Term Structure and Volatility:

Lessons from the Eurodollar Markets

Ruslan Bikbov
Columbia Business School

Mikhail Chernov
Columbia Business School

First Draft: August 2002

This Revision: July 7, 2004

We would like to thank Andrew Ang, Darrell Due, Steve Figlewski, Mike Johannes, Sergei Levendorskii,
Tano Santos, Ken Singleton as well as participants at several Columbia workshops for helpful comments.

Division of Finance and Economics, 6T Uris Hall, 3022 Broadway, New York, NY 10027, USA, Email:
rb2015@columbia.edu

Division of Finance and Economics, 413 Uris Hall, 3022 Broadway, New York, NY 10027, USA, Phone:
(212) 854-9009, Fax: (212) 316-9180, Email: mc1365@columbia.edu
Term Structure and Volatility:
Lessons from the Eurodollar Markets
Abstract
We evaluate the ability of several ane models to explain the term structure of the interest rates
and option prices. Since the key distinguishing characteristic of the ane models is the specication
of conditional volatility of the factors, we explore models which have critical dierences in this respect:
Gaussian (constant volatility), stochastic volatility, and unspanned stochastic volatility models. We
estimate the models based on the Eurodollar futures and options data. We nd that both Gaussian and
stochastic volatility models, despite the dierences in the specications, do a great job matching the
conditional mean and volatility of the term structure. When these models are estimated using options
data, their properties change, and they are more successful in pricing options and matching higher
moments of the term structure distribution. The unspanned stochastic volatility (USV) model fails to
resolve the tension between the futures and options ts. Unresolved tension in the ts points to addi-
tional factors or, even more likely, jumps, as ways to improve the performance of the models. Our results
indicate that Gaussian and stochastic volatility models cannot be distinguished based on the yield curve
dynamics alone. Options data are helpful in identifying the dierences. In particular, Gaussian models
cannot explain the relationship between implied volatilities and the term structure observed in the data.
1 Introduction
Ane Term Structure Models (ATSM) have been at the center of xed income research for over
a decade.
1
The popularity of these models is not hard to understand: their tractability and
exibility have proven useful in many diverse applications.
2
An important remaining issue is
the particular specication of the factors driving the variation in the yield curve. The only way
ATSM dier from each other is in their specication of the conditional volatility of the factors.
Thus, it is natural to assume that they would have dierent implications for the volatility of the
yield curve dynamics. In this paper, we show that, despite the initial appearance, these models
in practice produce very similar behavior of the volatility of the yield curve. Therefore, other
data should be used to dierentiate these models. Option prices are such source of additional
information.
We consider three most promising three factor models: (i) Gaussian, or multi-factor Va-
sicek, model (Langetieg, 1980); (ii) an extension of the BDFS (Balduzzi, Das, Foresi, and Sun-
daram, 1996) stochastic volatility model; and (iii) the unspanned stochastic volatility (USV)
model (Collin-Dufresne and Goldstein, 2002 CDG henceforth). These models have critically
dierent specications for the volatility of the spot rate process. Since the models dier in the
volatility specication, we depart from most existing studies by incorporating option prices,
which are by their nature sensitive to volatility, into our analysis.
3
We use a rich dataset based on the Eurodollar futures and options. These data have a
number of unique qualities which makes them more attractive than traditional data sources for
1
Due and Kan (1996) is the rst full analysis of these models. Dai and Singleton (2000) provide the ATSM
classication and sucient conditions for identication and existence of solutions. Piazzesi (2005) provides a
comprehensive review.
2
See, among others, Ang and Piazzesi (2003) in macroeconomics, Buraschi and Jiltsov (2004) in general
equilibrium, and Singleton and Umantsev (2002) in option pricing.
3
Umantsev (2001) is the only existing study that uses truly joint dataset for empirical analysis. Also see
Attari (2001), Chen and Scott (2002), Heidari and Wu (2002), Jagannathan, Kaplin, and Sun (2003) for other
studies using options data.
1
term structure studies. These markets are extremely liquid (the daily dollar turnover exceeds
that of the swaps markets tenfold) and provide actual traded prices as opposed to approxima-
tions, which are typically used in the U.S. Treasury-based studies.
4
Quarterly maturity cycle
of the futures allows us to use many points along the term structure.
Since it is not hard for any three-factor model to t the cross-section of yields (see the
principal component analysis of Litterman and Scheinkman, 1991), we concentrate on the time
series performance of the models of interest. We relax the traditional tight link between the
cross section and the time series of yields by not imposing restrictions which arbitrarily require
certain futures or option prices to be observed exactly. We use quasi-maximum likelihood with
Kalman lter for model estimation because the state variables cannot be inverted from prices
in this setting.
5
There is always a concern about the validity of asymptotic inference because of
the persistence of interest rates. For this reason, we use parametric bootstrap (Conley, Hansen,
and Liu, 1997) to compute standard errors and implement tests throughout the paper.
We nd that irrespective of the dataset used futures only or both futures and options
the Gaussian and BDFS models match identically many important aspects of the data, such
as pricing errors, or conditional means and volatilities of the yield curve. Moreover, we show
that while conditional volatilities of futures implied by the two models are highly correlated,
the associated underlying model factors are not. This result emphasizes that specication of
the interest rate volatility may not necessarily have a direct connection with the variability in
yields in practice. We show that the latent nature of the state variables and the yield data
(un)informativeness account for the ndings.
Since the models have important theoretical dierences, there must be a dimension along
which they dier in practice. One of our tests establishes that options data are required if one
wants to understand the dierences between the two models. Indeed, we nd that, once the
4
Dai, Singleton, and Yang (2004) discuss potential pitfalls of using spline methods to construct the zero yield
curve.
5
There is a number of subtle properties of the Eurodollar futures which make them attractive as a data
source for the employed estimation approach. We point these out throughout the text when relevant.
2
models are estimated based on the joint futures and options data, they produce dierent
matches to higher moments of the yields distribution and have dierent implications for the
relationship between option prices and the term structure studied by CDG.
Indeed, we nd an R
2
of about thirty percent when regressing implied volatilities on the
yields in line with CDG. However, only stochastic volatility models can replicate this nding.
Gaussian models produce an R
2
, which is close to zero. This dierence between the two
model classes is drastic, in the light of their similarity in many other respects. The dierence
is explained by the strong sensitivity of the implied volatility to the square-root factor. In
contrast, none of the (conditionally) Gaussian factors in either model class are related to the
implied volatility values.
Incorporating the options data into estimation also identies a tension between models
abilities to t term structure versus options, and to t lower order moments versus higher
order moments. One promising way to relax this tension is to use the USV model. The USV
restrictions on the BDFS model simultaneously weaken the term structure dependence on spot
volatility, and free up that same volatility factor to focus on option tting. We nd, consistent
with the motivation, that the factors of the USV model estimated from the joint futures and
options dataset coincide with the intuitive interpretation assigned to them at the specication
stage. In particular, the stochastic volatility factor is highly correlated with option-implied
volatility.
6
However, we formally reject the USV model in favor of the unrestricted BDFS
model. We also show that USV models fail along a number of important dimensions, such as
pricing errors, and ability to match conditional mean and variance.
In a recent paper, Collin-Dufresne, Goldstein, and Jones (2003) (CDGJ henceforth)
discuss related issues by contrasting the BDFS and USV models. They compare these two
models based on the swap data only. Their estimation approach allows stochastic volatility
to be a latent, or non-invertible, factor (it is impossible to study the USV model otherwise),
however the other two factors are inverted from the zero yields interpolated from the swap
6
This result is not obvious since we do not invert factors from some exactly observed prices, but lter them.
3
rates. In contrast to our results, CDGJ nd virtually no dierence between the USV and
BDFS models. Since the volatility factor from the USV model is highly correlated with the
GARCH estimate of volatility, while the one from BDFS is not, CDGJ select USV as the most
successful model. We argue that, while this intuitive feature is appealing, the most important
metric of models evaluation pertains to their ability to generate accurate cross-sectional and
dynamic representations of the data. Our results clearly show that USV fails in this respect.
The paper is organized into six sections. The next section describes the essence of
the debate on the volatility modelling. Section 3 describes the institutional features of the
Eurodollar markets, which we nd important for our study, and describes the specics of the
dataset. Section 4 develops ane models of Eurodollar futures and options prices and describes
our econometric approach. Section 5 discusses all of the results. The nal section makes
concluding observations on the ndings. All technical details are delegated to six appendixes.
2 Interest rate volatility and the term structure
Since the specication of conditional volatility of the factors is the characteristic which allows
researchers to distinguish various ATSM, they dedicate a lot of work to estimating the volatility
factors. The standard assumption used at the estimation stage that three yields are observed
exactly, and, therefore, the latent factors can be inverted from these yields. Since the yields
have a linear relationship with the factors, one can compute the volatility factor v as a linear
combination of the three yields (CDGJ and Piazzesi, 2005 make related points).
This observation has two implications. First, it is not appropriate to compare the em-
pirical properties of v computed from the yields cross-section to the variance of yields because
the former estimates the instantaneous volatility of the factors. Second, it is not clear that v
computed from the cross-section of yields is guaranteed to be the instantaneous variance of the
spot rate r. If a model is misspecied, assuming exactly observed yields could lead to severe
errors, a negative v in particular. These implications call for careful distinction between time
4
series and cross-sectional information in the term structure data, and for reconsideration of
the traditional estimation strategies.
Indeed, we know that it is not hard to t the cross-section of yields with a three factor
model, especially when one uses the factor inversion techniques. Inverting factors from a
misspecied model by construction implies time-series properties of state variables which are
inconsistent with the model specication. Hence, we do not have much evidence on the time-
series properties of the competing models. The notable exception are papers which use the
Ecient Method of Moments (EMM, Gallant and Tauchen, 1996) for estimation because it
does not require the inversion.
7
Reprojection (Gallant and Tauchen, 1998) is the only EMM
diagnostic which allows to study the time-series properties of models.
8
As a result, descriptive
studies focussing exclusively on the properties of short interest rate are more informative about
the time-series, but typically do not contrast multiple models.
9
A natural way to relax the tension between the time series and cross-sectional implica-
tions of ane models is to assume that all yields are observed with an error.
10
Suppose, we
observe the yields, y, and model them as a function of the factors S. As a result, the yields
can be represented by the so-called measurement equation:
y
t
() = g(S
t
, , ) +
t
(2.1)
where the function g is dictated by a no-arbitrage model (it is linear in the ane case), is
the vector of the model parameters, is time to maturity, and represents the observation
errors. Because of such a relationship, the factors are truly latent. This strategy does not
impose a strong cross-sectional relationship between the variance of the spot rate and the
7
In practice, EMM applicability is still limited to three yields. See, for instance, Dai and Singleton (2000),
and Ahn, Dittmar, and Gallant (2002) among others.
8
Also, see Duee (2002) for related points.
9
For instance, At-Sahalia (1996), Ball and Torous (1999), Johannes (2004).
10
Thompson (2004) is concerned with the time series versus cross-section tension as well. He resolves this
issue by eectively introducing more state variables than observable yields (he assumes that only three yields
are observable).
5
cross-section of yields. Moreover, the measurement equation, combined with the maximum
likelihood estimation based on the dynamics of the state variables (the state equation), forces
the model to t the time series aspect of the data.
11
Even if the data is not very informative
about the variance v, this strategy still provides the best way to identify this state variable.
In addition to the improved methodology, one could rely on additional data sources.
Indeed, Litterman, Scheinkman, and Weiss (1991) were able to establish the link between the
term structure curvature and volatility once they used implied volatility from the options on
T-bond futures. Heidari and Wu (2003) show that variability of swaption implied volatility
contains signicant information beyond that in the standard three principal components ex-
tracted from the yield curve. This evidence indicates that derivatives are a powerful source of
additional information about the data generating process.
Unfortunately, there is not much work done on xed-income derivatives in the time series
context. The existing studies focus on dierent modelling aspects and, therefore, consider
models from dierent branches of the ane class.
12
As a result, while these papers provide
important insights regarding the capability of various ane models to explain derivative prices,
the provided evidence is hardly conclusive.
In particular, Umantsev (2001) makes a strong case for the ability of conventional three-
factor models he evaluates the A
1
(3) class to generate reasonable volatility patterns: as
long as swaptions data are used at the estimation stage, an ane model will adjust to generate
volatility estimates closely aligned with volatility implied from derivatives. In sharp contrast,
Jagannathan, Kaplin, and Sun (2003) show that parameter estimates of the A
3
(3) model do
not change after adding the caps prices to swap rates. As a result, the models they consider
produce poor t to the derivatives prices. It appears that, apart from dierent models that
11
Lamoureux and Witte (2002), in a Bayesian setting, distinguish the cross-sectional and time-series impli-
cations as well. However, they focus on the appropriate number of factors in the CIR model, as opposed to
comparing dierent models with the same number of factors.
12
Following Dai and Singleton (2000), we use the notation A
m
(N) for a Nfactor branch with m factors
aecting the conditional volatility.
6
these authors consider, estimation strategy could be responsible for the divergence in ndings.
Umantsev assumes that prices of one of the option contracts are observed exactly, which allows
him to imply the volatility factor. Jagannathan, Kaplin, and Sun assume that only swap rates
can be observed exactly. We will avoid the preferential treatment of some of the series by
assuming that all the data are observed with errors as in (2.1).
Heidari and Wu (2002) t an A
0
(3) model to swap rates, and then use additional three
Gaussian factors to explain the swap rates residuals in conjunction with the caps prices. Attari
(2001) argues, based on calibrating a A
2
(3) model to Eurodollar options, that a fourth factor
should be added to explain the upward sloping at short horizons term structure of implied
volatilities. Finally, Chen and Scott (2002) add jumps to a A
1
(2) model, which is estimated
assuming that Eurodollar rate and option implied volatilities are proxies for the spot interest
rate and spot volatility respectively, in order explain the shape of implied volatility smiles.
These three studies point out the need to extend the regular ATSM. However, the dierent
models and inference approaches make it hard to reconcile these works.
We will attempt to clarify the issues related to our interest in the volatility dynamics
by incorporating the estimation strategy discussed above and using a joint options-underlying
dataset. Eurodollar markets provide a convenient data source in this last respect. They have
a number of useful features, which are not shared by the swap markets. Therefore, we discuss
these features before proceeding with the discussion of the models and results.
3 The Eurodollar data
In this paper, we follow a dierent strategy regarding data choice by estimating models based
on Eurodollar futures and options instead of US interest rate swaps and swaptions. The reason
is that the Eurodollar markets have a number of important advantages. This section outlines
these advantages and provides particular details about our dataset.
7
3.1 The Markets
Eurodollar futures are traded at CME and constitute the largest xed-income market in the
world. According to the CFTC, the open interest and average daily dollar turnover in Eurodol-
lar futures were 3 million contracts and $426 billion respectively at the end of 1998. Compare
this to 7 thousand contracts and $592 million for the 13-week T-bill futures, which are also
traded at CME.
13
Finally, according to the CME 2002 gures, the daily turnover in Eurodol-
lar futures was $850 billion, while according to the Federal Reserve, the daily turnover in the
US interest rate swaps was $82 billion in 2001. These numbers indicate that the Eurodollar
markets are highly liquid, and, therefore, the market prices are the most reliable.
Similarly to the swap rates, and in contrast to the T-bill futures, the Eurodollar futures
are contracts written directly on the interest rate, as opposed to a traded asset. The under-
lying interest rate is the 90-day LIBOR, which we denote by
t
(d), where d = 90/360 the
LIBOR horizon in years according to the market conventions. The Eurodollar futures rate
with maturity is quoted as
F
t
() = 100 f
t
() (3.1)
where f
t
() is the future LIBOR rate. The contract is settled in cash based on 100
t+
(d).
Eurodollar futures are issued every quarter. However, when the Eurodollar contract was
introduced in 1981, the available maturities, , went out only up to two years. The CME
gradually introduced longer maturities, and by the end of 1993 futures with maturities of up
to ten years were available. As a result, there are now eectively forty points of the term
structure, compared to six available swap rates.
Note, that the swap-based studies typically use a 6-month LIBOR to pin down the short
end of the curve. In practice, this one additional point demands an additional factor in a term
13
Interestingly, Amin and Morton (1994) report that in 1983, in the second year of the Eurodollar contract
existence, the situation was reversed: the Eurodollar turnover was $3 million, and the T-bill futures one was
$78 million.
8
structure model (Due and Singleton, 1997, and Jagannathan, Kaplin, and Sun, 2003). Due
and Singleton (1997) observe that the distinct institutional and credit characteristics of the
LIBOR market could explain such a result. On the contrary, the Eurodollar futures market
is homogeneous and therefore, one can be condent that one and the same data generating
process is responsible for both short and long end of the curve.
14
One of the perceived advantages of swaps data is the simplicity of valuation because of
the par-rate representation. This representation is based on the assumption that contractual
payments have AA credit quality, which is the same as that of the underlying LIBOR. However,
recently, Collin-Dufresne and Solnik (2001), and Johannes and Sundaresan (2003) argue that
because of the swap contract collateralization, which removes counterparty risk, and because of
the costs associated with posting and maintaining the collateral, the swap payments should be
discounted at a rate not higher than the risk-free rate. This observation implies that estimation
of a AA-quality term structure model based on the swap rates entails modelling of the risk-
free term structure, modelling the cost of collateral process, and estimation which is based on
T-bond data in addition to the swaps data.
In contrast, the mark-to-market feature implies that futures do not have to be adjusted
for the distortions implied by the need for holding the collateral. The futures price, f
t
(), is a
martingale under the risk-neutral measure Q, and valuation does not require any discounting:
f
t
() = E
Q
t
(
t+
(d)) (3.2)
As a result, the modelling costs are reduced, as we can exclusively focus on the AA quality
interest rate.
Yields on zero-coupon bonds are linear functions of the state variables in the ane
framework, simplifying estimation of such models. However, in the case of swaps, the mapping
between the observed rates and factors is nonlinear even if the simple par rate representation is
14
Moreover, avoiding the direct use of the LIBOR data allows us to avoid the inconsistency pointed out by
Heidari and Wu, 2002 which pertains to the simultaneous pricing of xed income derivatives and the underlying
securities in the estimation framework allowing for observation errors. We elaborate on this issue in section 4.2.
9
valid. The futures prices, similarly to zero-coupon bonds, have exponentially linear relationship
with the factors. Indeed, LIBOR is linked to bond prices in the following way:

t
(d) =
1
d
_
1
P
t
(d)
1
_
(3.3)
where P
t
(d) is the price of a hypothetical AA-quality zero-coupon bond which matures d
years from time t. It is clear from expressions (3.2) and (3.3) that if the bond price has an
exponentially linear form, the futures price also has a similar expression, which we show below.
Derivatives on the Eurodollar futures are American puts and calls, which lend themselves
for an easy valuation in the ane framework after conversion to their European counterparts.
15
Finally, options data have been available since 1985 as opposed to swaptions and caps, which
have been available only since 1997.
3.2 Details of the dataset
In order to be consistent with previous studies, we use weekly Eurodollar futures and options
data. Recall that CME gradually introduced futures of longer maturities, beginning with the
maximum maturity of two years in 1981, and ending with a maximum maturity of ten years in
1993. While it is possible to handle varying longest maturity (e.g. Jegadeesh and Pennacchi,
1996), we would like to use the full range of maturities to be comparable with the existing
interest rate swaps studies and start our sample on January 5, 1994. Conveniently, this date
coincides with the start of Allan Greenspans term as the Chairman of the Board of Governors
of the Federal Reserve System. One can argue, that our sample represents one monetary
policy regime (see Piazzesi, 2003 for details). The ending date of our sample, June 27, 2001,
is determined by the date the project was initiated.
16
As a result, we have T = 391 weeks in
our sample.
Since futures mature every quarter, we have potentially forty points on the Eurodollar
yield curve. This amount is much more rich than the six available swap rates plus one
15
The details of this simple procedure are provided below.
16
We obtained the data from the Institute for Financial Markets.
10
six-month LIBOR rate used to pin down the short end of the curve. In order to reduce the
computational burden associated with using all 40 datapoints, we select only 11 maturities.
The rst maturity is six months, and the rest are annual maturities from 1 to 10 years.
Note that futures, in contrast to the swap rates, do not have constant maturities. There-
fore, when we mention a maturity of n years, we mean the maturity closest to n years. We
track the same n-year contract until it has ten business days left to switching into the n1-
year category. Therefore, the actual number of years to maturity is not a random number,
but follows a predictable seesaw pattern.
We also narrow down the available options data. Options have maturities of up to two
years. The ones with maturities of up to one year are the most liquid. Moreover, CME oers
both standard quarterly (
c
=
f
) and serial (
c
<
f
) options. Hence, quarterly options are
eectively options on the cash Eurodollar rates. As a result, it is easier to value these options.
Also, quarterly options are typically more liquid. For these reasons we use only quarterly
options of maturities equal to six months and one year.
Since the focus of our paper is on learning about interest rate volatility, and not accurate
option valuation per se, we further limit our choice of options to the ones with short maturity
and close to being at-the-money. We compute options moneyness as ratio of the strike to the
futures price, and select options with moneyness as close to 1 as possible. As a result, 90%
of all the moneynesses is between 0.98 and 1.02 with median being exactly 1; the smallest
moneyness is 0.97 and the largest is 1.03.
4 Ane Models of Eurodollar Markets
4.1 No-Arbitrage Models
The foundation of our analysis is the hypothetical AA-quality zero-coupon bond with ninety
days to maturity. In the absence of arbitrage opportunities, its price is given by:
P
t
(d) = E
Q
t
_
e

_
t+d
t
r
s
ds
_
(4.1)
11
where d is time to maturity expressed in years, i.e. 90/360, Q denotes the risk-neutral prob-
ability measure, and r is the AA-quality spot interest rate. We have to assume a dynamic
model of r and the structure of Q.
As Dai and Singleton (2000) (henceforth DS) point out, there is a trade-o between the
exibility of modelling time-varying volatility with multiple square-root factors, and the ability
of a model to generate exible conditional and unconditional correlations between the state
variables. Based on this metric, it makes sense to consider only two out of the four possible
three-factor ane term structure models: a Gaussian model, often denoted as A
0
(3), and a
model with one square-root stochastic volatility factor, A
1
(3). This conclusion is supported by
the results in DS and Brandt and Chapman (2003) who nd that ane models with multiple
volatility factors underperform the simpler ones.
We x the ltered probability space (, {F
t
}, F, P) and specify the considered models
in the Ar (ane in r, see DS) form. It means that, for the ease of interpretation of the state
variables, we select r to be one of them. In this respect, we follow the tradition of many ATSM
studies, which assign statistical interpretation to the latent factors.
17
The advantage of this
approach is that it provides an extra diagnostic tool for a model: if the estimated factors do
not correspond to the assigned labels, the model could be misspecied.
The three-factor Gaussian model, A
0
(3), is specied as follows:
dr
t
=
P
r
(
t
+s
t
r
t
) dt +
r
dW
r
t
(P) +
r

dW

t
(P) +
rs
dW
s
t
(P) (4.2)
d
t
=
P

t
_
dt +

dW

t
(P) +
s
dW
s
t
(P) (4.3)
ds
t
=
P
s
s
t
dt +
s
dW
s
t
(P) (4.4)
where traditionally represents the central tendency factor. To the best of our knowledge,
we are the rst to specify the A
0
(3) model in the Ar form, hence the third state variable, s
17
See, for instance, Balduzzi, Das, Foresi, and Sundaram (1996), Dai and Singleton (2000), Due, Pedersen,
and Singleton (2003), Jegadeesh and Pennacchi (1996), Piazzesi (2003) as examples, and Piazzesi (2005) for a
discussion.
12
does not have a traditional interpretation. We will refer to it as the spread factor. This
factor could be interpreted as the impact of the Federal Reserve interest rate targeting (see
Attari, 2001 and Piazzesi, 2003 for related specications). Jegadeesh and Pennacchi (1996)
used the two-factor version of this model (with s switched o) to capture the dynamics of the
Eurodollar futures.
The stochastic volatility model, A
1
(3), is specied as follows:
dr
t
=
P
r
(
t
r
t
) dt +
P
rv
( v v
t
) dt
+
_

2
r
+v
t
dW
r
t
(P) +
r
_

v
t
dW

t
(P) +
rv

v
t
dW
v
t
(P) (4.5)
d
t
=
P

t
_
dt
+
r
_

2
r
+v
t
dW
r
t
(P) +
_

v
t
dW

t
(P) +
v

v
t
dW
v
t
(P) (4.6)
dv
t
=
P
v
( v v
t
) dt +
v

v
t
dW
v
t
(P) (4.7)
Here is again the central tendency factor, and v is the stochastic volatility factor. This model
represents a more exible extension of the BDFS model. DS nd this model to be the most
successful in capturing the dynamics of US interest rate swaps.
We introduce notations associated with the probability measure transformation from P
to Q in order to implement the bond formula (4.1). If we denote the vector of risk premia by
, then, according to the Girsanov theorem, we can link the two probability measures via:
W
t
(Q) = W
t
(P) +
_
t
0

s
ds (4.8)
We consider the essentially ane risk premia specication of Duee (2002). In the case
of A
0
(3) we have:

r
t
=
r
+
rr
r
t
+
r

t
+
rs
s
t
(4.9)

t
=

+
r
r
t
+

t
+
s
s
t
(4.10)

s
t
=
s
+
sr
r
t
+
s

t
+
ss
s
t
(4.11)
13
If the model is A
1
(3) then:

r
t
=
1
_

2
r
+v
t
_

2
r
+
rr
r
t
+
r

t
+ (
r
+
rv
)v
t
_
(4.12)

t
=
1
_

v
t
_

+
r
r
t
+

t
+ (

+
v
)v
t
_
(4.13)

v
t
=
v

v
t
(4.14)
For completeness, appendix A lists the dynamics of the models under the Q measure.
18
Cheredito, Filipovic, and Kimmel (2003) have recently shown that essentially ane
risk premia specications can be extended to the cases where the price of risk is inversely
proportional to the square-root volatility factor, as in, for example, (4.12) with
r
= 0. For
this reason we will not constrain
r
and

to be greater than zero at the estimation stage.

19
However, we still omit one possible additional term in (4.14), because its value binds a regularity
condition based on the results reported in Cheredito, Filipovic, and Kimmel (2003). Further
exploration of the contribution of the additional risk premia terms is an exciting direction of
future research.
We have made all the assumptions necessary to compute the bond price (4.1). The ane
specication implies:
P
t
(d) = e
A
P
(d)+B
P
r
(d)r
t
+B
P

(d)
t
+B
P

(d)
t
e
A
P
(d)+B
P
(d)S
t
(4.15)
where generically denotes either s in A
0
(3), or v in A
1
(3), and, for compactness, S
t
=
(r
t
,
t
,
t
)

. The coecients A
P
and B
P
solve well-known Ricatti ODEs (see DPS, for in-
18
We ensure stationarity of both models under both probability measures by requiring the positivity of the real
part of all the eigenvalues of the matrices constructed from the respective s (the Routh-Hurwitz criterion). We
also impose the generalized Feller condition (condition A of Due and Kan, 1996) to guarantee strict positivity
of volatility in A
1
(3).
19
Note that in order to satisfy the canonical representation in DS1, the parameters
r
and

in A
1
(3) can
not be equal to zero. On the other hand, the model with these parameters set at zero is an admissible and
identiable model (see Due and Kan (1996) for the regularity conditions). At-Sahalia and Kimmel (2002)
make a similar observation.
14
stance).
20
The boundary conditions are: A
P
(0) = 0, B
P
(0) = 0.
It is possible to impose constraints on the Q parameters of the A
1
(3) such that the
volatility factor v disappears from the bond pricing equation (the USV feature).
21
More
precisely, the markets are incomplete in the three-factor models if the loadings on the factors
and v are linearly dependent, i.e. one can nd coecients u

and u
v
(at least one of them
should be non-zero) such that:
u

B
P

(d) +u
v
B
P
v
(d) = 0 d (4.16)
CDG derive the parameter constraints implied by this requirement. We list these constraints
in the appendix B. CDGJ nd empirical support for the USV restrictions based on the swaps
data.
We now have enough structure to price instruments from the Eurodollar markets. We
combine the denitions of the Eurodollar futures (3.2), the LIBOR rate (3.3), and the LIBOR
bond price (4.15) to obtain the futures LIBOR rate:
f
t
() =
1
d
E
Q
t
_
e
A
P
(d)B
P
r
(d)r
t+
B
P

(d)
t+
B
P

(d)
t+
1
_
=
1
d
e
A
P
(d)
E
Q
t
_
e
B
P
r
(d)r
t+
B
P

(d)
t+
B
P

(d)
t+
_

1
d
(4.17)
=
1
d
e
A
P
(d)+A
f
()+B
f
r
()r
t
+B
f

()
t
+B
f

()
t

1
d

1
d
e
A
P
(d)+A
f
()+B
f
()S
t

1
d
where A
f
and B
f
solve almost the same Ricatti ODEs with slightly dierent boundary condi-
tions as compared to (4.15): A
f
(0) = 0, B
f
(0) = B
P
(d).
20
We would like to point out a delicate issue associated with the ODE-based bond valuation formula. This
approach relies on the Feynman-Kac theorem, which is easily applicable only when r is bounded from below
(see Due, Filipovic, and Schachermayer (2003)). For example, the model A
3
(3) would satisfy this requirement.
In the case of the A
0
(3) model, the justication of the use of the Feynman-Kac formula can be made based on
formula (17.4) in Sato (1999). Levendorskii (2004) derives the consistency conditions for the general payos
in the A
1
(N) models. It turns out that these conditions require boundedness of the securitys payo and the
regularity conditions in Due and Kan (1996). Hence Feynman-Kac is applicable for our models. We are
grateful to Darrell Due and Sergei Levendorskii for discussions regarding this issue.
21
The A
0
(3) model does not admit the USV property.
15
It is natural to work with yields, as opposed to bond prices. The zero-coupon bond yield
is:
y
P
t
(d) =
1
d
log P
t
(d) =
A
P
(d)
d

B
P
(d)
d
S
t
(4.18)
By analogy with the bond (also see (3.3)), we can construct an object based on the futures
price, which we will refer to as the futures yield:
y
f
t
() =
1
d
log(1 +d f
t
()) =
A
P
(d) +A
f
()
d
+
B
f
()
d
S
t
(4.19)
Note that the factor loadings on the bond and futures yields are dierent. Therefore, both
types of assets might not simultaneously satisfy the USV condition (4.16).
22
Indeed, zero-
coupon bond prices are directly related to the forward rates. The forward rates dier from the
futures rates by the convexity adjustment, which is sensitive to volatility. Hence, the futures
data is informative about volatility even in the presence of the USV. This feature allows us to
identify the volatility risk premium, which would not be possible with the regular swap data
(see CDGJ).
23
A European call option on Eurodollar futures can be evaluated via similar techniques:
C
t
(
f
,
c
, K) = E
Q
t
_
e

_
t+
c
t
r
s
ds
(F
t+
c
(
f

c
) K)
+
_
= E
Q
t
_
e

_
t+
c
t
r
s
ds
(k f
t+
c
(
f

c
))
+
_
(4.20)
where k = 100 K. Hence, a call on Eurodollar futures rate can be considered as a put on
LIBOR futures rate.
24
We follow Due, Pan, and Singleton (2000) in computing the call price.
Appendix C provides the details.
22
We have veried that our intuition is correct for the estimated models. The details are available upon
request.
23
Moreover, as we detail in the appendix E, this sensitivity to volatility allows us to estimate this latent factor
accurately, which would be much harder to do using the traditional data sources, such as bond yields or swap
rates.
24
Note that we are using the AA-quality instantaneous rate r
t
to discount the option payo at maturity. Since
options on Eurodollar futures are traded on CME and are subject to the usual regulations, in principle, one
should discount the payo at the risk-free rate. Such discounting would entail modelling and estimation of the
16
In reality, the Eurodollar options are American. Flesaker (1993) reports that the early
exercise premium for options not substantially in-the-money is generally less than a single basis
point, which is the tick size in this market. Nonetheless, we use a very accurate approximation
from Broadie, Chernov, and Johannes (2004) to compute the early exercise premium. The
procedure is straightforward. First, we compute the Black implied volatility using the binomial
tree, which takes into the account the early exercise option. Second, we obtain the European
option price by plugging the implied volatility into the European version of the Black formula.
Since we are looking at at-the-money short maturity options, the exercise premium is expected
to be small. Indeed, in terms of the Black implied volatility, the median premium is equal to
0.05%, and the largest premium is 0.16% (as a reference, the smallest implied volatility in our
sample is 6.7%).
25
4.2 Econometric Models
As we discussed in the section 2, the state-space system is the most convenient way to think
of our models econometrically. We can specialize the general observation equation (2.1) to our
particular case as:
y
f
t
(
ft
) = g
f
(S
t
,
m
,
ft
) +
t
(
ft
),
ft
= 0.5
t
, 1
t
, 2
t
,...,10
t
(4.21)
C
t
(
ct
,
ct
, K) = g
c
(S
t
,
m
,
ct
) +
t
(
ct
),
ct
= 0.5
t
, 1
t
(4.22)
credit spread, and modication in pricing which takes this into an account (see, for instance, Collin-Dufresne
and Solnik, 2001 and Johannes and Sundaresan, 2003). In practice, we expect to see a small dierence between
the instantaneous risk-free and refreshed AA rates. Alternatively, one could think that trading options involves
small counterparty risk, which is comparable to the default risk of AA-rated companies. Further modelling of
these ne features is beyond the scope of this paper.
25
In principle, in order to be absolutely accurate, we should have computed the exercise premium for the
particular model under consideration, i.e. A
1
(3). However, the impact of the correct model on the computation
is an order of magnitude smaller than the premium captured by the simple adjustment via the Black model.
Since we nd this adjustment to be very small already, it would be impractical to use more complicated models.
17
where S follows the state equation (4.2)-(4.4) or (4.5)-(4.7) depending on whether we estimate
the A
0
(3) or A
1
(3) models, and the functions g
f
() and g
c
() are given in the equations (4.19)
and (4.20), respectively. The number of years to maturity with subindex t emphasizes the fact
that actual maturity is dierent on each day.
The index m = 0, 1 of the parameter vector
m
signies the two models that we consider
A
0
(3) and A
1
(3) respectively. The full parameter vectors are:

0
=
_

P
r
,
P

,
P
s
,

,
r
,

,
s
,
r
,
rs
,
s
,
r
,

,
s
,
rr
,
r
,
rs
,
r
,

,
s
,
sr
,
s
,
ss
_

(4.23)

1
=
_

P
r
,
P

,
P
v
,
P
rv
,

, v,
r
,

,
v
,
r
,
rv
,
r
,
v
,

,
r
,

,
v
,
rr
,
r
,
rv
,
r
,

,
v
_

The parameters in boxes are restricted in the USV model (see Appendix B). Not all of the
parameters are necessarily identied. We discuss our identication strategy in the appendix
D.
The terms , and are often referred to as measurement errors. Indeed, it is natural to
assume that because of the bid-ask spread, the traded price does not represent the equilib-
rium price prescribed by the model. An alternative reason for including these errors into the
econometric specication of the model comes from an opposite view that an econometrician
observes the correct prices. This assumption leads to a stochastic singularity problem, which
can be resolved by introducing the error terms. Whatever is the correct motivation for the
error terms, we expect them to have small variance if the model is correctly specied. We
intentionally assume that the errors are independent across various maturities, and all have
the same variance,
2

, irrespective of maturity. We allow a dierent variance,

2

, for the op-

tion pricing errors. Such a restrictive specication puts more pressure on our original model
to t the data. It will be harder to detect that a model is misspecied if the error terms are
exible.
26
26
Heidari and Wu (2002) critique existing studies of LIBOR and swap rates, and swap-based derivatives in
the ane framework because the pricing formulas implicitly assume that the contracts will be settled based on
the theoretical LIBOR rates, i.e. they do not take into an account the measurement error. Our methodology
is not subject to their criticism because both futures and quarterly options contracts are settled based on the
actual LIBOR rate at maturity. Since, in contrast to all other studies, we do not use the LIBOR data in our
18
We use quasi-maximum likelihood with Kalman lter to estimate the models. Appendix
E describes the details of the implementation.
5 Results
5.1 Preliminary observations
We estimated two versions, depending on the dataset, of each of the three models: A
0
(3),
A
1
(3), and A
USV
1
(3). We assign a superscript f to the models estimated based on futures
only, and a superscript fo to the ones estimated on the joint dataset. We initially estimated
the maximal models. However, some of the parameters were insignicant according to the
asymptotic standard errors. Following the strategy of Dai and Singleton (2002), we restricted
some of these parameters to zero, if such a restriction did not lead to a notable decline in the
value of the log-likelihood function. We report the resulting parameter estimates in tables 1
and 2.
Our sample spans a relatively short period in the light of the well-known persistence of
the interest rate. Therefore, asymptotic standard errors may overstate the signicance of the
estimated parameters (see, e.g. Pritsker, 1998). We follow Conley, Hansen, and Liu (1997)
and construct nite sample condence intervals via the parametric bootstrap. Specically, we
simulate 800 state variables paths with 391 observations (the length of our sample) each. Next,
we generate futures and options prices according to our models from these paths. Finally, we
reestimate the models based on a dataset along each of the paths. This procedure generates
nite sample distribution of the parameters estimates. We report the 95% condence intervals,
which turn out to be asymmetric relative to the estimates. We follow similar strategy for all
statistical inference conducted in the paper.
The parameter estimates allow us to take a rst glance at the models properties. First,
study, we can simply assume that actual LIBOR rates are observed without an error without sacricing the
internal consistency.
19
we check if the initial interpretation of the state variables corresponds to the actual role they
play in respective models. We correlate the factors, ltered conditional on the estimated
parameters values, with some interesting observables, such as the six-month futures yield, or
the short rate; the ten-year futures yield, or the long rate; the dierence between the two,
or the slope of the term structure; the buttery spread, which corresponds to a long position
in the six-month and ten-year futures and double short position in the two-year futures; and,
nally, the absolute implied variance.
27
Table 3 reports the results. In many cases, there
are two observables that are highly correlated with a particular state variable. We will focus,
somewhat arbitrarily, on the observables with the highest correlation.
The correlations of the factors r, , and s with the observables are consistent across the
models and make intuitive sense: they are correlated with either the short rate, the slope, or
the long rate.
28
The only exception is the factor in the A
1
(3)
fo
model, which is correlated
with the buttery spread. The most intriguing result is for the factor v : it is correlated with
dierent observables for each avor of the A
1
(3) model.
The volatility factor is highly correlated with the buttery in case of A
1
(3)
f
. This re-
sult makes sense because the buttery spread should be sensitive to volatility (see, for in-
stance, Litterman, Scheinkman, and Weiss (1991)). Another model that makes perfect sense
is A
USV
1
(3)
fo
: v is highly correlated with implied variance, and other factors have correlations
similar to those of A
1
(3)
f
. A combination of the USV restriction and options data cleanly
separates the contribution of each of the factors.
In the case of A
1
(3)
fo
, v is correlated with the slope, which, in combination with our
27
The absolute implied variance refers to the variance of absolute changes in the interest rates according to
the Black model. The rationale for such volatility measure is that interest rate in the Black model follows the
Geometric Brownian Motion with constant volatility. As a result, the implied volatility coecient corresponds
to the volatility of the relative interest rates. The notion of absolute volatility, or variance, is closer to the
volatility factor v in our ane models.
28
The correlation of these factor with the respective observables deteriorates when the options data is added
reecting the tension between the requirements to t the term structure and option prices simultaneously.
20
observation regarding in this model, raises the suspicion about the model misspecication.
It appears that factor v has rotated into the central tendency, and factor has rotated into
the volatility (though correlations are weaker than for the A
1
(3)
f
counterparts). However,
has virtually zero correlation with the implied volatility, while its counterpart v in A
1
(3)
f
has
a respectable correlation of 0.24. It appears that when the A
1
(3) model is confronted with the
joint dataset, the factors substantially deviate from their intended roles in order to generate
realistic distribution of prices.
In the case of A
USV
1
(3)
f
, v is correlated with the long rate the same observable that
is correlated with in this model. This puzzling result is explained by the USV feature.
Formally, in the case of USV, a linear combination of two factors acts like one factor in the
cross-section.
29
The dataset involving the futures only is not informative enough to make a
signicant distinction between volatility and central tendency in the time series.
To summarize, we nd that estimated latent state variables are not necessarily behaving
according to the initial model specication. This fact is especially true for the factor v. The
evidence emphasizes the fact that an ane model does not necessarily have an intuitive re-
lationship between the continuous-time properties of the factors and discrete-time properties
of the yields. However, these observations do not mean a failure of the ane models: while
individual factors may not be easy to interpret, in combination, they could still produce a
successful representation of the data.
A complementary way to investigate further the impact of various factors on the term
structure is to plot the response of the term structure to one standard deviation shocks in the
state variables. The plots in Figure 1 show the changes in yields:
y
t
() =
B
r
()

r
t

B

()

t

B

()

t
(5.1)
where r, , and (v, or s) are taken to be equal to one standard deviation of the
respective state variable. Each panel plots the product of the factor loading in (5.1) and the
29
More precisely, this combination can be rotated into to the risk-neutral expected value of the spot interest
rate (see CDGJ). Hence, high correlation with the long rate is natural.
21
one standard deviation of the respective factor (this case is denoted by ess Q on the legend).
In addition, in order to disentangle the contributions of the state dynamics versus the ane
and essentially ane risk premia, we have recomputed the factor loadings, setting all risk
premia to zero (this case is denoted by P on the legend), and setting only the essential risk
premia (the ones with double index) to zero (ane Q on the legend). In this respect, our
approach is similar to Umantsev (2001).
We see, consistent with the correlations evidence in the Table 3, that the individual
factors adjust in their behavior and change roles from model to model and dataset to dataset.
In the next section, we will study how this variation in the factors impact aects the ability
of models to explain the data.
Finally, we observe, in respect to the risk premia, that there is no dierence between
the P and the Q measures across all models and all factors (with the exception of v) in the
absence of the essentially ane components. Hence, despite insignicance of many essentially
ane risk premia, they play important role in models ability to t the data.
5.2 Models ts
We report pricing errors in Table 4 as a starting point of models t evaluation. We compute the
average of the absolute dierences between the model-implied and actual futures yields dened
in (4.19). Also, we compute the absolute dierence between absolute Black implied volatilities.
The futures errors are very small and similar across the models (with some deterioration for
the USV models). Not surprisingly, the errors generally increase for the models that were
estimated on futures and options because of the increased pressure on the same model to t
a more elaborate dataset.
30
This increase in futures errors is associated with an impressive
30
This result is dierent from Jagannathan, Kaplin, and Sun (2003), who report no change in the swaps t
once the caps data is added. Since their parameter values do not change, we suspect that they obtained a local
optimum.
22
improvement in option pricing.
31
The A
USV
1
(3)
f
model is the exception that does particularly
badly in option pricing. This outcome is to be expected because, as we saw in Table 3, none of
its factors are correlated with volatility-related measures. Overall, it is hard to tell the models
apart based on the pricing errors because they appear to do an equally good job tting the
futures and option prices, with an exception of USV.
32
Perhaps, the good pricing performance is not surprising because three factor models
should be able to t the cross-section of yields well based on the principal components results
of Litterman and Scheinkman (1991). Hence, an ability to simultaneously t the time series
well is more relevant for the model evaluation. Our estimation framework, which relaxes the
connection between the cross-section and time-series allows us to do implement the time series
diagnostics.
In order to distinguish the time series properties of the models more eectively, we per-
form hypotheses testing associated with relative models performance based on the conditional
likelihood. Typically, a test of nested models is quite informative in locating a parsimonious
model which has a good data representation. In our case, the models from the A
0
(3) and A
1
(3)
classes are non-nested. Moreover, usual tests assume that one of the models is correct. We
suspect that all our models are misspecied, and we would like to take this complication into
an account.
The concept of encompassing (see Appendix F) is very useful in our situation. In the
context of this paper, the model A
0
(3) encompasses a model A
1
(3) if the former is able to
explain characteristics, i.e. certain statistics, or likelihood values, of the latter. The encom-
passing principle applies to the nested models as well. For example, a restricted model A
USV
1
(3)
encompasses a larger model A
1
(3) if the associated restrictions cannot be rejected. It is im-
31
Umantsev (2001) also observes a dramatic improvement in the options t of the A
1
(3) model once the
options data is used for estimation. He, however, does not acknowledge the deterioration of t in the underlying
security.
32
We have conducted a more detailed analysis of pricing errors for all maturities with essentially the same
results. Details are available upon request.
23
portant to emphasize that these tests in no way tell us that one model is better than the
other one when the null is rejected. The rejection shows that the two models are suciently
dierent from each other given the dataset.
It is informative to see if A
0
(3) encompasses A
1
(3) since it is perceived by many re-
searchers as a simpler model because volatility of the short interest rate is not stochastic.
Table 5 presents the tests for the model based on futures only and on futures and options
jointly. The results are striking. In the futures only case, the Gaussian model clearly en-
compasses the stochastic volatility one. The data is not suciently informative to distinguish
between the two models. However, when the options data are added, the null hypothesis is
rejected. Hence there is hope to determine the importance of the two models distinguishing
characteristics based on the options data.
The results perfectly match our initial intuition from section 2 that options data will
be helpful in telling the models apart. Importantly, the futures only based test provides
preliminary formal evidence that A
0
(3) model can capture term structure data as well as the
more sophisticated A
1
(3). This nding has very important modelling implications; therefore,
we will explore this further in the subsequent sections.
We also want to check if the USV restrictions are supported by the data. Similarly to
the A
0
(3) model, we test whether A
USV
1
(3) encompasses A
1
(3). The results are reported Table
5 as well. In contrast to the Gaussian model, the null hypothesis of encompassing is rejected
based on both types of datasets. Given the poor pricing performance of the USV models, this
test clearly highlights problems with the USV restrictions.
We also directly check the signicance of the individual restrictions on the parameters

,
P
v
,
P
rv
,

, and
r
via the Wald-type test. Table 6 reports the results. In the case of
A
1
(3)
f
, we cannot reject the restriction on the risk premium
r
because of zero correlation
between r and (see (B.10)). The remaining USV restrictions are rejected.
33
Only two of the
USV restrictions,
P
v
and
P
rv
, cannot be rejected for the A
1
(3)
fo
model.
33
The restriction (B.9),

= 0 is rejected at the 10% level.

24
These ndings are stronger than the ones in CDGJ. These authors compare A
1
(3)
f
and
A
USV
1
(3)
f
and nd little dierence between them. Consistent with our ndings, their strongest
evidence against A
1
(3)
f
is that the factor v has a much weaker correlation with observable
measure of volatility than the one from the USV version (see Table 3). However, our direct
tests clearly show the need for a specication more exible than the USV one.
5.3 Conditional mean and volatility
We would like to understand whether the results of the encompassing test based on futures
only could be attributed to weakness of the test in certain dimensions. Therefore, instead of
testing the whole likelihood function, we will focus on particular, practically relevant, aspects
of the data distribution and see if we can nd dierences and similarities between the models.
In this section, we explore abilities of the models to t conditional means and volatilities.
The recent trend in the literature is to use GARCH as a model-free estimate of con-
ditional volatility (see, for instance, CDGJ, Dai and Singleton, 2003, and Umantsev, 2001).
The idea is to estimate the GARCH on the actual data, and on the model-implied series, and
to compare the implied properties of the conditional volatilities. In spirit, such a diagnos-
tic is similar to the reprojection technique of Gallant and Tauchen (1998).
34
However, since
GARCH is still a model, one has to be careful with the interpretation of results. To streamline
the discussion, consider the basic GARCH(1,1) model:
Y
t
=
t1
+
t
e
t
(5.2)

2
t
= c +
2
t1
e
2
t1
+
2
t1
(5.3)
where Y generically denotes the series of interest. The estimated values of parameters governing
the volatility dynamics (c, , and ) will be sensitive to the estimated conditional mean .
34
More precisely, this exercise is similar to the initial use use of the technique in the Gallant and Tauchen
paper. More recently, the reprojection is also used to lter latent state variables (i.e. the dierence is in the
conditioning set: historical information only in the former applications versus historical and contemporaneous
information in the latter applications).
25
CDGJ focus directly on the dynamics of the state variable r, i.e. Y
t
= r
t
r
t1
, and compute
from the estimated model. As a result, the GARCH volatility becomes model-dependent even
when estimated based on the dataset: dierent specications of the ane models invariably
produce dierent estimates of r and . It appears that it is more appropriate to focus on the
volatility of the yields themselves because this leads to a truly ATSM-independent estimates
of conditional volatility. For instance, Dai and Singleton (2003) estimate univariate GARCH
model for 5-year yields.
Despite its shortcomings, the advantage of the CDGJ approach is in the potentially
rich specication of , which should lead to a more accurate estimate of the noise term. We
combine this exibility with the robustness of the actual yields in the role of Y by estimating
conditional means and volatilities from a VAR(1)-GARCH(1,1) model applied to a trivariate
series, which include the short rate, the slope, and the buttery spread. These three series
are highly correlated with the rst three principal components, and therefore, contain an
informative summary of the entire term structure.
After accounting for the VAR(1) conditional mean, we nd, based on the Portmanteau
test, the strongest evidence for heteroscedasticity of residuals in the slope, and the weakest
one in the short rate. Nonetheless, we were able to estimate the GARCH models for all three
series. Figures 2 and 3 show means and volatilities from the observed futures yields and the
ones implied by our models. Table 7 complements the gures by reporting correlations between
the means and volatilities implied from the data and those implied from the models.
We see that, when the dataset type is xed, the A
0
(3) and A
1
(3) models produce nearly
identical estimates of the conditional means and volatilities. Moreover, both models do a
pretty good job in tting the rst two moments, especially for the level and slope. USVs t
deteriorates relative to the two models, especially when one starts adding options data and
looks at the curvature. These results are surprising because these are the aspects of the data
that USV was designed to match the best. The fact that the USV model generates volatilities
that are similar or worse than the other models contradicts CDGJ, who nd that only USV
26
volatility correlates well with the GARCH volatility. The dierence in the ndings comes from
the way GARCH volatility is estimated.
Interestingly, despite a clear diculty in identifying the role of the instantaneous vari-
ance of the short rate (see our discussion of Table 3), the models estimated on the futures,
irrespective of their factor structure, do a remarkable job in explaining the variance of the
yields. Moreover, even when options data are added, the t of conditional mean remains very
good, despite the drop in the correlation between the factor r and the short rate (see Table 3).
This nding is quite surprising because the A
0
models can generate only constant con-
ditional volatility.
35
Our results can be explained by a dierence in the conditioning sets in
theory and in practice. The state variables in Gaussian models, being Markov processes, have
constant volatilities conditional on the entire vector of state variables r, , and s. However,
in practice, we lter the state variables, i.e. we condition on the observable yields and their
history. The state variables are not Markov in the yields, and as a result the ltered series do
not have to have constant conditional volatility. Indeed, as the equation (E.8) shows explicitly,
the variance of the ltered state variables is not constant. Interestingly, as we add options
data and the conditioning set becomes closer to the full state, the GARCH volatility becomes
more constant (see gure 3).
In the light of these observations, the natural question is whether one can distinguish
dierent models when the states are unobservable. Since, one has to rely on various data
sources to estimate models and lter state variables, the answer to this question will depend
on the informativeness of the particular dataset. As we have seen with the USV model, the
data is informative enough to unequivocally show that the model is misspecied. However,
it is denitely not enough to distinguish the Gaussian and stochastic volatility model: the
diagnostic is not powerful enough given the data. These conclusions are fully consistent with
the results of encompassing tests.
We see two ways to resolve this problem. One way, as the encompassing tests hint, is to
35
We are grateful to Ken Singleton for an extensive discussion and suggestions on this issue.
27
rely on the options data and nd a new diagnostic which is suciently powerful. We explore
this direction in the subsequent sections. Alternatively, we can explore whether a model
can in principle generate the amount of heteroscedasticity found in the data by simulating
from the estimated model. We proceed in the same fashion as with the encompassing tests.
We simulate 800 samples from the respective models, and reestimate GARCH along each of
the paths.
36
This procedure allows us to construct nite sample condence intervals for the
GARCH coecients.
First of all, one can not estimate a GARCH model on the data simulated from the A
0
class. Since conditional volatility is constant, parameters governing the GARCH dynamics
in (5.3) are not identied: both sets = = 0, and c = = 0 with = 1 produce
constant volatility. Therefore, we restricted to zero to allow the model to pick up some
heteroscedasticity due to small sample noise via the ARCH component. We found that the
95% condence bounds constructed based on 800 simulated paths cover the values of obtained
from the sample, however they also cover zero.
37
This nding is not surprising because we
simulated from a constant volatility model.
The more interesting question is whether the A
1
(3) model can replicate the levels of
GARCH in the data. Our nding is basically the same: the condence intervals are wide
enough to cover the value of the parameters obtained from the real data and to cover zero. We
believe that the intervals are so wide because the data are not suciently informative about
this aspect of the model.
Our ndings indicate that in practice, the Gaussian and stochastic volatility models are
not distinguishable based on the low order conditional moments of the term structure. This
implies that for many purposes, such as intra-regime modelling or studying of low frequency
data, the Gaussian models are sucient.
38
However, one has to use additional data sources
36
Because of the computational complexity, we estimated univariate GARCH models for each of the principal
components. To be consistent, we also estimate univariate GARCH in the data for the purposes of this exercise.
37
The values of in the sample ranged from 0.07 for the curvature to 0.27 for the slope.
38
In particular, this nding addresses the concern expressed in Bansal, Tauchen, and Zhou (2003) regarding the
28
and more elaborate diagnostics in order to understand whether the distinguishing models
characteristics are relevant. We proceed with this latter task below.
5.4 Term Structure of Volatility
An alternative way to evaluate how well a model can replicate volatility patterns is to study the
term structure of unconditional volatilities. The left panel of Figure 4 shows the unconditional
volatilities of yield changes computed as standard deviation directly from the data (solid line)
and population values implied by various models. In the data, we see the hump that was
documented for the swaps and Treasury rates. The A
0
(3)
f
and A
1
(3)
f
models generate very
similar term structures which have magnitudes on par with the data, however the hump is
much less pronounced. The same models estimated on options and futures data can generate
the hump, but overall magnitude is higher than in the data. Also, the Gaussian and stochastic
volatility models produce markedly dierent volatility patterns for maturities over 2 years.
However, it is hard to determine which one is preferred. Both USV models produce volatility
patterns that are quite dierent from the ones in the data.
Since we are using options data, we can evaluate the term structure of implied volatilities
as well. Backus, Li, and Wu (1998) document the hump-shaped term structure of cap-implied
volatilities. They propose univariate ARMA models to capture this eect. Attari (2001)
proposes four-factor ane models to replicate the upward-sloping term structure at short
horizons. It would interesting to see if our three-factor models are able to generate the hump
in implied volatilities.
The right panel of Figure 4 shows the population values of average implied volatilities
implied from the models and average implied volatilities computed from the data (solid line).
Dai, Singleton, and Yang (2003) regime-switching model: Given the nature of yields data, it would seem that
allowing for within regime volatility to be stochastic is quite important. It remains to be seen if the specication
which assumes a constant within regime volatility can account for the observed time-varying volatility and
conditional cross-correlation of yields.
29
The data is not very informative because we used only 6-month and 1-year options, which
capture only the upward slope at the short horizons. Therefore, we can only see how the models
dier from each other. The story is basically identical to the one for standard deviations. The
A
0
(3) and A
1
(3) models are very similar to each other within each dataset type, and they
generate fairly reasonable volatility values. The models estimated using options generate the
hump. The USV models do a poor job in both respects.
It is clear that we cannot distinguish either f or fobased models based on the rst
two moments. Both, especially when estimated using the options data, are quite successful in
reproducing the stylized shape of the volatility term structure. The next natural step is to
evaluate the ability of the models to t higher order moments.
5.5 Higher order moments
The stochastic volatility models are unconditionally not normal. Therefore, we might be able
to distinguish them from the Gaussian models if we focus on their ability to match skewness
and kurtosis. We continue with the same three series from the previous subsection: level,
slope, and curvature. We compute their skewness and kurtosis in the sample, and their nite
sample distribution via the parametric bootstrap. The skewness of all three series is very close
to zero, and this property is consistent with what all models generate. Therefore, we do not
report the skewness results and focus on kurtosis in Table 8.
It turns out that it is extremely hard for all models, even non-Gaussian ones, to generate
the relatively high degree of kurtosis observed in the term structure. Only the A
1
(3)
fo
model
and both USV ones succeed in matching the kurtosis of the level. The A
USV
1
(3)
f
model matches
the kurtosis of the slope as well.
Contrasting these results with the rst and second moments performance, we observe a
clear tension: the models tted to the term structure only do an excellent job tting VAR(1)-
GARCH(1,1), but they fail with kurtosis. The models that are tted to the joint dataset have
an opposite performance. We made a similar observation while evaluating the pricing errors.
30
Unfortunately, the USV models do not resolve the tension despite its good performance for
kurtosis: pricing errors increase for both types of the dataset; the performance based on rst
two moments is not impressive either.
We still nd that we can barely distinguish the A
0
(3) and A
1
(3) models even using the
options data. We did not yet explore the the option pricing implications in the time series
context. A natural issue to study is the premise for the USV models: the inability of term
structure to explain the option-implied volatility. Since USV feature is prohibited by the
A
0
(3) specication, maybe the implied volatility regressions will be able to shed light on the
dierences between the models.
5.6 Implied volatility regressions
We regress the volatilities implied from 6-month and 1-year options on the three principal
components of the term structure. We report the R
2
in the data. They are fairly low and
consistent with the values reported by CDG. We bootstrap the nite sample distribution by
simulating 800 paths for each model. Table 9 reports the results.
We nd that A
0
(3) models fail in replicating the R
2
in the data. However, the A
1
(3)
models (including the USV ones) do a reasonable job. Especially successful are the fo models:
pointwise, they closely match the data. Thus, we nd that the A
1
(3) model can replicate the
pattern in the data even without the USV restrictions.
The question is why the R
2
from A
0
(3) are so low. Figure 5 provides the answer. We
plot how at-the-money absolute implied volatility changes in response to the changes in the
state variables. As we can see, there is virtually no reaction to the changes in either r or
across all models. However, there is a roughly linear relationship between the factor v and
implied volatility in the A
1
(3) models, while there still no reaction to s the third factor in the
A
0
(3) models. Regressing implied volatility on the three principal components is the same as
regressing it on the three factors in the ane models. For this reason, we nd no relationship
between implied volatility and the term structure generated by the Gaussian model.
31
We have nally identied a dimension, which allows us to clearly distinguish the prop-
erties of the A
0
(3) and A
1
(3) models. As predicted by the encompassing tests, we were able
to do so only after the options data was incorporated into our analysis. It turns out that the
only factor that implied volatility is sensitive to is the stochastic volatility factor v in the A
1
(3)
models. Hence, purely Gaussian models have no hope of capturing all the aspects of option
prices.
6 Conclusion
In this paper we study the ability of dierent ane models to explain variability of the term
structure. In particular, we try to understand how the properties of model factors translate into
their ability to generate realistic yields and option prices and their dynamics. We estimate a
Gaussian (constant volatility), stochastic volatility, and unspanned stochastic volatility models
using a joint dataset of Eurodollar futures and options in our analysis. We subject the models
to a battery of diagnostic tests.
We nd that factor specication is not directly related to the model ability to match
a particular aspect of the data. For instance, as long as one is concerned with modelling
term structure only, a simple Gaussian model can do a superb job (at least within a regime).
Nonetheless, it is possible to identify dimensions that can distinguish properties of the mod-
els factors, especially the dierent exibility in generating conditional volatility. Option data
are such source of distinguishing information. We show that dierent models have dierent
implications for matching higher moments of the yields distribution, and explaining the re-
lationship between option-implied volatility and the term structure. A stochastic volatility
model is required for such tasks.
Introduction of the options data to estimation introduces a tension between models
abilities to t term structure versus options, and to t lower order moments versus higher
order moments. One way to relax this tension was proposed by Collin-Dufresne and Goldstein
32
(2002), who introduced the unspanned stochastic volatility. These restrictions simultaneously
weaken the term structure dependence on spot volatility, and free up that same volatility factor
to focus on option tting. Despite some initial support for this intuition, all major tests and
tting exercises favor the regular stochastic volatility model.
Despite some clear evidence for the misspecication of all studied models, we believe
that overall our ndings are good news for the ane class. These models can match key
aspects of the term structure very well and price options accurately. Despite the tension in
the simultaneous t to the yields and options, they still do a good job overall. However, it is
natural to ask the question of what should be the next step in the improvement of the ane
models.
One way to proceed would be to maintain the USV structure to allow clear separation
between the factors dedicated to tting term structure, and the factors dedicated to tting
options. Heidari and Wu (2002) make a less formal attempt in this direction.
39
The reported
tension in pricing and the parallel tension in matching lower and higher order moments point
into the direction of adding jump components. Jumps in interest rates have a clear economic
interpretation, e.g. Piazzesi (2003) and descriptive studies, such as Andersen, Benzoni, and
Lund (2004), Chen and Scott (2002), Das (2003), and Johannes (2004) show that this idea is
a promising direction to pursue. This endeavor is the subject of our future research.
39
These authors use Gaussian models which do not allow the USV feature. Nonetheless, Heidari and Wu
(2002) impose restrictions that are in spirit similar to the USV ones.
33
A Ane Models of the Interest Rate Under Q
Equations (4.2)-(4.4), (4.8), and (4.9)-(4.11) imply that under the measure Q the A
0
(3) model
evolves according to:
dr
t
=
_

Q
r

Q
r
r
t

Q
r

t

Q
rs
s
t
_
dt +
r
dW
r
t
(Q) +
r

dW

t
(Q) +
rs
dW
s
t
(Q)(A.1)
d
t
=
_

Q
r
r
t

Q

t

Q
s
s
t
_
dt +

dW

t
(Q) +
s
dW
s
t
(Q) (A.2)
ds
t
=
_

Q
s

Q
sr
r
t

Q
s

t

Q
s
s
t
_
dt +
s
dW
s
t
(Q) (A.3)
where

Q
r
=
r

rs
(A.4)

Q
r
=
P
r
+
rr

r
+
r

+
sr

rs
(A.5)

Q
r
=
P
r
+
r

r
+

+
s

rs
(A.6)

Q
rs
=
P
r
+
rs

r
+
s

+
ss

rs
(A.7)

=
P

s
(A.8)

Q
r
=
r

+
sr

s
(A.9)

=
P

+
s

s
(A.10)

Q
s
=
s

+
ss

s
(A.11)

Q
s
=
s

s
(A.12)

Q
sr
=
sr

s
(A.13)

Q
s
=
s

s
(A.14)

Q
s
=
P
s
+
ss

s
(A.15)
If the risk premia are completely ane, as in DS1, the parameters in boxes disappear, and
underlined parameters become equal to P parameters.
Equations (4.5)-(4.7), (4.8), and (4.12)-(4.14) imply that under the measure Q the A
1
(3)
34
model evolves according to:
dr
t
=
_

Q
r

Q
r
r
t

Q
r

t

Q
rv
v
t
_
dt
+
_

2
r
+v
t
dW
r
t
(Q) +
r
_

v
t
dW

t
(Q) +
rv

v
t
dW
v
t
(Q) (A.16)
d
t
=
_

Q
r
r
t

Q

t

Q
v
v
t
_
dt
+
r
_

2
r
+v
t
dW
r
t
(Q) +
_

v
t
dW

t
(Q) +
v

v
t
dW
v
t
(Q) (A.17)
dv
t
=
_

P
v
v
Q
v
v
t
_
dt +
v

v
t
dW
v
t
(Q) (A.18)
where

Q
r
=
P
rv
v
r

2
r

r
(A.19)

Q
r
=
P
r
+
rr
+
r

r
(A.20)

Q
r
=
P
r
+
r
+

r
(A.21)

Q
rv
=
P
rv
+
r
+
rv
+

r
+
v

r
+
v

rv
(A.22)

=
P

2
r

(A.23)

Q
r
=
rr

r
+
r
(A.24)

=
P

+
r

r
+

(A.25)

Q
v
=
r

r
+
rv

r
+

+
v
+
v

v
(A.26)

Q
v
=
P
v
+
v

v
(A.27)
As in the A
0
(3) model, if the risk premia are completely ane, the parameters in boxes
disappear, and underlined parameters become equal to P parameters.
B The USV Constraints
We use USV conditions (38)-(41) from CDGJ. In addition, CDGJ impose the condition
2
r

=
0 to ensure that A
1
(3) does not reduce to a two-factor model. These conditions can be rewritten
in terms of the Qparameters:
35

Q
r
+
Q
r

r
+ 2
Q
r

2
r
= 0 (B.1)
3
Q
r

r
+ 2
Q
r

Q

= 0 (B.2)

Q
rv

v

Q
r
(
rv

r

v
) = 0 (B.3)

Q
rv
(
Q
v

Q

) +
Q
r

Q
v
1
2
rv
= 0 (B.4)

= 0 (B.5)
Next, we replace the Qparameters by the combination of the risk premia and the
Pparameters based on the expressions in the appendix A and obtain the following restrictions
on the parameters to be estimated:

= 3
r
(
P
r
+
r

) + 2(
P
r
+
rr
+
r

r
) + 2
r

(B.6)

P
v
=
P

+
r

r
+

v
+
1 +
2
rv
+ (
P
r

r

r

)(
r

r
+
r

rv
+
v
+
v

v
)
k
P
rv
+
r
+
rv
+
r

v
+
rv

v
(B.7)

P
rv
=
1

v
(
P
r
+
r
+
r

)(
rv

r

v
)
r

rv

r

v

rv

v
(B.8)

= 0 (B.9)

r
=

r
1 +
r

r
_
k
P
r
+ 2
r
(
P
r
+
r
+
r

) + 2
rr
_
(B.10)
C Option Pricing Formula
First, we need an additional computation. Using (4.17) we get:
(k f
t+
c
(
f

c
))
+
=
_
k +
1
d

1
d
e
A
P
(d)+A
f
(
f

c
)+B
f
(
f

c
)S
t+
c
_
+
(C.1)
=
1
d
e
A
P
(d)+A
f
(
f

c
)
_
_
(dk + 1)e
A
P
(d)A
f
(
f

c
)
. .
X
e
B
f
(
f

c
)S
t+
c
_
_
+
Substituting this back into (4.20) we obtain:
C
t
(
f
,
c
, K) =
1
d
e
A
P
(d)+A
f
(
f

c
)
E
Q
t
_
e

_
t+
c
t
r
s
ds
_
X e
B
f
(
f

c
)S
t+
c
_
1
{B
f
(
f

c
)S
t+
c
log X}
_
36
=
1
d
e
A
P
(d)+A
f
(
f

c
)
(C.2)

_
XG
0,B
f
(
f

c
)
(log X, S
t
, t +
c
, Q) G
B
f
(
f

c
),B
f
(
f

c
)
(log X, S
t
, t +
c
, Q)
_
where
G
a,b
(y, S
t
, T, M) = E
M
t
_
e

_
T
t
r
s
ds
e
aS
T
1
{bS
T
y}
_
(C.3)
which is computed via the extended characteristic function:

M
(u, S
t
, t, T) = E
M
t
_
e

_
T
t
r
s
ds
e
uS
T
_
= e
A(Tt)+B(Tt)S
t
(C.4)
with boundary conditions A(0) = 0, B(0) = u. Given ,
G
a,b
(y, S
t
, T, M) =

M
(a, S
t
, t, T)
2

1

_

0

M
(a +ivb, S
t
, t, T)
_
v
dv (C.5)
Standard numerical integration techniques allow to compute the integral.
40
D Identication
Global identication requires a unique set of parameters, which maximizes the log-likelihood
L. However, the most operational notion of identication is that of local identication which
stipulates that the information matrix has to be of full rank. Unfortunately, even in the case
of QML, it is dicult to explore this matrix analytically because the expression involves the
Ricatti ODEs solutions. Numerical double dierentiation is very inaccurate and is required
to be performed at multiple points if the rank is not full at one point, there is no way to
ascertain that it will not be full at another one.
40
We note that it is harder to verify the applicability of the Feynman-Kac formula in the case of option pricing
because Levendorskii (2004) requires boundedness of the payo function. Though Sergei Levendorskii told us
that his approach is extendible to options payos, no results are available in print. Deriving full consistency
conditions is beyond our expertise. Nonetheless, we derive conditions necessary for the integral (C.5) to converge.
However, in principle, these conditions do not have to be sucient for the ODE-based formula to coincide with
the expected value in (4.20). It turns out that the conditions are simple:
2
r
0, and
2

0. This explains our

notation, which eectively guarantees the fulllment of the condition.
37
Ane term structure literature proposed another strategy based on factor rotations.
The idea is that all factor rotations that produce the same observable yield produce equivalent
models, so one has to select the smallest model. Unfortunately, it is hard to disentangle
parameters
P
m
and
Q
m
because they are combined in the yield equations. Things become
easier if one assumes observability of some of the state variables, such as yield at zero maturity
(spot interest rate) as in DS1, or the rst three derivatives of the term structure at zero
maturity (spot interest rate, risk-neutral mean, and volatility) as in CDGJ. Such assumptions
allow us to say something about identication of
P
m
and

m
separately from identication of

Q
m
.
These assumptions are reasonable because short yields are in practice very close to
implied spot rate irrespective if the model. Nonetheless, in practice such identication scheme
would work best for the studies which explicitly assume that observed yield is a proxy for
the spot rate because this assumption is equivalent to assuming zero prices of risk. If a
researcher wants to estimate the risk premia, such methodology may not be able to determine
all unidentied parameters.
Therefore, we use a supplemental methodology, which does not require separation of P
and Q parameters. As a rst step, we focus on a more specic identication problem. Some of
the parameters will not be identied if one can nd combinations of the original parameters,
which produce identical value of the likelihood, i.e. at time t the log-likelihood L, as a function
of
m
, could be represented as:
L(t,
m
) = G (t,
1
(
m
), . . . ,
n
(
m
)) t (D.1)
where n < dim(
m
). For example, if the price of risk
v
is not identied, then the vector (
P
v
,
v
)
can be replaced by =
P
v
+
v

v
(see (A.27)). In order to see if this is the case, we check if the
matrix L(t,
m
) /
m
, t = 1, . . . , T has a full rank for at least one set of parameters values.
Continuing our example, if
v
is not identied, then L(t,
m
) /
P
v
=
v
L(t,
m
) /
v
for
all t. This will decrease the rank of the matrix by one.
38
To implement this procedure we rst simulate time series of yields of length T = 391 from
a model m given an arbitrary choice of the parameter
m
. Then we take numerical derivatives
with respect to each parameter at each point in time. The derivatives matrices had a full rank
for both models.
41
The above procedure does not allow us to say anything if unidentiability occurs because
there exist at least two sets of parameters (of the same size) which yield the same likelihood
value. We deal with this at the estimation stage by performing global optimum robustness
checks.
42
We do not encounter typical manifestations of unidentied parameters, such as at
objective function. However, theoretically our identication procedure is not complete.
E Models Estimation via Kalman Filter
The preferred estimation methodology is maximum likelihood. However, with the exception
of the A
0
(3) model estimated on yields only, the likelihood is not available in analytical form.
Therefore, we implement quasi-maximum likelihood (QML), which entails approximating non-
Gaussian states by the Gaussian ones with the mean and variance (see, for instance, Fisher
and Gilles, 1996).
An additional complication arises from the fact that the states S are not observable in
the presence of measurement errors. In the case of the A
0
(3) model and futures rates only, one
can estimate the system with errors via Kalman lter. Indeed, the state evolves according to
the Gaussian system, the function g
f
is linear, which implies that the ML based on Kalman
41
This result may be surprising in the light of DS comments about unidentiability of some risk premia
parameters associated with Gaussian factors. What helps us in identication of all parameters is the availability
of more yields than factors (see also de Jong, 2000).
42
In particular, we use a very large and ecient set of starting value to ensure that the global optimum is
found. The grid search is extremely costly in a multi-dimensional space, and, in practice, limits the extent of
the global search. We reduce the computational costs by using the Sobol quasi-random sequences to generate
the starting points (see, e.g. Press, Teukovsky, Vetterling, and Flannery, 1992).
39
lter is the optimal estimation methodology.
43
However, if we consider non-linear function in
the measurement equation, such as g
c
, or non-Gaussian states such as the v factor in A
1
(3),
exact lter is generally not computationally feasible, and the Kalman lter is no longer optimal.
When the measurement equation is non-linear, one could use Extended Kalman lter
(EKF), which is eectively the regular Kalman lter applied to the rst order Taylor expansion
of the measurement functions (g
c
). When the state is non-Gaussian, one could use Quasi
Kalman lter (QKF), which replaces such a state (v) with a Gaussian state with identical rst
two moments (Duan and Simonato, 1995, and Chen and Scott, 2003). Both lters are the best
linear lters, but QML estimates are conditionally biased and inconsistent.
Despite the shortcomings, we expect our choice of estimation method to perform well.
EKF applied to high signal-to-noise ratio systems is known to have good properties asymptotic
in

(Picard, 1991). Monte-Carlo studies of the QKF applied to multi-factor square-root

processes in de Jong (2000), Duan and Simonato (1995) and Chen and Scott (2003) show
that, in practice, this procedure introduces minimal biases.
44
Moreover, studies attempting to
improve upon QKF in a theoretically sound way nd little dierence in practice (see, e.g. Lund
(1997) and Fr uhwirth-Schnatter and Geyer (1998)). Since we have at most one square-root
factor and only two non-linear measurement equations, these results suggest that we might
obtain reasonable estimates despite suboptimal asymptotic properties of the lters. In order
to guard ourselves against potential pitfalls of the methodology, we conduct nite sample
inference by bootstrapping the standard errors and by computing nite sample bias in the
estimates (see Conley, Hansen, and Liu, 1997). The procedure entails simulating multiple
paths from the model given the estimated parameters, and reestimating it along each path.
If the measurement equation (2.1) is linear, as is the case with g
f
in (4.21), and the
43
This observation highlights another advantage of using Eurodollar futures versus the swap rates. Since the
par representation formula used for swaps is non-linear in the state vector, Kalman lter will not be directly
applicable.
44
Duee and Stanton (2004) provide additional favorable evidence for this methodology and contrast it with
EMM.
40
state, S, equation is Gaussian, as is the case with (4.2)-(4.4), then we can apply the linear
Kalman lter directly. We start the recursion with the initial estimated value of the state and
its variance:

S
1|0
= E(S
1
) (E.1)

1|0
= E
_
(S
1
E(S
1
))(S
1
E(S
1
))

_
(E.2)
Then, assuming that we computed

S
t|t1
and
t|t1
at time t, we proceed with the update
step. We need to compute

S
t|t
= E(S
t
|y
f
t
, F
t1
) (E.3)
In order to do this we need to know the joint distribution of S
t
and y
t
conditional on F
t1
.
Under our normality assumptions, this distribution is normal with mean:
E(S
t
|F
t1
) =

S
t|t1
(E.4)
E(y
f
t
|F
t1
) = g
f
(

S
t|t1
) = A+B

S
t|t1
(E.5)
where A and B are vectors (A
P
(d)+A
f
(
t
))/
t
and B
f
(
t
)/
t
respectively for eleven dierent
maturities, and covariance matrix:
_
_
_
E
_
(S
t


S
t|t1
)(S
t


S
t|t1
)

|F
t1
_
E
_
(y
f
t
g(

S
t|t1
))(S
t


S
t|t1
)

|F
t1
_

E
_
(y
f
t
g(

S
t|t1
))(S
t


S
t|t1
)

|F
t1
_
E
_
(y
f
t
g(

S
t|t1
))(y
f
t
g(

S
t|t1
)

|F
t1
_
_
_
_
=
_
_
_

t|t1

t|t1
B
B

t|t1
B

t|t1
B +
_
_
_
(E.6)
where is the diagonal matrix of the measurement error variances. Now we can compute:

S
t|t
=

S
t|t1
+
t|t1
B(B

t|t1
B +)
1
(y
f
t
AB

S
t|t1
) (E.7)

t|t
=
t|t1

t|t1
B(B

t|t1
B +)
1
B

t|t1
(E.8)
41
based on the properties of bivariate normal distribution.
45
We propagate the lter forward
using the forecasting step, where we compute

S
t+1|t
and
t+1|t
based on the update equations
(E.7) and (E.8), and the normality of the state equation.
Extended Kalman lter (EKF) is used when the measurement equation, such as g
c
in
(4.22), is non-linear. In this case, g
c
is approximated by a linear function:
g
c
(S
t
,
m
) g
c
(

S
t|t
,
m
) +
_
g
c
(S
t
,
m
)
S
t
|
S
t
=

S
t|t
_

(S
t


S
t|t
) (E.9)
See Heidari and Wu (2002) for the option pricing application for Gaussian states.
Quasi Kalman lter (QKF) is used when the state equation, such as v in (4.7), is not
Gaussian. Duan and Simonato (1985) and Chen and Scott (2003) propose the procedure for
the particular case of the square-root processes (see Lund (1997) for extensions). The update
relationships (E.7) and (E.8) rely on the normality of the state equation. QKF approximates
the square-root state variable by a Gaussian state variable with the same conditional mean
and variance.
F Encompassing
We largely follow Gourieroux and Monfort (1994) in our exposition of encompassing. Also, see
Dhaene (1997) for more details.
Suppose the true data-generating process is associated with the log-likelihood L
d
. We are
trying to evaluate the proximity between two non-nested and potentially misspecied models
with log-likelihoods L
0
, i.e. A
0
(3), and L
1
, i.e. A
1
(3). Maximum likelihood estimation of
the respective sets of parameters
0
and
1
(see (4.23)) using the sample y is equivalent to
45
Note, that in the case of the USV models the right-hand side of the equation (E.5) involves only two of
the three state variables if bond yields or swap rates are used in place of y
f
t
. Therefore, based on the equation
(E.7), this feature could lead to very noisy estimates of the unspanned volatility factor. However, the futures
prices are sensitive to volatility even in the presence of USV, and, therefore, the Kalman lter produces accurate
estimates.
42
minimization of the Kullback-Leibler information criterion (KLIC):

i
= arg max

i
E
d
L
i
(y,
i
) = arg min

i
E
d
[L
d
(y) L
i
(y,
i
)] (F.1)
where superscript d emphasizes the true distribution. That is why this estimate is often referred
to as the pseudo-true value.
A binding function links the two misspecied models similar to how pseudo-true value
links a misspecied model to the true distribution:
b
10
(

0
) = arg min

1
E
0
_
L
0
(y,

0
) L
1
(y,
1
)
_
= arg max

1
E
0
[L
1
(y,
1
)] (F.2)
where superscript 0 emphasizes distribution associated with with A
0
(3). In other words, a
binding function is the value of parameter
1
such that the distribution of A
1
(3) approximates
the distribution of A
0
(3) in the best way possible (according to KLIC), or, equivalently, this
is an estimate of
1
if L
0
were the true distribution.
We say that the true distribution L
d
is such that A
0
(3) encompasses A
1
(3), A
0
(3)EA
1
(3),
if and only if the pseudo-true value coincides with the binding function,

1
= b
10
(

0
). Put less
formally, A
0
(3) encompasses A
1
(3) if the objects of interests associated with the latter behave
as they should were L
0
the true distribution. Thus, similar to nested hypothesis testing, the
encompassing model is preferred if it is more simple.
It is natural to consider a likelihood ratio test of encompassing. The implementation of
the tests was hindered by complicated asymptotic distributions and the explicit knowledge of
the binding functions. Since in this paper we rely on nite sample inference, it is easy to avoid
these complications. The nite sample inference is conducted by simulating 800 samples from
the A
0
(3) models and estimating the respective A
1
(3) models based on these samples. This
procedure allows to compute the LR statistic and its distribution.
43
References
Ahn, Dong-Hyun, Robert F. Dittmar, and A. Ronald Gallant, 2002, Quadratic term structure
models: Theory and evidence, Review of Financial Studies 15, 243288.
At-Sahalia, Yacine, 1996, Testing continuous-time models of the spot interest rate, Review of
Financial Studies 9, 38542.
, and Robert Kimmel, 2002, Estimating Ane multifactor term structure models using
closed-form likelihood expansions, Working paper, Princeton University.
Amin, Kaushik, and Andrew Morton, 1994, Implied volatility functions in arbitrage-free term
structure models, Journal of Financial Economics 35, 141180.
Andersen, Torben, Luca Benzoni, and Jesper Lund, 2004, Stochastic volatility, mean drift and
jumps in the short term interest rate, Working Paper, Northwestern University.
Ang, Andrew, and Monika Piazzesi, 2003, A no-arbitrage vector autoregression of term struc-
ture dynamics with macroeconomic and latent variables, Journal of Monetary Economics
50, 745787.
Attari, Mukkaram, 2001, Testing interest rate models: What does futures and options data
tell us?, Working Paper, University of Wisconsin-Madison.
Backus, David, Kai Li, and Liuren Wu, 1998, The hump-shaped mean term structure of
interest rate derivative vols, Working Paper, New York University.
Balduzzi, Pierluigi, Sanjiv Das, Silverio Foresi, and Rangarajan K. Sundaram, 1996, A simple
approach to three-factor ane models of the term structure, Journal of Fixed Income 6,
4353.
Ball, Cliord A., and Walter N. Torous, 1999, The stochastic volatility of short-term interest
rates: Some international evidence, Journal of Finance 54, 23392359.
44
Bansal, Ravi, George Tauchen, and Hao Zhou, 2003, Regime-shifts in term structure, expec-
tations hypothesis puzzle, and the real business cycle, forthcoming, Journal of Business and
Economic Statistics.
Brandt, Michael, and David Chapman, 2003, Comparing multifactor models of the term struc-
ture, Working Paper, Duke University.
Broadie, Mark, Mikhail Chernov, and Michael Johannes, 2004, Model specication and risk
premia: Evidence from futures options, Working Paper, Columbia University.
Buraschi, Andrea, and Alexei Jiltsov, 2004, Time-varying infaltion risk premia and the expec-
tation hypothesis: A monetary model of the Treasury yield curve, forthcoming, Journal of
Financial Economics.
Chen, Ren-Raw, and Louis Scott, 2002, Stochastic volatility and jumps in interest rates: An
empirical analysis, Working Paper, Rutgers University.
, 2003, Multi-factor CIR models of the term structure: Estimates and tests from a
state-space model using a Kalman lter, Journal of Real Estate Finance and Economics 27.
Cheredito, Patrick, Damir Filipovic, and Robert Kimmel, 2003, Market price of risk specica-
tions for ane models: Theory and evidence, Working Paper, Princeton University.
Collin-Dufresne, Pierre, and Robert Goldstein, 2002, Do bonds span the xed income markets?
Theory and evidence for unspanned stochastic volatility, Journal of Finance 57, 16851730.
, and Christopher Jones, 2003, Identication and estimation of maximal ane term
structure models: An application to stochastic volatility, Working Paper, Washington Uni-
versity.
Collin-Dufresne, Pierre, and Bruno Solnik, 2001, On the term structure of default premia in
the swap and LIBOR markets, Journal of Finance 56, 10951115.
45
Conley, Timothy, Lars P. Hansen, and W.F. Liu, 1997, Bootstrapping the long run, Macroe-
conomic Dynamics 1, 279311.
Dai, Qiang, and Kenneth Singleton, 2000, Specication analysis of ane term structure models,
Journal of Finance 55, 19431978.
, 2002, Expectation puzzles, time-varying risk premia, and dynamic models of the term
structure, Journal of Financial Economics 63, 415441.
, 2003, Term structure modeling in theory and reality, Review of Financial Studies 16,
631678.
, and Wei Yang, 2003, Regime shifts in a dynamic term structure model of U.S. Treasury
bond yields, Working Paper, Stanford University.
, 2004, Predictability of bond risk premia and ane term structure models, Working
Paper, Stanford University.
Das, Sanjiv, 2002, The surprise element: Jumps in interest rates, Journal of Econometrics 106,
2765.
de Jong, Frank, 2000, Time series and cross-section information in ane term-structure models,
Journal of Business and Economic Statistics 18, 300314.
Dhaene, Geert, 1997, Encompassing: Formulation, Properties and Testing (Springer: Heidel-
berg).
Duan, Jin-Chuan, and Jean-Guy Simonato, 1995, Estimating and testing exponential-ane
term structure models by Kalman lter, Working Paper, CIRANO.
Duee, Gregory, 2002, Term premia and interest rate forecasts in ane models, Journal of
Finance 57, 405443.
46
, and Richard Stanton, 2004, Estimation of dynamic term structure models, Working
paper, University of California at Berkeley.
Due, Darrell, Damir Filipovic, and Walter Schachermayer, 2003, Ane processes and appli-
cations in nance, Annals of Applied Probability 13, 9841053.
Due, Darrell, and Rui Kan, 1996, A yield-factor model of interest rates, Mathematical Finance
6, 379406.
Due, Darrell, Jun Pan, and Kenneth Singleton, 2000, Transform analysis and asset pricing
for ane jump-diusions, Econometrica 68, 13431376.
Due, Darrell, Lasse Pedersen, and Kenneth Singleton, 2003, Modeling sovereign yield spreads:
A case study of Russian debt, Journal of Finance 58, 119160.
Due, Darrell, and Kenneth Singleton, 1997, An econometric model of the term structure of
interest-rate swap yields, Journal of Finance 52, 12871321.
Fisher, Mark, and Christian Gilles, 1996, Estimating exponential-ane models of the term
structure, Working Paper, Federal Reserve.
Flesaker, Bjorn, 1993, Testing the Heath-Jarrow-Morton/Ho-Lee model of interest rate con-
tingent claims pricing, Journal of Financial and Quantitative Analysis 28, 483495.
Fr uhwirth-Schnatter, Sylvia, and Alois L.J. Geyer, 1998, Bayesian estimation of econometric
multi-factor Cox-Ingersoll-Ross models of the term structure of interest rates via MCMC
methods, Working Paper, Vienna University of Economics and Business Administration.
Gallant, A. Ronald, and George Tauchen, 1996, Which moments to match?, Econometric
Theory 12, 657681.
, 1998, Reprojecting partially observed systems with application to interest rate diu-
sions, Journal of American Statistical Association 93, 1024.
47
Gourieroux, Christian, and Alain Monfort, 1994, Testing non-nested hypothesis, in R. F.
Engle, and D.L. McFadden, ed.: Handobook of Econometrics, Vol. IV (Elsevier Science:
Amsterdam).
Heidari, Massoud, and Liuren Wu, 2002, Term structure of interest rates, yield curve residuals,
and the consistent pricing of interest rates and interest rate derivatives, Working Paper,
Baruch College.
, 2003, Are interest rate derivatives spanned by the term structure of interest rates?,
Journal of Fixed Income 13, 7586.
Jagannathan, Ravi, Andrew Kaplin, and Steve Sun, 2003, An evaluation of multi-factor CIR
models using LIBOR, swap rates, and cap and swaption prices, Journal of Econometrics
116, 113146.
Jegadeesh, Narasimhan, and George G. Pennacchi, 1996, The behavior of interest rates implied
by the term structure of Eurodollar futures, Journal of Money, Credit and Banking 28, 426
446.
Johannes, Michael, 2004, The statistical and economic role of jumps in interest rates, Journal
of Finance 59, 227260.
, and Suresh Sundaresan, 2003, Collateralized swaps, Working Paper, Columbia Uni-
versity.
Lamoureux, Christopher G., and H. Douglas Witte, 2002, Empirical analysis of the yield curve:
The information in the data viewed through the window of Cox, Ingersoll, and Ross, Journal
of Finance 57, 14791520.
Langetieg, Terence C., 1980, A multivariate model of the term structure, Journal of Finance
35, 7197.
48
Levendorskii, Sergei, 2004, Consistency conditions for ane term structure models, Stochastic
Processes and Their Applications 109, 225261.
Litterman, Robert, and Jose Scheinkman, 1991, Common factors aecting bond returns, Jour-
nal of Fixed Income 3, 3461.
, and Laurence Weiss, 1991, Volatility and the yield curve, Journal of Fixed Income
pp. 4953.
Lund, Jesper, 1997, Econometric analysis of continuous-time arbitrage-free models of the term-
structure of interest rates, Working Paper, The Aarhus School of Business.
Piazzesi, Monika, 2003, Bond yields and the Federal Reserve, forthcoming, Journal of Political
Economy.
, 2005, Ane term structure models, in Y. At-Sahalia, and L.P. Hansen, ed.: Hando-
book of Financial Econometrics (Elsevier Science: Amsterdam) fortchcoming.
Picard, Jean, 1991, Eciency of the extended Kalman lter for nonlinear systems with small
noise, SIAM Journal on Applied Mathematics 51, 843885.
Press, William H., Saul A. Teukovsky, William T. Vetterling, and Brian P. Flannery, 1992,
Numerical Recipies in C (Cambridge University Press: Cambridge) second edn.
Pritsker, Matthew, 1998, Nonparametric density estimation and tests of continuos time interest
rate models, Review of Financial Studies 11, 449487.
Sato, Kenichi, 1999, Levy processes and innitely divisible distributions (Cambridge University
Press: Cambridge).
Singleton, Kenneth, and Leonid Umantsev, 2002, Pricing coupon-bond options and swaptions
in ane term structure models, Mathematical Finance 12, 427446.
49
Thompson, Samuel, 2004, Identifying term structure volatility from the LIBOR-swap curve,
Working Paper, Harvard University.
Umantsev, Leonid, 2001, Econometric analysis of European LIBOR-based options within ane
term-structure models, Ph.D. dissertation, Stanford University.
50
Table 1 : Estimated parameters for model A
0
(3)
The model A
0
(3) dynamics is given by
dr
t
=
P
r
(
t
+ s
t
r
t
) dt +
r
dW
r
t
(P) +
r

dW

t
(P) +
rs
dW
s
t
(P)
d
t
=
P

t
_
dt +

dW

t
(P) +
s
dW
s
t
(P)
ds
t
=
P
s
s
t
dt +
s
dW
s
t
(P)
and the risk premia are:

r
t
=
r
+
rr
r
t
+
r

t
+
rs
s
t

t
=

+
r
r
t
+

t
+
s
s
t

s
t
=
s
+
sr
r
t
+
s

t
+
ss
s
t
The bootstraped 95% condence intervals are reported in square brackets.
Model A
0
(3)
f
A
0
(3)
fo
Estimate Condence interval Estimate Condence interval

P
r
0.90 [0.59, 1.49] 1.18 [0.48, 2.36]

1.84 [1.17, 3.63] 1.15 [0.64, 1.76]

P
s
0.16 [0.07, 0.46] 0.88 [0.41, 1.26]

10
2
0.30 [0.00, 0.73] 3.43 [0.64, 5.89]

r
10
2
0.87 [0.73, 0.95] 0.70 [0.58, 0.82]

0.02 [0.01, 0.04] 0.02 [0.01, 0.07]

s
0.03 [0.02, 0.05] 0.04 [0.02, 0.05]

r
0.06 [0.01, 0.18] -0.15 [0.30, 0.05]

rs
10
2
0.35 [0.24, 0.48] 0.10 [0.05, 0.41]

s
-0.01 [0.03, 0.00] -0.02 [0.05, 0.00]

r
-3.88 [4.91, 3.24] -2.85 [5.39, 2.11]

0 0

s
-0.14 [0.16, 0.13] 0

rr
-11.54 [81.78, 24.80] 12.27 [109.95, 59.70]

r
75.46 [38.04, 148.32] 0

rs
67.21 [39.89, 131.80] 34.37 [4.91, 81.21]

r
0 -10.17 [12.87, 3.69]

-56.74 [168.65, 31.87] 0

s
0.86 [5.24, 3.30] 7.40 [12.52, 20.89]

sr
0 14.71 [2.15, 25.79]

s
0 -20.64 [35.89, 6.11]

ss
-5.00 [14.54, 2.66] -35.45 [46.36, 22.06]
log-likelihood 73.66 85.09
Table 2 : Estimated parameters for model A
1
(3)
The model A
1
(3) dynamics is given by
dr
t
=
P
r
(
t
r
t
) dt +
P
rv
( v v
t
) dt +
_

2
r
+ v
t
dW
r
t
(P) +
r
_

v
t
dW

t
(P) +
rv

v
t
dW
v
t
(P)
d
t
=
P

t
_
dt +
r
_

2
r
+ v
t
dW
r
t
(P) +
_

v
t
dW

t
(P) +
v

v
t
dW
v
t
(P)
dv
t
=
P
v
( v v
t
) dt +
v

v
t
dW
v
t
(P)
and the risk premia are:

r
t
=
1

2
r
+ v
t
_

2
r
+
rr
r
t
+
r

t
+ (
r
+
rv
)v
t
_

t
=
1
_

v
t
_

+
r
r
t
+

t
+ (

+
v
)v
t
_

v
t
=
v

v
t
The bootsraped 95% condence bounds are reported in square brackets. An asterisk (*) denotes parameters
signicant at the 10% level. A dagger () denotes parameters restricted by the USV conditions (B.6)-(B.10).
Model A
1
(3)
f
A
1
(3)
fo
A
USV
1
(3)
f
A
USV
1
(3)
fo
Est. Conf. int. Est. Conf. int. Est. Conf. int. Est. Conf. int.

P
r
0.34 [0.32, 0.36] 1.19 [0.94, 1.39] 0.26 [0.21, 0.63] 0.11 [0.11, 0.12]

0.74 [0.50, 1.40] 1.06 [0.78, 1.64] 0.40

[0.39, 0.41] 0.20

[0.09, 0.92]

P
v
0.71 [0.59, 0.88] 0.41 [0.24, 1.06] 0.25

[0.14, 0.44] 1.03

[0.65, 2.27]

P
rv
10
2
-0.57 [10.17, 1.10] -4.99 [41.98, 18.11] 0.44

[0.05, 6.16] 0.87

[0.06, 1.24]

10
2
3.39 [0.10, 6.64] 4.37 [2.52, 5.08] 5.75 [0.56, 10.62] 3.90 [0.02, 9.16]
v 10
4
0.69 [0.52, 0.85] 0.11 [0.04, 0.21] 6.50 [3.82, 11.43] 0.97 [0.65, 1.70]

v
10
2
0.25 [0.09, 0.40] 0.31 [0.26, 0.39] 1.82 [1.55, 1.95] 1.35 [0.97, 2.18]

r
0 0 0 0

10
2
2.81 [1.97, 3.51] 1.60 [1.42, 1.83] 2.76 [1.36, 3.86] 9.43 [5.60, 14.26]

4.15

[0.00, 11.98] 5.28

[0.00, 6.11] 0

r
0.11 [0.07, 0.16] -0.34 [0.42, 0.26] -0.15 [0.41, 0.05] -0.00 [0.01, 0.01]

rv
-0.21 [0.39, 0.01] 0.16 [2.01, 0.34] 0.29 [0.14, 0.44] 0.10 [0.05, 0.30]

r
0 5.59 [3.60, 7.44] 0 0

v
5.06 [1.33, 8.04] -7.91 [17.36, 0.14] 2.98 [0.00, 45.59] -0.71 [1.33, 0.06]

r
10
2
-3.31 [4.44, 2.74] -1.05 [3.18, 0.72] -0.10 [0.27, 0.10] 0

10
2
-0.91 [1.26, 0.51] 2.76 [2.21, 3.48] 0.58 [0.26, 1.42] 0.02 [0.01, 0.07]

v
10
2
0.36 [0.28, 0.92] 0.15 [1.87, 0.72] -0.13 [0.24, 0.08] -0.76 [1.29, 0.46]

rr
0 -0.13 [0.17, 0.10] -0.06 [0.46, 0.01] 0

r
0.32 [0.29, 0.41] 0.05 [0.01, 0.08] 0.05 [0.08, 0.36] 0.07 [0.04, 0.09]

rv
0 0 0 -81.82 [112.96, 0.10]

r
0 0 0

-0.74 [1.40, 0.50] -0.69 [1.21, 0.37] 0 0.03 [0.70, 0.14]

v
10
2
16.98 [11.12, 22.79] -19.57 [34.34, 4.95] -0.29 [0.61, 0.02] -0.66 [1.13, 0.29]
log-likelihood 73.67 86.35 72.30 81.95
52
Table 3 : Correlations of state variables with observables
We want to asses which aspects of the term structure are reected in the model state variables. We, therefore,
compute correlations between the model implied state variables and their most plausible proxies in the data.
Correlations in boldface denote the highest correlation for the particular factor.
Observables Short rate Long rate Slope Buttery Implied Variance
Factor Model y
f
(0.5) y
f
(10) y
f
(10) y
f
(0.5) y
f
(0.5) + y
f
(10) (
imp
(0.5) y
f
(0.5))
2
2y
f
(2)
r A
0
(3)
f
0.94 0.29 -0.74 0.01 -0.01
A
0
(3)
fo
0.69 0.14 -0.61 0.45 -0.05
A
1
(3)
f
0.94 0.29 -0.74 0.02 -0.01
A
1
(3)
fo
0.75 0.14 -0.69 0.37 -0.08
A
USV
1
(3)
f
0.98 0.38 -0.70 -0.15 0.02
A
USV
1
(3)
fo
0.98 0.38 -0.70 -0.20 0.04
A
0
(3)
f
0.32 -0.36 -0.69 -0.52 -0.31
A
0
(3)
fo
-0.02 -0.77 -0.68 -0.05 -0.57
A
1
(3)
f
0.30 0.83 0.45 -0.77 0.45
A
1
(3)
fo
0.72 0.47 -0.34 -0.77 0.04
A
USV
1
(3)
f
0.42 0.98 0.45 -0.28 0.53
A
USV
1
(3)
fo
0.04 0.83 0.72 -0.52 0.44
s A
0
(3)
f
0.48 0.99 0.40 -0.38 0.53
A
0
(3)
fo
0.48 0.98 0.38 -0.32 0.55
v A
1
(3)
f
0.02 0.41 0.36 -0.93 0.24
A
1
(3)
fo
-0.55 0.31 0.89 0.24 0.43
A
USV
1
(3)
f
0.51 0.96 0.35 -0.16 0.50
A
USV
1
(3)
fo
-0.02 0.45 0.44 0.22 0.75
53
Table 4 : Pricing Errors
We report average absolute error (reported in basis points) in futures yields, |y
f
(, data) y
f
(, model)|
(for selected maturities ) and in absolute options-implied volatilities (reported in percent)
|[
imp
(, data)
imp
(, model)] y
f
(, data)|.
Futures errors (b.p.s.) Options errors (%)
Maturity 0.5 2 7 10 0.5 1
A
0
(3)
f
3.91 3.70 2.40 5.06 0.1598 0.1723
A
0
(3)
fo
7.44 6.76 2.07 7.83 0.0353 0.0484
A
1
(3)
f
3.84 3.75 2.43 5.00 0.1303 0.1218
A
1
(3)
fo
5.32 5.37 2.02 7.69 0.0359 0.0381
A
USV
1
(3)
f
5.97 5.05 2.36 3.91 1.5930 1.5722
A
USV
1
(3)
fo
7.42 9.84 2.94 10.94 0.0873 0.2448
Table 5 : Encompassing Tests
We test whether models from the A
0
(3) branch, or the A
USV
1
(3) branch encompass models from the A
1
(3)
branch based on the nite sample likelihood-ratio test. The inference is conducted by simulating 800 samples
from the A
0
(3), or the A
USV
1
(3) models and estimating the respective A
1
(3) models based on these samples.
This procedure allows to compute the LR statistic and its distribution.
Test LR Conf. int. pvalue
A
0
(3)
f
EA
1
(3)
f
0.008 [-0.011,0.047] 0.490
A
0
(3)
fo
EA
1
(3)
fo
1.385 [1.408,7.304] 0.025
A
USV
1
(3)
f
EA
1
(3)
f
1.365 [0.002,0.019] 0.000
A
USV
1
(3)
fo
EA
1
(3)
fo
4.415 [0.001,0.071] 0.000
54
Table 6 : The USV restrictions tests
We test whether the USV conditions hold in the data. In order to do this, we have to check whether restrictions
(B.6)-(B.10) hold in the model A
1
(3) estimated without imposing such restrictions. We simulated 800 paths,
estimated the model for every path and constructed the 5% condence bounds for the null hypothesis that
LHS-RHS in (B.6)-(B.10) is equal to zero.
Restriction A
1
(3)
f
A
1
(3)
fo

[0.72, 0.65] [8.26, 16.72]

P
v
[0.94, 5.11] [35.20, 48.68]

P
rv
[1367.50, 233.55] [1036.00, 3344.30]

[0.00, 12.18] [0.00, 6.14]

r
[0.00, 0.00] [34.33, 122.43]
55
Table 7 : Conditional mean and volatility correlations
This table complements Figures 2 and 3 by reporting correlations between conditional mean and volatility
computed from the data and the respective moments from the models.
Level Slope Curvature
y
f
(0.5) y
f
(10) y
f
(0.5) y
f
(0.5) +y
f
(10) 2y
f
(2)
VAR(1) GARCH(1,1) VAR(1) GARCH(1,1) VAR(1) GARCH(1,1)
A
0
(3)
f
0.9882 0.9001 0.9379 0.8974 0.8595 0.8663
A
0
(3)
fo
0.9701 0.8030 0.9003 0.8478 0.8170 0.5048
A
1
(3)
f
0.9878 0.9010 0.9375 0.8952 0.8621 0.8665
A
1
(3)
fo
0.9811 0.8133 0.9082 0.7979 0.8590 0.5209
A
USV
1
(3)
f
0.9786 0.8919 0.9062 0.8614 0.7763 0.7246
A
USV
1
(3)
fo
0.9723 0.7354 0.8087 0.7673 0.7282 0.2125
56
Table 8 : Kurtosis Tests
We test whether the estimated models can match unconditional kurtosis of changes in the principal components
of the log-futures (y
f
()). We construct nite sample 95% condence intervals by simulating 800 sample paths
from each model, and evaluating the kurtosis along each path. We compare the intervals with the sample
values. Boldfaced statistics indicate a failure to reject the null that a model can match a statistic.
Level Slope Curvature
y
f
(0.5) y
f
(10) y
f
(0.5) y
f
(0.5) + y
f
(10) 2y
f
(2)
Sample 4.31 4.70 5.95
Model Value Conf. int. Value Conf. int. Value Conf. int.
A
0
(3)
f
2.98 [2.38, 3.51] 2.99 [2.38, 3.56] 2.98 [2.36, 3.53]
A
0
(3)
fo
3.00 [2.60, 3.54] 2.98 [2.58, 3.55] 2.98 [2.58, 3.54]
A
1
(3)
f
3.02 [2.60, 3.59] 3.00 [2.61, 3.58] 2.99 [2.58, 3.55]
A
1
(3)
fo
3.46 [2.77, 4.77] 3.17 [2.66, 4.03] 3.00 [2.58, 3.60]
A
USV
1
(3)
f
4.05 [2.93, 6.30] 3.66 [2.83, 5.21] 3.00 [2.59, 3.58]
A
USV
1
(3)
fo
4.23 [2.68, 5.17] 3.00 [2.59, 3.59] 2.98 [2.57, 3.52]
57
Table 9 : R
2
of regression of implied volatility on term structure principal
components
We compute the R
2
of the regression of implied volatility on the three principle components. Then we construct
nite sample distribution of this statistics by simulating 800 paths from each of the models, and recomputing
the R
2
along each path. Boldfaced values indicate a rejection of the null that a particular model can replicate
the R
2
observed in the data.
Option Maturity 6 months 1 year
Sample R
2
0.34 0.46
Model R
2
Conf. int. R
2
Conf. int.
A
0
(3)
f
0.03 [0.00, 0.06] 0.01 [0.00 0.02]
A
0
(3)
fo
0.06 [0.01, 0.12] 0.00 [0.00,0.02]
A
1
(3)
f
0.16 [0.02, 0.42] 0.30 [0.14,0.58]
A
1
(3)
fo
0.47 [0.22, 0.77] 0.45 [0.10, 0.83]
A
USV
1
(3)
f
0.81 [0.38, 0.97] 0.74 [0.13, 0.97]
A
USV
1
(3)
fo
0.39 [0.23, 0.70] 0.27 [0.08, 0.65]
58
Figure 1. Term structure response to shocks in the state variables.
On this gure we plot how term structure changes in response to one standard deviation shock in one
of the three state variables. In order to disentangle the role of the risk premia (in particular, ane vs
essentially ane) from the role of the state variables, we evaluate the term structure assuming zero risk
premia (denoted by P on the plot), assuming zero essentially ane components of the risk premia (ane
Q on the plot), and nally using the full risk premia (ess Q).
0 5 10
0.2
0.3
0.4
0.5
0.6
0.7
c
h
a
n
g
e

i
n

z
e
r
o

y
i
e
l
d
s
(
%
)
shock in r
0 5 10
0.1
0.15
0.2
0.25
c
h
a
n
g
e

i
n

z
e
r
o

y
i
e
l
d
s
(
%
)
shock in
0 5 10
0.4
0.6
0.8
1
1.2
maturity
c
h
a
n
g
e

i
n

z
e
r
o

y
i
e
l
d
s
(
%
)
shock in s
0 5 10
0.1
0.2
0.3
0.4
0.5
shock in r
0 5 10
0.3
0.4
0.5
0.6
0.7
0.8
shock in
0 5 10
0.4
0.6
0.8
1
maturity
shock in s
0 5 10
0.3
0.4
0.5
0.6
0.7
0.8
shock in r
0 5 10
0.2
0.4
0.6
0.8
shock in
0 5 10
0.2
0.1
0
0.1
0.2
0.3
maturity
shock in v
0 5 10
0.1
0.2
0.3
0.4
0.5
0.6
shock in r
0 5 10
0.2
0.3
0.4
0.5
0.6
0.7
shock in
0 5 10
0.1
0.2
0.3
0.4
0.5
maturity
shock in v
0 5 10
0.4
0.5
0.6
0.7
0.8
shock in r
0 5 10
1
1.5
2
shock in
0 5 10
2
1.5
1
0.5
maturity
shock in v
0 5 10
0.5
0.55
0.6
0.65
0.7
0.75
shock in r
0 5 10
0.5
1
1.5
shock in
0 5 10
10
8
6
4
2
0
maturity
shock in v
ess Q
affine Q
P
A
0
(3)
f
A
0
(3)
fo
A
1
(3)
f

A
1
(3)
fo
A
1
(3)USV
f
A
1
(3)USV
fo

59
Figure 2. VAR conditional means of the principal components.
We plot conditional means estimated from the VAR(1)-GARCH(1,1) model estimated on the three prin-
cipal components constructed from the futures from the data and the models. The rst row reports the
results for the level, y
f
(0.5), the second row reports for the slope, y
f
(10) y
f
(0.5), and nally the last
row reports for the buttery spread y
f
(0.5) + y
f
(10) 2y
f
(2).
1994 1996 1998 2000
0.05
0
0.05
0.1
0.15
Futures
1994 1996 1998 2000
0
0.02
0.04
0.06
0.08
1994 1996 1998 2000
0.04
0.02
0
0.02
1994 1996 1998 2000
0.05
0
0.05
0.1
0.15
Futures and options
1994 1996 1998 2000
0
0.02
0.04
0.06
0.08
1994 1996 1998 2000
0.04
0.02
0
0.02
data
A
0
(3)
A
1
(3)
A
1
(3)USV
60
Figure 3. GARCH conditional variances of the principal components.
We plot conditional variances estimated from the VAR(1)-GARCH(1,1) model estimated on the three
principal components constructed from the futures from the data and the models. The rst row reports
the results for the level, y
f
(0.5), the second row reports for the slope, y
f
(10) y
f
(0.5), and nally the last
row reports for the buttery spread y
f
(0.5) + y
f
(10) 2y
f
(2).
1994 1996 1998 2000
0
2
4
6
x 10
6
Futures
1994 1996 1998 2000
0
2
4
6
8
x 10
6
1994 1996 1998 2000
0
2
4
6
x 10
6
1994 1996 1998 2000
0
2
4
6
x 10
6
Futures and options
1994 1996 1998 2000
0
2
4
6
8
x 10
6
1994 1996 1998 2000
0
2
4
6
x 10
6
data
A
0
(3)
A
1
(3)
A
1
(3)USV
61
Figure 4. Term Structure of unconditional and implied volatilities.
We plot population standard deviation and average implied volatility evaluated at the estimates of the
respective models. We compare these with the standard deviation and average implied volatility computed
from our sample.
0 2 4 6 8 10
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
Standard deviation
maturity
v
o
l
a
t
i
l
i
t
y
,
%
0 2 4 6 8 10
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
Implied volatility
maturity
v
o
l
a
t
i
l
i
t
y
,
%
A
0
(3)
f
A
0
(3)
fo
A
1
(3)
f
A
1
(3)
fo
A
1
(3)USV
f
A
1
(3)USV
fo
data
x 0.66
x 0.66
62
Figure 5. At-the-money implied volatility as a function of the state variables.
We plot how at-the-money implied volatilities is aected by the state variables. In order to disentangle
the role of the risk premia (in particular, ane vs essentially ane) from the role of the state variables,
we evaluate the term structure assuming zero risk premia (denoted by P on the plot), assuming zero
essentially ane components of the risk premia (ane Q on the plot), and nally using the full risk
premia (ess Q).
0.040.060.08
0.9
1
1.1
1.2
r
a
b
s
o
l
u
t
e

i
m
p
l
i
e
d

v
o
l
a
t
i
l
i
t
y
(
%
)
A
0
(3)
f
0.05 0.1 0.15
1
1.1
1.2

a
b
s
o
l
u
t
e

i
m
p
l
i
e
d

v
o
l
a
t
i
l
i
t
y
(
%
)
0 0.05 0.1
1
1.1
1.2
s
a
b
s
o
l
u
t
e

i
m
p
l
i
e
d

v
o
l
a
t
i
l
i
t
y
(
%
)
0.04 0.06 0.08
0.85
0.9
0.95
r
A
0
(3)
fo
0.05 0.1 0.15
0.86
0.88
0.9
0.92
0.94

0.05 0 0.05 0.1

0.86
0.88
0.9
0.92
0.94
s
0.040.060.08
0.9
0.95
1
1.05
1.1
1.15
r
A
1
(3)
f
0.05 0.1 0.15
0.9
0.95
1
1.05
1.1
1.15

5 10
x 10
5
0.8
0.9
1
1.1
1.2
1.3
v
0.04 0.06 0.08
0.9
0.95
1
r
A
1
(3)
fo
0.05 0.1 0.15
0.9
0.95
1

0.5 1 1.5 2 2.5

x 10
5
0.6
0.8
1
1.2
1.4
v
0.04 0.06 0.08
3.1
3.15
3.2
3.25
3.3
r
A
1
(3)USV
f
0.05 0.1 0.15
3.1
3.15
3.2
3.25
3.3

0.5 1 1.5 2
x 10
3
2
3
4
5
v
0.04 0.06 0.08
1
1.05
1.1
r
A
1
(3)USV
fo
0.05 0.1 0.15
1
1.05
1.1

0.5 1 1.5 2
x 10
4
0.8
1
1.2
1.4
1.6
v
ess Q
affine Q
P
63