Upload
others
View
11
Download
0
Embed Size (px)
Citation preview
Max-Correlation Test and Max-Causality Testfor Economic Time Series
Kaiji Motegi1
1Graduate School of Economics, Kobe University
UNC Econometrics SeminarFebruary 17, 2017
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 1 / 49
Introduction
Y
X
1. Are past values of Y useful for
predicting future Y?
-- White Noise Tests
2. Do past values of X serve as extra
information for predicting future Y?
-- Granger causality Tests
Present
Past Future
time
time
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 2 / 49
Introduction
White Noise
Granger
Causality
Hill and Motegi (2017a, WP)
Max-Correlation Test
Hill and Motegi (2017b, WP)
Stock Market Efficiency
Ghysels, Hill, and Motegi (2016, JoE)
Mixed Frequency Granger Causality
Ghysels, Hill, and Motegi (2017, R&R, JBES)
Max-Causality Test
Methodology Application
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 3 / 49
Table of Contents
1 Max-Correlation White Noise Test
1 Literature Review of White Noise Tests
2 Max-Correlation Test
3 Monte Carlo Simulations
2 Testing for Weak Form Efficiency of Stock Markets
1 Periodicity of Confidence Bands
2 Case Study (S&P 500)
3 Max-Causality Test (Overview)
4 Future Work
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 4 / 49
White Noise Tests: Motivation
We propose a new white noise test based on the largest sampleautocorrelation.
Two major fields where white noise tests are of use:
1 Testing for weak form efficiency of stock markets (cf. Fama,1965; Fama, 1970). Under weak form efficiency, stock returnsshould be white noise.
2 Residual diagnostics.
Establishing a formal white noise test is a relatively new andchallenging topic.
1 Infinitely many zero restrictions – H0 : ρ(h) = 0 for any h ≥ 1.2 No more than uncorrelatedness – weaker than IID or MDS.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 5 / 49
White Noise Tests: Literature Review
Consider univariate covariance stationary {y1, . . . , yn}.Assume that E[yt] = 0 for notational simplicity.
Population quantities are
γ(0) = E[y2t ], γ(h) = E[ytyt−h], ρ(h) =γ(h)
γ(0).
Sample quantities are
γn(0) =1
n
n∑t=1
y2t , γn(h) =1
n
n∑t=h+1
ytyt−h, ρn(h) =γn(h)
γn(0).
White noise hypothesis is
H0 : ρ(h) = 0 for h ≥ 1.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 6 / 49
White Noise Tests: Literature Review
Consider the Q-test (or portmanteau test):
Qn = nL∑
h=1
wn(h)ρ2n(h)
d→ χ2L.
wn(h) = (n+ 2)/(n− h) in Ljung and Box (1978).
Two reasons why the Q-test is not a white noise test:
1 It cannot capture autocorrelations beyond lag L.2 Asymptotic χ2 property requires the asymptotic independence
of {ρn(1), . . . , ρn(L)}, which holds when {yt} is IID.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 7 / 49
White Noise Tests: Literature Review
The issue of fixed lag length L is addressed by Hong’s (1996)spectral test.
A version of Hong’s test based on the Bartlett kernel is:
Nn =1√2Ln
Ln∑h=1
wn(h)[nρ2n(h)− 1
] d→ N(0, 1),
where Ln → ∞ as n→ ∞ and Ln = o(n).
Asymptotic normality still requires serial independence.
Shao’s (2010, 2011) dependent wild bootstrap (DWB)enables us to construct a formal white noise test.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 8 / 49
White Noise Tests: Literature Review
1 Set a block size bn.2 Generate iid {ξ1, ξ2, . . . , ξn/bn}. Define an auxiliary variable:
ω = [ ξ1, . . . , ξ1︸ ︷︷ ︸bn terms
, ξ2, . . . , ξ2︸ ︷︷ ︸bn terms
, . . . , ξn/bn , . . . , ξn/bn︸ ︷︷ ︸bn terms
]′.
3 Compute bootstrapped autocorrelations:
ρ(dw)n (h) =
1
γn(0)× 1
n
n∑t=h+1
ωt[ytyt−h − γn(h)], h = 1, . . . ,Ln,
and N(dw)n = (2Ln)
−1/2∑Ln
h=1wn(h)[n{ρ(dw)n (h)}2 − 1].
4 Repeat Steps 2-3 M times and compute the bootstrappedp-value p
(dw)n,M = 1
M
∑Mi=1 I(|N
(dw)n,i | ≥ |Nn|).
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 9 / 49
White Noise Tests: Literature Review
Block 2 Block 3
ξ1
y3y4, y4y5, y5y6 y6y7, y7y8, y8y9 y57y58, y58y59, y59y60y1y2, y2y3
・・・
・・・
Block 1 ・・・ Block 20
Compute bootstrapped autocorrelation
Repeat M times
Compute bootstrapped p-value
Generate i.i.d. random numbers ξ1, …, ξ20 ~ N(0, 1)
Examle:
h = 1; n = 60; bn = 3.
Preserved dependence within each block.
No dependence across different blocks.
Asymptotically correct size and consistency
under weak dependence (e.g. GARCH, bilinear).
ξ2ξ3 ξ20
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 10 / 49
Max-Correlation Test
We propose a max-correlation test statistic:
Tn =√n max {|ρn(1)|, . . . , |ρn(Ln)|} ,
where Ln → ∞ as n→ ∞ and Ln = o(n).
We use DWB for p-value computation.
’Sum of Squares’ versus ’Maximum’.
The max-approach is likely more robust against autocorrelationsat remote lags (e.g. seasonality).
Example: Lags 12, 24, 36, . . . in monthly data.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 11 / 49
Max-Correlation Test
Classical max-approach (e.g. Berman, 1964) attempts to derivean asymptotic distribution of the largest autocovariance.
Xiao and Wu (2014) prove that
an
[√n max1≤h≤Ln |γn(h)− γ(h)|√∑∞
h=0 γ(h)2
− bn
]
converges to the Gumbel distribution, where an, bn ∼√2 lnn.
Xiao and Wu (2014) use Horowitz, Lobato, Nankervis, and Savin’s
(2006) blocks-of-blocks bootstrap for p-value computation (without
proving its validity under their assumptions).
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 12 / 49
Max-Correlation Test
Innovation #1
The previous max-approach exploits extreme value theory inorder to derive the asymptotic distribution of max-covariance.
We bypath it by proving the asymptotic validity of DWB.
It suffices to prove that the actual test statistic andbootstrapped test statistics converge to the same distribution(no matter what it is).
Our approach allows for more general memory/momentproperties than the previous approach (e.g. NED vs. mixing).
Innovation #2
Xiao and Wu (2014) allow for observed data only.
We allow for filtered data (i.e. residuals).
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 13 / 49
First-Order Expansion for Filtered Series
Suppose that the true DGP is ARMA(1,1):
yt = 0.5yt−1 + νt + 0.5νt−1, νt ∼ WN(0, 1).
Fit AR(1) model: yt = θyt−1 + ut.
Least squares estimator: θn =∑n
t=2 ytyt−1/∑n
t=2 y2t−1.
Filtered series (i.e. residual): ut(θn) = yt − θnyt−1.
Since ut = νt + 0.5νt−1, population autocorrelation of {ut} atlag 1 is
ρ(1) = E[utut−1]/E[u2t ] = 0.4.
Sample autocorrelation of {ut(θn)} at lag 1 is
ρn(1) =1n
∑nt=2 ut(θn)ut−1(θn)
1n
∑nt=2 u
2t−1(θn)
.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 14 / 49
First-Order Expansion for Filtered Series
By the mean value theorem,
ρn(1) = ρ(1)−E[ut−1yt−1][E[y
2t−1]]
−1E[utyt−1]
E[u2t ]+ op(1)
= ρ(1)− D(1)×A× E[mt]
γ(0)+ op(1)
= 0.4− 1.5× 0.429× 0.5
1.25+ op(1)
= 0.4− 0.257+ op(1)
= 0.143 + op(1).
ρn(1) underestimates ρ(1), which lowers the test power.
Bias correction is desired in order to raise power.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 15 / 49
First-Order Expansion for Filtered Series
Define sample counterparts:
1 Dn(h) =1n
∑nt=h+1 ut−h(θn)yt−1.
2 An =(1n
∑nt=2 y
2t−1
)−1.
3 mt(θn) = ut(θn)yt−1.
Define a bias-corrected cross term of filtered series:
En,t,h(θn) = ut(θn)ut−h(θn)− Dn(h)× An ×mt(θn)
and its sample mean:
gn(h, θn) =1
n
n∑t=h+1
En,t,h(θn).
Execute DWB for {En,t,h(θn)− gn(h, θn)}.Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 16 / 49
Dependent Wild Bootstrap for Filtered Series
1 Set a block size bn.2 Generate iid {ξ1, ξ2, . . . , ξn/bn}. Define an auxiliary variable:
ω = [ξ1, . . . , ξ1︸ ︷︷ ︸bn terms
, ξ2, . . . , ξ2︸ ︷︷ ︸bn terms
, . . . , ξn/bn , . . . , ξn/bn︸ ︷︷ ︸bn terms
]′.
3 Compute bootstrapped autocorrelations:
ρ(dw)n (h) =
1
γn(0)× 1
n
n∑t=h+1
ωt
[En,t,h(θn)− gn(h, θn)
], h = 1, . . . ,Ln,
and test statistics T (dw)n =
√nmax1≤h≤Ln |ρ
(dw)n (h)|.
4 Repeat Steps 2-3 M times and compute the bootstrappedp-value p
(dw)n,M = 1
M
∑Mi=1 I(T
(dw)n,i ≥ Tn).
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 17 / 49
Monte Carlo Simulations
Two tests:
1 Max-correlation test with DWB.
2 Hong’s test with DWB.
Sample size is n = 500.
Ln = 5,[0.5× n
ln(n)
],[
nln(n)
].
L500 = 5, 40, 80.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 18 / 49
Simulation Results (Selected) – Size
Simple yt = et, Mean Filter, et ∼ IID.Ln = 5 Ln = 40 Ln = 80
1%, 5%, 10% 1%, 5%, 10% 1%, 5%, 10%Max .008, .047, .114 .004, .029, .078 .000, .027, .068Hong .007, .061, .117 .000, .013, .050 .000, .005, .022
AR(2) yt = 0.3yt−1 − 0.15yt−2 + et, AR(2) Filter, et ∼ IID.Ln = 5 Ln = 40 Ln = 80
1%, 5%, 10% 1%, 5%, 10% 1%, 5%, 10%Max .018, .064, .126 .004, .028, .075 .003, .030, .074Hong .016, .068, .135 .001, .012, .039 .000, .005, .022
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 19 / 49
Simulation Results (Selected) – Power
AR(2) yt = 0.3yt−1 − 0.15yt−2 + et, AR(1) Filter, et ∼ GARCH(1,1).Ln = 5 Ln = 40 Ln = 80
1%, 5%, 10% 1%, 5%, 10% 1%, 5%, 10%Max .356, .650, .776 .175, .403, .565 .127, .329, .469Hong .321, .632, .758 .002, .096, .235 .000, .013, .073
Simple yt = et, Mean Filter, et ∼ MA(48).Ln = 5 Ln = 40 Ln = 80
1%, 5%, 10% 1%, 5%, 10% 1%, 5%, 10%Max .013, .082, .154 .007, .055, .141 .764, .925, .957Hong .017, .064, .136 .002, .030, .117 .013, .210, .448
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 20 / 49
Max-Correlation Test: Summary
Shao’s (2010, 2011) dependent wild bootstrap enables us toconstruct a formal white noise test.
We propose a max-correlation test statistic.
Max-correlation test has sharper size and higher power thanHong’s test.
The advantage of max-correlation test is particularly prominentunder remote autocorrelations (seasonality).
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 21 / 49
Weak Form Efficiency: Introduction
A rejection of the white noise hypothesis might serve as helpful
information for arbitragers, because it indicates the presence of
non-zero autocorrelations at some lags.
IID
MDS
White Noise
xx
xx
Testing for Weak Form Efficiency of Stock Markets
= Testing for Unpredictability of Stock Returns (cf. Fama, 1970)
Many applicationscf. Lim and Brooks (2009)
REASON: It is hard to establish a formal white noise test.
BREAKTHROUGH: Shao’s (2010, 2011) dependent wild bootstrap.
Only few applications
This paper tests for the white noise hypothesis of stock returns,
using the dependent wild bootstrap.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 22 / 49
Weak Form Efficiency: Introduction
Adaptive Market Hypothesis
(cf. Lo, 2004; Lo, 2005)
Dependent Wild Bootstrap
(Shao, 2010; Shao, 2011)
Full Sample
Shao (2010): Temperature
Shao (2011): Stock returns
Rolling Window
New!
PERIODICITY IN
CONFIDENCE BANDS
REMEDY: Randomizing a block size across
bootstrap samples and windows.
REASON: Fixed block size produces similar
bootstrapped autocorrelations in every windows.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 23 / 49
Weak Form Efficiency: Introduction
S&P 500
White noise hypothesis is often rejected
during Iraq War and the subprime crisis.
We observe significantly negative
autocorrelations during crisis periods.
Campbell, Grossman, and Wang (1993)
cf. Fama and French (1988)
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 24 / 49
Review of Dependent Wild Bootstrap
Consider full sample analysis as a benchmark.
Consider univariate covariance stationary {y1, . . . , yn}.Assume E[yt] = 0 for notational simplicity.
Population quantities:
γ(0) = E[y2t ], γ(h) = E[ytyt−h], ρ(h) =γ(h)
γ(0).
Sample quantities:
γn(0) =1
n
n∑t=1
y2t , γn(h) =1
n
n∑t=h+1
ytyt−h, ρn(h) =γn(h)
γn(0).
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 25 / 49
Review of Dependent Wild Bootstrap
White noise hypothesis: ρ(h) = 0 for all h ≥ 1.
As a starting point, fix h and consider testing for ρ(h) = 0.
How can we construct a confidence band for ρn(h), assuminglittle more than serial uncorrelatedness?
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 26 / 49
Review of Dependent Wild Bootstrap
1 Set a block size bn (typically bn =√n).
2 Generate iid {ξ1, ξ2, . . . , ξn/bn}. Define an auxiliary variable:
ω = [ ξ1, . . . , ξ1︸ ︷︷ ︸bn terms
, ξ2, . . . , ξ2︸ ︷︷ ︸bn terms
, . . . , ξn/bn , . . . , ξn/bn︸ ︷︷ ︸bn terms
]′.
3 Compute a bootstrapped autocorrelation:
ρ(dw)n (h) =
1
γn(0)× 1
n
n∑t=h+1
ωt[ytyt−h − γn(h)].
4 Repeat Steps 2-3 M times and sort ρ(dw)n,(1)(h) < · · · < ρ
(dw)n,(M)(h).
5 The 95% confidence band is C(h) = [ρ(dw)n,(0.025M)(h), ρ
(dw)n,(0.975M)(h)].
6 If ρn(h) ∈ C(h), then we do not reject ρ(h) = 0.If ρn(h) /∈ C(h), then we reject ρ(h) = 0.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 27 / 49
Hidden Pitfall: Periodic Confidence Bands
Now consider rolling window analysis.
Suppose that window size is n = 60 and block size is bn = 3.
In window #1 (y1, . . . , y60), we have
[ y1y2, y2y3︸ ︷︷ ︸Block 1 (×ξ1)
, y3y4, y4y5, y5y6︸ ︷︷ ︸Block 2 (×ξ2)
, y6y7, y7y8, y8y9︸ ︷︷ ︸Block 3 (×ξ3)
, . . . , y57y58, y58y59, y59y60︸ ︷︷ ︸Block 20 (×ξ20)
].
In window #4 (y4, . . . , y63), we have
[ y4y5, y5y6︸ ︷︷ ︸Block 1 (×ξ1)
, y6y7, y7y8, y8y9︸ ︷︷ ︸Block 2 (×ξ2)
, . . . , y57y58, y58y59, y59y60︸ ︷︷ ︸Block 19 (×ξ19)
, y60y61, y61y62, y62y63︸ ︷︷ ︸Block 20 (×ξ20)
].
Similar bootstrapped autocorrelations appear in windows #1, #4, #7, . . . .
=⇒ Periodicity with bn = 3 cycles.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 28 / 49
Hidden Pitfall: Periodic Confidence Bands
2 4 6 8 10 12-0.5
0
0.5
y1, . . . , y71i.i.d.∼ N(0, 1).
Window size is n = 60. There are 71− 60 + 1 = 12 windows.
Block size is bn = 3.
We plot ρn(1) and 95% confidence band for each window.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 29 / 49
Remedy: Randomized Block Size
Block size is bn = [c√n].
Conventional choice that c = 1 produces periodicity.
We propose to draw c ∼ U(1− δ, 1 + δ) independently acrossrolling windows and bootstrap samples.
Randomness across windows removes periodicity.
Randomness across bootstrap samples reduces the volatility ofconfidence bands.
We choose δ = 0.5 (i.e. c ∼ U(0.5, 1.5)).
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 30 / 49
Remedy: Randomized Block Size
Illustrative Example:
y1, y2, . . . , y400i.i.d.∼ N(0, 1).
Window size is n = 240.
There are 400− 240 + 1 = 161 windows.
Block size is bn = [c√n] = [c
√240] = [c× 15.5].
We choose either c = 1 or c ∼ U(0.5, 1.5).
We plot ρn(1) and 95% confidence band for each window.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 31 / 49
Remedy: Randomized Block Size
50 100 150-0.2
-0.1
0
0.1
0.2
a. c = 1
50 100 150-0.2
-0.1
0
0.1
0.2
b. c ∼ U
When c = 1, confidence bands have clear periodicity.
When c ∼ U(0.5, 1.5), the periodicity evaporates dramatically.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 32 / 49
Stock Price Data: S&P 500
Jan05 Jan10 Jan150
100
200
300
Subprime
Shanghai
AA+
Iraq
a. Level
Jan05 Jan10 Jan15-0.2
-0.1
0
0.1
0.2
Subprime AA+Shanghai
Iraq
b. Log Return
Daily data of S&P 500.
January 1, 2003 – October 29, 2015 (3230 days).
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 33 / 49
Autocorrelations at Lag 1
Jan05 Jan10-0.5
0
0.5
a. c = 1
Jan05 Jan10-0.5
0
0.5
Iraq Subprime AA+
b. c ∼ U
Window size is n = 240 (roughly a year).
Block size is bn = [c√n] = [c× 15.5].
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 34 / 49
Cramer-von Mises White Noise Test
White noise requires that ρ(h) = 0 for all h ≥ 1.
Following Shao (2011), we use the Cramer-von Mises statistic:
Cn = n
∫ π
0
{n−1∑h=1
γn(h)ψh(λ)
}2
dλ, ψh(λ) =sin(hλ)
hπ.
Bootstrapped p-values are computed based on the dependentwild bootstrap (with a randomized block size).
We observe similar results after using Hill and Motegi’s (2017a)max-correlation test and Andrews and Ploberger’s (1996)sup-LM test.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 35 / 49
P-Values of Cramer-von Mises Test
Jan05 Jan100
0.5
1
P-values over rolling windows.
S&P has significant autocorrelations during Iraq War and thesubprime mortgage crisis.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 36 / 49
Weak Form Efficiency: Summary
1. Outline
2. Contributions
3. Empirical Finding
Test for white noise hypothesis of stock returns
Perform rolling window analysis with dependent wild bootstrap
Find that a fixed block size results in periodic confidence bands
Reveal that the periodicity stems from repeated block structures
Propose randomizing a block size to remove the periodicity
White noise hypothesis is rejected for S&P during crisis periods
-- Significantly negative autocorrelations at lag 1
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 37 / 49
Granger Causality Tests: Motivation
Consider two time series {x1, . . . , xn} and {y1, . . . , yn}.
Definition: {xt} does not Granger cause {yt} if
E[ yt+1 | yt, yt−1, . . . ]︸ ︷︷ ︸Univariate prediction
= E[ yt+1 | yt, yt−1, . . . , xt, xt−1, . . . ]︸ ︷︷ ︸Prediction with extra information of x
.
Interpretation: Knowing past and present values of x does notimprove the prediction accuracy of y.
How to test for Granger non-causality from {xt} to {yt}?
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 38 / 49
Max-Causality Test
Classical approach: formulate a regression model
yt = α0 + α1yt−1 + · · ·+ αpyt−p + β1xt−1 + · · ·+ βqxt−q + ut,
Then perform a Wald test with respect to β21 + · · ·+ β2
q = 0 .
Our approach (cf. Ghysels, Hill, and Motegi, 2017):
Model 1 : yt = α0 + α1yt−1 + · · ·+ αpyt−p + β1xt−1 + ut,
......
......
Model q : yt = α0 + α1yt−1 + · · ·+ αpyt−p + βqxt−q + ut.
Then test for max{|β1|, . . . , |βq|} = 0 .
We show via local power analysis and Monte Carlo simulationsthat our test achieves sharper size and higher power.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 39 / 49
Mixed Frequency Granger Causality Test
MIDAS allows for
mixed frequency
� Higher precision
Classical approach requires
single frequency
� Lower precision
InflationAnnounced monthly
GDPAnnounced quarterly
InflationQuarterly average
Temporal aggregation (= Loss of information)
GDPQuarterly
Ghysels, Hill, and Motegi (2016, 2017) expand Granger causality tests
into the mixed data sampling (MIDAS) literature (cf. Ghysels, Santa-
Clara, and Valkanov, 2004).
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 40 / 49
Mixed Frequency Granger Causality Test
x1,t x2,t x3,t x1,t+1 x2,t+1 x3,t+1
yt yt+1
Quarter t Quarter t+1
Classical approach only uses
MIDAS treats as if they were three distinct
quarterly variables.
time
time
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 41 / 49
Mixed Frequency Granger Causality Test
Classical approach: formulate a regression model
yt = α0 +
p∑k=1
αkyt−k +
q∑k=1
βkxt−k + ut
= α0 +
p∑k=1
αkyt−k +
q∑k=1
βk(13x1,t−k + 1
3x2,t−k + 13x3,t−k) + ut.
Then perform a Wald test with respect to∑q
k=1 β2k = 0 .
Our approach (cf. Ghysels, Hill, and Motegi, 2016):
yt = α0 +
p∑k=1
αkyt−k +
q∑k=1
(β1kx1,t−k + β2kx2,t−k + β3kx3,t−k) + ut.
Then perform a Wald test with respect to∑q
k=1
∑3j=1 β
2jk = 0 .
If β1k = β2k = β3k, our model reduces to the classical model.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 42 / 49
MF Causality Tests: Application
We analyze monthly oil prices, monthly inflation,
and quarterly GDP in the U.S.
MIDAS and the classical approach produce
remarkably different results.
The MIDAS-based results seem more reasonable
(e.g. significant causality from oil prices to inflation).
Ghysels, Hill, and Motegi (2016)
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 43 / 49
MF Causality Tests: Application
Ghysels, Hill, and Motegi (2017)
We analyze weekly interest rate spread
and quarterly GDP in the U.S.
MIDAS detects a longer period of significant causality
from spread to GDP than the classical approach.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 44 / 49
Future Work
1 Expanding max tests into general hypothesis testing.
Linear regression
Nonlinear regression
Parameters on the boundary (e.g. no ARCH effects)
2 Elaborating max-variance ratio tests for random walk.
3 Testing for Granger causality between atmosphere and ocean?
Lagged causality due to seasonality?
Mixed frequency data?
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 45 / 49
References
Andrews, D. W. K. and W. Ploberger (1996). Testing for SerialCorrelation against an ARMA(1, 1) Process. Journal of the AmericanStatistical Association, 91, 1331-1342.
Berman, S. M. (1964). Limit Theorems for the Maximum Term inStationary Sequences. Annals of Mathematical Statistics, 35,502-516.
Campbell, J. Y., S. J. Grossman, and J. Wang (1993). TradingVolume and Serial Correlation in Stock Returns. Quarterly Journal ofEconomics, 108, 905-939.
Fama, E. F. (1965). The Behavior of Stock-Market Prices. Journalof Business, 38, 34-105.
Fama, E. F. (1970). Efficient Capital Markets: A Review of Theoryand Empirical Work. Journal of Finance, 25, 383-417.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 46 / 49
References
Fama, E. F. and French, K. R. (1988). Permanent and TemporaryComponents of Stock Prices. Journal of Political Economy, 96, 246-273.
Ghysels, E., J. B. Hill, and K. Motegi (2016). Testing for Granger Causalitywith Mixed Frequency Data. Journal of Econometrics, 192, 207-230.
Ghysels, E., J. B. Hill, and K. Motegi (2017). Simple Granger CausalityTests for Mixed Frequency Data. Revise-and-resubmit stage, Journal ofBusiness and Economic Statistics.
Ghysels, E., P. Santa-Clara, and R. Valkanov (2004). The MIDAS Touch:Mixed Data Sampling Regression Models. Working paper at the UNCChapel Hill and UCLA.
Hill, J. B. and K. Motegi (2017a). A Max-Correlation White Noise Test forWeakly Dependent Time Series. Working paper at the UNC Chapel Hilland Kobe University.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 47 / 49
References
Hill, J. B. and K. Motegi (2017b). Testing for Weak Form Efficiencyof Stock Markets. Working paper at the UNC Chapel Hill and KobeUniversity.
Hong, Y. (1996). Consistent Testing for Serial Correlation ofUnknown Form. Econometrica, 64, 837-864.
Horowitz, J. L., I. N. Lobato, J. C. Nankervis, and N. E. Savin(2006). Bootstrapping the Box-Pierce Q Test: A Robust Test ofUncorrelatedness. Journal of Econometrics, 133, 841-862.
Lim, K.-P. and R. Brooks (2009). The Evolution of Stock MarketEfficiency over Time: A Survey of the Empirical Literature. Journalof Economic Surveys, 25, 69-108.
Ljung, G. M. and G. E. P. Box (1978). On a Measure of Lack of Fitin Time Series Models. Biometrika, 65, 297-303.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 48 / 49
References
Lo, A. W. (2004). The Adaptive Market Hypothesis: MarketEfficiency from an Evolutionary Perspective. Journal of PortfolioManagement, 30, 15-29.
Lo, A. W. (2005). Reconciling Efficiency Markets with BehavioralFinance: The Adaptive Market Hypothesis. Journal of InvestmentConsulting, 7, 21-44.
Shao, X. (2010). The Dependent Wild Bootstrap. Journal of theAmerican Statistical Association, 105, 218-235.
Shao, X. (2011). A Bootstrapped-Assisted Spectral Test of WhiteNoise under Unknown Dependence. Journal of Econometrics, 162,213-224.
Xiao, H. and W. B. Wu (2014). Portmanteau Test and SimultaneousInference for Serial Covariances. Statistica Sinica, 24, 577-600.
Motegi (Economics, Kobe U.) Max-Correlation and Max-Causality Tests February 17, 2017 49 / 49