Quantile Granger Causality in the Presence of

Instability
Alexander Mayer∗1 , Dominik Wied2 , and Victor Troster3
1
Università Ca’ Foscari
2 University of Cologne
arXiv:2402.09744v2 [econ.EM] 6 Dec 2024

3 Universitat de les Illes Balears

9th December 2024

Abstract

We propose a new framework for assessing Granger causality in quantiles in

unstable environments, for a fixed quantile or over a continuum of quantile
levels. Our proposed test statistics are consistent against fixed alternatives,
they have nontrivial power against local alternatives, and they are pivotal
in certain important special cases. In addition, we show the validity of a
bootstrap procedure when asymptotic distributions depend on nuisance pa-
rameters. Monte Carlo simulations reveal that the proposed test statistics
have correct empirical size and high power, even in absence of structural
breaks. Moreover, a procedure providing additional insight into the timing
of Granger causal regimes based on our new tests is proposed. Finally, an
empirical application in energy economics highlights the applicability of our
method as the new tests provide stronger evidence of Granger causality.

Keywords: Granger causality, Quantile regression, Parameter instability,

Structural breaks, Bootstrap.
JEL classification: C12, C22, C52.

∗
Corresponding author. E-mail address: alexandersimon.mayer@unive.it. We would like to thank
Lukas Hoesch for helpful comments and highly appreciate the suggestions and comments made by par-
ticipants at the Econometrics Workshop 2024 at the Goethe University Frankfurt, the research seminar
at Ca’ Foscari University, Venice, and the IAAE 2024, Thessaloniki. We would like to thank the associate
editor and two referees for valuable comments and suggestions that helped to improve the paper.
1 Introduction
The definition of Granger (1969) causality is a fundamental concept in time series
econometrics. Accordingly, let zi denote a series contained within an information
set that gathers all relevant information available up to time i, then zi is said to
Granger-cause yi if zi provides information relevant to predicting yi . Although
Granger causality is uncovered by the conditional distributions of yi , applied re-
search focuses often on Granger causality in mean because it entails easily testable
implications. However, by solely testing the significance of zi in a conditional mean
regression of yi on zi , one runs the risk of neglecting possible tail relationships or
nonlinearities.

It is for this reason that more recent research is also concerned with Granger
causality in quantiles, which allows for an equivalent characterization of Granger
causality in distribution. This implies that the conditional quantile function of yi
depends on zi for some quantiles of interest, given all the available information until
time i. For instance, Lee and Yang (2012) found fragile evidence of Granger causal-
ity between augmenting monetary policies and national income at the conditional
mean; nevertheless, the authors reported strong evidence of Granger causality at
extreme quantiles of the distribution.

One way to elicit potential evidence for Granger causality is by means of quantile
regressions. Koenker and Machado (1999) developed a parametric significance
test of quantile regression coefficients, which is frequently employed in empirical
work to test for Granger causality in quantile regressions (see e.g. Chuang et al.,
2009 or Yang et al., 2014). Troster (2018) extended the method of Koenker and
Machado (1999) by providing a semiparametric omnibus test for Granger causality
in quantiles that allows for nonlinear specifications of the quantile regressions under
the null hypothesis of no Granger causality. On the other hand, Jeong et al. (2012),
Taamouti et al. (2014) and Candelon and Tokpavi (2016) derived nonparametric
tests for Granger causality in quantiles. Bouezmarni et al. (2024) proposed such
a test for expectiles.

What all these papers have in common, is, however, that they implicitly assume
the pattern of Granger (non)causality to be stable over time. In this paper, we

1
therefore propose tests for Granger causality in quantiles that are robust against
temporal instabilities. This is of importance because financial and macroeconomic
data structures, where Granger causality is frequently tested, are subject to strong
fluctuations and volatility (Clark and McCracken, 2006; Rossi, 2005, 2013; Rossi
and Wang, 2019; Stock and Watson, 1996, 1999, 2003, 2006; Baum et al., 2021,
forthcoming). Rossi (2006), for instance, provides evidence of failure of traditional
Granger causality tests to detect Granger causality from certain macroeconomic
fundamentals to exchange rate fluctuations due to parameter instabilities in the
models over time. Chen et al. (2010), on the other hand, do report evidence
of Granger causality from exchange rates to commodity prices, when applying
Granger causality tests that allow for structural breaks. Caporin and Costola
(2022) have made similar arguments.

In addition, Giacomini and Rossi (2010), Rossi (2013), and Rossi (2021), among
others, show that instabilities in the parameters of the models can affect the perfor-
mance of Granger causality tests in different ways; hence, these authors recommend
incorporating structural breaks in Granger causality tests rather than testing for
instability in the parameters. Following this idea, it is important to apply meth-
ods that are robust to structural breaks or instabilities for correctly performing a
Granger causality analysis in macroeconomic or financial time series.

To address potential temporal instabilities, we resort to the work of Rossi (2005),

who –by extending the earlier work by Sowell (1996)– developed tests for nested
model selection with underlying parameter instability. Although the methods of
Rossi (2005) and, in particular, Rossi and Wang (2019) can be used to test for
Granger causality in mean between two time series, they fall short to capture
Granger causality in the tails or other parts of the conditional distribution not
captured by the mean. The same is true for the time-varying Granger causality
in-mean tests like the one employed, for example, by Chen et al. (2010) or Caporin
and Costola (2022).

We thus extend the method of Rossi (2005) and Rossi and Wang (2019) in a
consolidated way for testing for Granger causality in quantiles under structural
instabilities. We do so by drawing from results on structural break testing in

2
quantile regressions by Qu (2008) and Oka and Qu (2011); see also Hoga (2024) for
a recent extension to predictive quantile and CoVaR regression. To our knowledge,
no test for Granger causality in quantiles with structural instability has been
developed so far in the literature.

More specifically, the main idea is to consider under the null the intersection of
two sub-hypotheses: First, we hypothesize that the effect of the potential Granger-
causing variate is constant over time and over quantiles. Second, we assume that
this effect is zero. Likewise, the alternative in a local neighbourhood around the
null consists of two parts: one that specifies local deviations from the null hypoth-
esis of no Granger causality in quantiles and another that specifies local deviations
from the null hypothesis of constant parameters (over one or more quantiles of the
distribution). Thus, we construct our test statistics in such a way that ensure non-
trivial local power against the union of these two alternative hypotheses. For this
reason, we do not require a priori knowledge of whether any of the two alternative
sub-hypotheses holds (or whether both hold). Finally, we propose tests that nei-
ther involve trimming over time nor require the specification of tuning parameters.
Here we distinguish two cases, on the one hand, we have a single quantile, on the
other hand, we consider a continuum of quantiles (as a subset of the interval [0, 1]).

This idea goes back to, inter alia, Sowell (1996) and Rossi (2005) who have devel-
oped similar test procedures in a general generalized method of moments (GMM)
framework. Our extension is non-trivial. Unlike the GMM framework considered
by Rossi (2005), we demonstrate, for example, that the limiting distribution might
not be pivotal in certain cases. It is pivotal in certain important special cases such
as homoskedasticity or conditional mean independence between regressors. In
cases where the limiting distribution depends on nuisance parameters, we propose
alternatives, including a new bootstrap procedure whose validity is established.
This extends the semiparametric bootstrap used in Rothe and Wied (2013), where
the estimated quantile functions are applied to uniformly distributed random vari-
ables. Therefore, we provide (bootstrap) test statistics with correct asymptotic
size, which are consistent against fixed alternatives and possess nontrivial power
against local alternative hypotheses. This is also corroborated by our Monte Carlo
simulations; the finite-sample evidence shows that the test has appealing size and

3
power properties in finite samples. In all cases (structural breaks and no structural
breaks under the alternative), our tests are more powerful than the existing sup
Wald test. Given that our new test rejects, the question arises how to interpret the
result. To address this question, we propose an additional two-step procedure to
further characterize the data generating process. After applying a plain CUSUM
test for detecting and dating a structural break, we apply our new test in the
detected regimes. So, we get precise information about in which parts of the time
period we have Granger causality.

We illustrate the applicability of our tests by performing an empirical applica-

tion. We revisit an analysis about the causal relationships between crude oil and
stock returns from Ding et al. (2016), who consider the interplay between stock
returns and crude oil returns. In the application, we find several scenarios which
demonstrate the higher efficiency of the new test. Moreover, using our sequential
procedure, we are able to identify economically interpretable regimes of Granger
causality.

The rest of the paper proceeds as follows. In Section 2, we propose our test statis-
tics for jointly testing for Granger causality in quantiles and parameter instability.
In Section 3, we derive the asymptotic distribution of our test statistics; we also
propose and justify a bootstrap method for implementing our test statistics. In
Section 4, we perform Monte Carlo simulations to validate the finite-sample per-
formance of our test statistics. In Section 5, we present an empirical application
of our proposed tests. Finally, we conclude the paper in Section 6.

Throughout the paper, we use the following notation: Bm (λ), λ ∈ [0, 1], is a vector
of m independent Brownian motions, and BB m (λ) := Bm (λ) − λBm (1) is a vector
of m independent Brownian bridges. For a positive definite matrix A, A−1/2 is
defined as the Cholesky factor of its inverse A−1 , so that A−1 = (A−1/2 )′ A−1/2 .
Notation “⇒” and “→d ” indicates weak convergence and convergence in distribu-
tion, respectively. Notation T represents a closed interval such that T ⊂ [0, 1].
For an m × 1 vector z, we define ∥z∥∞ := max |zj |.
1≤j≤m

4
2 Granger Causality
Suppose we suspect that the p × 1 vector zi Granger causes the dependent variable
yi , and, at the same time, we have reasons to question temporal stability. In other
words, we expect Granger causality, but we are unsure whether its pattern persists
over time.

Within the framework of a linear quantile regression, these considerations amount

to parametrise the τ quantile of yi via

Qyi (τ | xi ) := x′i βi (τ ), xi := (zi′ , wi′ )′ , βi (τ ) := (γi (τ )′ , α(τ )′ )′ ∈ Rm , (1)

where w is a k × 1 vector of additional controls so that m = p + k, γi (τ ) ̸= 0

for some τ ∈ [0, 1], and i ∈ {1, . . . , n} under Granger causality. For simplicity, we
assume that all the available information up to time i can be represented by vector
xi . For example, xi might be equal to (yi−1 , ai , ai−1 ) for some univariate time series
ai . Two important models, which lead to such a structure, are the location-scale
model, yi = x′i δ + (x′i ρ)εi , and the random coefficient model, yi = x′i β(Ui ), with
a standard uniformly distributed random variable Ui that includes, among others,
quantile autoregressive distributed lag models (see e.g. Koenker, 2005 and Galvao
et al., 2013).

We formulate the following (joint) null hypothesis H0 := H0,1 ∩ H0,2 :

H0,1 := {γi (τ ) = γ0 (τ ), ∀ i ∈ {1, . . . , n}, τ ∈ T }

H0,2 := {γ0 (τ ) = 0p , ∀ τ ∈ T }

against the alternative hypothesis H1 := {¬H0,1 } ∪ {¬H0,2 }. More specifically, in

a local neighbourhood around H0 in the direction of H1 , the following sequence of
local alternatives is investigated

δ(τ ) i ∆(τ )
 
γi,n (τ ) := γ0,n (τ ) + √ g , γ0,n (τ ) := √ , (2)
n n n

where τ 7→ ∆(τ ) and τ 7→ δ(τ ) are deterministic continuous vector-valued and

scalar-valued functions, respectively, and v 7→ g(v) is a deterministic vector-valued

5
Riemann–Stieltjes integrable function.

Our setup is essentially similar to the approach of Rossi (2005) that extends the
earlier procedure of Sowell (1996), in which the null hypothesis consists of two
different restrictions. On the one hand, the parameter γi (τ ) is constant over i and
τ ; on the other hand, this constant is equal to 0. Analogously, the alternative
hypothesis also consists of two parts. The alternative H1,1 := ¬H0,1 specifies local
deviations from the null hypothesis of constant parameters, whereas the alternative
H1,2 := ¬H0,2 specifies local deviations from the null hypothesis of no Granger
causality in quantiles. Our tests are constructed in such a way that they have
power against the union of these alternatives. For this purpose, it is not required
to know a priori which of the two alternatives (or both) holds.

3 Test Statistics
Consider the following sequential process based on the subgradient of the uncon-
strained quantile regression

⌊λn⌋
Sn (λ, τ, t) := n−1/2 xi ψτ (yi − x′i t), t ∈ Rm ,
X
(3)
i=1

where λ ∈ [0, 1] indexes the time fraction, and ψτ (u) := 1{u ≤ 0} − τ . Moreover,
introduce the (constrained) estimator β̃n (τ ) := (0, αn (τ )′ )′ , where
n
ρτ (yi − wi′ α),
X
αn (τ ) := arg min ρτ (u) := u(1{u ≤ 0} − τ ). (4)
α∈Rk i=1

Our tests are based on the following process

Hn (λ, τ, t) := (Xn′ Xn /n)−1/2 Sn (λ, τ, t), λ, τ ∈ [0, 1], t ∈ Rm , (5)

where Xn := (x′1 , . . . , x′n )′ is n × m. As pointed out by Qu (2008), the process in

(5) is asymptotically pivotal when evaluated at the true parameter vector; see also
Parzen et al. (1994) for a similar argument. This allows us to construct tests that
do not require trimming over time.

6
The main idea behind our test statistic is to combine two detectors that are re-
spectively designed to find deviations from H0,1 and H0,2 . More specifically, a
CUSUM-type statistic
 
Ip 
LM1 (λ, τ ) := R′ ∆Hn (λ, τ, β̃n (τ )) , R :=  ,
∞ 0k×p

with
∆Hn (λ, τ, β̃n (τ )) := Hn (λ, τ, β̃n (τ )) − λHn (1, τ, β̃n (τ ))

is used to test H0,1 , while the restriction of H0,2 is verified using the LM (Lagrange
Multiplier) statistic
LM2 (τ ) := R′ Hn (1, τ, β̃n (τ )) .
∞

Since LM1 has no power against constant deviations from the null, and LM2 lacks
power if Granger causality is unstable, our tests will be of the form
 
φ LM1 + LM2 , (6)

for some weighting function φ : [0, 1] × [0, 1] 7→ R specified below.

In what follows, we distinguish between situations where our interest lies in detect-
ing deviations from the null (i) at a given quantile or (ii) across various quantiles.
To derive the properties of theses tests, we impose the following assumptions that
are similar to those in Qu (2008) and Oka and Qu (2011).

Assumption A Let ui,n (τ ) := yi − βi,n (τ )′ xi . Then 1{ui (τ ) ≤ 0} − τ is a martin-

gale difference array with respect to Fi−1 := σ({yj−1 , xj : j ≤ i}) for any τ ∈ [0, 1].

Let fi (·), Fi (·) and Fi−1 (·) denote the conditional density, conditional distribution,
and conditional quantile function, respectively, of yi given wi .

Assumption B .

B.1 The distribution functions Fi (·) are absolutely continuous, with continuous
densities fi (·) satisfying 0 < u ≤ fi (Fi−1 (τ )) ≤ ū < ∞ for all i.
¯
B.2 For any ϵ > 0, there exists a σ(ϵ) > 0 such that |fi (Fi−1 (τ )+s)−fi (Fi−1 (τ ))| ≤

7
 ϵ for all |s| < σ(ϵ) and all 1 ≤ i ≤ n.

Assumption C The regressors are assumed to satisfy: .

C.1 The vector w contains a constant.

P⌊λn⌋
C.2 plimn→∞ n1 i=1 fi (Fi−1 (τ ))xi x′i = λH(τ ) uniformly in λ ∈ [0, 1], where H(τ )
is a m × m non-random positive definite matrix.

C.3 There exists a > 0 and A < ∞ such that E[∥xi ∥4+a ] ≤ A.

C.4 There exists b > 0 and B < ∞ such that for any n:
n n
1X 3(1+b) 1X
E[∥xi ∥ ] ∨ E[ ∥xi ∥3 ]1+b ≤ B.
n i=1 n i=1

P⌊λn⌋
C.5 plimn→∞ n1 i=1 xi x′i = λJ uniformly in λ ∈ [0, 1], where J is a m × m non-
random positive definite matrix.

These assumptions are standard in the context of tests for structural breaks in
quantile models and of tests for Granger causality in quantiles. Assumption A
restricts the dependence over time. Serial independence is not required, instead
we have a martingale difference assumption on the innovations. Assumption B
introduces positivity and smoothness assumptions on the conditional density of yi
given xi . Assumption C imposes restrictions on the regressors xi , in particular on
the existence of moments. This assumption rules out trends in the regressors, but
it allows for heteroscedasticity. It might be interesting to relax this assumption
in future work to allow also for frequency-dependent regressors as, for example,
in Li (2012) in order to test for Granger causality at different frequencies akin to
Breitung and Candelon (2006).

3.1 Granger Causality at a Given Quantile

Let β0 (τ ) := (0′p , α0 (τ )′ )′ be the true coefficient under the null and define C(τ ) :=
J−1/2 H(τ ). Note that C(τ ) is a square root of the inverse of the variance-covariance
matrix of the limiting distribution of the estimator that solves the unrestricted
quantile regression problem (see, e.g. Koenker, 2005). Moreover, let Hα (τ ) and

8
Jα , denote the lower-right k × k block of H(τ ) and J, respectively, and partition
β0 (τ ) = (γ0 (τ )′ , α0 (τ )′ )′ , where, under the null, γ0 (τ ) = 0p .

The limiting distribution of the restricted quantile estimator, and the two detectors
can now be summarized as follows.

Proposition 1 Assume that Assumptions A, B, and C hold. For a given τ ∈ T ,

√
C(τ ) n(β̃n (τ ) − β0 (τ )) = −P(τ )J−1/2 Sn (1, τ, β0 (τ )) + op (1),

where  
0p×k 
P (τ ) := C(τ )R̄(R̄′ C(τ )R̄)−1 R̄′ , R̄ :=  .
Ik
Moreover, it holds that uniformly in λ ∈ [0, 1]
Z λ !
−1/2
h(τ )J Sn (λ, τ, β0 (τ )) ⇒ Bm (λ) + h(τ )C(τ )R λ∆(τ ) + δ(τ ) g(v)dv ,
0

h2 (τ ) := 1/(τ (1 − τ )), so that

h(τ )R′ ∆Hn (λ, τ, β̃n (τ ))

Z λ Z 1 !
′
⇒ BB p (λ) + h(τ )δ(τ )R C(τ )R (1 − λ) g(v)dv − λ g(v)dv
0 λ

=: Z (1) (λ, τ ),

and

h(τ )R′ Hn (1, τ, β̃n (τ ))

 Z 1 
′
⇒ B̃p (1, τ ) + h(τ )R T (τ )C(τ )R ∆(τ ) + δ(τ ) g(v)dv =: Z (2) (τ ),
0

where B̃m (λ, τ ) := T (τ )Bm (λ), with T (τ ) denoting the inverse of the m × m matrix
of eigenvectors of I − P (τ ).

Interestingly, and contrary to the corresponding GMM result in Rossi (2005), the
limiting distribution of the LM statistic Hn (1, τ, β̃n (τ )) is not pivotal because the
projection matrix P (τ ) is oblique (i.e. idempotent of rank k but not symmetric).

9
An important exception is given if the following additional condition is satisfied:

Assumption D The p × k matrix Q(τ ) := R′ C(τ )R̄H−1 1/2

α (τ )Jα is zero.

Assumption D ensures that the oblique projection matrix P (τ ) defined in Propo-

sition 1 is equal to the orthogonal projection R̄(R̄′ R̄)−1 R̄′ = R̄R̄′ , which follows
from observing that P (τ ) decomposes into an orthogonal projection perturbed by
a nilpotent matrix
 
0p×p Q(τ )
P (τ ) = C(τ )R̄(R̄′ C(τ )R̄)−1 R̄′ = R̄R̄′ +  . (7)
0k×p 0k×k

A sufficient condition for Assumption D is Hc(τ ) = J(τ ) for some scalar c(τ ) ∈
(0, ∞), which holds under homoscedasticity. Alternatively, Assumption D is satis-
fied if w is just a constant or, more generally, under conditional mean independence
of z with respect to w (i.e. E[z | w] = E[z]) as both of theses conditions ensure
under the null Q(τ ) = 0p×k .

Corollary 1 Under the null hypothesis and Assumptions A–D, we get for a given
τ ∈ T and uniformly in λ ∈ [0, 1]
 
Bp (λ) 
h(τ )Hn (τ, λ, β̃n (τ )) ⇒  ,
BB k (λ)

where BB k (·) and Bp (·) are independent.

The p-dimensional Brownian motion Bp and the k-dimensional Brownian bridges

BB k arise due to the restricted and unrestricted components of the process tow-
parameter process Hn (λ, τ, β(τ )), respectively. The limiting random variable under
the null is independent of τ . While this is true if Assumption D holds, violations
from this assumption introduce dependence on nuisance parameters (cf. Proposi-
tion 2).

Based on the previous result, we will now introduce our first test statistic, suited

10
to test H0 at a given quantile τ
!
LM(τ ) := h(τ ) sup LM1 (λ, τ ) + LM2 (τ ) .
λ∈[0,1]

For fixed τ , the test statistic essentially consists of the sum of two individual test
statistics, which reflect the two parts of the alternative hypothesis. Both statistics
are based on the standardized subgradient of the unconstrained quantile regression
through the process Hn (λ, τ, t) from (5). The first part, LM1 (λ, τ ), is the CUSUM
part that detects structural breaks in the parameter γi (τ ). Typically, for CUSUM
statistics, one considers the supremum over the potential breakpoints λ ∈ [0, 1].
The second part, LM2 (τ ), is essentially the LM statistic for the hypothesis H0,2 .

Corollary 1 states why it makes sense to consider the sum of the two individual
test statistics: The first statistic does have local power against structural breaks,
but it has no power if there is Granger causality with constant parameters. The
second part has power if there is Granger causality with constant parameters, but
it has no power if ∆(τ ) = 0 and 01 g(v)dv = 0. The interpretation of the latter
R

would be that there are structural breaks that lie in opposite directions over time.

Corollary 2 For a given τ ∈ T , we get under the assumptions of Proposition 1

LM(τ ) →d sup Z (1) (λ, τ ) + Z (2) (τ ) ,

λ∈[0,1] ∞ ∞

while under the null

LM(τ ) →d sup ∥BB p (λ)∥∞ + B̃ p (1, τ ) .

λ∈[0,1] ∞

where BB p and B̃p (λ, τ ) are independent. If Assumption D holds, then B̃p (λ, τ ) =
Bp (λ).

Thus, unless Assumption D is satisfied, the limiting distribution is not pivotal due
to the second element LM2 of our test statistic that induces dependence on the

11
quantile level τ via Q(τ ). More specifically, it can be shown that

LM2 (τ ) = R′ (I − P (τ ))J−1/2 Sn (1, τ, β0 (τ )) + op (1),

∞

where the oblique projection P (τ ) causes quantile dependece because it cannot be

diagonalized; cf. Eq. (7). The distribution of LM2 can be viewed as a maximum
of p-scaled absolute standard normals, where–similar to the discussion in Hansen
(2021)–the scaling differs in general from unity, thereby capturing deviations from
Assumption D (e.g. from homoskedasticity to heteroskedasticity).

3.2 Granger Causality at all Quantiles

To avoid multiple testing issues when performing inference across various quantiles,
we extend the test statistics from the previous section to allow uniform inference
across both λ and τ . Following Andrews and Ploberger (1994) and Hansen (1996),
we consider the following test statistics:
!
supLM := sup sup LM1 (λ, τ ) + LM2 (τ ) ,
τ ∈T λ∈[0,1]
" !# (8)
Z
1
expLM := exp sup LM1 (λ, τ ) + LM2 (τ ) dτ.
T 2 λ∈[0,1]

While both weighting schemes (over τ ) direct power against relatively distantly
located alternatives, expLM can be considered optimal (see also Rossi, 2005).

Similar to Proposition 1, we first derive the properties of the process Eq. (5) that
serves as the building block of our test statistics. To this end, define Sem (τ, λ) :=
T (τ )Sm (τ, λ), T (τ ) defined in Proposition 1, Sm (λ, τ ) := (S1m (λ, τ ), . . . , Smm (λ, τ ))′
is an m × 1 vector of independent Gaussian processes with

cov[Sim (λ1 , τ1 ), Sim (λ2 , τ2 )] = (λ1 ∧ λ2 )(τ1 ∧ τ2 − τ1 τ2 )

and SS m (λ, τ ) := Sm (λ, τ ) − λSm (1, τ ) so that

cov[SS im (λ1 , τ1 ), SS im (λ2 , τ2 )] = (λ1 ∧ λ2 − λ1 λ2 )(τ1 ∧ τ2 − τ1 τ2 ).

12
Proposition 2 Assume that Assumptions A, B, and C hold uniformly in τ . Then,
uniformly in τ ∈ T , we have
√
C(τ ) n(β̃n (τ ) − β0 (τ )) = −P(τ )J−1/2 Sn (1, τ, β0 (τ )) + op (1).

Moreover, it holds that uniformly in (τ, λ) ∈ T × [0, 1]

Z λ !
J−1/2 Sn (λ, τ, β0 (τ )) ⇒ Sm (λ, τ ) + C(τ )R λ∆(τ ) + δ(τ ) g(v)dv ,
0

so that
Z λ Z 1 !
′ ′
R ∆Hn (λ, τ, β̃n (τ )) ⇒ SS p (λ, τ ) + δ(τ )R C(τ )R (1 − λ) g(v)dv − λ g(v)dv
0 λ

=: Y (1) (λ, τ )

and
 Z 1 
′ ′
R Hn (1, τ, β̃n (τ )) ⇒ Se p (1, τ ) + R T (τ )C(τ )R ∆(τ ) + δ(τ ) g(v)dv =: Y (2) (τ ).
0

Again, the weak limit of Hn (1, τ, β̃(τ )) is affected by nuisance parameters un-
less Assumption D holds, in which case Sep (λ, τ ) = Sp (λ, τ ). Moreover, note that
S(1, τ ) = B(τ ), while the Gaussian process SS(λ, τ ) is also referred to as a Brown-
ian pillow or a pinned Brownian sheet; see also Qu (2008, Sec 4). From the above,
the limiting distribution of the test statistics follows readily by the continuous
mapping theorem:

Corollary 3 Uniformly in (τ, λ) ∈ T × [0, 1], we have under the conditions of

Proposition 2
!
(1) (2)
supLM →d sup sup Y (λ, τ ) + Y (λ, τ ) ,
τ ∈T λ∈[0,1] ∞ ∞
" !#
Z
1 (1) (2)
expLM →d exp sup Y (λ, τ ) + Y (λ, τ ) dτ,
T 2 λ∈[0,1] ∞ ∞

while, under the null, Y (1) (λ, τ ) = SS p (λ, τ ) and Y (2) (λ, τ ) = Sep (1, τ ). If Assump-

13
tion D holds, then Sep (λ, τ ) = Sp (λ, τ ).

3.3 Practical Implementation

If Assumption D is satisfied, then it is easy to simulate the limiting distributions
because they are free of unknown nuisance parameters (cf. Corollaries 2 an 3).
More specifically, using numerical techniques we can arbitrarily well approximate
the Brownian motion B(·) and the Brownian bridge BB(·) for the fixed-τ case
or, if instead a continuum of quantiles is considered, the limiting processes S(·)
and SS(·) (see the discussion in Andrews (1993) and Qu (2008) for details on the
numerical computation).

If Assumption D is violated, then we could still simulate the limiting distributions.

But since the weak limits depend in this case on the characteristics of the dgp (cf.
Corollaries 2 an 3), we need to tabulate critical values for each application sepa-
rately. In principle, one could proceed as follows: Firstly, based on a consistent esti-
mator Hn (τ ), say, of H(τ ), we estimate Q(τ ) using Qn (τ ) := Jn−1/2 Hn (τ )Hn,α (τ )Jn,α
1/2
,
′
with Jn := X X/n, and where Hn,α (τ ) and Jn,α denote, respectively, the sample
analogues of Hα (τ ) and Jα from Assumption D. For instance, a consistent estimator
can be obtained via
n
1 X
Hn (τ ) = 1{ûi (τ ) ≤ cn }xi x′i ,
2ncn i=1
√
where cn → 0, ncn → ∞ (see Powell, 1991 and Koenker, 2005, Sec 3.4). Secondly,
we obtain from Qn (τ ) and Eq. (7) the inverse matrix of eigenvectors Tn (τ ), which,
by the continuous mapping theorem, is a consistent estimator. Finally, we simulate
the limiting stochastic processes similarly to the case where Assumption D is met,
but we substitute B(·) (S(·)) with B(·)
e (S(·)).
e Clearly, this procedure becomes
very time-consuming when testing at many quantiles. Nevertheless, an important
exception is given for p = 1 < k,1 where proper scaling of the test statistics ensures
1
Note that the case k = 1 is trivial because then w = 1 and assumption D is automatically
satisfied.

14
a pivotal limiting distribution; e.g. for a given τ ∈ T

LM2 (τ ) q
h(τ ) q →d |B1 (1)| ≡ χ2 (1).
1 + Qn (τ )Qn (τ )′

Unfortunately, a similar re-scaling does not work for other values p > 1. Therefore,
we propose an additional resampling procedure that especially for the ‘many-τ ’
case is significantly less time consuming.

In particular, we propose a bootstrap procedure which is inspired by Rothe and

Wied (2013) and that is valid both in the ‘fixed-τ ’ case and the ‘many-τ ’ case.
Note, however, that we actually need the bootstrap only in the former case because
the computational burden of the procedure described above for the ‘fixed-τ ’ case
is manageable. The algorithm for obtaining one bootstrap sample {(ybi,b , xi,b ), 1 ≤
i ≤ n}, b ∈ {1, . . . , B}, for a large value of B is as follows:

Algorithm 1.

Step 1 Draw with replacement {xi,b , 1 ≤ i ≤ n} from the realized regressors

{xi , 1 ≤ i ≤ n}

Step 2 For each 1 ≤ i ≤ n, set

ybi,b = αn (Ui,b )′ wi,b ,

where {Ui,b , 1 ≤ i ≤ n} is a simulated IID sequence of standard uniformly

distributed random variables on the interval (0, 1), αn is the restricted
′ ′ ′
quantile estimator and xi,b = (zi,b , wi,b ) with the same dimensions as in
the realized data.

Step 3 Use the bootstrap data {(ybi,b , xi,b ), 1 ≤ i ≤ n} to obtain bootstrap esti-
mates Hn,b (λ, τ, β̃n,b (τ )), say, of Hn (λ, τ, β̃n (τ )) and construct the corre-
sponding test statistics for τ ∈ T .

Step 4 Our tests reject if they exceed the corresponding bootstrap critical values
cb(α), say, for some α ∈ (0, 1).

The algorithm above means that, for generating a bootstrap sample, we first draw
with replacement from the regressors, where random sampling is justified by As-

15
sumption A (i.e. Corollary 2 and Corollary 3 are unaffected by the temporal
dependence of xi ). The corresponding y-values are obtained by applying the esti-
mated quantile function on randomly chosen standard uniformly distributed ran-
dom variables. Thus, it is ensured that, under the null hypothesis, we asymptoti-
cally generate data from the distribution of (y, x) with Qy (τ |x) = x′ β(τ ) so that,
in Step 3, no centring of Hn,b (λ, τ, β̃n,b (τ )) is needed. Here, it is crucial to draw
from a uniform distribution on the whole interval (0, 1) in Step 2 to get simulated
data from the whole conditional distribution of y given x, although the interval T
is a strict subset of (0, 1). So, the restriction is stronger than actually necessary,
but the null distribution is still enforced. Under the alternative hypothesis, the
critical values remain stochastically bounded as the validity of the null hypothesis
is enforced within the generation of the bootstrap sample.

These considerations are summarized in Proposition 3. For simplicity, consider

the ‘many-τ ’ case and let us generically represent our test statistics in Eq. (8) as
φ(LM1 + LM2) using the weighting function φ : [0, 1] × T 7→ R from (6).

Proposition 3 . Let α ∈ (0, 1) and assume that Assumptions A, B, and C hold

uniformly in τ ∈ T .

(i) Under the null hypothesis

P(φ(LM1 + LM2 ) ≥ ĉ(α)) → α.

(ii) Under fixed alternatives

P(φ(LM1 + LM2 ) ≥ ĉ(α)) → 1.

3.4 Identifying Regimes of Granger Causality

If the null hypothesis H0 is rejected by the new tests, it might be interesting to
get more information about the reason for the rejection. It is possible that H0,1
is violated, i.e. that the parameters are not constant over time, and it is possible
that H0,2 is violated, i.e. that the parameters are not 0 for all time points. To get
more information, we suggest to apply an iterative procedure, in the case that H0

16
is rejected in the first step.

To be more specific, we follow the discussion in Oka and Qu (2011) and introduce
to this end the scaled subgradient
 −1
⌊bn⌋ ⌊λn⌋
xi x′i  xi φτ (yi − x′i t),
X X
Ha,b,n (λ, τ, t) := 
i=⌊an⌋+1 i=⌊an⌋+1

defined on λ ∈ [a, b] ⊂ [0, 1], while β̃a,b,n (τ ) := (0, αa,b,n (τ )′ )′ denotes the con-
strained estimator on the restricted sample, where

⌊bn⌋
ρτ (yi − wi′ α).
X
αa,b,n (τ ) := arg min
α∈Rk i=⌊an⌋+1

Next, define (supLMa,b could be defined analogously)

" !#
Z
1
expLMa,b := exp sup LMa,b,1 (λ, τ ) + LMa,b,2 (τ ) dτ, (9)
T 2 λ∈[a,b]

where the subsample detectors for violations of H0,1 and H1,1 are respectively given
by
LMa,b,1 (λ, τ ) := R′ Ha,b,n (λ, τ, β̃a,b,n (τ )) − λR′ Ha,b,n (b, τ, β̃a,b,n (τ ))
∞

and LMa,b,2 (τ ) := R′ Ha,b,n (b, τ, β̃a,b,n (τ )) , while we define the exponentially

∞
weighted CUSUM statistic (a supCUSUM statistic is analogously defined)

1
Z  
expCUSUMa,b := sup exp LMa,b,1 (λ, τ ) dτ.
λ∈[a,b] τ ∈T 2

In the following, we describe an algorithm which identifies different regimes, in

which Granger causality (GC, henceforth) holds or not. For ease of exposition, we
allow to identify at most three different regimes, i.e. at most two break points.
But the algorithm could straightforwardly be extended for the case of more than
two breaks.

17
Algorithm 2.

Step 1 Perform the expLM test at a significance of α. If the test rejects go to Step 2,
otherwise terminate.
1
Step 2 Perform the expCUSUM0,1 test at significance level 1 − (1 − α) 2 . If the test fails
to reject, GC is detected
 on [0, 1];terminate. Otherwise, the first breakpoint
λ1 := arg max τ ∈T exp 12 LM1 (λ, τ ) dτ is announced; go to Step 3.
R
λ∈[0,1]

1
Step 3 Perform the expLM0,λ1 and expLMλ1 ,1 tests at the significance level 1 − (1 − α) 4 :
1. If both tests fail to reject, the procedure is inconclusive on [0, 1]; terminate.
2. If the tests fail to reject on [0, λ1 ] but reject on [λ1 , 1], no GC is detected on
1
[0, λ1 ]. Perform the expCUSUMλ1 ,1 test at a significance level of 1 − (1 − α) 5 .
a) If the test fails to reject, GC is detected on [λ1 , 1].
 
R 1
b) If the test rejects, then a second break λ2 := arg max τ ∈T exp 2 LMλ1 ,1 (λ, τ ) dτ
λ∈[λ1 ,1]
is announced. Perform the expLMλ1 ,λ2 and expLMλ2 ,1 test at a significance
1
level of 1 − (1 − α) 7 .
i) If both tests fail to reject, conclude that the procedure is inconclusive
on [λ1 , 1].
ii) If both tests reject, GC is detected on both segments.
iii) If only one test rejects, GC is detected on the respective segment.
3. The tests reject on [0, λ1 ] but fail to reject on [λ1 , 1]: See Case 2 with the roles
of [0, λ1 ] and [λ1 , 1] inverted.
4. If both tests reject, compute max expCUSUMλk−1 ,λk ; set the significance level
1≤k≤2
1
to 1 − (1 − α) 5 .
a) If the test fails to reject, then GC is detected on [0, λ1 ] and [λ1 , 1].
b) If the test rejects, a new breakpoint λ2 is announced as the respective max-
imizer. Suppose λ1 ≤ λ2 and compute expLMλ1 ,λ2 and expLMλ2 ,1 ; the test
1
on [0, λ1 ] has already rejected; set the significance level to 1 − (1 − α) 7 .
i) If both tests fail to reject, the procedure is inconclusive on [λ1 , 1].
ii) If both tests reject, GC is detected on all segments.
iii) If only one test rejects, GC is detected on the respective segment.

Before proceeding, a few points are worth discussing: First, the choice of the
individual significance levels at each step corresponds essentially to a Šidák cor-
rection aimed at counteracting excessive accumulation of false rejections (see also

18
Galeano and Wied, 2017, pp. 339–340). This allows us to identify potential
regimes of Granger causality with a probability of committing a type one error
bounded by α. Second, it can be conjectured, based on the results collected in
Oka and Qu (2011), that the estimated break locations are n-consistent and their
effect on the limiting distributions of the tests in Step 3 will thus be asymptotically
negligible. This conjecture is supported by the finite sample evidence of the next
section. Third, in practice, the application of the CUSUM statistic in Step 1 can
be accompanied by plotting the CUSUM curve; a point that will be illustrated in
our empirical application. Fourth, regarding Step 2, note that the CUSUM test
also has non-trivial power, if there is more than one break point in a given inter-
val (see, e.g. the discussion in Qu, 2008, Section 5). Fifth, regarding Step 3.4,
note that we calculate the maximum of the CUSUM statistics in order to reduce
the total number of applied tests to two; see also Borsch et al. (2024) for details
about these two comments. The limiting distribution of the maximum of the two
CUSUM statistics is that of the maximum of two independent Brownian pillow
processes and can be simulated under the null of no additional break (see Oka
and Qu, 2011, Theorem 3). Fifth, and contrary to the CUSUM statistics, that
are nuisance parameter free even if Assumption D fails, we suggest to apply the
bootstrap procedure of Algorithm 1 in Step 1 to get critical values for the various
expLM statistics in case violation of Assumption D is suspected.

4 Monte Carlo Simulations

In the small sample simulations, we consider a location-scale model

yi = wi + γi zi + (1 + αwi )εi ,

where (w1 , z1 , ε1 ), . . . , (wn , zn , εn ) are IID copies2 of

w ∼ χ2 (3), z, ε ∼ N (0, 1), and ε ⊥ (w, z).

2
Appendix B contains additional Monte Carlo simulation results based on a quantile autore-
gressive distributed lag model with time trends. Here, no substantial differences are observed.

19
First, we investigate the size properties setting γi = 0 for all i. We distinguish
between homoscedasticity (α = 0) and heteroscedasticity (α = 3). In the latter
case, Assumption D will only be satisfied if E[z | w] = E[z]. Therefore, we consider
three scenarios: (1) α = 0 & z ⊥ w, (2) α = 3 & z ⊥ w, (3) α = 3 & cov[z, w] ≃
−3/4. Thus, it is only in scenario (3) where the asymptotic critical values based
on the asymptotic approximation under Assumption D are wrong.

We perform tests at five selected quantiles τ ∈ {0.05, 0.25, 0.50, 0.75, 0.95} as well
as across the complete interval [0.05, 0.95]. We consider the case where test statis-
tics are compared to critical values obtained under Assumption D (labelled ‘asy’).
Alternatively, we adjust the statistics as explained in Section 3.3 when testing at
a given τ or, when testing across all τ ∈ [0, 1], use the bootstrap (both labelled
‘adj’). For comparison, we compute also the supWald test of Koenker and Machado
(1999) given by

H
supWald = sup nh2 (τ )γn (τ )′ Ω−1 0 2 ′
n (τ )γn (τ ) →d sup h (τ )BB p (τ ) BB p (τ ) (10)
τ ∈T τ ∈T

where we equip the statistic with (pairs) bootstrap standard errors Ωn (·) imple-
mented using the quantreg package of R (Koenker et al., 2018). Critical values are
easily obtained from a discrete approximation of the Bessel limiting process. For
all bootstrap procedures we use B = 499 replications. All test decision are carried
out at the five per cent significance level.

Table 1 contains the Monte Carlo results under the null hypothesis of Granger non-
causality based on 2,000 Monte Carlo repetitions. As can be seen from panel a) and
b) of Table 1, size is controlled if n is moderately large irrespective of conditional
homoscedasticity/heteroscedasticity because conditional mean independence and
thus Assumption D is satisfied. As suggested by our theory, the performance
of the tests using the asymptotic approximation derived under Assumption D
deteriorates if cov[z, w] ̸= 0 (cf. panel c) of Table 1). However, in this case the
adjustment/bootstrap alternatives do their job by effectively keeping size. The
empirical size of the supWald test is in all scenarios in line with the nominal
significance level.

Turning to the power properties, we set, for better comparison, α = 0 and z ⊥ w

20
 a) Homosc. & uncorr. x b) Heterosc. & uncorr. x c) Heterosc. & corr. x
.05 .025 .50 .75 .95 [.05;.95] .05 .025 .50 .75 .95 [.05;.95] .05 .025 .50 .75 .95 [.05;.95]
n = 150 LM supLM expLM LM supLM expLM LM supLM expLM
adj .050 .045 .051 .048 .027 .053 .054 .060 .043 .054 .046 .033 .056 .054 .091 .074 .076 .062 .035 .048 .066
asy .064 .047 .051 .052 .038 .051 .046 .068 .043 .055 .047 .039 .053 .050 .124 .094 .091 .083 .057 .112 .100
Wald supW Wald supW Wald supW
asy .063 .045 .054 .058 .068 .067 .041 .032 .043 .047 .049 .043 .052 .049 .034 .038 .048 .040

n = 300 LM supLM expLM LM supLM expLM LM supLM expLM

adj .049 .048 .043 .047 .039 .044 .042 .060 .044 .048 .043 .038 .053 .048 .072 .058 .073 .060 .034 .047 .055
asy .064 .052 .044 .048 .047 .037 .039 0.065 .047 .048 .043 .043 .052 .045 .103 .090 .093 .090 .067 .115 .094
Wald supW Wald supW Wald supW
asy .058 .052 .047 .048 .060 .062 .043 .039 .041 .046 .039 .038 .047 .041 .041 .047 .049 .039

21
n = 1,000 LM supLM expLM LM supLM expLM LM supLM expLM
adj .046 .049 .049 .049 .033 .048 .051 .042 .058 .055 .045 .034 .054 .053 .044 .061 .058 .048 .033 .044 .053
asy .050 .050 .049 .049 .038 .048 .052 .046 .059 .055 .045 .041 .055 .049 .099 .100 .091 .086 .081 .115 .095
Wald supW Wald supW Wald supW
asy .051 .047 .050 .051 .046 .067 .048 .059 .044 .053 .041 .058 .057 .052 .047 .046 .048 .055

n = 2,000 LM supLM expLM LM supLM expLM LM supLM expLM

adj .050 .054 .054 .057 .044 .046 .049 .049 .056 .054 .050 .041 .044 .050 .033 .052 .050 .047 .026 .045 .051
asy .053 .054 .054 .057 .046 .042 .052 0.054 .056 .054 .050 .047 .045 .052 .097 .095 .089 .094 .082 .117 .091
Wald supW Wald supW Wald supW
asy .058 .052 .051 .050 .049 .063 .050 .042 .051 .052 .047 .060 .046 .050 .047 .050 .051 .055

Table 1: Rejection frequencies at a nominal size of five per cent of various tests statistics constructed over a grid [0.05, 0.06, . . . , 0.95]
based on 2,000 Monte Carlo repetitions.
(i.e. Assumption D is satisfied) and evaluate tests over the complete quantile
interval [0.05, 0.95]. First, let us consider the following three break scenarios:
 
γ i ≤ ⌊n/2⌋ 0 i ≤ ⌊n/2⌋
A : γi = B : γi = C : γi = γ.

−γ otherwise 
γ otherwise

This means that, in Scenario A, we have a structural break in the Granger pa-
rameter and the parameters sum up to zero over time (i.e. 01 g(v)dv = 0). Also
R

in Scenario B, there is a structural break, but the sum over time is not zero. In
Scenario C, the Granger parameter is constant and equal to γ (i.e. ∆(τ ) ̸= 0).

We consider our two statistics supLM, expLM (bootstrapped version) and the
supWald test. It is expected that our tests have power against all alternatives,
which increases in n, while supWald has no power in Scenario A, also for large n
because 01 g(v)dv = 0. This is indeed the empirical result. Somewhat surprisingly,
R

the power curves in Figure 1 suggest that in all scenarios, the break-robust tests
are more powerful than the supWald; in particular, also in Scenario C, where no
break is present. Among our new tests, the expLM test has more power than the
supLM test.

As a final analysis, we implement Algorithm 2 in order to identify regimes that

are potentially subject to Granger causality. Due to its superior performance (see
the discussion surrounding Figure 1), we use expLMλj ,λj+1 in conjunction with the
CUSUM statistic. For this, we consider up to three regimes of the (normalized)
sample [0, 1], i.e. [λj , λj+1 ], 0 ≤ j ≤ 2, with λ0 = 0, λ3 = 1, with 1/3 = λ1 <
λ2 = 2/3. We distinguish between four different scenarios: Two cases with one
break (0, 0, 1/2), (1/2, 0, 0), and two cases with two breaks (0, 1/2, 0), (1/2, 0, 1/2),
where (a, b, c) means that γi , i = 1, . . . , n, takes the value a, b, and/or c in regime
1, 2, and 3, respectively. As can be seen from Table 2, Algorithm 2 appears to
be able to consistently date the breaks and to correctly identify the regimes of
Granger causality with a probability of committing a type-one error not larger
than five percent. Once again, the the exponentially weighted test appears to
have a superior performance. Consistent with earlier results in the literature on
change point detection (see, e.g. Wied et al., 2012), correct detection rates are

22
 1
A n = 150 B n = 150 C n = 150
0.8

0.6

0.4

0.2

0
1
A n = 300 Bs n = 300 C n = 300
0.8

0.6

0.4

0.2

0
1
A n = 500 B n = 500 C n = 500
0.8

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0 0.1 0.2 0.3 0 0.1 0.2 0.3
Figure 1: Rejection frequencies under the alternative scenarios A, B, C as a function of γ ∈ [0, 0.3]
for supLM (orange), expLM (blue), and supW (black), with B = 499 using 2,000 Monte Carlo
iterations over a grid τ ∈ [0.05, 0.06, . . . , 0.95].

lower in cases of more than just one break.

5 Empirical Illustration: Crude Oil and Returns

We revisit an analysis about the causal relationships between crude oil and stock
returns from Ding et al. (2016) to illustrate the advantages of our new test. Ding
et al. (2016) consider the daily returns of West Texas Intermediate (WTI) and
Dubai crude oil as well as five major (mainly Asian) stock index returns, S&P
500 (SNP), Nikkei (NIK), Hang Seng (HAN), Shanghai (SHA), and KOSPI (KOS),

23
 % rejection frequencies
given LM & CUSUM reject % detected λ1,n λ2,n
n LM CUSUM [λ0 , λ1 ] [λ1 , λ2 ] [λ2 , λ3 ] = < > ? mean median var mean median var
500 sup 97.30 90.35 1.94 100.00 87.75 11.70 0.55 0.00 0.6253 0.6480 0.0048
exp 99.45 96.55 1.52 100.00 95.70 3.80 0.50 0.00 0.6406 0.6580 0.0032
1,000 sup 100.00 99.85 1.83 100.00 98.65 0.15 1.20 0.00 0.6428 0.6570 0.0019
exp 100.00 99.95 1.46 100.00 99.25 0.05 0.70 0.00 0.6511 0.6610 0.0011
(0, 0, γ)
2,000 sup 100.00 100.00 1.38 100.00 98.05 0.00 1.95 0.00 0.6546 0.6625 0.0005
exp 100.00 100.00 1.26 100.00 98.90 0.00 1.10 0.00 0.6589 0.6645 0.0003
4,000 sup 100.00 100.00 1.37 100.00 98.25 0.00 1.75 0.00 0.6601 0.6643 0.0002
exp 100.00 100.00 1.22 100.00 98.50 0.00 1.50 0.00 0.6623 0.6653 0.0001
500 sup 97.60 90.85 100.00 1.82 88.20 11.05 0.75 0.05 0.3761 0.3520 0.0052
exp 99.35 97.05 100.00 1.25 95.95 3.45 0.60 0.05 0.3582 0.3400 0.0032
1,000 sup 100.00 99.95 100.00 1.67 98.65 0.05 1.30 0.00 0.3583 0.3420 0.0022
exp 100.00 100.00 100.00 1.56 99.10 0.00 0.90 0.00 0.3474 0.3370 0.0012
(γ, 0, 0)
2,000 sup 100.00 100.00 100.00 1.88 98.30 0.00 1.70 0.00 0.3457 0.3370 0.0006
exp 100.00 100.00 100.00 1.31 99.35 0.00 0.65 0.00 0.3412 0.3355 0.0003
4,000 sup 100.00 100.00 100.00 1.58 98.05 0.00 1.95 0.00 0.3400 0.3355 0.0002
exp 100.00 100.00 100.00 1.31 98.95 0.00 1.05 0.00 0.3373 0.3343 0.0001
500 sup 100.00 30.10 100.00 2.61 100.00 19.15 80.85 0.00 0.00 0.3331 0.3320 0.0010 0.6649 0.6660 0.0012
exp 100.00 50.80 100.00 2.16 100.00 41.60 58.40 0.00 0.00 0.3288 0.3300 0.0012 0.6731 0.6680 0.0013
1,000 sup 100.00 75.80 100.00 1.57 100.00 73.40 26.60 0.00 0.00 0.3320 0.3330 0.0007 0.6668 0.6660 0.0006
exp 100.00 93.80 100.00 1.29 100.00 93.25 6.75 0.00 0.00 0.3284 0.3320 0.0006 0.6702 0.6670 0.0006
(γ, 0, γ)
2,000 sup 100.00 99.40 100.00 2.16 100.00 99.40 0.60 0.00 0.00 0.3313 0.3330 0.0003 0.6683 0.6665 0.0003
exp 100.00 100.00 100.00 1.20 100.00 100.00 0.00 0.00 0.00 0.3298 0.3320 0.0002 0.6695 0.6675 0.0002
4,000 sup 100.00 100.00 100.00 1.40 100.00 100.00 0.00 0.00 0.00 0.3322 0.3333 0.0001 0.6679 0.6668 0.0001
exp 100.00 100.00 100.00 1.05 100.00 100.00 0.00 0.00 0.00 0.3316 0.3330 0.0001 0.6685 0.6670 0.0001
500 sup 90.90 34.40 2.61 100.00 4.58 22.95 77.05 0.00 0.00 0.3260 0.3300 0.0012 0.6684 0.6680 0.0012
exp 96.20 53.10 1.58 100.00 3.04 44.45 55.55 0.00 0.00 0.3305 0.3320 0.0011 0.6679 0.6660 0.0012
1,000 sup 99.95 78.45 1.81 100.00 1.94 77.25 22.75 0.00 0.00 0.3294 0.3320 0.0007 0.6709 0.6670 0.0007
exp 100.00 93.95 1.23 100.00 1.18 93.55 6.45 0.00 0.00 0.3299 0.3330 0.0006 0.6698 0.6670 0.0005
(0, γ, 0)
2,000 sup 100.00 99.90 1.65 100.00 1.65 99.90 0.10 0.00 0.00 0.3299 0.3320 0.0003 0.6694 0.6670 0.0003
exp 100.00 100.00 1.20 100.00 1.75 100.00 0.00 0.00 0.00 0.3310 0.3325 0.0002 0.6686 0.6670 0.0002
4,000 sup 100.00 100.00 1.70 100.00 1.60 100.00 0.00 0.00 0.00 0.3307 0.3325 0.0001 0.6686 0.6668 0.0001
exp 100.00 100.00 1.55 100.00 1.45 100.00 0.00 0.00 0.00 0.3316 0.3329 0.0001 0.6678 0.6668 0.0001

Table 2: The various tests statistics constructed over a grid [0.05, 0.06, . . . , 0.95] of quantile levels;
the initial significance level is α = 0.05; 2,000 Monte Carlo are used. Rejection frequencies in
columns labelled ‘LM’ and ‘CUSUM’ are relative to total number of repetitions. The columns
labelled ‘% detected’ give the rate of detected breaks relative to the total number repetitions
while labelled ‘=’, ‘< / >’, and ‘?’ gives the rate of correctly detected breaks, too few/many
detected breaks, and the rate of inconclusive test decision, respectively.

from January 1, 1996, to October 12, 2012.

Following Ding et al. (2016), we consider the following autoregressive distributed

lag models
q
X q
X
QROILi (τ | Fi−1 ) = β1 (τ ) + α1,j (τ )ROILi−j + γ1,j (τ )RSi−j
j=1 j=1
Xq q
X
QRSi (τ | Fi−1 ) = β2 (τ ) + α2,j (τ )RSi−j + γ2,j (τ )ROILi−j
j=1 j=1

for Fi := σ({ROILj , RSj , j ≤ i}). Here, ROILi and RSi denote the oil and stock
returns at time point i, respectively, QYi (· | Fi−1 ) denotes the conditional quantile

24
functions of a given variable Yi . Of main interest are the coefficients γk,j (τ ), which
describe if one series helps to predict the conditional quantile of the other.

One of their main findings is a considerable Granger influence in quantiles of the

WTI returns on the stock returns, which is much stronger compared with the
other direction (stocks on WTI). The application of our test mainly supports this
finding and, in particular, gives substantially stronger evidence for this kind of
relationship. Moreover, by the use of Algorithm 2, we are able to give more insight
on the timing of violations of no Granger causality. We follow their discussion
(Ding et al., 2016, Section 4.3), where quantile ranges on [0.05, 0.95] corresponding
to the complete interval, the lower (i.e. [0.05, 0.2]) and upper tails (i.e. [0.80, 0.95]),
as well as the range around the median (i.e. [0.40, 0.60]) are considered.

Table 3 presents the p-values of our bootstrapped expLM and supLM tests, as well
as the supWald test in Eq. (10) equipped with bootstrap standard errors, each
computed over different quantile intervals using a step size of 0.01 for the grids. In
simulations, these three bootstrap-based tests display best size and power proper-
ties. In particular, the latter test equipped with bootstrap standard errors outper-
forms in our simulations the test equipped with a kernel based plug-in estimator
for the asymptotic covariance matrix used by Ding et al. (2016), so that we refrain
from using that test. As number of lags, we consider the values q ∈ {1, 3, 6, 9}.3

Out of 160 scenarios in total, the p-values of our test are smaller than or equal to
these in Ding et al. (2016) in 160 − 17 = 143 times. The cases in which our p-
values are higher, mostly concern the index KOSPI, where our p-values are larger
seven out of 16 times. But this mainly concerns cases, where the p-values are large
anyway. So, we have a robust finding that, if there is some evidence for Granger
causality, our test strengthens this evidence.

Having identified models, for which the null hypothesis is rejected, we aim at
identifying regimes of Granger causality with the multi-step procedure from Algo-
rithm 2. In order to limit computational complexity, we only consider the interval
[0.05, 0, 95] and the same lag lengths as in the first step. Table 4 yields the results
3
The results do not change substantially when using the sequential lag-length procedure of
Ding et al. (2016), where we consider the highest number of lags (with 9 as the maximum number)
for which the tests reject the null hypothesis.

25
for all stocks but SHA that did not display evidence of Granger causality.

[.05,.95] [.05,.20] [.40,.60] [.80,.95]

q supWald expLM supLM supWald expLM supLM supWald expLM supLM supWald expLM supLM
9 25.150 0.650 0.800 8.850 2.751 1.451 10.070 1.301 0.600 25.665 14.007 4.302
6 84.800 8.304 14.257 66.310 24.762 30.665 67.340 9.205 10.805 54.160 48.174 34.567
SHA

3 23.800 2.401 10.205 69.755 17.359 30.715 32.325 4.502 6.303 17.880 18.409 9.955
1 15.000 0.400 1.901 40.705 5.153 4.352 5.470 0.900 1.201 14.680 11.956 1.751
9 12.350 0.000 0.000 5.445 0.000 0.000 48.680 1.351 1.151 36.140 4.202 4.752
6 6.100 0.000 0.000 2.440 0.000 0.000 9.475 0.650 0.450 38.460 5.103 2.901
HAN

3 0.000 0.000 0.000 0.105 0.000 0.000 1.480 0.500 0.100 15.945 2.301 0.950
1 0.150 0.000 0.000 0.150 0.000 0.000 4.855 0.000 0.000 15.205 0.250 0.050
9 27.100 9.405 10.305 9.560 0.200 0.250 94.870 40.770 25.713 78.880 85.343 93.947
6 4.200 5.553 9.155 1.325 0.100 0.800 85.910 26.013 18.759 67.860 70.585 91.196
KOS

3 0.350 2.951 2.551 0.540 0.100 0.800 47.815 12.456 6.653 99.625 78.339 75.088
1 3.350 1.401 2.501 0.965 0.050 0.350 45.500 4.952 7.454 82.990 39.020 51.876
9 23.050 0.000 0.000 8.005 0.000 0.000 21.185 0.000 0.000 91.845 9.105 5.303
6 4.200 0.000 0.000 6.755 0.000 0.000 6.165 0.000 0.000 97.065 6.503 6.703
NIK

3 0.200 0.000 0.000 0.125 0.000 0.000 0.495 0.000 0.000 79.005 2.051 0.500
1 0.000 0.000 0.000 0.005 0.000 0.000 0.560 0.000 0.000 66.245 0.450 0.300
9 49.600 7.554 29.115 41.165 10.605 11.056 58.780 30.515 35.018 19.355 1.401 1.151
6 85.900 51.576 72.636 64.760 13.007 6.653 90.725 92.096 84.692 50.935 19.460 41.521
SNP

3 23.700 54.927 33.217 28.275 9.805 10.305 98.700 93.397 96.298 86.275 45.773 74.387
1 62.300 66.183 64.932 95.250 87.794 96.648 59.255 61.281 60.130 30.890 11.056 13.457

Table 3: p-values (percentages) for detecting Granger causality from lagged WTI returns (zi ) to
different stock index returns (yi ).

q SHA HAN KOS NIK

(0.650) (0.000) (9.405) (0.000)
1996-01-03 2009-03-06 1996-01-03 2009-02-20 1996-01-03 2009-03-17 1996-01-03 2008-09-03 2009-08-20
9 2009-03-06 2012-10-12 2009-02-20 2012-10-12 2009-03-17 2012-10-12 2008-09-03 2009-08-20 2012-10-12
✗ ✓ ✗ ✓ ✗ ✓ ✗ ✓ ✓
(8.304) (0.000) (5.553) (0.000)
1996-01-03 2009-03-06 1996-01-03 2009-02-20 1996-01-03 2009-03-31 1996-01-03 2008-09-26 2009-08-20
6 2009-03-06 2012-10-12 2009-02-20 2012-10-12 2009-03-31 2012-10-12 2008-09-26 2009-08-20 2012-10-12
✗ ✓ ✗ ✓ ✗ ✓ ✗ ✓ ✓
(2.401) (0.000) (2.951) (0.000)
1996-01-03 2009-02-17 1996-01-03 2009-02-20 1996-01-03 2009-03-31 1996-01-03 2008-09-18 2009-08-24
3 2009-02-17 2012-10-12 2009-02-20 2012-10-12 2009-03-31 2012-10-12 2008-09-18 2009-08-24 2012-10-12
✗ ✓ ✗ ✓ ✗ ✓ ✗ ✓ ✓
(0.400) (0.000) (1.401) (0.000)
1996-01-03 2008-10-19 1996-01-03 2008-10-13 1996-01-03 2008-10-13 1996-01-03 2008-09-18 2009-08-24
1 2008-12-19 2012-10-12 2008-10-13 2012-10-12 2008-10-13 2012-10-12 2008-09-18 2009-08-24 2012-10-12
✗ ✓ ✗ ✓ ✗ ✓ ✗ ✓ ✓

Table 4: Identified regimes of Granger causality (checkmark) or no Granger causality (cross) for
the interval τ ∈ [0.05, 0.95] based on Algorithm 2 using an initial significance level of ten percent
together with the p values (percentage) of the expLM test over the complete time series {1, . . . , n}
for lag length q ∈ {1, 3, 6, 9}.

In case of the SHA, HAN, and KOS, we identify two regimes: Before the onset

26
of the financial crisis of 2008 there seems to be no evidence of Granger causality,
while Granger causality is detected thereafter. This also applies to the NIK, where,
however, Algorithm 2 identifies three regimes 0 < λ1,n < λ2,n < 1. To rule out the
possibility of false rejection, we apply an additional refinement step by computing
the CUSUM statistics over [0, λ2,n ] and [λ1,n , 1], both of which clearly reject with
p-values given by 0.001 and 0.003, respectively, thereby confirming our findings.
Finally, Figure 2 succinctly summarizes the preceding discussion graphically by
plotting the respective CUSUM plots for all four stock indices where Step 1 rejects
in case of q = 1.
SHA λ1,n HAN λ1,n
0.6

0.4

0.2

KOS λ1,n NIK λ1,n λ2,n

0.6

0.4

0.2

0
1996-01

1998-09

2001-06

2004-03

2006-12

2009-09

2012-06
1996-01

1998-09

2001-06

2004-03

2006-12

2009-09

2012-06

Figure 2: Highlighted regimes of Granger causality detected using Algorithm 2, with q = 1,

B = 1,999 on τ ∈ [0.05, 0.95], and initial significance level of ten percent. The solid black/grey
lines are the CUSUM curves calculated over the whole sample [0, 1] and the sub-sample [λ1,n , 1],
respectively; the solid/dashed horizontal line represents the critical value at a significance level
of 1 − (1 − α)1/2 and 1 − (1 − α)1/5 , respectively.

We corroborate our findings with an analysis about the structural stability of

the correlation ρ(i, j) = corr[yi , zi−1 ] between the lagged WTI (z) and the stock
index returns (y). An application of the test for constant correlations from Wied
et al. (2012), whose assumptions are typically plausible in the context of (stock)
returns, shows that all p-values are smaller than 0.05 in the four cases, where the

27
new procedure detects a break. For HAN and NIK, they are smaller than 0.001,
for Shanghai, the p-value is 0.009, for KOS, the p-value is 0.014.

6 Conclusion
We have proposed new tests for Granger causality that are robust against struc-
tural breaks and compete very well against existing tests. For future research, it
might be interesting to consider nonlinear quantile models such as in Troster (2018)
instead of pure linear ones. Moreover, one could consider systems of equations: In
our setting, we have one cross-section regression equation and test for restrictions
in this equation. In our empirical application on stock returns, this leads to five
different tests for each interval of quantile levels. Merging the information to one
system of equations might increase the power even further because cross-sectional
dependence would be taken into account then.

References
Abadir, K. M. and J. R. Magnus (2005): Matrix Algebra, vol. 1, Cambridge:
Cambridge University Press.

Andrews, D. W. (1993): “Tests for Parameter Instability and Structural Change

with Unknown Change Point,” Econometrica: Journal of the Econometric So-
ciety, 821–856.

Andrews, D. W. K. and W. Ploberger (1994): “Optimal Tests when a

Nuisance Parameter is Present only under the Alternative,” Econometrica, 62,
1383–1414.

Baum, C., S. Hurn, and J. Otero (2021, forthcoming): “The dynamics of U.S.
industrial production: A time-varying Granger causality perspective,” Econo-
metrics and Statistics.

Borsch, M., A. Mayer, and D. Wied (2024): “Consistent Estimation of Mul-

tiple Breakpoints in Dependence Measures,” Journal of Business & Economic
Statistics, 42, 695–706.

28
Bouezmarni, T., M. Doukali, and A. Taamouti (2024): “Testing Granger
Non-Causality in Expectiles,” Econometric Reviews, 43, 30–51.

Breitung, J. and B. Candelon (2006): “Testing for short- and long-run

causality: A frequency-domain approach,” Journal of Econometrics, 363–378.

Candelon, B. and S. Tokpavi (2016): “A Nonparametric Test for Granger

Causality in Distribution With Application to Financial Contagion,” Journal of
Business and Economic Statistics, 34, 240–253.

Caporin, M. and M. Costola (2022): “Time-varying Granger Causality Tests

in the Energy Markets: A Study on the DCC-MGARCH Hong test,” Energy
Economics, 111, 106088.

Chen, Y.-C., K. S. Rogoff, and B. Rossi (2010): “Can Exchange Rates

Forecast Commodity Prices?” The Quarterly Journal of Economics, 125, 1145–
1194.

Chuang, C.-C., C.-M. Kuan, and H.-Y. Lin (2009): “Causality in Quantiles
and Dynamic Stock Return-Volume Relations,” Journal of Banking & Finance,
33, 1351–1360.

Clark, T. E. and M. W. McCracken (2006): “The Predictive Content of the

Output Gap for Inflation: Resolving In-Sample and Out-of-Sample Evidence,”
Journal of Money, Credit and Banking, 1127–1148.

Ding, H., H.-G. Kim, and S. Y. Park (2016): “Crude Oil and Stock Markets:
Causal Relationships in Tails?” Energy Economics, 59, 58–69.

Galeano, P. and D. Wied (2017): “Dating Multiple Change Points in the

Correlation Matrix,” Test, 26, 331–352.

Galvao, A. F., G. Montes-Rojas, and S. Y. Park (2013): “Quantile Autore-

gressive Distributed Lag Model with an Application to House Price Returns,”
Oxford Bulletin of Economics and Statistics, 75, 307–321.

Gebka, B. and M. E. Wohar (2013): “Causality between Trading Volume

and Returns: Evidence from Quantile Regressions,” International Review of
Economics & Finance, 27, 144–159.

29
Giacomini, R. and B. Rossi (2010): “Forecast Comparisons in Unstable Envi-
ronments,” Journal of Applied Econometrics, 25, 595–620.

Granger, C. W. (1969): “Investigating Causal Relations by Econometric Mod-

els and Cross-Spectral Methods,” Econometrica: Journal of the Econometric
Society, 424–438.

Hansen, B. E. (1996): “Stochastic Equicontinuity for Unbounded Dependent

Heterogeneous Arrays,” Econometric Theory, 12, 347–359.

——— (2021): “Criterion-Based Inference Without the Information Equality: The

Weighted Chi-Square Distribution,” working paper.

Hoga, Y. (2024): “Persistence-Robust Break Detection in Predictive Quantile

and CoVaR Regressions,” arXiv: 2410.05861.

Jeong, K., W. K. Härdle, and S. Song (2012): “A Consistent Nonparametric

Test for Causality in Quantile,” Econometric Theory, 28, 861–887.

Koenker, R. (2005): Quantile Regression, vol. 38, Cambridge university press.

Koenker, R. and J. A. F. Machado (1999): “Goodness of Fit and Related

Inference Processes for Quantile Regression,” Journal of the American Statistical
Association, 94, 1296–1310.

Koenker, R., S. Portnoy, P. T. Ng, A. Zeileis, P. Grosjean, and B. D.

Ripley (2018): “Package ‘quantreg’,” Reference manual available at R-CRAN:
https://cran. rproject. org/web/packages/quantreg/quantreg. pdf.

Lee, T.-H. and W. Yang (2012): “Money-Income Granger-Causality in Quan-

tiles,” in 30th Anniversary Edition (Advances in Econometrics, Volume 30),
ed. by D. Terrell and D. Millimet, Emerald Group Publishing Limited, vol. 30,
383–407.

Li, T.-H. (2012): “Quantile periodograms,” Journal of the American Statistical

Association, 107, 765–776.

Oka, T. and Z. Qu (2011): “Estimating Structural Changes in Regression Quan-

tiles,” Journal of Econometrics, 162, 248–267.

30
Parzen, M., L.-J. Wei, and Z. Ying (1994): “A Resampling Method based
on Pivotal Estimating Functions,” Biometrika, 81, 341–350.

Powell, J. L. (1991): “Estimation of monotonic regression models under quan-

tile restrictions,” in Nonparametric and Semiparametric Models in Economet-
rics, ed. by W. Barnett, J. Powell, and T. G., Cambridge: Cambridge University
Press.

Qu, Z. (2008): “Testing for Structural Change in Regression Quantiles,” Journal

of Econometrics, 146, 170–184.

Rossi, B. (2005): “Optimal Tests for Nested Model Selection with Underlying
Parameter Instability,” Econometric Theory, 21, 962–990.

——— (2006): “Are Exchange Rates really Random Walks? Some Evidence Ro-
bust to Parameter Instability,” Macroeconomic dynamics, 10, 20–38.

——— (2013): “Advances in Forecasting under Instability,” in Handbook of Eco-

nomic Forecasting, ed. by G. Elliott and A. Timmermann, Amsterdam: Elsevier,
vol. 2B of Handbook of Economic Forecasting, 1203–1324.

——— (2021): “Forecasting in the Presence of Instabilities: How We Know

Whether Models Predict Well and How to Improve Them,” Journal of Eco-
nomic Literature, 59, 1135–1190.

Rossi, B. and Y. Wang (2019): “Vector Autoregressive-Based Granger Causal-

ity Test in the Presence of Instabilities,” The Stata Journal, 19, 883–899.

Rothe, C. and D. Wied (2013): “Misspecification Testing in a Class of Condi-

tional Distributional Models,” Journal of the American Statistical Association,
108, 314–324.

Sowell, F. (1996): “Optimal Tests for Parameter Instability in the Generalized

Method of Moments Framework,” Econometrica: Journal of the Econometric
Society, 1085–1107.

Stock, J. H. and M. Watson (2003): “Forecasting Output and Inflation: The

Role of Asset Prices,” Journal of Economic Literature, 41, 788–829.

31
Stock, J. H. and M. W. Watson (1996): “Evidence on Structural Instability
in Macroeconomic Time Series Relations,” Journal of Business & Economic
Statistics, 14, 11–30.

——— (1999): “Forecasting Inflation,” Journal of Monetary Economics, 44, 293–

335.

——— (2006): “Chapter 10 Forecasting with Many Predictors,” in Handbook of

Economic Forecasting, ed. by G. Elliott, C. W. J. Granger, and A. Timmermann,
Amsterdam: Elsevier, vol. 1, 515–554.

Taamouti, A., T. Bouezmarni, and A. El Ghouch (2014): “Nonparamet-

ric Estimation and Inference for Conditional Density based Granger Causality
Measures,” Journal of Econometrics, 180, 251–264, publisher: Elsevier.

Troster, V. (2018): “Testing for Granger-Causality in Quantiles,” Econometric

Reviews, 37, 850–866.

Wied, D., W. Krämer, and H. Dehling (2012): “Testing for a change in

correlation at an unknown point in time using an extended functional delta
method,” Econometric Theory, 28, 570–589.

Yang, Z., A. Tu, and Y. Zeng (2014): “Dynamic Linkages between Asian
Stock Prices and Exchange rates: New Evidence from Causality in Quantiles,”
Applied Economics, 46, 1184–1201.

A Proofs
Lemma A.1 Recall β0 (τ ) = (0′p , α0 (τ )′ )′ is the parameter value under H0 , and
suppose Assumptions A-C are satisfied.

(i) For a given τ ∈ T it holds uniformly in λ ∈ [0, 1]

Z λ !
−1/2
h(τ )Hn (λ, τ, β0 (τ )) ⇒ Bm (λ) + J H(τ )R λ∆(τ ) + δ(τ ) g(v)dv .
0

32
 (ii) If Assumptions A-C hold uniformly in τ , then uniformly in (τ, λ) ∈ T × [0, 1]
Z λ !
−1/2
Hn (λ, τ, β0 (τ )) ⇒ Sm (λ) + J H(τ )R λ∆(τ ) + δ(τ ) g(v)dv .
0

Proof of Lemma A.1. To begin with, we note that (by Qu, 2008, Lem A.1,A.2)
the process Sn (λ, τ, β0 (τ )) is stochastically equicontinuous on [0, 1] × T equipped
with the norm ρ({λ1 , τ1 }, {λ2 , τ2 }) = |λ2 − λ1 | + |τ2 − τ1 |. We now prove first
Part (i). Fix some τ ∈ T , and let βi,n (τ ) := (γi,n (τ )′ , α0 (τ )′ )′ denote the m × 1
parameter vector under the local alternatives given by Eq. (2). Now, under the
sequence of local alternatives in Eq. (2), one gets

Sn (λ, τ, β0 (τ ))
1 ⌊λn⌋
xi ψτ (yi − x′i β0 (τ ))
X
=√
n i=1
1 ⌊λn⌋
xi ψτ (yi − x′i βi,n (τ ))
X
=√
n i=1
1 ⌊λn⌋
X h
+√ xi 1{yi ≤ x′i β0 (τ )} − 1{yi ≤ x′i βi,n (τ )}
n i=1
i
− Fi (x′i β0 (τ )) + Fi (x′i βi,n (τ ))
1 ⌊λn⌋
xi (Fi (x′i βi,n (τ )) − Fi (x′i β0 (τ ))) =: I + II + III,
X
+√ (A.1)
n i=1

say. We know that, by Assumption A, xi ψτ (yi −x′i βi,n (τ )) is a martingale difference

array under the local alternatives so that by the FCLT h(τ )J−1/2 I ⇒ Bm (λ).
Moreover, the stochastic equicontinuity of Sn (·) yields II = op (1) (cf. Qu, 2008,
Lem A.1/A.2). Finally, using a Taylor-series expansion and Eq. (2), yields

1 ⌊λn⌋ Z λ
xi x′i fi (x′i βi,n (τ ))Rγi,n (τ ) + op (1) ⇒ λHR∆(τ ) + δ(τ )HR
X
III = √ g(v)dv,
n i=1 0

which, as Xn′ Xn /n = J + op (1) (cf. Ass. B), proves the claim. Part (ii) follows
analogously noting that J−1/2 I ⇒ S(λ, τ ). □

33
Proof of Proposition 1. Recall that αn (τ ) denotes a solution to the constrained
optimization defined in Eq. (4) (i.e. assuming γi,n (τ ) = 0). Moreover, recall that
R̄ is an m × k selection matrix such that
 
0p 
R̄α0 (τ ) = β0 (τ ) =  .
α0 (τ )

Now, using the same argument that lead to Eq. (A.1), we get

Sn (λ, τ, β̃n (τ )) = Sn (λ,τ, β0 (τ ))

1 ⌊λn⌋
xi (Fi (wi′ αn (τ ) − Fi (wi′ α0 (τ ))) + op (1)
X
+√
n i=1
(1)
= Sn (λ, τ, β0 (τ ))
1 ⌊λn⌋ √
xi wi′ fi (wi′ α0 (τ )) n(αn (τ ) − α0 (τ )) + op (1)
X
+
n i=1
(2)
= Sn (λ, τ, β0 (τ ))
1 ⌊λn⌋ √
xi x′i fi (x′i β0 (τ ))(R̄ n(αn (τ ) − α0 (τ ))) + op (1)
X
+
n i=1
(3) √
= Sn (λ, τ, β0 (τ )) + λH(τ ) n(β̃n (τ ) − β0 (τ )) + op (1),

where β̃n (τ ) = R̄αn (τ ) = (0′p , αn′ (τ ))′ . Equation (1) uses a first order Taylor-series
expansion (see Qu, 2008, proof of lem. 1), (2) is due to xi x′i R̄ = xi wi′ , and (3)
is due to Ass. C that also defines the m × m matrix H(τ ). Moreover, recall that
Hα (τ ) and Jα denote the lower-right k × k block of H(τ ) and J, respectively, and
partition

Sn (λ, τ, t) = (Sn,γ (λ, τ, tγ )′ , Sn,α (λ, τ, tα )′ )′ , tγ ∈ Rp , tα ∈ Rk

to conform with the partitioning β0 (τ ) = (γ0 (τ )′ , α0 (τ )′ )′ for γ0 (τ ) = 0p . Then,

 
√ 0p
n(β̃n (τ ) − β0 (τ )) = −  −1  + op (1),
Hα Sn,α (1, τ, α0 (τ ))

34
see also Koenker and Machado (1999, Proof of Thm 1) for the partitioning. More-
over, some algebra reveals
 

0p 
−1
Hα Sn,α (1, τ, β0 (τ ))
  
0p×p 0p×k 0p
= −1 −1/2 −1

−1/2

0k×p Hα (Jα ) Jα Sn,α (1, τ, α0 (τ ))
 
0p×p 0p×k  J−1/2 Sn (1, τ, β0 (τ )) =: A(τ )J−1/2 Sn (1, τ, β0 (τ )),
= −1 −1/2 −1
0k×p Hα (Jα )

say, where the m × m matrix A(τ ) has been implicitly defined. Hence,

Hn (λ, τ, β̃(τ )) = J−1/2 Sn (λ, τ, β0 (τ )) − λC(τ )A(τ )J−1/2 Sn (1, τ, β0 (τ )) + op (1),

where we recall C(τ ) = J−1/2 H(τ ). The matrix, P (τ ) = C(τ )A(τ ) is idempotent
of rank k; cf. Eq (7) so that, by Lemma A.1, the distribution of the restricted
quantile estimator follows.

Next, by lemma A.1

h(τ )R′ H̃n (λ, τ,β̃(τ )) ⇒ BBp (λ)

Z λ Z 1 !
′
+ h(τ )δ(τ )R C(τ )R (1 − λ) g(v)dv − λ g(v)dv .
0 λ

In particular, if λ = 1, then

Hn (1, τ, β̃(τ )) = J−1/2 Sn (1, τ, β0 (τ ))

− C(τ )A(τ )J−1/2 Sn (1, τ, β0 (τ )) + op (1) (A.2)
= (I − P (τ ))J−1/2 Sn (1, τ, β0 (τ )) + op (1),

where we recall that under Ass. D, Im − P (τ ) = RR′ . □

Moreover, if T (τ ) is the inverse of the m × m matrix of eigenvectors of Im − P (τ ),

then, by the Jordan decomposition (see e.g. Abadir and Magnus, 2005, Ex 8.60),

35
Im − P (τ ) = RR′ T (τ ) (where T (τ ) = Im if Ass. A holds). Therefore,
Z 1
h(τ )R′ Hn (1, τ, β̃(τ )) ⇒ T (τ )Bp (1) + h(τ )R′ T (τ )C(τ )R(∆(τ ) − δ(τ ) g(v)dv).
0

Noting that B(1) and BB(λ) are independent because cov[B(1), BB(λ)] = 0 by
construction, this completes the proof. □

Proof of Corollary 1. Follows immediately from Proposition 1. □

Proof of Corollary 2. The claim follows from the continuous mapping theorem
and Proposition 1. □

Proof of Proposition 2. The proof follows analogously to that of Proposition 1

and Corollary 2 using Lemma A.1 (ii) and noting that

1 ⌊λn⌋
xi ψτ (yi − x′i βi,n (τ )) ⇒ Sm (λ, τ )
X
√
n i=1

uniformly on ℓ∞ ([0, 1] × T ) as mentioned in the proof of Lemma A.1. □

Proof of Corollary 3. The claim follows from the continuous mapping theorem
and Proposition 2. □

Proof of Proposition 3. Set Xn := {(yi , xi ), 1 ≤ i ≤ n} and let P∗ (·) denote

the probability measure induced by the empirical distribution of Xn . Begin by
observing that αn,b (τ ) solves the constrained quantile regression problem Eq. (4)
based on the rescaled data

{(πi,b wi′ αn (Ui,b ), πi,b wi ), 1 ≤ i ≤ n},

IID
for bootstrap weights πi,b ∼ π =d Multinomial(n, 1/n). Based on this observation,
we show first
√ √
n(αn,b (τ ) − αn (τ )) = n(αn (τ ) − α(τ )) + oP∗ (1). (A.3)

36
To see that this is true, observe that
√
n(αn,b (τ ) − αn (τ )) = − H−1
α (τ )Sn,b,α (τ, αn (τ )) + oP∗ (1),

where the sub-gradient of the constrained quantile regression problem is given by

1 ⌊λn⌋ ′
tα ∈ Rp .
X
Sn,b,α (λ, τ, tα ) := √ wi,b (1{ŷi,b ≤ wi,b tα } − τ ),
n i=1

Note that Sn,b,α (λ, τ, αn (τ )) is centred

1 ⌊λn⌋
X
′
E[Sn,b,α (λ, τ, αn (τ )) | Xn ] = √ E[wi,b (1{ŷi,b ≤ wi,b tα } − τ ) | Xn ]
n i=1
1 ⌊λn⌋
E[π]wi (E[1{wi′ αn (Ui,b ) ≤ wi′ αn (τ )} | Xn ] − τ ).
X
=√
n i=1

Now, using E[π] = 1 and that wi′ αn (τ ) is the τ -quantile of yi conditional on Xn ,

one gets E[Sn,b,α (λ, τ, αn (τ )) | Xn ] = 0. Similarly, it follows

cov[Sn,b,α (λ1 , τ1 , αn (τ1 )), Sn,b,α (λ2 , τ2 , αn (τ2 )) | Xn ] = (λ1 ∧ λ2 )(τ1 ∧ τ2 − τ1 τ2 ).

We can conclude by the FCLT for IID data that, uniformly in (λ, τ ) ∈ [0, 1] × T ,

Sn,b,α (λ, τ, αn (τ )) = −Hα (τ )Sp (λ, τ ) + oP∗ (1).

This proves Eq. (A.3). Next, define

1 ⌊λn⌋
xi,b ψτ (ybi,b − x′i,b β̃n,b (τ )),
X
Sn,b (λ, τ, β̃n,b (τ )) = √
n i=1

where β̃n,b (τ ) = (0′p , αn,b (τ )′ )′ . We can then follow the proof of Proposition 1,
using the above argument in conjunction with the stochastic equicontinuity of the
two-parameter process Eq. (5) in (λ, τ ) ∈ [0, 1] × T , and obtain
√
Sn,b (λ, τ, β̃n,b (τ )) = Sn,b (λ, τ, β̃n (τ )) + λH(τ ) n(β̃n,b (τ ) − β̃n (τ )) + oP∗ (1).

37
 √ √
By Eq. (A.3), n(β̃n,b (τ ) − β̃n (τ )) = n(β̃n (τ ) − β̃(τ )) + oP∗ (1) and, using similar
arguments, Sn,b (λ, τ, β̃n (τ )) = Sn (λ, τ, β̃(τ )) + oP∗ (1). In view of Proposition 2 and
Corollary 3, this proves Part (i). For Part (ii), note that under fixed alternatives
ĉ(α) = Op (1) because the way the bootstrap sample is generated enforces the null
hypothesis. Therefore, following the same argument used in the proof of Theorem
3 (ii) in Rothe and Wied (2013), Part (ii) is proven. □

B Additional Monte Carlo Results

We consider the following quantile autoregressive distributed lag model

′
yi = γi,n zi + α′ wi + ui , ui ∼ N (0, 1),

where the potentially Granger-causal regressor is given by the p × 1 vector zi :=

(z1,i , z2,i )′ (i.e. p = 2), with z1,i = (1/3)z1,i−1 + vi , vi ∼ N (0, 1), and z2,i ∼ χ2 (4),
while wi is k × 1 and given by wi := (1, yi−1 , yi−2 , i/n, (i/n)2 , w1,i )′ (i.e. k = 5 so
that m = 7) for w1,i ∼ χ2 (3) and α = (0, 1/3, 1/4, 1/2, 1/2, 1/2)′ . This design is
reminiscent of models considered elsewhere in empirical work; see, e.g. Chuang
et al. (2009) or Gebka and Wohar (2013).
√
We conduct a size study (γi,n = (0, 0)′ ) and a (local) power study ( nγi,n = (1, 1)′
for i < ⌊n/2⌋ and zero otherwise) for our new exp/supLM tests (i.e. γ = 02 ), setting
n ∈ {150, 300, 1,000, 2,000} and considering the quantile range [0.05, 0.95]. We use
both the asymptotic critical values (valid only under Assumption D) as well as our
bootstrap procedure from Algorithm 1. As in our simulation study of the main
text, we use 2,000 Monte Carlo iterations and B = 499 bootstrap replications.
The results collected by Table 5 confirm the good finite sample properties in terms
of size as well as power, with, in case of power, a superior performance of the
exponentially weighted statistic.

38
 size power
n supLM expLM supLM expLM
asy 4.50 4.30 17.1 20.3
150
boot 4.30 4.50 16.0 22.0
asy 5.15 4.55 18.1 21.3
300
boot 4.91 5.01 17.5 20.1
asy 4.75 4.85 20.8 25.6
1,000
boot 4.37 4.40 19.0 29.1
asy 4.60 4.95 20.4 26.5
2,000
boot 4.50 4.60 18.5 24.0

Table 5: Monte Carlo rejection rates (percentages) where ‘size’ and ‘power’ refer to rejection rates
at a five percent significance level under the null and local alternatives, respectively. The labels
‘asy’ and ‘boot’ refer, respectively, to the use of critical values obtained from the asymptotic
approximation and the bootstrap procedure from Algorithm 1.

39