Computational Statistics & Data Analysis 51 (2006) 2278 – 2294

www.elsevier.com/locate/csda

Testing the martingale difference hypothesis using

integrated regression functions
J. Carlos Escancianoa,∗ , Carlos Velascob
a Facultad de Económicas, Universidad de Navarra, Edificio Biblioteca (Entrada Este), Pamplona, 31080, Navarra, Spain
b Universidad Carlos III de Madrid, Spain

Available online 24 August 2006

Abstract
An omnibus test for testing a generalized version of the martingale difference hypothesis (MDH) is proposed. This generalized
hypothesis includes the usual MDH, testing for conditional moments constancy such as conditional homoscedasticity (ARCH effects)
or testing for directional predictability. A unified approach for dealing with all of these testing problems is proposed. These hypotheses
are long standing problems in econometric time series analysis, and typically have been tested using the sample autocorrelations or
in the spectral domain using the periodogram. Since these hypotheses cover also nonlinear predictability, tests based on those second
order statistics are inconsistent against uncorrelated processes in the alternative hypothesis. In order to circumvent this problem
pairwise integrated regression functions are introduced as measures of linear and nonlinear dependence. The proposed test does not
require to chose a lag order depending on sample size, to smooth the data or to formulate a parametric alternative model. Moreover,
the test is robust to higher order dependence, in particular to conditional heteroskedasticity. Under general dependence the asymptotic
null distribution depends on the data generating process, so a bootstrap procedure is considered and a Monte Carlo study examines
its finite sample performance. Then, the martingale and conditional heteroskedasticity properties of the Pound/Dollar exchange rate
are investigated.
© 2006 Elsevier B.V. All rights reserved.

Keywords: Nonlinear time series; Martingale difference hypothesis; Empirical processes; Exchange rates

1. Introduction

The concept of martingale or martingale difference sequence (MDS) is central in many areas of economics, finance
and other sciences. The martingale difference hypothesis (MDH) states that the best predictor, in the sense of least
mean square error, of a real time series given some information set is just the unconditional mean. In this paper we
are concerned with the case in which the time series to be predicted is a measurable real-valued transformation of a
stationary vector time series Xt ∈ Rd , d ∈ N, t = 0, ±1, ±2, . . . , and the information set is just the past values of the
time series. That is, we would like to test
 
H0 : E Yt | Xt−1, Xt−2, . . . = ,  ∈ R, (1)
where Yt is a measurable real-valued transformation of Xt and  = E [Yt ].

∗ Corresponding author. Tel.: +34 948 425600/2321; fax: +34 948 425626.
E-mail address: [email protected] (J.C. Escanciano).

0167-9473/$ - see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.csda.2006.07.039
 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294 2279

This hypothesis, referred to as the generalized MDH or simply the MDH, contains interesting testing problems as
special cases. For instance, when Yt is a power transformation of Xt , we are testing for conditional moments constancy;
the usual MDH (when Yt coincides with Xt ) or testing for ARCH effects of a MDS are examples of this case. When
Yt = I (Xt > c), c ∈ Rd , where the indicator I (A) is one when the event A occurs and zero otherwise, we are testing
for no directional predictability, see Hong and Chung (2003) or Linton and Whang (2004). If Xt = (X1t , X2t , X3t ), X1t
and X2t are random variables, Yt = X1t − X2t and  = 0, we are testing for equality of regression curves, see Ferreira
and Stute (2004) for a recent reference. All these hypotheses are important in economics and finance.
Usually, these hypotheses have been typically tested using the autocorrelations or autocovariances or in the spectral
domain using the periodogram. For instance, Cochrane (1988) proposed a variance ratio test for uncorrelatedness that
has been widely used in finance. Durlauf (1991) proposed a spectral distribution based test, using the fact that under the
MDH, the standardized spectral distribution function is a straight line. Recently, Deo (2000) has robustified Durlauf’s
(1991) test against certain forms of conditional heteroskedasticity.
However, all these tests are suitable for testing lack of serial correlation but not the MDH. In fact, they are not consistent
against nonmartingale difference sequences with zero autocorrelations, that is, when only nonlinear dependence is
present, as commonly happens with economic and financial data, see e.g. our application to exchange rates. These tests
are inconsistent because they only employ information contained in the second sample moments of the process.
Roughly speaking, there have been two main approaches in the literature for designing consistent tests for (1). The
first approach is based on checking an infinite number of orthogonality conditions, see for instance, Bierens (1984),
Stute (1997), Bierens and Ploberger (1997), Koul and Stute (1999), Whang (2000), Dominguez and Lobato (2003),
Escanciano and Velasco (2006) or Escanciano (2006a), among many others. The second line of research employs
nonparametric estimates of the conditional expectation function, see for instance Härdle and Mammen (1993), Zheng
(1996), Li (1999), Guerre and Lavergne (2005) and references therein.
Tests based on the second methodology have standard asymptotic null distributions, but they usually require strong
assumptions on the data generating process (DGP), see e.g. Li (1999). More importantly, they require subjective choices
of smoothing parameters and kernel functions, and usually statistical inferences can be sensitive to these selections.
On the other hand, tests statistics based on the first methodology do not demand in general the selection of any
user-chosen parameters. They are consistent against Pitman’s alternatives converging at the parametric rate and do not
require strong conditions about the DGP, and although they have nonstandard asymptotic null distributions, these can
be well approximated by bootstrap methods,  see Section  5. We propose a MDH test based on this methodology that
maintains such desirable properties. If It = Xt, Xt−1, . . . is the information set at time t and Ft is the -field generated
by It , the first methodology exploits the following equivalence principle:
   
E Yt | It−1 =  a.s.,  ∈ R ⇐⇒ E (Yt − ) f (It−1 ) = 0, (2)

 Ft−1 -measurable
for all bounded  weighting function f (·). Tests are usually based on the discrepancy of the sample
analog of E (Yt − ) f (It−1 ) to zero.
The problem of testing over all possible weighting functions can be reduced to testing the orthogonality condition
over a parametric family of functions. That is
   
E (Yt − ) f (It−1 ) = 0 ∀f (·) ⇐⇒ E (Yt − ) W (It−1 , x) = 0 ∀x ∈ ,

for a suitable family W (It−1 , x) and set , see Stinchcombe and White (1998). The most used weighting functions
have been either exponential functions, see e.g. Bierens (1984), or indicators functions, as in Stute (1997).
Most of the above references test the MDH conditioning on a finite-dimensional information set, and therefore, they
test for a particular Markov property and not the MDH, which involves an infinite number of lags. This solution is
unsatisfactory because there could be structure in the conditional mean at omitted lags. Often, the maximum power
could be achieved by using the correct lag order of the alternative. However, prior information on the conditional mean
structure is usually not available. In addition, when a large number of conditioned variables is considered, the empirical
power of those tests could be seriously affected by the curse of dimensionality.
There have been some proposals considering infinite-dimensional information sets. de Jong (1996) generalized
Bierens’ test to time series, and although this test has the appealing property of considering an increasing number of
lags as the sample size increases, it requires numerical integration with dimension equal to the sample size, which
makes this test unfeasible in applications where the sample size is usually large, e.g. financial applications. Using a
2280 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294

different methodology based on spectral methods, Hong (1999) proposed a MDH test, but his test is not robust to
conditional heteroskedasticity and higher order dependence because it is based on an independence assumption. This
drawback has been recently overcome in Hong and Lee (2005). Hong’s (1999) approach involves the choice of a kernel,
a bandwidth parameter and an integrating measure and, in general, statistical inferences are not robust to these choices.
Escanciano and Velasco (2006) avoid bandwidth choices by means of a generalized spectral distribution function. In
this paper we propose a MDH test which preserves the good properties of Escanciano and Velasco’s (2006) test and that
complements it because it leads to different power properties. The main methodology is assisted with the individual use
of new nonlinear dependence measures defined in the paper, which can play an important role in explaining well-known
stylized facts of economic and financial data, see Section 6 for examples.
Summarizing, the aims of this paper are two: first, to develop a methodology for testing the MDH that (i) does not
involve the choice of any lag order or kernel function; (ii) avoid the curse of dimensionality; (iii) is consistent against
a large class of nonlinear and linear deviations from (1); (iv) is robust to higher order dependence, such as conditional
heteroskedasticity; (v) is simple to compute; and (vi) performs quite well in finite samples as will be shown below, and
secondly, to propose new measures of nonlinear dependence which can provide new useful tools for studying nonlinear
serial dependence in applied work, as it is illustrated in the exchange rate application of Section 6.
The layout of the article is as follows, in Section 2 we define the integrated pairwise regression functions and the
integrated pairwise autoregression functions as our measures of dependence. In Section 3 we use these new dependence
measures to set a general methodology to test (1). In Section 4 we study the asymptotic distribution of our test under
the null. In Section 5 we propose a bootstrap approach and we present a simulation exercise comparing with competing
tests. In Section 6 we present an empirical application of our test and dependence measures to exchange rates. In the
Appendix we collect some asymptotic results on the new dependence measures. In the sequel, C is a generic constant
that may change from one expression to another. Let ∧ denotes the minimum, i.e, a ∧ b = min{a, b}. Unless indicated,
all convergences are taken as the sample size n −→ ∞.

2. Generalized dependence measures

In this section we define a generalization of the usual autocovariances and crosscovariances to a nonlinear framework.
It is well known that in the presence of nonlinearity (or non-Gaussianity) these measures do not characterize the
dependence in  the conditional
 mean and the practitioner needs more reliable measures such as the pairwise regression
functions E Yt | Xt−j = x . In general, estimation of these functions involves nonparametric smoothing estimation
with subjective bandwidth choices.
 Robinson
 (1983) has studied the large sample properties of kernel estimators of
lagged conditional means E Xt | Xt−j for various lags j , see also Auestad and TjZstheim (1990). A natural way to
overcome the smoothing approach is to consider cumulative measures. Assume that X ∈ Rd , and the random variable
Y is integrable, so that the regression function

m(x) = E[Y −  | X = x], x ∈ Rd ,

is well defined (up to a null set). By a measure-theoretic argument, the regression function m(·) can be characterized
by the integrated regression function (x) given by

 
(x) = E[(Y − )I (X x)] = E Y −  | X = z F (dz),
(−∞,x]d

where the second equality follows by the law of iterated expectations and F (·) is the cumulative distribution function
of X. In a time series context is particularly interesting the case when Y = Yt and X = Yt−j , in that case, the measures
  
j (x) = E (Yt − ) I Yt−j x

can be called the integrated pairwise autoregression functions (IPAF). In the general case in which Yt is different from
Xt , the measures
  
j (x) = E (Yt − ) I Xt−j x (3)
 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294 2281

sare the integrated pairwise regression functions (IPRF). These measures are useful for testing interesting hypotheses
in a nonlinear time series framework as in (1). They are able to pick out both linear and nonlinear dependence in the
conditional mean. The natural estimator of j (x) based on a sample {Yt , Xt }nt=1 is

1 
n
   

j (x) = Yt − Y n−j I Xt−j x , (4)
n−j
t=1+j

with

1 
n
Y n−j = Yt .
n−j
t=1+j

In the Appendix we develop the asymptotic theory for  j (x) under the null hypothesis (1). As a consequence of
Proposition 1 in the Appendix and the continuous mapping theorem we have that under H0 ,


KS Y |X (j ) := sup (n − j )1/2
j (x) = max (n − j )1/2
j (Xt )
1t n
x∈[−∞,∞]d

converges to the supremum of a Gaussian process, where the subscript Y | X in KS Y |X (j ) indicates that Y is the
dependent variable and X the conditioning variable at lag j. In particular, under homoscedasticity and d=1, a standardized
version ofj (x) has a standard Brownian bridge as the limiting distribution and asymptotic inference is possible because
the asymptotic quantiles are readily available, see Shorack and Wellner (1986). In the general case, the quantile can
be approximated via a bootstrap approach, see Section 5. With the bootstrap critical values we can calculate uniform
confidence bands for  j (x) and the significance of j (x) can be tested, see Section 6 below. Note that  j (x) can be
useful for detecting nonlinearities graphically (see Tong, 1990, p.12). In contrast with kernel estimators,  j (x) does

AUTOCORRELOGRAM
0.1

0.05
r (j)

-0.05

-0.1
0 5 10 15 20 25 30 35
Lag j

NON-LINEAR IPRF PLOT

2

1.5
KS (j)

0.5

0
0 5 10 15 20 25 30 35
Lag j

Fig. 1. Linear correlogram and nonlinear IPRF plot for the NLMA data.
2282 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294

not depend on kernel and bandwidth choices, so tests are straightforward to implement. Our test statistic for the MDH
j (x) simultaneously to test a pairwise version of (1). The plot of a standardization of KS Y |X (j )
uses all the measures
against the lag parameter j 1 can be viewed as generalization of the usual autocovariance plot in linear dependence
to nonlinear conditional mean dependence. We call this plot for the indicator case, the IPRF plot. We illustrate its use
with an example.

Example 1. Consider the following nonlinear moving average model (NLMA):

Yt = εt−1 εt−2 (εt−2 + εt + 1) , εt ∼ i.i.d N (0, 1).

 
2
It is easy to show that this DGP is uncorrelated but not a MDS with respect its past values since E Yt Yt−1  = 0, so
the usual correlogram is useless. However, the IPRF PLOT can display some useful information. In Fig. 1 we compare
both plots for a simulated series with n = 500 of the NLMA process. Considering heteroskedasticity
√ robust correlations
as in Deo (2000), the confidence bands under the null of MDH are, respectively, ±1/ n for Deo’s linear correlogram
and those derived from the bootstrap approach for the IPRF PLOT, see Section 5 below.

3. Pairwise consistent hypothesis testing

Our test is founded on a pairwise approach that leads to computationally feasible tests which are consistent against
a broad class of alternatives, cf. Escanciano and Velasco (2006). We test that all the IPRF are identically zero, i.e.

j (x) = 0 ∀j 1, ∀x ∈ Rd , a.s. (5)

If we define −j (·) = j (·) for j 1, we can consider the Fourier transform of the IPRF j (·),
∞
1 
f (w, x) = j (x)e−ijw ∀w ∈ [−, ], x ∈ Rd , (6)
2
j =−∞

which contains all the information about (5). Because we are taking into account linear and nonlinear dependence,
f (w, x) is a generalization of the spectral (or cross-spectral) density function, cf. Hong (1999). We can also consider
the generalized spectral distribution function as the integral of f (w, x),


H (, x) = 2 f (w, x) dw ∀ ∈ [0, 1], x ∈ Rd ,
0

that is,
∞
 sin j 
H (, x) = 0 (x) + 2 j (x) . (7)
j
j =1

Note that, both f (w, x) and H (, x) exist as functions in an appropriate Hilbert space. The generalized spectral
distribution function contains all the information about the pairwise regressions functions and could be viewed as a
generalization of the test statistic used in Durlauf (1991) and Deo (2000) because they considered the autocorrelations,
instead of the pairwise regression functions, and only took into account the second moment implications of the MDH.
By using H (, x), we consider all the pairwise implications of the MDH, including both linear and nonlinear conditional
dependencies. Notice that a more flexible weighting scheme on the measure j (x) is possible via a kernel function
and a lag-bandwidth parameter, but this approach would introduce some arbitrariness in the test via the kernel and
bandwidth choices.
Our test is based on the sample analogue of (7),
n−1

 1/2
(, x) = j sin j 
H 0 (x) + 2 1− j (x) ,
n j
j =1
 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294 2283

with (1 − j/n)1/2 a finite sample correction factor which delivers a better finite sample performance and where
 n−1
j (x) j =1 are given by (4). Notice that all the n − 1 lags in the sample are used, and therefore, there is no need to
choose a lag order. Therefore, all of the information in the conditional moment restriction is eventually accounted for.
Because (5) is equivalent to H (, x) = 0 (x), the test is based on the discrepancy between H (, x) and H 0 (, x) :=

0 (x). That is, we consider the process
 n 1/2  √
n−1
2 sin j 
Sn (, x) =  
H (, x) − H0 (, x) = (n − j )1/2
j (x)
2 j
j =1

to test H0 .
Informally speaking, under the null (1) the generalized sample spectral distribution H (, x) will be approximately
equal to H 0 (, x) and then, the process Sn (, x) will converge in distribution as n increases, while under the alternative
it is expected that H (, x) will differ from H0 (, x) and hence, Sn (, x) will diverge to infinity when n increases.
In order to evaluate the distance of Sn (, x) to zero, a norm has to be chosen. In this context the natural norm is
the Cramér–von Mises (CvM) norm. If Fn (x) is the usual empirical distribution function based on {Xt }nt=1 , the CvM
norm is
  (n − j )  2
n−1 n
Dn2 := (Sn (, x))2 Fn (dx) d = 
j (Xt ). (8)
j =1
n(j )2 t=1

Our test rejects the null hypothesis for large values of the test statistic Dn2 .

4. Asymptotic null distribution

In this section, we first establish the null limit distribution of the empirical process Sn (, x) under (1). The null limit
distribution of the new test is the limit distribution of a functional of Sn (, x). To further elaborate the foregoing points
we need some notation. Let  be the product measure of the probability distribution F and the Lebesgue measure on
[0, 1]. Also, denote = [0, 1] × [−∞, ∞]d and
:= (, x) ∈ . We consider Sn (
) as a random element on the
Hilbert space L2 ( , ) of all square integrable functions (with respect to the measure ) with the inner product
 
f, g := f (
)g(
) d(
) = f (, x)g(, x)F (dx) d.

For recent applications on convergence results on Hilbert spaces in the econometric literature see Politis and Romano
(1994), Chen and White (1996, 1998), Chen and Fan (1999), Escanciano and Velasco (2006) and Escanciano (2006b). If
Z is a L2 ( , )-valued random variable, we say that Z has mean m ∈ L2 ( , ) if E[ Z, f ] = m, f , ∀f ∈ L2 ( , ).
If EZ2 < ∞ and Z has zero mean, then the covariance operator of Z, CZ (·) say, is a continuous, linear, symmetric,
positive definite operator from L2 ( , ) to L2 ( , ) defined by CZ (h)=E[ Z, hZ]. Let ⇒ denote weak convergence
in the Hilbert space L2 ( , ) endowed with the norm metric. Define

wt (x) := {I (Xt x) − F (x)} , t ∈ Z, x ∈ Rd

and
√
2 sin j 
j () = , j 1.
j

Assumption A1. A1(a):{Xt }∞ t=1 is a stationary ergodic vector process in R with probability distribution function F (·)
d

absolutely continuous with respect to Lebesgue measure. Also, Yt is a measurable real-valued function of Xt .
A1(b):E|Y1 |2 < ∞.

Note that Assumption A1 is mild, in particular it allows us to consider conditional heteroskedastic processes. As-
sumption A1(b) is much weaker than the moment assumptions used in Durlauf (1991) and Deo (2000) for testing the
2284 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294

MDH, who assumed finite eight moments. The next theorem shows the null limit distribution of the L2 ( , )-valued
random element Sn (
). The reader is referred to Escanciano (2006b) for its proof.

Theorem 2. Under Assumption A1 and (1), the process Sn (

) converges weakly to S(
) on L2 ( , ), where S(
) is a
Gaussian process with mean zero and covariance operator

∞ 
 ∞   
       
CS (f ) = E (Y1 − )2 f
1 f
2 w1−j (x1 ) j (1 ) w1−k (x2 ) k (2 ) d
1 d
2 ,
j =1 k=1 ×

where f ∈ L2 ( , ),
1 = (1 , x1 ) and
2 = (2 , x2 ).

Remark 3. It is easy to show that under independence and using the equality

∞ 
 ∞
   
j ()k  =  ∧  −  , (9)
j =1 k=1

      
the covariance function of the process S(
) is equal to 2 F x ∧ x  − F (x)F x   ∧  −  where 2 =
 
E (Y1 − )2 . Thus, in this case the asymptotic null distribution is nuisance parameter-free after standardization
by a consistent estimate of −1 , and the quantiles of the asymptotic distribution of norms of S(
) can be tabulated.

Next corollary shows the asymptotic distribution of our test statistic, and follows from the continuous mapping
theorem (Billingsley’s, 1968, Theorem 5.1) and from Chang’s (1990), Theorem 1 and Lemma 3.1.

Corollary 4. Under (1) and Assumption A1


d
Dn2 −→ D∞
2
:= (S(, x))2 F (dx) d.

It can be shown that the asymptotic distribution of Dn2 can be expressed as a weighted sum of independent 21 random
variables with weights depending on the DGP. This could be a basis to approximate the asymptotic distribution of Dn2
and obtain critical values of tests. We do not pursue this possibility here, on the contrary we approximate the asymptotic
critical values of the test with the assistance of a bootstrap procedure in the next section.
The consistency and local power properties for Dn2 can be discussed following the lines of Section 3.2 in Escanciano
and Velasco (2006). The reader is referred to the latter reference for the asymptotic power local properties of the CvM
test based on Dn2 . In particular, under A1

∞ 
1 2 P  1
Dn −→ |j (x)|2 F ( dx).
n
j =1
(j ) 2
Rd

Thus, the test is consistent provided there exists at least one j 1 such that j (x)  = 0 for some subset of Rd with
positive Lebesgue measure. Note however, that there could exist non-MDS such that j (x) = 0, ∀j 1, and our test will
not be able to detect such alternatives. An example of such alternatives is Yt = εt−1 εt−3 + εt , where εt ∼ i.i.d N (0, 1).
It is interesting to formally characterize the class of non-MDS that satisfies (5), but this is beyond the scope of this
paper.
Moreover, it can be shown that under mild assumptions our test based on Dn2 is asymptotically admissible, that is,
there does not exist a test that is uniformly more powerful than our test for a general class of local alternatives to
H0 , see Corollary 4 in Escanciano and Velasco (2006). Thus, the test proposed in this paper should be viewed not as
competing with but as a complement to that considered in Escanciano and Velasco (2006) since they have different
power properties and both are asymptotically admissible.
 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294 2285

5. Bootstrap approximation and finite sample performance

In this section we estimate the distribution of Sn (

) by that of
√

n−1
2 sin j 
Sn∗ (
) = ∗j (x)
(n − j )1/2 ,
j
j =1

with

n
   
∗j (x) = (n − j )−1
 Yt − Y n−j I Xt−j x − Fn−j (x) Wt ,
t=1+j


n
 
Fn−j (x) = (n − j )−1 I Xt−j x ,
t=1+j

and where {Wt } is a sequence of independent random variables with zero mean, unit variance, bounded support and
also independent of the sequence {Xt }nt=1 . We construct the bootstrap CvM norm as

∗2
 ∗ 2
Dn := Sn (, x) Fn (dx) d.

This procedure is similar to the wild bootstrap, see, e.g. Härdle and Mammen (1993). Its theoretical justification can
be found in Escanciano and Velasco (2006). In particular, we can simulate the critical values for the tests statistics Dn2
and computing uniform confidence bands for j (x) by the following algorithm:

1. Calculate the test statistic Dn2 with the original sample.

2. Generate Wt , a sequence of independent random variables with zero mean, unit variance, bounded support and that
is independent of the sample {Xt }nt=1 .
3. Compute ∗j (x), Sn∗ (
) and Dn∗2 .
4. Repeat steps 2 and 3, B times and compute the empirical (1 − )th sample quantile of Dn∗2 with the B values, Dn,
∗2 .

∗2 .
The proposed test rejects the null hypothesis at the significance level if Dn2 > Dn,

In order to examine the finite sample performance of the proposed test we carry out a simulation experiment with
several DGPs under the null and under the alternative. The emphasis in these simulations is on the comparison between
indicator-based tests for testing the classical MDH, that is, when Yt =Xt and d =1. The first block of models considered
here have been used by Dominguez and Lobato (2003) and thus, they will be useful for comparing both tests. We also
compare the tests with the conditional heteroskedasticity robust Durlauf’s (1991) test proposed by Deo (2000). 
We briefly describe our simulation setup. We denote Dn2 the new CvM test statistic defined in (8). Let Y =n−1 nt=1 Yt
be the usual sample mean.
Dominguez and Lobato (2003) have considered a MDH test taking into account a fixed number of lags. We denote
by CvM P and KS P the CvM and Kolmogorov–Smirnov statistics, respectively, with P as the number of lags used.
These statistics are based on the multivariate integrated regression function, i.e.
 n 2
1  
n
  
CvM P = 2 Yt − Y I zt,P 
zj,P
n
j =1 t=1

and


1  n
   
KS P = max √ Yj − Y I  zi,P ,
zj,P 
1i n n
j =1
2286 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294

where 
zt,P = (Yt−1 , . . . , Yt−P ) is the P-lagged values of the series. To save space, only results for P = 1 and 2 are
presented. Note that, CvM 1 and KS 1 are functionals of a process studied by Koul and Stute (1999) in a more general
context.
Recently, Deo (2000) has proposed a correction of the Durlauf’s (1991) test to take into account conditional het-
eroskedasticity. The corrected statistic is


n−1
2
1
DU RC := aj2
n ,
j
j =1

where
⎡ ⎤−1/2

n−j
 2  2
 j ⎣(n − j )−1
aj =  Yt − Y Yt+j − Y ⎦
t=1

and

n−j
  
−1

j = (n − j ) Yt − Y Yt+j − Y .
t=1

We have considered the factor (n − j )−1 in 

j instead of n−1 as used in Deo (2000) because it gives a better finite
sample performance. Note that we do not consider a kernel or weighting function. Under the null hypothesis of the
MDS and some additional assumptions (see Deo, 2000),
 1
d
DU RC −→ B 2 (t) dt as n −→ ∞,
0

where B(t) is the standard Brownian bridge on [0,1]. The 10%, 5% and 1% asymptotic critical values are obtained
from Shorack and Wellner (1986, p. 147) and are 0.347, 0.461 and 0.743, respectively, although we have also used in
the simulations empirical critical values. In the sequel εt ∼ i.i.d N (0, 1). The first block of models considered in the
simulations are two MDSs:

1. A sequence of i.i.d N (0, 1) variates.

2. GARCH(1,1) processes:
Yt = εt t , 2t = w + Yt−1
2
+ 2t−1 ,
with w = 0.001 and the following combinations for ( , ): (0.01, 0.97), (0.09, 0.89) and (0.09, 0.90), we call these
processes GARCH1, GARCH2 and GARCH3, respectively.

And the following nonmartingale difference sequences:

3. Nonlinear moving average (NLMA) process: Yt = εt−1 εt−2 (εt−2 + εt + 1).

4. Bilinear Processes:
Yt = εt + b1 εt−1 Yt−1 + b2 εt−1 Yt−2 ,
with (b1 , b2 ): (0.15, 0.05) and (0.25, 0.15), we call these process BIL-I and BIL-II, respectively.

Note that the second and third GARCH models have unbounded eighth and sixth moment, respectively. Also note
that the NLMA process is uncorrelated and, therefore, tests based on autocorrelations, see for instance Box and Pierce
(1970), Durlauf (1991) or Hong (1996), have no asymptotic power against this model.
We consider for the experiments under the null a sample size of n = 100 and under the alternative n = 100, 200
and 300. The number of Monte Carlo experiments is 1000 and the number of bootstrap replications is B = 500. In
all the replications 200 pre-sample data values of the processes were generated and discarded. Random numbers were
 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294 2287

Table 1
Size of tests
IID GARCH1

10% 5% 1% 10% 5% 1%

Dn2 9.9 5.0 0.8 10.2 4.7 0.9

CvM 1 9.3 4.7 0.8 9.7 5.1 0.7
KS 1 10.8 5.6 0.8 11.2 5.9 0.7
CvM 2 10.3 6.1 0.9 10.6 4.9 1.3
KS 2 11.5 6.5 1.9 10.4 6.2 1.2
10.8 5.0 1.1 10.8 4.8 0.9
DURC
[9.6] [4.2] [1.0] [9.1] [4.1] [0.9]

Table 2
Size of tests
GARCH2 GARCH3

10% 5% 1% 10% 5% 1%

Dn2 9.3 5.2 1.1 10.4 4.9 1.2

CvM 1 9.2 4.6 1.1 10.4 4.9 1.1
KS 1 10.1 5.4 1.3 11.0 6.1 1.3
CvM 2 10.8 5.9 0.9 9.5 4.9 0.9
KS 2 10.3 5.5 1.1 10.3 6.1 0.8
12.0 5.0 1.0 10.2 4.1 0.7
DURC
[8.8] [3.9] [0.6] [8.4] [4.1] [0.6]

Table 3
Power of tests
NLMA n = 100 n = 200 n = 300

10% 5% 1% 10% 5% 1% 10% 5% 1%

Dn2 28.0 17.9 3.7 37.5 24.5 9.3 49.1 35.5 14.2
CvM 1 26.4 16.1 3.6 37.1 24.5 9.9 49.6 35.6 14.1
KS 1 26.9 16.3 3.5 41.4 26.7 10.5 52.9 40.1 17.4
CvM 2 20.6 12.5 2.2 28.5 16.9 4.3 35.0 23.0 8.4
KS 2 21.5 11.3 2.4 32.9 21.9 6.4 41.9 28.8 11.8
15.0 6.6 0.9 14.2 7.1 0.9 14.0 7.1 1.5
DURC
[12.8] [6.1] [1.0] [11.8] [5.7] [0.8] [12.9] [6.2] [1.1]

generated
√ using
√ IMSL√ ggnml subroutine. We employ
√ √{Wt }√
a sequence of i.i.d Bernoulli variates where P (Wt = 0.5(1 −
5)) = (1 + 5)/2 5 and P (Wt = 0.5(1 + 5)) = 1 − (1 + 5)/2 5. Note that the third moment of Wt is equal to
1 and hence the first three moments of the bootstrap series coincide with the three moments of the original series, see
Stute et al. (1998).
In Tables 1 and 2 we show the empirical rejection probabilities (RP) associated with the three nominal levels
10%, 5% and 1%. The results for Dn2 , CvM 1 , KS 1 , CvM 2 , KS 2 and DURC show that all the tests have good size
properties and are robust to thick tails given their behavior with GARCH models. For DURC we also present in
brackets the RP using the empirical critical values based on simulations with i.i.d standard normal variables and 10 000
replications.
In Table 3 we report the empirical power against the NLMA process. It increases with the sample size n, as expected.
Generally, the Kolmogorov–Smirnov test has more empirical power than CvM tests. There is no test which dominates
uniformly the others, although the statistics CvM P and KS P for P 2 have less power as P increases. Deo’s (2000)
statistic, DU RC, has no power against this alternative, as expected, because this NLMA model is uncorrelated. Among
2288 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294

Table 4
Power of tests
BIL-I n = 100 n = 200 n = 300

10% 5% 1% 10% 5% 1% 10% 5% 1%

Dn2 18.6 9.9 1.9 30.2 18.5 4.4 45.8 26.3 6.7
CvM 1 19.2 9.8 1.8 29.2 17.2 4.0 46.9 26.5 7.1
KS 1 21.1 10.2 2.8 28.4 17.8 4.8 43.9 29.2 8.1
CvM 2 13.0 7.3 1.9 17.8 9.2 1.9 24.9 12.2 2.2
KS 2 14.4 7.3 1.5 19.5 10.7 2.6 26.5 15.9 5.1
13.7 6.1 1.7 11.8 6.1 1.1 15.8 9.0 2.4
DURC
[12.2] [5.3] [1.6] [13.5] [6.5] [1.8] [14.1] [8.5] [2.2]

Table 5
Power of tests
BIL-II n = 100 n = 200 n = 300

10% 5% 1% 10% 5% 1% 10% 5% 1%

Dn2 39.3 23.6 7.1 75.1 56.1 21.4 90.0 78.9 37.4
CvM 1 41.4 23.9 7.5 76.3 58.5 21.1 92.9 79.5 33.9
KS 1 43.8 28.5 10.5 72.9 56.8 26.5 88.3 77.1 42.6
CvM 2 22.6 12.1 2.1 40.8 22.0 4.3 61.1 38.1 9.3
KS 2 27.6 17.1 5.5 48.1 33.8 14.2 67.2 53.7 26.9
21.4 11.9 2.8 23.1 14.2 4.9 35.3 24.0 9.8
DURC
[19.7] [11.2] [2.5] [25.4] [17.0] [5.7] [33.6] [22.2] [9.3]

the statistics Dn2 , CvM 1 and KS 1 the difference is not substantial, because the NLMA process has dependence in the
conditional mean only for the first lags.
In Tables 4 and 5 we show the RP for the bilinear models. As in the case of the NLMA, there is no test which
dominates uniformly the others and, again, the statistics CvM P and KS P for P 2 have less power in almost all cases.
In the BIL-II case, CvM 1 and KS 1 perform slightly better than Dn2 , but the difference is not substantial. The empirical
power of DURC is lower in all cases.
In these examples the dependence is present only at the first lags, and therefore there is not too much practical
difference between Dn2 , CvM 1 and KS 1 . For models with higher order dependence structure is expected that Dn2 will
have more empirical power than CvM 1 and KS 1 because Dn2 considers all lags. To illustrate the latter fact, we consider
the following second block of models:

5. A linear ARMA(1,2) model:

(1 − 0.3L)Yt = (1 − 0.5L2 )εt ,

where L is the lag shift operator, i.e. LY t = Yt−1 .

6. A non-Gaussian moving average model (NGMA): Yt = exp(εt ) − 0.7 exp(εt−3 ).
7. A threshold autoregressive model (TAR)

Yt = 0.1Yt−3 − 0.5Yt−4 + εt if Yt−3 1,

Yt = −0.5Yt−3 + 0.4Yt−4 + εt if Yt−3 < 1.

The empirical power for a sample size of n = 100 and level 5% and the same design as Tables 3–5 is shown in
Table 6. Now, Dn2 has more empirical power than CvM P and KS P , P = 1 and 2, because of the higher lag dependence.
DURC has better empirical power properties than Dn2 for the ARMA(1,2) alternative, as expected, since there is no
gain in considering nonlinear dependence when testing against this linear model. On the contrary, our test is superior
to DURC for the NGMA and TAR alternatives. The conclusion for these higher order dependence cases is that is more
 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294 2289

Table 6
Power of tests
n = 100 ARMA(1,2) NGMA TAR

Dn2 74.9 23.0 16.6

CvM 1 35.2 8.3 7.9
KS 1 32.0 7.9 7.9
CvM 2 32.8 1.7 8.6
KS 2 36.7 3.7 11.9
DURC 83.0 14.1 14.2

preferred to sum-up pairwise information at various lags than consider many of them simultaneously. Though our
additive approach restricts the hypothesis to be tested (from (1) to (5)), it is able to break the curse of the dimensionality
that affects the statistics CvM P and KS P for moderate P.
Summarizing, Dn2 has omnibus good empirical power against all linear and nonlinear dependencies, independently
if the dependence is at high or low lags. In many cases it presents better empirical power properties than competing
tests while showing and overall power against all the alternatives considered for moderate sample sizes.

6. Exchange rates dynamics

In this section we investigate by means of our generalized spectral distribution MDH-test and the IPAF measures the
dynamics of the daily log price changes of the British Pound exchange rate in terms of the US Dollar exchange rate
(BPUSD). This problem has been explored in, e.g., Hsieh (1989), Gallant et al. (1991), Bera and Higgins (1997) and
examined recently by Dominguez and Lobato (2003), among others. The data consist in two samples, first from January
2nd, 1974 to December 31st, 1983 (BPUSD1) and the second period from December 12th, 1985 to February 28th,
1991 (BPUSD2). For a better comparison, we discard the 10% final observations in BPUSD2 as in Bera and Higgins
(1997). Then, the number of total observations for BPUSD1 and BPUSD2 are 2505 and 1210, respectively. The rates of
change are calculated by taking the logarithmic differences between successive trading days, i.e., Yt =100 log(rt /rt−1 ),
where rt denote the US Dollar price of a Pound at time t. Table 7 provides summary statistics of the data. The sample
distribution of the data has heavy tails in the two periods, specially in BPUSD1, and the sample kurtosis coefficients
are substantially larger than that of the standard normal distribution (which is 3). Earlier investigations focused on
the linear predictability (or lack thereof) of the exchange rates. Usually, exchange rates data are serially uncorrelated.
Here we use Deo’s (2000) test statistic to check if the data are uncorrelated. Previous results have shown that BPUSD1
have little linear dependence, see e.g. Table 2 in Hsieh (1989). This is in agreement with the p-values of DURC test in
Table 8. These results also show that BPUSD2 has not linear dependence, which is in agreement with the findings of
Bera and Higgins (1997).
There has been some evidence in the literature supporting that exchange rate changes exhibit nonlinear dependence,
see e.g. Hsieh (1989). We now investigate the type of the serial dependence presented in the BPUSD by means of
the IPAF and the generalized spectral distribution test. An important problem is to distinguish whether the serial
dependence affects the conditional mean or the conditional variance. The solution of this problem is crucial because it
has important implications in economics. Hsieh (1989) proposed a test based on third order cumulants to discriminate
between both types of dependence and found evidence in favor of the multiplicative dependence, that is, evidence
that the serial dependence affects the conditional variance, and rejecting the dependence in the conditional mean. To
test these hypotheses we use our generalized spectral test and present the results for the two periods in Table 8. To
facilitate interpretations we show the p-values for Dn2 , CvM P and KS P for P = 1 and 2. For BPUSD1, Dominguez
and Lobato (2003) found evidence for P = 1 against the MDH, but for P = 2 they supported the MDH. Note that this
is a contradictory result. The statistic Dn2 shows strong evidence against the MDH, which disagrees with the findings
of Hsieh (1989). These results can be explained by the fact that the data has third order cumulants equal to zero, but is
not a MDS. For the second period we find evidence supporting the MDH, with all the tests. Then, a nonlinear model
for the conditional mean can not explain the nonlinear dependence in the BPUSD2, in particular the bilinear model
used in Bera and Higgins (1997).
2290 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294

Table 7
Summary statistics of log price changes for 100 log(St /St−1 )
1974–1983 1985–1991

Mean −0.0185 −0.0227

Median 0.0000 0.0000
SD 0.5572 0.4766
Skewness −0.5145 0.0315
Kurtosis 8.4885 4.7584
Maximum 3.4344 1.9597
Minimum −3.8427 −2.2520

Table 8
p -values for the BPUSD

1974–1983 1985–1991

n 2505 1210
Dn2 0.002 0.943
CvM 1 0.004 0.896
KS 1 0.000 0.706
CvM 2 0.016 0.870
KS 2 0.000 0.673
DURC 0.230 0.860

Table 9
Generalized dependence measures for BPUSD1 1974–1983

Lag j KS Y |Y (j ) [q0.95 ] KS Y 2 |Y (j ) [q0.95 ] KS Y 2 |Y 2 (j ) [q0.95 ] KS Y 3 |Y (j ) [q0.95 ]

1 1.95a [1.31] 2.02a [1.29] 2.74a [1.24] 1.48a [1.21]

2 0.91 [1.39] 1.56a [1.31] 2.91a [1.35] 0.78 [1.24]
3 0.71 [1.39] 2.23a [1.26] 2.78a [1.34] 0.55 [1.18]
4 1.14 [1.34] 1.59a [1.26] 2.30a [1.29] 0.53 [1.25]
5 1.46a [1.37] 2.76a [1.38] 3.50a [1.41] 0.94 [1.32]
6 0.56 [1.41] 1.59a [1.32] 2.77a [1.31] 0.86 [1.17]
7 0.91 [1.31] 1.35a [1.29] 2.41a [1.28] 0.52 [1.21]
8 0.88 [1.32] 1.93a [1.26] 2.44a [1.26] 1.08 [1.20]
9 1.09 [1.33] 1.71a [1.30] 2.17a [1.25] 0.99 [1.21]
10 0.84 [1.38] 1.47a [1.25] 2.45a [1.26] 0.46 [1.18]
20 1.03 [1.34] 1.62a [1.30] 2.16a [1.27] 0.82 [1.16]
30 0.66 [1.33] 1.21 [1.26] 1.98a [1.37] 0.57 [1.16]
40 0.74 [1.38] 1.26 [1.31] 1.56a [1.26] 0.86 [1.19]
50 1.18 [1.31] 1.91a [1.38] 1.28a [1.26] 1.41a [1.20]
a Significantly different from zero at the 5% level (bootstrap test).

To gain insight in the serial dependence properties of the data we consider the IPAF for the two periods and also for
the squares of the data. In Tables 9 and 10 we show the Kolmogorov–Smirnov tests statistics for the IPAF at different
lags and different combinations of variables, and the bootstrap 95% quantile to indicate the significance of the IPAF for
the BPUSD1 and BPUSD2, see Section 2. Table 9 reveals that the nonlinearity in the conditional mean is significative
at lags 1 and 5 for the BPUSD1, confirming a weekly effect in this daily data set. The BPUSD1 also seems to be highly
heteroskedastic at all lags and conditionally asymmetric at lag j = 1. The results in Table 10 display a different behavior
for the BPUSD2, in which all the IPAF are not significant except for some integrated conditional variances.
We also plot in Figs. 2 and 3 the IPRF at lag j = 1 for the BPUSD1 data levels and squares, respectively (Yt and
Yt2 as dependent variables and Yt−1 as the conditioning variable). We also plot the corresponding uniform confidence
bands under the null of the MDH and under the alternative. The pattern of the IPAF for the data confirms some stylized
 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294 2291

Table 10
Generalized dependence measures for BPUSD2 1985–1991

Lag j KS Y |Y (j ) [q0.95 ] KS Y 2 |Y (j ) [q0.95 ] KS Y 2 |Y 2 (j ) [q0.95 ] KS Y 3 |Y (j ) [q0.95 ]

1 0.50 [1.37] 1.40a [1.33] 0.82 [1.27] 0.76 [1.33]

2 0.78 [1.31] 0.51 [1.23] 0.67 [1.31] 0.63 [1.18]
3 0.86 [1.36] 1.11 [1.27] 1.63a [1.29] 0.82 [1.25]
4 0.68 [1.35] 0.82 [1.27] 1.31 [1.33] 0.69 [1.26]
5 0.46 [1.31] 1.19 [1.32] 1.66a [1.30] 0.77 [1.21]
6 0.70 [1.33] 1.17 [1.38] 1.89a [1.37] 0.66 [1.26]
7 0.83 [1.34] 0.82 [1.31] 1.43a [1.34] 0.66 [1.24]
8 0.55 [1.30] 0.77 [1.33] 1.20 [1.32] 0.79 [1.27]
9 1.20 [1.28] 0.71 [1.30] 0.94 [1.25] 0.89 [1.20]
10 0.88 [1.33] 1.15 [1.31] 1.55a [1.30] 0.82 [1.29]
20 0.91 [1.31] 0.89 [1.33] 1.06 [1.34] 1.27 [1.30]
30 0.96 [1.29] 0.69 [1.31] 1.16 [1.33] 0.96 [1.30]
40 0.96 [1.30] 0.88 [1.31] 1.13 [1.34] 0.71 [1.28]
50 1.07 [1.34] 1.23 [1.37] 1.18 [1.30] 0.72 [1.30]
a Significantly different from zero at the 5% level (bootstrap test).

0.08

0.06

0.04

0.02

-0.02

-0.04

-0.06

-0.08

-0.1
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Fig. 2. IPAF for Yt BPUSD1 at lag j = 1 (conditioned on Yt−1 ).

facts about exchange rate changes. For instance, Fig. 3 shows the well-known “Leverage effect” in the exchange rates,
in which volatility is higher when past rates changes are negative. The IPAF for Yt3 as dependent variable, which has
not been shown for the sake of space, reveals that large changes in exchange rates are often negative.
On the other hand, Fig. 2 shows the IPAF for the BPUSD1 data and confirms that the MDH is rejected at lag j = 1,
although the autocorrelation at lag j = 1 is not significatively different from zero. This reveals that the conditional mean
at this lag cannot be linear and there is nonlinear dependence. Since the slope of the IPAF at each point is proportional
to the regression function we can conclude that the regression function has the same sign as the lagged exchange rate
change. This feature supports the well-known fact that the sample autocorrelation at lag j = 1 of exchange rates is
usually positive.
Summarizing, our new test finds nonlinear dependence in the conditional mean of the BPUSD1 exchange rate changes,
contrasting with some previous studies which assume that exchange rate changes are very nearly to be unpredictable
2292 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294

0.1

0.08

0.06

0.04

0.02

-0.02

-0.04

-0.06

-0.08
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Fig. 3. IPAF for Yt2 BPUSD1 at lag j = 1 (conditioned on Yt−1 ).

given past prices, and in agreement with more recent findings by Escanciano and Velasco (2006) and Hong and Lee
(2003). The nonlinearity in the conditional mean of BPUSD1 suggests that additional effort has to be dedicated to
investigate the form of such nonlinearity in the conditional mean before modeling the conditional variance.

Acknowledgments

We would like to thank W. Stute and M.A. Delgado for helpful comments and I. Lobato for his help with his
program and the data. The paper has also benefited from the comments of two referees. Research funded by the Spanish
Ministerio de Educación y Ciencia, Ref. SEJ 2004-04583/ECON and Ref. SEJ 2005-07657/ECON.

Appendix: Asymptotic results for the IPRF

j (·) in (4) under H0 (and mild regularity conditions.) Throughout

In this section we show the weak convergence of
the Appendix j 1 is fixed.

Assumption B1. B1(a):{Xt }∞ t=1 is a stationary ergodic vector process in R with probability distribution function F (·)
d

absolutely continuous
 with respect to Lebesgue measure. Also, Yt is a measurable real-valued function of Xt .
1+ 
4
B1(b):E Yt Xt−j < ∞, for some  > 0.
B1(c):The conditional density of Yt given Xt−j is (uniformly) bounded and continuous.

j (·).
The next proposition shows the null limit distribution of

1
Proposition 5. Suppose that Assumption B1 holds. Then, under (1) the process (n − j ) 2j (·) converges weakly to
Bj (·) on the Skorohod space D[−∞, ∞] , where Bj (·) is a Gaussian process with zero mean and covariance function
d

    
Kj x, x  = E (Y1 − )2 w1−j (x)w1−j x  ,

where wt (x) := {I (Xt x) − F (x)}.

 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294 2293

Proof of Proposition 1. First, write

1 
n


rj (x) = (Yt − ) I (Xt−j x) − F (x) .
n−j
t=1+j

Note that
⎡ ⎤⎡ ⎤
1 
n
1 
n
   
 rj (x) − ⎣
j (x) =  (Yt − )⎦ ⎣ I Xt−j x − F (x) ⎦ . (10)
n−j n−j
t=1+j t=1+j

Now, it is straightforward to show that under (1) for fixed K, 1 < K < n,

n
(n − j )−1/2 (Yt − ) = OP (1) ∀1 j < K,
t=1+j

whereas by stationarity, ergodicity and monotonicity of Fn−j (x) we have that the Glivenko–Cantelli Theorem holds,
i.e.

sup Fn−j (x) − F (x) = op (1) ∀1 j < K.
x∈Rd

Hence, for each fixed j, 1 j < K,

j (x) − 
sup | (n − j )1/2 { rj (x)} | =oP (1) as n → ∞.
x∈Rd

Therefore, it is sufficient to show the weak convergence of the process

rj (x) which follows from Lemma 2 in Dominguez
and Lobato (2004).

References

Auestad, B., TjZstheim, D., 1990. Identification of nonlinear time series: First order characterization and order determination. Biometrika 77,
669–687.
Bera, A.K., Higgins, M.L., 1997. ARCH and bilinearity as competing models for nonlinear dependence. J. Business and Economic Statist. 15,
43–50.
Bierens, H.J., 1984. Model specification testing of time series regressions. J. Econometrics 26, 323–353.
Bierens, H., Ploberger, W., 1997. Asymptotic theory of integrated conditional moment tests. Econometrica 65, 1129–1151.
Billingsley, P., 1968. Convergence of Probability Measures. Wiley, New York.
Box, G., Pierce, D., 1970. Distribution of residual autocorrelations in autoregressive integrated moving average time series models. J. Amer. Statist.
Assoc. 65, 1509–1527.
Chang, N.M., 1990. Weak convergence of a self-consistent estimator of a survival function with doubly censored data. Ann. Statist. 18, 391–404.
Chen, X., Fan, Y., 1999. Consistent hypothesis testing in semiparametric and nonparametric models for econometric time series. J. Econometrics
91, 373–401.
Chen, X., White, H., 1996. Laws of large numbers for Hilbert space-valued mixingales with applications. Econometric Theory 12, 284–304.
Chen, X., White, H., 1998. Central limit and functional central limit theorems for Hilbert-valued dependent heterogeneous arrays with applications.
Econometric Theory 14, 260–284.
Cochrane, J.H., 1988. How big is the random walk in GNP? J. Political Economy 96, 893–920.
de Jong, R.M., 1996. The Bierens’ test under data dependence. J. Econometrics 72, 1–32.
Deo, R.S., 2000. Spectral tests of the martingale hypothesis under conditional heroscedasticity. J. Econometrics 99, 291–315.
Dominguez, M.A., Lobato, I.N., 2003. A consistent test for the martingale difference hypothesis. Econometric Rev. 22, 351–377.
Dominguez, M.A., Lobato, I.N., 2004. Consistent estimation of models defined by conditional moment restrictions. Econometrica 72, 1601–1615.
Durlauf, S., 1991. Spectral based testing of the martingale hypothesis. J. Econometrics 50, 355–376.
Escanciano, J.C., 2006a. Model checks using residual marked empirical processes. Statist. Sin., forthcoming.
Escanciano, J.C., 2006b. Goodness-of-fit tests for linear and non-linear time series models. J. Amer. Statist. Assoc. 101, 531–541.
Escanciano, J.C., Velasco, C., 2006. Generalized spectral tests for the martingale difference hypothesis. J. Econometrics 134, 151–185.
Ferreira, E., Stute, W., 2004. Testing for differences between conditional means in a time series context. J. Amer. Statist. Assoc. 99, 169–174.
Gallant, A.R., Hsieh, D.A., Tauchen, G., 1991. On fitting a recalcitrant series: the pound/dollar exchange rate, 1974–1983. In: Barnett, W.A., Powell,
J., Tauchen, G. (Eds.), Nonparametric and Semiparametric Methods in Econometrics and Statistics. Cambridge University Press, Cambridge,
pp. 199–240.
2294 J.C. Escanciano, C. Velasco / Computational Statistics & Data Analysis 51 (2006) 2278 – 2294

Guerre, E., Lavergne, P., 2005. Rate-optimal data-driven specification testing for regression models. Ann. Statist. 33, 840–870.
Härdle, W., Mammen, E., 1993. Comparing nonparametric versus parametric regression fits. Ann. Statist. 21, 1926–1974.
Hong, Y., 1996. Consistent testing for serial correlation of unknown form. Econometrica 64, 837–864.
Hong, Y., 1999. Hypothesis testing in time series via the empirical characteristic function: A generalized spectral density approach. J. Amer. Statist.
Assoc. 84, 1201–1220.
Hong, Y., Chung, J., 2003. Are the directions of stock price changes predictable? Statistical theory and evidence. Working paper.
Hong, Y., Lee, T.H., 2003. Inference on predictability of foreign exchange rate changes via generalized spectrum and nonlinear time series models.
Rev. Econom. Statist. 85, 1048–1062.
Hong, Y., Lee, Y.J., 2005. Generalized spectral tests for conditional mean models in time series with conditional heteroskedasticity of unknown form.
Review Econom. Stud. 72, 499–541.
Hsieh, D.A., 1989. Testing for nonlinear dependence in daily foreign exchange rates. J. of Business 62, 339–368.
Koul, H.L., Stute, W., 1999. Nonparametric model checks for time series. Ann. Statist. 27, 204–236.
Li, Q., 1999. Consistent model especification tests for time series econometric models. J. Econometrics 92, 101–147.
Linton, O., Whang, Y-J., 2004. The quantilogram: with an application to evaluating directional predictability. J. Econometrics, forthcoming.
Politis, D., Romano, J., 1994. Limit theorems for weakly dependent Hilbert space valued random variables with application to the stationary bootstrap.
Statist. Sin. 4, 461–476.
Robinson, P.M., 1983. Nonparametric estimators for time series. J. Time Series Anal. 4, 185–207.
Shorack, G., Wellner, J., 1986. Empirical processes with applications to statistics. Wiley, New York.
Stinchcombe, M., White, H., 1998. Consistent specification testing with nuisance parameters present only under the alternative. Econometric Theory
14, 295–325.
Stute, W., 1997. Nonparametic model checks for regression. Ann. Statist. 25, 613–641.
Stute, W., Gonzalez-Manteiga, W., Presedo-Quindimil, M., 1998. Bootstrap approximations in model checks for regression. J. Amer. Statist. Assoc.
93, 141–149.
Tong, H., 1990. Nonlinear Time Series: A Dynamic System Approach. Clarendon Press, Oxford.
Whang, Y-J., 2000. Consistent bootstrap tests of parametric regression functions. J. Econometrics 98, 27–46.
Zheng, X., 1996. A consistent test of functional form via nonparametric estimation technique. J. Econometrics 75, 263–289.