Do Bonds Span Volatility Risk in the

U.S. Treasury Market? A Specication Test for

Ane Term Structure Models
TORBEN G. ANDERSEN and LUCA BENZONI

Andersen is at the Kellogg School of Management, Northwestern University, NBER, and the Center for Research in Econo-
metric Analysis of Time Series (CREATES). Benzoni is at the Federal Reserve Bank of Chicago. We are grateful to Jeerson
Duarte, Darrell Due, Michael Fleming, Bob Goldstein, Mike Johannes, Chris Jones, Rick Nelson, Jun Pan, Rob Stambaugh
(the Editor), Sam Thompson, an anonymous referee, and seminar participants at Carnegie Mellon University, the Chicago Fed,
HEC Montreal, Indiana University, the University of Chicago, the Federal Reserve Board of Governors, the St. Louis Fed, the
University of Illinois at Chicago, the University of Wisconsin at Madison, as well as the Third T.N. Thiele Symposium on
Stochastic Volatility, the 2005 International Conference on Capital Markets, Corporate Finance, Money and Banking at the
Cass Business School, London, the 2006 Econometric Society Winter Meeting, the 2006 CIREQ-CIRANO-MITACS Financial
Econometrics Conference, the 2006 Bank of Canada Conference on Fixed Income Markets, the Multivariate Modeling and Risk
Management Conference, Sandbjerg, Denmark, the March 2006 NBER Asset Pricing Meeting, the 2007 AFA conference, and
the 2007 Chicago Conference on Volatility and High Frequency Data for helpful comments and suggestions. Further, we thank
Mitch Haviv of GovPX for providing useful information on their data. Andrea Ajello and Huiyan Qiu provided outstanding
research assistance. Of course, all errors remain our sole responsibility. Andersens work is supported by a grant from the
National Science Foundation to NBER and from CREATES funded by the Danish National Research Foundation. The views
expressed herein are those of the authors and not necessarily those of the Federal Reserve Bank of Chicago or the Federal
Reserve System.
Do Bonds Span Volatility Risk in the
U.S. Treasury Market? A Specication Test for
Ane Term Structure Models
ABSTRACT
We propose using model-free yield quadratic variation measures computed from intraday data as a tool
for specication testing and selection of dynamic term structure models. We nd that the yield curve
fails to span realized yield volatility in the U.S. Treasury market, as the systematic volatility factors are
largely unrelated to the cross-section of yields. We conclude that a broad class of ane diusive, quadratic
Gaussian, and ane jump-diusive models cannot accommodate the observed yield volatility dynamics.
Hence, the Treasury market per se is incomplete, as yield volatility risk cannot be hedged solely through
Treasury securities.
1
Virtually all variation in U.S. Treasury rates is captured by three factors, interpreted by Litterman
and Scheinkman (1991) as changes in level, steepness, and curvature. This fact has directly motivated
an extensive body of work on reduced-form term structure models in which bond yields are expressed as an
ane (or quadratic) function of a low-dimensional state vector; see, for example, Due and Kan (1996),
Due, Pan, and Singleton (2000), and Piazzesi (2003). These models are successful at capturing the cross-
sectional properties of bond yields as argued in Ahn, Dittmar, and Gallant (2002), Ahn et al. (2003), Brandt
and Chapman (2002), and Dai and Singleton (2000), among others. However, some of their implications
remain controversial. One major practical concern is how to hedge positions in Treasury securities. A
key feature of standard ane term structure models is that the quadratic variation of bond yields at any
maturity is a linear combination of the concurrent term structure of yields. Consequently, interest rate
volatility risk is spanned and can be hedged by trading solely in bonds. In this paper, we empirically
examine this prediction.
Previous studies investigate the issue using data on London Interbank Oered Rate (LIBOR) yields,
swap rates, and associated derivatives, and nd conicting evidence. Collin-Dufresne and Goldstein (2002)
conclude that swap rates have limited explanatory power for returns on at-the-money straddles, that is,
portfolios mainly exposed to volatility risk. Hence, they propose an ane term structure model in which
bond prices are unaected by changes in volatility, labeling this an unspanned stochastic volatility (USV)
restriction. Similarly, Li and Zhao (2006) nd that many prominent multi-factor dynamic term structure
models have serious diculties in hedging caps and cap straddles, even though they capture bond yields well.
In contrast, Fan, Gupta, and Ritchken (2003) nd that LIBOR bonds alone can hedge positions in swaptions
and even swaption straddles, supporting the notion that bond markets are complete. The latter results are
consistent with Litterman, Scheinkman, and Weiss (1991), who nd the yield spreads of volatility-sensitive
buttery combinations to be highly correlated with the curvature factor.
More recently, several studies examine term structure models that explicitly embed the USV restriction.
However, a direct comparison of their ndings is dicult due to dierences in model specication, estimation
method, data, and sample period. For instance, Collin-Dufresne, Goldstein, and Jones (2008, CDGJ) show
2
that the LIBOR volatility implied by an ane multi-factor specication from the swap rate curve can be
negatively correlated with the time series of volatility obtained from a standard GARCH approach. Jacobs
and Karoui (2007) nd similar results when using a short sample of swap rates, but with a longer data
set of U.S. Treasury yields they obtain a high correlation between model-implied and GARCH volatility
estimates. In contrast, Thompson (2004) rejects the Collin-Dufresne and Goldstein (2002) USV restriction
using swap rates data. Jagannathan, Kaplin, and Sun (2003) emphasize the importance of using derivatives
to evaluate term structure models. There are two dimensions to this extension. First, the models have strong
implications for yield volatility and these can best be tested via data on volatility-sensitive instruments.
Second, measurement errors may render the theoretical link between yield levels and volatility hard to
identify from observed bond prices alone, so derivatives prices may be necessary for ecient inference.
Accordingly, Bikbov and Chernov (2004), Han (2007), Jarrow, Li, and Zhao (2007), Joslin (2007), Trolle
and Schwartz (2007b), and Umantsev (2001) rely on derivatives prices and underlying interest rates to better
identify the volatility dynamics. In this case, the evidence on the volatility spanning condition is also mixed.
We argue that the preceding literature has not focused on the fundamental yield volatility implications
that characterize the ane model class. Two core predictions are 1) that the instantaneous yield volatility
is spanned by the contemporaneous cross-section of yields and 2) that no other economic variables can
improve on volatility forecasts extracted from the current yield curve. Within the diusive model class,
natural tests of these properties are, rst, to directly relate measures of realized quadratic variation to the
concurrent term structure of yields over short (say, daily, weekly, or monthly horizons) and, second, to explore
whether alternative predictor variables are able to improve signicantly on volatility forecasts implied by
the current yield curve. Of course, yield volatility is fundamentally unobserved, or latent, hampering direct
testing. However, an extensive recent literature documents theoretically and empirically that daily realized
volatility also can be measured with good precision from intraday price data; see, for example, Andersen and
Bollerslev (1998b), Andersen et al. (2001), and Barndor-Nielsen and Shephard (2002b, 2004a).
1
Following
these insights, we form model-free daily realized yield volatility series for Treasuries with a maturity of three
and six months, as well as one, two, ve, and 10 years, by cumulating squared intradaily yield changes.
3
We directly test whether bonds span volatility by relating our realized yield volatility measures to the
cross-section of daily bond yields. At each date, we compute average daily bond yields and then use
orthogonal principal components from those yields as explanatory variables in linear regressions, seeking to
span the realized yield volatility. In stark contrast to the notion that the yields quadratic variation is a linear
combination of the bond yields, the explanatory power of these regressions is, in most cases, nearly zero.
For instance, the R
2
for the realized volatilities with maturity in excess of two years is less than 0.6%. When
the dependent variable is the realized volatility of shorter maturity yields, the R
2
shows little improvement,
ranging from 1% to approximately 4%. Interestingly, the rst three principal components (i.e., level, slope,
and curvature) typically have insignicant coecients, while higher-order principal components may enter
signicantly. We conrm that these results continue to apply when the spanning condition is tested using
weekly and monthly realized volatility. Moreover, the volatility spanning condition is violated consistently
across subsamples. Finally, there is a substantial degree of predictable yield variation that is captured by
simple time-series models but is largely unrelated to the yield curve. This nding violates the second basic
property of the ane spanning condition, which should also apply within the general ane jump-diusion
setting.
There are several advantages to our approach. First, we test the generic ane yield volatility spanning
condition directly. Hence, the analysis is independent of specic modeling assumptions. In contrast, most
prior studies rely on a particular ane term structure specication, which results in a joint test of the
USV restriction and a certain interest rate model. Second, the basic spanning restriction follows from the
ane structure under the equivalent martingale or pricing measure, so it is invariant to whether the
representation under the actual measure is non-ane (Duarte (2004)). Third, we alleviate the impact
of measurement errors on the extracted daily zero-coupon bond yield series by exploiting intraday data.
Fourth, we avoid stipulating a specic time-series model for the conditional yield variance process. This
approach has been previously adopted to gauge the coherence between the volatility dynamics implied by
the model and the data; see, for example, CDGJ, Dai and Singleton (2003), and Jacobs and Karoui (2007).
Nonetheless, the realized volatility series allow us to construct simple, yet eective, reduced-form time-
4
series forecasts of future quadratic yield variation. In fact, such forecasts typically outperform predictions
generated by standard volatility models based on daily or lower frequency data (e.g., Andersen et al. (2003a)).
We exploit this type of reduced-form model in testing whether implied ane volatility predictions from the
current yield curve may be improved signicantly by alternate predictors. Fifth, we obtain realized yield
variation measures for multiple maturities, enabling us to study the volatility dynamics across the term
structure. This approach enhances the power of our tests as the spanning condition should hold for each
individual maturity. Sixth, our tests allow for as many underlying ane factors as there are reliably observed
yields, J. In practice, most successful and empirically tractable ane term structure models rely only on N
factors, with N being three or four, so we have N < J. Moreover, usually only a few factors, k < N, govern
the conditional yield variance. Dai and Singleton (2000) refer to the maximal version of this model as
the A
k
(N) specication. Our tests of the ane spanning property directly cover the full range of dierent
dimensional representations of these models. Seventh, we are free to test the spanning restriction at any
horizon, say, daily, weekly, or monthly, or, theoretically, even at an intraday level. Eighth, we are able to
expand our specication analysis beyond the traditional ane model class to encompass the quadratic term
structure model studied by, for example, Ahn, Dittmar, and Gallant (2002), Beaglehole and Tenney (1991,
1992), Constantinides (1992), and Longsta (1989).
The presence of jumps changes the model restrictions qualitatively as the strict spanning condition now
fails. However, the conditionally expected future quadratic variation is spanned, as before, by the yield
cross-section for the general ane jump-diusive model class. Hence, our second set of specication tests,
which exploit past realized volatility measures as additional forecast variables, remains valid. This approach
also oers robustness against the impact of standard measurement errors in the realized volatility measures:
even if spanning fails due to noise in the volatility proxies, no alternative predictor variables should be able
to improve on the forecasts provided by the yield curve. Hence, our test methodology accommodates both
jumps and measurement errors in the yield volatility measures. Moreover, in another robustness check, we
show that our ndings are largely unaected by measurement errors in the zero-coupon yields themselves
(rather than the yield volatilities), since the basic predictive regression coecients are close to invariant to
5
the use of suitable instruments for the yields.
The power of our tests and the interpretation of our ndings are related to the quality of our realized yield
volatility measures, so we undertake a variety of robustness checks to assess the reliability of these volatility
proxies. Most signicantly, we t a well-calibrated EGARCH-type semi-nonparametric (SNP) model to the
daily three-month maturity yield series. This model provides one-day-ahead volatility forecasts, which we
contrast to the corresponding realized volatility series. We conrm that the properties of these series are
consistent with the typical relationships between volatility forecasts and subsequent realizations. Moreover,
the general properties of the realized yield volatility series, both in terms of dynamic dependencies and
unconditional term structure features, are consistent with the extant literature that employs volatility
measures based on lower frequency returns. Thus, even if the yield volatility estimates inevitably contain a
noise component, the associated errors do not induce discernable systematic biases.
In conclusion, our results indicate that interest rate volatility cannot be extracted from the cross-section
of bond yields in the U.S. Treasury market. This nding underscores the importance of relaxing the volatility
spanning conditionwithin or outside the ane settingin term structure models,
2
especially for appli-
cations that require a good t to the yield volatility dynamics such as hedging interest rate volatility risk
or pricing xed-income derivatives. More generally, the pricing of yield risk across the business cycle is of
great interest in monetary economics. In ane term structure models the yield risk premium is proportional
to the yield factor volatility. Systematic errors in model-implied factor volatility are thus likely to distort
measurement and inference regarding the pricing of yield factor risk.
The remainder of the paper is organized as follows. Section I discusses the link between Treasury yields
and their quadratic variation in the context of ane diusive term structure models. We clarify how this
relation is aected by the presence of jumps and we introduce our realized volatility measures of the yields
quadratic variation. In Section II we describe the U.S. Treasury market data and document the salient
features of our volatility series and forecasts relative to the existing literature. Section III presents our main
empirical ndings. Concluding remarks are provided in Section IV.
6
I. Ane Term Structure Models
This section discusses the empirical implications of the continuous-time ane model class for the yield
volatility of zero-coupon bonds. These models provide testable restrictions that apply not only to standard
ane multi-factor diusions, but also to the recently popular quadratic-Gaussian models. Moreover, these
restrictions arise directly from the ane specication of the diusion coecient, which is invariant across the
equivalent martingale (risk-neutral) and the physical (actual) probability measures. Hence, they remain valid
for the essentially ane class proposed by Duee (2002) and also cover models that allow for non-ane
drift under the physical measure, as proposed by Duarte (2004) and further analyzed in Cheridito, Filipovic,
and Kimmel (2007). There are also interesting predictions for the ane jump-diusion representations of
the term structure that we distinguish from the pure diusion case. These linkages between yield levels and
concurrent as well as future yield variation form the basis for our specication analysis of the entire model
class through spanning conditions involving nonparametric realized volatility measures.
A. Bond Yields and Yield Volatility in Ane Diusion Models
We rst review some basic features of ane diusive term structure models. Following Due and Kan
(1996) and Dai and Singleton (2000), the short-term interest rate, y
0
(t), is an ane (linear-plus-constant)
function of a vector of state variables, X(t) = {x
i
(t), i = 1, . . . , N }:
y
0
(t) =
0
+
N

i=1

i
x
i
(t) =
0
+

X
X(t) , (1)
where the state vector X has risk-neutral dynamics
dX(t) = K(X(t))dt +

S(t)dW
Q
(t) . (2)
In equation (2), W
Q
is an N-dimensional Brownian motion under the pricing (Q) measure, K and are
N N matrices, is an N 1 vector, and S(t) is an N N diagonal matrix with the ith diagonal element
given by [S(t)]
ii
=
i
+

i
X(t), where
i
is a scalar and
i
is an N 1 vector.
Within this setting, one can nd (eectively) closed-form expressions for the time-t price of a zero-coupon
7
bond with time-to-maturity (measured in years):
P(t, ) = e
A()B()

X(t)
, (3)
where the scalar function A() and the N 1 vector of functions B() = {B
k
(), k = 1, . . . , N } solve a
system of ordinary dierential equations.
This result establishes a fundamental link between the state vector X(t) and the term structure of bond
yields. Specically, the time-t yield y

(t), expressed in percent per annum, on a zero-coupon bond with

time-to-maturity is dened by the relationship P(t, ) = e
y

(t)
. Thus, in view of equation (3), we have,
y

(t) =
A()

+
B()

X(t) . (4)
The J 1 vector of (observed) zero-coupon bond yields is denoted Y (t) = {y

j
(t), j = 1, . . . , J }, where we
assume that we observe more yields than there are state variables, that is, J N. Dening the J 1 vector
A =

A(
j
)

j
, j = 1, . . . , J

and the N J matrix B =

B(
j
)

j
, j = 1, . . . , J

, we write the above relation

as a system of equations:
Y (t) = A+B

X(t) . (5)
Typically, the B matrix is assumed to be of full row rank, N, which implies that all state variables aect the
contemporaneous cross-section of bond prices and yields. In this case we may invert the system to express
the state vector as an ane function of the yields. We therefore have, for C an N 1 vector,
X(t) = (BB

)
1
B(Y (t) + A) = C + (BB

)
1
BY (t) . (6)
Meanwhile, by Itos Lemma, the yield, y

, in equation (4) follows a diusion process:

dy

(t) =
y
(X(t), t) dt +
B()

S(t)dW
Q
(t) . (7)
Consequently, the (instantaneous) variation of the yield is
V
y

(t) =
B()

S(t)

B()

. (8)
Since by denition the elements of the S(t) matrix are ane in the state vector X(t) (as stated below
equation (2)) and in turn X(t) is an ane function of Y (t) from equation (6), it follows that, for any , we
8
can nd a set of constants a
, j
, j = 0, . . . , J, such that
V
y
(t) = a
,0
+
J

j=1
a
, j
y

j
(t) . (9)
Hence, the instantaneous variation of (constant maturity) yields is tied to the contemporaneous yield level
and thus to the cross-section of bond prices through the ane mapping in equation (9). Since the yield
variation is also directly related to the time series properties of the yields, it plays a dual role in standard
ane diusive term structure models. CDGJ highlight the implied link between the bond yields and the
short rate variation. The above shows that this relationship is valid for any xed maturity yield, implying
a range of joint restrictions across the yield volatility spectrum.
3
The volatility spanning condition (9) does not hold only for Ane Term Structure Models (ATSM) but
also for the so-called Quadratic Term Structure Models (QTSM) considered in Ahn, Dittmar, and Gallant
(2002), Beaglehole and Tenney (1991, 1992), Constantinides (1992), and Longsta (1989). In fact, the
QTSM is isomorphic to the ATSM in its mechanism for generating volatility, as the yield variation remains
proportional to the level of the state variables. Ahn, Dittmar, and Gallant (2002) and Cheng and Scaillet
(2007) formally show how the quadratic models may be embedded in an ane model with an extended state
vector. Hence, as long as we allow for a suciently large-dimensional state vector, our analysis automatically
covers the quadratic models as well. This is of particular interest as one of the primary motivations behind
these models is that they should better replicate the yield volatility than the regular ane models.
4
We now derive the implications of equation (9) for discrete-time data to enable testing on the basis of
bond prices observed at a reasonably high frequency. First, we recall the denition of the quadratic variation
process for the constant-maturity yield y

initiated at time t
0
= 0:
QV
y

(t)

t
0
V
y

(s) ds . (10)
It is convenient to cast the model implications in terms of increments in quadratic variation and yields over
daily or intraday periods [ t, t +h], h > 0. Denoting these quantities QV
y

(t +h, h) QV
y

(t +h)QV
y

(t)
and y

j
(t +h, h)
1
h

t+h
t
y

j
(s) ds, it follows from equation (9) that
QV
y

(t +h, h) = a
,0
+
J

j=1
a
, j
y

j
(t +h, h) . (11)
9
To simplify notation, in equation (11) we have redened the coecients a
, j
, j = 0, 1, ..., J, to be the
coecients a
, j
, j = 0, 1, ..., J, in equation (9) multiplied by a scaling factor h. We term this relation the
contemporaneous ane yield variation spanning condition. This is a strict realization-by-realization identity
in which shifts in the term structure match shocks to the yield volatility over any arbitrary period [t, t +h].
The yield levels on the right-hand side of equation (11) are readily approximated using empirical ob-
servations on the intraday yieldsor more crudely, the yields at the close of tradingacross the maturity
spectrum. The quadratic variation increment on the left-hand side is slightly more delicate, as it cannot be
measured with precision from daily bond price data. Perhaps as a consequence, the quadratic yield varia-
tion has not been the focus of direct measurement or testing within the term structure literature. Instead,
most existing studies rely on parametric conditional yield variance estimates or implied volatility measures
backed out from derivatives prices. However, while this approach can be rigorously justied, as we illustrate
below, the substitution of an alternative volatility proxy in place of the realized quadratic variation is not
innocuousit inevitably entails a loss of power in tests of the ane spanning condition.
B. Spanning Restrictions for the Conditional Yield Variance
The pure diusive semi-martingale representation of bond prices implies that the predictable yield changes
over short (daily or intraday) periods are negligible (order dt
2
) in terms of their contribution to the quadratic
yield variation compared to the yield innovations (order dW
Q
(t)
2
= dt). Hence, in practice, we may ignore
the conditional mean of the yield changes in computing the conditional yield variance over short horizons.
Letting the integer n 1 denote the number of equidistant yield changes sampled over the (short) interval
[ t, t +h], we thus invoke the following approximate martingale relation:
E
P
t

t +
i h
n

= y

(t) , i = 1, . . . , n, (12)
where the subscript t indicates that the expectation is evaluated conditional on time-t information, and the
superscript P indicates the so-called actual or physical probability measure.
10
It follows that the conditional yield variance is given by
V ar
P
t
[ y

(t +h) ] = E
P
t

i=1,...,n

t +
i h
n

t +
(i 1)h
n

. (13)
Equation (13) holds for an arbitrary n, so by letting n increase towards innity we have, by basic properties
of the quadratic variation process, that
V ar
P
t
[ y

(t +h) ] = E
P
t
[ QV
y

(t +h, h) ] . (14)
This relation highlights important distinctions between expected and realized yield volatility. The condi-
tional variance is a forward-looking expectation of the future sample-path variation, and thus fundamentally
an ex-ante concept. In contrast, the quadratic variation denotes the actual realized sample-path variation,
so it is an ex-post (realization) measure. Only if volatility is (conditionally) deterministic, as when it is
constant, do the two notions of yield variation coincide. In general, however, the yield variation has a large
and unpredictable innovation component that renders volatility stochastic. As such, the sample variability
of the quadratic yield variation process inevitably exceeds that of the conditional yield variance process
because sample realizations, by construction, uctuate more than their a priori expectations. The spanning
condition in equation (11) requires the yield levels to uctuate in concert with the realized yield volatility.
This is a much more stringent requirement than the corresponding prediction based on the (lower degree of)
variation in the expected conditional yields. In order to formally derive the latter implication of the ane
term structure models, we rst substitute equation (11) into equation (14) to obtain
V ar
P
t
[ y

(t +h) ] = a
,0
+
J

j=1
a
, j
E
P
t

j
(t +h, h)

. (15)
This prediction is valid only under the P measure, as it is related directly to the observed time-series
variation of the yields. The conditional moments over discrete (non-innitesimal) horizons will dier across
P and Q due to discrepancies in the volatility drift specication, even if the instantaneous volatility (and
quadratic variation) is identical under the two measures. Assuming that the diusion model is also ane
under the physical measure, which still allows for the essentially ane model of Duee (2002) but excludes
the extension by Duarte (2004), the future expected yields will be a linear combination of the current cross-
11
section of yields, so that
V ar
P
t
[ y

(t +h) ] = b
,0
+
J

j=1
b
, j
y

j
(t) . (16)
This is a formal representation of the type of spanning condition investigated in prior empirical studies.
We label it the predictive ane yield variation spanning condition. We stress again that it is much less
powerful in terms of testing the underlying model than equation (11) and it requires the model to be ane
under both the P and Q measures. A nal caveat is that it is valid only if the true conditional variance
process appears on the left-hand side of (16). Using an ad hoc time-series model to generate these volatility
forecasts will inevitably induce a degree of measurement error into any testing procedure.
A similar logic applies if we use implied volatility forecasts derived from derivatives prices in lieu of the
forecasts based on the time-series model, but the forecasts are now formed under the pricing measure. This
approach assumes that the derivatives pricing model is correctly specied and quality data on derivatives
are available. In that case, the implied conditional variance forecasts (under Q) should also be spanned by
the cross-section of yields. As for the fundamental spanning condition (11), this reasoning hinges only on the
model being ane under Q, so it applies also for Duarte-style extensions of the ane model. On the other
hand, the forecast horizon must equal the maturity of the derivatives contracts, which typically necessitates
monthly volatility predictions rather than daily or weekly forecasts and thus reduces the forecast comparison
sample and in turn lowers the tests power.
We may also explicitly relate daily changes in the conditional yield variance to the evolution of the yield
cross-section. Letting, generically, c(t) = c(t) c(t h), we have from equation (16) that
V ar
P
t
[ y

(t +h) ] =
J

j=1
b
, j
y

j
(t) . (17)
Of course, we can derive an equivalent expression for changes in the implied volatility forecasts under
the Q measure. Several studies employ such a specication of the ane spanning condition. This approach
is obviously closely related to the specication in equation (16), so we focus on the latter in the rest of the
paper. However, we conrm that the ndings are qualitatively similar, albeit even less attering, for the
basic ane model restrictions when we use the representation in (17).
12
C. A Framework for Testing the Fundamental Ane Spanning Condition
The previous sections outline the basic volatility spanning conditions implied by standard ane diusive
term structure models. In this section we develop a unied framework for testing these spanning conditions.
We rst focus on the ideal scenario in which the quadratic yield variation and the contemporaneous mean
yield levels are measured perfectly. We then extend the analysis to account for the presence of measurement
error.
C.1. Regression-based Tests of the Spanning Condition
The fundamental spanning condition in equation (11) directly motivates the regression
QV
y
(t +h, h) = a
,0
+
J

j=1
a
, j
y

j
(t +h, h) +(t +h, h) . (18)
If all variables are measured perfectly, the error term is identically zero and the associated R
2
should be
unity.
In contrast, the preceding literature invokes largely diagnostic procedures based on the predictive relation
in equations (16) and (17). A popular approach is to gauge the correlation between model-implied volatility
forecasts and alternative benchmark forecasts based on time-series models such as GARCH or implied
volatility measures extracted from derivatives prices. Denoting one such alternative volatility forecast for
the time period [t, t + h] by F(t + h, h), one may directly study the coherence between the alternative
forecasts through the regression
F(t +h, h) = c
,0
+
J

j=1
c
, j
y

j
(t) +
f
(t, h) . (19)
If the standard ane diusive model is valid, one would expect both a reasonable degree of correlation
between the alternative forecasts and a sizeable R
2
from the regression (19). Nonetheless, even in the
absence of measurement error, misspecication and estimation errors associated with the auxiliary forecast
procedure would induce noise in the volatility benchmark. Thus, there is no sharp prediction concerning
the actual degree of correlation but rather an expectation of qualitative coherence.
13
In view of the informal character of the diagnostic regression in equation (19), we seek an alternative
formulation that leads to a sharper test. In the traditional ane setting, under the null hypothesis the
conditional variance is spanned by the yield cross-section, so any auxiliary forecast variable should not
have predictive power for the future yield variation that is over-and-above the information conveyed by the
contemporaneous yields. Taking advantage of equations (14) and (16) we can test this property via the
regression
QV
y

(t +h, h) = d
,0
+
J

j=1
d
, j
y

j
(t) +d
F
F(t +h, h) +
af
(t +h, h) . (20)
The coecient d
F
should be insignicant and the associated R
2
should not be signicantly higher than
when the F(t +h, h) variable is excluded from the regression. However, in contrast to regression (18), there
is no presumption that the explanatory power should be close to perfect as this is a forecast regression
for the future yield variation, which is likely to possess a genuinely stochastic (unpredictable) component.
Equation (18), on the other hand, states that changes in the future yield variation must go hand-in-hand
with corresponding shifts in the future yield curve, thus producing a (near) perfect association.
C.2. Measurement Error
So far, we have focused on the predictions of ane term structure models in a frictionless market.
In practice, the extraction of xed-maturity zero-coupon yields from observed bond prices involves both
smoothing and extrapolation. Moreover, microstructure eects (e.g., bid-ask bounce) introduce noise in
observed bond prices. Therefore, there is inevitably some degree of error in the empirical proxies for the
variables on the left- and right-hand sides of equations (18) and (20). The question, then, is whether this is
likely to bias our tests and complicate interpretation of the empirical results. Here we extend the framework
for testing the spanning condition to accommodate some of these non-systematic eects, and we point to
the relevant empirical sections for additional robustness checks. We discuss the implications of errors in the
yields and in the yield volatilities separately, starting with the dependent variable, the realized quadratic
yield variation.
Any empirical proxy for the quadratic yield variation is subject to discretization error, as the continuous
14
sample-path variation is unobserved and must be approximated from measures constructed from discretely
sampled high frequency intraday (squared) yield changes. However, these errors are uncorrelated across
days. Thus, they simply insert noise into the contemporaneous spanning condition (18). This noise does
not aect the consistency of the OLS estimates but it reduces the R
2
of the regression so that it falls below
unityby some undetermined amounteven if volatility is truly spanned by the yields.
5
As a result, the
R
2
measure becomes unsuitable as a formal test statistic. Nevertheless, we still expect a robust coherence
between the yields and the yield volatility measure as long as a sensible volatility proxy is employed. In this
sense the explanatory power of the regression provides an informal diagnostic for model adequacy within
the ane diusive setting. Moreover, since the deterioration in explanatory power is directly related to the
amount of noise in the dependent variable, it is important to employ an ecient realized yield volatility
estimator in order to obtain a more powerful diagnostic. We study the empirical properties of our yield
volatility measures carefully in Section II.C.
The presence of (substantial) noise in the explanatory yield variables of regressions (18) and (20) will
create an error-in-variables problem and could result in inconsistent OLS estimates. This type of error
is mitigated as we reduce the impact of noisy measurements dramatically by obtaining the average daily
yield from a large number of intraday yield observations. Nevertheless, we also explore the traditional
remedy of estimating the system via instrumental variables. Following Stambaugh (1988), we construct
instruments using linear combinations of lagged yields. In Section III.A.2 we show that estimation by
instrumental variable techniques has little impact on the regression coecients while the explanatory power
of the regressions is virtually unchanged. As such, the traditional errors-in-variable problem does not appear
signicant in this context.
These observations also have direct implications for our test based on the predictive spanning regression
(20). If the errors-in-variable problem for the yield regressors is minor and the measurement errors for
yield volatility are serially uncorrelated, then it is still the case that no other variable should improve on
volatility forecasts extracted from the current yield curve. Hence, our predictive test is robust to these types
of measurement errors. We present the empirical results obtained from this approach in Section III.B.
15
D. Ane Jump-diusion Term Structure Models
It is common in the term structure literature to cast the analysis of interest rates in a diusive setting. To
our knowledge, existing studies of the volatility spanning condition rule out the presence of jumps. However,
there is compelling evidence that, for example, macroeconomic announcements induce instantaneous jumps
in the yields upon release (Andersen et al. (2007), Balduzzi, Elton, and Green (2001), Bollerslev, Cai,
and Song (2000), Fleming and Remolona (1999), Johannes (2004), and Piazzesi (2005)). Moreover, jumps
improve the t of the spot rate considerably (Andersen, Benzoni, and Lund (2004), Das (2002), and Johannes
(2004)). Here we show how the presence of jumps in state variables and yields changes the volatility spanning
property and we discuss how to test the relevant spanning conditions in this extended setting.
D.1. Bond Yields and Yield Volatility in Ane Jump-diusion Models
Following Due, Pan, and Singleton (2000), the state vector X in ane jump-diusion models has
Q-dynamics
dX(t) = K(X(t))dt +

S(t)dW
Q
(t) + Z dq
Q
(t) . (21)
Conditional on the path of X, q
Q
is a Poisson jump-arrival process that is independent of W
Q
and has
intensity (X) =
0
+

X
X. The jump size Z is an N 1 vector process. At a jump time t, the jump size Z
is independent of {X(s) : 0 s < t} and has a xed probability distribution
Q
, with E
Q
[Z(t)Z(t)

] = V
ZZ
.
The corresponding jump vector process, J(t) X(t) = Z(t)dq
Q
(t), is non-zero only if a jump actually
occurs. Under these assumptions, instead of equation (7) we have
dy

(t) =
y
(X(t), t) dt +
B()

S(t) dW
Q
(t) +Z dq
Q
(t)

, (22)
which shows that the quadratic variation for an ane jump-diusion model is the sum of a diusive and a
jump component:
QV
y
(t, h) =

t+h
t
B(T s)

T s
S(s)

B(T s)
T s
ds +

thst

B(T s)

T s
J(s)J(s)

B(T s)
T s

. (23)
The rst term in equation (23) is identical to the quadratic variation for the ane diusion model without
jumps, given by equations (8) and (10). Consequently, the diusive quadratic variation component is spanned
16
by the bond yields. In contrast, the jump component cannot be reduced to a linear combination of bond
yields. Thus, for an ane jump-diusion model the strict realization-by-realization spanning constraint in
equation (11) no longer applies.
D.2. A Framework for Testing the Spanning Condition with Jumps
When there are jumps, a version of the regression model (18) still holds, but with the error term
embedding the (compensated) jump component:
QV
y

(t +h, h) = a
,0
+
J

j=1
a
, j
y

j
(t +h, h) +(t +h, h) +(t +h, h)
(t +h, h) =

thst

B(T s)

T s
J(s)J(s)

B(T s)
T s

E
Q
t

tst+h
B(T s)

T s
J(s)J(s)

B(T s)
T s

.
(24)
In this case, we can still examine the volatility spanning condition by estimating a contemporaneous yield-
volatility regression. However, the R
2
associated with regression (24) would be below unity, even in the ideal
scenario in which the quadratic yield variation and the contemporaneous mean yield levels are measured
perfectly so that (t + h, h) is identically zero. In other words, under the null hypothesis of an ane
jump-diusive model, a test of the spanning condition based on this approach lacks power.
Nevertheless, as the label ane jump-diusion indicates, it is still the case that the state variables
span the rst two yield moments. Indeed, upon taking conditional expectations we nd
E
Q
t

tst+h
J(s)J(s)

= E
Q
t

t+h
t
(
0
+

X
X(s)) ds

V
ZZ
. (25)
Hence, the expected jump contribution to the quadratic variation is an ane function of the state variables.
Since the state variables may be written as a linear combination of the yields under the Q measure, this
property carries over to the full expected quadratic yield variation process. Thus, following the reasoning
in Section I.B, we obtain the following spanning condition for the (short) future time interval [t, t +h]:
E
Q
t
[ QV
y

(t +h, h) ] = b
,0
+
J

j=1
b
, j
y

j
(t) . (26)
17
This spanning condition is identical to equation (16) as the latter also holds under the Q measure, and it
is straightforward to derive the corresponding version of equation (17) as well. Consequently, the extension
that includes jumps has no impact on the predictive spanning condition under the Q measure. The identical
spanning restriction for conditional yield variance forecasts under the actual probability measure, P, applies
in the ane jump-diusion setting as well, if the expected jump distribution and jump intensity continue
to be ane functions of the state variables under P, as described in Section I.B. In terms of the regression
framework developed in Section I.C, equation (20) remains valid and provides a viable testing procedure
without the need for additional assumptions.
6
Moreover, as discussed previously, this approach also readily
accommodates standard measurement errors in the yield volatility measures.
E. Realized Yield Volatility Measurement
The recent volatility measurement and forecasting literature has advocated using realized volatility as
a way to approximate the daily realizations from the return quadratic variation process. Early applications
of this idea appear in Andersen and Bollerslev (1997b, 1998b), while the associated (continuous record)
asymptotic theory is provided in Barndor-Nielsen and Shephard (2002a, 2002b, 2004a), Andersen et al.
(2003a), and Andersen, Bollerslev, and Diebold (2004). The approach is fully nonparametric, and hence
model-free, and utilizes the cumulative squared high frequency intraday returns to obtain feasible quadratic
variation measures. Specically, we compute the annualized realized volatility of the yield y

over [ t, t +h],
given a sampling frequency
h
n
, by
v
2
y

(t +h, h) =
1
h
n

i=1

t +
i h
n

t +
(i 1)h
n

2
. (27)
The realized yield volatility converges, for ever more frequent sampling, to the actual realization of the
quadratic yield variation. Moreover, the errors are asymptotically uncorrelated for non-overlapping time
intervals. Equation (11) links the quadratic yield variation for maturity zero-coupon bonds to the cross-
section of yields. We use the realized yield volatility measure in (27) to approximate the annualized form of
the quadratic yield variation QV
y
(t +h, h) in equation (11).
Our analysis focuses on the volatility of bond yields. There are only a few realized volatility studies
18
of U.S. Treasury securities (for instance, Andersen et al. (2007)), and these invariably rely on the realized
volatility of bond returns, so it is useful to clarify the link between equation (27) and the realized bond
return volatility. Letting p(t, ) log(P(t, )) denote the time-t zero-coupon log bond price, the continuously
compounded return on a zero-coupon bond with time-to-maturity = T t h over [ t, t +h] is
r(t +h, h, ) = p(t +h, ) p(t, ), 0 t t +h T . (28)
Since P(t, ) = e
y (t)
, this implies the following expression for the intraday return during trading day t,
where, by convention, is constant:
r

t +
i h
n
,
h
n

t +
i h
n

t +
(i 1)h
n

. (29)
Thus, the sum of the squared intraday changes in yields is proportional to the sum of the intraday squared
returns, with a constant of proportionality of
2
, that is, the square of the time-to-maturity. It follows that
the annualized realized volatility of the return on a bond with maturity during [ t, t +h] is
v
2
r

(t +h, h) =
1
h
n

i=1

t +
i h
n

t +
(i 1)h
n

2
. (30)
Hence, the realized yield volatility and return volatility series are proportional for a given maturity zero-
coupon bond. Nonetheless, equation (30) shows that the realized volatility incorporates a scaling factor,

2
, directly linked to time-to-maturity. As such, by focusing on yield volatility we facilitate comparison of
results across the term structure.
II. Measuring Realized Treasury Yield Volatility
A. Intraday Yield Data
We rely on the GovPX database to construct intraday series of bond yields. GovPX consolidates and
posts real-time quote and trade data from most of the major interdealer Treasury securities brokers with a
notable exception being Cantor Fitzgerald Inc. Taken together, these brokers account for about two-thirds
of the interdealer broker market, a fraction that declined to 42% in the rst quarter of 2000. In turn, the
19
interdealer market is approximately one-half of the total market; see Fleming (1997, 2003). Hence, while
the estimated bills coverage exceeds 90% in every year of the GovPX sample, the availability of 30-year
bond data is limited because of the prominence of Cantor Fitzgerald at the long-maturity segment of the
market. Therefore, we use only data on the three-month, six-month, and one-year bills, as well as the two-,
ve-, and 10-year notes in our analysis. We rely exclusively on quotes for on-the-run contracts, which are
signicantly more liquid than o-the-run Treasuries.
7
Our sample starts at the inception date of GovPX, June 17, 1991, and ends on June 15, 2001. More
recent data are available, but we do not use them for several reasons. First, the one-year Treasury bill was
no longer auctioned beginning March 2001. Second, after the end of our sample period the GovPX coverage
of the U.S. Treasury market started to decline (Fleming (2003)). Third, the period following the September
11, 2001, terrorist attacks was tumultuous for bond markets (Fleming and Garbade (2002)).
The U.S. Treasury market is most active during business days from early morning through the late
afternoon. Thus, we start the intraday transaction record at 7:30am Eastern Time (ET) and we end it at
5:00pm ET. This window includes the regular macroeconomic and monetary policy announcements,
8
which
are among the most important determinants of yield changes; see, for example, Andersen et al. (2003b),
Balduzzi, Elton, and Green (2001), Fleming and Remolona (1997, 1999), Green (2004), and Li and Engle
(1998). Moreover, since the vast majority of trading takes place during these hours, we capture the associated
price discovery process, often seen to reect the aggregation of heterogeneous private information as well
as heterogeneous interpretation of public information; see, for example, Brandt and Kavajecz (2004) and
Pasquariello and Vega (2007).
The GovPX quote frequency for specic maturities is signicantly lower than for, say, stocks in the
Dow Jones 30 index. Hence, the recent literature on selecting an optimal intraday sampling frequency for
computing the quadratic variance process in the presence of market microstructure noise (for example, At-
Sahalia, Mykland, and Zhang (2005a, 2005b), Barndor-Nielsen et al. (2004), Hansen and Lunde (2006),
and Bandi and Russell (2006, 2008)) is not readily applicable. Instead, we follow the earlier literature in
using a fairly sparse and xed sampling frequency. A sensible compromise between adding information
20
regarding the strength of the yield movements and incorporating high frequency microstructure noise is
attained around the 10-minute sampling interval where the induced serial correlation in the yield changes
is minor. Thus, we use the immediately preceding on-the-run quote to construct the relevant bid and ask
prices at the end of each 10-minute interval and then dene the log-price, log(P(t)), as the mid-point of
the logarithmic bid and ask. We convert bond prices into zero-coupon yields according to the procedure in
Appendix A. Finally, we compute the intraday yield changes for each Treasury security in our sample.
There are a few days with very subdued trading activity. We discard days without any trading activity
for a period exceeding three hours.
9
This produces a series of 56 intraday 10-minute yield changes over 2,322
business days, for a total of 130,032 observations for each of the six Treasuries in our sample, as explained
in Appendix B. We compute the average daily trading period yield from the 57 intraday observations so
that the impact from any weakly dependent noise process should be negligible.
Market microstructure frictions such as price discreteness and bid-ask spread positioning due to dealer
inventory control may induce negative autocorrelation in the recorded series. We mitigate the impact of such
institutionally driven short-term price bouncing by applying an MA(1) lter to the yield change series.
10
B. Daily Constant-maturity Yield Data
As mentioned above, the GovPX coverage of the 30-year bond is limited and thus we exclude this security
from our analysis. However, data on the 30-year bond at the daily frequency are available from other sources
during our sample period. Although such information is not useful for computing accurate realized volatility
series, it can still be used to construct a proxy for the longer-term zero-coupon yield and serve as a regressor
in the volatility spanning condition (11). Consequently, such auxiliary daily yield series may be used to
provide an additional robustness check for our results based on the GovPX quotes.
We therefore consider a panel of daily yields from a constant-maturity series released by the Federal
Reserve Board of Governors. We focus on maturities of three and six months, and at one, two, three, ve,
seven, 10, and 30 years.
11
These series contain theoretical yields for coupon bonds sold at par. Hence, prior
to analysis we convert them into zero-coupon yields, as discussed in Appendix B.
21
C. Yield Volatility Measures
The quality of the empirical analysis of the spanning conditions hinges on the reliability of the realized
volatility measures. However, assessing their accuracy is not trivial because the true (realized) volatility
is latent. Thus, before embarking on an extensive analysis of the spanning restrictions, we document the
salient features of our volatility estimates in some detail. One objective is to assess their general properties
relative to the extant literature. Secondly, we seek to clarify the dierence between our ex-post realized
yield volatility measures and the more common (ex-ante) volatility forecasts. We pay particular attention
to the degree of predictability of yield volatility as this is of primary interest for the subsequent empirical
evaluation of the ane model class.
C.1. Realized Yield Volatility
We construct the realized yield volatility series in equation (27) from intraday quote data on the three-
month, six-month, and one-year bill, as well as on the two-, ve-, and 10-year note. These realized volatility
estimates constitute measures of the zero-coupon yield quadratic variation during business hours (7:30am
to 5pm ET) alone. In order to relate them to comparable term-structure studies based on daily or lower
frequency data they are expressed in units reecting a yearly percentage. However, we must also convert
them from trading-day (business hours) to actual calendar-time yield volatility measures.
The main issue is how to account for the yield volatility outside of business hours. For each maturity
series, we compute the interdaily yield change from the closing at 5pm on each trading day to the opening
at 7:30am on the next trading day. We then measure total daily realized volatility by adding the squared
interdaily yield change to the sum of squared intraday yield changes. This total daily realized volatility
measure is then used to compute the sample mean of the annual realized yield variation. Finally, we rescale
each intraday-based realized volatility measure proportionally so that, across the entire sample, the daily
annualized realized volatility measure based only on intraday yield changes matches that estimated from
the total realized volatility series. We emphasize that any arbitrariness in the choice of this scaling constant
is inconsequential for our empirical analysis of the fundamental spanning restriction.
22
The empirical volatility literature usually adopts a daily volatility scaling. Thus, to facilitate comparison,
we present results in this format in the remainder of this section. Towards this end, we dene the daily
scaled realized volatility measure as follows:
v
2
d,y

(t +h, h) = h
D
v
2
y

(t +h, h) , (31)
where h
D
=
1
252
and v
2
y

(t +h, h) is given in equation (27).

Figure 1 depicts the square root of the rescaled daily realized volatility series, v
d,y

, for the zero-coupon

yield series we analyze. Casual inspection shows a great deal of covariation in the yield volatilities across
the maturity spectrum. However, there are some striking dierences, particularly among the more extreme
outliers. Although these are manifest in all the series, they often dier distinctly across the maturities. For
example, there is a pronounced spike in the realized volatility of short-maturity yields on October 8, 1998,
which is much attenuated for the longer yields. On that date, investors appear to have reacted negatively to
concerns related to a slowdown in the world economy, a weakening dollar (it lost 12% against the Japanese
Yen in less than two days), and the initiation of an impeachment inquiry against President Bill Clinton.
Investors were increasingly attracted to short-term bonds, pushing their prices up in a volatile trading
session.
12
In contrast, the longer-maturity yields volatility spiked up on June 2, 1995, when a dramatic
drop in the payroll employment number seems to have raised fears of a recession, sparking a powerful rally
that sent prices of longer-term bonds to record highs.
13
Similarly, on March 8, 1996, the employment report
revealed that over 700,000 new jobs were added to the payroll, lowering the unemployment rate from 5.8%
to 5.5%. This event was contrary to the general perception that the economy was bordering on recession
and likely reversed expectations that the Federal Reserve might plan to cut interest rates.
14
These informal accounts are consistent with the widespread nding that macroeconomic announcement
eects are prevalent in Treasury securities, often inducing a jump in the yields and an associated burst
of volatility; see, for example, Johannes (2004) and Andersen et al. (2007). They also suggest that the
reaction across the maturity spectrum is a function of the content of news and current economic conditions.
In particular, the volatility response is highly correlated for nearby maturities, as one would expect if the
23
economic eects were deemed stronger at the shorter-, medium-, or longer-term maturities. The variance
of the measurement error for daily realized volatility increases with the level of the underlying volatility;
see, for instance, Barndor-Nielsen and Shephard (2002b). Hence, the coherent response across nearby
maturities during extreme events suggests that the realized volatilities capture the relative size of the eects
adequately in spite of potential measurement problems.
[ FIGURE 1 ABOUT HERE ]
In Figure 2 we plot the average (sample) daily realized volatility for the zero-coupon yields for the
dierent maturities. For comparison, we also include an alternative measure of yield volatility, obtained by
computing the standard deviation of the daily changes in the zero-coupon yields (measured by the yields
average towards the end of the day, from 4:10pm to 5pm). These graphs provide informal estimates of the
unconditional term structure of volatility. Both plots exhibit the characteristic snake shape documented
in, for example, Piazzesi (2003, 2005). Since our analysis is limited to yields with maturities of at least three
months, our volatility measures are largely unaected by short-lived deviations of the short rate from the
target zone, which can push the head of the snake further up. Moreover, the hump in the back of the
snake appears less pronounced than what has been reported in some studies, including Dai and Singleton
(2000, 2003). Such discrepancies may arise from varying degrees of policy inertia across the sample periods
covered by the studies, as suggested by Piazzesi (2003). Overall, our realized volatility series replicates the
qualitative features of prior studies along this dimension nicely. Moreover, the correspondence between the
graph constructed from the intraday yield-based measures and the daily yield-based measures adds further
support to the reliability of the realized volatility series.
[ FIGURE 2 ABOUT HERE ]
A nal critical feature of the volatility yield series is the degree of temporal dependency. This feature may
be gauged from the sample correlogram for the daily logarithmic realized volatility in Figure 3. The gure
shows a distinct hyperbolically declining pattern that is readily, albeit informally, assessed in quantitative
terms by reference to the superimposed tted hyperbolic curves. The ndings are again remarkably similar
across maturities. In terms of the standard coecient for fractional integration, d, the tted curves (going
24
from the shorter to longer maturities) imply values of 0.35, 0.34, 0.35, 0.33, 0.31, and 0.30. This result
suggests stationary but long-memory persistent volatility processes for each of these series, consistent with
prior evidence from analysis of realized volatilities on equities (e.g., Andersen and Bollerslev (1997a) and
Andersen et al. (2001)) and currencies (e.g., Andersen (2000), Andersen and Bollerslev (1997a, 1997b,
1998b)).
[ FIGURE 3 ABOUT HERE ]
In conclusion, we nd that the realized yield volatility series are consistent with prior evidence in terms
of the overall level of (unconditional) yield volatility across the term structure, they capture the volatility
bursts associated with the release of macroeconomic announcements in a credible fashion, and they display
the type of temporal dependencies documented in prior studies of equity, foreign exchange, and xed income
markets. Combined with the consistent features observed across nearby maturities, these ndings suggest
that the series are highly informative regarding the underlying true yield volatility realizations. There is no
indication that the measurement errors associated with the use of 10-minute yield changes or the impact of
microstructure noise seriously impair the quality of our volatility proxies.
C.2. Daily EGARCH Volatility Forecasts
Although the realized volatility series appear credible and consistent with evidence from daily yield
series along various dimensions, the time-series behavior is critical for our subsequent volatility spanning
regressions. Thus, we now directly explore the coherence between standard yield volatility forecasts obtained
from daily data and the realized yield volatility measures based on intraday data. A thorough exploration
of this relation across the entire yield curve would require a full-length study so here we focus on the daily
three-month Treasury bill yields only. This series has been analyzed in many prior studies so the ndings
serve both as a robustness check for the (ex-post) realized yield volatility measures and as a reference point
for comparison to the existing literature. A number of subtle points are involved in the specication and
selection of the appropriate daily model. These are largely unrelated to the main issues of this paper so we
defer the details to Appendix C. In Section III, when we need some candidate yield volatility forecasts, we
25
turn to simpler models that exploit the superior information in the observed intraday-based realized yield
volatility series directly.
We obtain daily yield volatility estimates from an ARMA-EGARCH-type model for the short-term zero-
coupon yields. The ARMA structure allows for a exible conditional mean process while the literature
documents that an EGARCH representation for the conditional yield volatility provides a convenient and
parsimonious model of the conditional heteroskedasticity in the series. In addition, we introduce an addi-
tional source of dynamic interaction via an interest level eect, that is, the EGARCH conditional variance
term is scaled by the power of the yield level, as is common in the preceding time-series studies. Finally,
we stipulate that the conditional innovations belong to the family of semi-nonparametric (SNP) densities
introduced by Gallant and Nychka (1987), so that the conditional density may accommodate remaining non-
normality and time-series dependence in the residuals. The former consideration is particularly important
given the evidence of jump-like outliers in the daily yield series.
The model is applied to the three-month Treasury bill yield series.
15
A long sample is required to identify
the dynamics of a highly persistent series so we use data covering July 1, 1983 to June 30, 2005, for a total
of 5,498 observations. Earlier data are available, but we exclude the period involving the Feds monetary
experiment as this arguably represents a regime shift. Table I provides summary statistics for the daily yield
series. The rst two columns correspond to the full sample used in estimation of the SNP density outlined
in Appendix C, while the last two columns represent the June 17, 1991 to June 15, 2001 sample period
that corresponds to our realized volatility series. The basic summary statistics are comparable across the
two sample periods. In particular, the interest rate levels are close and the kurtosis of the yield changes is
exceedingly high for both samples. However, the yield volatility is relatively lower over the 1991 to 2001
period.
[ TABLE I ABOUT HERE ]
We estimate the model by (quasi-)maximum likelihood (QML). The Bayesian (BIC) and Hannan-Quinn
(H-Q) information criteria are used to guide model selection while additional information regarding the
choice of the ARMA and EGARCH terms is obtained from Ljung-Box tests for the autocorrelation of the raw
26
and squared residuals. This analysis leads us to a so-called ARMA(6,1)-Level-EGARCH(2,1)-Kz(8)-Kx(0)
specication. More information on the SNP model and its nomenclature is in Andersen and Lund (1997)
and Andersen, Benzoni, and Lund (2002, 2004). Estimation results are available from the authors upon
request. The representation implies that the conditional heteroskedasticity is captured by the EGARCH-
Level representation while the Kz(8) term accommodates strong departures from conditional normality of
the standardized daily yield innovations.
The high orders of the ARMA and EGARCH terms imply persistent and complex dynamics for the con-
ditional mean and volatility of the yield series, while the parameter estimates indicate extremely persistent
rst- and second-order conditional moments, even if the stationarity conditions are satised. Specically,
the inverse of the dominant root for the conditional mean polynomial is 0.9998, while the inverse of the
roots for the conditional variance polynomial are 0.9970 and 0.9385. Obviously, the mean dynamics are hard
to distinguish from the unit root case. The volatility roots fall within the range that produces high-order
autocorrelations consistent with the long-memory-type persistence documented in Figure 3 for the realized
volatility series.
Figure 4 depicts the EGARCH-Level-SNP model-implied one-day-ahead volatility forecast,

V (y

|x; ),
along with the corresponding daily realized volatility series, v
d,y

, where = 3M. This is literally a plot

of daily volatility forecasts versus subsequent realizations, in which we use v
d,y

as a proxy for the ex-post

volatility realizations in the spirit of equation (14). The smoothing associated with the formation of ex-
ante expectations within the EGARCH-Level model is readily apparent, in contrast to the jagged nature
of the realized volatility series. Moreover, it is evident that the extreme positive outliers in the realized
volatility series are (almost by denition) not a priori predictable. Nonetheless, there is good qualitative
coherence between the two series constructed from distinct data sources, as the long-run movements in
the yield volatility forecasts clearly correspond to shifts in the overall intensity of the realized volatility
measures. The correlation between the two series is about 44%. Moreover, the overall degree of explanatory
power of the forecasts for future realized volatility is only slightly lower than reported for futures on the
30-year U.S. Treasury yield in Andersen et al. (2007). We also note that the latter study nds the volatility
27
predictability to be considerably lower for the xed-income markets than for the major equity and currency
markets. Finally, at 0.0448% and 0.0470% per day, respectively, the sample means of the SNP forecast and
realized volatility series are close.
[ FIGURE 4 ABOUT HERE ]
In sum, our short-term realized yield volatility series is coherent with the volatility forecasts obtained
from the extended EGARCH model estimated at the daily frequency. The relation between the two series
is analogous to what has been found in prior empirical work and is consistent with the theoretical results in
Andersen, Bollerslev, and Meddahi (2004, 2005). In turn, this nding suggests that our daily realized yield
volatility series provide useful measures of the underlying quadratic yield variation realizations.
III. Evidence on the Ane Spanning Conditions
This section presents our main empirical ndings. Section III.A focuses on the contemporaneous spanning
condition for diusive ane models, while Section III.B concentrates on the predictive version of the spanning
condition that accounts for the presence of jumps.
A. Testing the Contemporaneous Spanning Condition
A.1. Baseline Case
We test the contemporaneous version of the spanning condition for diusive ane models by estimating
the empirical version of equation (18). For now we abstract from measurement errors and jumps. We return
to these important issues in subsequent sections.
The dependent variable in the regression is the realized yield volatility. To construct a proxy for the
independent variables, we compute the daily average of the intraday zero-coupon yield series with three-
month, six-month, one-year, two-year, ve-year, and 10-year maturities. To facilitate interpretation and
alleviate the multi-collinearity problem we extract orthogonal principal components from the panel of average
daily yields. As usual, the rst three components correlate very highly with empirical measures of the level,
28
slope, and curvature of the yield curve. Although these three components capture almost all of the yields
variation over time, we include all six in our regressions to ensure that our right-hand-side variables capture
the entire variation contained in the original yield series. This approach also provides some robustness by
including yields that may be relevant for the quadratic model class, which can involve a large number of
ane yield factors. We denote the principal components, or factors, by PC
j
, j = 1, . . . , 6. We use ordinary
least squares to estimate the empirical version of equation (18) via the regression model
v
2
y
(t, h) =
0
+
6

j=1

j
PC
j
(t, h) +(t, h) . (32)
We focus on the results for the daily realized volatility measures, that is, h = 1/252. For robustness, we
also discuss ndings for the weekly (h = 1/52) or monthly (h = 1/12) volatility measures in Section III.A.2.
Table II.A reports results based on the full sample. We initially focus on the regression with the shortest
maturity (three-month) realized yield volatility, that is, v
2
y
, = 3M, as the dependent variable. Remarkably,
the R
2
of this regression is only 4%. In theory, and ignoring measurement error, this regression should
explain yield volatility perfectly in a standard ane diusive model. Furthermore, it is noteworthy that
the coecients on the rst three principal components are all insignicant, while a few of the higher-order
components appear to enter signicantly in the regression.
16
Hence, the usual factors identied in dynamic
term structure models appear to be entirely unrelated to the yields volatility. This evidence is at odds with,
for example, Litterman, Scheinkman, and Weiss (1991), who conclude that interest rate volatility is linked
to the curvature factor.
[ TABLE II ABOUT HERE ]
The results from estimating equation (32) with the realized volatility of longer-maturity yields as the
dependent variable are reported in rows 2 to 6 of Table II.A. The explanatory power is now even weaker
as, for maturities beyond two years, the R
2
s are less than 0.6%. Hence, the failings of the ane diusive
models are not linked to idiosyncratic features at the short end of the maturity spectrum.
We also estimate the covariance matrix of the residuals from the six regressions reported in Table II.A
and perform a principal component analysis. This is in the spirit of Collin-Dufresne and Goldstein (2002).
29
They explore the residuals from regressions of returns on at-the-money straddles (portfolios mainly exposed
to volatility risk) against changes in swap rates. As emphasized previously, however, our test procedure
should be more powerful. Consistent with their results, we nd that the rst principal component explains
nearly 80% of the variation in the model residuals (Table II.B). This evidence suggests that the limited
explanatory power reported in Table II.A is not due to noisy data. Instead, there is a dominant common
factor, unrelated to the level of the yields, driving the yield volatility of the zero-coupon bonds at dierent
maturities. However, the yield volatility innovations are clearly not perfectly correlated across the maturity
spectrum so multiple stochastic volatility factors appear necessary to accommodate these features.
A.2. Robustness of the Results
Subsample Analysis As a robustness check, we estimate the spanning regression for dierent subsamples
and over longer horizons. For each subsample, we t six separate regressions corresponding to each of the
available yield volatility series. To conserve space, Panels A and B of Table III report only ndings for the
rst and second parts of our sample and for the three-month, two-year, and 10-year realized volatility series,
as these are representative of the full set of results.
[ TABLE III ABOUT HERE ]
The explanatory power of regressions with short-maturity realized yield volatility as the dependent
variable improves slightly compared to the results for the full sample. However, bond yields still have
virtually no power to explain the realized volatility of longer-maturity rates. Further, while the loading on
the rst principal component now is generally signicant, it is positive during the rst sample period and
negative during the second. Such abrupt sign changes suggest that the improvement reects in-sample over
tting.
17
An alternative interpretation is that there is a change in regime during the sample period but
additional analysis does not support this conjecture.
18
Monthly Volatility Spanning The analysis in Section II.C suggests that our realized volatility series pro-
vide quite accurate measurements of the true underlying yield variation. Nevertheless, there may be residual
concern regarding the impact of measurement errors in the variables entering the spanning regressions. A
30
simple robustness check involves aggregation to a lower sampling frequency, since measurement errors tend
to cancel and thus become less important relative to the accumulated signals in the yield changes and vari-
ation measures over longer horizons. Consequently, we explore the spanning condition (11) at the monthly
frequency. For that purpose, the daily realized volatility measures are aggregated into overlapping monthly
series. Similarly, we construct average monthly yields and extract the corresponding principal components.
These series are then used in estimating equation (32) by OLS. The results, in Table IV.A, are in line with
our previous ndings. The increase in overall explanatory power is consistent with the substantially lower
idiosyncratic yield variability at the monthly compared with daily frequency. Nonetheless, the explanatory
power remains very low for the longer-maturity series. In addition, for the shortest maturity yield volatility
series none of the rst four principal components, typically associated with the factors in standard ane
diusive models, are signicant. Finally, and perhaps most tellingly, the residual variation of the yield
volatilities continues to have a very strong common component as documented in Table IV.B. Hence, the
monthly yield variation also displays a strong covariation across the maturity spectrum that is unrelated to
the yield levels.
19
[ TABLE IV ABOUT HERE ]
Spanning with Alternative Yield Series We also explore whether our results are robust to the set of
proxies used for the bond yields on the right-hand side of the regression model. Specically, we consider a
second panel of daily yields consisting of the constant-maturity series released by the Board of Governors,
discussed in detail in Section II.B. In this case, we have a larger number of maturities available, namely,
three- and six-month, and one-, two-, three-, ve-, seven-, 10-, and 30-year Treasuries. After converting
the par coupon yields into zero-coupon yields, we extract principal components from the series and we
employ them in regression (32), with the dependent variables being the usual six realized volatility measures
constructed from intradaily observations. The results, in Table V, are consistent with those in Table II.
Most importantly, although the R
2
s improve marginally, we still nd little or no evidence that a portfolio of
bonds can span (hedge) volatility risk. This reinforces our earlier conclusions. First, our previous ndings
31
are seemingly not hampered by the fact that we were not using the yields on the 30-year bond. Second, the
ineectiveness of increasing the number of yields on the right-hand side of our regressions suggests that the
spanning condition is violated even in higher-order ane (or quadratic) models.
[ TABLE V ABOUT HERE ]
Instrumental Variables Regressions A noisy measure of the yields principal components on the right-
hand side of regression (32) could create an error-in-variables problem, resulting in inconsistent OLS esti-
mates of the model coecients. To address this potential problem, we estimate the model with instrumental
variables along the lines of Stambaugh (1988). In Table VI we report estimates for the case in which the
instruments are the principal components of the average daily yields lagged by one day. Alternative choices
of the instruments, for example, an average of the daily yields computed over a longer lagged time window,
produce similar results. The main result is that the instrumental variable estimates of the model coecients
remain close to the OLS values in Table II. Moreover, the explanatory power of the regressions is virtually
unchanged.
20
[ TABLE VI ABOUT HERE ]
B. The Predictive Spanning Condition
The presence of jumps breaks the realization-by-realization linkage between volatility and the cross-
section of yields, as documented in Section I.D. Version (24) of the contemporaneous spanning condition
still applies although the error term now embeds the (compensated) jump component. Therefore, the empiri-
cal analysis reported in Tables II to VI remains meaningful, but the R
2
associated with the contemporaneous
spanning regressions will be below unity. Hence, in the general setting where jumps (or signicant mea-
surement errors) are present, we do not have a rm benchmark for gauging the explanatory power of the
contemporaneous regressions.
Consequently, we turn to the predictive version of the spanning condition given in equation (26), which is
robust to the presence of jumps. Under weak auxiliary assumptions, equation (26) holds for the conditional
32
expectation evaluated under the P measure, and this is the version we examine.
The regression model (20) provides a natural framework for testing this predictive spanning restriction.
In its general form, it includes auxiliary forecast variables in addition to the cross-section of yields. In the
traditional ane setting, these auxiliary forecasts should not improve the yield volatility forecasts extracted
from the current yield curve. Below, we reject this implication: the predictive power of the yields is very
low and lagged realized volatility measures have signicant incremental predictive power. Hence, the yields
fail to capture a signicant and predictable component of the future yield variation.
B.1. Baseline Case
We rst consider the regression (20) including only the current yields (principal components) as regres-
sors. Under the null hypothesis, the conditional expectation of the future quadratic yield variation is an
ane combination of these yields and the regression mechanically maximizes the in-sample R
2
. Hence, the
explanatory power of the regression should serve as an upper bound on the degree of predictability inherent
in the yield volatility. Table VII.A reports OLS results for
v
2
y
(t +h, h) =
0
+
6

j=1

j
PC
j
(t, h
D
) +(t +h, h) , (33)
where the dependent variable is the daily (h = 1/252), overlapping weekly (h = 1/52), and overlapping
monthly (h = 1/12) realized volatility for zero-coupon yields with maturity , and PC
j
(t, h
D
), j = 1, . . . , 6,
are the principal components extracted from the sample of average past daily yields.
21
[ TABLE VII, Panel A ABOUT HERE ]
It is evident that the yield cross-section contains very limited information regarding the future yield
volatility. The explanatory power of the regressions is small at the one-day-ahead forecasting horizon
(h = 1/252). At the weekly and monthly horizons the regressions R
2
improves, but the predictive power
for volatility of long-term yields remains negligible. For example, at the 10-year maturity, the R
2
for the
h = 1/52 and h = 1/12 cases is below 3% and 6%, respectively. Moreover, the results are remarkably similar
to those reported in Table II for the daily contemporaneous spanning regressions. This nding suggests that
the daily innovations to the yield curve are not systematically related to the corresponding innovations in
33
yield volatility. In order to assess whether these volatility residuals are truly unpredictable, we turn to an
alternative ad hoc forecasting procedure for realized volatility.
B.2. Realized Volatility Component Model Forecasts
We consider a simple forecasting method that relies on time-series models built directly on the history of
ex-post volatilities, denoted HAR-RV. This approach has been studied in Andersen, Bollerslev, and Diebold
(2007), Corsi (2003), and M uller et al. (1997), and provides an elementary yet powerful alternative to the
more highly parametric EGARCH-style forecasts explored earlier. The HAR-RV model takes the form
v
2
y

(t +h, h) =
0
+
D
v
2
y

(t, h
D
) +
W
v
2
y

(t, h
W
) +
M
v
2
y

(t, h
M
) +(t +h, h) , (34)
where the dependent variable is the daily (h = 1/252), overlapping weekly (h = 1/52), and overlapping
monthly (h = 1/12) realized volatility for zero-coupon yields with maturity , while the right-hand side
variables are the daily, weekly, and monthly realized volatility (h
D
= 1/252, h
W
= 1/52, and h
M
= 1/12).
The mixing of three volatility components allows for a slow volatility autocorrelation decay that closely
approximates a hyperbolic pattern, which is consistent with the properties of our realized volatility series
documented in Section II.C.1. In addition, the forecast procedure ts readily within the regression-based
diagnostic tests developed in the preceding sections.
In Table VII.B we report OLS estimation results for the HAR-RV model (34). The estimates for
D
,

W
, and
M
conrm the existence of highly persistent volatility dependence. Moreover, the R
2
s improve
considerably with the forecasting horizon, consistent with the theory in Andersen, Bollerslev, and Meddahi
(2004) and the evidence documented in, for example, Andersen et al. (2003a). Intuitively, the pronounced
right skew in the quadratic variation process induced by occasional short-lived volatility bursts renders
the predictive power for future ex-post sample variation low, but this eect is somewhat mitigated at
intermediate forecast horizons.
[ TABLE VII, Panel B ABOUT HERE ]
All in all, there is reasonable coherence between the HAR-RV forecasts and the ex-post volatility realiza-
tions. This is illustrated in Figure 5, which depicts the one-day-ahead HAR-RV forecasts, v
d,y

, along with
34
the corresponding realized volatility series, v
d,y
. It is evident that long-run movements in the forecasts are
related to corresponding shifts in the overall intensity of the realized yield volatility measures. In this re-
gard, the performance of these simple predictors appears roughly on par with the more complex augmented
EGARCH model presented earlier.
22
[ FIGURE 5 ABOUT HERE ]
The HAR-RV regressions in Table VII.B provide a natural benchmark for the volatility forecasts in
Section III.B.1, constructed from the yields alone. It appears that the predictive power of the HAR-RV
volatility components is substantially higher than that implied by the zero-coupon yields. For example, the
R
2
for the one-day-ahead forecast at the three-month and 10-year maturities are 4.7% and 0.6% in Panel
A, versus 8.4% and 2.1% in Panel B. The improvement in predictive performance suggests that the ad hoc
time-series forecasts provide information over-and-above that embedded in the yields cross-section.
B.3. Volatility Spanning with Jumps and Measurement Error
A more direct test of the predictive spanning restriction is given by the encompassing regression in Table
VII.C. Here, the set of explanatory variables includes the conditional HAR-RV yield volatility forecasts
formed at time t as well as the principal components of the time-t zero-coupon yields,
v
2
y
(t +h, h) =
0
+
6

j=1

j
PC
j
(t, h
D
) +
D
v
2
y
(t, h
D
) +
W
v
2
y
(t, h
W
) +
M
v
2
y
(t, h
M
) +(t +h, h) . (35)
The three HAR-RV related regressors enter almost uniformly with positive and mostly statistically signicant
coecients. In contrast, the coecients on the yields (not reported) are generally insignicant. Hence, there
is evidence of systematic predictability in yield volatility for all the maturities and across daily, weekly, and
monthly forecast horizons that is captured by the time-series forecasts but not by the yield cross-section. In
addition, it is noteworthy that the regression coecients on the past realized volatility components invariably
change very little with the addition of the yield principal components as regressors and the increase in the
overall explanatory power of the regressions is very small.
[ TABLE VII, Panel C ABOUT HERE ]
35
The evidence in Panels B and C shows that a non-trivial component of the yield volatility is predictable
at the daily, weekly, and monthly frequency. At the same time, it runs counter to the hypothesis that the
conditional expectation of the yield volatility may be expressed as an ane combination of current yields. In
fact, the information in the yield curve seems largely irrelevant for volatility prediction relative to forecasts
generated from a simple time-series model.
IV. Conclusions
In recent years, the study of the term structure of interest rates has relied predominantly on continuous-
time multi-factor models and, in particular, on ane specications. A general implication of these models
is that the (expected) quadratic yield variation for any xed-maturity zero-coupon bond is spanned by the
contemporaneous yield cross-section. That is, standard ane models predict that interest rate volatility can
be extracted from current bond prices.
We rely on model-free, yet ecient, yield volatility measures constructed from high frequency intraday
data to directly test the yield volatility implications for a very broad class of term structure models. We
show that neither realized nor expected future quadratic yield variation is spanned by the term structure of
zero-coupon yields. In fact, we nd a pronounced and systematic covariation in yield volatility across the
maturity spectrum that appears unrelated to the state of the term structure.
Our results shed new light on related empirical ndings in the literature. First, there is extensive
evidence that short-term interest rates display pronounced stochastic volatility features that are largely
unrelated to the level of the short-term interest rate itself; see, for example, Brenner, Harjes, and Kroner
(1996) and Andersen and Lund (1997). It does seem natural, in theory, to relate the short-rate volatility
to the curvature factor in the yield curve, as originally proposed by Litterman, Scheinkman, and Weiss
(1991). This specic linkage has subsequently been questioned by various studies, and we conrm that there
is little empirical support for it. Indeed, our results imply more generally that there is, at best, a very weak
relation between the short-rate volatility factor and the term structure. Moreover, any such link becomes
virtually non-existent when we relate long maturity yield volatility to the yield curve. Hence, the model
36
failings are even more glaring at the long end of the term structure.
Second, we also gain a fresh perspective on the recent conicting evidence regarding the ane USV
restriction. If we accept that standard ane models are seriously decient in terms of their implications
for the yield volatility dynamics, it is inherently dicult to interpret hypothesis tests concerning the USV
restriction within a broader ane model setting. In fact, rejecting the USV hypothesis in favor of an
encompassing ane model leads directly back to the empirical conundrum highlighted in our results.
Taken together, our ndings suggest that further extensions to the term structure modeling framework
are warranted. The literature provides a variety of interesting directions to pursue. One natural and popular
approach involves the linkage of the underlying yield curve factors to macroeconomic variables. This may
set the stage for a better understanding of the interaction between the term structure dynamics on the
one hand and monetary policy, the general economic environment, and inationary expectations on the
other.
23
A related strand of literature explores the yield curve reaction to the release of regularly scheduled
macroeconomic announcements. These news releases contribute signicantly to the observed total quadratic
yield variation through both jumps and the associated short-lived volatility bursts in the xed-income
markets. Our realized volatility measures conrm this evidence. Further, they suggest that the reaction
across the maturity spectrum is a function of the content of news as well as the prevaling economic conditions.
In particular, the volatility response is highly correlated for nearby maturities, as one would expect if the
economic eects vary in terms of their impact for the shorter, medium, or longer-term maturities. Hence,
models with general linkages of the term structure of yields and volatilities to macroeconomic and monetary
policy variables, and associated time-varying reactions to scheduled announcements, may be particularly
promising.
At the more methodological level, we expect that the use of high frequency intraday bond data for
improved measurement of the real-time evolution in yield volatility will be informative regarding the speci-
cation of future candidate term structure models.
37
Appendix A: Computing Zero-coupon Yields
There is an array of methods to extract zero-coupon yields from a cross-section of bond prices and a
vast literature has studied the relative advantages of dierent approaches. The key feature underlying the
dierent methods is the choice of a specic interpolation to t the discount rate (or forward rate) curve. As
noted by Dai, Singleton, and Yang (2007), on one extreme is the unsmoothed Fama-Bliss approach, which
iteratively extracts forward rates from coupon bond prices by building a piecewise linear discount rate
function. As such, the implied discount rates exhibit kinks at the maturities of the coupon bonds used. At
the opposite extreme the Nelson-Siegel-Bliss and smoothed Fama-Bliss methods approximate the discount
rates with exponential functions of time to maturity, which yields a smooth term structure interpolation.
In between these two extremes are methods that rely on a spline curve interpolation for the discount or
forward rates.
Here, we use a spline-based method to extract the zero-coupon yields. More specically, we t the curve
of zero-coupon yields with a cubic spline interpolation y(, ), where is the vector of spline coecients,
is time-to-maturity, and
min

max
, with
min
and
max
denoting the nearest and farthest maturities,
respectively. We dene P

k
, k = 1, . . . , N, to be the time-t market price of a bond with maturity
k
24
and
dene

P

k
to be the price of the same bond computed by discounting its coupon and principal payments at
the discount rate y. Next, we choose to solve the problem
min

k=1
(P

k


P

k
)
2
+

max

min
() y

(, )
2
d

. (A1)
This is the approach preferred by Waggoner (1997) and used in, for example, Dai, Singleton, and Yang
(2007), except that we t the smoothed cubic spline directly on the zero-coupon yields curve (similar to
McCulloch (1975)), while Waggoner (1997) and Dai, Singleton, and Yang (2007) t the smoothed spline
on the forward rates curve. We follow Fisher, Nychka, and Zervos (1995) and impose a penalty for the
roughness of the spline interpolation, that is, we t a smoothed spline. As in Waggoner (1997) and Dai,
Singleton, and Yang (2007), we calibrate the penalty function () to be more severe at long maturities,
which stiens the spline interpolation and reduces its oscillations at the long end of the term structure. In
38
contrast, we impose a small penalty for roughness at short maturities, where more exibility is necessary to
t the prices of short-term securities, which are available in larger number than long-term bonds. Waggoner
(1997) recommends using a number of knots approximately equal to one-third the number of bonds used
in the construction of the spline. Dai, Singleton, and Yang (2007) instead use a larger number of knots
(as many as 50 to 60). In our application, we take an approach that falls in between these two by using a
number of knots approximately equal to the number of bonds used in the estimation.
It is useful to illustrate the properties of the zero-coupon yields that we extract from bond prices through
the spline interpolation. To this end, we construct three series of zero-coupon yields using on-the-run GovPX
quotes at the 10-minute frequency for three-month, six-month, and one-year Treasury bills. These series do
not rely on the spline term structure interpolation method. As such, depending on the Treasurys issuing
calendar there is a small mismatch in the yields maturities compared with the zero-coupon yields series
that we use in our analysis. Further, we construct three series of coupon yields using on-the-run GovPX
quotes at the 10-minute frequency for two-year, ve-year, and 10-year Treasury notes. Also in this case,
these series do not rely on the term structure interpolation procedure. Thus, in addition to being subject to
a small maturity mismatch, these series dier from those used in our analysis because they contain coupon
rather than zero-coupon yields.
We compute the correlation between each one of these six series and the corresponding series of zero-
coupon yields constructed by solving problem (A1). The results are in Table A.I.A. The sample correlations
are nearly perfect at all maturities, which suggests that our spline interpolation method does not dramatically
change the nature of the yield series that we investigate in the paper. We repeat this analysis for the data
set of daily constant-maturity (TCM) par yields released by the Federal Reserve Board. Again, we nd that
the sample correlations between the series of TCM par yields and the series of zero-coupon yields extracted
through our term structure interpolation are nearly perfect (Table A.I.B).
[TABLE A.I ABOUT HERE]
Next, we measure the sample correlations between realized volatility series constructed using the GovPX
zero-coupon yield series and the series of raw yields. Also in this case, for Treasury bills the raw yields
39
are zero-coupon yields with maturities closest to three months, six months, and one year. For Treasury
notes, the raw yields are coupon yields with maturities closest to two years, ve years, and 10 years. The
results, reported in Table A.II, show that these sample correlations are nearly perfect for maturities up to
ve years, and are approximately 97% at the 10-year horizon.
[TABLE A.II ABOUT HERE]
Finally, we use the realized volatility series constructed from raw GovPX yields to conduct the entire
analysis that we perform in the paper with the zero-coupon yields. The results, available from the authors
upon request, are very close to those reported in the paper for zero-coupon yield series. For short-maturity
yields, they are nearly identical.
Overall, this evidence suggests that the conclusions of our analysis are not biased by the noise due to
the procedure that we use to extract the series of zero-coupon yields.
Appendix B: Quotes Updates and Sampling Scheme
In our analysis, at the end of each 10-minute interval we use the immediately preceding on-the-run quote
to construct the relevant bid and ask prices. Due to infrequent trading, a quote may not be updated for
more than 10 consecutive minutes. In these cases, our approach is to retain the last observed quote until a
new quote becomes available.
Such a sampling scheme is economically meaningful. If there is no arrival of new information, brokers do
not feel the need to update their quotes. As such, the previous quote still reects the brokers assessment
of the securitys value. This is the price that we use in our analysis. In some cases, quotes could be stale
and therefore misaligned with the securities values. Eventually, however, such deviations will be corrected
when the broker adjusts the quote. Such an adjustment will be reected in our realized volatility measures.
When activity is very subdued this delay could create some bias in our volatility estimates. Thus, we discard
those days for which we could not nd any trading activity for a period longer than three hours.
In Table B.I below, we report the percentage number of unique yield observations in the intraday GovPX
data set. The gures in Panel A are computed using the full sample period, from June 17, 1991 to June 15,
40
2001. In Panel B we report the same percentage ratios computed after discarding trading days with three or
more hours of inactivity. In all cases, more than 80% of the yields observations are unique quotes. This
ratio increases to nearly 100% for the two-year notes (which are among the most liquid Treasury contracts)
and is well in excess of 90% for most other maturities. As expected, discarding days with sparse trading
activity results in an even higher percentage of unique quotes.
[TABLE B.I ABOUT HERE]
Appendix C: The SNP Model for the Daily Three-month Yields
In summary form, the ARMA-EGARCH-Level-SNP model for the daily three-month yield takes the
general form
y
t
=
t
+y

t1

h
t
z
t
, (C1)

t
=
0
+
s

i=1

i
y
ti
+
u

i=1

i
(y
ti

ti
) , (C2)
ln h
t
= (1
p

i=1

i
) +
p

i=1

i
ln h
ti
+ (1 +
1
L +... +
q
L
q
) [
1
z
t1
+
2
(b(z
t1
)

2/) ] ,
(C3)
where the index t is measured in daily units. The conditional density for the yield innovation, z
t
, is given
by a exible SNP representation. The full density takes the form
f
K
(y
t
|x
t
; ) =

+ (1 )
[P
K
(z
t
, x
t
)]
2

R
[P
K
(z
t
, x
t
)]
2
(u)du

(z
t
)
y

t1

h
t
, = 0.01 , (C4)
where:
z
t
=
y
t

t
y

t1

h
t
,

t
=
0
+
s

i=1

i
y
ti
+
u

i=1

i
(y
ti

ti
) ,
lnh
t
= (1
p

i=1

i
) +
p

i=1

i
lnh
ti
+ (1 +
1
L +... +
q
L
q
) [
1
z
t1
+
2
(b(z
t1
)

2/) ] ,
b(z) = |z| for |z| /2K, b(z) = (/2 cos(Kz))/K for |z| < /2K ,
P
K
(z, x) =
Kz

i=0
a
i
(x)z
i
=
Kz

i=0

Kx

|j|=0
a
ij
x
j

z
i
, a
00
= 1 ,
41
where j is a multi-index vector, x
j
(x
j
1
1
, . . . , x
j
M
M
), and |j |

M
m=1
j
m
.
References
Ahn, Dong-Hyun, Robert F. Dittmar, and A. Ronald Gallant, 2002, Quadratic term structure models:
Theory and evidence, Review of Financial Studies 15, 243-288.
Ahn, Dong-Hyun, Robert F. Dittmar, A. Ronald Gallant, and Bin Gao, 2003, Purebred or hybrid?: Re-
producing the volatility in term structure dynamics, Journal of Econometrics 116, 147-180.
At-Sahalia, Yacine, Per A. Mykland, and Lan Zhang, 2005a, How often to sample a continuous-time
process in the presence of market microstructure noise, Review of Financial Studies 18, 351-416.
At-Sahalia, Yacine, Per A. Mykland, and Lan Zhang, 2005b, Ultra high-frequency volatility estimation
with dependent microstructure noise, NBER Working paper #11380.
Andersen, Torben G., 2000, Some reections on analysis of high-frequency data, Journal of Business and
Economic Statistics 18, 146-153.
Andersen, Torben G., Luca Benzoni, and Jesper Lund, 2002, An empirical investigation of continuous-time
equity return models, Journal of Finance 57, 1239-1284.
Andersen, Torben G., Luca Benzoni, and Jesper Lund, 2004, Stochastic volatility, mean drift and jumps in
the short-term interest rate, Working paper, Northwestern University, University of Minnesota, and
Nykredit Bank.
Andersen, Torben G., and Tim Bollerslev, 1997a, Intraday periodicity and volatility persistence in nancial
markets, Journal of Empirical Finance 4, 115-158.
Andersen, Torben G., and Tim Bollerslev, 1997b, Heterogeneous information arrivals and return volatility
dynamics: Uncovering the long-run in high frequency returns, Journal of Finance 52, 975-1005.
42
Andersen, Torben G., and Tim Bollerslev, 1998a, Answering the skeptics: Yes, standard volatility models
do provide accurate forecasts, International Economic Review 39, 885-905.
Andersen, Torben G., and Tim Bollerslev, 1998b, Deutsche Mark-Dollar volatility: Intraday activity pat-
terns, macroeconomic announcements, and longer run dependencies, Journal of Finance 53, 219-265.
Andersen, Torben G., Tim Bollerslev, and Francis X. Diebold, 2004, Parametric and nonparametric volatil-
ity measurement, in Lars P. Hansen and Yacine At-Sahalia, eds. Handbook of Financial Econometrics
(North-Holland, Amsterdam), forthcoming.
Andersen, Torben G., Tim Bollerslev, and Francis X. Diebold, 2007, Roughing it up: Including jump
components in the measurement, modeling and forecasting of return volatility, Review of Economics
and Statistics 89, 701-720.
Andersen, Torben G., Tim Bollerslev, Francis X. Diebold, and Heiko Ebens, 2001, The distribution of
realized stock return volatility, Journal of Financial Economics 61, 43-76.
Andersen, Torben G., Tim Bollerslev, Francis X. Diebold, and Paul Labys, 2003a, Modeling and forecasting
realized volatility, Econometrica 71, 579-625.
Andersen, Torben G., Tim Bollerslev, Francis X. Diebold, and Clara Vega, 2003b, Micro eects of macro
announcements: Real-time price discovery in foreign exchange, American Economic Review 93, 38-62.
Andersen, Torben G., Tim Bollerslev, Francis X. Diebold, and Clara Vega, 2007, Real-time price discovery
in stock, bond and foreign exchange markets, Journal of International Economics 73, 251-277.
Andersen, Torben G., Tim Bollerslev, and Nour Meddahi, 2004, Analytical evaluation of volatility forecasts,
International Economic Review 45, 1079-1110.
Andersen, Torben G., Tim Bollerslev, and Nour Meddahi, 2005, Correcting the errors: Volatility forecast
evaluation using high-frequency data and realized volatilities, Econometrica 73, 279-296.
43
Andersen, Torben G., and Jesper Lund, 1997, Estimating continuous-time stochastic volatility models of
the short-term interest rate, Journal of Econometrics 77, 343-377.
Ang, Andrew, Monika Piazzesi, and Min Wei, 2006, What does the yield curve tell us about GDP growth?
Journal of Econometrics 131, 359-403.
Balduzzi, Pierluigi, Edwin J. Elton, and T. Clifton Green, 2001, Economic news and bond prices: Evidence
from the U.S. Treasury market, Journal of Financial and Quantitative Analysis 36, 523-543.
Bandi, Federico M., and Jerey R. Russell, 2006, Separating microstructure noise from volatility, Journal
of Financial Economics 79, 655-692.
Bandi, Federico M., and Jerey R. Russell, 2008, Microstructure noise, realized variance, and optimal
sampling, Review of Economic Studies 75, 339-369.
Barndor-Nielsen, Ole E., Peter R. Hansen, Asger Lunde, and Neil Shephard, 2004, Regular and modied
kernel-based estimators of integrated variance: The case with independent noise, Working paper,
University of Aarhus, Stanford University, and University of Oxford.
Barndor-Nielsen, Ole E., and Neil Shephard, 2002a, Econometric analysis of realized volatility and its
use in estimating stochastic volatility models, Journal of the Royal Statistical Society, Series B 64,
253-280.
Barndor-Nielsen, Ole E., and Neil Shephard, 2002b, Estimating quadratic variation using realized vari-
ance, Journal of Applied Econometrics 17, 457-477.
Barndor-Nielsen, Ole E., and Neil Shephard, 2004a, Econometric analysis of realized covariation: High
frequency based covariance, regression and correlation in nancial economics, Econometrica 72, 885-
925.
Barndor-Nielsen, Ole E., and Neil Shephard, 2004b, Power and bipower variation with stochastic volatility
and jumps, Journal of Financial Econometrics 2, 1-37.
44
Beaglehole, David, and Mark Tenney, 1991, General solutions of some interest rate-contingent claim pricing
equations, Journal of Fixed Income 1, 69-83.
Beaglehole, David, and Mark Tenney, 1992, Corrections and additions to a nonlinear equilibrium model
of the term structure of interest rates, Journal of Financial Economics 32, 345-353.
Bikbov, Ruslan, and Mikhail Chernov, 2004, Term structure and volatility: Lessons from the Eurodollar
markets, Working paper, Columbia Business School.
Bikbov, Ruslan, and Mikhail Chernov, 2005, No-arbitrage macroeconomic determinants of the yield curve,
Working paper, Columbia Business School.
Bliss, Robert R., and David C. Smith, 1998, The elasticity of interest rate volatility: Chan, Karolyi,
Longsta, and Sanders revisited, Journal of Risk 1, 21-46.
Bollerslev, Tim, Jun Cai, and F. M. Song, 2000, Intraday periodicity, long-memory volatility, and macroe-
conomic announcement eects in the U.S. Treasury bond market, Journal of Empirical Finance 7,
37-55.
Boudoukh, Jacob, Christopher Downing, Matthew Richardson, Richard Stanton, and Robert F. Whitelaw,
2007, A multifactor, nonlinear, continuous-time model of interest rate volatility, Working paper, New
York University, and University of California, Berkeley.
Brandt, Michael W., and David A. Chapman, 2002, Comparing multifactor models of the term structure,
Working paper, Duke University, and Boston College.
Brandt, Michael W., and Kenneth A. Kavajecz, 2004, Price discovery in the U.S. Treasury market: The
impact of order ow and liquidity on the yield curve, Journal of Finance 59, 2623-2654.
Brenner, Robin J., Richard H. Harjes, and Kenneth F. Kroner, 1996, Another look at models of the
short-term interest rate, Journal of Financial and Quantitative Analysis 31, 85-107.
45
Casassus, Jamie, Pierre Collin-Dufresne, and Robert S. Goldstein, 2005, Unspanned stochastic volatility
and xed income derivatives pricing, Journal of Banking and Finance 29, 2723-2749.
Cheng, Peng, and Olivier Scaillet, 2007, Linear-quadratic jump-diusion modeling, Mathematical Finance
17, 575-598.
Cheridito, Patrick, Damir Filipovic, and Robert L. Kimmel, 2007, Market price of risk specications for
ane models: Theory and evidence, Journal of Financial Economics 83, 123-170.
Collin-Dufresne, Pierre, and Robert S. Goldstein, 2002, Do bonds span the xed income markets? Theory
and evidence for unspanned stochastic volatility, Journal of Finance 57, 1685-1730.
Collin-Dufresne, Pierre, Robert S. Goldstein, and Christopher S. Jones, 2008, Can interest rate volatility
be extracted from the cross section of bond yields? Journal of Financial Economics, forthcoming.
Constantinides, George M., 1992, A theory of the nominal term structure of interest rates, Review of
Financial Studies 5, 531-552.
Corsi, Fulvio, 2003, A simple long memory model of realized volatility, Working paper, University of
Southern Switzerland.
Dai, Qiang, and Thomas Philippon, 2004, Fiscal policy and the term structure of interest rates, Working
paper, University of North Carolina, and New York University.
Dai, Qiang, and Kenneth J. Singleton, 2000, Specication analysis of ane term structure models, Journal
of Finance 55, 1943-1978.
Dai, Qiang, and Kenneth J. Singleton, 2003, Term structure dynamics in theory and reality, Review of
Financial Studies 16, 631-678.
Dai, Qiang, Kenneth J. Singleton, and Wei Yang, 2007, Regime shifts in a dynamic term structure model
of U.S. Treasury bond yields, Review of Financial Studies 20, 1669-1706.
Das, Sanjiv, 2002, The surprise element: Jumps in interest rates, Journal of Econometrics 106, 27-65.
46
Diebold, Francis X., Monika Piazzesi, and Glenn Rudebusch, 2005, Modeling bond yields in nance and
macroeconomics, American Economic Review: Papers and Proceedings 95, 415-420.
Duarte, Jeerson, 2004, Evaluating an alternative risk preference in ane term structure models, Review
of Financial Studies 17, 370-404.
Duee, Greg R., 2002, Term premia and interest rate forecasts in ane models, Journal of Finance 57,
405-443.
Due, Darrell, and Rui Kan, 1996, A yield-factor model of interest rates, Mathematical Finance 6, 379-406.
Due, Darrell, Jun Pan, and Kenneth J. Singleton, 2000, Transform analysis and asset pricing for ane
jump diusions, Econometrica 68, 1343-1376.
Fan, Rong, Anurag Gupta, and Peter Ritchken, 2003, Hedging in the possible presence of unspanned
stochastic volatility: Evidence from swaption markets, Journal of Finance 58, 2219-2248.
Fisher, Mark, Douglas Nychka, and Daniel Zervos, 1995, Fitting the term structure of interest rates with
smoothing splines, Working paper, Federal Reserve Board.
Fleming, Michael J., 1997, The round-the-clock market of U.S. Treasury securities, Federal Reserve Bank
of New York Economic Policy Review 3, 9-32.
Fleming, Michael J., 2003, Measuring Treasury market liquidity, Federal Reserve Bank of New York Eco-
nomic Policy Review 9, 83-108.
Fleming, Michael J., and Kenneth D. Garbade, 2002, When the back oce moved to the front burner:
Settlement fails in the Treasury market after 9/11, Economic Policy Review 8, 35-57.
Fleming, Michael J., and Eli M. Remolona, 1997, What moves the bond market? Federal Reserve Bank of
New York Economic Policy Review 3, 31-50.
Fleming, Michael J., and Eli M. Remolona, 1999, Price formation and liquidity in the U.S. Treasury market:
The response to public information, Journal of Finance 54, 1901-1915.
47
French, Kenneth R., G. William Schwert, and Robert F. Stambaugh, 1987, Expected stock returns and
volatility, Journal of Financial Economics 19, 3-29.
Gallant, A. Ronald, and Douglas W. Nychka, 1987, Semi-nonparametric maximum likelihood estimation,
Econometrica 55, 363-390.
Green, T. Clifton, 2004, Economic news and the impact of trading on bond prices, Journal of Finance 59,
1201-1234.
Han, Bing, 2007, Stochastic volatilities and correlations of bond yields, Journal of Finance 62, 1491-1524.
Hansen, Peter R., and Asger Lunde, 2006, Realized variance and market microstructure noise, Journal of
Business and Economic Statistics 24, 127-218.
Hsieh, David A., 1991, Chaos and nonlinear dynamics: Application to nancial markets, Journal of Finance
46, 1839-1877.
Jacobs, Kris, and Lot Karoui, 2007, Conditional volatility in ane term structure models: Evidence from
Treasury and swap markets, Journal of Financial Economics, forthcoming.
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, 113-146.
Jarrow, Robert, Haitao Li, and Feng Zhao, 2007, Interest rate caps smile too! But can the LIBOR
market models capture the smile? Journal of Finance 62, 345-382.
Johannes, Michael, 2004, The statistical and economic role of jumps in continuous-time interest rate models,
Journal of Finance 59, 227-260.
Joslin, Scott, 2006, Can unspanned stochastic volatility models explain the cross section of bond volatilities?
Working paper, Massachusetts Institute of Technology.
Joslin, Scott, 2007, Pricing and hedging volatility risk in xed income markets, Working paper, Mas-
sachusetts Institute of Technology.
48
Kimmel, Robert L., 2004, Modeling the term structure of interest rates: A new approach, Journal of
Financial Economics 72, 143-183.
Li, Haitao, and Feng Zhao, 2006, Unspanned stochastic volatility: Evidence from hedging interest rate
derivatives, Journal of Finance 61, 341-378.
Li, Li, and Robert F. Engle, 1998, Macroeconomic announcements and volatility of Treasury futures,
Working paper, University of California, San Diego.
Litterman, Robert, and Jose A. Scheinkman, 1991, Common factors aecting bond returns, Journal of
Fixed Income 1, 54-61.
Litterman, Robert, Jose A. Scheinkman, and Laurence Weiss, 1991, Volatility and the yield curve, Journal
of Fixed Income 1, 49-53.
Longsta, Francis A., 1989, A nonlinear general equilibrium model of the term structure of interest rates,
Journal of Financial Economics 23, 195-224.
McCulloch, J. Huston, 1975, The tax-adjusted yield curve, Journal of Finance 30, 811-830.
Meddahi, Nour, 2002, A theoretical comparison between integrated and realized volatility, Journal of
Applied Econometrics 17, 475-508.
M uller, Ulrich A., Michel M. Dacorogna, Rakhal D. Dave, Richard B. Olsen, Olivier V. Pictet, and Jacob
E. von Weizsacker, 1997, Volatilities of dierent time resolutionsAnalyzing the dynamics of market
components, Journal of Empirical Finance 4, 213-239.
Pasquariello, Paolo, and Clara Vega, 2007, Informed and strategic order ow in the bond markets, Review
of Financial Studies 20, 1975-2019.
Piazzesi, Monika, 2003, Ane term structure models, in Lars P. Hansen and Yacine At-Sahalia, eds.
Handbook of Financial Econometrics (North-Holland, Amsterdam), forthcoming.
Piazzesi, Monika, 2005, Bond yields and the Federal Reserve, Journal of Political Economy, 113, 311-344.
49
Poterba, James M., and Lawrence H. Summers, 1986, The persistence of volatility and stock market
uctuations, American Economic Review 76, 1142-1151.
Schwert, G. William, 1989, Why does stock market volatility change over time? Journal of Finance 44,
1115-1153.
Stambaugh, Robert F., 1988, The information in forward rates: Implications for models of the term
structure, Journal of Financial Economics 21, 41-70.
Thompson, Samuel, 2004, Identifying term structure volatility from the LIBOR-swap curve, Working paper,
Harvard University.
Trolle, Anders B., and Eduardo S. Schwartz, 2007a, Unspanned stochastic volatility and the pricing of
commodity derivatives, Review of Financial Studies, forthcoming.
Trolle, Anders B, and Eduardo S. Schwartz, 2007b, A general stochastic volatility model for the pricing of
interest rate derivatives, Review of Financial Studies, forthcoming.
Umantsev, Leonid A., 2001, Econometric analysis of European LIBOR-based options within ane term-
structure models, Ph.D. dissertation, Stanford University.
Waggoner, Daniel F., 1997, Spline methods for extracting interest rate curves from coupon bond prices,
Working paper, Federal Reserve Bank of Atlanta.
50
Table I
Summary Statistics for the U.S. Three-month Treasury Bill Yield Data
Summary statistics are computed using daily U.S. three-month Treasury bill yield data.
07/01/198306/30/2005 06/17/199106/15/2001
y y y y
Mean 5.0876 -0.0010 4.6441 -0.0009
Std. Dev. 2.3041 0.0568 0.9198 0.0467
Skewness 0.0475 -0.5392 -0.5512 -0.7656
Kurtosis 2.5134 15.6448 2.2427 19.8250
Table II
Evidence on the Contemporaneous Spanning Condition
OLS estimates for the regression v
2
y
(t, h) =
0
+

6
j=1

j
PC
j
(t, h) + (t, h) . The dependent variable is
the annualized daily realized volatility (h = 1/252) for zero-coupon yields with maturity = 3M, 6M, 1Y,
2Y, 5Y, and 10Y. The explanatory variables, PC
j
, j = 1, . . . , 6, are the six principal components extracted
from the six daily averaged zero-coupon yields. The coecient t-ratios in square brackets are based on
(Newey-West) heteroskedasticy and autocorrelation robust standard errors.
Panel A: Daily observations from 06/17/1991 to 06/15/2001
Dep. variable

0

1

2

3

4

5

6
R
2
Adj.
[t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio]
v
2
y

, = 3M
0.5563 0.0287 -0.0890 -0.2616 2.3036 -3.7902 2.6120 4.02%
[12.56] [ 1.02] [-1.38] [-1.32] [ 2.28] [-2.88] [ 2.51]
v
2
y
, = 6M
0.5381 0.0193 -0.0090 0.3128 1.8899 -1.3124 2.1202 1.96%
[14.10] [ 0.89] [-0.18] [ 1.51] [ 2.47] [-1.67] [ 2.07]
v
2
y

, = 1Y
0.6814 0.0177 0.0406 0.4573 1.6214 -0.4782 2.6390 1.30%
[15.84] [ 0.83] [ 0.83] [ 2.00] [ 2.13] [-0.58] [ 2.11]
v
2
y

, = 2Y
0.8914 0.0066 0.0067 0.5495 1.2593 0.5477 2.6234 0.55%
[15.97] [ 0.26] [ 0.13] [ 1.80] [ 1.47] [ 0.48] [ 1.61]
v
2
y
, = 5Y
0.9508 0.0085 -0.0014 0.1453 1.2338 0.7189 2.0102 0.24%
[17.98] [ 0.36] [-0.03] [ 0.49] [ 1.69] [ 0.64] [ 1.38]
v
2
y

, = 10Y
0.8745 0.0000 0.0621 0.2904 0.8540 0.0408 2.3428 0.47%
[20.96] [ 0.00] [ 1.34] [ 1.24] [ 1.31] [ 0.05] [ 1.88]
Panel B: Percent of residual variation explained by principal components
extracted from the regressions error terms
1st 2nd 3rd 4th 5th 6th
78.16% 10.46% 7.26% 2.40% 1.08% 0.63%
51
Table III
Evidence on the Contemporaneous Spanning Condition for Subsamples
OLS estimates for the regression v
2
y

(t, h) =
0
+

6
j=1

j
PC
j
(t, h) + (t, h) . The dependent variable is
the annualized daily realized volatility (h = 1/252) for zero-coupon yields with maturity = 3M, 2Y, and
10Y. The explanatory variables, PC
j
, j = 1, . . . , 6, are the six principal components extracted from the six
daily averaged zero-coupon yields. The coecient t-ratios in square brackets are based on (Newey-West)
heteroskedasticy and autocorrelation robust standard errors.
Panel A: Daily observations from 06/17/1991 to 06/14/1996
Dep. variable

0

1

2

3

4

5

6
R
2
Adj.
[t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio]
v
2
y
, = 3M
0.4940 0.0922 0.0963 -0.4211 1.8954 0.1035 4.6589 8.31%
[15.75] [ 7.17] [ 2.44] [-2.64] [ 3.63] [ 0.10] [ 2.94]
v
2
y

, = 2Y
0.9628 0.0623 -0.0475 0.4105 0.2968 4.1564 0.0655 0.45%
[11.72] [ 2.39] [-0.35] [ 1.21] [ 0.35] [ 1.40] [ 0.02]
v
2
y
, = 10Y
0.9216 0.0328 0.0335 0.2419 1.1689 2.0494 -0.0190 0.09%
[14.99] [ 1.45] [ 0.39] [ 0.83] [ 1.76] [ 0.92] [-0.01]
Panel B: Daily observations from 06/17/1996 to 06/15/2001
v
2
y

, = 3M
0.6193 -0.2056 -0.1275 -0.0197 3.5546 -5.4825 -1.3821 6.79%
[ 9.13] [-2.13] [-1.41] [-0.04] [ 1.54] [-3.27] [-0.43]
v
2
y
, = 2Y
0.8214 -0.1366 -0.1198 0.3876 1.1159 -1.5966 2.0268 2.06%
[12.27] [-2.13] [-1.03] [ 0.64] [ 0.79] [-1.46] [ 1.15]
v
2
y

, = 10Y
0.8286 -0.0618 -0.0103 0.0476 -0.0449 -1.4946 2.1762 0.63%
[15.75] [-1.09] [-0.13] [ 0.11] [-0.03] [-1.44] [ 1.50]
52
Table IV
Evidence on the Contemporaneous Volatility Spanning Condition for Monthly RV Series
OLS estimates for the regression v
2
y

(t, h) =
0
+

6
j=1

j
PC
j
(t, h) + (t, h) . The dependent variable is
the annualized overlapping monthly realized volatility (h = 1/12) for zero-coupon yields with maturity =
3M, 6M, 1Y, 2Y, 5Y, and 10Y. The explanatory variables, PC
j
, j = 1, . . . , 6, are the principal components
extracted from overlapping monthly averages of the six zero-coupon yields. The coecient t-ratios in square
brackets are based on (Newey-West) heteroskedasticy and autocorrelation robust standard errors.
Panel A: Overlapping monthly observations from 06/17/1991 to 06/15/2001
Dep. variable

0

1

2

3

4

5

6
R
2
Adj.
[t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio]
v
2
y
, = 3M
0.5577 0.0384 -0.0840 -0.1945 2.1229 -3.7441 4.7233 23.45%
[11.76] [ 1.38] [-1.17] [-1.04] [ 2.03] [-2.77] [ 3.02]
v
2
y

, = 6M
0.5389 0.0267 -0.0003 0.3246 1.7930 -1.5646 2.8996 15.48%
[13.14] [ 1.17] [-0.00] [ 1.61] [ 2.20] [-1.44] [ 2.16]
v
2
y

, = 1Y
0.6825 0.0246 0.0497 0.4604 1.4809 -0.6541 2.6565 12.14%
[14.54] [ 1.06] [ 0.87] [ 1.99] [ 1.69] [-0.56] [ 1.92]
v
2
y
, = 2Y
0.8936 0.0144 0.0171 0.5528 0.9117 0.5478 2.4063 5.84%
[14.91] [ 0.51] [ 0.29] [ 1.72] [ 0.96] [ 0.35] [ 1.41]
v
2
y

, = 5Y
0.9540 0.0147 0.0101 0.1401 1.0023 0.5942 1.8336 2.54%
[16.44] [ 0.53] [ 0.20] [ 0.43] [ 1.21] [ 0.40] [ 1.10]
v
2
y

, = 10Y
0.8680 -0.0045 0.0511 0.2256 0.7394 0.3736 3.4187 7.77%
[20.28] [-0.22] [ 1.18] [ 0.96] [ 1.09] [ 0.33] [ 2.48]
Panel B: Percent of residual variation explained by principal components
extracted from the regressions error terms
1st 2nd 3rd 4th 5th 6th
80.32% 13.74% 3.96% 1.11% 0.59% 0.27%
53
T
a
b
l
e
V
E
v
i
d
e
n
c
e
o
n
t
h
e
C
o
n
t
e
m
p
o
r
a
n
e
o
u
s
V
o
l
a
t
i
l
i
t
y
S
p
a
n
n
i
n
g
C
o
n
d
i
t
i
o
n
f
o
r
D
a
i
l
y
C
o
n
s
t
a
n
t
-
m
a
t
u
r
i
t
y
Y
i
e
l
d
s
O
L
S
e
s
t
i
m
a
t
e
s
f
o
r
t
h
e
r
e
g
r
e
s
s
i
o
n
v
2 y

(
t
,
h
)
=

0
+

9 j
=
1

j
P
C
j
(
t
,
h
)
+

(
t
,
h
)
.
T
h
e
d
e
p
e
n
d
e
n
t
v
a
r
i
a
b
l
e
i
s
t
h
e
a
n
n
u
a
l
i
z
e
d
d
a
i
l
y
r
e
a
l
i
z
e
d
v
o
l
a
t
i
l
i
t
y
(
h
=
1
/
2
5
2
)
f
o
r
z
e
r
o
-
c
o
u
p
o
n
y
i
e
l
d
s
w
i
t
h
m
a
t
u
r
i
t
y

=
3
M
,
6
M
,
1
Y
,
2
Y
,
5
Y
,
a
n
d
1
0
Y
.
T
h
e
e
x
p
l
a
n
a
t
o
r
y
v
a
r
i
a
b
l
e
s
,
P
C
j
,
j
=
1
,
.
.
.
,
9
,
a
r
e
t
h
e
p
r
i
n
c
i
p
a
l
c
o
m
p
o
n
e
n
t
s
e
x
t
r
a
c
t
e
d
f
r
o
m
t
h
e
n
i
n
e
c
o
n
s
t
a
n
t
-
m
a
t
u
r
i
t
y
d
a
i
l
y
z
e
r
o
-
c
o
u
p
o
n
y
i
e
l
d
s
,
w
i
t
h
m
a
t
u
r
i
t
i
e
s
o
f
3
M
,
6
M
,
1
Y
,
2
Y
,
5
Y
,
7
Y
,
1
0
Y
,
a
n
d
3
0
Y
.
T
h
e
c
o
e

c
i
e
n
t
t
-
r
a
t
i
o
s
i
n
s
q
u
a
r
e
b
r
a
c
k
e
t
s
a
r
e
b
a
s
e
d
o
n
(
N
e
w
e
y
-
W
e
s
t
)
h
e
t
e
r
o
s
k
e
d
a
s
t
i
c
y
a
n
d
a
u
t
o
c
o
r
r
e
l
a
t
i
o
n
r
o
b
u
s
t
s
t
a
n
d
a
r
d
e
r
r
o
r
s
.
P
a
n
e
l
A
:
D
a
i
l
y
o
b
s
e
r
v
a
t
i
o
n
s
f
r
o
m
0
6
/
1
7
/
1
9
9
1
t
o
0
6
/
1
5
/
2
0
0
1
D
e
p
.
v
a
r
i
a
b
l
e

9
R
2 A
d
j
.
[
t
-
r
a
t
i
o
]
[
t
-
r
a
t
i
o
]
[
t
-
r
a
t
i
o
]
[
t
-
r
a
t
i
o
]
[
t
-
r
a
t
i
o
]
[
t
-
r
a
t
i
o
]
[
t
-
r
a
t
i
o
]
[
t
-
r
a
t
i
o
]
[
t
-
r
a
t
i
o
]
[
t
-
r
a
t
i
o
]
v
2 y

=
3
M
0
.
5
5
6
3
0
.
0
1
6
5
-
0
.
0
6
9
1
-
0
.
1
3
7
1
0
.
2
1
7
6
-
2
.
5
9
5
0
3
.
7
3
9
1
-
0
.
2
1
7
5
-
1
.
9
3
3
9
6
.
6
5
6
9
5
.
1
7
%
[
1
2
.
9
3
]
[
0
.
6
1
]
[
-
1
.
8
9
]
[
-
1
.
0
5
]
[
0
.
7
5
]
[
-
2
.
7
4
]
[
3
.
2
9
]
[
-
0
.
2
8
]
[
-
1
.
7
8
]
[
1
.
4
6
]
v
2 y

=
6
M
0
.
5
3
8
1
0
.
0
1
5
9
-
0
.
0
0
6
2
0
.
2
1
4
4
-
0
.
0
0
3
1
-
2
.
0
2
1
2
1
.
8
8
6
3
1
.
3
5
8
9
-
1
.
6
9
2
0
5
.
3
7
9
6
2
.
9
8
%
[
1
4
.
5
4
]
[
0
.
7
5
]
[
-
0
.
2
1
]
[
1
.
5
5
]
[
-
0
.
0
1
]
[
-
3
.
0
5
]
[
2
.
2
9
]
[
2
.
0
0
]
[
-
1
.
5
8
]
[
1
.
7
2
]
v
2 y

=
1
Y
0
.
6
8
1
4
0
.
0
1
9
4
0
.
0
2
9
2
0
.
2
9
3
8
-
0
.
0
3
1
4
-
1
.
8
5
1
3
0
.
8
7
1
1
1
.
1
4
6
0
-
2
.
6
7
1
1
5
.
8
0
1
8
1
.
9
9
%
[
1
6
.
1
9
]
[
0
.
9
8
]
[
0
.
9
9
]
[
1
.
7
3
]
[
-
0
.
1
1
]
[
-
2
.
8
0
]
[
0
.
9
7
]
[
1
.
3
7
]
[
-
2
.
1
0
]
[
2
.
0
9
]
v
2 y

=
2
Y
0
.
8
9
1
4
0
.
0
0
6
8
0
.
0
1
1
7
0
.
3
6
1
4
-
0
.
0
4
5
2
-
1
.
2
3
9
1
0
.
1
6
2
0
1
.
4
6
5
2
-
2
.
7
8
6
9
6
.
5
5
8
1
0
.
9
2
%
[
1
6
.
1
1
]
[
0
.
3
1
]
[
0
.
3
4
]
[
1
.
4
8
]
[
-
0
.
1
4
]
[
-
1
.
6
4
]
[
0
.
1
4
]
[
1
.
2
1
]
[
-
1
.
6
6
]
[
2
.
2
6
]
v
2 y

=
5
Y
0
.
9
5
0
8
0
.
0
0
8
3
0
.
0
0
1
2
0
.
1
0
4
0
0
.
1
8
8
2
-
1
.
1
6
3
9
-
0
.
0
8
9
5
1
.
3
2
9
3
-
2
.
2
6
3
4
5
.
6
2
9
3
0
.
6
5
%
[
1
8
.
1
7
]
[
0
.
4
0
]
[
0
.
0
4
]
[
0
.
4
4
]
[
0
.
6
1
]
[
-
1
.
7
1
]
[
-
0
.
0
8
]
[
1
.
1
9
]
[
-
1
.
4
4
]
[
2
.
6
4
]
v
2 y

=
1
0
Y
0
.
8
7
4
5
0
.
0
0
8
1
0
.
0
4
4
5
0
.
0
7
6
4
-
0
.
4
1
8
4
-
1
.
4
0
4
7
0
.
3
7
2
2
1
.
6
6
1
7
-
1
.
9
6
1
1
7
.
0
6
3
9
1
.
5
4
%
[
2
1
.
8
3
]
[
0
.
4
6
]
[
1
.
5
2
]
[
0
.
4
7
]
[
-
1
.
3
3
]
[
-
2
.
3
6
]
[
0
.
4
2
]
[
1
.
8
4
]
[
-
1
.
4
6
]
[
3
.
1
5
]
P
a
n
e
l
B
:
P
e
r
c
e
n
t
o
f
t
h
e
r
e
s
i
d
u
a
l
v
a
r
i
a
t
i
o
n
e
x
p
l
a
i
n
e
d
b
y
p
r
i
n
c
i
p
a
l
c
o
m
p
o
n
e
n
t
s
e
x
t
r
a
c
t
e
d
f
r
o
m
t
h
e
r
e
g
r
e
s
s
i
o
n
s

e
r
r
o
r
t
e
r
m
s
1
s
t
2
n
d
3
r
d
4
t
h
5
t
h
6
t
h
7
8
.
0
8
%
1
0
.
5
4
%
7
.
2
3
%
2
.
4
3
%
1
.
0
9
%
0
.
6
4
%
54
Table VI
Evidence on the Contemporaneous Spanning Condition with Measurement Error
Instrumental variables estimates for the regression v
2
y

(t, h) =
0
+

6
j=1

j
PC
j
(t, h)+(t, h) . The dependent
variable is the annualized daily realized volatility (h = 1/252) for zero-coupon yields with maturity = 3M,
6M, 1Y, 2Y, 5Y, and 10Y. The explanatory variables, PC
j
, j = 1, . . . , 6, are the six principal components
extracted from the six daily averaged zero-coupon yields. The instruments are the principal components
extracted from the corresponding zero-coupon yields lagged by one day. The coecient t-ratios in square
brackets are based on (Newey-West) heteroskedasticy and autocorrelation robust standard errors.
Panel A: Daily observations from 06/17/1991 to 06/15/2001
Dep. variable

0

1

2

3

4

5

6
R
2
Adj.
[t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio] [t-ratio]
v
2
y

, = 3M
0.5564 0.0332 -0.0936 -0.2345 2.4317 -4.6381 2.9651 3.92%
[12.60] [ 1.26] [-1.39] [-1.21] [ 2.41] [-3.21] [ 2.73]
v
2
y
, = 6M
0.5383 0.0226 -0.0113 0.3366 2.0892 -2.0162 1.9132 1.88%
[14.11] [ 1.08] [-0.22] [ 1.62] [ 2.63] [-2.08] [ 1.86]
v
2
y
, = 1Y
0.6816 0.0211 0.0383 0.4888 1.8286 -1.0797 2.1436 1.25%
[15.83] [ 1.00] [ 0.78] [ 2.08] [ 2.31] [-1.11] [ 1.88]
v
2
y

, = 2Y
0.8918 0.0106 0.0059 0.5827 1.4688 -0.1130 2.0212 0.51%
[15.95] [ 0.41] [ 0.11] [ 1.89] [ 1.68] [-0.09] [ 1.35]
v
2
y
, = 5Y
0.9511 0.0101 -0.0032 0.1618 1.4610 0.2005 1.7120 0.21%
[17.96] [ 0.41] [-0.07] [ 0.55] [ 2.02] [ 0.17] [ 1.19]
v
2
y

, = 10Y
0.8644 -0.0072 0.0410 0.2260 1.1802 -0.2646 3.0316 0.52%
[21.13] [-0.37] [ 0.97] [ 1.00] [ 1.86] [-0.25] [ 2.34]
Panel B: Percent of residual variation explained by principal components
extracted from the regressions error terms
1st 2nd 3rd 4th 5th 6th
79.37% 10.25% 6.25% 2.39% 1.09% 0.64%
Table VII
Evidence on the Predictive Spanning Condition: Daily, Weekly, and Monthly Volatility
Forecasts
We report OLS estimates for v
2
y

(t + h, h) =
0
+
D
v
2
y

(t, h
D
) +
W
v
2
y

(t, h
W
) +
M
v
2
y

(t, h
M
) +

6
j=1

j
PC
j
(t, h
D
) + (t + h, h) , where the dependent variable is the annualized daily (h = 1/252), over-
lapping weekly (h = 1/52), and overlapping monthly (h = 1/12) realized volatility for zero-coupon yields
with maturity = 3M, 6M, 1Y, 2Y, 5Y, and 10Y. The right-hand side variables are the daily, weekly, and
monthly realized volatility (h
D
= 1/252, h
W
= 1/52, and h
M
= 1/12). PC
j
, j = 1, . . . , 6, are the principal
components extracted from the six daily averaged zero-coupon yields. The sample period is 06/17/1991 to
06/15/2001. The coecient t-ratios in square brackets are based on (Newey-West) heteroskedasticy and
autocorrelation robust standard errors.
55
Panel A: Estimation results for the constrained model. The set of explanatory variables contains the
principal components of the past yields, that is, the coecients
D
,
W
, and
M
are xed at zero.
v
2
y
, = 3M v
2
y
, = 6M v
2
y
, = 1Y
h 1 5 21 1 5 21 1 5 21

0
0.5570 0.5582 0.5606 0.5371 0.5385 0.5404 0.6825 0.6833 0.6841
[12.57] [11.72] [11.33] [14.16] [13.20] [13.08] [15.76] [14.70] [14.59]

1
0.0400 0.0442 0.0525 0.0280 0.0313 0.0395 0.0249 0.0283 0.0347
[ 1.52] [ 1.59] [ 2.22] [ 1.35] [ 1.42] [ 2.04] [ 1.17] [ 1.25] [ 1.62]

2
-0.0832 -0.0819 -0.0755 -0.0004 0.0003 0.0055 0.0489 0.0513 0.0573
[-1.17] [-1.05] [-0.99] [-0.01] [ 0.00] [ 0.09] [ 0.96] [ 0.90] [ 1.01]

3
-0.2260 -0.2242 -0.1541 0.3507 0.3441 0.3623 0.5152 0.5021 0.4960
[-1.17] [-1.16] [-0.91] [ 1.70] [ 1.66] [ 1.99] [ 2.19] [ 2.11] [ 2.33]

4
2.3311 2.2252 1.8070 2.0321 1.9445 1.7576 1.7757 1.7174 1.6606
[ 2.35] [ 2.18] [ 2.14] [ 2.59] [ 2.49] [ 2.60] [ 2.27] [ 2.15] [ 2.22]

5
-4.4249 -4.1794 -2.6315 -2.0016 -1.8397 -1.3122 -1.1193 -1.0415 -0.9379
[-3.26] [-3.06] [-2.61] [-2.19] [-1.82] [-1.60] [-1.21] [-1.01] [-1.14]

6
2.9977 2.7765 1.6889 1.8634 1.3717 0.6833 1.8862 1.3138 0.5145
[ 2.83] [ 2.48] [ 1.54] [ 1.89] [ 1.36] [ 0.68] [ 1.64] [ 1.23] [ 0.52]
R
2
Adj.
4.67% 12.97% 16.28% 2.50% 8.16% 15.29% 1.51% 5.89% 14.18%
v
2
y
, = 2Y v
2
y
, = 5Y v
2
y
, = 10Y
h 1 5 21 1 5 21 1 5 21

0
0.8951 0.8958 0.8969 0.9556 0.9571 0.9592 0.8701 0.8710 0.8724
[15.82] [14.89] [14.97] [17.89] [16.67] [16.53] [20.99] [20.01] [20.03]

1
0.0148 0.0179 0.0250 0.0158 0.0172 0.0188 -0.0047 -0.0037 -0.0015
[ 0.56] [ 0.63] [ 0.87] [ 0.62] [ 0.62] [ 0.66] [-0.24] [-0.17] [-0.07]

2
0.0155 0.0191 0.0292 0.0100 0.0127 0.0177 0.0498 0.0507 0.0507
[ 0.29] [ 0.32] [ 0.50] [ 0.22] [ 0.25] [ 0.36] [ 1.14] [ 1.10] [ 1.17]

3
0.6125 0.6132 0.6274 0.1935 0.2102 0.2438 0.2641 0.2618 0.2977
[ 1.97] [ 1.88] [ 2.09] [ 0.65] [ 0.66] [ 0.81] [ 1.16] [ 1.11] [ 1.37]

4
1.4046 1.2707 1.3589 1.3933 1.2442 1.3609 1.1000 0.9531 0.9215
[ 1.63] [ 1.48] [ 1.59] [ 1.94] [ 1.70] [ 1.83] [ 1.74] [ 1.54] [ 1.52]

5
-0.2263 0.0494 -0.1319 0.0553 0.3644 0.1716 -0.3213 -0.0345 -0.1790
[-0.19] [ 0.04] [-0.13] [ 0.05] [ 0.31] [ 0.17] [-0.32] [-0.03] [-0.22]

6
1.6289 1.2584 0.5114 1.3114 0.8321 -0.3055 2.5786 2.8332 0.8219
[ 1.06] [ 0.89] [ 0.41] [ 0.89] [ 0.57] [-0.23] [ 1.96] [ 2.10] [ 0.76]
R
2
Adj.
0.56% 2.73% 7.92% 0.23% 1.25% 3.95% 0.58% 2.98% 5.52%
56
Table VIIContinued
Panel B: Estimation results for the constrained model. The set of explanatory variables contains
volatility components only, that is, the coecients
j
, j = 1, . . . 6, are xed at zero.
v
2
y

, = 3M v
2
y

, = 6M v
2
y

, = 1Y
h 1 5 21 1 5 21 1 5 21

0
0.1983 0.2464 0.3433 0.2134 0.2422 0.3319 0.3090 0.3326 0.4148
[ 4.29] [ 5.20] [ 7.17] [ 4.87] [ 5.60] [ 7.28] [ 6.03] [ 6.22] [ 6.87]

D
0.0542 0.0465 0.0255 -0.0068 0.0237 0.0121 0.0238 0.0120 0.0066
[ 0.83] [ 2.54] [ 2.95] [-0.25] [ 1.79] [ 2.20] [ 1.07] [ 1.17] [ 1.57]

W
0.2394 0.1788 0.1345 0.2364 0.1578 0.1314 0.0961 0.0845 0.0823
[ 3.34] [ 3.32] [ 3.10] [ 2.73] [ 2.97] [ 3.14] [ 1.82] [ 2.00] [ 2.97]

M
0.3552 0.3388 0.2331 0.3779 0.3748 0.2478 0.4305 0.4204 0.3078
[ 2.60] [ 3.27] [ 3.10] [ 4.49] [ 4.54] [ 3.31] [ 5.70] [ 5.07] [ 4.07]
R
2
Adj.
8.40% 17.81% 17.73% 5.57% 15.53% 16.96% 3.44% 11.40% 16.53%
v
2
y

, = 2Y v
2
y

, = 5Y v
2
y

, = 10Y
h 1 5 21 1 5 21 1 5 21

0
0.4541 0.4848 0.6174 0.4483 0.4931 0.6374 0.5013 0.5293 0.6485
[ 6.73] [ 6.80] [ 7.79] [ 7.94] [ 8.02] [ 8.94] [ 8.02] [ 8.34] [ 8.98]

D
0.0292 0.0054 0.0058 0.0469 0.0155 0.0090 0.0156 0.0163 0.0043
[ 1.54] [ 0.62] [ 1.55] [ 1.44] [ 1.63] [ 2.17] [ 0.47] [ 2.05] [ 1.10]

W
0.0643 0.0927 0.0918 0.1315 0.1394 0.0985 0.1709 0.1186 0.0742
[ 1.38] [ 2.32] [ 3.44] [ 2.76] [ 4.28] [ 3.85] [ 2.19] [ 2.10] [ 3.45]

M
0.4018 0.3635 0.2163 0.3554 0.3334 0.2311 0.2390 0.2593 0.1799
[ 4.40] [ 3.92] [ 3.11] [ 4.81] [ 4.88] [ 4.16] [ 3.32] [ 3.41] [ 2.68]
R
2
Adj.
2.50% 8.82% 11.11% 3.76% 11.08% 12.82% 2.11% 6.70% 7.56%
57
Table VIIContinued
Panel C: Estimation results for the unconstrained model. The set of explanatory variables contains
both volatility components and past yields. The coecients
j
, j = 1, . . . 6, are not reported.
v
2
y

, = 3M v
2
y

, = 6M v
2
y

, = 1Y
h 1 5 21 1 5 21 1 5 21

0
0.2618 0.3152 0.4024 0.2563 0.2835 0.3768 0.3534 0.3754 0.4617
[ 5.83] [ 6.74] [ 8.43] [ 5.23] [ 5.96] [ 8.07] [ 6.07] [ 6.40] [ 7.43]

D
0.0478 0.0399 0.0206 -0.0094 0.0211 0.0096 0.0220 0.0103 0.0049
[ 0.74] [ 2.17] [ 2.59] [-0.34] [ 1.66] [ 1.95] [ 0.99] [ 1.00] [ 1.22]

W
0.2100 0.1506 0.1141 0.2136 0.1378 0.1126 0.0799 0.0693 0.0670
[ 2.90] [ 2.89] [ 2.90] [ 2.61] [ 2.84] [ 2.95] [ 1.60] [ 1.78] [ 2.65]

M
0.2694 0.2442 0.1477 0.3161 0.3154 0.1812 0.3806 0.3720 0.2537
[ 2.17] [ 2.59] [ 2.09] [ 3.70] [ 3.75] [ 2.35] [ 5.29] [ 4.44] [ 3.14]
R
2
Adj.
9.67% 22.48% 24.76% 6.29% 18.54% 24.95% 3.92% 13.86% 24.47%
v
2
y

, = 2Y v
2
y

, = 5Y v
2
y

, = 10Y
h 1 5 21 1 5 21 1 5 21

0
0.4788 0.5097 0.6457 0.4570 0.5005 0.6422 0.5369 0.5701 0.6733
[ 6.66] [ 6.88] [ 8.11] [ 7.88] [ 8.12] [ 9.16] [ 7.98] [ 8.38] [ 9.29]

D
0.0277 0.0040 0.0044 0.0459 0.0146 0.0080 0.0149 0.0157 0.0037
[ 1.45] [ 0.47] [ 1.16] [ 1.40] [ 1.52] [ 1.91] [ 0.46] [ 1.98] [ 0.96]

W
0.0539 0.0817 0.0798 0.1212 0.1290 0.0875 0.1590 0.1088 0.0652
[ 1.19] [ 2.04] [ 3.21] [ 2.63] [ 3.92] [ 3.69] [ 2.13] [ 2.01] [ 3.30]

M
0.3843 0.3453 0.1953 0.3538 0.3329 0.2352 0.2054 0.2191 0.1588
[ 4.54] [ 3.98] [ 2.87] [ 4.88] [ 5.05] [ 4.31] [ 2.71] [ 2.71] [ 2.28]
R
2
Adj.
2.67% 10.14% 16.50% 3.77% 11.67% 15.93% 2.24% 7.99% 11.14%
58
Table A.I
Sample Correlations Between Zero-coupon and Raw Bond Yields
For Treasury bills, the raw yields are zero-coupon yields with maturities closest to three months, six
months, and one year. For Treasury notes, the raw yields are coupon yields with maturities closest to two
years, ve years, and 10 years.
Panel A: Correlations between intradaily zero-coupon and raw yields
3M 6M 1Y 2Y 5Y 10Y
100.00% 100.00% 99.97% 99.98% 99.83% 99.51%
Panel B: Correlations between daily constant-maturity zero-coupon and par yields
3M 6M 1Y 2Y 3Y 5Y 7Y 10Y 30Y
100.00% 100.00% 100.00% 100.00% 99.98% 99.86% 99.74% 99.52% 96.96%
Table A.II
Sample Correlations Between RV Measures Constructed from Intradaily Zero-coupon Yields
and from Raw Bond Yields
For Treasury bills, the raw yields are zero-coupon yields with maturities closest to three months, six
months, and one year. For Treasury notes, the raw yields are coupon yields with maturities closest to two
years, ve years, and 10 years.
3M 6M 1Y 2Y 5Y 10Y
99.81% 99.93% 99.84% 99.99% 99.89% 96.48%
59
Table B.I
Percentage Number of Unique Yield Observations in the Intraday GovPX Sample
Panel A reports the percentage ratio between unique and total yield observations during the full June 17,
1991 to June 15, 2001 sample period. Panel B reports the same percentage ratios computed after discarding
trading days with three or more hours of inactivity.
Panel A: Full sample (June 17, 1991 to June 15, 2001)
3M 6M 1Y 2Y 5Y 10Y
80.75% 81.33% 91.05% 97.67% 94.22% 87.63%
Panel B: Trading days with three or more hours of inactivity are discarded
3M 6M 1Y 2Y 5Y 10Y
81.77% 82.54% 92.69% 99.10% 97.05% 91.20%
1995 2000
0
0.1
0.2
0.3
0.4
3
M
1995 2000
0
0.1
0.2
0.3
0.4
2
Y
1995 2000
0
0.1
0.2
0.3
0.4
6
M
1995 2000
0
0.1
0.2
0.3
0.4
5
Y
1995 2000
0
0.1
0.2
0.3
0.4
Date
1
Y
1995 2000
0
0.1
0.2
0.3
0.4
Date
1
0
Y
Figure 1. Realized volatility series. The plots depict the square root of the daily percentage realized
volatility measures (in percent per day), v
d,y
, for the three-month, six-month, one-year, two-year, ve-year,
and 10-year maturity yields (sample period: 06/17/1991 to 06/15/2001).
60
0 1 2 3 4 5 6 7 8 9 10
0.04
0.045
0.05
0.055
0.06
0.065
0.07
Maturity (years)
Y
i
e
l
d

v
o
l
a
t
i
l
i
t
y

(
p
e
r
c
e
n
t

p
e
r

d
a
y
)


std. dev. y
mean ( v
2
d,y
)
1/2
Figure 2. The term structure of yield volatility. The continuous line depicts the sample standard
deviation of daily changes in zero-coupon yields (in percent per day). The dashed line depicts the average
daily realized volatility v
d,y

(in percent per day). In both cases, the plots are constructed using yields with
three-month, six-month, one-year, two-year, ve-year, and 10-year maturities (sample period: 06/17/1991
to 06/15/2001).
10 20 30 40 50
0
0.5
1
3
M
10 20 30 40 50
0
0.5
1
2
Y
10 20 30 40 50
0
0.5
1
6
M
10 20 30 40 50
0
0.5
1
5
Y
10 20 30 40 50
0
0.5
1
Lags
1
Y
10 20 30 40 50
0
0.5
1
Lags
1
0
Y
Figure 3. Sample autocorrelations for the realized volatility series. The continuous line plots the
sample autocorrelations for daily logarithmic realized volatility. The dotted line depicts the minimum-
distance estimates of the hyperbolic decay rate, c Lag
2d1
. Sample period: 06/17/1991 to 06/15/2001.
61
1992 1994 1996 1998 2000
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Date
3
M

y
i
e
l
d

v
o
l
a
t
i
l
i
t
y

(
p
e
r
c
e
n
t

p
e
r

d
a
y
)


v
d,y
SNP vol
Figure 4. U.S. three-month Treasury bill yield volatility. The plots depict daily realized volatility
(i.e., v
d,y

, in percent per day) versus one-day-ahead volatility forecasts based on the SNP model (i.e.,

V (y

|x; ), in percent per day), where = 3M. Sample period: 06/17/1991 to 06/15/2001.
1995 2000
0
0.1
0.2
0.3
0.4
3
M
1995 2000
0
0.1
0.2
0.3
0.4
2
Y
1995 2000
0
0.1
0.2
0.3
0.4
6
M
1995 2000
0
0.1
0.2
0.3
0.4
5
Y
1995 2000
0
0.1
0.2
0.3
0.4
Date
1
Y
1995 2000
0
0.1
0.2
0.3
0.4
Date
1
0
Y
Figure 5. Volatility forecasts and realized volatility series. The plots depict one-day-ahead HAR-
RV forecasts, v
d,y

, and realized volatility series, v

d,y

(in percent per day). Sample period: 07/17/1991 to

03/27/2001.
62
Notes
1
On occasion, earlier studies employ direct low-frequency volatility measures obtained by cumulating
higher frequency squared returns. These are viewed as intuitive model-free measures, akin to historical
sample variance statistics, and there is no articulation of a theoretical foundation. Such studies include
French, Schwert, and Stambaugh (1987), Hsieh (1991), Poterba and Summers (1986), and Schwert (1989).
2
In particular, the evidence is not inconsistent with the USV class of models of Collin-Dufresne and
Goldstein (2002), CDGJ, and Joslin (2006).
3
As noted, an exception is the USV class of models of Casassus, Collin-Dufresne, and Goldstein (2005),
Collin-Dufresne and Goldstein (2002), CDGJ, Joslin (2006), and Trolle and Schwartz (2007a), in which the
full rank condition on the B matrix in equation (5) is violated and the inversion argument behind equation
(6) does not apply. A related class of models is explored in Kimmel (2004).
4
There is a caveat, akin to the invertibility condition of standard ane models. It is possible that the
quadratic models naturally embed a form of the USV restriction. This is not readily determined by analytic
means, but we have conrmed that the model estimated by Ahn et al. (2003) induces a near perfect linkage
between the concurrent yield cross-section and yield volatility, as should be the case in an ane model. We
accomplish this by simulating the model and then ex-post regressing the associated quadratic yield variation
on the concurrent yield curve shifts. These results are available from the authors upon request.
5
Measurement errors in the yield quadratic variation resulting from microstructure noise will have the
same eect, provided that such error is uncorrelated across days. We introduce our measures for the
quadratic yield variation in Section I.E, while in Section II.C we provide a detailed account of our procedure
to adapt this measure to noisy yield data and we discuss the relevant theoretical and empirical properties
of our estimates.
6
As suggested by a referee, an alternative approach is to use bipower variation, which is robust to the
presence of jumps (Barndor-Nielsen and Shephard (2004b)) to separate the diusive and jump quadratic-
variation components as implemented in, for example, Andersen et al. (2007). This analysis warrants a
63
separate treatment, which we plan to pursue in future work.
7
Fleming (2003) points out that the GovPX raw data les need to be cleaned due to some interdealer
brokers posting errors that are not ltered out by GovPX. Hence, prior to our analysis we implement the
error correction procedures recommended by Fleming in the Appendix of his paper.
8
Most scheduled macroeconomic announcements take place at 8:30am or 10am. The statements from
the Federal Open Market Committee meetings are typically released around 2:15pm. Further, the window
includes the Federal Reservess customary intervention times (11:30am before 1997, 10:30am from 1997 to
1999, and 9:30am as of 1999) and the Treasury auctions announcement times (1:30pm to 2pm).
9
Although there are no xed trading hours for Treasuries, the Bond Market Association (BMA) makes
recommendations regarding holiday closures and early closing times. Specically, the BMA often recom-
mends that the market close early (usually at 2pm) before a holiday, which typically results in low trading
activity for those days. As a robustness check, we also consider eliminating days during which we cannot
nd any trading activity for a period longer than either two or four hours. Neither approach changes the
conclusions discussed below.
10
The MA coecient estimates are as follows. For the three-month series: 0.059; six-month series:
0.060; one-year series: 0.043; two-year series: 0.006; ve-year series: 0.003; and 10-year series: 0.051.
As a robustness check, we also apply the realized volatility estimator proposed in Hansen and Lunde (2006),
which is designed to accommodate the eects of market-microstructure noise. This alternative approach
produces results similar to those reported below.
11
Specically, we use the tcm3m, tcm6m, tcm1y, tcm2y, tcm3y, tcm5y, tcm7y, tcm10y, and tcm30y series
from the web site http://federalreserve.gov/releases/h15/data.htm.
12
There has been a dramatic steepening of the yield curve, said Michael Boss, a bond trader with IBJ
Lanston Futures in Chicago to CNN, referring to the growing dierential between yields on long-term and
short-term Treasury issues. Fear is the overriding factor hereit has just been really ugly.
13
Economists fear that downward momentum could feed on itself, wrote Christopher Georges in the
June 5, 1995 issue of the Wall Street Journal. Once it gets going, the downward spiral is hard to stop,
64
said Sung Won Sohn, chief economist at NorWest Corp. in Minneapolis. The correction could go on for
longer than anticipated.
14
A paradigm shift on Wall Street today. It all started with a report showing that a stunning number
of people got jobs last month, the fastest improvement in the employment picture in 12 years. That started
the stronger economy neon sign ashing, and gone went hopes of lower interest rates. The bond market
fell o a cli... with the 30 year Treasury falling a heart-stopping 3 points, reported David Brancaccio at
CNN.
15
We use the H.15 series of daily three-month T-bill bank discount rates from the Federal Reserve Bank
of St. Louis web site, http://research.stlouisfed.org/fred2/. Prior to analysis, we convert the H.15 bank
discount rate data into continuously compounded yields.
16
We report Newey-West standard errors that are robust to the presence of heteroskedasticity and serial
correlation in the regression residuals. We use 20 lags in the computation of standard errors in all regressions
that rely on daily data, 30 lags for the overlapping weekly data, and 40 lags for the overlapping monthly
data. We have conrmed that the point estimates of the t-ratios are robust to the choice of a higher number
of lags.
17
We conrm that such sign changes are not due to a rotation of the yields principal components. That
is, we conrm that in each subsample the rst three yields principal components correlate positively with
the level, slope, and curvature of the term structure.
18
Bliss and Smith (1998) provide evidence that the elasticity of interest rate volatility is subject to
structural shifts, while Boudoukh et al. (2007) argue that the volatility of interest rates is increasing in the
level of interest rates mostly for sharply upward-sloping term structures. Generally, in regime-switching ane
models, shorter sample periods should correspond more closely to the presence of one dominant (persistent)
regime, which would result in a better in-sample t. In order to informally test for this possibility, we split
the full sample into ve distinct two-year periods and run the regressions for each subsample. Although the
explanatory power is considerably higher for one of these samples, it is still extremely low in most cases.
Moreover, the signs of the coecients associated with dierent principal components are highly unstable
65
across the subsamples, indicating problems with overtting. In summary, the spanning condition appears
to fail uniformly.
19
All these ndings apply for the weekly frequency as well. The detailed results are available from the
authors upon request.
20
Noise in the left-hand-side realized yield volatility measures would not aect consistency of the OLS
estimates but would lower the R
2
of the regression so that it would fall below unityby some undetermined
amounteven if volatility is spanned by the yields. While this is, to some extent, inevitable, we conrm in
unreported simulations from an ane A
1
(3) model that measurement error of plausible magnitude cannot
lower the regression R
2
to the levels that we nd in our regressions.
21
These tests may use either the zero-coupon yields extracted at the end of the day (with the potential
advantage of providing more current information regarding future yield changes) or the past daily average of
the 10-minute yields (likely measured with less error). We report results based on the latter approach, but
we have conrmed that near identical results obtain when using only yield observations sampled towards
the end of the trading day (2pm to 5pm). This nding provides further assurance that idiosyncratic yield
measurement errors are not contaminating our results.
22
Additional diagnostics (available from the authors upon request) conrm that the quality of t is
reasonable and further show that, if anything, the HAR-RV forecasts are more accurate than the EGARCH-
type forecasts introduced in Section II.C.2.
23
Previous studies along these lines include Ang, Piazzesi, and Wei (2006), Bikbov and Chernov (2005),
Dai and Philippon (2004), Diebold, Piazzesi, and Rudebusch (2005), and Piazzesi (2005).
24
For the GovPX data set, N = 6 and
k
= 3M, 6M, 1Y, 2Y, 5Y, and 10Y. For the data set of daily
constant maturity Treasuries N = 9 and
k
= 3M, 6M, 1Y, 2Y, 3Y, 5Y, 7Y, 10Y, and 30Y.