneous short rate, and hence bond prices, un-
changed. We refer to such transformations as “invariant transformations.” More
precisely, consider an ATSM with state vector Y~t!, Brownian motions W~t!,

1972
The Journal of Finance
and parameter vector f ϭ ~d0, dy,K, ⌰, ⌺,$ai, bi : 1 Յ i Յ N, l!. An invariant af-
fine transformation TA is defined by an N ϫ N nonsingular matrix L and an
N ϫ 1 vector q, such that TAY~t! ϭ LY~t! ϩ q, TA f ϭ ~d0 Ϫ dy' LϪ1q, L'Ϫ1dy,
LKLϪ1, q ϩ L⌰, L⌺,$ai Ϫ bi' LϪ1q, L'Ϫ1bi : 1 Յ i Յ N, l! are the state vector and
the parameter vector, respectively, under the transformed model. The Brown-
ian motions are not affected. Such transformations are generally possible, be-
cause of the linear structure of ATSMs and the fact that the state variables are
not observed. A diffusion rescaling TD rescales the parameters of @S~t!#ii and
the ith entry of l by the same constant. That is, for any N ϫ N nonsingular
matrix D, TDf ϭ ~d0, dy,K, ⌰, ⌺DϪ1,$Di2i ai , Di2i bi : 1 Յ i Յ N, Dl! is the param-
eter vector for the transformed model. The state vector and the Brownian mo-
tions are not affected. Such rescalings may be possible, because only the
combinations ⌺S~t!⌺' and ⌺S~t!l enter the pricing equations ~5!, ~6!, and ~7!.
ABrownian motion rotation TO takes a vector of unobserved, independent Brown-
ian motions and rotates it into another vector of independent Brownian mo-
tions. That is, for any NϫN orthogonal matrix O ~i.e., OϪ1 ϭOT! that commutes
with S~t!, TOW~t! ϭ OW~t! and TOf ϭ ~d0, dy,K, ⌰, ⌺OT,$ai, bi : 1 Յ i Յ N,Ol!
are the Brownian motions and the parameter vector, respectively, for the trans-
formed model. The state vector is not affected. Finally, a permutation TP sim-
ply reorders the state variables, which has no observable consequences. It is
easily checked that any two ATSMs linked by any combination of the above in-
variant transformations are equivalent in the sense that the implied bond prices
~including the short rate! and their distributions are exactly the same.
B. Admissibility of the Canonical Model
For an arbitrary affine model, deriving sufficient conditions for admissi-
bility is complicated by the fact that admissibility is a joint property of the
drift ~K and ⌰! and diffusion ~⌺ and B! parameters in equation ~9!. A key
motivation for our choice of canonical representations is that we can treat
the drift and diffusion coefficients separately in deriving sufficient condi-
tions for admissibility. Therefore, verification of admissibility is typically
straightforward. In this Appendix, we provide sufficient conditions for our
canonical representation of Am~N ! to be well defined.
The canonical representation of Am~N ! has the conditional variances of
the state variables controlled by the first m state variables:
Sii~t! ϭ Yi~t!, 1 Յ i Յ m,
~A1!
m
Sjj~t! ϭ aj ϩ
@bj #k Yk~t!, m ϩ 1 Յ j Յ N,
~A2!
(
kϭ1
where aj Ն 0, @bj#i Ն 0.25 Therefore, as long as Y B~t! [ ~Y1,Y2,...,Ym!' is
non-negative with probability one, the canonical representation of Y~t! ϭ
'
'
~Y B ~t!,Y D ~t!!', where Y D~t! [ ~Ymϩ1,Ymϩ2,...,YN!, will be admissible.
25
Any model within Am~N ! can be transformed to an equivalent model with this volatility
structure through an invariant transformation.