1. 1
The Relationship between sectoral
stock market indices and foreign
exchange market
Evidence from U.S. Equity Market
Thesis of Stavros Gkogkos
University of Macedonia
Department of Economics
Supervisor: Panagiotidis Theodore
February 2016
Thessaloniki
2. 2
Contents
Abstract ........................................................................................................................................................ 3
1. Introduction.............................................................................................................................................. 4
2. Theoretical Literature and Review .......................................................................................................... 5
3. Data and Econometric Methodology..................................................................................................... 10
3.1 Data.................................................................................................................................................. 10
3.2 Unit Root Tests................................................................................................................................. 13
3.2.1 Augmented Dickey Fuller Test .............................................................................................. 13
3.2.2 Phillips Perron Test................................................................................................................ 14
3.3 Standard Granger Causality Test...................................................................................................... 15
3.4 Toda and Yamamoto approach for Granger Causality..................................................................... 16
3.5 Non-parametric Causality Test......................................................................................................... 17
4. Empirical Results .................................................................................................................................... 19
4.1 Unit Root Tests................................................................................................................................. 19
4.2 Standard Granger Causality Test...................................................................................................... 24
4.3 Toda and Yamamoto approach for Granger Causality.....................................................................34
4.4 Non-parametric Causality Test......................................................................................................... 44
5. Conclusion .............................................................................................................................................. 52
6. References.............................................................................................................................................. 54
3. 3
Abstract
The aim of this study is to investigate the nature of the causal relationship between stock prices and
exchange rates, using data from 7 January 2005 to 7 January 2015 about the U.S. equity market. The
research covers a sample of 25 sectoral stock market indices along with the NASDAQ Composite and
S&P 500 index, used as the stock prices (SP) variables. Moreover, we used 12 nominal exchange rates of
major currencies against the U.S. dollar and the two traded weighted U.S. dollar indices (broad and
major currencies) in order to examine the causal linkages between these two financial variables. The
existence of the causation linkage between stock prices and exchange rates have preoccupied the minds
of the economists and econometricians for the last two decades, increasing the number of the empirical
work, occupied with the causality between stock prices and exchange rates.
In this thesis, we tried to explore the possible causal relationships, utilizing standard Granger
Causality (Granger, 1969) for the short-run analysis and the Causality test suggested by Toda &
Yamamoto (1995) for the long-run. However, Bekiros and Diks (2007) mention that tests based on
residuals will be sensitive only to causality in the conditional mean while co-variables may influence the
conditional distribution of the response in nonlinear ways. Furthermore, Baek and Brock (1992) noted
that linear Granger causality tests have low power against certain nonlinear alternatives, because
nonlinear tests place direct emphasis on prediction without imposing the existence of a linear functional
form (Bekiros and Diks, 2007, p.4). In this paper, we employed the nonlinear Granger Causality test
suggested by Diks and Pachenko (2006) to examine the nonlinear effects between the variables. In the
linear analysis, the results indicate strong causal relationship, running from stock prices to exchange
rates, while weak causality of the reverse direction is found in several cases. The results of nonlinear
tests differ from those of parametric tests. The Diks & Pachenko (DP) test shows strong bidirectional
causality in the most of cases. In summary, there is no consensus of the results, leading us to mixed
conclusions, with the exception of some sectors.
4. 4
1. Introduction
Many factors, such as enterprise performance, dividends, stock prices of other countries, gross domestic
product, exchange rates, interest rates, current account, money supply, employment, various external
information (political news) etc. have an important impact on daily stock prices (Kurihara, 2006).
Especially, the continuing increases in the world trade and capital movements in conjunction with the
emergence of new capital markets, the relaxation of foreign capital controls and the adoption of more
flexible exchange rate regimes, have made the exchange rates as one of the main determinants of
business profitability and equity prices (Amalendu, 2012, p.47; Kim, 2003).
The gradual abolition of foreign exchange controls in emerging economies has opened the road to
international investors, creating new opportunities of investments and portfolio diversification. The
adoption of flexible exchange rate regimes has increased the risk associated with the international
investments. This means that the choice of the currency denomination plays a crucial role to the
portfolio decision, as it affects the value of the investorβs portfolio. Hence, the dynamic relationship
between exchange rates and stock prices has drawn the attention of many researchers and
professionals who are interested in the prices of financial assets.
This study focuses on the interaction between stock prices and exchange rates, using both
theoretical and empirical analysis. The innovation of this paper is that U.S. sectoral stock market indices
were used as the stock prices (SP) variables, in order to examine the importance of the major industries
of U.S. economy in prediction the exchange rate movements and vice versa. If exchange rate market is
found to lead the stock market indices, the emphasis of the government policy ought to be placed on
controlling the exchange rate. In contrast, domestic economic policies are the priority in stabilizing stock
market in the case where stock market leads the exchange rate market (Granger, Huang and Yang, 2000,
p.1).
From theoretical perspective, there are two sets of models, namely βflow-oriented modelβ or goods
market approach, first discussed by Dornbusch and Fisher (1980) and βstock-oriented modelβ or
portfolio balance model (Branson, 1983; Frankel, 1983). Flow-oriented models suggest that a countryβs
current account and trade balance performance are two crucial factors of exchange rate determination.
This means that stock prices and exchange rates are positively related. Portfolio balance model assumes
that the major factor of exchange rate determination is the capital account, thus the direction of the
causality runs from stock market to foreign exchange market and the two variables are negatively
related.
As for the empirical literature, the results of the past studies are mixed, while researchers used
mostly linear granger causality tests, cointegration techniques and impulse response functions to
investigate the short-run and long-run relationship between the variables.
The contribution of this thesis is as follows: (1) an extended dataset is used (27 sectoral stock
indices and 14 exchange rates), (2) two variations of Granger Causality tests are estimated (Toda &
Yamamoto and Standard Granger Causality test) and (3) the non-parametric causality test developed by
Diks and Pachenko (DP) is employed to capture the non-linear effects between the variables.
The study is organized as follows. Section 2 contains information about the theoretical framework
of the issue and reviews the results of some previous studies in this subject. In section 3, is explained the
methodology and described the data employed in the study. In section 4, are discussed the empirical
results. The last section concludes.
5. 5
2. Theoretical Literature and Review
As mentioned previously, there are two subsets of theoretical models, discussing the interaction
between exchange rates and equity market. According to Granger, Huang and Yang (2000), these
models are renowned as traditional (goods market approach) and portfolio approaches. The traditional
approach postulates that exchange rate movements cause movements in stock prices, while the
portfolio balance models assume the opposite to flow-oriented models, that is, that changes in stock
prices cause movements in exchange rates via capital account. Ramasamy and Yeung (2005) notes that a
possible reason for these divergent results is that the nature of the interaction between stock and
foreign exchange markets is sensitive to the stage of the business cycle and wider economic factors,
such as developments or changes in market structures within an economy.
The flow-oriented model, developed by Dornbusch and Fisher, posits that exchange rates
movements affect international competiveness and the trade balance which have an effect in the real
output of the economy. This affects the current and future cash flows of the companies and their stock
prices1
. At this stage, it is important to distinguish if the company is multinational or domestic (Frank
and Young et. al. 1972). In the case, we refer to multinational company, changes in the value of the
currency (used by the firm for international transactions); have an important impact to companyβs
foreign operations. The profits or losses generated by the foreign operations appear on the firmβs
balance sheet which will affect its market value and stock price. On the other hand, if the company is
domestic, it is not affected directly by the exchange rate fluctuations. This kind of entities may be
affected only in the case they import the means of production from foreign countries. The appreciation
or depreciation of the domestic currency will raise or reduce the cost of production which will have
impact to the companyβs sales and its earnings (prices of the products will increase or decrease
according to the cost of production). Consequently, the market value of the company and then its stock
price is affected indirectly by the exchange rate fluctuations in positive or negative way. In recent years,
that theoretical scenario has its critics, because the majority of the companies utilize hedging techniques
via financial derivatives2
.
On the other hand, portfolio balance models argue that the movements in stock markets may cause
changes in FX (foreign exchange) markets, emphasizing the role of the capital account as the main factor
of the exchange rate determination. In this category, there are two subsets, namely portfolio balance
models and monetary models (Alagidede, Panagiotidis and Zhang, 2011, p.67). Portfolio balance models
(Branson, 1983; Frankel, 1983) assume that the increase in stock prices attracts the interest of the
foreign investors who want to buy domestic assets. This action, cause an increase in the demand of the
domestic currency which leads to its appreciation. In other words, a bullish domestic stock market will
signal favorable domestic economic prospects, thereby inducing capital inflows and an appreciation of
the exchange rate (Caporale, Hunter and Ali, 2014, p.88). Monetary models (Gavin, 1989), posit that
there is no linkage between exchange rates and stock prices, unless the effects of the stock market are
large enough and if the money is not βtoo neutralβ (Gavin, 1989, p.196). If, however, both traditional
1
The stock prices represent the discounted present value of firmβs future cash flows.
2
The investor in order to reduce the risk associated with the investment buys a contract which serves as tool for
hedging.
6. 6
and portfolio approach are empirically relevant, a bidirectional relation between the two variables will
be found with an arbitrary correlation (Granger et. al. 2000).
The empirical literature is rich but the results provided by them are mixed. The main categories of
econometric methods used in order to examine the possible relationship between stock prices and
foreign exchange market are three. The first category includes the empirical research related with the
Granger Causality framework and its variations (linear and non-linear methods). The second one
contains the studies related with the long-run analysis and the cointegration techniques (Engle-Granger
two step procedure and Johahnsen cointegration test). The last category contains the empirical studies
utilized regression analysis (OLS) and different versions of GARCH models (UEDCC-GARCH).
Aggarwal (1981) was one of the first who investigated the relationship between stock prices and
exchange rates. His study provided some evidence in support of the traditional model. The study
examined the relationship between the variables by looking at the correlation between changes in the
U.S. trade weighted exchange rate and changes in U.S. stock market indices (NYSE, S&P 500, DC 500),
using simple regression analysis (OLS estimation) in monthly data for the period 1974 to 1978. The
results show that the traded-weighted exchange rate and the U.S. stock market indices were positively
correlated during this period. These empirical findings lead Aggarwal (1981) to support the flow model.
The theoretical explanation given is that the movements of the exchange rates could directly affect the
stock prices of multinational firmβs (influence the value of its multinational operations) and indirectly
effect domestic firms (the prices of its exports or imports are affected by the exchange rates
fluctuations).
One of the first studies, utilizing Granger causality and cointegration techniques, Bahmani-Oskooee
and Sohrabian (1992) examine the relationship between stock prices and exchange rates, processing
monthly data of the S&P 500 and the effective exchange rate of the U.S. Dollar from July 1973 to
December 1988 (186 observations). They found bidirectional causality between the variables, which
supports both traditional and portfolio approaches. As regard, the long-run analysis, they employed the
methodology suggested by Engle and Granger (1987) and found that the variables were not
cointegrated (no long-run relationship).
Ajayi and Mougoue (1996), searched the relationship between exchange rates and stock indices for
eight advanced economies (Canada, France, Germany, Italy, Japan, Netherlands, UK, US) using the Engle-
Granger two step procedure and ECM (Error Correction Model) for a sample period with daily data from
1985 to 1991. The results show that there are significant short-run and long-run feedback relations
between stock indices and exchange rates. An increase in stock price has a negative short-run effect as
well as a positive long-run effect on domestic currency value (Amalendu, 2012, p.49). Abdalla and
Murinde (1997) searched stock prices-exchange rates relationships in the Asian financial markets of
India, Korea, Pakistan and the Philippines using monthly data from 1985 to 1994 and found
unidirectional causality running from exchange rates to stock prices in India, Korea and Pakistan. On the
other hand, the reverse direction of causality was evidenced for the Philippines.
The Asian financial crisis of the late 1990s attracted the interest in the interaction between
currency and stock markets solely in developing countries. This crisis was characterized by huge
decreases in stock market indices and currency depreciations. Granger (2000, p.1) referred some
incidents of Asian crisis which are worthy of mention: βThe financial crisis sparked in Thailand in July,
1997 has sent shock waves throughout Southeast Asia, South Korea and Japan. On October 27, the
7. 7
short-run interest rate in Hong Kong took a huge jump in order to maintain its pegged exchange rate to
the U.S. dollars. As a result, the Hang Seng Index plummeted 1438 points, setting off the crash in the
U.S. with the Dow Jones Industrial Average down by 554.26 points. On November 24, Yamaichi-the
fourth largest financial corporation-filed for bankruptcy which gave rises to an 854-point drop in Nikkei
Index. In mid-December, the Korean won depreciated drastically from 888 wons (per dollar) in July 1 to
more than 2000 wons. The currency crisis in South Korea set off a financial avalanche in its stock
markets which witnessed a 50.3% freefall. Similar debacles also occurred in other Asian marketsβ.
Granger et al. (2000) focused in Asian stock and currency markets, especially in Hong Kong, Indonesia,
Japan, South Korea, Malaysia, the Philippines, Singapore, Thailand and Taiwan, using cointegration
models, granger causality tests coupled with impulse response functions. The sample period starts from
January 3, 1986 to November 14, 1997 (3097 daily observations). It is found that data from Japan, Hong
Kong and Thailand are in agreement with the flow oriented model, so that exchange rates leads stock
prices with positive correlation. On the other hand, data from Taiwan supports the portfolio approach
while data from Indonesia, Korea, Malaysia and the Philippines indicate strong bidirectional relations.
Only the data from Singapore fails to reveal any recognizable pattern.
Another study, which used cointegration techniques in order to examine the relationship between
stock and foreign exchange markets, was conducted by Nieh and Lee (2001). Their paper employed daily
data (618 observations) of closing stock market indices and foreign exchange rates from October 1993
to February 1996 for the G-7 Countries (Canada, France, Germany, Italy, Japan, UK and the US). Two
alternative methodologies were employed in this study: 1) Engle and Granger (1987) two-step
methodology and 2) Johansen multivariate maximum likelihood cointegration test. The researchers
support the Bahmani-Oskooee and Sohrabianβs (1992) finding that there is no long-run equilibrium
relationship between the variables for each G-7 country. However, short-run significant relationship has
only been found for one day for certain G-7 countries. To be more specific, the results from the VECM
estimation (Vector Error Correction Model) show the existence of short-run predictability (no more than
two consecutive trading days). The last conclusion of this paper is that the US fails to show any
significant correlation either in short-run or long-run.
Lean, Halim and Won (2003), in the same spirit with Granger (2000), examined the relationship
between exchange rates and stock prices on the eight Asian countries badly hit by Asian Financial Crisis.
The sample consists of weekly stock market indices and exchange rates for eight major Asian countries,
namely Hong Kong, Indonesia, Japan, Korea, Malaysia, the Philippines, Singapore, and Thailand. The
study utilized Engle and Granger two-step methodology and Granger Causality test. The empirical results
show that before the Asian Financial Crisis, all countries, except the Philippines and Malaysia,
experience no evidence of Granger Causality between exchange rates and the stock prices. However,
the causality but not the cointegration between capital and currency markets seems to become strong
during the Financial Crisis period (1997). After the 11 September-terrorist-attack, the causality
relationship between the two financial variables returns to normal as in the pre-crisis period and the
cointegration relationship between exchange rates and stock prices weakens.
Kim (2003) investigated the existence of long-run equilibrium relationships among the aggregate
stock price, industrial production, real exchange rate, interest rate, and inflation in the U.S. The study
employed the Johansenβs cointegration approach in monthly data from 1974 to 1998. It was found that
the S&P 500 stock index is positively related to the industrial production but negatively to the real
8. 8
exchange rate, interest rate, and inflation. Short-run analysis of error correction mechanism (VECM)
showed that the stock price, industrial production, and inflation adjusted to correct disequilibrium
among the five variables.
For five Pacific Basin countries (Hong Kong, Malaysia, Singapore, Thailand and Philippines),
Phylaktis and Ravazzolo (2005) investigate the dynamic linkages between stock prices and exchange
rates, using monthly data from 1980 to 1998 and conclude to the following: The authors employed
cointegration methodology (Johansen) and multivariate Granger causality tests for cointegrating
systems suggested by Dolado and Lutkepol (1996). They found that there is no long-run relationship
(cointegration) in these countries, except Hong Kong. Foreign exchange restrictions were not significant
and did not distort the results regarding the linkages between domestic stock and foreign exchange
markets. The multivariate causality tests showed that the U.S. stock market leads the system which
simply means that the U.S. market, with its dominance in the world market place, is the most influential
producer of information (Phylaktis & Ravazzolo, 2005, p.1044).
Hatemi and Roca (2005) examined the relationship between stock and currency markets, using a
sample of daily data from 1 January to 31 December 1997 for ASEAN 4 countries-Malaysia, Indonesia,
Philippines and Thailand. The sample was divided into two sub-periods, the one was from 1 January to 1
July 1997 (pre-crisis) and the second from 2 July to 31 December 1997 (representing the crisis period).
The estimations were conducted using bootstrap causality tests with leveraged adjustments as
introduced in the paper of Hacker and Hatemi-J (2003). This technique overcomes problems associated
with the non-normalities and ARCH effects in the data. The results of the estimations showed that
during the pre-crisis period with the exception of the Philippines, exchange and stock markets are
significantly related with the direction of causality running from exchange rates to stock prices in the
case of Indonesia and Thailand and from stock prices to exchange rates in the case of Malaysia. During
the Asian crisis period, this relationship did not exist in any of the countries.
The recent empirical literature has mainly focused on Granger causality framework and especially
on non-linear causality (Hiemstra-Jones, 1994 and Diks-Pachenko, 2006) and other methods such
bootstrap causality and UEDCC-GARCH. Alagidede, Panagiotidis and Zhang (2011) conducted one of the
recent studies which employ non-parametric Granger causality. Their article examined the relationship
between stock markets and foreign exchange markets in Australia, Canada, Japan, Switzerland and UK
from January 1992 to December 2005. The econometric methods used in their study are the following:
1) two variations of testing for cointegration (Johansen, 1995 and Saikkonen-Lutkepol test, 2001), 2)
three versions of testing for Granger causality (standard Granger causality, Hsiaoβs version of Granger
causality test and causality through a VECM approach), 3) non-parametric causality test proposed by
Hiemstra and Jones (1994). The research provided the following significant results: First, the data
provided evidence that there is no long-run equilibrium relationship between the two variables using
the two methods which mentioned previously. Second, the results from Granger causality tests are
similar. The causal linkage found from exchange rates to stock prices in Canada, UK and Switzerland. In
Hsiaoβs version test, causal relationship was found only for Switzerland and the direction running from
stock prices to exchange rates. As for standard and unequal lag length version of Granger causality test,
the null hypothesis of no causality from stock prices to exchange rates is rejected at 10% level of
significance in the case of Switzerland. The last result provided by this study related with the estimation
of non-linear causality test developed by Hiemstra-Jones. It was found that there is causal linkage from
9. 9
stock price to exchange rate for Japan, and for some lag combinations, weak causality from foreign to
stock market is found for Switzerland (Alagidede, Panagiotidis and Zhang, 2011, p.83).
Katechos (2011) proposed a new approach in order to investigate the relationship between
exchange rates and stock prices. The study introduced an alternative method where one global variable-
global equity market returns-is considered to influence the exchange rates changes, with the relative
interest rate level of a currency determining the sign of the relationship (Katechos, 2011, p.550). The
sample consists of weekly observations from January 1999 to August 2010. The econometric model
applied in this paper contains regression analysis (OLS estimation) as well as GARCH models specification
(estimation of variance equation), especially ML-GARCH (Maximum Likelihood GARCH model). The
empirical findings show that exchange rates and global stock market returns are significantly related.
The value of higher yielding currencies is positively related to global stock market returns, whereas the
value of lower yielding currencies is negatively related (Katechos, 2011, p.558). A stronger relationship is
evidenced when interest rate differentials are relatively large, while the explanatory power of the model
is weakened in the case where interest rate differentials are narrow.
Amalendu (2012) and Ghulam, Melati and Sayyed (2013) are two studies which utilized Toda and
Yamamoto (1995) approach for Granger causality. The first study used data from 2 April 2001 to 31
March 2011 about India and national, services, financials, industrials and technology indices are taken as
stock price indices. The empirical results show that there is bidirectional causality between exchange
rate and all sectoral indices. The second study employed weekly data from BRIC (Brazil, Russia, India and
China) countries from May 2003 to September 2010. The sample was divided into 3 sub-periods (pre-
crisis, during-crisis and post-crisis). In first sub-period, the results indicate that in Brazil was found
bidirectional causal relationships and Russia and India showed uni-directional causality running form
stock market to currency market. During crisis, bidirectional causality found in Russia, while India and
Brazil evidenced with uni-directional causality moving from stock prices to exchange rates. In post crisis
sub-period, Brazil and Russia showed bidirectional causality and India evidenced with uni-directional
causality running from exchange rates to stock prices. China did not provide significant results in any of
all three sub-periods.
Caporale, Hunter and Ali (2014) examined the nature of the linkages between stock prices and
exchange rates in six advanced economies (United States, UK, Canada, Japan, Euro area and Switzerland)
using weekly data (Wednesday to Wednesday) from August 2003 to December 2011 (441 observations).
The sample was parted in a pre-crisis sub-sample and a crisis-period sub-sample. The researchers
employed a bivariate VAR-GARCH model. To be more specific, the first step was a differenced VAR
specification (conditional mean equation), while the second step was the estimation of the UEDCC-
GARCH model developed by Conrad and Karanasos (2010) to capture the joint volatility dynamics
between the variables. The results of the UEDCC-GARCH estimation showed unidirectional Granger
causality from stock returns to exchange rate changes in the US and the UK, in the opposite direction in
Canada and feedback relationship in the euro area and Switzerland. Moreover, causality-in-variance was
found from stock returns to exchange rate changes in the US and in the opposite direction in the euro
area and Japan, while bidirectional causality was evidenced in Switzerland and Canada.
Nazlioglu, Kar and Akel (2014) investigated the causal linkages between stock and currency markets
using linear and non-linear Granger causality tests in nine transition economies (Bulgaria, Czech
Republic, Estonia, Hungary, Latvia, Lithuania, Poland, Romania and Russia). The dataset used in this
10. 10
paper consists of daily data from 1995 to 2011. The econometric models used for the estimations are
the long-run Toda-Yamamoto (1995) causality procedure and the non-linear causality test developed by
Diks-Pachenko (2006). The results for the linear causality test showed that there is unidirectional
causality running from exchange rates to stock prices in the cases of Czech Republic, Hungary, Poland,
Romania and bidirectional causal relationship only in the case of Russia. As for the Diks and Pachenko
(DP) test, bidirectional relationship was found in the case of Russia, while all the other countries
provided insignificant results.
One of the most recent studies related with the causal relationship between stock prices and
exchange rates, is the study of Liang, Chen and Yang (2015). In this paper, bootstrap panel Granger
causality approach (Konya, 2006) is employed in order to investigate the relationship between the
variables in an inter-country framework. The dataset used in the research covered the ASEAN-5
countries (Indonesia, Malaysia, Singapore, Philippines and Thailand) during January 2000-August 2013
period. The empirical findings indicated unidirectional causality from stock prices to exchange rates in
Malaysia, the Philippines and Thailand while causation in the opposite direction was found in Indonesia.
There is no significant relationship between the two variables in Singapore.
In summary, the empirical literature provided us with mixed results. To be more specific, it is not
found a clear pattern which could explain the interaction between foreign exchange and equity markets.
The next sections of my thesis try to shed a light on this important topic, through empirical analysis and
methods.
3. Data and Econometric methodology
In this thesis, we employed daily data (5 days the week) to analyze the dynamic linkages between stock
market indices and exchange rates. As for the stock price variables, we used the U.S. sectoral stock
market indices and the nominal exchange rates of twelve basic currencies expressed in U.S. dollars as
the exchange rate variables. Furthermore, the traded-weighted indices (broad and major currencies)
were used in the study, because they provide a useful measure of competiveness of U.S. economy. The
sample period starts from 7 January 2005 to 7 January 2015 (2515 observations).
As for the econometric methodology followed in this study, three versions of Granger causality are
applied. The first step is to determine whether the series used in the study are stationary or non-
stationary. For this purpose, Augmented Dickey Fuller and Phillips-Perron tests are applied. In short-run
analysis, standard Granger causality test (Granger, 1969) is used in differenced data (stationary) in a VAR
framework, while in long-run analysis, Toda and Yamamoto (1995) is applied on the levels (non-
stationary) of the variables. The third version of Granger causality is a non-linear test proposed by Diks
and Pachenko (2006), via which we examine the non-linear nature of the relationship between stock
and currency markets.
3.1 Data
As referred previously, the innovation of this study is the fact that sectoral analysis is employed, using
data from U.S. equity market. The data consists of twenty five sectoral stock indices, two market indices,
twelve nominal exchange rates against U.S. Dollar (all currencies expressed in U.S. Dollars) and two
11. 11
traded weighted indices of U.S. Dollar. The stock price variables divided into three categories. The first
category consists of S&P 500 index and its ten major sectors, namely consumer discretionary3
, consumer
staples4
, health and care, industrials, information technology, materials, telecommunication services,
utilities5
, financials and energy sector. The second category includes the NASDAQ sectors (bank,
industrial, insurance, computer, telecommunications, biotechnology, other finance and transportation)
and the NASDAQ Composite stock index. The last category includes the NASDAQ PHLX sectoral indices,
namely defense, drug, gold and silver, housing, oil service, semiconductor and utility sector. As for
exchange rate variables, we used the following nominal exchange rates against U.S. Dollar: Japanese
Yen, British Pound, Euro, Canadian Dollar, Swiss Francs, Swedish Kronor, Australian Dollar, Brazilian Real,
Mexican New Peso, Malaysian Ringgit, Thai Baht and Turkish Lira.
Furthermore, we used two traded weighted U.S. dollar indices. The first one is called traded
weighted U.S. dollar index: Broad. This index represents the weighted average of the foreign exchange
value of the U.S. dollar against the currencies of a broad group of major U.S. trading partners. Broad
currency index includes the Euro Area, Canada, Japan, Mexico, China, United Kingdom, Taiwan, Korea,
Singapore, Hong Kong, Malaysia, Brazil, Switzerland, Thailand, Philippines, Australia, Indonesia, India,
Israel, Saudi Arabia, Russia, Sweden, Argentina, Venezuela, Chile and Colombia. The second one is called
traded weighted U.S. dollar index: Major currencies. This specific index, as shown by its name, consists
of a subset of the broad index currencies that circulate widely outside the country of issue. Major
currencies index includes the Euro Area, Canada, Japan, United Kingdom, Switzerland, Australia, and
Sweden (the information obtained from FRED database).
As we refer above, the two traded weighted indices will play a significant role in our empirical
analysis, because they gather all the information from all the countries which have commercial and
financial transactions with United States into one variable. The nominal exchange rates as
macroeconomic variables attract the interest of policy makers and international investors, because the
fluctuations of exchange rates determine the monetary policy, the profitability and the investorsβ
portfolio decisions.
The sectoral analysis could provide us with useful information about the performance of specific
industries of the economy. Moreover, this type of analysis could benefit the portfolio decision making,
through the predictability which can be achieved using causality measures, in the case we found a
recognizable pattern of causal relationship between the two variables.
The Tables 1 and 2 provide some useful information about the variables used in this study and the
names given to the variables (see next page). The Table 1 presents the stock price variables while the
Table 2 provides with information about exchange rates variables.
3
The sector of the economy that consists of businesses which produce nonessential goods and services.
Companies in this sector include retailers, media companies, consumer durables, automobiles and components
companies.
4
This sector includes companies that sell essential products such as food, beverages and household items.
Consumer staples sector stocks are considered to be non-cyclical which means that they are always in demand
regardless the performance of the economy.
5
This sector includes stocks for utilities such as gas and power. The utilities sector contains firms such as electric,
gas and water firms.
12. 12
Table 1. Sectoral Stock Market indices
Sectoral Index Name Sectors Variable Name
S&P 500 Consumer Discretionary Consumer Discretionary COND
S&P 500 Consumer Staples Consumer Staples CONS
S&P 500 Health and Care Health and Care HLTH
S&P 500 Industrials Industrials INDU
S&P 500 Information Technology Information Technology INFT
S&P 500 Materials Materials MATR
S&P 500 Telecommunication Services Telecommunication Services TELS
S&P 500 Utilities Utilities UTIL
S&P 500 Financials Financials FINL
S&P 500 Energy Energy ENER
S&P 500 Index - SP500
NASDAQ Bank sector Financials IXBK
NASDAQ Industrial Industrials IXID
NASDAQ Insurance Financials IXIS
NASDAQ Computer Information Technology IXCO
NASDAQ Telecommunications Telecommunication Services IXTC
NASDAQ Biotechnology Biotechnology NBI
NASDAQ Other Finance Financials IXFN
NASDAQ Transportation Transportation IXTR
NASDAQ Composite - IXIC
PHLX Defense Sector Defense DFX
PHLX Drug Sector Pharmaceuticals RXS
PHLX Gold and Silver Sector Gold and Silver XAU
PHLX Housing Sector Housing HGX
PHLX Oil Services Sector Oil Services OSX
PHLX Semiconductor Sector Semiconductors SOX
PHLX Utility Sector Utilities UTY
Table 2. Exchange Rates and traded weighted indices of U.S. Dollar
Exchange Rates Variable Name
Japanese Yen against USD JPY
British Pound against USD GBP
Euro against USD EUR
Canadian Dollar against USD CAD
Swiss Francs against USD CHF
Swedish Kronor against USD SEK
Australian Dollar against USD AUD
Brazilian Real against USD BRL
Mexican Pesos against USD MXN
Malaysian Ringgit against USD MYR
Thai Baht against USD THB
Turkish Lira against USD TRY
Traded weighted US Dollar Index, Major currencies TWEXM
Traded weighted US Dollar Index, Broad TWEXB
13. 13
The data were downloaded from four different databases. The sectoral indices were obtained from SPDJ
McGraw & Hill and NASDAQ database while the exchange rates were downloaded from Quandl
platform. The traded weighted indices were received from St. Louis Federal Reserve (FRED) database.
Lastly, logarithmic form of the data is used in Toda and Yamamoto approach (levels) while the returns
(first logarithmic differences) were used in standard Granger causality tests and Diks & Pachenko
approach.
3.2 Unit root tests
3.2.1 Augmented Dickey Fuller Tests
The first step of our methodology is to determine whether the time series in our model are stationary or
non-stationary. It is very important to employ stationarity tests, because time series models assume that
the variables used in them are stationary. Non-stationary series may lead to spurious results (spurious
regression). In our analysis, standard Granger causality acquires stationary time series, while Toda and
Yamamoto procedure allows the estimation of a VAR model formulated in levels of the variables. A
series is said to be integrated of order k, denoted I(k), if k is the number the series must be differenced
to be stationary (Ajayi and Mougoue, 1996, p.196). Thus, an I(1) series is non-stationary and first
differences of the same series is stationary, namely I(0). I(1) series indicate that the series contains one
unit root.
In this paper two unit root tests are implemented. The first one is the stationarity test proposed by
Dickey and Fuller (1981). Specifically, we employed the ADF (Augmented Dickey Fuller) test based on
OLS regression. Dickey-Fuller test assume that the error term (ut) of the Autoregressive model (AR) is
white noise and there is no autocorrelation. However, if the dependent variable of the model (Ξyt) is
serially correlated with the past observations of the same variable (autocorrelation) then will have
autocorrelation in the error term ut. The solution to this problem is to introduce p lags of the dependent
variable in our model. The Dickey and Fuller equation is based on the regression,
π₯π¦π‘ = π0 + π0π‘ + π1π¦π‘β1 + π’π‘ (1)
Where, c0 is a constant, and t denotes time trend. The equation can also be estimated in the case c0
or/and a0 are equal to 0. The ADF equation has the following form,
π₯π¦π‘ = π0 + π0π‘ + π1π¦π‘β1 + ππ
π
π=1 π₯π¦π‘βπ + π’π‘ (2)
Where, c0 is a constant, t denotes time trend, Ξyt-j are the j-lags of the dependent variable, and ut the
error term.
The augmented Dickey-Fuller test is a pseudo t-statistic for the null hypothesis (H0) that c1=0
against the alternative hypothesis (H1) c1<0. In other words, H0: I(1) or integration of higher order
14. 14
against the alternative H1: I(0). However, in case we cannot reject the null hypothesis, we must continue
to the estimation of the following ADF regression and perform the test once more,
π₯2
π¦π‘ = π0 + π0π‘ + π1π₯π¦π‘β1 + ππ
π
π=1 π₯2
π¦π‘βπ + π’π‘ (3)
Where Ξ2
yt denotes the second differences of the yt. The null hypothesis H0: ytΛ·I(2) against the
alternative H1: ytΛ·I(1). In case we reject the null hypothesis we conclude that the series are I(1) and we
must difference them one time to turn them into stationary series.
3.2.2 Phillips Perron Test
The second unit root test employed in this paper is the one developed by Phillips and Perron (1988)
which have become very popular in the analysis of financial time series. The Phillips-Perron (PP) unit
root tests differ from the ADF tests mainly in how they deal with serial correlation and
heteroskedasticity in the errors. In particular, where the ADF tests use a parametric autoregression to
approximate the ARMA structure of the errors in the test regression, the PP tests ignore any serial
correlation in the test regression. The test regression for the PP tests is the following,
π₯π¦π‘ = π0 + ππ‘ + πΌπ¦π‘β1 + π’π‘ (4)
Where ut is I(0) and may be heteroskedastic. The Phillips Perron test corrects for any serial correlation
and heteroskedasticity in the error terms ut of the test equation by modifying tΟ=0 and T π. These
modified statistics denoted Zt and ZΞ± are given by the following formulas:
ππ‘ =
π2
π2
1/2
. π‘π=0 β
1
2
π2βπ2
π2
.
π΅.ππΈ π
π2
ππ = ππΌ β
1
2
π΅2
. ππΈ(πΌ)
π2
(π2
β π2
)
The terms π2
and π2
are consistent estimates of the variance parameters, where:
π2
= lim
π΅ββ
π΅β1
πΈ π’π‘
2
π
π=1
π2
= lim
π΅ββ
πΈ πβ1
ππ
2
π
π=1
15. 15
Where ππ = π’π‘
π
π‘=1 , the sample variance of the least squares residuals π’π‘ is a consistent estimate of
π2
, and the Newey-West long-run variance estimate of π’π‘ using π’π‘ is a consistent estimate of π2
.
Under the null hypothesis that Ο = 0, the PP Zt and ZΟ statistics have the same asymptotic distributions
as the ADF t-statistic and normalized bias statistics. One advantage of the PP tests over the ADF tests is
that the PP tests are robust to general forms of heteroskedasticity in the error term ut. Another
advantage is that the user does not have to specify a lag length for the test regression.
3.3 Standard Granger causality test
Two or more variables can be interdependence on one another if the occurrence of one causes the
other to take place or vice versa, then one can talk of unidirectional causality and feedback causality.
Changes in the first variable precede changes in the other. That is one Granger causes another knowing
fully that the present with its lagged values can only predict the future but the future cannot predict the
past (Oguntade, Olanrewaju and Ojeniyi, 2014, p.16). Assume two strictly stationary time series, X and Y.
If variable X contains useful information for predicting variable Y, then X causes Y. That is X granger
causes or predicts the changes of Y.
Testing for Granger causality (GC) in a VAR framework can be written as:
π₯π¦π‘ = π1 + ππ
π
π=1
π₯π¦π‘βπ + π£π
π
π=1
π₯π₯π‘βπ + π’π¦π‘
π₯π₯π‘ = π2 + ππ
π
π=1
π₯π₯π‘βπ + πΏπ
π
π=1
π₯π¦π‘βπ + π’π₯π‘
Where π1, π2 are constant terms, π₯π¦π‘ , π₯π₯π‘ are stationary variables, q and m are the lag order
(q=m) and uyt , uxt are error terms and assumed to be serially uncorrelated. The optimal lag length q=m
is determined using the usual information criteria (Akaike, Schwarz and Hannan-Quinn). In this study
Schwarz information criterion (SIC) is used to determine the optimal lag length (the lag order which
minimizes the SIC). However, it is possible for the error terms to be evidenced with autocorrelation
when we estimate a VAR model, using the optimal lag length. For this purpose diagnostic tests such as
autocorrelation LM (Lagrange Multiplier) test is applied to find possible anomalies. The problem is
solved by adding extra lags until autocorrelation in VAR residuals disappears. The next step is to
estimate the restricted equations of the model in order to obtain the RSSRY and RSSRX (residual sum of
squares) which will be used in the F-test. From the unrestricted model (the equations above) we obtain
the RSSUY and RSSUX. The restricted model is as follows:
π₯π¦π‘ = π + ππ
π
π=1
π₯π¦π‘βπ + π’π¦π‘
π₯π₯π‘ = π + πππ₯π₯π‘βπ
π
π=1
+ π’π₯π‘
16. 16
The step following after the model estimations is to set the null and alternative hypotheses. The
hypotheses are as follows:
H0 : π£π
π
π=1 = 0 or Xt does not cause Yt H0 : πΏπ
π
π=1 = 0 or Ξ₯t does not cause Ξ§t
H1 : π£π
π
π=1 β 0 or Xt causes Yt H1 : πΏπ
π
π=1 β 0 or Ξ₯t causes Ξ§t
The final step of standard Granger causality procedure is the calculation of the F βstatistic for the
normal Wald tests. The general formula using for the computation of the F-test is the following:
πΉ β π‘ππ π‘ =
π πππ β π ππππ
π ππππ
Γ·
ππ β πππ
πππ
Where RSSR and RSSUR are respectively, sum of squared errors for the restricted and the unrestricted
versions of the equations, while d is the degree of freedom. If the computed F- value exceeds the critical
F- value, reject the null hypothesis and conclude that Xt Granger causes Yt, or Yt Granger causes Xt.
3.4 Toda and Yamamoto approach for Granger causality
According to Toda and Yamamoto (1995), economic series could be either integrated of the different
orders or non-cointegrated or both. In these cases, the ECM cannot be applied for Granger causality
tests. Hence, they developed an alternative test, irrespective of whether Yt and Xt are I(0), I(1) or I(2),
non-cointegrated or cointegrated of an arbitrary order. This is widely known as the Toda and Yamamoto
(1995) augmented Granger causality. This procedure provides the possibility of testing for causality
between integrated (non-stationary) variables based on asymptotic theory. In our analysis Toda &
Yamamoto test used in the levels of the series to indentify the underlying relationship between stock
and currency markets, while Standard GC and Diks and Pachenko (2006) test is used on the returns of
the variables.
Toda and Yamamoto (1995) proposed the modified WALD (MWALD) test to analyze the linear
restriction parameters. This methodology allows the asymptotic x2
distributions while estimating VAR
(k+dmax) where k is used for lag-length proposed by information criteria and dmax is maximum order of
integration in the system determined by unit root tests (for example if the series are I(1) and the k=2
then we specify a VAR model with 3 lags). If VAR residuals are found to be serially correlated with the
use of LM autocorrelation test, extra lags will be added until autocorrelation problem disappears.
The differences between Standard GC and Toda and Yamamoto (TY) are in the unrestricted and
restricted form of the VAR model, the type of the series used (I(1) in TY and I(0) in GC) and the in F-test
formula (MWALD test). The unrestricted VAR model specified in TY procedure is the following:
π¦π‘ = π1 + ππ
π+π
π=1
π¦π‘βπ + π£π
π+π
π =1
π₯π‘βπ + π’π¦π‘
π₯π‘ = π2 + ππ
π+π
π=1
π₯π‘βπ + πΏπ
π+π
π =1
π¦π‘βπ + π’π₯π‘
The restricted form of the VAR model is the following:
17. 17
π¦π‘ = π1 + ππ
π+π
π=1
π¦π‘βπ + π£π
π
π=1
π₯π‘βπ + π’π¦π‘
π₯π‘ = π2 + ππ
π
π=1
π₯π‘βπ + πΏπ
π+π
π =1
π¦π‘βπ + π’π₯π‘
The set of the null and alternative hypotheses and the calculation of the F-statistic for the
modified Wald Test are the following:
H0 : π£π
π
π=1 = 0 or Xt does not cause Yt H0 : ππ
π
π=1 = 0 or Ξ₯t does not cause Ξ§t
H1 : π£π
π
π=1 β 0 or Xt causes Yt H1 : ππ
π
π=1 β 0 or Ξ₯t causes Ξ§t
The general form of the F-test used in TY approach is the following:
πΉ β π‘ππ π‘ =
π πππ β π ππππ
π ππππ
Γ·
π β πΎ
π
Where K is the number of the estimated coefficients and N the number of the observations. If the
computed F-value exceeds the critical F-value, reject the null hypothesis and conclude that Xt causes Yt,
or Yt causes Xt.
3.5 Non-parametric Causality Test
The linear Granger causality tests (GC and TY) do not account for nonlinearities among the variables,
thus the next step in our analysis is to relax the assumption of the linearity and investigate for nonlinear
relationships between sectoral stock indices and exchange rates. In recent empirical literature various
non parametric methods have been developed. Baek and Brock (1992) proposed a non-linear method
for detecting non-linear Granger causality by using correlation integral between time series (Nazlioglu,
Kar and Akel, 2014, p.395). Hiemstra and Jones (1994) developed a non-linear statistical test which
allows each series to display short-term temporal dependence. Although, Diks and Pachenko (2006)
showed that the test proposed by Hiemstra and Jones (1994) may over reject the null hypothesis of non-
causality in the cases the sample size is increasing because it does not take account the possible
variations in conditional distributions (Nazlioglu, Kar and Akel, 2014).
In my thesis, Diks and Pachenko (2006) nonlinear Granger causality test is employed since it
provides a solution in the over-rejection problem in Hiemstraβs Jones test. In the remaining of this
section, following Diks and Pachenko (2006) and Bekiros and Diks (2008, p.2676), I analyze the details of
the Diks and Pachenko non-linear procedure.
Assume that we have two strictly stationary and weakly dependent time series Xt and Yt, let St
m
be
the m-length lead vector of Xt and ππ‘
βπ¦
the βy-length lag vector of ππ‘ and ππ‘
βπ₯
the βx-length lag vector of
ππ‘ (βy, βxβ₯1). To keep the notation compact and to take account the fact that the null hypothesis of no
causality is a statement about the invariant distribution of (βy +βx+m)-dimensional vector Qt= (ππ‘
βπ¦
, ππ‘
βπ₯
,
St
m
), as Bekiros and Diks (2008) and Bampinas and Panagiotidis (2015), we drop the time index and
assume that βy=βx=1. The conditional distribution of S given (Y,X)=(y,x) is the same as that of S given Y=y
18. 18
under the null hypothesis. The joint probability density function ππ,π,π(π₯, π¦, π ) and its marginals must
satisfy the following equation:
ππ,π,π(π₯, π¦, π )
ππ(π¦)
=
ππ,π(π₯, π¦)
ππ(π¦)
β
ππ,π(π¦, π )
ππ(π¦)
This relationship states that X and S are independent conditionally on Y=y for each fixed value of y.
Diks and Pachenko (2006) then reformulated the null hypothesis of no nonlinear Granger causality
which implies:
π β‘ πΈ ππ,π,π π, π, π ππ π β ππ,π π, π ππ,π π, π = 0
Where an estimator for q (proposed by Diks and Pachenko)is :
ππ ππ =
(2ππ)βππ βππβππ
π π β 1 (π β 2)
πΌππ
πππ
πΌππ
π
β πΌππ
ππ
πΌππ
ππ
π,πβ π
π,πβ π,π
π
Where πΌππ
π€
= πΌ ππ β π
π < ππ and ππ is the bandwidth depending on the sample size n. If we denote
local density estimators of a dw-variate random vector W at Wi by,
π
π€ ππ =
(2π)βππ€
π β 1
πΌππ
π€
π,πβ π
Given this estimator, the test statistic simplifies to,
ππ ππ =
π β 1
π(π β 2)
ππ,π,π ππ, ππ, ππ ππ ππ β ππ,π ππ, ππ ππ,π ππ, ππ
π
For one lag (βy=βx=1) and the bandwidth calculated by ππ = πΆπβπ
(C>0, 1/4<ΞΈ<1/3), Diks and
Pachenko (2006) proved that the test statistic in the equation above satisfies:
π
ππ ππ β π
π»π
π·
β π(0,1)
Where
π·
β denotes convergence in standard normal distribution and Hn is an estimator of the asymptotic
variance of ππ β . The last but most important step of DP test is the choice of the bandwidth value.
According to Diks and Pachenko (2006, p.1658) the computation of bandwidth is a very important
matter. ARCH coefficients denote that c=1 and a=0.4 and for theoretically Ξ²=2/7 the authors chose
Cβ8.62. Taking account these finding, it is assumed that bandwidth is a function of the sample size. In
this study, bandwidth was calculated approximately 0.9 while other studies (Bekiros & Diks, 2008,
Bampinas & Panagiotidis, 2014) set bandwidth to unity.
19. 19
4. Empirical Results
The empirical methodology comprises of four steps. The first step is the ADF and PP unit root tests
estimation. The other three steps are the three variations of Granger causality. Standard Granger
causality and Toda Yamamoto are linear approaches while Diks and Pachenko test are nonlinear
methodology. The following sections present the studyβs empirical results.
4.1 Unit root tests
This section refers to the results of the stationarity tests. Two tests were employed, namely the
augmented Dickey Fuller (1981) and Phillips Perron (1988) test. The ADF tests for the logarithmic levels
and log-daily returns of stock indices and exchange rates are presented in Tables 3,4,5,6,7,8,9 and 10.
Tables 3 and 4 report the results of unit root tests on exchange rate variables, while the rest of the
tables refer to the estimations of unit root tests on stock market indices.
All the variables were found non-stationary with order of integration I(1)-one unit root-in the log
levels of the series. To transform the time series to stationary processes (I(0)), we differenced all the
variables once. In the tables where ADF results reported, we contain information about the lag lengths
6
and t-statistics (two models, one with trend and one with no trend employed) while to those where we
put the results of PP tests, we include information related with Newey-West bandwidth and the test
statistics (two models, one with trend and one with no trend employed).
Table 3. ADF Tests on Exchange Rates Series
ADF (no trend) ADF (with trend)
Variables Difference Lag Statistic Lag Statistic
CAD Level 3 -2.345186 3 -2.056768
AUD Level 3 -1.802087 3 -1.621580
BRL Level 4 -1.599957 4 -1.444684
CHF Level 4 -1.396851 5 -2.396561
EUR Level 3 -2.274581 3 -2.215199
GBP Level 4 -1.455124 3 -1.990259
JPY Level 1 -0.714447 1 0.074746
SEK Level 3 -2.144428 3 -2.087430
TRYUSD Level 4 -0.720056 4 -2.582430
MXNUSD Level 3 -1.302205 3 -2.670198
MYRUSD Level 4 -1.763126 4 -0.777922
THBUSD Level 2 -1.865745 2 -1.460668
TWEXM Level 0 -1.622665 0 -1.129556
TWEXB Level 0 -1.523692 0 -0.810105
CADUSD D1 2 -24.07780*** 2 -24.10542***
AUDUSD D1 2 -23.85580*** 2 -23.87166***
BRLUSD D1 3 -23.91293*** 3 -24.01941***
CHFUSD D1 3 -24.04328*** 3 -24.04245***
EURUSD D1 2 -23.96848*** 3 -22.87428***
GBPUSD D1 3 -23.51232*** 3 -23.50765***
JPYUSD D1 0 -41.69554*** 0 -41.73223***
6
The lag length determination is performed using the Schwarz information Criterion (SIC).
20. 20
SEKUSD D1 2 -25.80751*** 2 -25.80968***
TRYUSD D1 3 -23.14142*** 3 -23.14985***
MXNUSD D1 2 -24.24444*** 2 -24.24857***
MYRUSD D1 3 -24.07381*** 3 -24.13834***
THBUSD D1 1 -29.47505*** 1 -29.50119***
TWEXM D1 0 -50.60991*** 0 -50.63300***
TWEXB D1 0 -48.53526*** 0 -48.57343***
Notes: The table reports the results from ADF tests on exchange rates. SIC is used for the optimal lag selection. The
test performed in both log-levels and first differences of the series. The *, **, *** denotes the rejection of the null
hypothesis in 10%, 5% and 1% significance levels respectively.
Table 4. PP tests on exchange rate series
PP (no trend) PP (with trend)
Variables Difference Bandwidth Statistic Bandwidth Statistic
CADUSD Level 4 -2.203764 3 -1.736230
AUDUSD Level 10 -1.685612 10 -1.385554
BRLUSD Level 7 -1.609939 5 -1.442026
CHFUSD Level 18 -1.497487 18 -2.454210
EURUSD Level 15 -2.168513 15 -2.108361
GBPUSD Level 9 -1.424161 9 -1.874206
JPYUSD Level 12 -0.733425 11 0.070025
SEKUSD Level 12 -2.055503 12 -1.990841
TRYUSD Level 0 -0.004083 1 -2.336062
MXNUSD Level 13 -1.186054 14 -2.612641
MYRUSD Level 10 -1.764262 10 -0.782266
THBUSD Level 11 -1.853562 11 -1.397799
TWEXM Level 12 -1.635155 12 -1.135002
TWEXB Level 14 -1.690349 13 -1.042642
CADUSD D1 18 -26.73810*** 19 -26.64778***
AUDUSD D1 16 -26.94062*** 16 -26.92837***
BRLUSD D1 17 -28.97692*** 18 -28.92438***
CHFUSD D1 6 -27.82589*** 6 -27.82207***
EURUSD D1 6 -28.16929*** 6 -28.16992***
GBPUSD D1 12 -26.64072*** 12 -26.63405***
JPYUSD D1 6 -41.74320*** 6 -41.76858***
SEKUSD D1 4 -30.38022*** 5 -30.12006***
TRYUSD D1 28 -26.59608*** 29 -26.60372***
MXNUSD D1 4 -32.00142*** 4 -32.00191***
MYRUSD D1 5 -36.99980*** 6 -36.98002***
THBUSD D1 7 -40.81992*** 6 -40.78930***
TWEXM D1 11 -50.60788*** 11 -50.63187***
TWEXB D1 12 -48.62429*** 12 -48.64866***
Notes: The table reports the results from PP tests on exchange rates. The bandwidth of PP tests is selected by
Newey and West (1994). The test performed in both log-levels and first differences of the series. The *, **, ***
denotes the rejection of the null hypothesis in 10%, 5% and 1% significance levels respectively.
21. 21
Table 5. ADF tests on Sectoral Stock Market Indices (S&P 500) Series
ADF (no trend) ADF (with trend)
Variables Difference Lag Statistic Lag Statistic
SP500 Level 1 -0.084577 1 -0.959829
COND Level 0 0.684130 0 -1.074878
CONS Level 1 0.963757 1 -1.121445
ENER Level 1 -2.409388 1 -2.708212
FINL Level 1 -1.271387 1 -0.780484
HLTH Level 1 2.031293 1 0.249634
INDU Level 0 -0.512881 0 -1.259527
INFT Level 0 0.121569 0 -1.559844
MATR Level 0 -1.652577 0 -2.289533
TELS Level 2 -1.649450 2 -1.753009
UTIL Level 2 -1.112497 2 -1.442125
SP500 D1 0 -55.62734*** 0 -55.65017***
COND D1 0 -51.54003*** 0 -51.60923***
CONS D1 0 -55.44592*** 0 -55.49240***
ENER D1 0 -54.95870*** 0 -54.95746***
FINL D1 0 -56.33875*** 0 -56.35128***
HLTH D1 0 -53.00227*** 0 -53.12617***
INDU D1 0 -52.81534*** 0 -52.83144***
INFT D1 0 -52.85388*** 0 -52.88153***
MATR D1 0 -52.40042*** 0 -52.39209***
TELS D1 1 -39.03028*** 1 -39.02269***
UTIL D1 1 -39.23121*** 1 -39.22862***
Notes: This table reports the results of the ADF tests on S&P 500 sectors (levels and first differences). SIC is used
for the optimal lag selection. The *, **, *** denotes the rejection of the null hypothesis in 10%, 5% and 1%
significance levels respectively.
Table 6. ADF tests on Sectoral Stock Market Indices (NASDAQ) Series
ADF (no trend) ADF (with trend)
Variables Difference Lag Statistic Lag Statistic
IXBK Level 1 -1.555886 1 -1.122630
IXCO Level 0 0.463586 0 -1.469993
IXFN Level 2 -1.527604 2 -1.632596
IXHC Level 0 2.099083 0 0.068102
IXIC Level 0 0.367808 0 -1.221443
IXID Level 0 -0.007202 0 -1.350253
IXIS Level 2 0.099027 2 -1.255182
IXTC Level 0 -2.429528 0 -2.602142
IXTR Level 0 -1.127399 0 -1.491812
NBI Level 3 3.018927 3 0.656922
IXBK D1 0 -58.04175*** 0 -58.05542***
IXCO D1 0 -52.25724*** 0 -52.29335***
IXFN D1 1 -40.39101*** 1 -40.38645***
22. 22
IXHC D1 2 -31.07880*** 2 -31.26023***
IXIC D1 0 -52.86336*** 0 -52.90086***
IXID D1 0 -51.32548*** 0 -51.34717***
IXIS D1 1 -41.42315*** 1 -41.44236***
IXTC D1 0 -52.12204*** 0 -52.11326***
IXTR D1 0 -52.13705*** 0 -52.14812***
NBI D1 2 -32.27812*** 2 -32.50393***
Notes: This table reports the results of the ADF tests on NASDAQ sectors (levels and first differences). SIC is used
for the optimal lag selection. The *, **, *** denotes the rejection of the null hypothesis in 10%, 5% and 1%
significance levels respectively.
Table 7. ADF tests on Sectoral Stock Market Indices (NASDAQ-PHLX) Series
ADF (no trend) ADF (with trend)
Variables Difference Lag Statistic Lag Statistic
DFX Level 0 0.366748 0 -0.961791
RXS Level 0 0.483659 0 -0.767670
XAU Level 0 -2.011181 0 -2.026971
HGX Level 0 -2.206654 0 -1.658725
OSX Level 0 -2.693333* 0 -2.530762
SOX Level 0 -0.830168 0 -1.145601
UTY Level 2 -1.438168 2 -1.652535
DFX D1 0 -52.02161*** 0 -52.04513***
RXS D1 0 -52.91335*** 0 -52.96488***
XAU D1 0 -50.72170*** 0 -50.74046***
HGX D1 0 -48.66223*** 0 -48.69965***
OSX D1 0 -52.41428*** 0 -52.42368***
SOX D1 0 -51.20466*** 0 -51.22273***
UTY D1 1 -39.42401*** 1 -39.41885***
Notes: This table reports the results of the ADF tests on PHLX sectors (levels and first differences). SIC is used for
the optimal lag selection. The *, **, *** denotes the rejection of the null hypothesis in 10%, 5% and 1%
significance levels respectively.
Table 8. PP (Philips Perron) tests on Sectoral Stock Market Indices (S&P 500) Series
PP (no trend) PP (with trend)
Variables Difference Bandwidth Statistic Bandwidth Statistic
SP500 Level 17 0.055933 17 -0.843556
COND Level 9 0.892435 9 -0.948569
CONS Level 12 1.170355 10 -0.946378
ENER Level 10 -2.383630 7 -2.676246
FINL Level 27 -1.183301 28 -0.573737
HLTH Level 14 2.337338 14 0.478198
INDU Level 9 -0.357412 9 -1.126687
INFT Level 21 0.330515 20 -1.387528
MATR Level 10 -1.459825 9 -2.089841
TELS Level 7 -1.646016 7 -1.762779
UTIL Level 10 -1.121573 9 -1.473271
23. 23
SP500 D1 16 -56.17811*** 16 -56.26381***
COND D1 8 -51.67559*** 10 -51.79923***
CONS D1 9 -56.00689*** 11 -56.20173***
ENER D1 10 -55.24524*** 10 -55.24951***
FINL D1 25 -57.69056*** 26 -57.82276***
HLTH D1 10 -53.22471*** 13 -53.50343***
INDU D1 8 -52.89393*** 9 -52.93121***
INFT D1 19 -52.98256*** 20 -53.04258***
MATR D1 11 -52.66595*** 11 -52.65869***
TELS D1 9 -53.07327*** 9 -53.06226***
UTIL D1 11 -55.43640*** 11 -55.43544***
Notes: The table reports the results from PP tests on S&P 500 sectors. The bandwidth of PP tests is selected by
Newey and West (1994). The test performed in both log-levels and first differences of the series. The *, **, ***
denotes the rejection of the null hypothesis in 10%, 5% and 1% significance levels respectively.
Table 9. PP (Philips Perron) tests on Sectoral Stock Market Indices (NASDAQ) Series
PP (no trend) PP (with trend)
Variables Difference Bandwidth Statistic Bandwidth Statistic
IXBK Level 20 -1.437261 20 -0.900758
IXCO Level 18 0.665011 17 -1.310865
IXFN Level 19 -1.441012 19 -1.551061
IXHC Level 3 2.309261 2 0.142176
IXIC Level 13 0.635350 13 -1.009292
IXID Level 6 0.102954 6 -1.264546
IXIS Level 5 -0.021702 7 -1.341499
IXTC Level 0 -2.429528 0 -2.602142
IXTR Level 10 -0.857773 10 -1.236069
NBI Level 4 2.800346 4 0.513145
IXBK D1 19 -59.79632*** 20 -59.94002***
IXCO D1 17 -52.37198*** 18 -52.44696***
IXFN D1 20 -58.40717*** 20 -58.41225***
IXHC D1 4 -48.98139*** 2 -49.01183***
IXIC D1 12 -53.08383*** 13 -53.16482***
IXID D1 6 -51.39745*** 7 -51.45302***
IXIS D1 9 -60.26641*** 8 -60.35295***
IXTC D1 3 -52.17176*** 3 -52.16291***
IXTR D1 11 -52.35916*** 11 -52.38410***
NBI D1 8 -49.32483*** 5 -49.51257***
Notes: The table reports the results from PP tests on NASDAQ sectors. The bandwidth of PP tests is selected by
Newey and West (1994). The test performed in both log-levels and first differences of the series. The *, **, ***
denotes the rejection of the null hypothesis in 10%, 5% and 1% significance levels respectively.
24. 24
Table 10. PP (Philips Perron) tests on Sectoral Stock Market Indices (NASDAQ-PHLX) Series
PP (no trend) PP (with trend)
Variables Difference Bandwidth Statistic Bandwidth Statistic
DFX Level 10 0.515610 10 -0.834962
RXS Level 7 0.680208 7 -0.615091
XAU Level 10 -1.915920 11 -1.911342
HGX Level 4 -2.204766 4 -1.651855
OSX Level 6 -2.611478* 5 -2.412588
SOX Level 15 -0.643480 16 -0.971400
UTY Level 7 -1.494850 7 -1.712560
DFX D1 9 -52.10355*** 10 -52.13612***
RXS D1 4 -52.94808*** 6 -53.03316***
XAU D1 12 -50.83791*** 13 -50.90295***
HGX D1 3 -48.64409*** 2 -48.68955***
OSX D1 7 -52.46914*** 8 -52.49091***
SOX D1 16 -51.35698*** 17 -51.39564***
UTY D1 9 -55.72540*** 9 -55.71869***
Notes: The table reports the results from PP tests on PHLX sectors. The bandwidth of PP tests is selected by Newey
and West (1994). The test performed in both log-levels and first differences of the series. The *, **, *** denotes
the rejection of the null hypothesis in 10%, 5% and 1% significance levels respectively.
4.2 Standard Granger causality Test
The first test we employ in our analysis is the standard Granger causality test proposed by Granger
(1969). Bivariate VAR models were specified in order to perform pairwise linear Granger causality tests
(378 pairs in total). Firstly, we selected the lag length of the VAR processes using Schwarz Information
Criterion, however in most cases more lags were added in the models in order to avoid the
autocorrelation problem in VAR residuals. Secondly, we specified VAR models using the log-differences
of the variables and receiving from them the residual sum of squares (RSS). The last step in this
procedure is the calculation of the Wald tests and the rejection or non-rejection of the null hypothesis of
non causality.
The Table 11 presents the optimal lag lengths which we used in the VAR specifications of the
pairwise Granger causality tests.
Table 11. *Optimal lag length that suggested by information criteria (Schwartz Bayesian Information Criterion) in
Standard Granger Causality Test.
Variables EUR GBP JPY CAD CHF SEK AUD BRL MXN MYR THB TRY TWEXM TWEXB
COND 5 4 2 3 4 3 3 4 3 3 2 3 1 2
CONS 5 6 2 4 4 3 4 4 2 4 2 4 3 3
HLTH 5 6 2 4 4 3 4 4 3 4 2 3 3 3
INDU 5 5 2 4 4 3 4 4 2 3 2 3 1 1
INFT 5 5 2 4 4 3 4 4 2 3 2 3 1 1
MATR 5 6 2 4 4 3 3 3 2 4 4 3 1 1
25. 25
TELS 5 5 2 4 5 3 5 4 3 3 2 3 1 1
UTIL 5 4 2 4 4 3 3 4 2 3 2 3 2 3
FINL 6 6 2 5 5 5 5 5 7 5 2 5 3 1
ENER 5 5 2 4 4 3 4 4 3 5 2 3 2 2
SP500 5 5 2 4 4 3 4 4 2 4 2 3 1 3
IXBK 6 6 2 5 5 5 3 5 3 5 2 5 3 4
IXID 5 5 2 4 4 2 3 4 2 3 2 3 2 2
IXIS 5 6 2 3 4 2 3 4 2 3 2 4 3 3
IXTC 5 5 2 3 4 3 4 4 2 3 2 3 2 2
NBI 5 4 2 3 4 3 3 4 3 3 2 3 3 3
IXCO 5 5 2 3 4 2 4 4 3 3 2 3 1 1
IXFN 6 5 2 5 5 5 4 5 7 5 5 5 2 3
IXTR 5 5 2 4 4 3 3 3 3 4 2 3 1 1
IXIC 5 5 2 4 4 3 4 4 2 4 2 3 2 2
DFX 5 5 2 4 5 2 3 4 2 3 2 3 1 1
RXS 5 5 2 4 4 2 3 4 3 3 2 3 1 1
XAU 5 5 1 4 6 3 5 4 4 4 3 3 3 1
HGX 5 4 3 3 4 3 3 3 3 3 2 3 1 1
OSX 5 4 2 4 4 4 4 4 7 4 2 3 4 7
SOX 5 5 2 4 4 3 4 4 3 4 2 3 1 1
UTY 5 4 2 4 4 2 3 4 2 3 2 3 2 3
*In most cases, the lag length is higher than the optimal, in order to avoid the autocorrelation problem in VAR
residuals (Autocorrelation LM Test employed for this purpose).
As it can be easily noticed the values of the lag length are different from VAR to VAR model. In most
cases we have to add more lags to avoid misspecifications in our models.
The tables 12 to 19 present the results of the pairwise Granger causality tests. The tables contain
information related with the p-values and the rejection or non-rejection of the null hypothesis of the
causality tests. In order to overcome the difficulty of presenting larger tables than we already present,
we used the following simplifying notations for the level of significance: β*β, β**β, β***β. The Tables 12
and 13 show the results of the Granger causality tests between traded weighted indices of the U.S.
dollar and the sectoral stock market indices, while the rest of them report the empirical results from the
causality tests between nominal exchange rates and sectoral indices.
Table 12. Linear Causality (Standard Granger Causality) between Traded weighted U.S. Dollar Indices (Broad and
Major currencies) and sectoral stock indices.
Variables TWEXB SP SP TWEXB
Stock Prices (SP) p-values p-values
COND 0.5088 0.0000***
CONS 0.0172** 0.0000***
HLTH 0.0153** 0.0000***
INDU 0.4329 0.0000***
INFT 0.3043 0.0000***
MATR 0.5116 0.0000***
TELS 0.0160** 0.0000***
26. 26
UTIL 0.1371 0.0000***
FINL 0.8776 0.0000***
ENER 0.0012*** 0.0000***
SP500 0.0041*** 0.0000***
IXBK 0.3368 0.0000***
IXID 0.2346 0.0000***
IXIS 0.0015*** 0.0000***
IXTC 0.0555* 0.0000***
NBI 0.0096*** 0.0000***
IXCO 0.2889 0.0000***
IXFN 0.0018*** 0.0000***
IXTR 0.7470 0.0000***
IXIC 0.0034*** 0.0000***
DFX 0.4589 0.0000***
RXS 0.5872 0.0000***
XAU 0.6678 0.0000***
HGX 0.8500 0.0000***
OSX 0.0001*** 0.0000***
SOX 0.2277 0.0000***
UTY 0.1181 0.0000***
Notes: denotes the no-causality null hypothesis on first differences. The Schwarz Information Criterion was
used to determine the optimal lag lengths for VAR(p) models. However, in most cases we add more lags to avoid
autocorrelation in residuals, according to LM test. The p-values with *, **, *** represent the significance at 10%,
5% and 1% level respectively.
Table 13. Linear Causality (Standard Granger Causality) between Traded weighted U.S. Dollar Indices (Broad and
Major currencies) and sectoral stock indices.
Variables TWEXM SP SP TWEXM
Stock Prices (SP) p-values p-values
COND 0.5657 0.0000***
CONS 0.2315 0.0000***
HLTH 0.2264 0.0000***
INDU 0.7922 0.0000***
INFT 0.7745 0.0000***
MATR 0.8296 0.0000***
TELS 0.1414 0.0000***
UTIL 0.2001 0.0000***
FINL 0.0003*** 0.0000***
ENER 0.0608* 0.0000***
SP500 0.0263** 0.0000***
IXBK 0.1070 0.0000***
IXID 0.4402 0.0000***
27. 27
IXIS 0.0481** 0.0000***
IXTC 0.2877 0.0000***
NBI 0.1033 0.0000***
IXCO 0.7088 0.0000***
IXFN 0.0074*** 0.0000***
IXTR 0.5249 0.0000***
IXIC 0.0403** 0.0000***
DFX 0.7140 0.0000***
RXS 0.1542 0.0000***
XAU 0.4531 0.0000***
HGX 0.8797 0.0000***
OSX 0.0325** 0.0000***
SOX 0.4139 0.0000***
UTY 0.2351 0.0000***
Notes: denotes the no-causality null hypothesis on first differences. The Schwarz Information Criterion was
used to determine the optimal lag lengths for VAR(p) models. However, in most cases we add more lags to avoid
autocorrelation in residuals, according to LM test. The p-values with *, **, *** represent the significance at 10%,
5% and 1% level respectively.
The results presented in Tables 12 and 13 allow for the following remarks: In the pairwise
implementation of the linear Granger causality tests, strong (1% and 5% level of significance)
bidirectional Granger causality between traded weighted index of U.S. dollar, Broad (TWEXB) and
sectoral indices (CONS, HLTH, TELS, ENER, IXIS, NBI, IXFN and OSX) was evidenced. Strong feedback
relationships have also evidenced between TWEXB and S&P 500 and NASDAQ Composite market
indices. On the other hand, weak (10% level of significance) bidirectional causal relationship was found
between TWEXB and the NASDAQ Telecommunications (IXTC). As for the rest of the sectoral indices, it
was found that they cause changes to the traded weighted index of U.S. dollar: broad (unidirectional
causality from stock variables to exchange rate variable).
As for the traded weighted index of U.S. dollar, major currencies (TWEXM), it is evidenced strong
bidirectional relationships between TWEXM and financial sectors (S&P 500 Financials, NASDAQ
Insurance, NASDAQ other Finance) and PHLX Oil Services sector (OSX). Strong feedback relationships
were also found between the two market indices (S&P 500 and NASDAQ Composite) while weak bi-
causal relationship was evidenced for the energy sector (S&P 500 energy sector). Unidirectional causal
linkages were found between TWEXM and the rest of the sectoral stock market indices with the
direction from stock indices to exchange rates.
The Tables 14 to 19 present the results of pairwise standard Granger causality tests between
nominal exchange rates and sectoral stock market indices.
33. 33
Table 19. Linear Causality (Standard Granger Causality, differences, short-run), NASDAQ-PHLX sectors
DFX >EUR >GBP >JPY >CAD >CHF >SEK >AUD >BRL >MXN >MYR >THB >TRY
p-values 0.000*** 0.000*** 0.000*** 0.000*** 0.0466** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000***
RXS >EUR >GBP >JPY >CAD >CHF >SEK >AUD >BRL >MXN >MYR >THB >TRY
p-values 0.000*** 0.000*** 0.000*** 0.000*** 0.0231** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000***
XAU >EUR >GBP >JPY >CAD >CHF >SEK >AUD >BRL >MXN >MYR >THB >TRY
p-values 0.000*** 0.000*** 0.1110 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000***
HGX >EUR >GBP >JPY >CAD >CHF >SEK >AUD >BRL >MXN >MYR >THB >TRY
p-values 0.000*** 0.000*** 0.000*** 0.000*** 0.7957 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000***
OSX >EUR >GBP >JPY >CAD >CHF >SEK >AUD >BRL >MXN >MYR >THB >TRY
p-values 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000***
SOX >EUR >GBP >JPY >CAD >CHF >SEK >AUD >BRL >MXN >MYR >THB >TRY
p-values 0.000*** 0.000*** 0.000*** 0.000*** 0.6116 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000***
UTY >EUR >GBP >JPY >CAD >CHF >SEK >AUD >BRL >MXN >MYR >THB >TRY
p-values 0.000*** 0.000*** 0.000*** 0.000*** 0.1063 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000***
Notes: denotes the no-causality null hypothesis on first differences. The Schwarz Information Criterion was used to determine the optimal lag lengths for
VAR(p) models. However, in most cases we add more lags to avoid autocorrelation in residuals, according to LM test. The p-values with *, **, *** represent the
significance at 10%, 5% and 1% level respectively
34. 34
In Summary, the results presented in the Tables 14 to 19 do not show a recognizable causal pattern.
Overall the results, except from JPY and CHF, show unidirectional relationship running from stock prices
to exchange rates and bi-causal linkages for all sectoral stock indices. Unidirectional causality was found
running from Swiss Franc (CHF) to consumer discretionary (CONS), health and care (HLTH) and
biotechnology sector (NBI) while the reverse direction was evidenced in the cases of the S&P 500
sectors (Materials, Telecommunications, Industrials, Utilities and Energy) and PHLX sectors (Defense,
Gold and Silver and Oil services sector). Bidirectional causal linkages were found from CHF to FINL
(financials) and RXS (Pharmaceuticals). For the rest of the sectors the GC test failed to provide any
evidence of causation between the variables.
For the Japanese Yen, weak causal relationship was found from exchange rates to stock prices in
the case of the PHLX Gold and Silver sector (XAU) and bidirectional causality evidenced in the Housing
sector (HGX).
4.3 Toda and Yamamoto approach for Granger causality
The second approach for Granger causality testing used in this paper is the linear causality test proposed
by Toda and Yamamoto (1995). The main advantage of this test is that it does not require stationary
variables to produce non spurious results. Toda and Yamamoto (1995) suggested the increasing of the
lag length of the VAR models, used in GC tests, from k-lags to k+dmax where dmax is the maximum order
of integration among the series used in the model (e.g. if we have two variables, one is I(2) and the
other is I(1), the dmax=2). The TY test is used to examine the long-run causal relationships, avoiding the
problem of the spurious regressions. Schwarz information criterion was used for the lag length selection
and in cases we detect autocorrelation in VAR residuals, extra lags were added to avoid distortion in the
VARβs results.
This section contains information about the empirical results provided by Toda and Yamamoto tests
implemented in sectoral stock indices and exchange rate variables. The difference between TY and
standard GC test is that in Toda and Yamamoto test detects possible causal relationships between
integrated variables (in our case in log-levels of the variables, not between the returns). Table 20
contains information about the optimal lag length used TY tests (k+dmax).
Table 20. *Optimal lag length that suggested by information criteria (Schwartz Bayesian Information Criterion) in
Granger Causality Test (Toda & Yamamoto).
Variables EUR GBP JPY CAD CHF SEK AUD BRL MXN MYR THB TRY TWEXM TWEXB
COND 6 5 4 5 6 5 5 6 5 6 4 5 4 4
CONS 6 7 4 6 6 6 6 6 9 6 4 5 7 5
HLTH 6 6 4 6 6 5 5 6 6 6 4 6 5 5
INDU 6 5 4 5 6 5 5 6 5 6 4 5 5 3
INFT 5 5 4 5 6 5 5 6 5 6 4 5 3 4
MATR 6 6 4 6 6 5 6 6 9 6 4 5 3 3
TELS 6 6 4 5 6 5 5 6 5 6 4 5 5 5
UTIL 6 5 4 5 6 5 5 6 6 6 4 6 4 5
FINL 7 7 4 12 7 7 7 7 9 7 4 7 7 7
ENER 6 5 4 6 6 5 6 6 11 6 4 5 4 4
SP500 6 5 4 6 6 5 5 6 9 6 4 6 4 5
IXBK 7 5 4 5 7 5 5 7 9 7 4 6 5 5
IXID 6 5 4 6 6 5 5 6 5 6 4 5 3 4
IXIS 6 5 4 5 6 5 5 6 5 6 4 6 4 5
35. 35
IXCO 6 5 4 5 6 5 5 6 5 6 4 5 3 4
IXTC 5 5 4 5 6 5 5 6 5 6 4 5 4 4
NBI 6 5 4 5 6 5 5 6 5 6 4 5 4 4
IXFN 7 7 4 7 7 7 7 7 9 7 4 6 3 7
IXTR 6 5 4 5 6 5 5 6 5 6 4 5 4 3
IXIC 6 5 4 5 6 4 5 6 5 6 4 5 4 4
DFX 6 5 4 5 6 5 5 6 9 6 4 6 3 3
RXS 6 7 4 6 6 4 6 6 6 6 4 6 5 5
XAU 6 5 3 6 6 5 6 6 5 6 4 5 4 3
HGX 5 5 5 5 6 5 5 6 5 6 4 5 3 3
OSX 6 6 4 6 6 6 6 6 9 6 6 6 4 4
SOX 6 5 4 6 6 5 6 6 5 6 3 6 3 3
UTY 6 5 4 5 6 5 5 6 6 6 4 6 4 5
*Note: The table shows the optimal lag length (lag*+dmax), used in Toda and Yamamoto (TY) version, where
βdmaxβ denotes the maximum order of integration and βlag*β the lag length, suggested by SBC. All the series found
I(1), using ADF and Phillips-Perron unit root tests, which means that dmax =1. In most cases, the lag length is higher
than the optimal, in order to avoid the autocorrelation problem in VAR residuals (Autocorrelation LM Test
employed for this purpose).
The tables 21 to 28 present the results of the augmented Granger causality tests (TY). The tables contain
information related with the p-values and the rejection or non-rejection of the null hypothesis of the
causality tests. The Tables 21 and 22 show the results of the Toda and Yamamoto Granger causality
tests between traded weighted indices of the U.S. dollar and the sectoral stock market indices, while the
rest of them report the empirical results from the causality tests between nominal exchange rates and
sectoral indices.
Table 21. Linear Causality (Toda & Yamamoto) between Traded weighted U.S. Dollar Indices (Broad and Major
currencies) and sectoral stock indices.
Variables TWEXM SP SP TWEXM
Stock Prices (SP) p-values p-values
COND 0.1674 0.0000***
CONS 0.0506* 0.0000***
HLTH 0.0537* 0.0000***
INDU 0.0320** 0.0000***
INFT 0.0651* 0.0000***
MATR 0.4918 0.0000***
TELS 0.1621 0.0000***
UTIL 0.3968 0.0000***
FINL 0.0002*** 0.0000***
ENER 0.1340 0.0000***
SP500 0.0926* 0.0000***
36. 36
IXBK 0.0194** 0.0000***
IXID 0.0617* 0.0000***
IXIS 0.0214** 0.0000***
IXTC 0.2898 0.0000***
NBI 0.0897* 0.0000***
IXCO 0.0650* 0.0000***
IXFN 0.0074*** 0.0000***
IXTR 0.6560 0.0000***
IXIC 0.1155 0.0000***
DFX 0.1160 0.0000***
RXS 0.0361** 0.0000***
XAU 0.4406 0.0000***
HGX 0.2476 0.0000***
OSX 0.1350 0.0000***
SOX 0.2112 0.0000***
UTY 0.3855 0.0000***
Notes: denotes the no-causality null hypothesis on first differences. The
Schwarz Information Criterion was used to determine the optimal lag lengths for
VAR(p) models. However, in most cases we add more lags to avoid autocorrelation in
residuals, according to LM test. The p-values with *, **, *** represent the
significance at 10%, 5% and 1% level respectively.
Table 22. Linear Causality (Toda & Yamamoto) between Traded weighted U.S. Dollar Indices (Broad and Major
currencies) and sectoral stock indices.
Variables TWEXB SP SP TWEXB
Stock Prices (SP) p-values p-values
COND 0.0629* 0.0000***
CONS 0.0077*** 0.0000***
HLTH 0.0069*** 0.0000***
INDU 0.0753* 0.0000***
INFT 0.0398** 0.0000***
MATR 0.2089 0.0000***
TELS 0.0326** 0.0000***
UTIL 0.2679 0.0000***
FINL 0.0011*** 0.0000***
ENER 0.0092*** 0.0000***
SP500 0.0051*** 0.0000***
IXBK 0.0985* 0.0000***
IXID 0.0257** 0.0000***
IXIS 0.0030*** 0.0000***
IXTC 0.1153 0.0000***
NBI 0.0072*** 0.0000***
IXCO 0.0365** 0.0000***
37. 37
IXFN 0.0064*** 0.0000***
IXTR 0.4055 0.0000***
IXIC 0.0114** 0.0000***
DFX 0.0556* 0.0000***
RXS 0.0045*** 0.0000***
XAU 0.7629 0.0000***
HGX 0.2602 0.0000***
OSX 0.0132** 0.0000***
SOX 0.0823* 0.0000***
UTY 0.2482 0.0000***
Notes: denotes the no-causality null hypothesis on first differences. The
Schwarz Information Criterion was used to determine the optimal lag lengths for
VAR(p) models. However, in most cases we add more lags to avoid autocorrelation in
residuals, according to LM test. The p-values with *, **, *** represent the
significance at 10%, 5% and 1% level respectively.
The results presented in Tables 21 and 22 allow for the following remarks: In the pairwise
implementation of the Toda and Yamamoto (1995) Granger causality tests, strong (1% and 5% level of
significance) bidirectional Granger causality between traded weighted index of U.S. dollar, Major
currencies (TWEXM) and sectoral indices (FINL, INDU, IXBK, IXIS, IXFN and RXS) was evidenced. Weak
(10%) feedback relationships were evidenced between TWEXM and S&P 500, consumer staples, health
and care, information technology, industrials, biotechnology and computer sector. On the other hand,
strong unidirectional causal relationship was found between TWEXM and the rest of the stock indices,
running from stock prices to exchange rates.
As for the traded weighted index of U.S. dollar, broad (TWEXB), it is evidenced strong unidirectional
causality between TWEXB and MATR (materials), UTIL (utilities), IXTC (telecommunications), IXTR
(transportations), XAU(gold and silver), HGX (housing sector) and UTY (PHLX utility sector), running from
stock prices to exchange rates. As for the rest of the sectors, feedback relationships were found
between TWEXM and the remaining sectors.
The Tables 23 to 28 present the results of Toda and Yamamoto Granger causality tests between
nominal exchange rates and sectoral stock market indices. These tables provided us with mixed
empirical results. Generally, the results with exception JPY and CHF indicate unidirectional relationship
running from stock prices to exchange rates and bi-causal linkages for all sectoral stock indices. On the
other side, unidirectional causality was found running from Swiss Francs (CHF) to consumer
discretionary (CONS), health and care (HLTH) and biotechnology sector (NBI). For the rest of the sectoral
indices the TY test failed to provide any evidence of causation between the variables. For the Japanese
Yen (JPY), bidirectional causal relationships were found for all the sectoral indices except from PHLX gold
and silver sector which found to be uncorrelated with JPY.
We conclude with the presentation of the empirical results reported by Tables 23 to 28.