Fractional Integration and Cointegration:

Testing the Term Structure of Interest Rates

Marco R. Barassi Dayong Zhang

Department of Economics Southwestern University
University of Birmingham of Finance and Economics
United Kingdom Chengdu, China

Abstract
The expectation hypothesis suggests there exists long run equilibrium of interest rate
term structure. Two theoretical approaches proposed by Campbell and Shiller (1987) and
Hall el al. (1992) suggest that the term spread of long-term and short-term interest rates
should be a stationary I(0) process. However, an empirically non stationary term spread
or rejection of cointegration between long and short interest rates, in the traditional sense
need not to be considered against the simple theoretical model. It is likely that the
dichotomy between I(1) or I(0) and/or integer values of cointegration are environments
which are too restrictive to model the term structure. In this paper, we evaluate and
apply some recent techniques on testing fractional integration and propose the use of a
residual based approach which uses the Exact Local Whittle Estimator. The method is
then used to investigate the term structure in the UK and the US.

JEL: C22; E43

Keywords: Term Structure, Long Memory, Fractional Integration, Fractional Cointegra-
tion, Local Whittle Estimation.

Corresponding author: Marco R. Barassi, Department of Economics, University of

Birmingham, Edgbaston, Birmingham B15 2TT, West Midlands, United Kingdom. Email:
[email protected].

1
1 Introduction
The expectation hypothesis suggests there is long run equilibrium between long-term and short-
term interest rates. Specifically, current long-term interest rates should be equal to the agents’
expectations about current and future short-term interest rates, plus some liquidity premium.
Campbell and Shiller (1987) suggested a present value type of relationship between the short-
term and long-term interest rates: the latter is a function of present discounted value of the
expected future short-term interest rate. They also tested this hypothesis using 20 years US
treasury bonds yield and one month T-Bill rate as an example. The results seems supportive
of a stationary term spread, however, the Dickey-Fuller test on the spread only rejects the null
of non-stationary marginally at 10% for the period between 1959-1978.
The term structure of interest rates contains critical information about market expectations
on interest rates, future inflation, and it also plays a role in transmission mechanism between
the real sector and money market (Mankiw and Summers, 1984).
After Campbell and Shiller’s (1987) seminal work, the simple equilibrium relation between
long and short-term interest rates has been investigated by many researchers using more and
more sophisticated techniques. Among others, Camarero and Tamarit (2002) modelled the term
structure in Spain using cointegration analysis allowing for structural changes and parameter
instability. Seo (2003) looked at non-linear mean reversion in the term structure with a thresh-
old vector correction model finding significant threshold effects. Kuo and Enders (2004) suggest
that the dynamic adjustment of term spread may incorporate an inherently asymmetric process
of inflation and other Business cycle variables, thus performing differently with positive and
negative deviations from equilibrium. They employed a MTAR model to test this hypothesis
for the Japanese interest rates.
With the help of some recently developed testing procedures, we intend to approach the
problem from a different angle. The simple long run equilibrium is theoretically attractive, how-
ever, as we mentioned, the term structure is potentially affected by many unexpected changes
such as monetary policies, inflation and other macroeconomic factors. Moreover, investor senti-
ment can affect the liquidity premium. Investors might not have rational or perfect expectations
especially in the relatively short/medium term. In theis sense, testing for stationarity of the
term spread or for cointegration between short and long term interest rates might fail to reveal
some important bits of information.
It is well known for example that standard unit root and cointegration tests may provide the
wrong results in the presence of structural breaks in either the series or in their cointegrating
relationship. Also, the tests used may have low power near the alternative, and moreover, they
cannot provide information about the level of persistence on the series. Here we consider the
possibility that the term spread is actually a long memory process and that is the reason why
standard unit root and cointegration analysis may fail to offer empirical support for the simple

2
theory of the term structure. The basic idea is that equilibrium in the market does exist,
however, due to unexpected frictions or asymmetric effects, the term spread does not need
to be stationary everywhere, however, it should revert to its mean in the long run. An early
characterisation of the term structure as a long memory process is provided by Backus and
Din (1993) who proposed it for an expectation based model of the term structure. Following
the argument put forward by Granger (1980), fractional integratedness of the term structure
may arise from the fractional integration of inflation which is the result of an aggregation of
individual beliefs.
From an empirical point of view, fractional integratedness of the term structure is an ap-
pealing way to model interest rates and the term stucture as it combines high persistence with
mean reversion in the long run, still maintaining the possiblity of a common stochastic term in
a multivariate setting. The use of fractional integration may also be motivated by the presence
of occasional breaks in a series which otherwise would be only weakly autocorrelated. This can
be very appealing in the case of interest rates as changes in the official discount rate from the
monetary authorities, although infrequent, may still cause a break in the interest rates series.
Granger and Hyung (2004) show that fractional integration and infrequent breaks may be vir-
tually impossible to distinguish from each other and adopting fractional integration modelling
strategies will produce good forecasts.
For these reasons, here we use a residual based test for fractional cointegration based on the
Exact Local Whittle (ELW) estimator of the fractional integration parameter and its estimated
variation (FELW) (Phillips and Shimotsu 2006) to test for the hypothesis that the term spread
in the US and UK is a long memory process which though highly persistent may revert to its
mean in the long run. In the past a similar approach has also been used by among others Cheung
and Lai (1993) who suggest using the Geweke and Porter-Hudak (1983) technique (GPH) on
OLS residuals to analyse fractional cointegration of real exchange rates. Also, Caporale and
Gil-Alana (2004) test the present value model estimating the fractional integration parameter
of OLS residuals using the Local Whittle (LW) estimator of Robinson (1995).
We will first derive the critical values of the ELW residual based test for fractional coin-
tegration under the null that the series are not cointegrated (i.e.the residuals from the OLS
regression are I(d = 1)). We then investigate the statistical properties of the ELW based tests
by means of a monte carlo simulation showing that the ELW and FELW estimators are superior
to rival techniques such as the tests proposed by Geweke and Porter-Hudak (GPH) (1987) and
the LW estimator of Robinson (1995). We briefly introduce the theoretical background and
introduce the relevant econometric techniques in the next two sections. Section 4 provides the
simulation results and section 5 discusses the empirical results. A summary concludes.

3
2 The term structure of interest rates
The linear expectations theory of the term structure which is used to represent the expectations
hypothesis (EH) as formulated by Campbell and Shiller (1987,1991) defines the equilibrium
relationship between a k-period interest rate R(k)t and an l-period interest rate R(l)t , where
k > l1 . Here R(k)t can be seen as the interest rate on a longer-term instrument relative to
a shorter-term rate R(l)t . The expectation hypothesis simply states that the expected return
from investing in a k-period rate will equal the return from investing in l-period rates up to
k − l future periods plus a constant risk premium c. This simple relationship can be expressed
as:
δ−1
X
R(k)t = (1/δ) Et [R(l)t+li ] + c. (1)
i=0

where δ = k/l, Et indicates expectation at time t and c is the constant risk premium. Basically,
the long-term rate is expressed as a weighted average between the current and expected future
short term rate R(l)t , plus a risk premium. Subtracting R(l)t from both sides in the above
equation we obtain an expression for the term spread between the k-period and the l-period
rates S(k, l)t :
δ−1
XX i
S(k, l)t = [R(k)t − R(l)t ] = (1/δ) Et [∆R(l)t+jl ] + c (2)
i=0 j=1

The equation above provides some interesting information on the possible characteristics of the
term spread. If it is true that interest rate series are non stationary I(1), the term spread of in-
terest rates should be a stationary I(0) process, and we should have a cointegrating relationship
(1,-1) linking the long and short-term interest rates. Thus testing term structure equilibrium,
or equivalently testing a present value model can be done by looking at the stationarity of
S(k, l)t or whether we have a cointegration relation between R(k)t and R(l)t .
However, as we mentioned above, frictions in the markets may affect expectations, the risk
premium may also be affected by a change in investors’ sentiment, and also occasional structural
breaks caused by intervention of the monetary authority may occur affecting the persistence of
the term spread. In terms of the empirical analysis, these events may cause failure to support
the hypothesis of an I(0) stationary term structure in standard stationarity and cointegration
tests. Still, we claim that failure in these types of test should not be taken as an argument
against the simple term structure model. In this paper we abandon the dichotomy between
I(1) and I(0) processes and consider the case where two non-stationary I(1) processes may be
cointegrated as a I(d) process where 0 ≤ d < 1. This is to say that the term structure can be
1
Hall et al. (1992) propose a more general relationship between yields with different maturities. They assume
that R(k, t) is a continuously compounded yield to maturity of a k period pure discount bond, and F (j, t) is the
forward rate, defined as the rate of return from contracting at time t to buy a one period pure discount bond
that matures at time t + k.

4
modelled as a mean reverting long memory process which may have finite or infinite variance
depending on the value of d. An account of long memory models and their estimation follows.

3 Econometric Methodology
3.1 Long Memory Models
A stationary time series {yt } has long memory (see Granger and Yojeux (1980)) if there is a non
zero d ∈ (−0.5, 0.5) such that its spectral density obeys a power law f (λ) ∼ kλ−2d as λ → 0+ .
Clearly, as λ → 0, f (λ) → ∞, if d > 0 or f (λ) → 0, if d < 0. In the case where d = 0, {yt } will
have short memory and (0 < f (0) < ∞) such as in the case of all the stationary and invertible
ARMA processes. Hosking (1981) established the relationship between long memory models
and fractional differencing, proposing a class of models denominated fractional ARIMA, where
the degree of differencing can be any real number. The simplest of these is the fractionally
integrated noise or ARIM A(0, d, 0) defined for d > −1, that is:

∆d yt = t ;  ∼ iid(0, σ 2 ), and, ∆d = (1 − L)d . (3)

To obtain an ARIM A(0, d, 0) for d that lies outside the range (−0.5, 0.5), it is sufficient to
start with an ARIM A(0, d, ¯ 0), with d¯ ∈ (−.05, 0.5), such that d − d¯ is an integer. If d > 0.5,
we need to integrate d − d¯ times, if d < 0.5, we difference d − d¯ times. The result will be an
ARIM A(0, d, 0) which however will not be stationary if d ≥ 0.5.
Using the spectral representation, we find that {yt } has spectral density,
−2d
σ2
f (λ) = 2π
1 − eiλ (4)
σ2 −2d
= 2π
|2 sin(λ/2)| , (5)

where σ 2 is the variance of t . For λ close to 0, sin λ ∼ λ, and

σ 2 −2d
f (λ) ∼|λ| , λ → 0. (6)
2π
Clearly, if d > 0 then limλ→0 f (λ) = ∞, and f (·) has a pole at zero frequency, on the other
hand, for d < 0, the limλ→0 f (λ) = 0, such that f (·) has a zero at zero frequency. Note that
negative d can arise in practice as a result of over-differencing.
The more general fractional ARIMA (p,d,q)can be written as:

φ(L)∆d yt = θ(L)t , (7)

where t is as before a zero mean white noise process, d ∈ (−0.5, 0.5), and φ(L) and θ(L) are
polynomial in the lag operator or order p and q respectively. The ARIMA(p,d,q) can also be
written as
φ(L)yt = θ(L)(∆−d t ), (8)

5
showing that the ARIMA(p,d,q) can be thought of as an ARMA(p,q) driven by a fractionally
integrated ARIMA(0,d,0) noise, ∆−d t . The spectral density for the ARIMA(p,d,q) is:

σ2 2 −2d
f (λ) = θ(e−iλ )/φ(e−iλ ) 1 − e−iλ (9)
2π
and as λ → 0+
σ2
f (λ) = |θ(1)/φ(1)|2 |λ|−2d . (10)
2π
Again, if d 6= 0, then yt is a long memory process and it can be shown that its autocovariances
obey the power law cr ∼ kr2d−1 as r → ∞.
Alternatively we could write the spectral density of the fractional ARIMA(p,d,q) as:

σ2 −2d
f (λ) = 1 − e−iλ g(λ) (11)
2π
where g(λ) is the spectral density of an ARMA(p,q).
The fractional ARIMA(p,d,q) is particularly convenient as it can be used to describe both
the long and the short term behaviour of the yt . In fact d determines the long term correlations
and the behaviour of the spectral density near zero frequency and the AR and MA parameters
describe the short term correlations and shape the spectral density for frequencies not near
zero, independently from d.
The ARMA parameters as well as d are unknown and must be estimated. Assuming that p
and q are known and that {t } are Gaussian, we can use maximum likelihood estimation (MLE)
to estimate the ARMA parameters as well as d. However the MLE of such type of models has
some drawbacks, namely the fact that it is computationally intensive to evaluate the likelihood
function as O(n3 ) operations are necessary. To avoid the intense computation involved in using
maximum likelihood, Fox and Taqqu (1986) suggest using the Whittle approximation to the
likelihood function, that is:
n/2
X
−2 log l(Θ) ≈ [log fΘ (ωj ) + Ij /fΘ (ωj )], (12)
j=1

where Θ is the vector of ARMA parameters plus d, and fΘ (ω) is the spectral density of the
ARIMA(p,d,q) process. The Whittle estimator (WE) is obtained by minimisation with respect
to Θ. Fox and Taqqu (1986) showed that the Whittle estimator of Θ, say Θ̂W , is asymptotically
equivalent to the exact MLE assuming that d > 0, the model is correct, and the order of p and
q is known. However the advantage of WE is that the Whittle approximation to the likelihood
function can be evaluated in O(n log n) rather than in O(n3 ) operations which are necessary
for the MLE to be computed, with a gain in terms of computational efficiency.
In practice, however, p and q are unknown, and as a result both the ML and W estimators
of d will be asymptotically biased. The same problem will arise if g(λ) is not ARMA but

6
it is assumed that it is. The way to circumvent the problem is to assume that in equation
(9) the spectral density of the ARMA part g(λ) is continuous at λ = 0 and that there exist
some positive constants C1 and C2 such that C1 < g(λ) < C2 for all λ ∈ [0, π], such that
g(·) corresponds to a stationary and invertible ARMA. In this way the model becomes semi-
parametric in that the long memory part of the model is parametrically described by d while
the short memory part g(λ) does not obey any parametric model.
Geweke and Porter-Hudak (1983)(GPH hereafter) use this framework and propose an es-
timator of d based on the OLS regression of the log periodogram on the log frequency. They
start by assuming that asymptotically, the first k normalised periodogram ordinates {Ij /fj }kj=1
are identically independently distributed. If this holds then E[{Ij /fj }] = −C = −0.577216,
that is, minus the Euler constant and its variance var[{Ij /fj }] = π 2 /6. This implies that
j = log(Ij /fj ) + C are i.i.d.(0, π 2 /6), for j = 1, . . . , m.
Using equation (9),

log Ij = log fj + log(Ij /fj )

= log fj − C + j (13)
= [log((σ 2 /2π)g(λ)) − C] − 2d log 1 − e−iωj + j .

when k is small relative to n, at λ = 0, continuity of g(λ) implies that g(ω) is constant at

g0 , such that we have that:

log Ij = [log((σ 2 /2π)g0 ) − C] − 2d log 1 − e−iωj + j , j = 1, . . . , m. (14)

GPH use this result to estimate d using least squares from the regression of {log Ij }m j=1 on
−iωj m
{−2 log |1 − e |}j=1 . The usefulness of the regression of log periodogram on the log frequency
rests on the fact that there is no need to specify a parametric model for g(λ). However if we
set the number of frequencies m too small the variance of dˆGP H will be extremely high, and if
we take m too large the assumption that g(ω) is approximately constant will break down and
dˆGP H will be biased.
However, this method has noticeable limits, discussed among others by Banerjee and Urga
(2005). Robinson (1995) proposed a semiparametric Gaussian estimator for d, which, unlike
GPH, is not defined in closed form but which has better properties than the GPH in that it is
asymptotically more efficient, and much weaker assumptions on gaussianity may be imposed.
This is known as the Local Whittle (LW) estimator, and it is developed under the assumption
that yt is stationary and its spectral density behaves like Gλ−2d as λ → 0+ . The frequency
domain Gaussian likelihood in the vicinity of the origin is:
m
λ2d
j
Qm (G, d) = m−1 [log(Gλ−2d
X
j )+ I(λj )], (15)
j=1 G

7
where as before m < n is an integer controlling the number of frequencies included in the local
likelihood. Estimates of G and d are then obtained through the minimisation of Qm (G, d) such
that
ˆ = arg min
(Ĝ, d) Qm (G, d), (16)
G∈(0,∞),d∈[∆1 ,∆2 ]

with ∆1 and ∆2 are numbers such that −∞ < ∆1 < ∆2 < 1/2. Concentrating the likelihood
with respect to G, it is found that dˆ satisfies
m m
1 X
Ĝ(d) = m−1 λ2d
X
arg mind∈[∆1 ,∆2 ] log Ĝ(d) − 2d log λj , j I(λj ). (17)
m j=1 j=1

Under appropriate assumptions and conditions, and for d0 ∈ (−1/2, 1/2), Robinson(1995)
√
and Shimotsu and Phillips (2006) show that m(dˆ − d0 ) → N (0, 1/4).
Phillips and Shimostu (2004), among others, show that when d > 1/2 the GPH and LW
estimators exhibit non standard behaviour. Although they are consistent for d ∈ (1/2, 1] and
asymtotically normal for d ∈ (1/2, 3/4), the LW has non normal asymptotic distribution for
d ∈ [3/4, 1], and d > 1, and also in that case both LP and LW converge to 1 in probability and
are inconsistent.
The solution to the problem is provided by Shimotsu and Phillips (2004, 2006) who devel-
oped the Exact Local Whittle (ELW) estimator and its variants, whose asymptotics are based
on the exact frequency domain (or its estimate which will give rise to FELW estimator) of the
data generating process which is obtained from the minimisation of the objective function:
m
λ2d
j
Q∗m (G, d) = m−1 [log(Gλj−2d ) +
X
I d (λj )]. (18)
j=1 G (1−L) y

The ELW is computationally more demanding than LW but is shown to be consistent and
asymptotically normally distributed for any value of d and therefore is valid under a wider
range of cases.
Shimotsu and Phillips (2006), discuss the importance of the choice of m, explaining that
as in the case of GPH, LW, and ELW estimation, m has to grow fast for dˆ to be consistent,
but also that a too large value of m may induce a bias to the estimator from the short run
dynamics. A rule of thumb suggests, that within m = nα , we should choose a value of α in the
interval [0.5, 0.65] (see Shimotsu and Phillips 2006). We bear this in mind when designing our
experiment

3.2 Residual based tests of Fractional Cointegration

A cointegration relationship CI(d, b) is defined in a system Yt if each of the components of the
system is integrated of order d and a linear transformation of the system with θ0 Yt is integrated
of order (d − b), where d > b > 0. In standard cointegration analysis, the parameters d and b

8
are usually taken as integers such as CI(1, 1), which suggests that a system of individually I(1)
series is cointegrated with one or more linear combinations of the variables in the system being
I(0). Starting from this definition, the concept of fractional cointegration has been introduced
to account for the cases where d and b are not necessarily integer values. Fractional cointegration
has been analysed extensively in the recent literature such as Marinucci and Robinson (2001),
Robinson and Yajima (2002), Davidson (2002) among others. In this paper, we follow the
approach of Cheung and Lai (1993) and use a simple residual based approach for fractional
cointegration in three simple steps. We first estimate the order of integration of the individual
interest rate series in our system, second, we perform an OLS regression between long and short
term interest rates thereby obtaining the OLS residuals which will in turn be tested for their
order of (possibly fractional) integration.
Cheung and Lai (1993) used just the GPH-log periodogram method to test the null of no
cointegration d = 1 against the alternative of a fractionally cointegrated system (i.e. 0 ≤ d < 1).
We follow their method and consider also the simple LW test, and propose the use of ELW
and FELW first using simulation methods to compare the properties of all these techniques.
The advantage of using ELW and FELW is that unlike the LW and the GPH estimators, they
are consistent even when d = 1, indeed under what will be the null hypothesis. Note that
in testing for fractional cointegration using the residuals from an OLS regression, the critical
values to be used are non standard ones and those derived from the standard distribution
cannot be used directly to evaluate the estimate of d (see Engle and Granger 1987 for the same
argument in standard cointegration analysis). This is because the residuals are not observed
but are obtained from minimising the residual variance of the cointegrating regression, and the
residuals series tends to bias toward being stationary, therefore causing over-rejection of the null
if using standard Whittle estimator’s critical values. Thus, in the same spirit as Cheung and
Lai (1993) we derive the critical values for the various test statistics for the different bandwidths
we are going to use and also evaluate the performance of the techniques described above by
means of a simple Monte Carlo experiment.

4 The Simulation
We use 10,000 replications to derive the critical values obtained by generating a bivariate system
of two independent I(1) processes of size n = 312, which is the sample size of our dataset, using
the following equation:
y1t + y2t = u1t (1 − L)u1t = 1t (19)
here the 1t are i.i.d. normal innovations, and L is the lag operator.
All the estimation techniques are sensitive to the selection of parameter α, which will affect
m as m = nα . Following the recommendation of Shimotsu and Phillips (2006) we shall consider

9
the cases of α = 0.55, 0.575, 0.60. The obtained critical values for GPH, LW, ELW and FELW
tests are presented in Tables 1-4.
Table 1 in particular contains the critical values for the GPH test. We report that our
critical values are very similar to those of Cheung and Lai (1993), with the different sample size
used not affecting the critical values that much. All GPH, LW, ELW, and FELW as expected
provide a distribution which is not centered around zero and has negative mean and skewness.
This is consistent with the fact that the critical values have been derived using the residuals
from an OLS regression and support the claim that if we used standard critical values the tests
would tend to reject the null of no cointegration too often.
Another important point to be mentioned is that we assume that the system contains I(1)
series, where the order of d in empirical applications might not necessarily have an integer value.
If that is the case, the critical values might be somehow not the appropriate ones. However, for
our requirement of analysing the interest rate term structure, as the interest rates in question
are indeed I(1) series, the critical values are appropriate.
In this paper, we are also interested in comparing the power of the various techniques against
the alternative of fractional integration for different values of d . For our simulations we use
the same data generating process as Cheung and Lai (1993), and generate two series y1t and
y2t as follows:
y1t + 2y2t = u2t (1 − L)d u2t = 2t . (20)
Again the 2t are i.i.d. normal innovations, and L is as usual the lag operator. The values of d
which are considered go from 0.05 to 0.95 with increments of 0.1, and the power of the tests is
evaluated at the 0.05 and 0.1 significance levels using exact critical values.
The results are presented on Tables 5-8, and show that the simple LW test outperforms
the GPH displaying a greater power in all circumstances. However, the simple LW is itself
outperformed by both the ELW and FELW estimators which, in the scenario designed by our
simulation, offer a very similar performance to each other. Overall, as expected, the power
of all the tests considered increases with the sample size n, and for each n, with m especially
in small and moderate samples (n = 150 and n = 300). Notice that, however, this increase
in power becomes negligible in larger samples. Furthermore, it should be noticed that all the
methods suffer from low power in the cases of higher order fractional alternatives, especially
near the null such as when d = 0.95. In the next section we apply the techniques described to
the term structure of interest rates for the UK and the US.

10
5 Empirical Analysis: Testing the term structure of in-
terest rates
In this section, we empirically evaluate the interest rate term structure for the UK and US
between the 10 years and the 3 months yield treasury securities. The interest rate data are
collected from the Bank of England and the Federal Reserve Bank for the UK and US rates
respectively. The samples cover the period between January 1992 and December 2007 with
monthly frequency for a total n = 312.
Before testing the order of integration of the term spread, we test the individual series for
a unit root using the NG and Perron (NP) MZt test, and for the null of a unit root against
the alternative of fractional integration using all GPH, LW, ELW and FELW estimators. The
results of the Ng and Perron test displayed in Table 9 clearly do not reject the hypothesis of
a unit root in all the short and long term UK and US rates suggesting that they are all non-
stationary I(1) series. In Table 10 the results for the null that dˆ = 1 for the individual rates
against fractional integration seem to cast some doubt on the MZt results as especially the
LW estimator seems to suggest with 95% confidence that the UK long-term rate and the US
short-term rate may be instead fractionally integrated processes. However, as the evidence is
not overwhelming, we choose to treat the interest rate series as I(1). This would be in line with
the theory of term structure and also constitutes the necessary basis for analysing the term
structure using the residual based tests for fractional cointegration which we have discussed
above and for which critical values have been derived.
The next step involves regressing the long term rates against the relative short term one and
obtaining the residuals which will be in turn tested for their order of (fractional) integration 2 .
We use all the GPH, LW, ELW and FELW techniques with different bandwidths for comparison.
Specifically in Table 11 we report estimates of the order of integration dˆ of the term spread for
different values of α = 0.55, 0.575, 0.6.
Generally speaking, the results support the claim that the term spread is indeed a mean
reverting process which displays long memory, thereby offering support for the theory of the
term structure of interest rates as in Campbell and Shiller (1987,1991) and Hall et al. (1992).
However, we notice that there are differences in the results provided by the different techniques.
The GPH in fact indicates that the UK term structure does not seem to be a mean reverting
long memory process with finite or even infinite variance reporting values of d that range from
a minimum of 0.884 to a maximum of 0.955 depending on the choice of m. On the other hand
LW, ELW and FELW strongly suggest that the UK term structure is a mean reverting process,
2
It should be also noticed that we have tested the term spread for a unit root using standard unit root tests
and we have also run standard cointegration tests between the series. All the tests conducted, which are not
reported here but can be made available on request, did not reject the null of a unit root on the term spread
and or the null of no cointegration between the interest rates series.

11
although its variance may not be finite. The implication of mean reversion is that a shock will
not have a permanent effect on the term spread. This is very different from the case where the
term spread would be an I(1) process which is both covariance non stationary and not mean
reverting, as a shock to it would persist forever. The estimates of d provided by LW and ELW
are very similar to each others ranging from 0.589 to 0.616 for the LW estimator and from 0.6
to 0.65 for both ELW and FELW which unsurprisingly provide identical results.
The estimation of the order of integration of the US term spread provides similar albeit
slightly weaker results. which seem to be a bit more sensitive to the value of m used in the
estimation. When m = mα has α = 0.55 or 0.575, all the LW, ELW and FELW tests point
toward mean reversion at least at the 10% level with dˆ which ranges between 0.704 and 0.744.
However, when α = 0.6, the null of no cointegration is supported by LW, ELW and FELW.
The story is completely different when we use the GPH test. When α = 0.55 and 0.575 the
null that the term spread is integrated of order 1 cannot be rejected, not even at the 10% level,
but using α = 0.6, the estimated order of integration of the US term spread is 0.784 which is
significantly different from 1 at the 10 % level.
As we have seen there is strong evidence that the term spread for the UK is a mean revrting
process and fairly good evidence that the same happens for the term spread in the US. The
mean reversion of the term spread is very attractive, even though the term spreads may be
covariance non-stationary, as it suggests that the equilibrium is restored but within a relatively
long time. The strong albeit not infinite persistence of shocks may be justified for example
by the existence of sentimental effects which can last for long periods such as the case where
agents being generally positive about future economic conditions, may ask for a smaller liquidity
premium, whereas, on the contrary, a higher liquidity premium could be required in the case
of a bearish attitude towards future economic conditions. In reality, these kind of sentimental
effects can last for quite a long period of time,implying that deviation from equilibrium in
term structure can last for quite a while before reverting back to its long run path. Long
memory of the term spread may also be related to business cycles, where expectations about
the future as well as the liquidity premium may change in different stages thus affecting the
time series property of the interest rates term structure. In any of these cases, clearly the
fractional cointegration environment proves to be a valid framework when analysing the term
structure of interest rates.

6 Summary and Conclusions

In this paper, we have used residual based tests for fractional cointegration to investigate the
term structure of interest rates of UK and US treasury securities. In testing the hypothesis
of fractional cointegration and estimating the order of d of the term spread, we have used

12
established techniques such as GPH and LW estimators but we have also extended the literature
on residual based fractional cointegration tests as we have used Exact Local Whittle and Feasible
Exact Local Whittle based tests (Phillips and Shimotsu 2004, Shimotsu and Phillips 2006). As
the distribution of the tests is non standard when used on residuals from an OLS regression,
we derived the critical values for the null of no cointegration and performed a simulation to
evaluate their power for different sample sizes and bandwidths. Interestingly we found that
ELW and FELW outperform their rival techniques showing consistently higher power under all
the alternatives for all the sample sizes considered and have used them to estimate the order
of fractional integration of the term spread in order to verify the validity of the expectation
hypothesis.
The simple expectation hypothesis models such as Campbell and Shiller (1987, 1991), Hall
et al. (1992) suggest there exists a long run equilibrium relationship between the short-term and
long-term interest rates. Standard unit root analysis and cointegration tests have often failed to
confirm the prediction of the theoretical model. The dichotomy between I(0) and I(1) implied
from standard cointegration analysis may prove to be too restrictive to be used for the analysis
of the term structure. Therefore, we hypothesized the existence of long-range dependence of
term structure and used a battery of residual based test for fractional cointegration of the UK
and US data. The results offer support for our hypothesis revealing that the term spread is
indeed a mean reverting process with long memory. The mean reverting phenomena found in
our series is very attractive, even though with the fact that the term spread are covariance
non-stationary, as it suggests that the equilibrium is restored but within a relatively long time.

13
Table 1: Quantiles of residual-based GPH test for fractional cointegration

n=1000
Quantile m = n0.55 m = n0.575 m = n0.6
0.0250 -2.2553 -2.3446 -2.2732
0.0500 -1.9318 -1.9381 -1.9019
0.1000 -1.4999 -1.5116 -1.5016
0.2500 -1.0103 -1.0357 -1.0133
0.5000 -0.1328 -0.1509 -0.1462
0.7500 0.3925 0.3536 0.3844
0.9000 1.1063 1.0872 1.1085
0.9500 1.4263 1.4170 1.4360
0.9750 1.7019 1.7162 1.7497
Mean: -0.2315 -0.2562 -0.2217
StDev: 1.0198 1.0254 1.0259

14
Table 2: Quantiles of residual-based LW test for fractional cointegration

n=312
Quantile m = n0.550 m = n0.575 m = n0.600
0.025 -1.777 -1.702 -1.620
0.050 -1.512 -1.440 -1.402
0.100 -1.225 -1.171 -1.144
0.250 -0.783 -0.739 -0.722
0.500 -0.315 -0.289 -0.271
0.750 0.107 0.142 0.153
0.900 0.478 0.508 0.525
0.950 0.698 0.720 0.747
0.975 0.889 0.922 0.937
Mean: -0.353 -0.314 -0.294
StDev: 0.676 0.666 0.649

15
Table 3: Quantiles of residual-based ELW test for fractional cointegration

n=312
Quantile m = n0.550 m = n0.575 m = n0.600
0.025 -1.719 -1.564 -1.522
0.050 -1.428 -1.332 -1.290
0.100 -1.129 -1.091 -1.033
0.250 -0.709 -0.655 -0.618
0.500 -0.250 -0.224 -0.201
0.750 0.171 0.197 0.199
0.900 0.537 0.573 0.552
0.950 0.805 0.776 0.766
0.975 0.013 0.965 0.932
Mean: -0.276 -0.243 -0.218
StDev: 0.675 0.654 0.622

16
Table 4: Quantiles of residual-based FELW test for fractional cointegration

n=312
Quantile m = n0.550 m = n0.575 m = n0.600
0.025 -1.688 -1.729 -1.673
0.050 -1.546 -1.468 -1.334
0.100 -1.243 -1.123 -1.031
0.250 -0.706 -0.639 -0.602
0.500 -0.208 -0.192 -0.151
0.750 0.233 0.250 0.271
0.900 0.606 0.637 0.630
0.950 0.842 0.846 0.868
0.975 1.064 1.023 1.060
Mean: -0.255 -0.222 -0.181
StDev: 0.709 0.692 0.662

17
 Table 5: Power of residual-based GPH test for fractional cointegration

T=150
Power at the 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.7048 0.6944 0.6644 0.5872 0.4616 0.3436 0.2208 0.1304 0.0688 0.0356
m = n0.575 0.7796 0.7588 0.7352 0.6468 0.5092 0.3732 0.2348 0.1336 0.0620 0.0344
m = n0.60 0.8928 0.8732 0.8584 0.7916 0.6536 0.4868 0.3120 0.1672 0.0756 0.0396
Power at the 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.7972 0.7960 0.7712 0.7124 0.6012 0.4720 0.3200 0.2068 0.1164 0.0656
m = n0.575 0.8740 0.8656 0.8396 0.7868 0.6712 0.5320 0.3636 0.2276 0.1184 0.0652
m = n0.60 0.9372 0.9288 0.9256 0.8808 0.7852 0.6412 0.4476 0.2656 0.1340 0.0636
T=300
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.8920 0.8976 0.8976 0.8500 0.7644 0.6008 0.3992 0.2060 0.0952 0.0400
m = n0.575 0.9228 0.9356 0.9276 0.8960 0.8168 0.6484 0.4268 0.2056 0.0884 0.0364
m = n0.60 0.9712 0.9744 0.9724 0.9520 0.8992 0.7628 0.5288 0.2688 0.1084 0.0412
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9276 0.9344 0.9376 0.9120 0.8568 0.7204 0.5332 0.3024 0.1460 0.0736
m = n0.575 0.9540 0.9640 0.9616 0.9492 0.9056 0.7872 0.5760 0.3240 0.1592 0.0772
m = n0.60 0.9832 0.9872 0.9856 0.9804 0.9468 0.8644 0.6652 0.3924 0.1848 0.0780
T=500
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9364 0.9492 0.9604 0.9364 0.9128 0.7600 0.5472 0.2928 0.1160 0.0428
m = n0.575 0.9680 0.9772 0.9800 0.9688 0.9536 0.8360 0.6052 0.3216 0.1212 0.0400
m = n0.60 0.9892 0.9964 0.9968 0.9892 0.9824 0.9164 0.7224 0.4128 0.1460 0.0448
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9580 0.9696 0.9792 0.9656 0.9564 0.8536 0.6724 0.4060 0.1896 0.0696
m = n0.575 0.9844 0.9908 0.9916 0.9836 0.9776 0.9228 0.7484 0.4608 0.2104 0.0756
m = n0.60 0.9936 0.9988 0.9980 0.9952 0.9928 0.9640 0.8312 0.5456 0.2408 0.0852
T=1000
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9796 0.9892 0.9936 0.9932 0.9856 0.9420 0.7704 0.4416 0.1596 0.0428
m = n0.575 0.9944 0.9964 0.9984 0.9984 0.9952 0.9744 0.8356 0.5016 0.1684 0.0444
m = n0.60 0.9996 0.9988 1.0000 1.0000 0.9992 0.9920 0.9188 0.6164 0.2216 0.0544
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9872 0.9944 0.9964 0.9968 0.9920 0.9684 0.8572 0.5736 0.2436 0.0748
m = n0.575 0.9984 0.9984 0.9996 0.9992 0.9984 0.9892 0.9180 0.6540 0.2780 0.0824
m = n0.60 0.9996 0.9996 1.0000 1.0000 1.0000 0.9980 0.9672 0.7420 0.3300 0.0916

18
 Table 6: Power of residual-based LW test for fractional cointegration

T=150
Power at the 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9870 0.9704 0.9376 0.8648 0.7214 0.5480 0.3364 0.2034 0.1166 0.0614
m = n0.575 0.9958 0.9886 0.9684 0.9198 0.8010 0.6170 0.3872 0.2238 0.1190 0.0636
m = n0.60 0.9996 0.9976 0.9924 0.9668 0.8982 0.7316 0.4890 0.2768 0.1382 0.0674
Power at the 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9970 0.9906 0.9768 0.9382 0.8594 0.7164 0.5114 0.3332 0.2036 0.1312
m = n0.575 0.9986 0.9960 0.9892 0.9719 0.9154 0.7888 0.5888 0.3748 0.2340 0.1370
m = n0.60 1.0000 0.9998 0.9978 0.9906 0.9572 0.8554 0.6654 0.4222 0.2448 0.1366
T=300
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 0.9996 0.9986 0.9918 0.9568 0.8378 0.6156 0.3460 0.1610 0.0674
m = n0.575 1.0000 1.0000 0.9998 0.9976 0.9766 0.8902 0.6750 0.3814 0.1644 0.0602
m = n0.60 1.0000 1.0000 1.0000 0.9998 0.9916 0.9460 0.7590 0.4450 0.1920 0.0656
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 1.0000 0.9998 0.9974 0.9820 0.9266 0.7726 0.5148 0.2708 0.1292
m = n0.575 1.0000 1.0000 1.0000 0.9998 0.9914 0.9560 0.8236 0.5582 0.2884 0.1322
m = n0.60 1.0000 1.0000 1.0000 1.0000 0.9970 0.9808 0.8786 0.6164 0.3142 0.1270
T=500
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 1.0000 1.0000 0.9998 0.9938 0.9494 0.7724 0.4360 0.1924 0.0682
m = n0.575 1.0000 1.0000 1.0000 0.9998 0.9984 0.9802 0.8558 0.5394 0.2200 0.0730
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 0.9918 0.9158 0.6184 0.2548 0.0760
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 1.0000 1.0000 1.0000 0.9978 0.9814 0.8890 0.6256 0.3258 0.1428
m = n0.575 1.0000 1.0000 1.0000 1.0000 0.9998 0.9940 0.9336 0.6946 0.3554 0.1476
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 0.9986 0.9634 0.7724 0.4028 0.1608
T=1000
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 1.0000 1.0000 1.0000 1.0000 0.9966 0.9488 0.6862 0.3064 0.0888
m = n0.575 1.0000 1.0000 1.0000 1.0000 1.0000 0.9990 0.9856 0.7972 0.3692 0.1008
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 0.9998 0.9952 0.8712 0.4400 0.1074
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 1.0000 1.0000 1.0000 1.0000 0.9994 0.9824 0.8256 0.4508 0.1670
m = n0.575 1.0000 1.0000 1.0000 1.0000 1.0000 0.9996 0.9956 0.8886 0.5242 0.1800
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9980 0.9288 0.5684 0.1842

19
 Table 7: Power of residual-based ELW test for fractional cointegration

T=150
Power at the 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9840 0.9824 0.9676 0.9280 0.8292 0.6852 0.4708 0.3000 0.1616 0.0808
m = n0.575 0.9936 0.9920 0.9844 0.9640 0.8916 0.7616 0.5272 0.3388 0.1680 0.0800
m = n0.60 0.9972 0.9984 0.9952 0.9824 0.9408 0.8220 0.5988 0.3588 0.1728 0.0780
Power at the 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9904 0.9908 0.9820 0.9600 0.9028 0.7892 0.5972 0.4032 0.2352 0.1272
m = n0.575 0.9940 0.9956 0.9912 0.9804 0.9372 0.8356 0.6396 0.4316 0.2368 0.1176
m = n0.60 0.9980 0.9988 0.9956 0.9924 0.9724 0.8908 0.7056 0.4688 0.2404 0.1148
T=300
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9976 0.9980 0.9984 0.9924 0.9640 0.8644 0.6636 0.3788 0.1572 0.0716
m = n0.575 0.9992 0.9992 0.9992 0.9968 0.9836 0.9160 0.7348 0.4284 0.1748 0.0760
m = n0.60 0.9996 0.9996 0.9996 0.9988 0.9944 0.9532 0.8004 0.4944 0.2000 0.0668
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9988 0.9992 0.9992 0.9980 0.9824 0.9284 0.7808 0.5148 0.2572 0.1160
m = n0.575 0.9992 0.9996 0.9996 0.9984 0.9908 0.9548 0.8320 0.5536 0.2744 0.1224
m = n0.60 1.0000 1.0000 1.0000 0.9996 0.9984 0.9784 0.8728 0.6120 0.2964 0.1208
T=500
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9996 1.0000 1.0000 0.9996 0.9940 0.9556 0.7864 0.4604 0.1896 0.0560
m = n0.575 0.9996 1.0000 1.0000 1.0000 0.9988 0.9808 0.8688 0.5404 0.2232 0.0608
m = n0.60 1.0000 1.0000 1.0000 1.0000 0.9996 0.9948 0.9280 0.6236 0.2496 0.0668
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9996 1.0000 1.0000 1.0000 0.9984 0.9824 0.8776 0.5952 0.2900 0.1052
m = n0.575 1.0000 1.0000 1.0000 1.0000 1.0000 0.9924 0.9240 0.6608 0.3216 0.1064
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 0.9992 0.9624 0.7268 0.3560 0.1128
T=1000
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 1.0000 0.9996 1.0000 1.0000 0.9956 0.9244 0.6216 0.2276 0.0452
m = n0.575 1.0000 1.0000 1.0000 1.0000 1.0000 0.9996 0.9688 0.7316 0.2744 0.0548
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9924 0.8244 0.3340 0.0580
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 1.0000 1.0000 1.0000 1.0000 0.9992 0.9688 0.7620 0.3436 0.0940
m = n0.575 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9888 0.8312 0.3976 0.1000
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9980 0.9024 0.4568 0.1008

20
 Table 8: Power of residual-based FELW test for fractional cointegration

T=150
Power at the 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9532 0.9432 0.9312 0.8868 0.7964 0.6552 0.4528 0.2808 0.1516 0.0804
m = n0.575 0.9780 0.9752 0.9740 0.9424 0.8736 0.7352 0.5156 0.3252 0.1620 0.0820
m = n0.60 0.9932 0.9916 0.9900 0.9776 0.9336 0.8164 0.5948 0.3560 0.1696 0.0804
Power at the 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9768 0.9748 0.9712 0.9456 0.8824 0.7716 0.5916 0.3952 0.2284 0.1208
m = n0.575 0.9904 0.9864 0.9884 0.9692 0.9264 0.8268 0.6292 0.4176 0.2264 0.1196
m = n0.60 0.9952 0.9956 0.9948 0.9872 0.9672 0.8844 0.7040 0.4524 0.2396 0.1164
T=300
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9932 0.9928 0.9948 0.9872 0.9568 0.8576 0.6576 0.3720 0.1560 0.0672
m = n0.575 0.9972 0.9988 0.9992 0.9936 0.9796 0.9092 0.7352 0.4264 0.1816 0.0728
m = n0.60 0.9996 0.9996 0.9996 0.9976 0.9928 0.9548 0.8076 0.5040 0.2072 0.0712
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9980 0.9980 0.9988 0.9956 0.9796 0.9256 0.7856 0.5032 0.2536 0.1128
m = n0.575 0.9992 0.9992 0.9996 0.9980 0.9892 0.9504 0.8248 0.5408 0.2616 0.1144
m = n0.60 1.0000 0.9996 0.9996 0.9996 0.9984 0.9768 0.8724 0.6092 0.2936 0.1156
T=500
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9984 0.9984 0.9992 0.9992 0.9932 0.9504 0.7820 0.4544 0.1796 0.0584
m = n0.575 0.9992 1.0000 1.0000 1.0000 0.9992 0.9812 0.8748 0.5420 0.2196 0.0596
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 0.9948 0.9324 0.6384 0.2608 0.0700
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 0.9996 1.0000 1.0000 0.9996 0.9976 0.9824 0.8776 0.5972 0.2880 0.1060
m = n0.575 0.9992 1.0000 1.0000 1.0000 1.0000 0.9928 0.9240 0.6592 0.3168 0.1000
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 0.9988 0.9656 0.7260 0.3572 0.1080
T=1000
Power at 5% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 1.0000 1.0000 1.0000 1.0000 0.9948 0.9228 0.6324 0.2336 0.0440
m = n0.575 1.0000 1.0000 1.0000 1.0000 1.0000 0.9996 0.9708 0.7424 0.2820 0.0564
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9944 0.8364 0.3476 0.0616
Power at 10% level
d 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
m = n0.55 1.0000 1.0000 1.0000 1.0000 1.0000 0.9996 0.9708 0.7652 0.3488 0.0960
m = n0.575 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9884 0.8312 0.3972 0.0956
m = n0.60 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9980 0.9016 0.4268 0.0976

21
Table 9: Empirical results: Tests for the null of a unit root on the UK and the US interest rates

UK 3 months UK 10 years US 3 months US 10 years

MZt statistic lags MZt statistic lags MZt statistic lags MZt statistic lags
-2.2170 1 -1.9470 1 -1.4176 1 -1.5996 2
MZt test critical values
1% level -3.4200
5% level -2.9100
10% level -2.6200

22
 Table 10: Empirical results: Tests for the null of d = 1 for the UK and the US interest rates

GPH test LW test ELW test FELW test

α=0.55 α=0.575 α=0.6 α=0.55 α=0.575 α=0.6 α=0.55 α=0.575 α=0.6 α=0.55 α=0.575 α=0.6
UK 3 months
dˆ 1.074 1.145 1.050 0.824 0.954 0.954 0.764 0.920 0.920 0.740 0.923 0.923
t-stat for H0 : d = 1 0.451 0.969 0.365 -0.844 -0.218 -0.218 -1.224 -0.413 -0.413 -1.442 -0.425 -0.426
UK 10 years
dˆ 0.739 0.743 0.695 0.700 0.678 0.673 0.894 0.897 0.904 0.894 0.892 0.904

23
t-stat for H0 : d = 1 -1.570 -1.710 -2.208 -1.437 -1.671 -1.820 -0.504 -0.557 -0.534 -0.504 -0.559 -0.534
US 3 months
dˆ 0.967 0.993 1.113 0.660 0.649 0.664 0.803 0.808 0.840 0.803 0.808 0.841
t-stat for H0 : d = 1 -0.197 -0.043 0.990 -1.630 -1.822 -1.865 -0.942 -0.994 -0.885 -0.942 -0.994 -0.885
US 10 years
dˆ 1.093 1.051 1.043 0.663 0.651 0.677 0.884 0.881 0.924 0.884 0.881 0.924
t-stat for H0 : d = 1 0.560 0.343 0.314 -1.612 -1.813 -1.796 -0.552 -0.614 -0.421 -0.552 -0.614 -0.421
95% Crit. Values -2.135 -2.091 -2.053 -1.104 -1.045 -1.008 -1.201 -1.170 -1.088 -1.273 -1.225 -1.098
90% Crit Values -1.736 -1.706 -1.668 -0.845 -0.803 -0.785 -0.913 -0.884 -0.828 -0.951 -0.904 -0.838
Table 11: Empirical results: testing the term structure

The US term structure

m = n0.55
GPH LW ELW FELW
dˆ 0.807 0.704 0.725 0.725
t-stat for H0 : d = 1 -1.159 -1.422 -1.318 -1.318
m = n0.575
GPH LW ELW FELW
dˆ 0.774 0.727 0.744 0.744
t-stat for H0 : d = 1 -1.506 -1.416 -1.326 -1.326
m = n0.60
GPH LW ELW FELW
dˆ 0.784 0.844 0.862 0.862
t-stat for H0 : d = 1 -1.563 -0.867 -0.764 -0.764
The UK term structure
m = n0.55
GPH LW ELW FELW
dˆ 0.898 0.589 0.600 0.600
t-stat for H0 : d = 1 -0.614 -1.971 -1.918 -1.918
m = n0.575
GPH LW ELW FELW
dˆ 0.884 0.592 0.618 0.618
t-stat for H0 : d = 1 -0.772 -2.117 -1.981 -1.980
m = n0.60
GPH LW ELW FELW
dˆ 0.955 0.616 0.650 0.650
t-stat for H0 : d = 1 -0.320 -2.135 -1.946 -1.946

24
7 References
[1] Backus, D. K. and S.E. Zin, 1993, Long memory inflation uncertainty: evidence from the
term structure of interest rates, Journal of Money Credit and Banking, 25, 681-700.
[2] Banerjeea, A., and G. Urga, 2005, Modelling structural breaks, long memory and stock
market volatility: an overview, Journal of Econometrics, 129, 1-34
[3] Camarero, M., and C. Tamarit, 2002, Instability tests in cointegration relationships. An
application to the term structure of interest rates, Economic Modelling 19, 783-799
[4] Campbell, J.Y., Shiller, R.J., 1987. Cointegration and tests of present value models. J.
Polit. Economy 95, 1062-1088.
[5] Caporale, G. M., and L. A. Gil-Alana, 2004, Fractional cointegration and tests of present
value models, Review of Financial Economics, 13, 3, 245-258
[6] Cheung, Yin-Wong and Kon S. Lai, 1993, A fractional cointegration analysis of purchasing
power parity, Journal of Business and Economic Statistics, Vol 11, 1, 103-112
[7] Davidson, J., 2002, A model of fractional cointegration, and tests for cointegration using
the bootstrap, Journal of Econometrics, 110, 187-212
[8] Diebold, F.X., Inoue, A., 2001. Long memory and regime switching. Journal of Econometrics
105, 131-159.
[9] Fox, R and M.S. Taqqu, 1986 Large-sample properties of parameter estimates for strongly
dependent stationary Gaussian time series, Annals of Statistics. 14, pp. 517532.
[10] Geweke, J. and S. Porter-Hudak, 1983, The estimation and application of long memory
time series models, Journal of Time Series Analysis, 4, 221-38.
[11] Granger, C. W. J. and R. Joyeux, 1980. An introduction to long-memory time series
models and fractional differencing. Journal of Time Series Analysis 1, 15-2
[11] Granger, C. W. J. 1980. Long-memory relationships and the aggregation of dynamic
models. Journal of Econometrics, 14, 227-238
[12] Granger, C.W.J. and N. Hyung, 2004. Occasional structural breaks and long-memory,
Journal of Empirical Finance, 11, pp. 399421.
[13] Hall, A.D., Anderson, H.M., Granger, C.W.J., 1992. A cointegration analysis of treasury
bill yields. Review of Economics and Statistics, 74, 116-126.
[14] Hosking, J. R. M. (1981). Fractional differencing. Biometrika 68, 165-76.
[15] Kuo, Shew-Huei and W. Enders, 2004, The term structure of Japanese interest rates:
The equilibrium spread with asymmetric dynamics, Journal of the Japanese and International

25
Economies 18, 84-98.
[16] MacKinnon, J., 1990, Critical values for cointegration tests, University of California at San
Diego, Economics Working Paper Series 90-4
[17] Mankiw, G., and J. Miron, 1986. The changing behavior of the term structure of interest
rates. Quarterly Journal of Economics 101, 211-228.
[18] Mankiw, G., and L. Summers, 1984. Do long-term interest rates overreact to short-term
interest rates? Brookings Papers on Economic Activity, pp. 223-242.
[19] Marinucci D. and P.M. Robinson, 2001, Semiparametric fractional cointegration analysis,
Journal of Econometrics, 105, 225-247
[20] Robinson, P.M., 1995, Gaussian semiparametric estimation of long range dependence,
Annals of Statistics, 23, 1630-1661.
[21] Robinson, P.M. and Y. Yajima, 2002, Determination of Cointegrating rank in fractional
systems, Journal of Econometrics, 106, 217-241
[22] Rudebusch, G.D., 1995. Federal reserve interest rate targeting, rational expectations, and
the term structure. Journal of Monetary Economics, 35, 245-274.
[23] Seo, B., 2003, Nonlinear mean reversion in the term structure of interest rates. Journal of
Economic Dynamics and Control 27, 2243 - 2265.
[24] Shimotsu, K., and P.C.B. Philllips, 2005, Exact Local Whittle Estimation of Fractional
Integration, Annals of Statistics 33(4), August 2005, pp. 1890-1933
[25] Shimotsu, K., and P.C.B. Philllips, 2006, Local Whittle estimation of fractional integration
and some of its variants, Journal of Econometrics, 103, 209-233.
[26] Shimotsu, K., and P.C.B. Phillips, 2009, Exact Local Whittle Estimation of Fractional
Integration with Unknown Mean and Time Trend, Journal of Econometrics, forthcoming.
[27] Velasco, C., 1999. Gaussian semiparametric estimation of non-stationary time series. Jour-
nal of Time Series Analysis 20, 87-127.

26