Professional Documents
Culture Documents
(1)
The mother wavelet must comply with the admissibility condition, which states that
( ) f
df
f
s,t
(t) by simply rescaling and translating :
27
Fig. 0. (a) Representation of the real part of a Morlet wavelet, with
Irequency e0=10. Conditions (1) and (2) ensure that the wavelet is
different from 0 but decays down to 0 fast enough. This wavelet is
the mother wavelet. The admissibility condition also ensures that
it is possible to recover x, the original variable, from its projection
on the wavelet family. (b) A daughter wavelet, rescaled and
translated with parameters s=4 and t2. (c) Another daughter
wavelet, with parameters s0.25 and t2.
where s is the scaling Iactor, and t the
translating one (the square root of s is used
here to keep the unit energy).
Fig. 0(b) and (c) shows two daughter
wavelets obtained by rescaling and
translating the Morlet mother wavelet. Once
the wavelet family is identified, it is possible
to obtain the wavelet transform by projecting
the signal x(t) on this set of functions. The
wavelet coefficients obtained are of the
following form:
is the complex conjugate wavelet oI
s,t
. This
transform is called the Continuous Wavelet
Transform (CWT) as parameters s and t can
take any real value (i.e. there are an infinite
number of wavelet coefficients). However, it
can be shown that information contained by
the wavelet coefficients WT(s,t) is not lost
when s and t are made discrete, which
considerably simplifies the computation of
the wavelet transform. The most common
method to discretize the parameters while keeping the relevant information is to use
parameters of the following form: s=2
j
and tk2
j
where j and k are integers. This new
transform is called the Discrete Wavelet Transform (DWT). So far, we have considered that
the signal was a function of a continuous time variable t. However, in economic time series,
observations are more often than not discrete, which brings us to the following expression of
the wavelet coefficients, strictly equivalent to the expression found in the continuous time
variable case:
where * is the complex conjugate wavelet oI and N is the number of observations. j is the
level of decomposition (the timescale) of the transform, and takes every integer value between
1 and J, where J is the highest level (also called the overall depth of the transform). For j=1,
28
N/2 coefficients are calculated which correspond to short cycles (high frequency components
of periods comprised between 2
1
and 2
2
months). For j=2, N/4 coefficients are calculated of
periods comprised between 2
2
and 2
3
months. Until we get to j=J, the coarsest decomposition
level (N/2
J
coefficients), corresponding to the largest intervals (i.e. long cycles of periods
superior to 2
J
months). It must be noted that N/2
J
>1, which implies that the highest resolution J
that can be obtained with a given discrete signal is necessarily inferior or equal to log
2
(N).
The coefficients' sub sampling method described above presents a drawback in the case of the
DWT: in order to exactly cover the sample of finite observations without having to pad the
sample with extra zeros or by using polynomial extrapolations, we must use time series with
size N equal to a multiple of a power of 2 (dyadic length vector of observations). Indeed, we
need to be able to divide by two the sample size as many times as the decomposition level
requires. In order to overcome this major drawback, as dyadic length vectors of observations
are rare, we can use another kind of transform, the Maximum Overlap Discrete Wavelet
Transform (MODWT). The difference between the MODWT and the DWT lies in the fact
that the translations are now oI the Iorm tk, k being an integer. For this kind of sampling,
there is no more need to extend the observation vector to obtain a dyadic size, as now every
level contains the same number of wavelet coefficients, equal to the number of observations
N. Moreover, the MODWT provides a wavelet variance estimator which is asymptotically
more efficient than the one based on a DWT analysis. As stated in condition (2), the mother
wavelet must integrate to 0. However, the signal we want to decompose does not necessarily
integrate to 0, hence the projection of x(n) on the set of wavelets will only provide
information about its fluctuations, not about its trend. Therefore we need another kind of
Iunction which integrates to 1, the scaling function on which to project the signal, and
which will serve to approximate the trend of the signal. We have therefore the final condition:
_ (t)Jt = 1
+
-
The projection of x on the mother wavelet and its daughters will provide the detail
coefficients, corresponding to the high frequency components of the signal, while its
projection on the scaling function and its daughters will provide the smooth coefficients,
corresponding to the trend. In the case of the MODWT, the projections of x(n) on these two
types of wavelets are of the following forms:
- Detail coefficients:
29
- Smooth coefficients:
These two kinds of coefficients provide an additive decomposition of the original signal,
thanks to the following formula: xt =
This additive decomposition is what we call a multi-resolution analysis, and
which we will use later in this thesis. As this thesis is aiming at better understanding the
relationship between variables, wavelet variance, correlation are very useful tools. The
wavelet variance at level j is given by:
Where: L
j
=(2
j
1)(L1)1 is the width oI the wavelet Iilter at level j, and L being the overall
width of the wavelet. M
j
=NL
j
+1, L is the number of wavelet coefficients unaffected by the
boundary conditions. Assuming that the signal x follows a Gaussian process, we can therefore
deduce the confidence interval for the variance. Then using the variance formula, we can
obtain the correlation which we will use later in this thesis.
5. Data
We collect daily data for the six member-countries of the OECD (Denmark, Finland,
France, Germany, Norway, United Kingdom) over the period from Oct 25, 2006 to Oct 18,
2011. Stock market indices are obtained from MSCI database, while oil price data are
extracted from the Energy Information Administration (EIA). The Brent spot prices are used
to represent the international crude-oil market since they usually serve as reference prices for
pricing crude oil and many other derivatives products using oil as underlying asset. Unlike the
majority of previous studies which employ low frequency data (yearly, quarterly, monthly,
and weekly), we use daily data in order to adequately capture the rapidity and intensity of the
dynamic interactions between oil and stock prices in the OECD. All price data are
denominated in US dollars to take into account the impacts of exchange rates and to ease the
comparison across countries. Daily returns are calculated from daily price data by taking the
natural logarithm of the ratio of two successive prices. The statistical properties of the data are
summarized in Table 1, Table 2.
30
Table 1 : Descriptive Statistics
BRENT DENMARK FINLAND France GERMANY NORWAY UK
MEAN 0.052 -0.003 -0.045 -0.033 -0.015 -0.007 -0.026
MINIMUM -16.832 -13.512 -10.188 -11.566 -9.637 -14.225 -10.421
MAXIMUM 18.130 10.713 11.569 11.844 11.589 15.394 12.161
Std. Dev 2.455 1.946 2.269 2.071 2.017 2.747 1.909
SKEWNESS 0.005 -0.202 -0.051 0.037 -0.049 -0.379 -0.071
KURTOSIS 9.013 8.106 6.199 8.018 7.606 6.988 9.344
Jarque-Bera 1883.428
[0.000]
1366.545
[0.000]
533.718
[0.000]
1312.009
[0.000]
1105.308
[0.000]
858.337
[0.000]
2096.979
[0.000]
ARCH(5) 14.731
[0.000]
82.921
[0.000]
20.874
[0.000]
41.349
[0.000]
41.031
[0.000]
60.683
[0.000]
67.972
[0.000]
ARCH(10) 9.7382
[0.000]
46.48
[0.000]
16.878
[0.000]
27.653
[0.000]
32.928
[0.000]
47.476
[0.000]
39.895
[0.000]
Q(10) 21.767
[0.016]
30.279
[0.000]
15.937
[0.101]
22.247
[0.013]
6.516
[0.770]
13.843
[0.180]
28.802
[0.000]
Q(10) 183.308
[0.000]
963.059
[0.000]
311.107
[0.000]
552.550
[0.000]
604.832
[0.000]
954.934
[0.000]
797.628
[0.000]
Note: This table provides the basic statistics of daily cride-oil returns for 6 countries of the OECD. Data are over the period from Oct 25,
2006 to Oct 18, 2011. They include mean (Mean), standard deviation (Std. dev.), minimum (Min.), maximum (Max.), skewness (Skew.),
kurtosis (Kurt.), and i Ljung-Box Q(10). JB refers to the empirical statistic of the JacqueBera test for normality. ARCH(5), ARCH(10) are
the empirical statistic of the Lagrange Multiplier test for ARCH in the residuals, which considers three lagged values chosen using
information criteria.
In Table 2, we test for the presence of unit roots in the levels (price series in logarithm)
and first differences (return series) of oil and stock market price indices. The results from the
Augmented DickeyFuller (ADF) tests indicate that the price series are integrated of order
one. All the return series do not have unit roots. To check whether our unit root test results are
robust to the presence of a potential structural break we implement the PP unit root test where
break dates are endogenously determined within the model. Our findings show that the null
hypothesis of a unit root cannot be rejected for all the price series with the exception of oil
prices, while it is clearly rejected for all the return series.
We report, in Table 1, basic statistics of return series. Average daily returns on OECD stock
market indices are all negative over our sample period under the effects of the recent global
financial crisis 20072009, sparked by the US subprime crisis. The stock market in Filand
realized the worst performance (0.045%), followed by those in the France and UK.
31
Inversely, the oil market experienced a positive average return,which is not surprising in view
of the increasing trend in the price of oil over the last decade. Skewness is negative for all
stock markets except for France, and positive for the oil market. This means that extreme
negative and positive returns are likely to be realized for stock and oil markets respectively.
Kurtosis coefficients are important in size and highly significant, indicating that outliers may
occur with a probability higher than that of a normal distribution (They have fatter tails and
longer left tails than a normal distribution). Accordingly, the JarqueBera test statistics
strongly reject the null hypothesis of normality for all series. As also shown in Table 1, all the
return series were found to have a leptokurtic behavior (i.e., their distributions have fatter tails
than corresponding normal distributions). This suggests that each of the mean equations
should be tested for the existence of conditional heteroskedasticity. By applying the Engle
(1982)'s test, we observe that the null hypothesis of no ARCH effects is rejected at
conventional levels in all cases, thus confirming that GARCH modeling is adequate for
capturing any persistence in the financial volatility of stock and oil markets we consider. All
these facts support our choice to use the quasi-maximum likelihood (QML) estimation
method.
Table 2 : Unit Root Tests
BRENT DENMARK FINLAND FRANCE GERMANY NORWAY UK
ADF -
34.783
***a
-34.427
***a
-35.950
***a
-
35.784
***a
-34.644
***a
-35.900
***a
-
36.056
***a
PP -
34.779
***a
-34.434
***a
-36.032
***a
-
35.913
***a
-34.645
***a
-35.942
***a
-
36.459
***a
(a) Refers to the Augmented Dickey-Fuller test without trend nor intercept. From these results, the data rejects unit root
Fig. 1 depicts the time-paths followed by different stock market returns & oil price series.
Some signs of volatility clustering and persistence can be observed. Further, the presence of
several sudden changes in the time series may indicate the occurrence of structural breaks. We
also perform the Ljung-Box and Engle (1982) LM ARCH tests to further analyze the
distributional characteristics of oil return series (Table 1). These tests provide clear indication
of autocorrelation and ARCH effects in the series considered. In contrast, the West and Cho
(1995) modified Ljung-Box test, which is robust to conditional heteroscedasticity, shows that
autocorrelation is significantly present at the conventional levels only in seven series .
32
Fig. 1 : Dynamic of OECD Stock markets indices & Brent oil prices
Fig. 2 : Returns
BRENT
2007 2008 2009 2010 2011 2012
50
100
150
BRENT
DENMARK
2007 2008 2009 2010 2011 2012
4000
6000
DENMARK
FINLAND
2007 2008 2009 2010 2011 2012
500
750
1000 FINLAND FRANCE
2007 2008 2009 2010 2011 2012
1000
1500
2000
2500
FRANCE
GERMANY
2007 2008 2009 2010 2011 2012
1500
2500
GERMANY
NORWAY
2007 2008 2009 2010 2011 2012
2000
3000
4000
5000
NORWAY
UK
2007 2008 2009 2010 2011 2012
1000
1500
UK
BRENT
2007 2008 2009 2010 2011 2012
-10
0
10
20
BRENT
DENMARK
2007 2008 2009 2010 2011 2012
-10
0
10
DENMARK
FINLAND
2007 2008 2009 2010 2011 2012
-10
0
10
FINLAND France
2007 2008 2009 2010 2011 2012
-10
0
10
France
GERMANY
2007 2008 2009 2010 2011 2012
0
10
GERMANY NORWAY
2007 2008 2009 2010 2011 2012
-10
0
10
20
NORWAY
UNITED
2007 2008 2009 2010 2011 2012
-10
0
10
UNITED
33
Fig. 3 : Squared returns
As to Figures 2 and 3, Figure 2 shows the time-variations of stock returns through time.
We observe that both Brent oil and stock markets in Europe have experienced alternatives
periods of overvaluation and undervaluation. Typically, all the return series display some
extreme positive and negative returns over the 2008-2009 period, indicating the potential
effects of the global financial crisis. Figure 3 depicts the evolution of the squared returns,
usually viewed as a measure of financial volatility. It can be seen that oil and stock returns are
particularly volatile between the end of 2008 and the beginning of 2009. Moreover, the
volatility dynamics seem to be sharply time-varying and clustered because the periods of high
volatility tend to be followed by periods of high volatility, and periods of low volatility by
periods of low volatility. Hence, a time-varying volatility model such as GARCH-family
processes should be appropriate for modeling the volatility dynamics of Brent oil and stock
market returns.
6. Empirical results
6.1 Return decomposition: A multi-resolution analysis of the oil price and stock returns
As variable for the oil price, we use the logarithm of the spot price Brent, provided by
the Energy Information Agency, on a daily basis, from Oct 2006 until Oct 2011. We therefore
BRENT
2007 2008 2009 2010 2011 2012
100
200
300
BRENT
DENMARK
2007 2008 2009 2010 2011 2012
50
100
150
200
DENMARK
FINLAND
2007 2008 2009 2010 2011 2012
50
100
150
FINLAND FRANCE
2007 2008 2009 2010 2011 2012
50
100
150
FRANCE
GERMANY
2007 2008 2009 2010 2011 2012
50
100
150
GERMANY NORWAY
2007 2008 2009 2010 2011 2012
100
200
NORWAY
UK
2007 2008 2009 2010 2011 2012
50
100
150
UK
34
have a vector of 1251 observations, which enables us to use 7 levels of wavelet
decomposition. As mentioned above, we use the MODWT transform for the decomposition.
Regarding the choice of the wavelet function, we follow the advice given by Genay et al.
(2002) which states that the longer the length of the wavelet, the better it approximates an
ideal band pass filter. Given the shape of the original series, which is rather smooth, we
therefore choose the Least Asymmetric wavelet of level 8, LA(8) (8 is the width of the filter),
as the wavelet function. The various decomposition levels we obtain correspond to time
scales: D1 (2 to 4 months), D2 (4 to 8 months), D3 (8 to 16 months), D4 (16 to 32 months),
D5 (2.7 to 5.3 years), D6 (5.3 to 10.6 years), D7 (10.6 to 21.3 years). We also obtain a smooth
S7 which corresponds to the trend of the signal, while the details Dj correspond to
fluctuations of various sizes. From Fig. 4 to Fig. 10 represent the details D1 to D7 and the
smooth S7, obtained by applying the LA(8) to the oil price & stock market returns variables.
Given the properties of the wavelet decomposition, each detail represents the contribution of
fluctuations of a specific time scale to the oil price variations & stock market returns
variations, while the smooth S7 represents its trend. From these figures, we can see that, in a
country each time scale has a dynamic of return and dynamics of six coutries are different at
the same levels. The shape of the smooth, which is continuously increasing & decreasing
throughout the whole sample. Turning to the wavelet details on Fig. 4-Fig. 10,we note that
they are able to capture the classical oil shock events & stock returns. At the smallest level D1
(2 to 4 month scale),we observe a clear rupture in 2009, corresponding to the economic crisis
2008-2009. At level D2, (4 to 8 month scale) we can see an additional significant rupture in
4/2010. At levels D5(app. 2.7 to 5.3 years scale) and D6(app. 5.3 to 10.6 years scale), smaller
shocks are merged into larger oil price movements. Another feature which can be observed
from Fig. 4-Fig. 10 is the cyclical shape of all details, especially of the larger ones, which
exhibit troughs and peaks in a regular manner. This fact does not constitute a surprise in itself,
as it is common for macroeconomic variables to follow cycles.
It is useful to note that the Multi-Resolution Analysis is an additive decomposition of the
original signal. When considering Fig. 4-Fig. 10, this additive decomposition property implies
that the dot - curve is equal to the sum of all blue curves (d(log(Oilt))=D1+D2+D7+smooth
S7). Therefore,
d],t
0It
the percentage contribution of detail j to the oil price variations at time t.
35
Fig 4 : Multi-resolution of oil price (Logarithm of the Brent, between October 25, 2006 and October 18, 2011 on a daily basis). The top right
cell represents the trend (S7) and the original oil price curve.
Fig 5 : Multi-resolution of stock market returns (Logarithm of DENMARK, between October 25, 2006 and October 18, 2011 on a daily
basis). The top right cell represents the trend (S7) and the original oil price curve..
S BRENT
2007 2008 2009 2010 2011 2012
4.0
4.5
5.0
S BRENT
D1
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
0.2
D1 D2
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D2
D3
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D3 D4
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D4
D5
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D5 D6
2007 2008 2009 2010 2011 2012
0.0
0.1
D6
D7
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
0.2
D7
S DENMARK
2007 2008 2009 2010 2011 2012
8.0
8.5
9.0
S DENMARK D7
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
0.10
D7
D6
2007 2008 2009 2010 2011 2012
0.0
0.1
D6 D5
2007 2008 2009 2010 2011 2012
-0.025
0.000
0.025
0.050
D5
D4
2007 2008 2009 2010 2011 2012
0.00
0.05
D4 D3
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D3
D2
2007 2008 2009 2010 2011 2012
-0.025
0.000
0.025
0.050
D2 D1
2007 2008 2009 2010 2011 2012
-0.025
0.000
0.025
0.050
D1
36
Fig 6 : Multi-resolution of stock market returns (Logarithm of FINLAND, between October 25, 2006 and October 18, 2011 on a daily
basis). The top right cell represents the trend (S7) and the original oil price curve
Fig 7 : Multi-resolution of stock market returns (Logarithm of FRANCE, between October 25, 2006 and October 18, 2011 on a daily
basis). The top right cell represents the trend (S7) and the original oil price curve
S FINLAND
2007 2008 2009 2010 2011 2012
6.0
6.5
7.0
S FINLAND
D7
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D7
D6
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D6 D5
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D5
D4
2007 2008 2009 2010 2011 2012
0.0
0.1
D4 D3
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D3
D2
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D2 D1
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D1
S FRANCE
2007 2008 2009 2010 2011 2012
7.0
7.5
8.0
S FRANCE
D6
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D6
D7
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D7
D4
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D4
D5
2007 2008 2009 2010 2011 2012
0.0
0.1
D5
D2
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
0.10
D2
D3
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
0.10
D3
D1
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D1
37
Fig 8 : Multi-resolution of stock market returns (Logarithm of GERMANY, between October 25, 2006 and October 18, 2011 on a daily
basis). The top right cell represents the trend (S7) and the original oil price curve.
Fig 9 : Multi-resolution of stock market returns (Logarithm of NORWAY, between October 25, 2006 and October 18, 2011 on a daily
basis). The top right cell represents the trend (S7) and the original oil price curve.
S GERMANY
2007 2008 2009 2010 2011 2012
7.0
7.5
8.0
S GERMANY D7
2007 2008 2009 2010 2011 2012
-0.05
0.05
D7
D6
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D6 D5
2007 2008 2009 2010 2011 2012
0.00
0.05
D5
D4
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D4 D3
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D3
D2
2007 2008 2009 2010 2011 2012
-0.025
0.000
0.025
0.050
D2 D1
2007 2008 2009 2010 2011 2012
-0.025
0.000
0.025
0.050
D1
S NORWAY
2007 2008 2009 2010 2011 2012
7.5
8.0
8.5
S NORWAY D7
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
0.2
D7
D6
2007 2008 2009 2010 2011 2012
-0.1
0.0
0.1
D6 D5
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D5
D4
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
0.10
D4 D3
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D3
D2
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D2 D1
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D1
38
Fig 10 : Multi-resolution of stock market returns (Logarithm of UK, between October 25, 2006 and October 18, 2011 on a daily basis).
The top right cell represents the trend (S7) and the original oil price curve.
6.2 Wavelet estimated variance
Fig 11 : Wavelet variance of the time series at all 7 levels with a 95% confidence interval. Time scales are represented in abscissa.
S UK
2007 2008 2009 2010 2011 2012
6.5
7.0
7.5
S UK
D7
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D7
D6
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
0.10
D6 D5
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D5
D4
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D4 D3
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D3
D2
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
D2 D1
2007 2008 2009 2010 2011 2012
-0.05
0.00
0.05
0.10
D1
39
As mentioned above, the method used to decompose the time series in this thesis is the
Maximum Overlap Discrete Wavelet Transform (MODWT), as it allows us to use a sample
which size is not necessarily a multiple of a power of 2, and because its wavelet variance
estimator is asymptotically more efficient than the one based on the DWT. Using an MODWT
variance analysis, with an LA(8) wavelet and 7 levels of decomposition, we obtain the
variance decomposition of our seven variables represented in Figs. 11. We can observe in
Figs. 11 that the variance of 7 variables rises when the level increases. This is not surprising
when considering Fig. 4-10, showing that the higher the level of decomposition, the larger the
dispersion of the wavelet details around their average. There seems to be an almost linear
relationship between the wavelet variance of the oil price variable and the wavelet scale,
suggesting a possible long term dependent structure in the oil price series (Elder and Serletis,
2008; Choi and Hammoudeh, 2009)
6.3 Wavelet correlation
When analysing the relationship between oil price and stock price, one major concern is
whether one is leading the other. To investigate this issue, Fig. 12 shows the wavelet
correlation between the two variables at all seven levels corresponding to 6 pairs ( France
Brent, Filand-Brent, Denmark-Brent, Gemany-Brent, Norway-Brent, UK Brent) . Regarding
Fig. 12, we can see that.
- Each country has a different correlation between oil price and stock price
- Most of the correlations between the two variables appear to be significant at level 6 except
for correlations Brent Denmark and Brent Norway are always positive ( Denmark and
Norway are net oil-exporting countries)
-The correlations between the two variables (France Brent and UK-Brent, Germany-Brent)
only appear to be significant at level 6, with negative values. At the other levels, the
correlation between the two variables is not significantly different from zero or positive. In
fact, it is suitable to France & UK are both oil-exporting and oil-importing countries and
Germany is oil-importing country
- The correlations between the two variables (Brent-Filand) is negative or is not significantly
different from zero. This is consistent with the oil-importing country.
40
Fig 12 : Wavelet correlation of BRENT and Countries Indices
7. Discussions and concluding remarks
While most previous studies rely on a VAR-type methodology to tackle the issue of the
stock market impact of oil price disruptions, this thesis uses a more innovative approach based
on the wavelet theory. This decomposition enables to decompose a signal into various time
scales without losing time related information and to capture the various time scales at which
the factors that influence the oil price operate. It is useful in revealing that the nature of the oil
price stock market relationship is not the same through all time scales, and that a multi-scale
analysis can help unravel the changes that can occur in such relationships when various times
cales are considered. Our main findings may be summarized as follows. First, In a country,
each time scale has a corresponsing dynamic of return. Second, the higher variance when the
level increases, it means that if bigger horizon is higher risk. Third, investigating the
relationship between the oil price and national stock markets indicators proxied by the MSCI
index, we show that at low wavelet levels (high frequency cycles), both the oil price and the
MSCI index seem to entertain a feedback relationship, where they are both leading and
lagging each other. At larger wavelet levels however, only the MSCI happens to be leading
Correlation BRENT-DENMARK
1 2 3 4 5 6 7 8
0.0
0.5
1.0
Correlation BRENT-DENMARK Correlation BRENT-FINLAND
0 1 2 3 4 5 6 7 8
-0.5
0.0
0.5
Correlation BRENT-FINLAND
Correlation BRENT-FRANCE
0 1 2 3 4 5 6 7 8
-0.5
0.0
0.5
1.0
Correlation BRENT-FRANCE Correlation BRENT-GERMANY
0 1 2 3 4 5 6 7 8
-0.5
0.0
0.5
1.0
Correlation BRENT-GERMANY
Correlation BRENT-NORWAY
0 1 2 3 4 5 6 7 8
0.0
0.5
1.0
Correlation BRENT-NORWAY Correlation BRENT-UK
0 1 2 3 4 5 6 7 8
-0.5
0.0
0.5
1.0
Correlation BRENT-UK
41
the oil price with a positive correlation. And this relationship is different corresponding each
coutry. So, the investor should adjust their portfolios at different countries.
A possible extension of this thesis is to investigate the long memory property of energy
prices through a wavelet approach. In particular, relying on the Discrete Wavelet Transform
(DWT) seems to be a promising way to improve the performance of usual estimates of the
fractional difference parameter of long memory processes. Indeed, the DWT provides a
decomposition of any given signal into orthogonal components corresponding to specific time
scales. This is this ability to decorrelate time series that makes DWT so relevant to analyse
long memory processes, and which could be applied to analyse energy prices. The second,
base on time scale we can built suitable invesment portfolios for each country. Third, from
wavelet composition we can find the volatility at any time.
42
References
[1] Grossmann A, Morlet J. Decomposition of hardy functions into square integrable wavelets of constant shape.
SIAM J Math Anal 1984;15:72336.
[2] Percival DB, Walden AT. Wavelet methods for time series analysis. Cambridge, MA: Cambridge University
Press; 2000.
[3] Mallat S. A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans Pattern
Anal Mach Intell 1989;11:67493.
[4] Daubechies I. Ten lectures on wavelets. Philadelphia: SIAM; 1992.
[5] Meyer Y, Coifman R. Wavelets. Cambridge: Cambridge University Press; 1997.
[6] Cao L, Hong Y, Zhao H, Deng S. Predicting economic time series using a nonlinear deterministic technique.
Comput Econom 1996;9(2):14978.
[7] Aussen A, Campbell J, Murthagh F. Wavelet-based feature extraction and decomposition strategies for
financial forecasting. J Comput Intell Finance 1998;6(2):512.
[8] Milidiu RL, Machado RJ, Rentera RP. Time-series forecasting through wavelet transform and a mixture of
expert models. Neurocomputing 1999;28:1456.
[9] Fryzlewicz P, Van Bellegem S, von Sachs R. Forecasting non-stationary time series by wavelet process
modelling. Technical report, Universite catholique de Louvain: Institut de Statistique, 2002.
[10] Ramsey JB. Wavelets in economics and finance: past and future. Stud Nonlinear Dynam Economet
2002;6(3):127.
[11] Daubechies I. Orthonormal bases of compactly supported wavelets. Commun Pure Appl Math
1988;41(7):90996.
[12] Brock WA. Scaling in economics. Indus Corporate Change 1999;8(3).
[13] Mohamed El Hedi Arouri. Does crude oil move stock markets in Europe? A sector investigation 2011
[14] Nguyen et al. Oil prices, stock markets and portfolio investment: Evidence from sector analysis in Europe
over the last decade 2009
[15] Tho Naccache. Oil price cycles and wavelets 2010
[16] Fama, Eugene F., 1981. Stock returns, real activity, inflation, and money. American Economic Review 71
(4), 545565.
[17] Fama, Eugene F., 1990. Stock returns, expected returns, and real activity. Journal of Finance 45 (4), 1089
1108.
[18] Gabor, D., 1946. Theory of communication. Journal of I.E.E. London 93 (3) Nov.
[19] Genay, Ramazan, Seluk, Faruk, Whitcher, Brandon, 2002. An introduction to Wavelets and Other
Filtering Methods in Finance and Economics. Academic Press, Elsevier.
[20] Hamilton, James D., 1989. A new approach to the economic analysis of nonstationary time series and the
business cycle. Econometrica
[21] Hamilton, James D., 1996. This is what happened to the oil pricemacroeconomy relationship. Journal of
Monetary Economics 38
[22] Hamilton, James D., 2008. Understanding Crude Oil Prices. Working Paper. University of California, San
Diego.
[23] Hodrick, R.J., Prescott, E.C., 1997. Post-war business cycles: an empirical investigation. Journal of Money,
Credit & Banking 29, 116.
[24] Hooker, Mark A., 2000. Are Oil Shocks Inflationary? Asymmetric and Nonlinear Specifications versus
Changes in Regime. Federal Reserve Board. Washington, D.C., August.
[25] Hotelling, Harold, 1931. The economics of exhaustible resources. Journal of Political Economy 39, 137
175.
[26] Huang, N.E., Shen, Z., Long, S.R., 1998. The empirical mode decomposition and the Hilbert spectrum for
nonlinear and non stationary time series analysis. Process of the Royal Society of London, A 454, 903995.
[27] Kaul, Gautam, Seyhun, H.N., 1990. Relative price variability, real shocks, and the stock market. The
Journal of Finance XLV (2),
43
[28] Kilian, Lutz, 2005. Exogenous Oil Supply Shocks, How Big are They and How Much Do They Matter for
the US Economy? University of Michigan and CEPR.
[29] Kilian, Lutz, 2006. Not All Oil Prices Shocks Are Alike: Disentangling Demand and Supply Shocks on the
Crude Oil Market. Discussion Paper. International Economics, Center for Economic Policy Research.
[30] Kilian, Lutz, Allessandro, Rebucci, Nikola, Spatafora, 2007. Oil shocks and external imbalances. IMF
Working Paper.
[31] Kitchin, Joseph, 1923. Cycles and trends in economic factors. Review of Economics and Statistics 5 (1),
1016.
[32] Kondratieff, N.D., Stolper, W.F., 1935. The long waves in economic life. The Review of Economic
Statistics 17 (6), 105115.
[33] Krautkraemer, Jeffrey A., 1998. Nonrenewable resource scarcity. Journal of Economic Literature 36 (4),
20652107.
[34] Lee, K., Ni, S., Ratti, R.A., 1995. Oil shocks and the macroeconomy: the role of price variability. The
Energy Journal 16 (4), 3956.
[35] Lucas, R.E., 1975. An equilibrium model of the business cycle. Journal of Political Economy 83, 1113
1144.
[36] Mork, Knut Anton, 1989. Oil and the macroeconomy when prices go up and down: an extension of
Hamilton's results. The Journal of Political Economy 97 (3), 740744 June.
[37] Naccache, Tho, 2010. Slow oil shocks and the weakening of the oil pricemacroeconomy relationship.
Energy Policy 38 (5)
[38] Percival, Walden, 2000. Wavelet methods for time series analysis. Cambridge Series in Statistical and
Probabilistic Mathematics, 2006. [15] Romer, Christina D., 1999. Changes in business cycles: evidence and
explanations. Journal of Economic Perspectives 13, 2344 Spring.
[39] Schmitt-Grohe, Stephanie, Uribe, Martin, 2008. What's news in business cycles. NBER: Working Paper,
W14215. August.
[40] Schumpeter, Joseph A., 1939. Business Cycles. Mc Graw Hill Book Company.