Université catholique de Louvain
Faculté des Sciences Economiques, Sociales et Politiques
Louvain School of Management
Volatility Dynamics Around Information:
Empirical Evidence from the Euro/Dollar
Currency Market
Thèse présentée pour l’obtention du grade de
Docteur en Sciences de Gestion
Walid BEN OMRANE
October, 2006
Dissertation committee:
Supervisor:
Prof. Luc BAUWENS
Jury:
Prof. Michel BEINE
Prof. Eric DE BODT
Prof. Pierre GIOT
Prof. Eric GIRARDIN
c Walid Ben Omrane, Thesis UCL, 2006.
°
To my Mother Badria, my father Abbés, my wonderful wife
Hédia, and my children Hamza and Sarah, with love
Contents
1 General Introduction
I
1
Volatility Dynamics Around Public Information
2 News Announcements, Market Activity and Volatility
13
15
2.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
2.2
News Announcements and Volatility
. . . . . . . . . . . . . . . .
17
2.3
Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
2.3.1
Euro/dollar Exchange Rate Data . . . . . . . . . . . . . .
21
2.3.2
News Announcement Data
. . . . . . . . . . . . . . . . .
24
2.3.3
The Intradaily Average Volatility . . . . . . . . . . . . . .
27
Models and Empirical Results . . . . . . . . . . . . . . . . . . . .
29
2.4.1
Impact of Announcements and Activity on Volatility . . .
29
2.4.2
Impact of Announcements on Intradaily Average Volatility
35
2.4
2.5
Conclusion
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3 The Information Content of Quoting Activity
39
41
3.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
3.2
The Quoting Activity Signal
. . . . . . . . . . . . . . . . . . . .
43
3.2.1
News Announcements and Quoting Activity . . . . . . . .
44
3.2.2
What announcements are common knowledge? . . . . . . .
46
3.2.3
Inter-dealer Interaction . . . . . . . . . . . . . . . . . . . .
46
3.3
Data and Descriptive Statistics . . . . . . . . . . . . . . . . . . .
48
3.4
Models and Results
. . . . . . . . . . . . . . . . . . . . . . . . .
53
3.4.1
Modeling Quote Arrival . . . . . . . . . . . . . . . . . . .
53
3.4.2
Results
55
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
iii
iv
II
CONTENTS
3.5
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
3.6
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
Volatility Dynamics Around Technical Signal Based
Trading
71
4 The Performance Analysis of Technical Chart Patterns
73
4.1
Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
4.2
Technical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . .
76
4.3
Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
79
4.3.1
Identification of Local Extrema . . . . . . . . . . . . . . .
79
4.3.2
Chart Pattern Quantitative Definitions . . . . . . . . . . .
81
4.3.3
The Performance Measures . . . . . . . . . . . . . . . . . .
85
4.3.4
Monte Carlo Simulation . . . . . . . . . . . . . . . . . . .
88
4.4
Data and Empirical Results . . . . . . . . . . . . . . . . . . . . .
89
4.5
Conclusion
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
4.6
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
95
5 The Technical Signal Based Trading Effects on Volatility
105
5.1
Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2
Technical Signal Based Trading . . . . . . . . . . . . . . . . . . . 107
5.2.1
Noise Trading and Volatility . . . . . . . . . . . . . . . . . 107
5.2.2
TSBT and News . . . . . . . . . . . . . . . . . . . . . . . 110
5.3
Methodology and Hypotheses . . . . . . . . . . . . . . . . . . . . 111
5.4
Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.4.1
The euro/dollar Exchange Rate and News Announcements
Data
5.4.2
5.5
5.6
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
The Chart Pattern Data . . . . . . . . . . . . . . . . . . . 116
Models and Empirical results . . . . . . . . . . . . . . . . . . . . 117
5.5.1
Models
5.5.2
Empirical Results
Conclusion
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
. . . . . . . . . . . . . . . . . . . . . . 118
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
List of Tables
1.1
Foreign Exchange Market Turnover, by counterparty . . . . . . .
3
2.1
Moments of the euro/dollar returns . . . . . . . . . . . . . . . . .
24
2.2
News categories . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
2.3
Daily news announcement frequencies . . . . . . . . . . . . . . . .
27
2.4
EGARCH model for impact of announcements and activity on
volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
2.5
Pre-announcement and total impact of announcements . . . . . .
33
2.6
Impact of announcements on intradaily average volatility . . . . .
38
3.1
Descriptive statistics of the number of quotes per 5-minute interval
of the first sample of banks for the period May 14 to September
10, 2001 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2
49
Descriptive statistics of the number of quotes per 5-minute interval
for the second sample of banks for the period August 24 to October
26 2001 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
3.3
News categories . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
3.4
Correlation matrix of the q estimated by the MDACP model for
the first sample of banks for the period May 14 to September 10,
2001 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5
56
Correlation matrix of the q estimated by the MDACP model for
the first sample of banks for the period August 24 to October 26,
2001 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6
Estimation results of MDACP models for sample 1, May 14 to
September 10, 2001.
3.7
56
. . . . . . . . . . . . . . . . . . . . . . . . .
58
Estimation results of MDACP models for sample 2, August 24 to
October 26, 2001.
. . . . . . . . . . . . . . . . . . . . . . . . . .
v
63
vi
LIST OF TABLES
3.8
Wald tests of equality for all banks of the effect of news
. . . . .
64
4.1
Detected chart patterns
. . . . . . . . . . . . . . . . . . . . . . .
89
4.2
Predictability of the chart patterns . . . . . . . . . . . . . . . . .
90
4.3
MAXIMUM profitability of the chart patterns . . . . . . . . . . .
91
4.4
Profitability of the trading strategy . . . . . . . . . . . . . . . . .
92
5.1
Detected chart patterns
5.2
The TSBT effects on quoting activity . . . . . . . . . . . . . . . . 119
5.3
The TSBT Effect on Volatility . . . . . . . . . . . . . . . . . . . . 120
5.4
The TSBT total impact on volatility . . . . . . . . . . . . . . . . 122
. . . . . . . . . . . . . . . . . . . . . . . 116
List of Figures
2.1
Intradaily, day-specific and overall, average volatilities (see Equation (2.4.6)).
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
3.1
Time-of-the-day effect . . . . . . . . . . . . . . . . . . . . . . . .
52
3.2
Correlogram of banks’ quoting and standardized residuals from
DACP models
3.3
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
60
Quantile plot of the Z statistic of individual banks . . . . . . . .
61
4.1
The Head and Shoulders (HS) theoretical chart pattern
. . . . .
82
4.2
The Head and Shoulders: Observed chart pattern . . . . . . . . .
83
5.1
The Geometric Configuration for the technical charts . . . . . . . 113
vii
Acknowledgements
First of all I would like to dedicate this dissertation to my mother Badria Haj
Kacem, my father abbés Ben Omrane, my wife Hédia Ben Aribi and my children
Hamza and Sarah for their huge support, encouragement, patience and love.
I am very grateful to my supervisor Luc Bauwens. It has been a privilege
to work with him. I have learned, and continue to learn, vastly from him about
econometrics. I am particulary grateful to Eric de Bodt for his support and many
very thoughtful comments.
I also thank Pierre Giot for providing me with useful comments, and I would
like also to thank my coauthors. It is a pleasure for me to work with them. Let
me present them chronologically: Luc Bauwens, Pierre Giot, Andréas Heinen and
Hervé Van Oppens.
I also thank the other members of my thesis jury: Michel Beine and Eric
Girardin for their helpful comments.
I would like to thank the families Ben Omrane and Ben Aribi, particulary
Mima Slima, my brothers Zied, Tarek and Khaled and my brother-in-law Mehdi
for giving me support.
Many thanks go also to the finance unit of the Louvain School of Management
(IAG) of Université catholique de Louvain for providing me with financial support. All my thesis has been done when I was teaching assistant in this business
school.
Chapter 1
General Introduction
The spot foreign exchange (FX) market is described by Lyons (2001) as a decentralized multiple-dealer market. Competition is provided via multiple competing
dealers, rather than limit orders as in auction markets. In addition, not all dealer
quotes are observable and simultaneous transactions can occur at different prices.
There is no physical location (or exchange) where dealers meet with customers,
nor is there a screen that consolidates all executable dealer quotes in the market. There is a screen called Reuters FXFX that displays dealer quotes, but
these quotes are firm only for a short period of time. On the contrary, quotes
displayed on the electronic brokerage screen (EBS) are firm until executed or
canceled. These quotes reflect only a subset of firm quotes in the market at any
given time.
Quotes displayed on Reuters or Telerate are firm for a short period of time
and for a relatively small volume. They are considered as indicative, since dealers
often get a slightly different quote from the one displayed. The difference is
justified by the period of time between visualization of the quote and the actual
trade. However, this is not the case in EBS since it is both a quotation and
a dealing screen. Quotes conveyed by EBS are anonymous but considered as
transaction prices. Although information displayed on EBS involves the bid, the
ask and the amount of the limit order corresponding to each of the prices, it does
not provide the identity of the bank dealer who submits the limit order. Reuters,
however, displays the bid-ask prices and the identity of the bank dealer who
submits the quotes. Therefore, Reuters is much more transparent than EBS in
terms of identifying the quoting dealer. Furthermore, being competitors, Reuters
1
2
Chapter1. General Introduction
and EBS display at any given time almost the same quotes.
The FX market is characterized by an enormous trading volume and interdealer trades account for two-thirds of this volume, and a low trade transparency.
It has moreover an uncommon information structure, and order flows in FX are
not as transparent as in other multiple-dealer markets. FX trades have no disclosure requirements, and they are not generally observable. Then, the trading
process is less informative and the information reflected in prices is reduced. The
interdealer trading is the most liquid part of the market, in the euro/dollar market, for instance, current spreads are one basis point. Dealers communicate with
each other through telephone or Reuters Dealing system. This system is different from the one that displays information. Reuters Dealing allows dealers to
communicate bilateral quotes and trade via computers rather than verbally via
telephones.
It is worth pointing out that an FX deal involves the time at which the
communication is initiated, the name of both bank counterparties, the currencies
(for example euro/dollar), the amount of the transaction, the bid and the ask
quotes, the confirmation of the deal (whether a buy or a sell), the transaction
price, and eventually the name of the banks to which both dealers have to transfer
and from which they receive currencies.
The main participants in the FX market are dealers, brokers and customers.
Each dealer represents a bank and provides two-way quotes to both customers
and other dealers. The customers consist of many institution categories, such
as financial and nonfinancial companies, and central banks. The latter intervene
within the market in order to stem the excess depreciation or appreciation of
the local money against the foreign currency. Companies order flows, however,
convey private information and are a source of asymmetric information among
dealers because they are not observable by all dealers. Each dealer has sole
knowledge of his own customer order flow.
The brokers in currency market do not trade for themselves (which is the
case for those who trade in the equity market). They only connect dealers that
might not otherwise find each other, so they facilitate the trades between dealers.
This facilitation role is important in the spot market. In turn, a dealer has two
ways to trade with another, he can call another for a quote and either buy at
3
the quoted ask or sell at the quoted bid. This way is called a direct interdealer
trade. The second way consists in trading through a broker. FX brokers do not
quote by themselves. They gather firm prices from dealers and then communicate
those prices back to dealers. A dealer might choose to post a price through a
broker because he prefers not to reveal his identity before the trade is executed
(revealing one’s identity before the trade is a necessary consequence of trading
directly). For example, one dealer may post with the broker a limit order to buy
10 millions euro against dollar at a price of 1.2130. Another dealer may post
with the same broker a limit order to sell 25 millions euro against dollar at a
price of 1.2135. If these are the best prices the broker has received on either side,
then the broker will display a two-way side, in basis points, of 30-35, and will
do so without identifying the dealers posting those prices. A third dealer can
choose to trade at one of those prices through the broker. If so, after the deal
the broker reveals the counterparty, and settlement occurs directly between the
counterparties. Both of them pay the broker a commission.
We turn now to display some data and statistics related to the FX market.
The 2005 Bank for International Settlements (BIS) survey reports that the foreign
exchange average daily turnover amounted to 1,880 billion dollars in April 2004.
Table 1.1 reports the FX market turnover by counterpart.
Table 1.1: Foreign Exchange Market Turnover, by counterparty
Reporting dealers
Other financial institutions
Non-financial customers
Estimated gaps in reporting
Total turnover
2001
689
329
156
26
1,200
2004
936
585
252
107
1,880
Daily average in April, in billions of US dollars.
Source: BIS, Triennial Central Bank Survey, March
2005.
The category ‘other financial institutions involves some investing banks, some
of which are important in dealing. According to Lyons (2001), the evidence that
dealers are included in such a category comes from the fact that this category
contains important brokered trading; FX brokers are strictly interdealer, so, these
trades belong in the interdealer category. If we add the turnover realized by the
4
Chapter1. General Introduction
other financial institutions to those of reporting dealers we get a growth of 41%
between 2001 and 2004, and interdealer trading presents the lion’s share of the
total FX trading, respectively 85% in 2001 and 81% in 2004.
One point is worth noting. Despite the consolidation of eleven European
cross markets into the Euro market, and the more efficient inventory management that is resulting from the electronic brokerage systems (which can dampen
the hot potato process described by Lyons (2001)), the FX market realized a
growth through an increase of its turnover by 57% between 2001 and 2004. The
BIS report mentions that the growth was particulary pronounced for trading
between banks and financial customers, raising total turnover by 77.8%. This
was reportedly driven to a large extent by the greater activity of institutional
investors, the leveraged investor community and corporate treasurers.
An increase in activity triggers probably a higher market liquidity but it
doesn’t implies automatically a rise in volatility. The link between volatility and
market activity was and continue to be a source of debate within the scientific
community. Volatility is a relevant feature of financial markets, and its presence
implies both risks and opportunities. Risks mean losses that could be carried out
by unexpected price movements. However, opportunities involve gains realized
when the price moves in the desired way. A large segment of the finance literature
seeks to model and forecast volatility. Moreover volatility forecasting has found
numerous applications in finance, such as portfolio management, option pricing
or risk management.
A large number of empirical studies that focus on the FX markets investigate
the link between information and volatility. They report evidence of a positive
effect of public and private information on volatility (Degennaro and Shrieves,
1997, Andersen and Bollerslev, 1998, and Cai, Cheung, Lee, and Melvin, 2001
amongst others).
Information as a determinant of asset prices plays a central role in many theoretical models. The recent literature on the microstructure of financial markets
has produced models that treat the arrival of public information or news as an
important factor with models differing in institutional characteristics such as the
presence of informed and uninformed traders. The empirical literature on the
effect of public information arrival on exchange rates has generally focused on
5
the effects of a limited number of announcements of key macroeconomic variables like unemployment or the gross national product. Such studies focus on
exchange rate changes or volatility around the time of announcements and so produce ‘snapshots’ of the market over brief intervals every month or week (Melvin
and Yin, 2000).
Empirical studies show, moreover, that volatility increases after the release
of some public information, essentially news announcements, or when there is
an unusual level of customers order flow. Andersen and Bollerslev (1998) conclude that scheduled releases occasionally induce large price changes, but the
associated volatility shocks appear short lived. The reason is probably their one
time character. Market participants may have different information sets, and
thus differ in their interpretation of the news, but the market typically settles on
a new equilibrium price after a brief period of hectic trading. Announcements
may constitute news arrivals with a well-defined content and clear termination
that endows them with a particularly short-lived impact, largely unrelated to the
strong volatility persistence observed at the daily level.
The issue of the source of volatility in the foreign exchange market is not
limited only to news announcement. Andersen and Bollerslev (1998) stipulate
that the volatility process is driven by the simultaneous interaction of numerous
components, some associated with economic news releases, some with predominantly predictable calendar effects, and some with persistent factors ascribed to
order flow. Behind the market order flow there are different bank customers and
the inter-dealer exchanges. The latter component is relatively huge compared to
the first one. It is triggered by dealer inventory control and technical trading
(i.e. trading based on technical signals)
Frankel and Froot (1990) and Carlson and Osler (2000) show that trading
based on technical signals leads to excessive volatility. Economists consider technical signals as noise and technical traders as noise traders. De Long, Shleifer,
Summers, and Waldmann (1990) and Jeanne and Rose (2002) demonstrate,
through a theoretical investigation, that the more noise traders there are the
higher the volatility created and the profits realized.
Dealers who trade on technical signals contribute to increase market liquidity,
trading activity, and trigger an excess of volatility. Technical signal based trading
6
Chapter1. General Introduction
is generated by speculators’ activity or chartists who base their trading on visual
examination of the past history of price movements. They take decisions to deal
according to technical signals provided by some specific patterns observed in
prices, but without regard to any underlying economic or fundamental analysis.
According to Black (1986) technical trading provides the essential missing
ingredient. Technical trading is trading on noise as if it were information. Dealers
who trade on noise are willing to trade even though from an objective point of
view they would be better off not trading, because they believe that the noise
they are trading on is information. The more technical trading there is, the
more liquid the markets are, in the sense of having frequent trades that allow
to observe prices. But technical trading puts noise into prices. The price of the
currency reflects both the information that information traders trade on and the
noise that technical traders trade on.
Black (1986) adds that the short term volatility of price will be greater than
the short term volatility of value. Since noise is independent from information
in this context, when the variance of the price moves caused by noise is equal
to the variance of the price moves caused by information, the total variance of
the price will be roughly twice the variance of the value. There will be an excess
volatility stemmed from technical trading.
Nevertheless, almost all the previous empirical research has investigated the
volatility dynamics during and after the release of public information.
Re-
searchers use ARCH-type or realized volatility models and they proxy public
information by market news announcements. Studies focusing on the effect of
noise or technical trading on volatility are limited to theoretical results without
any empirical evidence. As a consequence, the aim of this dissertation is to answer to the following question: how does foreign exchange volatility behave, in
the very short term, around public information and technical signals ? To answer to this question we try to study the volatility dynamics before, during and
after public news announcement and technical chart pattern signals. In order to
meet this objective, we implement different methodologies specific to the different chapters of the dissertation. Each chapter tries to answer to a sub-question
emerging from the main question of the thesis.
This thesis contains four chapters. Beyond this introduction, Chapter 2,
7
entitled ‘News announcements, market activity and volatility’, focuses on the
impact of public and private information on volatility. We analyze in a first
step the effect of nine categories of news announcements on the volatility in
the euro/dollar currency market. The second step consists in investigating the
volatility dynamics before, during and after scheduled and unscheduled news
announcements.
We focus more particulary on the reaction of volatility in the pre-announcement
period as this topic has not yet been dealt with extensively in the existing literature. Taking into account the specific features of public and private information,
the news impact on volatility can take place both before and after the announcement.
The econometric analysis is performed on a high frequency data set of 5minute regularly time-spaced euro/dollar quotes. The time period ranges from
May 15 to November 14, 2001. Our database also includes the news headlines
that were released on the news-alert screens. We implement an EGARCH model
where we control for intraday seasonality, news arrival (represented by dummy
variables) and quoting frequency. Our results show that the euro/dollar return
volatility increases before the announcement of scheduled news. We surmise that
speculative/anticipatory trades, the possible flow of private information or the
re-balancing of positions by traders who prefer to avoid announcement ‘surprises’
lead to this increase in volatility. When an announcement is not scheduled, the
evidence of a volatility increase during the pre-announcement phase is tenuous.
Moreover, for four categories of announcements, volatility increases in total over
the pre-announcement and post-announcement periods. We find also that market activity, proxied by the quoting activity, significantly impacts volatility as
expected by the theoretical literature on the order flow.
Chapter 3, entitled ‘The information content of quoting activity’, is a continuation of the first one. Quoting activity, treated as an exogenous variable in
Chapter 2, become an endogenous one. Some researchers, such as Melvin and
Yin (2000), consider the quoting activity as a proxy of volatility.
In this chapter, we investigate the information content of dealers’ quoting
activity as measured by the frequency of price revision in the euro/dollar currency market. The FX market is a market characterized by a very low degree
8
Chapter1. General Introduction
of transparency, in particular dealers have no information about the other ones.
However, FX dealers get real time information through electronic screens about
public news announcements, bid-ask quotes with the identity of the dealer who
submit them and other banks quoting activity. The latter constitutes the only
information that a bank dealer has about other banks quotes. On the other
hand, it is a well documented fact that dealers revise their quotes in order to reflect their reaction to public news announcements (Andersen, Bollerslev, Diebold,
and Vega, 2002) or their detention of private information coming from their customers’ order flow (Lyons, 1995, Evans and Lyons, 2002, Cao, Evans, and Lyons,
2003).
With respect to the previous literature, we study in this chapter the hypothesis according to which individual dealers’ quoting activity is sensitive to news,
and dealers try to infer other dealers’ private information and reaction to news
announcements through their quoting activity. Hence, dealers could use two information channels to build their reaction to news events. The first channel takes
place directly through the news broadcasters. This corresponds, in a sense, to
dealers’ spontaneous or direct reaction to the announcements. The second channel, which could be a source of noise for the first one, is carried out through the
quoting activity of the other dealers.
Using the same data set as in Chapter 2 and the multivariate autoregressive
double Poisson model we find firstly that the banks’ reaction to news is quite
heterogeneous. There are differences across banks not only in terms of which
news they react to, but also in terms of how they react to the announcements.
Some dealers increase their quoting activity whilst others decrease it or keep
it unchanged as a response to the same news. This suggests that banks react
differently, either because they have different private information or because they
interpret the public news announcements differently in terms of their implications
for the exchange rate.
Our second result consists in showing that there are significant inter-dealer
effects. This means that there are systematic lead-lag relations in the intensity
of quote revision of the various banks. We offer evidence suggestive of the fact
that dealers observe the quoting activity of others to infer useful information like
other dealers’ private information and reaction to public news announcements.
9
If a dealer increases (decreases) his quoting activity following the publication of
news this means that he raises (reduces) the frequency his price revision. Banks
typically react to some news announcements and not to others. They also respond
to the quoting activity of certain banks. We interpret this as meaning that when
banks are unsure about the interpretation of certain news announcements, they
tend to wait a little and watch other banks that they might perceive as being
better informed in order to infer their interpretation of the news content. This
means that a public news announcement could have no effect on a given bank’s
quoting activity either because they consider that news as irrelevant or they
prefer to wait and see how better informed dealers react. In that sense we argue
that quoting activity is a potentially useful information channel in an otherwise
very opaque foreign exchange market.
After our investigations about volatility dynamics around public information
like news announcement, we investigate this dynamics around noise trading, essentially technical chart signals. The next chapter, entitled ‘The performance
analysis of chart patterns’, involves an introduction to the basic tools of technical analysis, more precisely the chart pattern signals. The latter is considered by
a category of FX dealers, called also technical dealers, as the relevant information that has a significant impact on their trading decisions. They do not care
about public information, since according to them prices embody all aspects of
the market, balancing all the forces of supply and demand.
In Chapter 4, we start by checking the existence of twelve chart patterns in
the euro/dollar foreign exchange market. We use two identification methods for
detecting technical charts. The first method, also used in the literature, considers
only prices at the end of each time interval. The second method, which is new
compared to those used in the literature, takes into account both the highest and
the lowest price in each interval of time corresponding to a detected pattern.
The methodology adopted in this chapter consists in identifying regularities
in the time series of currency prices, corresponding also to the same data-set implemented in the previous chapters, by extracting nonlinear patterns from noisy
data. We take into consideration significant price movements which contribute
to the formation of a specific chart pattern and we ignore random fluctuations
considered as noise. We do this by adopting a smoothing technique in order to
10
Chapter1. General Introduction
reduce the noise. The smoothing technique allows to identify significant price
movements which are only characterized by sequences of extrema. The detected
extrema are analyzed through twelve recognition pattern algorithms, each of
them corresponding to a defined chart pattern. Our purpose is also to check the
predictability and profitability of each type of chart pattern. In addition, we
intend to test the usefulness of our contribution regarding the extrema detection
method. Although Osler (1998) and Chang and Osler (1999) briefly mention
these prices, most of the previous studies focused on chart patterns do not give
much interest to high and low prices. This is in sharp contrast to the majority of practitioners (in particular dealers) who use high and low prices in their
technical strategies through bar charts and candlesticks. In addition, Fiess and
MacDonald (2002) show that high, low and close prices carry useful information
for forecasting the volatility as well as the level of future exchange rates. Consequently, we investigate also the sensitivity of the chart patterns to both extrema
detection methods. To evaluate the statistical significance of our results, we run
a Monte Carlo simulation. We simulate a random walk process to construct artificial series. Each of them has the same length, mean, variance and starting
value as the original observations.
The results of this chapter show the apparent existence of some chart patterns in the euro/dollar intra-daily foreign exchange rate. More than one half
of the detected patterns, according to both detection methods, seem to have a
significant predictive success. Nevertheless, only two patterns from our sample of
twelve present a significant profitability prospect which is however too small to
cover the transaction costs. We show, moreover, that the second extrema detection method provides higher but riskier profits than those provided by the first
one. These findings are in accordance with those found by Levy (1971), Osler
(1998), Dempster and Jones (1998b), Chang and Osler (1999).
Technical chart signals present indeed an important predictive success, otherwise practitioners would not rely on them. According to Allen and Taylor
(1992) and Menkhoff (1998) more than 60% of FX dealers implement technical
signals to build their decisions. They show, through questionnaire evidence, that
technical signals are used either as the primary or the secondary information
source by more than 90% of currency dealers trading in London. Furthermore,
11
60% judge technical signals to be at least as important as fundamentals. Thus,
the questions that emerge at this stage are: How can technical analysis impact
volatility ? What about the volatility dynamics around the technical signals ?
The answers to these two questions are provided in Chapter 5 which is the last
one of the dissertation.
In this chapter, entitled ‘The technical signal based trading effects on volatility’, we examine the technical signal based trading (TSBT) effect on foreign
exchange (FX) rate volatility. We use the basic technical signals, essentially the
chart patterns trading signals and we implement the same methodology as Lo,
Mamaysky, and Wang (2000) and Ben Omrane and Van Oppens (2006) to recognize chart patterns. Based on the euro/dollar time series, news announcements
database, and chart pattern signals, we provide evidence about the volatility dynamics around TSBT. We find firstly that volatility drops during the completion
of the technical chart patterns, before the occurrence of the chart signal, where
the exchange rate moves within resistance and support levels. Pring (1985) defines the support level as a zone representing a concentration of demand, and the
resistance area as a zone corresponding to a concentration of supply. Murphy
(1986) defines the support level as the area on the chart where buying interest
is sufficiently strong to overcome selling pressure. As a result, a decline is halted
and prices turn back again. The resistance level is the opposite of support. We
also find that when the technical signal occurs, it generates TSBT triggering
volatility increases. A breakout takes place once the exchange rate crosses the
support or the resistance level just after the chart completion. It attracts the
attention of chartists and creates heterogeneity within the market participants.
To summarize, this thesis contributes to the empirical finance literature on
intradaily exchange rate volatility as follows. First we assess the previous results,
about the effect of public information on volatility, on the new euro/dollar market
and extend these by distinguishing between the impact on volatility of scheduled
and unscheduled news in three time periods centered around the release of the
news, i.e. the pre-announcement, contemporaneous, and post-announcement periods. We present evidence that volatility increases in the pre-announcement
period of scheduled news, which could be triggered by speculative or anticipatory trades. Second, we show that FX dealers quoting activity reacts to news
12
Chapter1. General Introduction
announcements and it conveys useful information. This information involves,
essentially, the reaction of other FX dealers to the public events. The third
contribution consists in presenting a new approach to recognize technical chart
patterns from a time series, and shedding light on the predictive success of the
technical chart signals. Finally, the last contribution consists in the finding that
technical information, considered by economists as ”noise”, has a significant effect on volatility. Volatility declines during the technical chart completion, when
the exchange rate moves within the support and resistance levels. A possible
explanation for such a volatility drop, is that the foreign exchange market is
dominated by fundamental traders, sharing the same expectations and featuring
a homogeneous behavior.
The last finding consists of volatility increase just
after the chart completion, when the exchange rate crosses the support or resistance level. In such an area, the attention of chartists is attracted and TSBT
occurs creating heterogeneity within the market participants, and triggering an
increase in volatility.
Part I
Volatility Dynamics Around
Public Information
13
Chapter 2
News Announcements, Market
Activity and Volatility
This chapter is based on a paper published in the Journal of International
Money and Finance, 2005, 24, 1108-1125, and co-authored by Luc Bauwens
and Pierre Giot.
2.1
Introduction
The impact of information on the volatility of foreign exchange (FX) returns has
been theoretically and empirically studied in several papers, e.g. Degennaro and
Shrieves (1997), Andersen and Bollerslev (1998), Evans and Lyons (1999), Melvin
and Yin (2000) and Cai, Cheung, Lee, and Melvin (2001). As stressed in the literature on market microstructure (see O’Hara, 1995), the ‘information’ variable
includes both a public and a private component. Regarding the FX market, both
public and private components are strongly related to currency market news announcements. The public component is made up of announcements which take
place at fixed times (which we call scheduled public announcements), or at random times (unscheduled public announcements). Regarding private information,
the most recent literature on the microstructure of exchange rates allows for two
types of private information. Firstly, some market participants could have access to yet unreleased information by central banks or government agencies (i.e.
payoff related private information in the terminology used by Lyons, 2001). Secondly, the notion of private information can be extended to include the so-called
15
16
Chapter2. News Announcements, Market Activity and Volatility
unrelated payoff information, i.e. private information that a dealer has regarding
interim states of the market (for example, a dealer knows that another dealer is
keen on selling a large euro/dollar position, which should depress prices in the
short run). Because this second possibility is the most probable type of private
information event in the FX market, private information is strongly related to
order flow between traders and their customers.1
With respect to the previous literature on FX volatility about the impact of
news announcements, the aim of this paper is twofold. Firstly, we analyze the
impact of a more refined and extended set of nine categories of news announcements on FX volatility in the new euro/dollar market. Secondly, we investigate
the volatility dynamics before, during and after scheduled and unscheduled news
announcements. The contribution of this chapter therefore consists in assessing the previous results on the new euro/dollar market and extending these by
distinguishing between the impact on volatility of scheduled and unscheduled
news in three time periods centered around the release of the news, i.e. the preannouncement, contemporaneous, and post-announcement periods. We focus
more particulary on the reaction of volatility in the pre-announcement periods as
this topic has not yet been dealt with extensively in the existing literature. Taking into account the specific features of public and private information, the news
impact on volatility can take place both before and after the announcement.
In the case of scheduled news announcements, volatility increases in the preannouncement period could be due to anticipatory trades by dealers who open
positions to profit from some personal beliefs, i.e. they hope that the actual news
outcome will coincide with their forecast of the outcome. A post-announcement
volatility increase can be attributed to heterogeneity of interpretations of the
contents of the news, surprised reactions and closing of positions based on prior
anticipations. On the other hand, in the case of unscheduled news announcements, an increase in volatility before the announcement is probably linked, as
suggested by Degennaro and Shrieves (1997), to the presence of informed traders
who exploit their privileged information.
The econometric analysis, implemented within the chapter, is performed on
a high frequency data set of 5-minute regularly time-spaced euro/dollar quotes.
1
See Lyons (2001) and references therein for additional information and recent developments.
2.2. News Announcements and Volatility
17
The time period ranges from May 15 to November 14, 2001. Our database also
includes the news headlines that were released on the news-alert screens. Regarding these news announcements, we consider a much larger set of news events
(classified into nine general categories) than those used in the previous literature. Furthermore, to highlight the effect of the possible ‘surprise’ contained
in the most important scheduled US macroeconomic figures, we distinguish between so-called positive and negative news (by computing the difference between
expected and realized values). As in the previous literature, we also take into
account the effect of private information (proxied by the deseasonalized quoting frequency) on the volatility of the euro/dollar returns. More generally, the
focus of this work is therefore on the economic determinants of the euro/dollar
return volatility with particular attention to the links between the information
flow and the market reactions measured by volatility and quoting frequency. We
use an EGARCH model where we control for intraday seasonality, news arrival
(represented by dummy variables) and quoting frequency.
Our results show that the euro/dollar return volatility increases before the announcement of scheduled news. We surmise that speculative/anticipatory trades,
the possible flow of private information or the re-balancing of positions by traders
who prefer to avoid announcement ‘surprises’ lead to this increase in volatility.
When an announcement is not scheduled, the evidence of volatility increase during the pre-announcement phase is tenuous, except for rumors of central bank
interventions. Moreover, for four categories of announcements, volatility increases
in total over the pre-announcement and post-announcement periods.
The follow-up of this chapter is divided in four sections. In Section 2.2 we
present a brief review of the empirical market microstructure issues related with
FX news announcements and volatility. In Section 2.3 we describe our data.
We present our models and we discuss the estimation results in Section 2.4. We
conclude in Section 2.5.
2.2
News Announcements and Volatility
The impact of news announcements on FX return volatility has been studied in
several papers (see the citations at the beginning of the introduction). These
18
Chapter2. News Announcements, Market Activity and Volatility
studies have focused on the most active currency markets (dollar/mark, dollar/yen, and sterling/dollar). Analyzing the factors, such as news announcements, which affect FX volatility is an important topic in empirical finance as
this provides an economic explanation of the volatility in currency markets. Regarding news announcements, it is important to distinguish between scheduled
announcements (the announcement is planned, e.g. US macroeconomic figures)
and unscheduled ones (e.g. rumors of central bank interventions). Unscheduled
announcements can lead to both public and private information, according to
their market effects. If the information is public, then an unscheduled event
should not be preceded by an abnormal increase in volatility. On the other hand,
an unexpected rise in volatility during the pre-announcement period indicates
that informed trades (due to private information) take place. For scheduled announcements, an abnormal increase in volatility during the pre-announcement
period can be linked to speculative trades initiated on the basis of anticipations,
or to traders who close their positions to avoid ‘surprises’ when the news is released.
The link between volatility and private information was suggested in the theoretical work of Admati and Pfleiderer (1988). In their model, informed traders
prefer to transact during periods of high trading activity in order to maximize
the potential profit that comes from their private information. In addition, liquidity motivated traders are also active during this period to profit from the
high market activity and the available liquidity. An important empirical implication of this model is that private information leads to increased price volatility
during periods of high trading activity. Set in the FX framework, this type of
analysis suggests that the dynamics of currency quotes will therefore be shaped
by the news announcements and by the traders’ private information. When no
private information is at play, FX quotes should adjust to events after the news
announcements. When there is information asymmetry, currency quotes should
adjust to private information before the news announcement and continue their
adjustment during the announcement, while volatility should decrease during the
post-announcement period. In their empirical study, Degennaro and Shrieves
(1997) test the theoretical results of Admati and Pfleiderer (1988). They analyze the pattern of volatility around news announcements and study the impact
2.2. News Announcements and Volatility
19
of the release of macroeconomic, economic policy and interest rate reports on
the volatility of the dollar/yen exchange rate. According to their results, a high
market activity leads to increases in volatility and spread, which they interpret
as private information effects. These results are in agreement with Lyons (1995,
1997) and Evans and Lyons (1999), who show that a significant part of the FX
order flow is made up of deals between traders and their customers (deals that
generate private information). Andersen and Bollerslev (1998) and Cai, Cheung,
Lee, and Melvin (2001) analyze the sensitivity of short and long term volatility
(for the dollar/mark and dollar/yen FX quotes) with respect to US macroeconomic announcements and seasonal factors. Andersen and Bollerslev (1998) show
that announcements have a significant positive effect, but for a very short period
of time. Seasonal factors, such as the opening of the local markets, lunch breaks,
some specific days of the week (Thursday and Friday, i.e. days of US macroeconomic announcements), also lead to volatility increases. Cai, Cheung, Lee, and
Melvin (2001) confirm the results of Andersen and Bollerslev (1998) and show
the importance of the positive effect on volatility of the order flow, in comparison
with the news announcements. Eddelbüttel and McCurdy (1998) do not classify
the news in categories but count instead the number of events in a given time
interval. Deseasonalized counts lead to increases in both market activity and
return volatility for dollar/yen quotes.
Next to return volatility, a second important variable is market activity. FX
market activity, measured by the number of quotes in a given time interval, is
considered in many papers (e.g. Goodhart and Figliuli, 1991, Bollerslev and
Domowitz, 1993, Degennaro and Shrieves, 1997, and Melvin and Yin, 2000) as
a proxy for volatility and should therefore be affected by news announcements.
Market activity is also a proxy for private information (Degennaro and Shrieves,
1997, Evans and Lyons, 1999, and Rime, 2000). A significant part of the order
flow is due to transactions between traders and their customers. Customer-dealer
trades are not observable by the other market participants. Therefore they are a
source of information asymmetry among dealers and can be considered as private
information. Consequently, when traders execute these orders, they play a role of
intermediation or transfer of information between their clients and other traders.
For example, Lyons (1997) shows that traders could be motivated to exploit the
20
Chapter2. News Announcements, Market Activity and Volatility
information contained in their own customer orders to take speculative positions
that distort this information intermediation. The empirical study of Degennaro
and Shrieves (1997) shows that the unexpected component of market activity
(i.e. market activity adjusted for its seasonal component) can be viewed as a
proxy for private information: unexpected market activity leads to an increase
in volatility and a widening of the spread.2
Regarding the methodological framework, the sensitivity of FX volatility to
exogenous factors is usually assessed and quantified using models, as for example
in Degennaro and Shrieves (1997) and Melvin and Yin (2000). The use of a
conditional volatility model is justified by the key features of the distribution of
FX returns (volatility clustering and fat tails). Note however that Andersen and
Bollerslev (1998) and Cai, Cheung, Lee, and Melvin (2001) set their analysis in
the framework of realized volatility, i.e. volatility computed as the sum of high
frequency absolute returns.
Finally, the literature on the volatility of FX returns stresses that market
openings and closings, week-ends and some days of the week lead to significant cyclical factors in the pattern of volatility (Bollerslev and Domowitz, 1993,
Andersen and Bollerslev, 1996 and 1998, Degennaro and Shrieves, 1997). For
example, Andersen and Bollerslev (1998) show that scheduled news announcements have a seasonal impact on volatility. These announcements induce both
a cyclical and a stochastic component, the latter stemming from incorrect anticipation by the market participants and thus a ‘surprise’ effect. To remove the
seasonal effects and to highlight the stochastic components, it is first necessary
to detect and identify the factors that are likely to generate periodic effects. A
number of methods have been put forward in the literature. Degennaro and
Shrieves (1997) introduce the ‘cross-sectional’ average of market activity in the
volatility equation. In order to model the seasonal impact, Andersen and Bollerslev (1998) introduce a flexible Fourier form (FFF in what follows), i.e. a sum
of sinusoids, to detect intraday cycles. Melvin and Yin (2000) divide returns by
their cross-sectional average to remove seasonality.
2
The relationship between private information and the size of the spread can be traced back
to Glosten and Milgrom (1985).
2.3. Data Description
2.3
2.3.1
21
Data Description
Euro/dollar Exchange Rate Data
The euro/dollar FX market is a market maker based trading system, where three
types of market participants interact around the clock (i.e. in successive time
zones): the dealers, the brokers and the customers from which the primary order
flow originates. The most active trading centers are New York, London, Frankfurt, Sydney, Tokyo and Hong Kong. A complete description of the FX market
is given by Lyons (2001).
To compute the returns used for the estimation of the models reported in
Section 2.4, we bought from Olsen and Associates a database made up of ‘tick-bytick’ euro/dollar quotes for the period ranging from May 15 to November 14, 2001
(i.e. 26 weeks and three days). This database includes 3,420,315 observations.
As in most empirical studies on FX data, these euro/dollar quotes are market
makers’ quotes and not transaction quotes (which are not widely available).3
More specifically, the database contains the date, the time-of-day time stamped
to the second in Greenwich mean time (GMT), the dealer bid and ask quotes, the
identification codes for the country, city and market maker bank, and a return
code indicating the filter status. According to Dacorogna, Müller, Nagler, Olsen,
and Pictet (1993), when trading activity is intense, some quotes are not entered
into the electronic system. If traders are too busy or the system is running at
full capacity, quotations displayed in the electronic system may lag prices by a
few seconds to one or more minutes. We retained only the quotes that have a
filter code value greater than 0.85.4
From the tick data, we computed mid-quote prices, where the mid-quote is the
average of the bid and ask prices. As we use five-minute returns, we have a daily
grid of 288 points. At the end of each interval, we used the closest previous and
next mid-quotes to compute the relevant price by interpolation. The mid-quotes
3
Danielsson and Payne (2002) show that the statistical properties of 5-minute dollar/DM
quotes are similar to those of transaction quotes.
4
Olsen and Associates recently changed the structure of their HF database. While they
provided a 0/1 filter indicator some time ago (for example in the 1993 database), they now
provide a continuous indicator that lies between 0 (worst quote quality) and 1 (best quote
quality). While a value larger than 0.5 is already deemed acceptable by Olsen and Associates,
we choose a 0.85 threshold to have high quality data. We remove however almost no data
records (Olsen and Associates already supplied us with data which features a filter value larger
than 0.5), as most filter values are very close to 1.
22
Chapter2. News Announcements, Market Activity and Volatility
are weighted by their inverse relative time distance to the interval endpoint. Next,
the return at time t is computed as the difference between the logarithms of the
interpolated prices at times t−1 and t, multiplied by 10,000 to avoid small values.
Because of scarce trading activity during the week-end, we excluded all returns
computed between Friday 22h05 and Sunday 24h. In addition, we excluded the
first return of each Monday to avoid possible biases due to the lack of activity
during the week-end. The total number of returns is 37,653.
The final data transformation consists of adjusting the returns for the intradaily component of volatility. The seasonally adjusted (SA) returns are obtained by dividing the returns by their cross-sectional intradaily average volatility.
An average value of volatility is computed and attributed to the endpoint of every 5 minute interval. The time series of these values constitutes an intradaily
‘seasonal index’ of volatility. This can be done by considering all days of the
week as similar (an overall index), or by computing a specific index for each day
of the week. In Section 2.4.2 we explain the details of the procedure we adopted
to compute these indices and to adjust the returns. Figure 2.1 displays these
indices. Comments on their pattern are provided in Section 2.3.3, since they are
related to news announcements.
It is worth pointing that we took care in making the seasonal adjustment to
account for the time change (to winter time) that occurred on October 29, 2001.
This concerns GMT hours from 6h until 21h (corresponding to market times in
Europe and the USA).
Table 2.1 presents summary statistics of the euro/dollar returns, before and
after seasonal adjustment. The mean of the SA returns is almost equal to zero
and their distribution has fatter tails than the normal, but it is almost perfectly
symmetric. The unadjusted returns are much more leptokurtic, and feature a
positive skewness coefficient. There is a highly significant negative autocorrelation of order one and of order two in both series of returns (although the level
of autocorrelation is much closer to 0 at the order 2). The negative autocorrelation in FX returns has been discussed in the academic literature. According
to Goodhart and Figliuli (1991), the negative autocorrelation stems from constraints in the control of positions, while according to Bollerslev and Domowitz
(1993) and Lo and MacKinlay (1990), this feature comes from the computation
2.3. Data Description
23
Figure 2.1: Intradaily, day-specific and overall, average volatilities (see Equation
(2.4.6)).
The panel of each day displays the average volatility of that day and the average volatility
when all days are treated as having the same pattern. This overall volatility is shifted upwards
by 4 units. It is also shown in the bottom right panel (without the shift)
24
Chapter2. News Announcements, Market Activity and Volatility
of asynchronous price series at the interval endpoints.
Table 2.1: Moments of the euro/dollar returns
Mean
Standard deviation
Skewness coefficient
Kurtosis coefficient
JB normality test
Autocorrelation of order 1
Autocorrelation of order 2
Autocorrelation of order 3
Returns SA returns
0.007
0.00006
4.41
1.01
0.34
0.02
15.2
4.41
233,879
5,922
-0.086
-0.12
-0.025
-0.027
0.000
-0.001
The SA (seasonally adjusted) returns are the returns divided
by their intradaily average volatility (see Section 2.3.1 and the
appendix for details). The 5-minute returns have been premultiplied by 10,000 (to avoid small values). The number of
observations is 37,653, corresponding to the period from May
15 to November 14, 2001.
2.3.2
News Announcement Data
Our news announcements database includes the news headlines that were released on the Reuters news-alert screens over the May 15 to November 14, 2001
period. These events are time stamped to the minute and are a key feature of
our news announcements analysis. In addition to the news headlines, Reuters
also provides an economic agenda which gives the day and time of some of the
announcements that are scheduled in the following week. We work with nine
categories of news announcements (Table 2.2 lists all the news categories). The
announcements can be classified into two broad groups: scheduled and unscheduled announcements. The first group contains US macroeconomic figures, more
specifically employment reports, producer and consumer price indices, gross domestic product and other important figures. This group also includes European
macroeconomic figures, scheduled speeches of senior officials of the government
and of public agencies, such as the Chairman of the Federal Reserve, the Chairman of the European Central Bank, and economy and finance ministers. We also
classify as scheduled the US and European interest rate reports, although this is
debatable. The meeting days of the ECB and the Fed are certainly scheduled,
and agents know when the meeting starts and approximately ends. However, the
2.3. Data Description
25
exact time of the announcements is not exactly known in advance. The second
group is made up of the forecasts of key institutes and specialized organizations,
such as the IMF, the World Bank, and the IFO institute (an influential servicebased research organization in Germany). This group also contains the declarations of OPEC members, the rumors of central bank interventions and other
extraordinary events (natural disasters, wars, terrorist attacks,. . . ). To highlight
the effect of the possible ‘surprise’ contained in the scheduled US macroeconomic
figures, we distinguish positive news from negative news by computing the difference between the expected and realized values: if the realization is larger than
the expectation and is a figure which corresponds to economic growth, the news
is classified as positive; if the actual figure implies instead higher-than-expected
inflation or a slowdown of the economy, it is regarded as negative. The expected
values are given on Reuters screens a few days before the news announcements.
26
Chapter2. News Announcements, Market Activity and Volatility
Table 2.2: News categories
Scheduled news announcements
Positive Negative
η1,τ
η2,τ
Employment report
+
ISM index(ex NAPM)
+
Whole sales
+
Gross domestic product (GDP)
+
Producer price index (PPI)
+
Retail sales
+
Housing starts
+
Consumer confidence index
+
Consumer price index (CPI)
+
Construction spending
+
Car sales
+
Business inventories
+
Housing completions
+
Import prices
+
Current account deficit
+
Non-farm productivity
+
Personal income
+
Real earnings
+
House sales
+
3-European macroeconomic figures
η3,τ
4-Speeches of senior officials of the government
η4,τ
1 and 2-US macroeconomic figures
and of public agencies
5-US and European interest rate reports
Unscheduled news announcements
6-Forecasts made by economic institutes
7-Declarations of OPEC members
8-Rumors of central bank interventions
9-Extraordinary events
η5,τ
η6,τ
η7,τ
η8,τ
η9,τ
The events are the news headlines released on the Reuters money news-alerts.
For US macroeconomic figures, we separate positive and negative news-alerts by
comparing the expected and the announced numbers. If the actual numbers are
larger than the expectations for economic variables that contribute to economic
growth, the announcements are classified as positive (+). If the actual news
release means more inflation or a forthcoming economic slowdown, it is classified
as a negative news announcement (-). The expected values are given on Reuters
screens a few days before the news announcements.
The employment report includes the unemployment figures.
ISM is the abbreviation for the Institute of Supply Management, ex NAPM,
National Association of Purchasing Management. It is a monthly composite
index and gives the earliest indication of the health of the manufacturing sector.
The symbol ηj,τ is the coefficient of the dummy variable dj,τ in the equations
reported in Tables 2.4 and 2.5. In Table 2.6, the coefficients are called ηj .
2.3. Data Description
27
The impact of the scheduled announcements should include both a deterministic (seasonal) component and a stochastic component. The latter reflects the
surprise effect due to the discrepancy between the actual contents of the news and
the expected contents before the release. US macroeconomic figures are usually
released at 12h30 and 14h00 GMT. European macroeconomic figures are mostly
released around 7h30 and 10h00 GMT. As far as the announcement days are
concerned, Table 2.3 presents the number of news belonging to each category for
each day of the week. The number of news announcements during the 6-month
period of our study is equal to 1,040. For example, 61% of US macroeconomic
figures are released on Thursday and Friday. Therefore, we expect that regularly
scheduled news announcements induce a seasonal effect in volatility.
Table 2.3: Daily news announcement frequencies
Monday Tuesday Wednes. Thursday Friday
Category Tot % Tot % Tot % Tot
% Tot %
d1
11 13
13 15
12 14
23
26
29 33
d2
5
7
12 17
8 12
25
36
19 28
d3
42 14
68 23
76 25
59
20
54 18
d4
45 22
39 19
46 22
41
20
36 17
d5
21 26
27 33
13 16
15
19
5
6
d6
24 27
16 18
25 28
14
16
11 12
d7
21 21
17 17
24 24
19
19
18 18
d8
5 42
1
8
2 17
2
17
2 17
d9
22 23
14 15
20 21
24
25
15 16
Total
196 19 207 20 226 22 222
21 189 18
Total
88
69
299
207
81
90
99
12
95
1040
Entries are the numbers of announcements, and in italics, the percentages per day.
d1: Positive US macroeconomic figures. d2: Negative US macroeconomic figures. d3:
European macroeconomic figures. d4: Speeches of senior officials of the government
and of public agencies. d5: US and European interest rate reports. d6: Economic
institute forecasts. d7: OPEC member declarations. d8: Central bank intervention
rumors. d9: Extraordinary events.
2.3.3
The Intradaily Average Volatility
As mentioned above, the euro/dollar currency is almost continuously traded in
FX markets that belong to different time zones, but the activity (measured by
the number of quotes per 5 minute interval) varies a lot over the 24 hours. It
is also well-known that the volatility varies. Figure 2.1 illustrates the intradaily
28
Chapter2. News Announcements, Market Activity and Volatility
seasonal pattern of the average volatility of five-minute returns. The bottom
right panel of that figure shows the average volatility when all days of the week
are assumed to have the same pattern. It shows some components which exist
more or less every day, although days differ from each other as can be concluded
informally by looking at the other panels.
Volatility increases after midnight, i.e. at the opening of the Singapore and
Hong Kong markets, one hour after the opening of the Tokyo market and two
hours after Sidney. Around 4h GMT, volatility decreases because of the lunch
break in the four Asian financial markets. Thereafter, volatility increases again
because trading activity resumes in the Asian markets and it reaches a local
maximum around 7h-8h GMT, i.e. right after the opening of the key European
markets such as London and Frankfurt. Volatility increases around the opening
and closing times of the regional markets are in agreement with Admati and
Pfleiderer (1988), who show that these periods are characterized by a sustained
level of market activity which attracts different categories of traders. In addition,
Lyons (1997) shows that, because traders have to control or close their positions
at the end of every day, they increase their activity right before the closing of
trading and just after the market opening to get rid of unwanted risky positions.
Because of the lunch break in Europe, volatility decreases around 11h30. A
rebound in volatility occurs at 12h GMT as New York opens for trading. The
big spike between 12h and 13h is due to US news announcements at 12h30 on
Friday. Between 12h and 16h GMT volatility is generally at its highest level
due to the simultaneous activity of the American and European markets. Just
before the New York lunch break (which is clearly visible in the figure around
17h GMT), there is a short volatility increase due to the closing of the European
markets. Volatility increases also around 21h GMT, i.e. when the New York
trading session ends. Starting at 21h GMT, a short period of stability is then
observed until the opening of the Sidney and Tokyo markets, which leads to an
increase in volatility.
The five day-specific panels of Figure 2.1 indicate that the seasonality of
volatility is also dependent on the day of the week: shocks are mainly observed
on Tuesday and Wednesday around 8h, 10h and 16h30 (interest rate reports
and European macroeconomic figures), Thursday and Friday around 12h, 12h30
2.4. Models and Empirical Results
29
and 16h30 (US macroeconomic figures and speeches of senior officials of the government). Therefore, the seasonality of volatility stems partly from the cyclical
component of scheduled news announcements. In addition, Figure 2.1 shows that
volatility increases for a short period of time in all markets at the beginning of
the trading sessions of each Monday. Volatility increases every Monday around
0h, 8h, 13h, and 22h, which are respectively the opening hours of the Asian,
European, American, and Australian markets. These increases during the first
minutes of trading of every week are linked to the control of positions: FX traders
who accumulate customers’ orders at the end of the Friday session and who could
not settle their positions have to keep them during the week-end. To minimize
the risk of these positions, dealers are keen on executing their remaining orders
in the first minutes of the Monday session. They do so by quoting attractive
prices to attract counterparts and quickly close their positions.
2.4
Models and Empirical Results
2.4.1
Impact of Announcements and Activity on Volatility
We want to study the sensitivity of volatility with respect to the news announcements and market activity . Following the discussion in the introduction and in
Section 2.2, we want to answer four broad questions:
1. Is there an increase of volatility, considering the reaction of the market
before announcements?
2. Is there an increase of volatility, considering the total reaction of the market
(before, during, and after announcements)?
3. Is there a difference between scheduled and unscheduled announcements in
the answers to the previous questions?
4. Is there an impact of ‘market activity’ on volatility, in particular of the
unexpected component of market activity , considered as a proxy of traders’
private information?
30
Chapter2. News Announcements, Market Activity and Volatility
We use the EGARCH model of Nelson (1990) to model SA returns (denoted
qt ) and their conditional variance, denoted ht (our measure of volatility).5 The
level of returns is modelled by a moving average process of order two to account
for the detected autocorrelation (see Table 2.1):
qt = θ0 + ut + θ1 ut−1 + θ2 ut−2 ,
(2.4.1)
and the error term ut by an EGARCH(2,2) process:6
ut =
ln ht
p
ht ²t ,
2 ³
h
i
´
X
p
= ω+
βi ln ht−i + αi | ²t−i | − 2/π + γi ²t−i
+
i=1
9
3
XX
(2.4.2)
(2.4.3)
ηj,τ dj,τ,t + φ ast−1 + δ amt−1 .
j=1 τ =1
The innovations ²t are assumed identically and independently distributed. We
proceed as if their distribution was normal, i.e. we estimate the model by the
quasi maximum likelihood (QML) method, thus accounting for lack of normality in computing asymptotic standard errors. Table 2.4 reports the estimation
results. The standardized residuals and squared residuals are not autocorrelated
(according to Q-statistics). The EGARCH coefficients are significant (except for
the asymmetry effects) and compatible with a stationary process. Next, we define the variables appearing in the second line of Equation (2.4.3). In conjunction
with hypothesis tests (discussed below), the coefficients of these variables allow
us to tackle the questions raised above.
5
Since we model SA returns, ht is the deseasonalized conditional variance.
An EGARCH (1,1) structure was not sufficient to clean the autocorrelation of the squared
standardized residuals.
6
2.4. Models and Empirical Results
31
Table 2.4: EGARCH model for impact of announcements and activity on volatility
(1)
(2)
(3)
qt = θ0 + ut + θ1 ut−1 + θ2 ut−2
√
ut = hth²t
³
i
´
p
P
ln ht = ω + 2i=1 βi ln ht−i + αi | ²t−i | − 2/π + γi ²t−i
P
P
+ 9j=1 3τ =1 ηj,τ dj,τ,t + φ ast−1 + δamt−1
Coefficient
θ0
θ2
α1
α2
γ1
γ2
β1
β2
η1,1
η1,2
η1,3
η2,1
η2,2
η2,3
η3,1
η3,2
η3,3
η4,1
η4,2
η4,3
Obs.
j
Q(j)
Q2 (j)
∗∗
Estimation
0.003
-0.028∗ ∗
0.305∗ ∗
-0.221∗ ∗
-0.003
0.006
1.264∗ ∗
-0.295∗ ∗
0.100∗
0.049
-0.083
0.212∗ ∗
-0.039
-0.007
0.105∗ ∗
-0.091∗
0.000
0.111∗ ∗
-0.171∗ ∗
0.133∗ ∗
37 650
1
1.17
2.88
P-Value
(%)
49
0.0
0.0
0.0
55
31
0.0
0.0
2.0
49
15
0.0
64
91
0.0
1.3
99
0.1
0.1
0.0
2
1.76
5.49
Coefficient
θ1
ω
η5,1
η5,2
η5,3
η6,1
η6,2
η6,3
η7,1
η7,2
η7,3
η8,1
η8,2
η8,3
η9,1
η9,2
η9,3
φ
δ
12
10.29
11.76
Estimation
∗∗
-0.134
-0.072∗ ∗
0.167∗ ∗
0.024
-0.024
-0.074
0.003
0.032
0.081∗
-0.106
0.096
0.484∗ ∗
-0.565∗
0.101
0.095
0.091
-0.013
0.005∗ ∗
0.000
W (ηj,τ = 0)
24
19.64
30.22
P-Value
(%)
0.0
0.0
0.0
73
70
21
97
63
1.2
8.9
9.0
0.0
2.4
61
7.4
22
84
0.0
11
114∗ ∗
and ∗ indicate respectively significance at 1% and 5%.
W (ηj,τ = 0) is the Wald statistic for the hypothesis of nullity of the 27 coefficients
ηj,τ .
Q(j) and Q2 (j) are the Ljung-Box statistics of order j respectively for standardized
residuals and for their square. Twelve lags correspond to 1 hour.
Variables in Equations (1)-(3): qt is the SA return (multiplied by 10,000); dj,τ,t is
a dummy variable for the event j announced during period τ relative to time t; ast
is the SA market activity, and amt the market activity seasonal index.
Estimation was done by the quasi maximum likelihood method.
32
Chapter2. News Announcements, Market Activity and Volatility
The variable denoted dj,τ,t is a dummy variable for the existence of an an-
nouncement of category j during the period τ , relative to the 5-minute return at
time t. It is equal to 1 if there is a news announcement during the time interval
τ and is equal to 0 otherwise. The index τ indicates an observation window: a
pre-announcement period (τ = 1), a period just after the announcement (τ = 2),
and a post-announcement period (τ = 3). The observation windows are equal
to 15 minutes before the announcement (τ = 1), five minutes just after the announcement (τ = 2), and 20 minutes after the announcement (τ = 3). Said
differently, we allow one lead and two lags in the effect of the dummy variable for
the event j. Since there are nine categories of events (see Table 2.2), there are 27
ηj,τ coefficients. The impact of the announcement dummy variables is significant,
since the Wald test for the null hypothesis that the 27 coefficients are equal to 0
delivers a P-value smaller than 1% (the Wald statistic is equal to 114).
Now we can answer questions 1, 2, and 3 regarding the news announcements
(see the list in Table 2.2). To answer question 1 for category j, we simply test the
hypothesis that ηj,1 = 0 against ηj,1 > 0 (since we presume a positive impact).
Rejection implies that volatility increases before the announcement. To answer
P
the second question, we test if 3τ =1 ηj,τ = 0 against the hypothesis that it is
greater than 0. Rejection implies that after an announcement volatility increases
(if the sum is positive) before reverting slowly to its initial level. Reversion to the
initial level is implied by the stationarity of the model: the persistence degree is
determined by the sum of the coefficients β1 and β2 estimated at 0.97. Table 2.5
presents the results of the tests.
2.4. Models and Empirical Results
33
Table 2.5: Pre-announcement and total impact of announcements
Announcement category (j)
η̂j,1
Scheduled:
1-Positive US macro figures
0.10∗
2-Negative US macro figures
0.21∗ ∗
3-European macro figures
0.11∗ ∗
4-Speeches of senior officials
0.11∗ ∗
5-Interest rate reports
0.17∗ ∗
Unscheduled:
6-Economic institutes forecasts
-0.07
7-OPEC member declarations
0.08∗
8-Central bank intervention rumors 0.48∗ ∗
9-Extraordinary events
0.10
P3
τ =1
η̂j,τ
0.07
0.17∗∗
0.01
0.07∗∗
0.17∗∗
-0.04
0.07
0.02
0.17∗∗
The second column gives the pre-announcement impact, the last
column the total impact. Estimates are taken in Table 2.4.
∗∗
and ∗ indicate respectively significance at 1% and 5%, for
one-tail tests.
Clearly, the answer to question 1 is positive for scheduled announcements:
these increase volatility during the pre-announcement period. Volatility increases
by 10 percent prior to the release of US positive and European macroeconomic
figures, and official speeches. It increases by around 20 percent prior to negative
US macroeconomic figures and interest rate reports. A possible explanation of
such volatility increases is that they are caused by anticipatory trades, i.e. by
traders who open positions hoping that their anticipations will coincide with
the contents of the news. However, another category of traders may also take
part in the trading. These traders, who are characterized by a high level of risk
aversion, prefer to execute their clients’ orders right before the news release to
avoid possible reversals of trends in the currency rate (Lyons, 1991).
The answer to question 1 is also positive for two types of unscheduled announcements: declarations of OPEC members (+8 percent), and especially rumors of central bank interventions (+48 percent!). These increases are probably
not due to anticipatory trades. For example, before being broadcasted by a specialized news agency, rumors of an intervention circulate for a certain period of
time from one dealer to another until they become widely disseminated. Then,
specialized agencies treat them seriously and announce them. It is during this
circulation phase that the market reacts to the news, through price adjustments.
34
Chapter2. News Announcements, Market Activity and Volatility
The answer to question 2 depends on the category of the announcements.
A significant total impact on volatility, of the order of 17 percent, is found for
interest rate reports and US negative announcements. For speeches of senior
officials, an impact of 7 percent is found on average. Unscheduled announcements
have a zero total impact, with the exception of extraordinary events. Thus,
for example, concerning rumors of central bank interventions, once the rumor
is refuted or confirmed, volatility drops immediately (η8,2 , estimated at -0.57,
significant at 5%, annihilates the pre-announcement positive effect estimated at
0.48).
Thus, with respect to question 3, there is no fully systematic difference between scheduled and unscheduled announcement effects on volatility (whether
before the announcement, or globally). Rather, we can conclude that regarding the pre-announcement period, almost all types of news increase volatility,
whereas regarding the total reaction (pre + during + post), only three categories
of scheduled news increase volatility.
A question of related interest, regarding the news announcements, is whether
there is a difference in the revealed volatility reaction between positive and negative US macroeconomic announcements. There is no strong evidence in favor
of a difference: for the pre-announcement period, the P-value of the test of the
hypothesis η1,1 = η2,1 is equal to 7 percent, and for the total effect, the P-value
P
P
for 3τ =1 η1,τ = 3τ =1 η2,τ is at 5.8 percent.
The variable ast in the EGARCH equation is the proxy for unexpected market activity. Market activity would be ideally measured by the flow of orders
between traders and their customers. This is not usually available on FX markets. However, market activity is usually proxied by the number of issued quotes.
Danielsson and Payne (2002) show that quote data tend to reflect the actual pace
of market activity. From the series of the number of quotes per five-minute interval of our database, we construct a series of average market activity for each day
of the week (like we did to build the average volatility of each day, see Section
2.4.2). This yields the variable amt (average market activity activity at point t
of the week). The variable ast is obtained by dividing the number of quotes by
the corresponding average.
2.4. Models and Empirical Results
35
With these two variables in the model, we can answer question 4 stated at
the start of this section, i.e. we can assess the impact of private information as
proxied by unexpected market activity. If φ is positive and δ is equal to zero, an
increase in SA volatility can only be due to unexpected trading activity, which
basically corresponds to the flow of orders (private information) between dealers
and their customers. The results in Table 2.4 show that φ is positive and highly
significant (with a P-value smaller than 1%) but that δ can be considered not to
differ from 0 (P-value = 11%). Thus, volatility rises in periods of unusually high
activity. This increase is fed by information asymmetry between traders. Each
trader benefits from his privileged information that originates from the order flow
with his own customers (Lyons, 1997).
2.4.2
Impact of Announcements on Intradaily Average
Volatility
In this section, we want to test whether news announcements have any impact
(altogether and individually) on the average volatility of the days of the week.
Therefore, we create a series (called mvt ) by stacking the intradaily volatility
series of each day of the week, from Monday to Friday, making up a total of
1,415 observations as explained below. We regress mvt on 9 variables zt,j (j
ranges from 1 to 9) corresponding to the nine categories of news announcements
we have defined. The variable zt,j is the number of news announcements belonging
to category j, which occurred during the five-minute time interval indexed by t.
For example, z1,1 is the number of news announcements corresponding to positive
US macroeconomic figures released on Monday between 00h05 and 00h10 over
the six-month sample of our database.
We control for autocorrelation in the regression by use of an ARMA(1,1)
model. Seasonal effects that originate from the structural characteristics of the
FX market (opening times, lunch breaks, closing times,. . . ) are controlled for
by a FFF of order 4 (this number was chosen according to the Akaike criterion).
36
Chapter2. News Announcements, Market Activity and Volatility
The model used for estimation is
mvt = c0 + β mvt−1 +
4
X
(δc,p cos xt,p + δs,p sin xt,p )
(2.4.4)
p=1
+
9
X
ηj zt,j + α ²t−1 + ²t ,
t = 1, . . . , 1415.
j=1
The variable xt,p in the FFF is defined by:
xt,p =
2πp nk
for nk = 1, 2, . . . , Nk , and k = 1, 2, . . . , 5,
Nk
(2.4.5)
where Nk is the number of time intervals per day (287 on Monday, 264 on Friday,
and 288 the other days). Therefore, there is a one-to-one correspondence between
each value of the index t of xt,p (and of mvt ) and a given five-minute interval
endpoint in a specific day of the week. The link is made via the relationship
t = f (1, k, nk ) (see Equation (2.4.7)).
To compute the intradaily average volatility at time nk of day k (called mvnk ),
we divide each day into 288 five-minute intervals. We assume for simplicity that
we have exactly S weeks of data. For each interval endpoint per day of the week
over the S week period, we have one euro/dollar return. We thus compute in
principle 288 values mvnk for each day of the week. Actually, as explained in
Section 2.3.1, we delete the first interval of Monday and the intervals from 22h05
to 24h of Friday. Hence, we have 287 points on Monday and 264 on Friday. That
makes a total of 1415 values over a week.
Each value mvnk is the square root of the average of the S squared returns
at time nk of day k (k = 1 is for Monday, k = 5 for Friday). For example, the
value of mvnk on Tuesday at 12h (k = 4 and n4 = 144) is the square root of the
average of the squared returns observed every Tuesday at 12h during the S week
period. Formally,
Ã
mvnk =
S
1X 2
r
S s=1 f (s,k,nk )
where
f (s, k, nk ) = 1415 (s − 1) +
!0.5
k−1
X
,
Nj + n k ,
(2.4.6)
(2.4.7)
j=1
for s = 1, . . . , S, k = 1, . . . , 5, N1 = 287, N2 = N3 = N4 = 288, n1 = 2, . . . , 288,
2.4. Models and Empirical Results
37
n2 = 1, . . . , 288, n3 and n4 likewise, and n5 = 1, . . . , 264 as stated above. Notice
that when s varies from 1 to S = 26 for example, the function f (s, k, nk ) takes
the values from 1 to 36,790. Actually, in our database, we have 37,654 price
observations (hence one return less), corresponding to 26 full weeks starting a
Tuesday and three more days of the twenty-seventh week. In Figure 2.1, each
day-specific panel displays the temporal profile of mvnk for that day (the lower
line). The last panel displays the temporal profile of average volatility when the
days are considered as identical, e.g. the average volatility at 12h is the same
for all days. For this, the average of squared returns is taken over all returns
corresponding to 12h in the sample (e.g. for 26 weeks, there are 5 × 26 squared
returns). That overall profile is also displayed in the day-specific panels, after
shifting it upwards to make it visible.
To adjust the returns for seasonality, we just divide the return at the endpoint
of each five minute interval by the corresponding value of the intradaily average
volatility (using the day-specific volatility). That means, for example, that all
returns at 12h on Thursday in the sample are divided by the same value (the
average volatility at 12h on Thursday).
38
Chapter2. News Announcements, Market Activity and Volatility
Table 2.6: Impact of announcements on intradaily average volatility
P
mvt = c0 + β mvt−1 + 4p=1 (δc,p cos xt,p + δs,p sin xt,p )
P9
+ j=1 ηj zt,j + α²t−1 + ²t
(4)
Coefficient Estimate P-Value Coefficient Estimate
P-Value
(%)
(%)
∗∗
c0
18.39
0.0
β
0.85∗ ∗
0.0
∗∗
α
-0.69
0.0
η1
7.59∗ ∗
0.0
∗∗
∗∗
δc,1
-10.97
0.0
η2
4.41
0.1
δc,2
-1.17
30
η3
1.30∗ ∗
0.9
∗∗
∗
δc,3
4.89
0.0
η4
-2.68
1.3
δc,4
-2.67∗ ∗
0.9
η5
3.36∗ ∗
0.0
∗∗
-3.75
η6
-1.66
30
δs,1
0.0
∗∗
δs,2
2.97
0.7
η7
-0.54
73
δs,3
-2.71∗
1.2
η8
-1.07
80
∗
∗
δs,4
2.08
4.4
η9
3.11
4.0
2
Obs.
1 415
R
0.39
W (ηj = 0)
321∗ ∗
j
1
2
12
24
Q(j)
0.88
1.04
22.75
42.85
∗∗
and ∗ indicate respectively significance at 1% and 5%.
W (ηj = 0) is the Wald statistic for the null hypothesis of nullity of the 9 coefficients
ηj .
Q(j) is the Ljung-Box statistic of order j of the residuals.
Variables in Equation (4): mvt is the intradaily average volatility for each time
interval t (see Section 2.4.2); zt,j is the number of announcements for the news
category j during five minutes before t; xt,p = 2πp nk /Nk where k = 1 . . . 5, Nk is
the daily time interval number (287 for Monday, 264 for Friday, 288 for the other
days), and nk = 1 . . . Nk . Estimation was done by the quasi maximum likelihood
method.
Table 2.6 gives the estimation results of Equation (2.4.4). By a Wald test,
we strongly reject the null hypothesis of joint nullity of the nine coefficients of
the announcement counts (the P-value is much lower than 1%, the statistic being
equal to 321). Clearly, announcements have an impact on the average volatility.
More precisely, the variables for the scheduled news announcements, in particular
US macroeconomic figures, European figures, speeches of senior officials of the
government, and interest rate reports have significant coefficients quasi at the
1% level. Thus these events have a seasonal impact on volatility. These results
confirm those of Andersen and Bollerslev (1998) with respect to the seasonal
impact of scheduled events on volatility. FFF coefficients are, with one exception,
significant at 5%. These coefficients take into account the seasonal effects not
due to the cyclical news releases.
2.5. Conclusion
2.5
39
Conclusion
Using 5-minute high frequency data for the May 15 - November 14, 2001 time
period combined with a news database, we shed light on the impact of news
announcements and private information on the volatility of euro/dollar FX returns. Regarding the news releases, we highlight the impact of both scheduled
and unscheduled news announcements and focus on the reaction of volatility in
the pre-announcement periods. We also take into account FX private information, proxied by the level of deseasonalized market activity. Our results show that
the release of scheduled news leads to a pre-announcement rise in volatility. This
increase in volatility does not occur in the case of announcements of unscheduled news, except for rumors of central bank intervention. Our interpretation of
volatility increases right before the announcement of scheduled news is that these
categories of news attract traders who wish to make anticipatory trades based on
their personal beliefs. Another result drawn from estimation is that the reaction
of volatility in the post-announcement period is in most cases muted. Indeed,
most of the news announcements considered in our study have not been followed
by significant volatility increases or decreases in the post-announcement period.
In addition, the volatility of euro/dollar returns is positively and significantly affected by market activity (adjusted from its seasonal component). This stresses
the importance of private information in FX markets as a key determinant of
volatility, the private information stemming primarily from the flow of orders
between traders and their clients.
Chapter 3
The Information Content of
Quoting Activity
This chapter is based on joint work with Andréas Heinen.
3.1
Introduction
In the previous chapter we considered the quoting activity as an exogenous variable that has an effect on volatility. In this chapter, we deepen our investigation
by studying the information content of quoting activity.
Unlike stock markets, foreign exchange markets are characterized by a very
low degree of transparency. The quantities exchanged or even whether a transaction took place is only known by the transacting parties, but is not known
by other market participants. Every dealer observes only his own trades. The
only systematic source of information on foreign exchange markets are electronic
screens, such as Reuters, Telerate and EBS (i.e. electronic brokerage system),
which transmit information about the prices at which the main participants, that
is large international banks, are willing to buy or sell currencies. Quotes displayed
on Reuters or Telerate are firm for a short period of time and for a relatively small
volumes. They are considered as indicative, since dealers often get a quote with
small differences from the one displayed. The difference is justified by the period
of time between visualization of the quote and the actual trade. However, this
is not the case in EBS since it is both a quotation and a dealing screen. Quotes
conveyed by EBS are anonymous and firm until executed or canceled and therefore considered as transaction prices. Although information displayed on EBS
involves the bid, the ask and the amount of the limit order corresponding to each
of the prices, it does not provide the identity of the dealer who submits the limit
41
42
Chapter3. The Information Content of Quoting Activity
order. Reuters, however, displays the prices and the identity of the dealer who
submits the quotes. Therefore, Reuters is much more transparent than EBS in
terms of identifying the quoting dealer. Furthermore, being competitors, Reuters
and EBS display at any given time almost the same quotes.
Whenever a bank revises a spot quote of a currency, the latest quote is flashed
instantaneously up on the main screen. There are other sources of information
available to dealers besides electronic screens. Some dealers choose to conduct
their transactions through brokers or directly via telephone. Others may prefer
not to exhibit quotes on any screen even though they are active on the foreign
exchange market (Goodhart and Demos, 1991). Nevertheless electronic screens
are the only systematic source of information that each dealer has on the other
dealers’ quotes. Information about brokered trades for instance, will typically
only consist of net volume and a price, which is not fully informative.
In this chapter we use the multivariate autoregressive double Poisson model
of Heinen and Rengifo (2003), which is a count model, to analyze the quoting
activity of several large banks jointly as a system and to evaluate the effect of
news announcements on quoting activity from a disaggregated point of view. By
disaggregated we mean that the effect of each announcement on every bank is
allowed to be different. This is new, as thus far, to the best of our knowledge,
there has been no work on the response of individual banks’ quoting activity to
news. Our framework allows us to identify the announcements that matter for
quoting activity, but also to see whether all banks react in the same way to the
same news. Finally we can analyze interactions between different banks’ quoting
activity.
By looking at a sample of major dealers on the euro/dollar exchange rate,
and by using the multivariate double autoregressive conditional Poisson model,
we offer prima facie evidence of the fact that dealers’ quoting activity reacts significantly to some events but differently to the same news announcements. In
particular, for some types of announcements, some banks increase their quoting
activity, whilst others decrease it or keep it unchanged. This can lead to an
ambiguous effect at the aggregate level implying that aggregate studies tend to
underestimate (or overestimate) the importance of public news announcements
for quoting activity. In addition, using quoting activity as a proxy for mar-
3.2. The Quoting Activity Signal
43
ket activity like in Degennaro and Shrieves (1997), Melvin and Yin (2000) and
Bauwens, Ben Omrane, and Giot (2005), allows us to classify news announcements according to Evans’ taxonomy (Evans, 2002). We find that scheduled
news are non-common knowledge (NCK) news whereas unscheduled news seem
to belong to the category of common knowledge (CK) news.
Moreover we find that banks quoting activity is typically affected by the quoting activity of some other dealers. This means that follower dealers observe the
frequency of price revision of leader dealers in order to infer some information.
This supports the view that dealers try to infer private information or the interpretation of the news content by other dealers.
The chapter is structured as follows. In Section 3.2 we present the literature
about news announcement and inter-dealer effects on quoting activity, in Section
3.3 the data, in Section 3.4 the models and results, and in the last section we
conclude. We present, moreover, some technical details about the model within
an appendix at the end of the chapter.
3.2
The Quoting Activity Signal
During the trading period, FX dealers get real time information through electronic screens about public news announcements as well as other banks’ quoting
activity. The latter constitutes the only information that a bank has about other
banks’ quotes. It is a well documented fact that dealers revise their quotes in order to reflect their reaction to public news announcements (Andersen, Bollerslev,
Diebold, and Vega, 2002) or their detention of private information coming from
their customers’ order flow (Lyons, 1995, Evans and Lyons, 2002, Cao, Evans,
and Lyons, 2003). In addition Degennaro and Shrieves (1997) and Bauwens,
Ben Omrane, and Giot (2005) show that quoting activity (assimilated to the activity of price revision) increases return volatility. They also use quoting activity
adjusted for its seasonal component as a proxy for private information occurring
from customers’ order flow.
Given these facts, we argue that dealers can try to infer other dealers’ private
information or reaction to news announcements through their quoting activity.
Hence, dealers could use two information channels to build their reaction to
news events. The first channel takes place directly through the news broadcast-
44
Chapter3. The Information Content of Quoting Activity
ers. This corresponds, in a sense, to dealers’ spontaneous or direct reactions to
the announcement. The second channel works through the quoting activity of
the other dealers. It could be a source of noise to the first one because dealers
spontaneous reactions to news announcements could be influenced by bank leaders reactions to the same events. In the remainder of this section we formulate
the questions that we investigate and review some of the related literature. In
the first subsection we deal with the link between quoting activity and news announcements, in the second we explain how we can identify which announcements
are common knowledge. In the third we provide a brief summary of a relatively
new literature which has been concerned with the analysis of individual banks
and we talk about inter-dealer interaction.
3.2.1
News Announcements and Quoting Activity
News announcements and quoting activity were analyzed in several studies. Degennaro and Shrieves (1997) use three categories of news announcements (scheduled and unscheduled macroeconomic news announcements as well as interest
rate reports) and six different periods around the event and analyze their impact on quoting activity. They find a significant effect of all three categories
of news, but at different times relative to the announcement. Melvin and Yin
(2000) work with a sample of US Dollar/Japanese Yen and US Dollar/Deutche
Mark data from December 1993 to April 1995 in hourly data. They take as
news variable the number of news events that happen within an hour and do not
make any distinction between different categories of news. They find a significant
impact of the number of news on quoting activity, working with deseasonalised
variables, and conclude that quoting activity is not self-generating. Evans and
Lyons (2003) identify two channels of transmission of macro news to exchange
rates: a direct effect and an indirect effect via order flow. The news variable is
the number of news announcements that occur within the period. Identification
of the various effects is done by the imposition of orthogonality conditions on the
various innovation terms in the model and estimation is carried out using the
generalized method of moments (GMM). Changes in mid-quotes are regressed on
order flow with two error terms, one with a constant variance, which represents
information directly impounded into prices, another whose variance depends on
3.2. The Quoting Activity Signal
45
the number of information events and represents the common knowledge effect of
macro news on the exchange rate. The order flow is also the sum of two shocks,
one of them having its variance depending on news. This shock is interpreted as
the indirect effect of news on exchange rates via induced order flow. In order to
justify that macroeconomic news affect the order flow, Evans and Lyons (2003)
mention differences in interpretation of the news or differences in opinion as to
the impact of the news on the exchange rate. Several studies have taken the
number of bank quotes as a proxy for the number of transactions, which is tantamount to assuming that a fixed proportion of posted quotes correspond to actual
trades. This assumption has been made, amongst others, by Goodhart and Figliuli (1991) and Bollerslev and Domowitz (1993), who prefer to use quote arrival
as a proxy for market activity, than transaction volume, because quotes signal a
willingness to trade. Degennaro and Shrieves (1997) use the same assumption, as
they consider the seasonal and stochastic parts of quoting activity to be a proxy
for the expected and surprise components of market activity. Furthermore, their
results are suggestive of the fact that the surprise part of market activity reflects
informed trading. Melvin and Yin (2000) have made the same assumption.
In this chapter we analyze the reaction of individual banks to a series of news
announcement. Thus far there has been to the best of our knowledge, no work
on the response of individual banks’ quoting activity to news. Degennaro and
Shrieves (1997) regress quoting activity on news and find a significant impact
of certain types of news announcements. In our analysis we allow for different
responses of individual banks to the same news and we compare the results to
those at the aggregate level. We find that the banks’ reaction to news is completely heterogeneous. There are differences across banks not only in terms of
which news they react to, but also in terms of how they react to the announcements. Some dealers increase their quoting activity whilst others decrease it or
keep it unchanged as a response to the same news. This suggests that banks act
differently, either because they have different private information or because they
interpret the public news announcements differently in terms of their implications
for the exchange rate.
46
Chapter3. The Information Content of Quoting Activity
3.2.2
What announcements are common knowledge?
Evans (2002) distinguishes two types of news: common knowledge (CK) and noncommon knowledge (NCK) news. Common Knowledge (CK) news is available
simultaneously to all market participants and is interpreted in the same way.
News that is not known by everybody at the same time or for which interpretations are different are termed non-common knowledge news (NCK). He considers,
instead of an equilibrium price, an equilibrium price distribution. He justifies this
by the lack of transparency of currency markets, which makes it possible for several transactions to happen simultaneously at different prices. This can also be
understood, if one considers that different dealers have different interpretations
of the events that influence the exchange rate. His result suggests that CK news
are not the predominant source of long term movements in the exchange rate.
In the empirical part, based on prices and order flow, CK and NCK shocks are
identified by the assumption that CK news leads to an immediate change in the
mean of the equilibrium price and have no effect on order flow, whereas NCK
news has an impact both on prices and order flow, which may take time.
Using Evan’s taxonomy and following the literature which takes quoting activity as a proxy for order flow, allows us to classify different categories of news
announcements according to whether they impact quoting activity or not. If they
do not, this means that they can be considered as CK news events, whereas public announcements, that have quoting activity implications, do so maybe because
of heterogeneous interpretations by dealers. Indeed some banks might have different degrees of understanding of the same news, which can lead them to act on
their anticipations or to stay away from the market, waiting for better-informed
banks to act first.
3.2.3
Inter-dealer Interaction
A relatively new literature has been concerned with the analysis of individual
banks. In this strand of the literature papers deal mainly with the identification of price leaders in the market around central bank interventions, but also in
normal trading. Peiers (1997) analyzes the mid-quotes of several banks on the
Dollar/Mark exchange rate around the European Central Bank interventions using a vector autoregression (VAR) model and Granger causality tests to identify
3.2. The Quoting Activity Signal
47
the price leading bank. The sample of banks includes Deutsche Bank, Société
Générale, Chemical Bank, Rabobank, Den Norske and BHF Bank. Deutsche
Bank is the first to react, 60 minutes prior to the announcement, followed by
other banks, 25 minutes before the announcement. Wang (2001) and Sapp (2002)
instead use cointegration analysis. They focus on a small subset of banks and
analyze their mid-quotes with a cointegrated VAR model. The midquotes of all
the banks are integrated of order one and they cannot deviate in the long run,
which means that they are cointegrated. The number of cointegrating relationships is equal to the number of banks minus one, which means that there is only
one stochastic trend driving the system, which can therefore be interpreted as the
fundamental market price. Wang (2001) analyzes price leadership amongst three
leading New York-based dealers on the US Dollar/Deutche Mark market: J.P.
Morgan, Chemical Bank and Citibank. Sapp (2002) works on the same market
and estimates a cointegrated VAR system and deduces measures of information
shares, for all the trading period as well as around central bank interventions.
This is used to identify the banks whose information share is largest around
central bank interventions.
In the empirical part of this chapter we use a vector autoregressive (VAR)
type model to analyze the interactions between individual dealers’ quoting activity. We establish that there are significant inter-dealer effects. This means that
there are systematic lead-lag relations in the intensity of quote revisions of the
various banks. We offer evidence suggestive of the fact that dealers observe the
quoting activity of others to infer useful information like other dealers’ private
information or their reaction to public news announcements. If a dealer increases
(decreases) his quoting activity following the publication of news this means that
he increases (reduces) his price revision. Banks typically react to some news
announcements and not to others. They also respond to the quoting activity of
certain banks. We interpret this as meaning that when banks are unsure about
the interpretation of certain news announcements, they tend to wait a little and
watch other banks that they might perceive as being better informed in order
to infer their interpretation of the news content. This means that a public news
announcement could have no effect on a given bank’s quoting activity either because it considers that news as irrelevant or it prefers to wait and see how better
48
Chapter3. The Information Content of Quoting Activity
informed dealers react. In that sense we argue that quoting activity is a potentially useful information channel in an otherwise very opaque foreign exchange
market.
3.3
Data and Descriptive Statistics
We use the same data as in Chapter 2. We work with two samples of banks
corresponding to two periods. Because the data comes from different quoting
systems. From May 14 until September 10, 2001, the data comes from Reuters,
and from August 24 until October 26, 2001, it comes from Tenfore System.
We take into consideration two electronic screens to eliminate the quoting activity bias shown by Goodhart and Demos (1991) (i.e. all dealers do not conduct
their quoting through only one electronic screen, but they choose different ones).
We therefore work with two samples of banks during different periods. We selected the most active banks in our sample. Tables 3.1 and 3.2 show that for the
first sample, the four banks we select cover about 24.4% of the overall quotes,
whereas the six banks of the second sample post about 45% of the total number
of quotes in the sample. The reason why we focus attention on these banks is
that they are the most active dealers in our data and the remaining quotes are
posted by a very large number of dealers with a very small contribution. Our
first sample of banks contains BG Bank, Copenhagen (BGFX), Berliner Handelsund Frankfurter Bank, Frankfurt (BHFX), Rabobank, London (RABO), Société
Générale, Paris (SGOX) for the period May 14 to September 10, 2001. This
corresponds to 9396 5-minute observations. The second sample of banks includes
Barclay’s Bank, London (BARL), Dresdner Bank, Frankfurt (DREF), Oolder &
de Jong, Amsterdam (OHVA), Oko Bank, Helsinki (OKOH), SHK Bank, Hong
Kong (SHKH) and Union Bank of Switzerland, Zurich (UBSZ), for the period
August 24 to October 26, 2001, for a total of 4968 5-minute quoting intervals.
Descriptive statistics for the first sample are shown in Table 3.1 and in Table
3.2 for the second sample. The minimum number of quotes is zero and the mean
is generally quite small, which justifies the use of discrete distributions like the
Poisson. Moreover, most series are over-dispersed (meaning that the variance is
larger than the mean), with the exception of BHFX, RABO and OHVA, which are
3.3. Data and Descriptive Statistics
49
Table 3.1: Descriptive statistics of the number of quotes per 5-minute interval of
the first sample of banks for the period May 14 to September 10, 2001
Mean Std. Dev. Dispersion
Maximum
Q(10)
BGFX
BHFX
RABO
SGOX
Rest
7.99
9.40
6.86
15.40
120.58
4.57
2.37
2.20
6.71
56.88
2.61
0.60
0.71
2.92
26.83
28 26404
17 2822.5
20 4608.3
46 14315
399 57444
Aggreg
162.25
64.03
25.27
472
50807
Table 3.2: Descriptive statistics of the number of quotes per 5-minute interval
for the second sample of banks for the period August 24 to October 26 2001
Mean Std. Dev. Dispersion
Maximum
Q(10)
BARL
DREF
OHVA
OKOH
SHKH
UBSZ
Rest
11.54
2.27
14.22
31.56
3.03
10.82
81.66
4.15
2.62
3.73
21.86
2.15
5.59
61.79
1.49
3.02
0.98
15.14
1.53
2.89
46.75
22 9571.9
20 8217.9
21 10809
82 36956
16 4228.3
44 8980.1
352 37913
Aggreg
163.2
78.81
38.06
486
34264
50
Chapter3. The Information Content of Quoting Activity
under-dispersed. This justifies the use of the double Poisson distribution, since,
unlike other count distributions, it allows for both over- and under-dispersion.
We use the same news announcements as in the previous chapter and we test the
impact of nine categories of news. News announcements, shown in Table 3.3 are
classified, as in the previous chapter, into two groups, scheduled and unscheduled
announcements.
As can be seen from Table 3.3, the total number of news announcements in
the first sample is 377, the most frequent type of news event is European macroeconomic figures with 105 events, but there are only 3 occurrences of rumors of
central bank interventions. In the second sample, there are 251 events, with 53
speeches of government officials and only 3 rumors of central bank intervention.
We compute averages of the quoting activity over 5-minute intervals for all
banks and divide them by their cross-sectional average in order to make them
comparable across banks. The seasonal patterns are shown in Figure 3.1. First
of all, we note that the seasonality of the banks in the sample is not the same
for all, which is not surprising. BGFX, SGOX, BARL, DREF, and UBSZ all
start with a small decrease in the morning until 10 AM GMT, and after that
quoting activity starts increasing from around 12 PM GMT, which corresponds
to the morning on the East Coast of the US, to a peak around 2 or 3 PM GMT,
and then the activity decreases until 5 PM GMT, when European offices start to
close. SHKH is somewhat different, as it starts the day with an increase, then
its pattern is almost stable. DREF is different from other banks, in that it starts
closing earlier, which means that its quoting activity decreases sharply shortly
before 4 PM GMT. A similar pattern is observed for other banks, but between
6 and 7 PM for most of them, which is why we chose to stop our sample at
5 PM. The remaining banks (BHFX, RABO, OHVA and OKOH) do not seem
to exhibit any particular seasonality over our sample period. This is confirmed
for RABO, OHVA and OKOH by the Wald test for joint significance of the
seasonality variables shown in Tables 3.6 and 3.7 (see Section 3.4.2 for more
details). Furthermore we note that DREF has a particular pattern of diurnal
seasonality, with very important spikes on or around the hour.
3.3. Data and Descriptive Statistics
51
Table 3.3: News categories
Scheduled news announcements
1 and 2-US macroeconomic figures
Employment report
ISM index(ex NAPM)
Whole sales
Gross domestic product (GDP)
Producer price index (PPI)
Retail sales
Housing starts
Consumer confidence index
Consumer price index (CPI)
Construction spending
Car sales
Business inventories
Housing completions
Import prices
Current account deficit
Non-farm productivity
Personal income
Real earnings
House sales
3-European macroeconomic figures
4-Speeches of senior officials of the
government and those of public agencies
5-US and European interest rate reports
Unscheduled news announcements
6-Forecasts made by economic institutes
7-Declarations of OPEC members
8-Rumors of Central Bank interventions
9-Extraordinary events
Total
Positive
Negative
sample1
sample2
η1
+
+
+
+
+
+
+
+
+
+
+
+
+
η3
η4
η2
+
+
+
+
+
+
-
98
54
105
78
51
53
η5
36
25
η6
η7
η8
η9
36
13
3
8
377
19
25
3
21
251
The events are the news headlines released on the Reuters money news-alerts.
For US macroeconomic figures, we separate positive and negative news-alerts by comparing the
expected and the announced numbers. If the actual numbers are larger than the expectations
for economic variables that contribute to economic growth, the announcements are classified
as positive (+). If the actual news release means more inflation or a forthcoming economic
slowdown, it is classified as a negative news announcement (-). The expected values are given
on Reuters screens a few days before the news announcements.
The employment report includes the unemployment figures.
ISM is the abbreviation for the Institute of Supply Management, ex NAPM, National Association of Purchasing Management. It is a monthly composite index and gives the earliest
indication of the health of the manufacturing sector.
The symbol ηj is the coefficient of the dummy variable dj in the equations reported in Tables
3.6 and 3.7.
52
Chapter3. The Information Content of Quoting Activity
Figure 3.1: Time-of-the-day effect
This figure presents the time-of-day effect of each bank of the two samples. The figure shows
the ratio of the 5-minute means over the day relative to the overall mean. BGFX, BHFX,
RABO and SGOX are observed from May 14 to September 10, 2001. BARL, DREF, OHVA,
OKOH, SHKH and UBSZ are observed from August 24 to October 26, 2001.
3.4. Models and Results
3.4
53
Models and Results
We turn to the empirical methodology used in this chapter. We model the quote
arrival of each bank in order to study its sensitivity to news announcements
and to check whether there are significant interbank effects. In order to do this
we use the Multivariate Double Autoregressive Conditional Poisson (MDACP)
model introduced by Heinen and Rengifo (2003). The first subsection draws on
that paper. For more details on the econometrics we refer the reader to that
paper. The second part of the section presents and discusses the results.
3.4.1
Modeling Quote Arrival
In the remainder of the chapter we work with the number of quotes of individual banks on the euro/dollar currency market. As the number of quotes for most
banks is a relatively small number, usual time series models, based on the normal
distribution, is not appropriate. Instead we work with time series models developed specifically for count data. In our sample, the data is not equidispersed,
but it is overdispersed for most series, and underdispersed for some. In order to
model the dispersion in a flexible way, we use the double Poisson distribution
of Efron (1986). The Multivariate Double Autoregressive Conditional Poisson
(MDACP) model assumes that the number of quotes of bank i in period t, Ni,t ,
follows a double Poisson distribution, conditionally on past information:
Ni,t |Ft−1 ∼ DP (µi,t , φi ) , ∀ i = 1, . . . , K.
(3.4.1)
where Ft−1 is the information set generated by the past of the series up to and
including time t − 1 and K is the number of banks. The parameters µi,t and φi
are respectively the mean and the coefficient of dispersion of the double Poisson.
Its conditional variance is equal to
2
V [Ni,t |Ft−1 ] = σi,t
=
µi,t
.
φi
(3.4.2)
See Appendix A.1, at the end of the chapter, for details about the double
Poisson distribution. The distribution is over- or underdispersed for values of
φi respectively less or greater than 1. When φi = 1 the distribution reduces
to the equidispersed Poisson. Whereas the double Poisson distribution leads to
54
Chapter3. The Information Content of Quoting Activity
overdispersion or underdispersion, the effect of autocorrelation is to increase this
dispersion unconditionally.
The vector µt = (µ1,t , . . . , µK,t )0 of conditional means is assumed to follow a
VARMA(1, 1) process
µt = ω + A Nt−1 + B µt−1 ,
(3.4.3)
where Nt = (N1t . . . Nkt ), A is a full rank matrix of coefficients capturing the
impact of the lagged inter-dealers’ quoting activity effects, and B is a diagonal
matrix of autoregressive coefficients of the own lagged conditional mean. More
explicitly, the equation for each mean µi,t is
µi,t = ωi +
K
X
αi,j Nj,t−1 + βi µi,t−1 .
(3.4.4)
j=1
We work with a (1, 1)-lag structure for the mean equation, as this is parsimonious and flexible enough.
We are interested in analyzing the impact of news announcements on individual banks’ quoting activity, allowing for diurnal seasonality. The news variables take the form of dummies for the presence of a certain announcement (dj,t ,
j = 1, . . . , 9.) as in Chapter 2. The seasonality is modeled using the flexible
Fourier form (FFF) introduced in the foreign exchange literature by Andersen
and Bollerslev (1998) at daily, half daily and hourly frequencies. We modify the
conditional mean in the following manner to include these exogenous regressors:
µ∗i,t
= µi,t exp
9
¡X
j=1
ηi,j dj,t +
X
p=1,2,12
(ψi,c,p cos
2πp Re[t, T ]
2πp Re[t, T ] ¢
+ ψi,s,p sin
) ,
T
T
where Re[t, T ] is the remainder of the integer division of t by T , the number of
periods in a trading session. The way we include the regressors separates the
autoregressive part from the effect of seasonality and news, and this functional
form guarantees the positivity of the conditional mean.
In order to estimate our model, we use a two-stage estimator as in Patton
(2002). In the first step we estimate parameters of the marginal models under
the assumption that conditionally on the past, the different series of individual
banks’ quoting activity are uncorrelated. This means that there is no contem-
3.4. Models and Results
55
poraneous correlation and that all the dependence between the series is assumed
to be captured by the conditional mean. Consequently we estimate our system
equation by equation by the maximum likelihood method. However, to capture
contemporaneous cross-correlation, we resort to a copula function in the second
step. A one-step estimation procedure which would estimate copula parameters
and parameters of the marginal models jointly is not numerically feasible due
to the very large number of parameters. As far as the choice of copula is concerned, we choose to work with the most easy one, which is the Gaussian copula.
It provides a very general way of introducing dependence among several series
with known marginals. It is noteworthy that the second step does not require
any optimization, as the maximum likelihood estimator of the multivariate normal copula covariance matrix is simply the sample variance-covariance matrix of
the vector of normal scores (see Appendix A.2 for details about copulas and the
likelihood function of our model).
In order to evaluate the quality of the model, we use tools developed in density
forecast evaluation by Diebold, Gunther, and Tay (1998). The main idea is to use
the cumulative distribution of the data under the estimated density and to check
whether this is uniformly distributed, as it should be according to the probability
integral transformation theorem (PITT) of Fisher (1932). The assumptions of
the theorem are that the density is continuous, which is violated in the case of
counts. We explain in the Appendix A.3, how we deal with this problem using
continued extensions of discrete variables.
We also test the standardized residuals
εi,t
Ni,t − µ∗i,t
Ni,t − µ∗i,t
=
= q
σi,t
µ∗i,t /φi
for autocorrelation, which would indicate a failure of the model to capture the
dynamics of the series, and for deviation of their variance from one, which would
indicate misspecification of the dispersion.
3.4.2
Results
Tables 3.4 and 3.5 show the copula correlation matrix which is capturing the
contemporaneous cross-correlation and the part of the lagged cross-correlation
between individual banks’ quoting activity, which does not go through the time-
56
Chapter3. The Information Content of Quoting Activity
varying mean. Cross-correlations vary between 0.04 and 0.48. This means that
contemporaneous effects between different banks depend on the importance of
the influence of some banks’ quoting activity on others.
Table 3.4: Correlation matrix of the q estimated by the MDACP model for the
first sample of banks for the period May 14 to September 10, 2001
BGFX
BHFX
RABO
SGOX
BGFX
BHFX
1.00
0.17
0.28
0.44
1.00
0.19
0.20
RABO SGOX
1.00
0.29
1.00
The table presents the correlation matrix q, based
on the probability integral transformation z, of the
continued count data under the marginal densities
estimated using the MDACP models by the two-step
procedure (see section 3.4.1). It shows the contemporaneous correlations of the aggregate system of
table 3.6.
Table 3.5: Correlation matrix of the q estimated by the MDACP model for the
first sample of banks for the period August 24 to October 26, 2001
BARL DREF
BARL
DREF
OHVA
OKOH
SHKH
UBSZ
1.00
0.35
0.21
0.18
0.07
0.48
1.00
0.15
0.14
0.04
0.39
OHVA OKOH SHKH
1.00
0.15
0.06
0.18
1.00
0.04
0.18
1.00
0.09
UBSZ
1.00
The table presents the correlation matrix q, based on the probability integral transformation z, of the continued count data under the marginal
densities estimated using the MDACP models by the two-step procedure
(see section 3.4.1). It shows the contemporaneous correlations of the aggregate system of table 3.7.
Estimation results for the conditional mean parameters are presented in Tables 3.6 and 3.7. There is evidence of diurnal seasonality in the activity of all
banks except three (see the end of Section 3.3). For the other banks, the three
pairs of trigonometric function at the daily, half-daily and hourly frequency are
jointly significant at least at the 5% level. The effect of news announcements is
3.4. Models and Results
57
generally significant (at 5%) for all banks, as can be seen from the Wald tests of
the joint significance of all announcements. What the dummy variables results of
individual banks show clearly, is that their reaction to the same news announcements are different. There is variation across banks, both in whether or not they
react to a certain category of news and in the way they react to it, by increasing
or decreasing their activity.
The use of the double Poisson distribution is justified by the fact that we have
estimated both overdispersed distributions (the majority of them) and some underdispersed distributions. The variance of the standardized residuals is within a
few percent of one for nearly all banks, except for OKOH, which means that the
dispersion is relatively well captured. Upon closer inspection of its time series,
we can see that there seems to be a change of regime in OKOH, which went from
heavy quoting to lower levels of activity after October 8, 2001. The autocorrelations of the standardized residuals of BARL and UBSZ, shown as representative
example in Figure 3.2, are often in the 95% confidence band. However, some autocorrelations are a little outside of the bands, resulting in significant Q-statistics
(the sample size is large). Nevertheless, the Q-statistics are very strongly reduced,
compared to the raw data, even though they are still significant. Another way of
testing the specification is to look at the density forecast tools. The probability
integral transformation Z (PIT) of the data under the estimated distributions
should be uncorrelated and uniformly distributed. Figure 3.3 shows the quantile plot of Z, which is very close to the 45-degree line for six banks, shown as
examples.
For both samples of banks, it seems that at least one type of scheduled news
event has an impact on every bank, except for OHVA. Positive and negative
surprises in U.S. and European figures (respectively η1 , η2 and η3 ) seem to have
the most important effects. This is in line with the findings of Andersen, Bollerslev, Diebold, and Vega (2002), that macroeconomic surprises have the most
significant impact on the level of the exchange rate. According to Evans (2002),
these types of announcements are therefore NCK news, as they impact order
flows. Given that they are simultaneously received by all dealers, it has to be
the case that they are interpreted differently. A lot of banks react to the first
three scheduled news announcements. In particular SGOX and DREF increase
58
Chapter3. The Information Content of Quoting Activity
Table 3.6: Estimation results of MDACP models for sample 1, May 14 to
September 10, 2001.
Parameters
η1
η2
η3
η4
η5
η6
η7
η8
η9
α
αBGF X
αBHF X
αRABO
αSGOX
ω
β
φ
W (η = 0)
W (ψ = 0)
W (αi = 0)
V ar(εt )
Q(10)
LL
BGFX
0.091
0.091
0.122∗ ∗
0.014
0.114
0.133
-0.186
-0.439
0.176
BHFX
-0.064∗
-0.060
-0.031
-0.005
-0.029
0.026
-0.060
0.251
0.027
RABO
-0.105∗
-0.088∗
0.050
-0.033
-0.017
0.056
0.004
-0.308
0.122
SGOX
0.193∗ ∗
0.185∗ ∗
0.135∗ ∗
0.025
0.048
0.030
0.003
-0.233
0.158
0.203∗ ∗
-0.001
-0.004
-0.007∗ ∗
0.199∗ ∗
0.789∗ ∗
0.687∗ ∗
26.18∗ ∗
60.70∗ ∗
44.92∗ ∗
0.96
111.45∗ ∗
-23795.62
-0.001
0.137∗ ∗
0.018∗ ∗
0.007∗ ∗
0.272∗ ∗
0.809∗ ∗
1.739∗ ∗
12.14
74.91∗ ∗
63.83∗ ∗
0.95
19.95∗
-21108.37
0.0008
0.030∗ ∗
0.132∗ ∗
0.009∗ ∗
0.071∗ ∗
0.796∗ ∗
1.490∗ ∗
21.04∗
10.62
176.3∗ ∗
0.90
52.08∗ ∗
-20223.68
0.025∗ ∗
0.078∗ ∗
0.038∗
0.271∗ ∗
0.616∗ ∗
0.608∗ ∗
0.504∗ ∗
59.09∗ ∗
90.88∗ ∗
89.40∗ ∗
0.97
74.39∗ ∗
-20291.97
Rest
0.039
0.093∗ ∗
0.125∗ ∗
0.017
-0.003
0.050
0.011
-0.089
0.045
0.626∗ ∗
- 0.303∗ ∗
-0.194∗
-0.456∗ ∗
0.062
13.76∗ ∗
0.347∗ ∗
0.201∗ ∗
155.82∗ ∗
49.74∗ ∗
3999.6∗ ∗
1.41
21.34∗ ∗
-42840.85
Aggregate
0.046
0.082∗ ∗
0.111∗ ∗
0.012
0.009
0.055
-0.020
-0.044
0.053
0.502∗ ∗
7.22∗ ∗
0.452∗ ∗
0.161∗ ∗
116.89∗ ∗
69.92∗ ∗
4213.78∗ ∗
1.026
93.12∗ ∗
-45356.05
The estimated model is the MDACP model, with the following mean (see Section 3.4.1 for
definition of variables):
¡ P9
P
2πp Re[t,T ]
] ¢
µ∗i,t = µi,t exp
+ ψi,s,p sin 2πp Re[t,T
) , and
j=1 ηi,j dj,t +
p=1,2,12 (ψi,c,p cos
T
T
PK
µi,t = ωi + j=1 αi,j Nj,t−1 + βi µi,t−1 ,
where K is the number of banks. W (ψ = 0) is the Wald statistic for joint significance of all the
seasonality variables (whose estimated coefficients are not reported). W (η = 0) is the Wald
statistic for the null hypothesis that all nine announcements are jointly non-significant, V ar(εt )
is the variance of the Pearson residual, and Q(10) is the Ljung-Box statistic of order 10 of the
residuals. Estimation was done by maximum likelihood. Estimates that are significant at the
1% and 5% level are indicated by two and one stars respectively, and they appear in bold font.
The number of observations is 9396.
3.4. Models and Results
59
their activity as a response to US and European macroeconomic figures, RABO
and OKOH decrease it as a response to US figures but increase it to react to
European figures. On the other hand, banks like BHFX, RABO and SHKH reduce their quoting in response to US figures. However, speeches of senior officials
of the government (η4 ) seem to pertain to the category of CK news, given that
this variable is not significant for any bank. It could of course also be that this
variable simply does not have any informational content, as perceived by foreign
exchange dealers, but in the previous chapter we find that it has an impact on
volatility which is significant at the 1% level. The remaining unscheduled news
(η6 , η7 , η8 and η9 ) hardly affects banks’ quoting activity and we can thus consider
them as CK news, unless markets don’t regard them to be informative at all.
Moreover, SGOX and DREF activities exhibit a significant effect respectively
on the other banks of each sample. We remind that the former banks react to
the first three much important news announcement whereas the other banks do
not react to all of them. This could be interpreted through the ”leader-follower”
perspective. leader banks like SGOX and DREF react rapidly (within five minutes) to news by increasing the frequency of their price revision, however follower
banks react afterward to guess more information from the quoting activity of the
leaders. This information often embodies the way of interpretation of news.
Table 3.8 shows for every type of announcement the result of a Wald test of
the null hypothesis that the announcements impact all banks in the same way.
The results show that US and European macroeconomic figures affect banks
differently in both samples, whereas the coefficients of interest rate reports are
only significantly different in sample 2. The remaining announcements have
impacts on different banks that are not significantly different, which is not a
surprise since their coefficients are much less significant in general.
In addition, we estimated a restricted DACP model, i.e. equation (3.4.4) when
αi,j = 0 for i 6= j, on the same banks’ quoting activity. Since we find almost the
same estimated dummy coefficients as for the more general model, the results are
not reported. We also estimated a DACP model on aggregate quoting activity,
(”Aggregate” in Tables 3.6 and 3.7), adopting the same seasonality variables,
a (1, 1)-lag structure and the same samples of banks, in order to compare the
obtained results with those generated by MDACP. We find, for instance in the
60
Chapter3. The Information Content of Quoting Activity
Figure 3.2: Correlogram of banks’ quoting and standardized residuals from
DACP models
This figure presents the correlogram of banks activity and of standardized residuals from DACP
Nt −µt
t
models. The standardized residuals (Pearson residuals) are defined as εt = Ntσ−µ
= √
.
t
µt /φ
The dashed line represents the autocorrelations of the raw series, and the solid line the autocorrelations of the Pearson residual. The 95% bounds of significance are also plotted.
3.4. Models and Results
61
Figure 3.3: Quantile plot of the Z statistic of individual banks
This figure presents the quantile plot of the Z statistic of individual banks. This statistic is
defined as the probability integral transform of the original data under the estimated density,
in our case, the double Poisson: Zt = F ∗ (Nt , µt ), see Appendix 2 for more details.
62
Chapter3. The Information Content of Quoting Activity
case of positive US macroeconomic figures, that there are both increases and
decreases in quoting activity of individual dealers. These effects offset each other,
which reduces the significance of news on quoting activity at the aggregate level.
Another example is European and US interest rate reports, which are significant
for three banks of sample 2, but not at the aggregate level. However, in the case
of negative US macroeconomic figures, there are both increases and decreases in
activity, but the increases seem to dominate at the aggregate level. This is thus
strong evidence that aggregate analysis of quoting activity can hide the fact that
individual banks have different reactions. In some cases, even though there is no
aggregate impact of news on quoting activity, individual banks do respond, but
their responses can offset each other, and in other cases, a positive coefficient at
the aggregate level can conceal a less unified picture at the level of individual
dealers.
Finally the results in Tables 3.6 and 3.7 show that there are significant interdealer effects. The quoting activity of each dealer increases or decreases in response to the lagged activity of some other dealers. Some banks do not influence
the quoting activity of other banks, they are clearly followers. In sample 1 each
bank’s quoting activity is sensitive to at least one other bank’s quotes. In sample 2, however, at least three banks’ quotes have a significant impact on every
dealers’ quotes. This supports the hypothesis that some dealers observe the frequency of price revision of some influential dealers’ quoting activity in order to
infer useful information. Moreover, the number of quotes of the rest of the banks
in both samples (”Rest” in Tables 3.6 and 3.7 ) is influenced by nearly all of
our large banks (except SGOX in sample 1 and BARL and DREF in sample 2).
The rest of the banks is made up of a very large number of small banks and
the fact that they are influenced by the quoting activity of the larger ones confirms that they are relatively more risk averse and simply follow the larger banks
in terms of price revision. Consequently, the results related to dealers’ quoting
activity sensitivity to both news announcements and quoting activity of some
other dealers, confirm the general hypothesis according to which quoting activity
provides an important informative signal. Indeed, during event periods, dealers
monitor the quoting activity of some others in order to infer their reaction to
news announcements before their immediate or afterwards reaction.
3.4. Models and Results
63
Table 3.7: Estimation results of MDACP models for sample 2, August 24 to
October 26, 2001.
Parameters BARL
DREF
OHVA
OKOH
SHKH
UBSZ
-0.300∗
Rest
Aggregate
η1
0.043
0.304∗ ∗
0.008
-0.165∗ ∗
0.124
0.036
0.007
η2
0.186∗ ∗
0.910∗ ∗
0.032
-0.088∗
-0.033
0.353∗ ∗
0.166∗ ∗
0.150∗ ∗
η3
0.068
0.602∗ ∗
0.009
0.074∗
-0.160
0.021
0.140∗ ∗
0.091∗ ∗
η4
0.039
-0.014
-0.143
0.059
0.028
0.012
η5
0.162∗ ∗
0.526∗ ∗
0.155∗
0.044
0.051
-0.043
-0.015
-0.010
-0.003
0.173
η6 -0.099
-0.217
0.041
0.008
-0.257
-0.023
-0.032
-0.013
η7 -0.034
-0.007
-0.035
-0.027
-0.369∗
-0.029
-0.002
-0.005
η8
0.221
0.592
0.023
-0.079
0.062
0.048
0.254
0.147
η9
0.069
0.149
0.049
0.047
0.045
0.098
0.053
0.068
0.429∗ ∗
0.663∗ ∗
α
αBARL
0.324∗ ∗
0.003
0.015∗
αDREF -0.042∗ ∗
0.323∗ ∗
αOHV A
0.044∗ ∗
0.008∗ ∗
0.426∗ ∗
αOKOH
0.002∗
0.003∗ ∗
-0.003∗ ∗
0.026
0.015∗
αSHKH
-0.007
0.015
0.043∗ ∗
0.162
0.059∗ ∗
0.015
0.006∗ ∗
0.057∗ ∗
0.019∗ ∗
0.659∗ ∗
0.001∗ ∗
0.012∗ ∗
0.231∗ ∗
0.085∗ ∗
0.231∗ ∗
0.030
0.004
0.261∗ ∗ -0.348∗ ∗
0.012
0.0112∗
0.113∗ ∗ -0.007
-0.017∗
-0.111∗
0.023∗ ∗ -0.009∗
0.017∗
ω
0.177∗ ∗
0.119∗ ∗
0.229∗ ∗
0.598∗ ∗
0.014
0.191∗ ∗
4.588∗ ∗
3.194∗ ∗
β
0.576∗ ∗
0.514∗ ∗
0.538∗ ∗
0.313∗ ∗
0.667∗ ∗
0.539∗ ∗
0.555∗ ∗
0.315∗ ∗
φ
0.958∗ ∗
0.609∗ ∗
1.34∗ ∗
0.455∗ ∗
0.830∗ ∗
0.559∗ ∗
0.208∗ ∗
0.165∗ ∗
W (η = 0)
22.88∗ ∗
268.21∗ ∗
3.12
27.66∗ ∗
21.08∗
46.85∗ ∗
80.74∗ ∗
65.63∗ ∗
W (ψ = 0)
24.85∗ ∗
91.57∗ ∗
8.84
6.30
79.3∗ ∗
51.42∗ ∗
34.29∗ ∗
W (αi = 0)
321.8∗ ∗
64.52∗ ∗
190.7∗ ∗
150.0∗ ∗
91.24∗ ∗
194.3∗ ∗
αU BSZ
V ar(εt ) 0.99
Q(10)
44.08∗ ∗
-0.013
1.11
1.11
1.64
10
91.36∗ ∗
77.11∗ ∗
LL -12888.70 -8699.02
12.46
17774.17∗ ∗
3752.56∗ ∗
0.95
1.02
1.002
1.01
17.07
23.47∗ ∗
71.68∗ ∗
120.91∗ ∗
-12611.31 -16239.46 -9480.34
-13937.07 -21612.71
-23911.72
The number of observations is 4968. See caption of Table 3.6 for the rest of the table description.
64
Chapter3. The Information Content of Quoting Activity
Table 3.8: Wald tests of equality for all banks of the effect of news
Announcement category
Sample 1
Sample 2
27.64∗ ∗
26.99∗ ∗
23.17∗ ∗
1.31
3.58
34.49∗ ∗
147.09∗ ∗
56.74∗ ∗
4.60
32.58∗ ∗
1.39
2.49
1.08
1.33
3.98
3.24
0.21
0.51
Scheduled:
η1
η2
η3
η4
η5
-
Positive US macro figures
Negative US macro figures
European macro figures
Speeches of senior officials
Interest rate reports
Unscheduled:
η6
η7
η8
η9
-
Economic institutes forecasts
OPEC member declarations
Central bank intervention rumors
Extraordinary events
This table shows the results of a Wald test for the hypothesis that ηi,1 =
ηi,2 = · · · = ηi,K , where the first index refers to the announcement and the
second to the bank, and K is number of banks.
One and two stars indicate rejection of the null hypothesis at the 5% and
1% respectively. The test statistic takes the following form:
0
0
W = (Rηi ) (RΣi R )−1 (Rηi ) ,
where for sample 1,

1
R=0
0
−1
1
0
0
−1
1

0
0 ,
−1
2
2
2
refers to the variance of coefficient ηi,k and
Σi = diag(σi,1
, . . . , σi,4
), σi,k
ηi = (ηi,1 , . . . , ηi,4 ). Σi is diagonal due to the structure of the likelihood
function (see Appendix 2).
3.5. Conclusion
3.5
65
Conclusion
In this chapter we show that foreign exchange dealers’ quoting activity can play
a significant role in conveying information to overall market participants. By
looking at a sample of major dealers on the euro/dollar exchange market, we offer
evidence of the fact that firstly, banks’ quoting activity reacts to certain news
announcements and that FX dealers’ quoting activity reacts differently to the
same news announcements. We take this as an indication of their heterogeneous
interpretation of the news content. Moreover, the differences in reaction are
more extreme than could be expected a priori, as it is not rare to see that some
banks increase their activity, while others decrease it in response to the same
announcement. Finally, there is significant inter-dealer interaction, as banks’
quoting activity is typically affected by the intensity of quote revision of some
others. This means that follower dealers observe the frequency of price revision
of leader dealers in order to infer some useful information. This offers support
for the hypothesis that most FX dealers monitor the quoting activity of some
influential or leader ones in order to infer other dealers’ private information,
stemming for instance from their customer order flow or their reaction to public
news announcements.
66
Chapter3. The Information Content of Quoting Activity
3.6
Appendix
A.1 The Double Poisson Distribution
The double Poisson distribution, for the integer valued positive random variable
y, with parameters µ and φ (µ > 0, φ > 0) has the following expression:
³ 1
´ µ e−y y y ¶ µ eµ ¶φy
−φµ
fDP (y|µ, φ) = c(µ, φ) φ 2 e
,
y!
y
where c(µ, φ) is a constant (with respect to y) such that the probabilities add
to one. The parameters µ and φ are respectively the mean and the coefficient of dispersion of the double Poisson. Efron (1986) shows that the value
of c(µ, φ) varies little with respect to µ and φ. He also provides the approximation
1
c(µ,φ)
≈ 1+
1−φ
(1
12µφ
+
1
)
µφ
and he suggests maximizing the approximate
likelihood (leaving out the highly nonlinear multiplicative constant) in order to
estimate the parameters and using the correction factor when making probability
statements using the density.
A.2 Copulas
Copulas provide a very general way of introducing dependence among several series with known marginals. Copula theory goes back to the work of Sklar (1959),
who showed that a joint distribution can be decomposed into its K marginal
distributions and a copula, that embeds the dependence between the variables.
This theorem provides an easy way to form valid multivariate distributions from
known marginals. A more detailed account of copulas can be found in Joe (1997)
and in Nelsen (1999). Let H(y1 , . . . , yK ) be a continuous K-variate cumulative distribution function with univariate marginals Fi (yi ), i = 1, . . . , K, where
Fi (yi ) = H(∞, . . . , yi , . . . , ∞). According to Sklar (1959), there exists a function
C, called copula, mapping [0, 1]K into [0, 1], such that:
H(y1 , . . . , yK ) = C(F1 (y1 ), . . . , FK (yK )) .
The joint density function is given by the product of the marginals and the
copula density:
K
∂C(F1 (y1 ), . . . , FK (yK ))
∂H(y1 , . . . , yK ) Y
fi (yi )
=
.
∂y1 . . . ∂yK
∂F1 (y1 ) . . . ∂FK (yK )
i=1
3.6. Appendix
67
While there are many alternative formulations for copulas in the bivariate
case, the number of possibilities for multivariate copulas is rather limited. We
choose to work with the most easy one, which is the Gaussian copula, obtained
by the inversion method (based on Sklar, 1959). This is the K-variate function
defined by
C(z1 , . . . , zk ; Σ) = ΦK (Φ−1 (z1 ), . . . , Φ−1 (zK ); Σ) ,
where ΦK is the K-variate standard normal distribution function, Φ−1 is the inverse of the standard univariate normal distribution function and Σ is a positivedefinite symmetric matrix. Its density is given by
µ
c(z1 , . . . , zK ; Σ) = | Σ |
−1/2
exp
1 0
(q (IK − Σ−1 )q
2
¶
,
q = (q1 , . . . , qK )0 , with scores qi = Φ−1 (zi ), (qi is called a normal score) . It can
be seen that if z1 , . . . , zK are mutually independent, the matrix Σ is equal to the
identity matrix IK and the copula density is then equal to 1.
The joint density of the counts Ni,t (i = 1, 2, . . . , K) in the Double Poisson
case with the Gaussian copula is:
ht = h(N1,t , . . . , NK,t , θ, Σ) =
K
Y
fDP (Ni,t |µ∗i,t , φi ) · c(qt ; Σ) ,
i=1
where
fDP (Ni,t |µ∗i,t , φi )
denotes the Double Poisson density of the observation Ni,t ,
with conditional mean µ∗i,t and dispersion parameter φi . The function c denotes
the multivariate normal copula density defined above, and θ = (ω, vec(A), vec(B)),
see equation (3.4.3). The qi,t , gathered in the vector qt are the normal quantiles
of the zi,t :
qt = (Φ−1 (z1,t ), . . . , Φ−1 (zK,t ))0 ,
and zi,t is the probability integral transformation (PIT) of the continuoused count
data, under the marginal densities (see Appendix A.3 for details).
By taking the log, one gets the contribution of the t-th observation to the
log-likelihood function of the sample as:
log(ht ) =
K
X
i=1
log(fDP (Ni,t , µ∗i,t , φi )) + log(c(qt ; Σ))).
68
Chapter3. The Information Content of Quoting Activity
We use a two-stage estimator as in Patton (2002). In the first step, we
maximize the log-likelihood function with respect to the parameter θ, which
appears in the first term only. It turns out clearly that this part of the likelihood
function can itself be maximized in K separate steps. Given that we use the
multivariate normal copula, the second step of the two-stage procedure does not
require any optimization, as the MLE of the variance-covariance matrix of a
multivariate normal with a zero mean, is simply the sample counterpart:
T
1X 0
qt q .
Σ̂ =
T t=1 t
It is important to realize that correct specification of the density in the
marginal models is crucial to the specification of the copula, as any mistake
would have as a consequence the fact that the uniformity assumption would be
violated, which would invalidate the choice of the copula.
A.3 Discrete Distributions and PITT
The problem with discrete distributions is that the probability integral transformation theorem (PITT) of Fisher (1932) does not apply, and the uniformity
assumption does not hold, regardless of the quality of the specification of the
marginal model. The PITT states that if Y is a continuous variable, with cumulative distribution F , then
Z = F (Y )
is uniformly distributed on [0, 1].
Denuit and Lambert (2002) use a ”continued extension” to overcome these
difficulties and apply copulas with discrete marginals. The main idea of continued
extensions of a discrete variable is to create a new random variable Y ∗ by adding
to a discrete variable Y a continuous variable U valued in [0, 1], independent of
Y , with a strictly increasing cdf, sharing no parameter with the distribution of
Y , such as the uniform on [0, 1] for instance:
Y ∗ = Y + (U − 1) .
As can be seen, knowing the value of Y ∗ , which is the new continuous variable,
is equivalent to knowing the value of the underlying count. If Y ∗ = 4.38, then
3.6. Appendix
69
we know that Y = 5. Hence we do not lose any information by creating this new
variable.
Using continued extension, Denuit and Lambert (2002) state a discrete analog
of the PITT. If Y is a discrete random variable with domain χ, in N, such that
fy = P (Y = y), y ∈ χ, continuoused by U, then
Z ∗ = F ∗ (Y ∗ ) = F ∗ (Y + (U − 1)) = F ([Y ∗ ]) + f[Y ∗ ]+1 U = F (Y − 1) + fy U
is uniformly distributed on [0, 1], and [Y ] denotes the integer part of Y .
It is worth pointing out that we use the continuoused version of the probability integral transformation in order to test the correct specification of the
marginal models. If the marginal models are well-specified, then Z ∗ , the PIT
of the series under the estimated distribution and after continued extension, is
uniformly distributed.
Part II
Volatility Dynamics Around
Technical Signal Based Trading
71
Chapter 4
The Performance Analysis of
Technical Chart Patterns
This chapter is based on a paper published in Empirical Economics, 2006, 30,
947-971, and co-authored by Hervé Van Oppens.
4.1
Introduction
The issue of market efficiency is the subject of a long and continuing debate since
the 1970s (Fama, 1970, Black, 1972, and Lucas, 1978, amongst others). The dominant viewpoint is that prices incorporate the information available about the fundamental values, returns are random walk, and there is no arbitrage opportunity
to make money by beating the market. However this viewpoint is based on the
general behavior of financial markets, and the supporting empirical evidence uses
low frequency data (monthly, weekly or daily returns). Black (1986) and Shiller
(2003) have mentioned several anomalies triggered by the market microstructure
and dealers behavior. Chordia, Roll, and Subrahmanyam (2005) present a model
where financial markets converge to efficiency according to a certain speed. By
regressing the returns on lagged order imbalances, they show that the latter are
often significant predictors of future returns over short intervals. In no more
than thirty minutes, order imbalances lose their predictive ability and returns
are no longer negatively dependent. Stock markets are inefficient over a very
short period of time and are pushed toward efficiency by the collective behavior
of traders.
The anomalies that feature the efficient market could be explained by the
behavior of non-fundamental traders, called technical traders and considered by
economists as noise traders. These traders build their decisions on technical analysis that provide another kind of information, different from the fundamental one
73
74
Chapter4. The Performance Analysis of Technical Chart Patterns
revealed for instance through the announcements. Technical analysis is the oldest method for analyzing market behavior. It is defined by Murphy (1999) as
the study of market action, primarily through the use of charts, for the purpose
of forecasting future price trends. The term ‘market action’ includes three main
sources of information available to the technician: price, volume and open interest. Béchu and Bertrand (1999) distinguish three categories of technical analysis.
Traditional analysis is entirely based on the study of charts and the location of
technical patterns like the head and shoulders pattern.1 Modern analysis is composed of more quantitative methods like moving averages, oscillators, etc. The
third category, qualified as philosophical, has the ambition to explain more than
the overall market behavior. One of the most famous examples is the Elliot wave
theory (for more details see Prost and Prechter, 1985) which assumes that every price movement can be decomposed into eight phases or waves: five impulse
waves and three corrective ones.
In this chapter, we focus on the traditional approach of technical analysis
and particularly on chart patterns. These patterns have been studied, among
others, by Levy (1971), Osler (1998), Dempster and Jones (1998b), Chang and
Osler (1999), and Lo, Mamaysky, and Wang (2000) who have mainly focused
on the profitability of trading rules related to chart patterns and also on the
informational content that could generate such patterns. All these investigations conclude to the lack of profitability of technical patterns. However, Lo,
Mamaysky, and Wang (2000) find that these patterns present an informational
content that affect stock returns.
We investigate twelve chart patterns in the euro/dollar foreign exchange market. Currency markets are especially appropriate for testing technical signals
because of their very high liquidity, low bid-ask spread, and round-the-clock decentralized trading (Chang and Osler, 1999). As our empirical evidence is built
upon high frequency data, we rather focus on the speed of convergence to market
efficiency than on the hypothesis of market efficiency per se. As argued by Chordia, Roll, and Subrahmanyam (2005), information takes a minimum of time to
be incorporated into prices so that markets may be at the same time inefficient
over a short-time (e.g. 5-minute) interval and efficient over a longer (e.g. daily)
1
This chart pattern is defined in Section 4.3.2.
4.1. Introduction
75
interval.
To test the existence of twelve chart patterns in the euro/dollar currency
market, we use two identification methods for detecting local extrema. The first
method (M1), also used in the literature, considers only prices at the end of each
time interval (they are called close prices). The second method (M2), which is
new compared to those used in the literature, takes into account both the highest
and the lowest price in each interval of time corresponding to a detected pattern.
The detected extrema are analyzed through twelve recognition pattern algorithms, each of them corresponding to a defined chart pattern. Our purpose is
to analyze the predictability and profitability of each type of chart pattern. In
addition, we intend to test the usefulness of our contribution regarding the extrema detection method M2. Although Osler (1998) and Chang and Osler (1999)
briefly mention these prices, most of the previous studies on chart patterns do
not give much interest to high and low prices. This is in sharp contrast to the
majority of practitioners (in particular dealers) who use high and low prices in
their technical strategies through bar charts and candlesticks. In addition, Fiess
and MacDonald (2002) show that high, low and close prices carry useful information for forecasting the volatility as well as the level of future exchange rates.
Consequently, in our framework, we investigate also the sensitivity of the chart
patterns to the extrema detection methods M1 and M2. Moreover, to evaluate
the statistical significance of our results, we run a Monte Carlo simulation. We
simulate a geometric Brownian motion to construct artificial series. Each of them
has the same length, mean, variance and starting value as the original observations. Then we look for the occurrence of the chart patterns in the artificial
series for which we compute the predictive success and profitability. Finally we
compare these results to those obtained in the observed data series.
Our results show the apparent existence of some chart patterns in the euro/dollar
intra-daily foreign exchange rate. More than one half of the detected patterns,
according to M1 and M2, seem to have a significant predictive success. Nevertheless, only two patterns from our sample of twelve present a significant profitability
which is however too small to cover the transaction costs. We show, moreover,
that the extrema detection method M2 provides higher but riskier profits than
those provided by M1. These findings are in accordance with those found by Levy
76
Chapter4. The Performance Analysis of Technical Chart Patterns
(1971), Osler (1998), Dempster and Jones (1998b), Chang and Osler (1999).
The chapter is organized as follows. In Section 4.2, we summarize the most
recent empirical studies which have focused on technical analysis, particularly
on chart patterns. Section 4.3 is dedicated to the methodology adopted for
both the extrema detection methods M1 and M2, and to the pattern recognition
algorithms. The section also includes details about the two criteria used for the
analysis of the observed technical patterns: predictability and profitability. The
data and empirical results are exposed in Section 4.4. We conclude in Section
4.5. The appendix, displayed at the end of the chapter, involves the definitions of
eleven chart patterns and technical details about the chart recognition algorithms.
4.2
Technical Analysis
Technical analysis is widely used in practice by several dealers, also called technical analysts or chartists. According to Cheung and Wong (1999), 25 to 30
percent of the foreign exchange dealers base most of their trade on technical
trading signals. More broadly, Allen and Taylor (1992) show, through questionnaire evidence, that technical analysis is used either as the primary or the
secondary information source by more than 90% of the foreign exchange dealers
trading in London. Furthermore, 60% judge charts to be at least as important
as fundamentals. Most of them consider also chartism and fundamental analysis
to be largely complementary. Menkhoff (1998) shows in addition that more than
half of foreign exchange market participants in Germany give more importance
to the information coming from non-fundamental analysis, i.e. technical analysis
and order flows. Moreover, Lui and Mole (1998a) show that technical analysis is
the most used method for short term horizon on the foreign exchange market in
Hong Kong.
Despite its broad use by practitioners, academics have historically neglected
technical analysis, mainly because it contrasts with the most fundamental hypothesis in finance, namely market efficiency. Indeed, the weak form of the market efficiency hypothesis implies that all information available in past prices must
be reflected in the current price. Then, according to this hypothesis, technical
analysis, which is entirely based on past prices (Murphy, 1999), cannot predict
future price behavior.
4.2. Technical Analysis
77
Recently, several studies have focused on technical analysis. Brock, Lakonishok, and LeBaron (1992) support the use of two of the simplest and most popular trading rules: moving average and trading range break (support and resistance
levels). They show that these trading rules help to predict return variations in
the Dow Jones index. These simple trading rules were studied, amongst others,
by Dooley and Shafer (1984), Sweeney (1986), Levich and Thomas (1993), Neely
(1997) and LeBaron (1999) in the context of the foreign exchange rate dynamics. Moreover, Andrada-Felix, Fernandez-Rodriguez, and Sosvilla-Rivero (1995),
Ready (1997) and Detry (2001) investigate the use of these rules in stock markets. Still with the moving average trading rules, Gençay and Stengos (1997),
and Gençay (1998, 1999) examine the predictability of stock market and foreign
exchange market returns by using past buy and sell signals, and they find an
evidence of nonlinear predictability of such returns.
In addition to these simple trading rules, technical analysis abounds in methods in order to predict future price trends. These methods have also been considered in empirical research. Jensen (1970) tests empirically the ’relative strength’
trading rule.2 The estimated profit provided by this trading rule is not significantly bigger than the one obtained by the ’Buy and Hold’ strategy.3 Osler
(2000) finds that the support and resistance technique provides a predictive success. Other studies make use of genetic programs to develop trading rules likely
to realize significant profits (e.g., Neely, Weller, and Dittmar, 1997, Dempster
and Jones, 1998a and Neely and Weller, 2003). Furthermore, Blume, Easley, and
O’Hara (1994) demonstrate that sequences of volume can be informative. This
would explain the widespread use by practitioners of technical analysis based
upon volumes.
The different studies mentioned above have mainly focused on linear price
relations. However, other researchers have oriented their investigations to nonlinear price relations. Technical chart patterns are considered as non-linear patterns. Both Murphy (1999) and Béchu and Bertrand (1999) argue that these
kinds of patterns present a predictive success which allows traders to acquire
2
Based on computing the ratio Pt /P̄t where P̄t corresponds to the mean of prices preceding
the moment t, the relative strength trading rule consists in buying the asset if the ratio is bigger
than a particular value and selling it when the ratio reaches a specific threshold.
3
This strategy consists in buying the asset at the beginning of a certain period and keeping
it until the end.
78
Chapter4. The Performance Analysis of Technical Chart Patterns
profit by developing specific trading rules. In most studies, technical patterns
are analyzed through their profitability. Levy (1971) focuses on the predictive
property of the patterns based on a sequence of five price extrema and conclude,
after taking into account the transaction costs, to the unprofitability of such
configurations. Osler (1998) analyzes the most famous chart pattern, the head
and shoulders pattern. She underlines that agents who adopt this kind of technical pattern in their strategy must be qualified as noise traders because they
generate important order flow and their trading is unprofitable. Dempster and
Jones (1998b) and Chang and Osler (1999) obtain the same conclusion regarding
the non profitability of the trading rules related to chart patterns. In contrast,
Lo, Mamaysky, and Wang (2000) show that the informational content of chart
patterns affects significantly future stock returns.
Some studies go beyond the scope of testing the performance of trading models. For example, Gençay, Ballocchi, Dacorogna, Olsen, and Pictet (2002) and
Gençay, Dacorogna, Olsen, and Pictet (2003) employ a widely used commercial
real-time trading model as a diagnostic tool to evaluate the statistical properties
of foreign exchange rates. They consider that the trading model on real data outperforms some sophisticated statistical models implying that the latter are not
relevant for capturing the data generating process. They add that in financial
markets, the data generating process is a complex network of layers where each
layer corresponds to a particular frequency.
In our study we choose to deal with high frequency data, believing that all
our results are sensitive to the time scale. The results obtained for one hour
or thirty minutes time scale are certainly different from those obtained for five
minutes frequency. However, the goal of our study is to analyze the performance
of some chart patterns at a specific time scale without generalizing our results to
other frequencies. Our choice of high frequency data, as emphasized by Gençay,
Dacorogna, Olsen, and Pictet (2003), is motivated by two main reasons. First,
any position recommended by our strategy (defined below in subsection 4.3.3)
have to be closed quickly within a short period following the chart completion.
The stop-loss objectives need to be satisfied and the high frequency data provides
an appropriate platform for this requirement. Second, the trading positions and
strategies, can only be replicated with a high statistical degree of accuracy by
4.3. Methodology
79
using high frequency data in a real time trading model.
However in practice, technical analysts often combine high and low time scales
in order to monitor their positions in the short (five-minutes to one hour) and
long run (one day to one month).
4.3
Methodology
The methodology adopted in this chapter consists in identifying regularities in
the time series of currency prices by extracting nonlinear patterns from noisy
data. We take into consideration significant price movements which contribute
to the formation of a specific chart pattern and we ignore random fluctuations
considered as noise. We do this by adopting a smoothing technique in order to
average out the noise. The smoothing technique allows to identify significant
price movements which are only characterized by sequences of extrema.
In the first subsection we present two methods used to identify local extrema.
Then, we explain the pattern recognition algorithm which is based on the quantitative definition of chart patterns. In the third subsection, we present the two
criteria chosen for the analysis of the detected charts: predictability and profitability. The last subsection is dedicated to the way we compute the statistical
significance of our results. It is achieved by running a Monte Carlo simulation.
4.3.1
Identification of Local Extrema
Each chart pattern can be characterized by a sequence of local extrema, that is
by a sequence of alternate maxima and minima. Two methods are used to detect
local extrema. The first method, largely used in the literature, is based on close
prices, i.e. prices which take place at the end of each time interval. The second
method is built on the highest and the lowest prices in the same time intervals.
We examine the usefulness of using high and low prices in the identification
process of chart patterns. Taking these prices into account is more in line with
practice as dealers use bars or candlestick charts to build their technical trading
rules.4 Moreover, Fiess and MacDonald (2002) show that high and low prices
carry useful information about the level of future exchange rates.
4
Béchu and Bertrand (1999) stipulate that line charts are imprecise because they do not
display all the information available as they are only based upon close prices of each time
interval. In contrast, bar charts and candlesticks involve the high, low, open and close prices
of each time interval.
80
Chapter4. The Performance Analysis of Technical Chart Patterns
The extrema detection method based on close prices (M1) works as follows.
The first step consists in smoothing the price curve to eliminate the noise in
prices and locate the different extrema on the smoothed curve. To smooth the
estimated curve we use the Nadaraya-Watson kernel estimator.5 We then determine different extrema by finding the moments at which the kernel first derivative
changes its sign. We therefore guarantee the alternation between maxima and
minima. This smoothing technique has been also used by Lo, Mamaysky, and
Wang (2000). Other methods to detect extrema have been adopted by Levy
(1971), Osler (1998), Dempster and Jones (1998b) and Chang and Osler (1999).
The second step involves projections of the smoothed extrema on the original
price curve. In other words, we deduce corresponding extrema on the original
curve.
The second method (M2) is based on high and low prices. Local maxima
must be determined on the high price curve and local minima on the low one.
We smooth both curves and we select the corresponding extrema when there
is a change of the sign for the kernel first derivative function. In such a case,
alternation between extrema is not automatically obtained. Thus, we start by
projecting the first extremum on the corresponding original price curve. If this
extremum is a maximum (minimum), we project it into the high price curve (low
price curve) and then we alternate between a projection of a minimum (maximum) on the low price curve (high price curve) and a projection of a maximum
(minimum) on the high price curve (low price curve).
To detect local extrema we use a rolling window which goes through all the
time periods with an increment of a single time interval. For each window, we
apply both extrema detection methods and the pattern recognition algorithms
in order to test if the detected sequence of extrema corresponds to one of our
twelve chart pattern definitions (see the following section). The advantage of a
rolling window is to concentrate on patterns that sequentially develop in the same
window and therefore to cancel the risk of look-ahead bias. This implies that the
future evolution of the price curve is not yet known at the time of detection of
technical patterns. A technical pattern is thus recorded only if all extrema have
been detected in windows of identical time duration. Furthermore, we add a
5
Details about this estimator are given in Appendix A (at the end of this chapter).
4.3. Methodology
81
filter rule to keep only one record of each detected chart pattern. We present in
Appendix B a detailed description of the two extrema detection methods.
4.3.2
Chart Pattern Quantitative Definitions
By looking at specialized books on technical analysis like Murphy (1999) and
Béchu and Bertrand (1999), which provide graphical descriptions of technical
patterns, we build twelve quantitative definitions corresponding to the most famous chart patterns. Only the Head and Shoulders (HS) definition is presented
in this section. This pattern, HS, is defined from a particular sequence of extrema
detected by the method presented in Appendix B. The other pattern definitions
and graphs are presented in Appendix C. The eleven remaining chart patterns
are the following: Inverse Head and Shoulders (IHS), Double Top (DT), Double
Bottom (DB), Triple Top (TT), Triple Bottom (TB), Rectangle Top (RT), Rectangle Bottom (RB), Broadening Top (BT), Broadening Bottom (BB), Triangle
Top (TRIT) and Triangle Bottom (TRIB).
From a series of price Pt , we denote by Ei (i = 1, .., I) the local extremum
i from a sequence composed of I extrema and tEi the moment when it occurs.
The slope, p(Ei , Ej ), of the line passing through Ei and Ej and the y-coordinate
at tk of a point of this line, Vtk (Ei , Ej ), are defined as follows:
Ej − Ei
tEj − tEi
Vtk (Ei , Ej ) = Ei + (tk − tEi ) × p (Ei , Ej ) .
p(Ei , Ej ) =
(4.3.1)
(4.3.2)
Figure 4.1 presents the theoretical Head and Shoulders chart pattern while
Figure 4.2 illustrates the observed pattern after implementing both extrema detection methods. The theoretical figure serves mainly to help in the comprehension of the definition of the HS chart pattern. The HS chart pattern is
characterized by a sequence of five extrema Ei (i = 1, .., 5) such that:
82
Chapter4. The Performance Analysis of Technical Chart Patterns
Figure 4.1: The Head and Shoulders (HS) theoretical chart pattern
4.3. Methodology
83
Figure 4.2: The Head and Shoulders: Observed chart pattern
0.8585
0.8565
0.8575
price
0.8595
Head and Shoulders: Method M1
2170
2180
2190
2200
time
price
0.857
0.858
0.859
0.860
Head and Shoulders: Method M2
2170
2180
2190
2200
time
This figure shows an observation window in which the Head and Shoulders chart pattern is
detected through both M1 and M2 methods (detailed in Appendix B). The dashed line in both
graphs illustrates the smoothed price curves and the solid line, for the first graph, presents the
original price curve. The second graph shows the original price series through bar charts. Each
of them involves the maximum, the minimum, the open and the close price for each five-minute
time interval.
84
Chapter4. The Performance Analysis of Technical Chart Patterns























hs ≡
E1 > E2
(a)
E 3 > E 1 , E 3 > E5
(b)
|p(E1 , E5 )| ≤ tg(10)
(c)
|p(E2 , E4 )| ≤ tg(10)
(d)
E1 −VtE (E2 ,E4)
0.9 ≤ E5 −Vt 1 (E2 ,E4) ≤ 1.1 (e)
E5




h

1.1 ≤ s ≤ 2.5
(f)




t 2 −td

1

≤ tfE−t
≤2
(g)

2

E4


t −t

1

≤ E4 m E2 ≤ 2
(h)

2



¡
¢

Ptd − Ptmin ≥ 23 × h
(i)
where
- h is the height of the head: h = E3 − VtE3 (E2 , E4 )
- s is the average height of the two shoulders: s =
(E1 −VtE (E2 ,E4 ))+(E5 −VtE (E2 ,E4 ))
1
2
5
¡
¢
- td is the starting time for the pattern: td = maxt Pt ≤ Vt (E2 , E4 ) , t < tE1
¡
¢
- tf is the ending time for the pattern: tf = mint Pt ≤ Vt (E2 , E4 ) , t > tE5
- td−(f −d) = td − (tf − td )
- tf +(f −d) = tf + (tf − td )
- m is the average time that the shoulders take for their total completion: m =
(tE2 −td )+(tf −tE4 )
2
- Ptmin is the smallest price observed in the [td−(f −d) , td ]:
¯
Ptmin = min(Pt ) ¯ td−(f −d) ≤ t ≤ td
If a sequence of five extrema satisfies the above conditions, they build up a Head
and Shoulders chart pattern. Theoretically, at the completion of this chart pattern, the price must go down for at least the height of the head, h. Furthermore,
the objective price detected by the chart pattern has to be reached within the
time interval [tf , tf +(f −d) ]. In other words, the price has to reach at least P (obj)
such that:
P (obj) = Ptf − h .
(4.3.3)
4.3. Methodology
4.3.3
85
The Performance Measures
Detected chart patterns are analyzed in terms of predictability and profitability.
In other words, we study the capability of each chart pattern to predict the
future price trend just after the chart completion and the profit that a dealer
could realize when he applies a trading rule.
Predictability
Following its completion, the chart pattern can be used to forecast the future price
trend. More precisely, it predicts the price objective which has to be reached.
We denote by h the predicted price variation, and by tf and td respectively, the
time at the end and at the beginning of the chart pattern. If the pattern predicts
a downward trend, the price objective is given by equation (4.3.3). This price
objective has to be reached within the time interval [tf , tf +(f −d) ]. In such cases,
we can measure the actual price reached in this time interval by computing Pa
such that:
© ¯
ª
Pa = min Pt ¯tf ≤ t ≤ tf +(f −d) .
(4.3.4)
The value of the observed trend is then:
trend = Ptf − Pa .
(4.3.5)
The predictability criterion is defined as follows:
pred =
trend
.
h
(4.3.6)
We distinguish three possible cases:
• 0 ≤ pred < 1 : the price does not reach its predicted objective. It goes
in the predicted direction but only for the fraction pred of the forecasted
objective.
• pred = 1 : the price reaches exactly its objective.
• pred > 1 : the price exceeds its objective by (pred − 1).
Consequently if pred ≥ 1, the chart pattern can be said to predict successfully
the future price trend.
86
Chapter4. The Performance Analysis of Technical Chart Patterns
Profitability
If a chart pattern presents a predictive success, is it sufficient to get a profit ? To
answer this question, we investigate the profitability that technical patterns could
imply. When the price evolves in the direction predicted by the chart pattern, a
trader who takes a position at a precise time could realize a profit. Nevertheless,
if the price evolves in the opposite direction, the position taken at the same
time would involve a loss. A profit or a loss is the result of the implementation
of a trading rule chosen by a chartist trader at a given time according to the
completion of the chart pattern.
We propose the following strategy: the trader opens a position at the end
of the pattern (at the moment of its completion) and closes it according to the
future price direction. We distinguish two cases for the future trend:
• If the price evolves in the predicted direction , the trader closes his position
when the price reaches 50% of the predicted price variation, h.
• If the price evolves in the opposite direction, the trader closes his position
after a loss corresponding in absolute value to 20% of the forecasted price
variation.
However, if at the end of the interval [tf , tf +(f −d) ], the trader position is not
yet closed, this latter is automatically closed at tf +(f −d) . In both cases, the trader
can be considered as risk averse. Indeed, he limits his eventual profit and accept
only small losses.
Once the predictability and the profitability criteria of each pattern are computed, we compare the results for the two extrema detection methods M1 and M2.
We adopt a test of difference of means in order to infer the statistical significance
of such comparisons. It consists in computing the statistic, t, as follows:
mM 1 − mM 2
t = q¡ 2
,
s2M 2 ¢
sM 1
+ nM 2
nM 1
(4.3.7)
where mMi and s2Mi are respectively the estimated mean and variance of the
outputs (i.e. the number of detected charts, the predictability or the profitability
criteria) obtained when method Mi (i=1,2) is adopted. The t-statistic follows a
Student distribution with nM 1 + nM 2 − 1 degrees of freedom, where nM 1 and nM 2
4.3. Methodology
87
are respectively the number of observations resulting from the methods M1 and
M2.
The last step for the profitability analysis consists in taking into consideration
the risk incurred by the strategy. The latter is measured by the variability of
the obtained profits. We compute three measures in order to gauge our strategy
performance. We start by computing the ratio of mean profit to its standard
deviation. However, according to Dacorogna, Gençay, Müller, and Pictet (2001)
and Dacorogna, Gençay, Müller, Olsen, and Pictet (2001), this kind of Sharpe
ratio is numerically unstable, exhibits a lot of deficiencies, and does not take into
consideration dealers risk aversion. For a robust performance evaluation we adopt
two other performance measures, proposed by these authors, which are directly
related to the utility of a strategy of risk averse dealer. Both are based on the
maximization of the expected utility of a dealer. The first measure, called Xef f ,
considers a constant risk aversion, while the second one, named Ref f , supposes
an asymmetric risk aversion (a higher risk aversion when there is a loss).
Let us assume that the return Rt corresponding to the period t follows a Gaussian random walk with mean Rt and the risk aversion parameter α is constant
with respect to Rt . The resulting utility u(Rt ) of an observation is −e(−αRt ) , with
an expectation value of u = u(Rt )e(α
2
2 σt )
2
, where σt2 is the variance of Rt . The
expected utility can be transformed back to the effective return, Xef f =
where Xef f = Rt −
− log(−u)
α
ασt2
.
2
The asymmetric effective returns measure, Ref f , corresponds to a high risk
aversion in the zone of negative returns and a low one in the zone of profits. A
high risk aversion in the zone of negative returns means that the performance
measure is dominated by the large drawdowns. It is assumed that dealers have
two risk aversion levels: a low one, α+ for positive Rt and a high one, α− ,
for negative Rt , where α+ < α− . The high value of α− reflects the high risk
aversion of typical market participants in the loss zone. Trading models may
have some losses but, if the loss observations strongly vary in size, the risk of
very large losses becomes unacceptably high. On the side of the positive profit
observations, a certain regularity of profits is also better than a strong variation in
size. However, this distribution of positive returns is never as vital for the future
of market participants as the distribution of losses. Therefore, α+ is smaller than
88
Chapter4. The Performance Analysis of Technical Chart Patterns
α− and we assume that α+ =
α−
4
and α− = 20% (Dacorogna, Gençay, Müller,
Olsen, and Pictet, 2001). The exponential-type utility function becomes:

 − e−α+ Rt
α+
u(Rt ) =
 1 − 1 −
α−
α+
for Rt ≥ 0
e−α− Rt
α−
for Rt < 0.
The inverse formula computes the asymmetric effective returns, Ref f as follow:
Ref f =

 − log(−α+ u)
α+
for u ≥
1
α+
 − log(1− α+ −α− u)
α−
for u <
1
.
α+
α−
These performance measures adjust the mean profit for a kind of risk premium. We have adapted these two Xef f and Ref f using the profit levels instead
of returns realized after the completion of the chart pattern.
4.3.4
Monte Carlo Simulation
In order to assess the statistical significance of the obtained results, we run a
Monte Carlo simulation. We create 200 artificial exchange rate series6 and we
implement both extrema detection methods and the pattern recognition algorithms. These series follow a geometric Brownian motion process and are characterized by the same length, mean, variance and starting value as the original
observations.7
There is an important difference between the simulated series and the original
one: the simulated series are built in such a way that any detected pattern is
fortuitous, whereas in the original exchange rate series, this may or may not be
true. The existence of technical patterns in the original series could be generated
by trader behaviors which induce a particular pattern in the prices.
The simulation consists in repeating the procedure of the chart detection, predictability and profitability computation 200 times in order to get a distribution
for each of them under the null hypothesis. The null hypothesis H0 states that
there is no chart patterns (or predictability, or profitability) in the observed series. Then, for example, the probability for observing a profit (in the simulation)
higher than the observed one is equivalent to the p-value of the observed profit.
6
We limit our simulation to 200 series because the recognition pattern algorithm needs a lot
of computer time.
7
The same methodology was adopted by Chang and Osler (1999), Osler (1998), Gençay
(1998), Gençay, Ballocchi, Dacorogna, Olsen, and Pictet (2002), and Gençay, Dacorogna, Olsen,
and Pictet (2003).
4.4. Data and Empirical Results
89
Thus, if the p-value is higher than 5% then we do not reject the hypothesis that
the trading rule is not profitable (at the 5% level).
4.4
Data and Empirical Results
We use the same database as in Chapter 2. Table 4.1 presents the number of detected chart patterns for the extrema identification methods (M1 and M2). The
results show the apparent existence of some chart patterns in the euro/dollar
foreign exchange series. Using the first detection method (M1), we have more
detected charts in the original price series than in the simulated one for six chart
patterns (out of twelve), at the 5% significance level. When we implement the
M2 method, we detect significantly only four chart patterns, which are also significantly detected by the method M1: DT, DB, RT and RB. By looking at the
last column which represents the total number of detections, we can see that we
have more detected chart patterns when only close prices are used (M1). These
results confirm the idea that the presence of such chart patterns does not occur
by chance, at least for some chart patterns, but it is due, amongst others, to a
determined behavior of the chartist dealers.
Table 4.1: Detected chart patterns
Meth HS IHS DT
M1 78
4
7**
DB
(0.42) (0.86) (0.00)
(0.00)
40%
25%
57%** 33.%**
(0.92) (0.84) (0.00)
M2
28
(0.00)
35**
44**
(1.00) (0.45) (0.00)
(0.00)
21%
14
12**
21%
29%** 43.%**
(0.44) (0.64) (0.00)
(0.00)
TT TB RT RB BT BB TRIT TRIB
5
12
107**
89**
57*
135
(0.96) (0.91) (0.00) (0.00) (0.02) (0.09)
60%
42%
58%
73%
72%
67%
(0.51) (0.90) (0.79) (0.17) (0.39) (1.00)
16
20
24**
33**
26
57
(0.36) (0.50) (0.00) (0.00) (0.92) (1.00)
19%
35%
46%
45%
69%**
49%
(0.69) (0.34) (0.19) (0.24) (0.00) (0.10)
Σ
38
73*
617
(0.18)
(0.02)
(0.12)
76%
74%
63%
(0.38)
(0.72)
(0.66)
15
23
335
(1.00)
(1.00)
(0.50)
53%
61%
42%
(0.35)
(0.60)
(0.23)
Entries are the number of detected chart patterns (first row) and the percentage of chart patterns that reached
their price objective (third row), according to the extrema detection methods M1 and M2 (described in Appendix
B). The p-values are computed according to the procedure described in Section 4.3.4, are given in parenthesis. The
last column presents results for the whole sample, whatever is the chart pattern. ∗ ∗ and ∗ indicate respectively
significance at 1% and 5%.
The percentage of successful chart patterns (i.e. charts for which the price
objective has been met) is given by the third row of each panel in Table 4.1. For
example, 40% of Head and Shoulders (HS) detected by M1 succeed to meet their
90
Chapter4. The Performance Analysis of Technical Chart Patterns
objective, but this result is not significant since for 92% of the simulated series
we obtain more successful HS. For M1, only two charts, DT and DB present a
significant successful percentage.8 For M2, in addition to DT and DB, the chart
pattern BT presents a significant percentage of success.
Nevertheless, this measure of predictive power, i.e. the percentage of charts
that succeed to meet their price objective, is too drastic. It does not allow to
capture to what extent the price objective is not met or to what extent the price
objective is outclassed. That is why we quantified the predictability through the
ratio pred.
Table 4.2 presents the average predictability pred for all detected chart patterns which succeed or fail to meet their objectives. For example, in the case
of M1, HS has an average predictive power of 1.12. This average ratio is not
significant at 5% since for 79% of the artificial series, we obtain a higher average
ratio. However, the table shows that whatever the extrema detection method
implemented, more than one half of the whole chart pattern sample presents a
statistically significant predictability success. At the 5% significance level, predictability varies from 0.86 to 9.45. The triangle chart patterns (TRIT and TRIB)
offer the best predictability.
Table 4.2: Predictability of the chart patterns
Meth
M1
HS IHS DT DB TT TB RT RB
1.12
0.88
1.93** 0.86**
1.72
1.33
2.56
BT
3.52** 4.38**
BB TRIT TRIB
4.00
(0.79) (0.70) (0.00) (0.00) (0.41) (0.80) (0.10) (0.00) (0.00) (0.06)
M2
0.70
0.87
0.88** 1.19**
0.74
1.16
1.05
1.46* 2.52** 1.68**
(0.14) (0.14) (0.00) (0.00) (0.42) (0.08) (0.17) (0.02) (0.00) (0.00)
M1-M2
+**
+
+
-
+
+
+**
+**
+**
+**
µ
9.35*
9.45**
4.15
(0.03)
(0.01)
(0.15)
2.58
3.42
1.50
(0.43)
(0.31)
(0.10)
+**
+**
+**
This table shows the predictability of different chart patterns according to the extrema detection methods M1 and
M2 (described in Appendix B). The predictability criterion is detailed in Section 4.3.3. The last column shows
the weighted average predictability for the whole sample of charts. The p-values are computed according to the
procedure described in Section 4.3.4, are given in parenthesis. The last line of the table reports the sign of the
difference between both method’s outputs and its statistical significance according to the difference of means test.
∗ ∗ and ∗ indicate respectively significance at 1% and 5%.
These results are consistent with those obtained in Table 4.1 in which M1
exhibits more predictability. This observation is even more striking in Table
4.2. The last column shows that M1 provide on average, a predicted value more
8
Both chart patterns DT and DB have not been detected in any simulated series, whatever
the extrema detection method implemented.
4.4. Data and Empirical Results
91
than twice larger than M2. This is confirmed by positive significant signs for the
difference of means test presented in the last line of the Table 4.2. Comparatively,
Table 4.1 shows a percentage of 63% of successful chart patterns using M1 against
42% provided by M2.
Table 4.3 gives the maximum profitability that can be achieved by the use
of chart patterns. It is computed in basis points (i.e.: 1/10,000) and provided
for each of the twelve chart patterns. It corresponds to the implementation of
the trading rule related to each chart pattern whatever its success level, i.e. we
assume that the trader closes his position once the price reaches the last level
before reverting its course, (thus, the trader realizes the maximum of profits).
The maximum profit is equal to the difference, in absolute value, between the
price at the end of the chart and the minimum/maximum9 of the prices occurring
¯
¯
after the chart pattern (¯Ptf − Pa ¯). To compute these profits, we suppose that
dealers are able to buy or to sell the currency at the mid price. The computed
profits vary between 3 and 52 basis points, but are significant for only three chart
patterns: DT, DB and BT.
Table 4.3: MAXIMUM profitability of the chart patterns
Meth HS IHS DT DB TT TB RT RB BT BB TRIT TRIB
M1
8
14
9**
3**
6
9
11
13
16
14
52
37
(1.00) (0.71) (0.00) (0.00) (0.98) (0.92) (0.99) (0.76) (0.42) (0.99)
M2
10
16
10**
12**
8
15
7
12
22*
16
(0.99) (0.51) (0.00) (0.00) (0.90) (0.49) (0.98) (0.84) (0.02) (0.72)
µ
17
(0.29)
(0.84)
(0.81)
28
51
16
(0.89)
(0.64)
(0.53)
This table shows the maximum computed profit, according to the extrema detection methods M1 and M2
(described in Appendix B), expressed in basis points. The p-values are computed according to the procedure
described in Section 4.3.4, are given in parenthesis. The last column shows the weighted average maximum
profitability for the whole charts. ∗ ∗ and ∗ indicate respectively significance at 1% and 5%.
However, this profit can not be realized surely by the chartists because they
can not precisely guess if the price is at the end of its right trend or not. That is
why they adopt a strategy for their intervention according to their risk aversion.
Table 4.4 presents the results for the strategy described in Section 4.3.3. Profits
are computed as the average profit realized after the completion of each chart
pattern. This profit is statistically significant for only two charts, DT and DB
whatever the detection method implemented. However, this profit more or less
9
We adopt the minimum if the price moves down, after the completion of the chart, and we
adopt the maximum when there is an upward trend.
92
Chapter4. The Performance Analysis of Technical Chart Patterns
equal to one basis point for three cases out of four, seems too small to cover
the transaction costs. Indeed, the transaction costs are often estimated as the
observed bid-ask spread which varies on average, in the euro/dollar currency
market, between 3 to 5 basis points (Chang and Osler,1999). Consequently, even
by choosing a particular risk averse trading rule, strategies using chart patterns
seem unprofitable.
Table 4.4: Profitability of the trading strategy
Meth HS
IHS
DT DB
M1
2.80
1.10** 0.60** -0.13
1.42
TT
(1.00) (0.70) (0.00) (0.00) (0.99)
M2
0.07
0.05
1.10** 3.10** 0.99
(0.99) (0.94) (0.00) (0.00) (0.85)
M1-M2 +*
+*
-
-**
-
M1
0.48
0.54** 0.37** -0.05
TB RT RB BT
Profit
1.54
(0.92)
3.23
(0.63)
-*
1.10
1.52
2.13
1.50
(1.00) (0.98) (1.00) (1.00) (0.82)
(1.00)
(0.95)
0.70
1.87
1.25
BB TRIT TRIB µ
4.15
1.39
2.20
2.77
5.18
2.16
(0.95) (0.84) (0.27) (0.99) (0.96)
2.44
(0.85)
(0.64)
+
-
-**
-*
+
-**
-**
0.61
0.98
0.78
0.79
0.74
0.76
0.71
(0.98) (0.57) (0.92) (1.00) (0.94)
(0.99)
(0.88)
0.19
0.78
0.50
(0.91) (0.57) (0.01) (0.92) (0.89)
(0.83)
(0.55)
0.93
1.78
1.26
(0.99)
(0.95)
Sharpe
0.44
(0.99) (0.65) (0.00) (0.00) (1.00)
M2
0.01
0.01
0.26** 0.73** 0.22
(1.00) (0.95) (0.00) (0.00) (0.78)
0.50
(0.91)
0.64
(0.33)
0.51
1.30** 0.53
0.42
Xef f
M1
0.92
1.56
0.89** 0.48** -0.43
(0.99) (0.64) (0.00) (0.00) (0.99)
M2
-1.06
-1.93
0.29** 2.20** 0.13
(0.94) (0.84) (0.00) (0.00) (0.70)
1.13
(0.91)
2.01
(0.35)
1.40
1.12
1.24
2.26
(0.99) (0.96) (1.00) (1.00) (0.84)
3.19
1.18
(0.87) (0.58) (0.04) (0.93) (0.81)
0.10
1.22
3.64*
1.37
0.88
(0.75)
(0.54)
0.96
Ref f
M1
0.99
1.90
0.94** 0.50** -0.65
(1.00) (0.46) (0.00) (0.00) (0.99)
M2
-1.32
-2.59
0.32** 2.40** 0.13
(0.94) (0.81) (0.00) (0.00) (0.55)
1.23
(0.91)
2.24
(0.29)
1.89
1.31
(0.99) (0.97) (1.00) (1.00) (0.82)
1.45
(0.99)
(0.95)
0.03
3.74
1.25
(0.59)
(0.51)
1.33
1.16
3.83
1.29
1.49
2.44
0.80
(0.80) (0.54) (0.06) (0.95) (0.80)
This table includes the average profits, expressed in basis point, realized after adopting the strategy detailed in
Section 4.3.3, according to the extrema detection methods M1 and M2 (described in Appendix B). M1-M2 indicates
the computed difference results between the two methods. It shows the sign of this difference and its statistical
significance through the difference of means test. The Sharpe ratio measure the profit adjusted for risk. However,
Xef f and Ref f are also a measure of profit adjusted for risk but they take into account respectively symmetric
and asymmetric dealers risk aversion (more details for the computation of these two measures are provided in
Dacorogna, Gençay, Müller, and Pictet (2001)). The p-values, computed through a Monte-Carlo simulation, are
given in parenthesis. The last column shows the weighted average profitability for the whole charts. ∗ ∗ and ∗
indicate respectively significance at 1% and 5%.
Furthermore, the difference of means test shows that M2 is more profitable
than M1. For the majority of charts, profitability computed by adopting M2 is
significantly larger than the one provided by M1. We observe in Table 4.4 five
significant negative signs versus two positive. This observation is confirmed by
4.5. Conclusion
93
the significant negative sign for the weighted average profitability for all chart
sample, presented in the last column.
This finding is quite important since at the light of the predictability results,
we might conclude that only close prices matter. However, when the profitability
is taken into consideration, the use of high and low prices seems to have an
importance which is more in accordance with what is observed in practice (dealers
use Bar charts and only profit matters).
Nevertheless, if we consider the profit adjusted for the inherent risk, the same
two mean profits of one basis point obtained for DT have different risk levels.
Taking into account the risk level by computing the three different performance
measures; Sharpe ratio, Xef f , and Ref f , we obtain a smaller value for M2. This
means that the second method M2 generates riskier profits than M1. Moreover,
Xef f and Ref f performance measures carry out the same outputs as the Sharpe
risk-adjusted profits which implies the robustness of our results in terms of performance assessment.
4.5
Conclusion
Using five-minutes euro/dollar mid-quotes for the May 15 through November 14,
2001 time period, we shed light on the predictability and profitability of some
chart patterns. We compare results according to two extrema detection methods.
The first method (M1), traditionally used in the literature, considers only prices
which occur at the end of each time interval (they are called close prices). The
second method (M2) takes into account both the highest and the lowest price of
each interval of time. To evaluate the statistical significance of the results, we
run a Monte Carlo simulation.
We conclude on the apparent existence of some technical patterns in the
euro/dollar intra-daily foreign exchange rate. More than one half of the detected
patterns, according to M1 and M2, seem to have some significant predictive
success. Nevertheless, only two out of twelve patterns present significant profitability, which is however too small to cover the transaction costs. We also show
that the extrema detection method using high and low prices provides higher but
riskier profits than those provided by the M1 method.
To summarize, chart patterns seems to really exist in the euro/dollar foreign
94
Chapter4. The Performance Analysis of Technical Chart Patterns
exchange market at the five minute level. They also show some power to predict
future price trends. However, trading rules based upon them seem unprofitable.
4.6. Appendix
4.6
95
Appendix
A. Price Curve Estimation
Before adopting the Nadaraya-Watson kernel estimator, we tested the cubic
splines and polynomial approximations but we conclude empirically that the appropriate smoothing method is the kernel. Because the two first methods carry
out too smoothed results and they are not flexible as the kernel method.
From the complete series of the price, Pt (t = 1, . . . , T ), we take a window k of l
regularly spaced time intervals,10 such that:
© ¯
ª
Pj,k ⊂ Pt ¯ k ≤ t ≤ k + l − 1 ,
(4.6.1)
j = 1, . . . , l and k = 1, . . . , T −l+1. For each window k, we consider the following
relation:
Pj,k = m(XPj,k ) + ²Pj,k ,
(4.6.2)
where ²Pj,k is a white noise and m(XPj,k ) is an arbitrarily fixed but unknown non
linear function of a state variable XPj,k . Like Lo, Mamaysky, and Wang (2000)
to construct a smooth function in order to approximate the time series of prices
Pj,k , we set the state variable equal to time, XPj,k = t. For any arbitrary x, a
smoothing estimator of m(x) may be expressed as:
l
1X
m̂(x) =
ωj (x)Pj,k ,
l j=1
(4.6.3)
where the weight ωj (x) is large for the prices Pj,k with XPj,k near x and small
for those with XPj,k far from x. For the kernel regression estimator, the weight
function ωj (x) is built from a probability density function K(x), also called a
kernel:
Z
+∞
K(x) ≥ 0 ,
K(u)du = 1 .
(4.6.4)
−∞
By rescaling the kernel with respect to a parameter h > 0, we can change its
spread:
1
Kh (u) ≡ K(u/h) ,
h
Z
+∞
Kh (u)du = 1
(4.6.5)
−∞
and define the weight function to be used in the weighted average (4.6.3) as:
10
We fix l at 36 observations.
96
Chapter4. The Performance Analysis of Technical Chart Patterns
ωj,h ≡ Kh (x − XPj,k )/gh (x)
(4.6.6)
l
1X
gh (x) ≡
Kh (x − XPj,k ) .
l j=1
(4.6.7)
Substituting (4.6.7) into (4.6.3) yields the Nadaraya-Watson kernel estimator
m̂h (x) of m(x):
l
1X
ωj,h (x)Pj,k =
m̂h (x) =
l j=1
Pl
j=1
Pl
Kh (x − XPj,k )Pj,k
j=1
Kh (x − XPj,k )
.
(4.6.8)
If h is very small, the averaging will be done with respect to a rather small
neighborhood around each of the XPj,k ’s. If h is very large, the averaging will
be over larger neighborhoods of the XPj,k ’s. Therefore, controlling the degree
of averaging amounts to adjusting the smoothing parameter h, also known as
the bandwidth. Choosing the appropriate bandwidth is an important aspect of
any local-averaging technique. In our case we select a Gaussian kernel with a
bandwidth, hopt,j , computed by Silverman (1986):
x2
1
Kh (x) = √ e− 2h2
h 2π
hopt,k =
³ 4 ´1/5
3
σk l−1/5 ,
(4.6.9)
(4.6.10)
where σk is the standard deviations for the observations that occur within the
window k. However, the optimal bandwidth for Silverman (1986) involves a fitted function which is too smooth. In other words this optimal bandwidth places
too much weight on prices far away from any given time t, inducing too much
averaging and discarding valuable information in local price movements. Like
Lo, Mamaysky, and Wang (2000), through trial and error, we found that an acceptable solution to this problem is to use a bandwidth equal to 20% of hopt,k :
h∗ = 0.2 × hopt,k .
(4.6.11)
B. Extrema Detection Methods
Technical details for both extrema detection methods and projection procedure
are presented below:
4.6. Appendix
97
B.1 M1
M1 is the extrema detection method using the close prices. After smoothing the
data by estimating the Nadaraya-Watson kernel function, m̂h (XPj,k ), we compute
maxima and minima respectively noted by maxm̂h (XPj,k ) and minm̂h (XPj,k ) :
¯ ¡
©
¢
¡ 0
¢
ª
0
¯
maxm̂h (XPj,k ) = m̂h (XPj,k )¯ S m̂h (XPj,k ) = +1, S m̂h (XPj+1,k ) = −1
¯ ¡
¢
¡ 0
¢
ª
©
0
¯
minm̂h (XPj,k ) = m̂h (XPj,k )¯ S m̂h (XPj,k ) = −1, S m̂h (XPj+1,k ) = +1 ,
where S(X) is the sign function, equal to +1 (-1) when the sign of X is positive
0
(negative), and m̂h (XPj,k ) is the first derivative of the kernel function m̂h (XPj,k ).
By construction we obtain alternate extrema. We denote respectively by tM (m̂h (XPj,k ))
and tm (m̂h (XPj,k )) the moments correspondent to detected extrema such that:
tM (m̂h (XPj,k )) =
ª
©
j | j ∈ maxm̂h (XPj,k )
(4.6.12)
tm (m̂h (XPj,k )) =
ª
©
j | j ∈ minm̂h (XPj,k ) .
(4.6.13)
After recording the moments of the detected extrema we realize an orthogonal
projection of selected extrema, from the smoothing curve, to the original one. We
deduce the corresponding extrema to construct the series involving both maxima,
maxPj,k and minima, minPj,k such that:
³
´
maxPj,k = max PtM (m̂h (XPj,k ))−1,k , PtM (m̂h (XPj,k )),k , PtM (m̂h (XPj,k ))+1,k
´
³
minPj,k = min Ptm (m̂h (XPj,k ))−1,k , Ptm (m̂h (XPj,k )),k , Ptm (m̂h (XPj,k ))+1,k .
For each window k we get alternate maxima and minima. This is assured by the
bandwidth h which provide at least two time intervals between two consecutive
extrema. The final step consists to scan the extrema sequence to identify an
eventual chart pattern. If the same sequence of extremum was observed in more
than one window, only the first sequence is retained for the recognition study to
avoid the duplication of results.
B.2 M2
M2 is the extrema detection method built on high and low prices. According
to this method, maxima and minima have to be detected onto separate curves.
Maxima on high prices curve and minima on the low one.
98
Chapter4. The Performance Analysis of Technical Chart Patterns
Let Ht and Lt (t = 1, .., T ), be respectively the series for the high and the how
prices, and k a window containing l regularly spaced time intervals such that:
© ¯
ª
Hj,k ⊂ Ht ¯ k ≤ t ≤ k + l − 1
(4.6.14)
© ¯
ª
Lj,k ⊂ Lt ¯ k ≤ t ≤ k + l − 1 ,
(4.6.15)
j = 1, . . . , l and k = 1, . . . , T − l + 1. We smooth these series through the kernel
estimator detailed in Appendix A to obtain m̂h (XHj,k ) and m̂h (XLj,k ). We detect
maxima on the former series and minima on the latter one in order to construct
two separate extrema series maxm̂h (XHj,k ) and minm̂h (XLj,k ) such that:
¯ ¡
©
¢
¡ 0
¢
ª
0
¯
m̂h (XHj,k ) ¯ S m̂h (XHj,k ) = +1 , S m̂h (XHj+1,k ) = −1
¯ ¡
¢
¡ 0
¢
ª
©
0
¯
= m̂h (XLj,k ) ¯ S m̂h (XLj,k ) = −1 , S m̂h (XLj+1,k ) = +1 ,
maxm̂h (XHj,k ) =
minm̂h (XLj,k )
where S(x) is the sign function defined in the previous Section.
We record the moments for such maxima and minima, denoted respectively by
tM (m̂h (XHj,k )) and tm (m̂h (XLj,k )) and we project them on the original high and
how curves to deduce the original extrema series maxHj,k and minLj,k , such that:
³
´
maxHj,k = max HtM (m̂h (XHj,k ))−1,k , HtM (m̂h (XHj,k )),k , HtM (m̂h (XHj,k ))+1,k
´
³
minLj,k = min Ltm (m̂h (XLj,k ))−1,k , Ltm (m̂h (XLj,k )),k , Ltm (m̂h (XLj,k ))+1,k .
However, this method does not guarantee alternate occurrences of maxima and
minima. It is easy to observe, in the same window k, the occurrence of two consecutive minima on the low series before observing a maximum on high series. To
resolve this problem we start by recording the moments for the selected maxima
on high curve, tM (Hj,k ), and minima in low curve, tm (Lj,k ). Then we select, for
window k the first extremum from these two series, E1,k , and its relative moment,
tE1,k , such that:
E1,k
¡
¢
tE1,k = min tM (Hj,k ) , tm (Lj,k )
t
¢
¡
= {maxHj,k } ∪ {minLj,k } | j = tE1,k .
(4.6.16)
(4.6.17)
If we meet a particular case such that a minimum and a maximum occur at
the same first moment, then we retain arbitrarily the maximum. To build the
4.6. Appendix
99
alternate series, we have to know the type of the last extremum introduced into
the series. If it is a maximum (minimum) then the next extremum has to be a
minimum (maximum) selected from the low (high) series such that:
¯
©
¡
¢
ª
¯
= minLj,k ¯ j = min tm (Lj,k ) , tm (Lj,k ) > tE(j−1),k
}
j,k
¯
©
¡
¢
ª
¯
= maxHj,k ¯ j = min tM (Hj,k ) , tM (Hj,k ) > tE(j−1),k ,
E ¯¯
j,k E(j−1),k ∈{maxH
E ¯¯
j,k E(j−1),k ∈{minL
j,k
}
where Ej,k is the extremum detected on original series.
Finally, the obtained series is scanned by the recognition patterns algorithms to
identify an eventual chart pattern.
C. Graphs and Definitions of Chart Patterns
We present in this appendix all the definitions corresponding to each chart pattern, however, we exhibit only the configurations corresponding to IHS, TRIB
and TRIT chart patterns. The configurations related to the rest of the technical
chart patterns are displayed in Chapter 5.
C.1 Inverse Head and Shoulders (IHS):
IHS is characterized by a sequence of 5 extrema Ei (i = 1, .., 5) such that:























IHS ≡
E3 < E1 , E3 < E5
|p(E2 , E4 )| ≤ tg(10)
|p(E1 , E5 )| ≤ tg(10)
VtE (E2 ,E4 )−E1
1
≤ 1.1
VtE (E2 ,E4 )−E5

5




1.1 ≤ hs ≤ 2.5




t 2 −td

1

≤ tfE−t
≤2

2

E4


t −t

1

≤ E4 m E2 ≤ 2

2



¢
 ¡
Ptmax − Ptd ≥ 23 × h
where
- h is the height of the head :
h = VtE3 (E2 , E4 ) − E3
E1 < E2
0.9 ≤
100
Chapter4. The Performance Analysis of Technical Chart Patterns
- s is the height average of the two shoulders : s =
(VtE (E2 ,E4 )−E1 )+(VtE (E2 ,E4 )−E5 )
1
2
5
- Ptmax is the highest price observed into the [td−(f −d) , td ] :
¯
Ptmax = max(Pt ) ¯ td−(f −d) ≤ t ≤ td
- td is the starting time for the pattern
- tf is the ending time for the pattern
- td−(f −d) = td − (tf − td )
- tf +(f −d) = tf + (tf − td )
- m is the average time which the shoulders take for their total completion
C.2 Double Top (DT):
DT is characterized by a sequence of 3 extrema Ei (i = 1, .., 3), such that11 :



E1 > E2





E1 −E2

=1

VtE (E1 ,E3 )−E2

2



E3 −E2

=1
VtE (E1 ,E3 )−E2
2
DT ≡
t 2 −td

1

≤ (tfE−t
≤2


2
d )/2



tf −tE2
1


 2 ≤ (tf −td )/2 ≤ 2



 ¡Pt − Pt ¢ ≥ 2 × ¡Vt (E1 , E3 ) − E2 ¢
min
E2
d
3
11
5.1.
Graphs corresponding to DT,DB, TT, TB, RT, RB, BT and BB are displayed in Figure
4.6. Appendix
101
C.3 Double Bottom (DB):
DB is characterized by a sequence of 3 extrema Ei (i = 1, .., 3), such that :



E1 < E2





E2 −E1

=1

E2 −VtE (E1 ,E3 )

2



E2 −E3

=1
E2 −VtE (E1 ,E3 )
2
DB ≡
t 2 −td

1

≤ (tfE−t
≤2


2
d )/2



t −t 2
1

≤ (tff−tdE)/2
≤2


2


¡
¢
¡
¢

 Pt
≥ 32 × E2 − VtE2 (E1 , E3 )
max − Ptd
C.4 Triple Top (TT):
T T is characterized by a sequence of 5 extrema Ei (i = 1, .., 5) such that:
























TT ≡
E1 > E2
| p(E1 , E5 ) | ≤ tg(10)
| p(E2 , E4 ) | ≤ tg(10)
0.9 ≤
h
E1 −VtE (E2 ,E4 )
≤ 1.1
1
0.9 ≤ E5 −Vt h (E2 ,E4 ) ≤ 1.1

E5


t
−t

E
d
1
2

≤ (tf −t
≤2


2
d )/3



t −t 2
1

≤ (tEf 4−tdE)/3
≤2


2



t −t 4
1


≤ (tff−tdE)/3
≤2

2


¢
¡

 Pt − Pt
≥ 23 × h
min
d
C.5 Triple Bottom (TB):
T B is characterized by a sequence of 5 extrema Ei (i = 1, .., 5) such that:
102
Chapter4. The Performance Analysis of Technical Chart Patterns
























TB ≡
E1 < E2
| p(E2 , E4 ) | ≤ tg(10)
| p(E1 , E5 ) | ≤ tg(10)
0.9 ≤
h
VtE (E2 ,E4 )−E1
≤ 1.1
1
0.9 ≤ Vt (E2h,E4 )−E5 ≤ 1.1

E5


tE2 −td

1

≤ (tf −td )/3 ≤ 2

2




t −t 2
1

≤ (tEf 4−tdE)/3
≤2


2



t −t 4
1


≤ (tff−tdE)/3
≤2

2


¡
¢

 Pt
≥ 23 × h
max − Ptd
C.6 Rectangle Top (RT):
RT is characterized by a sequence of 6 extrema Ei (i = 1, .., 6) such that:


E1 > E2






| p(E1 , E5 ) | ≤ 0.001





 | p(E2 , E6 ) | ≤ 0.001
RT ≡
VtE (E1 ,E5 )
3

=1


E3



E4

=1


VtE (E2 ,E6 )

4

¡
¢


Ptd − Ptmin ≥ 32 × h
C.7 Rectangle Bottom (RB):
RB is characterized by a sequence of 6 extrema Ei (i = 1, .., 6) such that:


E1 < E2






| p(E2 , E6 ) | ≤ 0.001





 | p(E1 , E5 ) | ≤ 0.001
RB ≡
E3

=1


VtE (E1 ,E5 )

3


VtE (E2 ,E6 )

4

=1

E4


¢ 2
¡

 P
tmax − Ptd ≥ 3 × h
4.6. Appendix
103
C.8 Broadening Top(BT):
BT is characterized by a sequence of 5 extrema Ei (i = 1, .., 5) such that:



E > E2

 1
BT ≡
E3 > E1 , E4 < E2 , E5 > E3



 ¡P − P ¢ ≥ 2 × h
td
tmin
3
C.9 Broadening Bottom(BB):
BB is characterized by a sequence of 5 extrema Ei (i = 1, .., 5) such that:



E < E2

 1
BB ≡
E3 < E1 , E4 > E2 , E5 < E3


¢

 ¡P
−P ≥ 2 ×h
tmax
td
3
C.10 Triangle Top(TRIT):
T RIT is characterized by a sequence of 4 extrema Ei (i = 1, .., 4) such that:



E1 > E2






p(E1 , E3 ) ≤ tg(−30)


1 ,E3 )|
T RIT ≡
0.9 ≤ |p(E
≤ 1.1
p(E2 ,E4 )




 tf ≤ tE1 + 0.75 × (tint − tE1 )




 ¡P − P ¢ ≥ 2 × h
tE1
tmin
3
where tint is the moment of support and
resistance lines interSection:
¡
¢
tint = mint Vt (E1 , E3 ) ≤ Vt (E2 , E4 ) , t > tE4 .
104
Chapter4. The Performance Analysis of Technical Chart Patterns
C.11 Triangle Bottom(TRIB):
T RIB is characterized by a sequence of 4 extrema Ei (i = 1, .., 4) such that:



E1 < E2






p(E2 , E4 ) ≤ tg(−30)


2 ,E4 )|
T RIB ≡
0.9 ≤ |p(E
≤ 1.1
p(E1 ,E3 )





tf ≤ tE1 + 0.75 × (tint − tE1 )



¡
¢ 2

 P
tmax − PtE1 ≥ 3 × h
where tint is the moment of support and resistance lines interSection:
¡
¢
tint = mint Vt (E2 , E4 ) ≤ Vt (E1 , E3 ) , t > tE4 .
Chapter 5
The Technical Signal Based
Trading Effects on Volatility
5.1
Introduction
Technical chart patterns, as shown in Chapter 4, may provide predictive success
and small profitability. They provide also the signals of buying or selling financial
assets, for instance foreign currency. This signal is the output of the technical
analysis considered as an important information that all technical traders rely
on. This kind of information could have, consequently, a significant impact on
the exchange rate movements.
There is a clear consensus that speculative activity triggered by technical
traders and built on technical analysis, is used mainly as a guide to short term
exchange rate behavior (Black, 1986, Allen and Taylor, 1990, Allen and Taylor,
1992, Osler, 2003, and Shiller, 2003). Technical traders generally employ a wide
variety of technical trading tools, involving visually identifying recurring patterns,
trend identification formulas, trend reversal signals, and genetic algorithms. They
use extrapolation to predict the future behavior of prices, since they have only
the time series in their information set, and they consider that prices embody all
aspects of the market, balancing all the forces of supply and demand. In efficient
markets, economists consider the chartists as noise traders since they base their
trades on noise considered by chartists as ”technical information”.
Using a theoretical investigation based on the misperception of the expected
risky asset price by the noise traders, De Long, Shleifer, Summers, and Waldmann (1990) show that noise traders’ price misperception affects future prices
(by driving them up if they are optimistic about the asset or down when they
are pessimistic) and increases volatility. Jeanne and Rose (2002) present a model
of exchange rate regimes based on the presence of noise traders. They find that
105
106
Chapter5. The Technical Signal Based Trading Effects on Volatility
in the floating exchange rate regimes, intense noise trading, stemmed from the
numerous noise trader entries in the market, increases exchange rate volatility.
In addition, and contrary to Friedman (1953)’s argument, Black (1986) and
Frankel and Froot (1990) show that trading based on technical signals leads to
excessive volatility. Carlson and Osler (2000) however, find that speculators’
impact on exchange rate volatility varies according to the type of shocks hitting
the market. Some shocks, such as changes in liquidity demand, do not increase
volatility. Other shocks, such as changes of interest rates or risk, induce a rise in
volatility.
This chapter presents empirical evidence on the theoretical results previously
found by De Long, Shleifer, Summers, and Waldmann (1990) and Jeanne and
Rose (2002). It examines the effects of noise trading or the technical signal based
trading (TSBT) on foreign exchange (FX) rate volatility.
We use the basic technical signals, essentially the chart patterns trading signals, and we implement the same methodology as Lo, Mamaysky, and Wang
(2000) and Ben Omrane and Van Oppens (2006) to recognize chart patterns (see
Chapter 4 for more details). Moreover, we check for the impact of four scheduled news announcements since they present a positive pre-announcement effect
highlighted in Chapter 2.
Based on our euro/dollar time series, news announcements database, and
chart pattern signals, we study the volatility dynamics around TSBT. We find
that (1) volatility drops during the completion of technical chart patterns, before the occurrence of the chart signal, where the exchange rate moves within
a resistance and support levels;1 (2) when the technical signal occurs, it generates volatility increases. A breakout takes place once the exchange rate crosses
the support or the resistance level just after the chart completion, attracting
the attention of technical traders and creating heterogeneity among the market
participants.
The remainder of the chapter contains four sections and a conclusion. Section
5.2 provides a review of the principal theoretical and empirical studies which
1
Pring (1985) defines the support level as a zone representing a concentration of demand,
and the resistance area as a zone corresponding to a concentration of supply. Murphy (1986)
defines the support level, otherwise, as the area on the chart under the market where buying
interest is sufficiently strong to overcome selling pressure. As a result, a decline is halted and
prices turn back again. Resistance level is the opposite of support.
5.2. Technical Signal Based Trading
107
have examined the interaction between rational and technical trading forces in
the foreign exchange market. We mean by rational trading, the trading based
on typical information rather than noise or technical one. Typical information
is information provided by electronic screens such as Reuters. In Section 5.3 we
explain our methodology to study the volatility dynamics around TSBT. Section
5.4 describes the data and gives some details about the chart pattern recognition
algorithms. Section 5.5 presents the model and provides a discussion about the
estimation results. We conclude in Section 5.6.
5.2
Technical Signal Based Trading
TSBT is considered by economists as noise trading, and is generated by the
technical traders. It has been widely discussed since the eighties by several studies
which examine the ability of such trading to increase volatility. Black (1986),
Frankel and Froot (1990), and recently Shiller (2003) focus on financial markets
anomalies that might be triggered by noise trading. Carlson and Osler (2000)
examine the speculators’ effect on exchange rate volatility with respect to the
types of the shocks hitting the market. Lo, Mamaysky, and Wang (2000) study
the effect of chart patterns on the stock return distribution. Based on order
clustering patterns in executed orders, Osler (2003) proposes an explanation for
the price behavior through support and resistance levels. Regarding the studies
which are focused on both rational trading and TSBT, Frankel and Froot (1986)
and Allen and Taylor (1990) present a comparison between typical information
and technical analysis. Allen and Taylor (1989,1992) and Lui and Mole (1998b)
report the results of questionnaire surveys on the influence of chartism and the use
by foreign exchange dealers, respectively in London and Hong Kong, of rational
and technical analysis to form their forecasts.
5.2.1
Noise Trading and Volatility
Black (1986) considers technical traders as noise traders and defines noise trading
as the trading on noise as if it were information. He notes that prices embody
information that both information and noise traders trade on. Thus, it is difficult
for both kinds of traders to know if they are trading on information or noise.
He adds that the short term volatility of price will increase since there is noise
trading. Frankel and Froot (1990) confirm Black’s arguments and show that
108
Chapter5. The Technical Signal Based Trading Effects on Volatility
trading volume is influenced by the importance of heterogenous expectations
and might be triggered by trading based on noise rather than news, leading in
turn to excessive volatility.
De Long, Shleifer, Summers, and Waldmann (1990) present a model that
considers two categories of traders, noise and rational traders. Noise traders
misperceive the expected price of the risky asset by an independent and identically distributed normal random variable of which first moment measures the
average misperceptions of the noise traders. Its second moment measures the
variance of noise traders’ misperception of the expected return, it gauges the risk
that noise traders’ beliefs will not revert to their mean and might become even
more extreme. Rational traders, though try to counter noise traders by betting
against them. They buy assets when noise traders depress prices and sell when
noise traders push prices up. Such active contrarian trading strategies may push
prices toward the fundamental value, but often fail to reach this objective. Since
rational traders are risk averse they limit their arbitrage against noise traders,
because they are not able to bear the noise traders’ risk in addition to the fundamental risk. The more the price is far from the fundamental value the more
there is excess volatility (Shiller, 2003). Still, after maximizing the expected utility functions for noise and rational traders the same authors get an equilibrium
price as a function of four factors:
• The fundamental value.
• The variance of noise traders’ misperceptions. The price varies substantially
as noise traders’ opinion shift. When the majority of noise traders are
bullish, they bid up the price, when they are bearish they bid it down.
When they hold their average misperception, they do not at this level affect
the price.
• The deviation of the price from its fundamental value. If the traders are
bullish on average, this ‘price pressure’ effect makes the price of the risky
asset higher that it would otherwise be.
• A factor interpreted as a compensation for rational traders to bearing the
risk that noise traders will become bearish (if they were bullish) and the
price of the risky asset will fall. Except the first above component of the
5.2. Technical Signal Based Trading
109
price, the three others are weighted by the proportion of noise traders
dealing within the market. As a consequence, the more numerous noise
traders are relative to rational ones, the more volatile asset prices are.
Jeanne and Rose (2002) uses the same model of De Long, Shleifer, Summers,
and Waldmann (1990) but they adapt it to the foreign exchange context. They
mix elements from the latter model with the macroeconomic theory of exchange
rate determination. By maximizing the expected utility functions of noise and
rational traders trading within the currency and bond markets, they get, at the
equilibrium, the following results: 1) an exogenous increase of the number of
noise traders in the market drives exchange rate volatility up which triggers a
raise in the risk premium. 2) When the market features a low volatility and low
risk premium then it does not offer noise traders enough gain to induce many of
them to enter the market. However, when volatility and the risk premium are
high, more traders are attracted to the market because it does offer large gains.
Shiller (2003) explains the notion of excess volatility through the feedback
theory. This theory is built on the mimicking behavior of different market participants. When prices go up, creating success for some traders, this may attract
other market participants attention and heighten expectations for further price
increases. This process in turn enhances the demand and thus generates another
round of price increases. If the feedback is not interrupted, it may produce after many rounds a speculative ”bubble”, in which high expectations for further
price increases support very high current prices. The high prices are ultimately
not sustainable, since they are high only because of the expectation of further
price increases. Then, the bubble bursts and the prices start falling down. The
same feedback may also produce a negative bubble, downward price movements
propelling further downward price drops, until the price reaches an unsustainable
low level.
Osler (2003) gives more details about the unsustainable levels in the exchange
rate market. She shows that downtrends and up-trends tend to reverse their
course respectively at support and resistance levels, which can be identified exante and which are often round numbers. She finds also that trends tend to be
unusually rapid after rates cross support and resistance levels. Based on price
contingent orders data-set, she offers a prima facie evidence that order clusters
110
Chapter5. The Technical Signal Based Trading Effects on Volatility
just before round numbers are dominated by take-profit orders. Thus, price
trends might reverse course when they hit take-profit dominated orders. However,
order clusters beyond round numbers are dominated by stop-loss orders. She
shows, furthermore, that the large stop-loss buy (sell) orders cluster just above
(below) round numbers, thus could explain why prices tend to move rapidly
once crossing round numbers. Osler (2002) provides evidence that the rapid
trends after round numbers are derived from stop-loss order clusters, and not
from central bank interventions or any other possible source suggested in the
literature.
Carlson and Osler (2000) show that the TSBT effect on exchange rate volatility varies according to the type of shocks hitting the market. Some shocks, such
as changes in liquidity demand, have no direct impact on speculators’ preferred
portfolio position. Other shocks, such as changes of interest rates or risk, do
directly change speculators’ preferred portfolio positions. Then, increasing speculators’ activity induces a raise in volatility.
Still on the TSBT literature, Lo, Mamaysky, and Wang (2000) propose a
systematic approach to recognize technical patterns using a nonparametric kernel
regression. They show that some technical chart patterns do provide incremental
information which affects the conditional distribution of returns, especially for
the Nasdaq stocks.
5.2.2
TSBT and News
Regarding the effect on prices of both typical information and TSBT, a number
of studies compare the performance of each of these sources in terms of forecasting. Other studies investigate the influence and the use of each of them
by foreign exchange traders. Allen and Taylor (1990) compare the accuracy of
chartist predictions with various economic and statistical approaches, using the
root mean square error (RMSE) of the forecasts of each as a performance measure. They find that chartist views generate a lower RMSE than the one carried
out by ARIMA, vector auto-regressions (VAR),2 and the random walk. Frankel
and Froot (1986) propose a model involving three categories of actors: rational
traders, chartists and portfolio managers. The latter form their expectations as a
2
They estimate two types of fourth-order VAR; 1) an ”economic ” VAR based upon the
exchange rate, the interest rate differential and relative stock market performance; 2) a VAR
involving only three currencies quoted against US-Dollar.
5.3. Methodology and Hypotheses
111
weighted average of the predictions of both former actors. They show that both
rational and technical analysis represent competitive forces within the mind of
a single representative agent. However, this finding is in contrast with those of
Allen and Taylor (1989,1992) and Lui and Mole (1998b). These researchers have
conducted questionnaire surveys in order to study the use of both rational and
technical analysis by foreign exchange traders. They found that more than 85% of
the respondents use both typical information and technical signals to form their
expectations. At short horizons, intra-day to one week, there exists a propensity
towards technical analysis, with 60% judging charts to be at least as important
as typical information. At longer forecast horizons, from one to three months or
six months to one year, there is a more pronounced propensity towards typical
information, with nearly 30% of respondents relying on pure typical information
and 85% judging it to be more important than charts. Allen and Taylor (1989)
point out also that less than 8% of respondents consider the two approaches to
be competing to the point of being mutually exclusive while the rest think the
approaches are complementary to some degree. They find moreover, that only
2% are pure chartists, and never use typical information at any horizon.
5.3
Methodology and Hypotheses
With respect to the previous literature on FX volatility, the aim of this chapter is
to study the volatility dynamics around TSBT. By using the same chart patterns
(except two) considered in the previous chapter we investigate volatility dynamics
during and after the chart completion.
We focus on four pairs of chart patterns. They only involve configurations
where the exchange rate clusters within two borders corresponding to support
and resistance levels. We consider two categories of technical traders. The first
one believes that the price reverses its course at support and resistance levels.
These levels correspond respectively to a concentration of demand and supply.
The second one involves the chartists who believe in the chart predictive success
and profitability. Technical chart patterns selected to deal with are: Broadening
Bottom (BB), Broadening Top (BT), Double Bottom (DB), Double Top (DT),
Rectangle Bottom (RB), Rectangle Top (RT), Triple Bottom (TB), and Triple
Top (TT). Figure 5.1 shows the configuration corresponding to each chart pat-
112
Chapter5. The Technical Signal Based Trading Effects on Volatility
tern. Just after the completion of the chart, the exchange rate can evolve in two
directions. It may cross the support or the resistance level. It could also reverse
its course and pursue its trend within the charts’ borders. In this case the chart
fails to meet its predictive goal. We point out that we consider only technical
chart patterns in which the exchange rate crosses the borders.
5.3. Methodology and Hypotheses
113
Figure 5.1: The Geometric Configuration for the technical charts
The four pairs of technical chart patterns: Broadening Bottom (BB), Broadening Top
(BT), Double Top (DT), Double Bottom (DB), Rectangle Bottom (RB), Rectangle
Top (RT), Triple Top (TT) and Triple Bottom (TB).
114
Chapter5. The Technical Signal Based Trading Effects on Volatility
Nevertheless, the objective of our study is not to check for the predictive success of different charts (for this see Chapter 4. We want to examine the volatility
dynamics during and after the completion of the chart, despite its success or
failure to fulfill its prediction.3 According to Osler (2002, 2003), stop-loss orders placed just after support and resistance levels trigger a rapid downtrend
(up-trend) in exchange rate. This means that volatility could increase once the
exchange rate crosses these levels. Black (1986) and Frankel and Froot (1986) attribute the swings in volatility to the heterogeneous expectations among traders,
whereas Admati and Pfleiderer (1988), Degennaro and Shrieves (1997), Andersen and Bollerslev (1998), Evans and Lyons (1999), Melvin and Yin (2000), and
Bauwens, Ben Omrane, and Giot (2005) argue about the positive impact of public
and private information.
Moreover, Shiller (2003) explains how market participants rely on mimicking
behavior to trade. In such a case, they create a homogeneous behavior which
could lead to a drop on volatility. De Long, Shleifer, Summers, and Waldmann
(1990) and Jeanne and Rose (2002) emphasize that the more noise traders there
are the more volatility created. Black (1986) notices that market participants
have to go along the herd, otherwise they lose money. Moreover, Friedman
(1953) claims that rational speculation reduces volatility, since irrational speculators regularly lose money and they will be driven out of the market by rational
speculators with more successful strategies. In turn, volatility drops, since rational speculation can’t be destabilizing.
Based on the above literature, our hypotheses regarding volatility dynamics
before and after the TSBT can be stated as follows:
H1 : If technical chart pattern signals attract the attention of technical traders
then the market activity should rise when the chart signal occurs.
When technical traders focus their attention on the technical chart signal, they
take positions according to the signal. Thus they increase their order flow and
amplify the activity within the market.
H2 : During the chart completion, before the occurrence of the technical chart
3
We recall that the technical signal is located at the intersection of the end of the chart and
the border, it exhibits information about the future currency trend.
5.4. Data Description
115
signal, the market is dominated by typical information trading, featuring homogeneous behavior that could trigger a volatility drop.
There is no excess volatility before the chart signal, because the market participants feature a homogeneous behavior. The market is dominated by rational
dealers and could involve a minority of technical dealers. However, the latter category of dealers have to go along the herd, otherwise they lose money. Moreover,
they behave homogeneously since there is no signal that could trigger divergent
objectives.
H3 : At the chart completion, as the technical signal occurs, it could attract the
attention of technical traders. TSBT takes place once the exchange rate crosses
the support or the resistance level triggering an increase in volatility.
When the signal occurs, the market becomes dominated by the two categories
of technical traders who don’t share the same objectives. The first category, who
believes in the price reverse course at the support and resistance levels, place
stop-loss orders beyond these levels and generate an acceleration of the exchange
rate trend (Osler, 2003). The second category, the chartists, who believe in
the basic chart patterns predictive success, place take-profit orders instead of
stop-loss ones (Murphy, 1999, Béchu and Bertrand, 1999, Lo, Mamaysky, and
Wang, 2000, and Ben Omrane and Van Oppens, 2006). Clearly both categories
of technical traders have divergent objectives and create heterogeneity within the
market. Moreover, there is obviously a third category involving mainly rational
traders who trade only on typical information. The meeting of three divergent
behaviors heightens the degree of heterogeneity and feeds a boost in volatility.
5.4
5.4.1
Data Description
The euro/dollar Exchange Rate and News Announcements Data
In Chapter 2 we showed that volatility increases in the pre-announcement periods,
particulary before scheduled events. In order to control for this effect and avoid
the omitted variable bias, we introduce news announcement variables in our
study. News variables involve only scheduled events (i.e. news for which the day
116
Chapter5. The Technical Signal Based Trading Effects on Volatility
and the time of announcement are displayed in advance on Reuters economic
agenda ).4
5.4.2
The Chart Pattern Data
The identification of technical chart patterns is done through the period ranging from May 15 to November 14, 2001. The goal is to recognize some exchange
rate movements which contribute to the formation of specific chart patterns. The
recognition process starts by identifying the sequences of local extrema, and ends
by doing the correspondence between the different sequences and the quantitative specification of the four pairs of chart patterns.5 Once the chart pattern is
detected, we identify its starting time, its completion time, and we compute its
duration. Then, we use this information to create the dummy variables corresponding to the pre- and post- completion periods relative to the chart pattern.
Table 5.1: Detected chart patterns
Detection
p-value
Mean Duration
BB
28
9%
25
BT
18∗
2.0%
25
DB
14∗ ∗
0.0%
19
DT
11∗ ∗
0.0%
16
RB
10∗ ∗
0.0%
29
RT
22∗ ∗
0.0%
20
TB
1
91%
35
TT
1
96%
23
Entries are the number of detected chart patterns. The p-values computation is showed
in Chapter 4. The last row presents the mean duration of each chart pattern. ∗ ∗ and ∗
indicate respectively significance at 1% and 5%.
Table 5.1 reports the number of detected chart patterns as well as their durations. Clearly the number of detected charts is lower than the one found in the
previous chapter, although the same recognition algorithms are applied to the
same data. In fact, the recognition algorithms implemented in this study involve
more constrains related to the chart pattern completion.6 That is why we detect
a lower number of charts.
The BB chart pattern presents the maximum number of detection. However,
the recognition process identifies only one TT and one TB chart pattern. Regarding the p-values, there is overwhelming significance for DB, DT, RB and RT
chart patterns, with p-values that are very close to zero. In contrast, the TB
4
See Chapter 2 for more details.
The quantitative specifications of the four pairs of chart patterns are presented in detail in
Chapter 4.
6
The constrains correspond to the support and resistance levels.
5
5.5. Models and Empirical results
117
and TT chart patterns carry out no statistical significance (at 5% significance
level). The duration of different charts (i.e. the time taken by the chart from its
beginning to its completion) ranges from 16 to 35 time intervals (of five minutes).
5.5
5.5.1
Models and Empirical results
Models
We start by studying the sensitivity of the quoting activity, taken as a proxy for
market activity, to the chart signals. Our goal is to check if the chart signals
attract indeed the attention of chartists. In such a case, chartists could initiate
TSBT and generate more activity within the market. The post-completion period
corresponding to the technical chart pattern j is represented through the dummy
variable Chartpost
j, t taking the value 1 five-minute after the signal and 0 otherwise.
The model to be estimated is
qt = c 0 +
8
X
j=1
ρj Chartpost
j, t
+
2
X
i=0
P
X
λi εt−i +
(δc,p cos yt,p + δs,p sin yt,p ) + κ ht−1 ,
p=1
(5.5.1)
where λ0 = 1 and εt is the error term. We use the quoting activity qt as a
proxy for the market activity. To control for the seasonality we introduce the
flexible Fourier form, and to control for the effect of volatility on the activity
we introduce the variable ht computed hereafter through Equation (5.5.4). This
variable is lagged to avoid the simultaneity problem.
The variable yt,p is equal to 2πp nk /Nk and Nk indicates the number of hours
per day: Nk = 288, for all open days of the week except for Fridays (N5 = 264),
and nk takes the values 1, 2, . . . , Nk . The FFF order is limited at P = 4.
The second step of our approach consists in studying the volatility dynamics
around TSBT. As in Chapter 2 we use the EGARCH model of Nelson (1990)
to model the SA returns (denoted rt ) and their conditional variance, denoted by
ht (to measure volatility). The level of returns is modeled by a moving average
process of order two to account for the detected autocorrelation, such that :
rt = θ0 + ut + θ1 ut−1 + θ2 ut−2 ,
(5.5.2)
118
Chapter5. The Technical Signal Based Trading Effects on Volatility
and the error term ut by an EGARCH(2,2) process:
p
ut = ht ²t ,
ln ht
2 ³
h
i
´
X
p
= ω+
βi ln ht−i + αi | ²t−i | − 2/π + γi ²t−i
+
i=1
8
2
XX
ηj,τ Chartj,τ,t +
j=1 τ =1
4
X
(5.5.3)
(5.5.4)
ϕn dn,t + φ ast−1 .
n=1
The innovations ²t are assumed identically and independently distributed. To
estimate the model by the quasi maximum likelihood (QML) method, we proceed
as if their distribution was normal. The variable denoted Chartj,τ,t is a dummy
variable corresponding to the chart pattern of category j identified during the
period τ , relative to the five-minute return time t. The index τ indicates an
observation window: a period corresponding to the chart pattern completion (τ =
1), and a period just after the completion (τ = 2). The variable ast corresponds
to quoting activity. The first observation window corresponds to the time period
in which the chart pattern evolves from its beginning until its completion, and
the second to the five minutes occurring just after the completion. The variable
denoted dn,t is also a dummy variable corresponding to the pre-announcement
scheduled news of category n.
The coefficients of Equation (5.5.1) allow to test our first hypothesis (H1 ).
The coefficients of Equation (5.5.4) allow to test the hypotheses H2 and H3.
5.5.2
Empirical Results
To evaluate H1, we test the hypothesis that ρj is equal to zero against positive
ρj (since we guess a positive impact). Rejecting the null hypothesis means that
the chart signals attract the attention of chartists who could generate TSBT
and consequently increase market activity. To test H2, the null hypothesis is
that ηj,1 = 0 for all j, and the alternative is that ηj,1 < 0 (since we guess a
negative impact). Rejecting the null implies that volatility drops, before TSBT,
during the chart pattern completion. Finally, to test H3, the null hypothesis is
ηj,2 = 0 for all j, and the alternative is ηj,2 > 0 (since we guess a positive impact).
Rejecting the null implies that the post-completion period for the technical chart
feeds TSBT that generates an increase in volatility.
5.5. Models and Empirical results
119
Table 5.2: The TSBT effects on quoting activity
qt = c0 +
+
Coefficient
c0
λ1
λ2
κ
ρ1
ρ2
Obs.
P4
P8
p=1 (δc,p
j=1 ρj
Chartpost
j, t +
P2
i=0 λi εt−i
cos yt,p + δs,p sin yt,p ) + κ ht−1
Estimation
P-Value
0.826∗ ∗
0.737∗ ∗
0.429∗ ∗
0.088∗ ∗
0.086∗
0.023
37 650
(%)
0.0
0.0
0.0
0.0
1.7
68
R2
Coefficient
ρ3
ρ4
ρ5
ρ6
ρ7
ρ8
83.30%
Estimation
(1)
P-Value
(%)
0.062
37
0.029
55
0.175∗ ∗
0.4
-0.025
44
-0.023
45
0.660∗ ∗
0.0
W (ρj = 0) 726.89∗ ∗
∗∗
and ∗ indicate respectively significance at 1% and 5%.
Estimation by the quasi maximum likelihood method with the Newey-West HAC
standard errors. To save space, coefficients for FFF are not shown. All the FFF
coefficients are statistically significant at 1% level. W (ρj = 0) is the Wald statistic
for the hypothesis of nullity of the 8 coefficients ρj . j = 1, 2, . . . , 8 corresponds
respectively to the chart patterns BB, BT, DB, DT, RB, RT, TB, and TT..
qt is the quoting activity. Chartpost
j, t is a dummy variable corresponding to the
post-completion period for the chart category j.
Tables 5.2 and 5.3 report the estimation results of the TSBT effects respectively on quoting activity and volatility. Table 5.2 shows positive estimated
coefficients for six dummies (out of eight) of which three are statistically significant (two at 1% and one at 5%). TSBT occurring just after the completion of
BB, RB and TT chart patterns triggers an increase in quoting activity. This
result is out of the effect of market volatility and means that some technical signals attract the attention of technical traders that trigger an increase of market
activity. Clearly, we do not reject H1, since the joint hypothesis of nullity of the
8 ρj coefficients is rejected at the 1% level.
Table 5.3 displays the estimation results for Equations (5.5.2)-(5.5.4). The
standardized residuals and squared residuals are not autocorrelated (according to
Q-statistics). The EGARCH coefficients are significant (except for the asymmetry effects) and compatible with a stationary process. The table shows, moreover,
negative significant estimated coefficients for dummies corresponding to the chart
completion period, and positive significant coefficients for the post-completion period. Clearly, H2 and H3 are not rejected. Volatility decreases by 25 percent,
during the completion of a DB chart pattern. It drops by 20 percent for DT,
120
Chapter5. The Technical Signal Based Trading Effects on Volatility
Table 5.3: The TSBT Effect on Volatility
rt = θ0 + ut + θ1 ut−1 + θ2 ut−2
√
ut = hth²t
³
i
´
p
P
ln ht = ω + 2i=1 βi ln ht−i + αi | ²t−i | − 2/π + γi ²t−i
P
P
P
+ 8j=1 2τ =1 ηj,τ Chartj,τ,t + 4n=1 ϕn dn,t + φ ast−1
Coefficient
θ0
θ2
α1
α2
γ1
γ2
β1
β2
η1,1
η2,1
η3,1
η4,1
η5,1
η6,1
η7,1
Obs.
j
Q(j)
Q2 (j)
∗∗
Estimation
P-Value
0.003
-0.028∗ ∗
0.307∗ ∗
-0.229∗ ∗
-0.005
0.007
1.292∗ ∗
-0.321∗ ∗
-0.039
0.076
-0.254∗ ∗
-0.196∗
-0.054
-0.166∗ ∗
-0.482∗ ∗
(%)
44
0.0
0.0
0.0
39
20
0.0
0.0
42
16
0.0
4.9
51
0.0
0.3
37 650
1
1.21
3.21
W (ηj,1 = 0)
2
1.88
4.94
Coefficient
θ1
ω
η8,1
η1,2
η2,2
η3,2
η4,2
η5,2
η6,2
η7,2
η8,2
ϕ1
ϕ2
ϕ3
ϕ4
φ
5.177∗ ∗
12
10.88
10.98
Estimation
(2)
(3)
(4)
P-Value
(%)
∗∗
-0.134
0.0
∗∗
-0.067
0.0
-0.063
85
0.041
40
-0.066
22
∗∗
0.249
0.0
0.179
6.5
0.058
48
∗∗
0.164
0.0
0.476∗ ∗
0.2
0.072
83
∗∗
0.149
0.0
0.032∗ ∗
3.1
∗∗
0.056
0.3
∗∗
0.170
0.0
0.005∗ ∗
0.0
W (ηj,2 = 0) 5.178∗ ∗
24
20.72
30.13
and ∗ indicate respectively significance at 1% and 5%. W (ηj,1 = 0) is the Wald
statistic for the hypothesis of nullity of the 8 coefficients ηj,1 for all j, and W (ηj,2 = 0)
is the same statistic corresponding to the 8 coefficients ηj,2 . Q(j) and Q2 (j) are the
Ljung-Box statistics of order j respectively for standardized residuals and for their
square. Twelve lags correspond to 1 hour. Variables in Equations (2), (3), and (4):
rt is the SA return (multiplied by 10,000). Chartj,τ,t is a dummy variable for the
chart j corresponding respectively to the period of its completion (τ = 1), and the
period that occurs just after the completion (τ = 2), relative to time t. j = 1, 2, . . . , 8
corresponds respectively to the chart patterns BB, BT, DB, DT, RB, RT, TB, and
TT. dn,t is a dummy variable corresponding to the category of news n. Estimation
was done by the quasi maximum likelihood method.
5.5. Models and Empirical results
121
17 percent for RT, and 48 percent during the completion of TB. In turn, during the chart pattern completion, there are no technical signals, deals could be
dominated by rational traders who build their trade on typical information and
feature homogeneous behavior. They share the same expectations that weigh
on the exchange rate movement. Such behavior dampens the price changes and
triggers a drop on volatility, since the important order flows are generated by
typical information trades.
However, just after the completion of the technical chart patterns, once the
exchange rate crosses the support or the resistance level, technical traders are
attracted to initiate trades triggering an increase in volatility: DB (25 percent),
DT (18 percent), RT (16 percent) and TB (48 percent). These empirical results agree with the theoretical results found by De Long, Shleifer, Summers,
and Waldmann (1990) and Jeanne and Rose (2002) who attribute this raise on
volatility to the numerous presence of noise traders compared to the rational
ones. These two kinds of dealers are indeed different in terms of their behavior
and beliefs. However, heterogeneity is also created by the divergent objectives
of the two main categories of technical traders. One category, believing only in
the price reverse course at support and resistance levels, has already placed stop
loss (take profit) orders after (before) the chart completion. However, the second
category which involves the chartists, trust the basic chart patterns, and has
already placed take profit orders after the chart completion. The difference of
objectives characterizing both categories of technical traders heightens the price
cluster and generates a boost in volatility.
Moreover, there are some chart patterns, for instance BB, BT, RB and TT
that have no direct significant impact on volatility. They have, besides, an indirect impact through the market activity, since this latter variable has a positive
impact on volatility (highlighted in Chapter 2). Nevertheless, BT chart patterns
have no significant impact on either volatility or market activity. This means that
this kind of chart patterns does not attract the attention of technical traders and
does not trigger TSBT.
Table 5.4 displays the results corresponding to the total impact (before and after the completion) of the chart patterns on volatility. Volatility decreases before
the completion of the chart since there is no TSBT. After the chart completion,
122
Chapter5. The Technical Signal Based Trading Effects on Volatility
and once the exchange rate crosses the support or the resistance level, TSBT occurs and volatility increases to revert slowly to its initial level. The total impact
of TSBT is null except for BT and DT which present respectively small positive
and negative (of the order of 1.1 and 1.7 percent) effects on volatility. Reversion
to the initial level is implied by stationarity of the EGARCH model: the persistence is carried out through the sum of the coefficients β1 and β2 estimated at
0.97.
Table 5.4: The TSBT total impact on volatility
Chart pattern category (j)
η̂j,1
1-Broadening Bottom (BB) -0.039
2-Broadening Top (BT)
0.076
3-Double Bottom (DB)
-0.254∗ ∗
4-Double Top (DT)
-0.196∗
5-Rectangle Bottom (RB)
-0.054
6-Rectangle Top (RT)
-0.166∗ ∗
7-Triple Bottom (TB)
-0.482∗ ∗
8-Triple Top (TT)
-0.063
P2
τ =1 η̂j,τ
0.002
0.01∗ ∗
-0.005
-0.017∗
0.004
-0.002
-0.006
0.009
The second column gives the technical chart pre-completion
impact and the last column the total impact. Estimates are
taken in Table 5.3.
∗∗
and ∗ indicate respectively significance at 1% and 5%, for
one-tail tests.
Regarding the news announcements impact on volatility, the estimation results displayed in Table 5.3 are consistent with those found in Chapter 2. Volatility increases by 15 percent prior to the release of US figures. It rises by 3 percent
prior to European figures, by 6 percent prior to official speeches and 17 percent
prior to the release of US and European interest rate reports. A possible explanation of such volatility increases is that they are caused by anticipatory trades, i.e.
by traders who open positions hoping that their anticipations will coincide with
the contents of the news. However, another category of traders may also take
part in the trading. These traders, who are characterized by a high level of risk
aversion, prefer to execute their clients’ orders right before the news release to
avoid possible reversals of trends in the currency rate (Lyons, 1991 and Bauwens,
Ben Omrane, and Giot, 2005).
5.6. Conclusion
5.6
123
Conclusion
This last chapter of the thesis focuses on technical signal based trading effect on
exchange rate volatility. Using five-minute euro/dollar returns, news announcements, and technical chart patterns trading signals, we shed light on volatility
dynamics around TSBT.
The chapter results show that volatility decreases during the technical chart
pattern completion, when the exchange rate moves within the support and resistance levels, corresponding respectively to a concentration of demand and supply.
A possible explanation for such a volatility drop, is that the foreign exchange
market is dominated by rational traders, sharing the same expectations and featuring a homogeneous behavior. The second result consists of volatility increase
just after the chart completion, when the exchange rate crosses the support or
resistance level. In such an area, the attention of technical traders is attracted
and TSBT occurs creating heterogeneity within the market participants and triggering a rise in volatility. This empirical result agree with the theoretical finding
carried out by De Long, Shleifer, Summers, and Waldmann (1990) and Jeanne
and Rose (2002).
124
Chapter5. The Technical Signal Based Trading Effects on Volatility
Bibliography
Admati, A., and P. Pfleiderer (1988): “A theory of intraday patterns:
volume and price variability,” The Review of Financial Studies, 1, 3–40.
Allen, H., and M. Taylor (1989): “Charts and fundamentals in the foreign
exchange market,” Bank of England Discussion Paper, 40.
(1990): “Charts, Noise and Fundamentals in the London Foreign Exchange Market,” The Economic Journal, 100, 49–59.
(1992): “The use of technical analysis in the foreign exchange market,”
Journal of International Money and Finance, 11, 304–314.
Andersen, T., and T. Bollerslev (1996): “Heterogeneous information arrivals and returns volatility dynamics: uncovering the long-run in high frequency rendements,” NBER Working paper 5752.
(1998): “Deutsche mark-dollar volatility: intraday volatility patterns,
macroeconomic announcements and longer run dependencies,” The Journal of
Finance, 1, 219–265.
Andersen, T., T. Bollerslev, F. Diebold, and C. Vega (2002): “Micro
effects of macro announcements: real-time price discovery in foreign exchange,”
NBER Working paper 8959.
Andrada-Felix, J., F. Fernandez-Rodriguez, and S. Sosvilla-Rivero
(1995): “Technical analysis in the Madrid stock exchange,” Foundation for
Applied Economic Research (FEDEA) Working Paper, No. 99-05.
Bauwens, L., W. Ben Omrane, and P. Giot (2005): “News Announcements, Market Activity and Volatility in the Euro/Dollar Foreign Exchange
Market,” Journal of International Money and Finance.
125
126
BIBLIOGRAPHY
Béchu, T., and E. Bertrand (1999): L’analyse technique: pratiques et mthodes, gestion. Economica.
Ben Omrane, W., and H. Van Oppens (2006): “The Performance Analysis of
Chart Patterns: Monte Carlo Simulation and Evidence from the Euro/Dollar
Foreign Exchange Market,” Empirical Economics, 30, 947–971.
Black, F. (1972): “Capital Market Equilibrium with Restricted Borrowing,”
The Journal of Business, 45, 444–455.
(1986): “Noise,” The Journal of Finance, 41, 529–543.
Blume, L., D. Easley, and M. O’Hara (1994): “Market statistics and technical analysis : The role of volume,” Journal of Finance, 49(1), 153–181.
Bollerslev, T., and I. Domowitz (1993): “Trading patterns and prices in
the inter-bank foreign exchange market,” The Journal of Finance, 4, 1421–
1443.
Brock, W., J. Lakonishok, and B. LeBaron (1992): “Simple technical
trading rules and the stochastic properties of stock returns,” The Journal of
Finance, 47, 1731–1764.
Cai, J., Y. Cheung, R. Lee, and M. Melvin (2001): “Once in a generation
yen volatility in 1998: fundamentals, intervention and order flow,” Journal of
International Money and Finance, 20, 327–347.
Cao, H., M. Evans, and R. Lyons (2003): “Inventory Information,” typescript, U.C. Berkeley, June.
Carlson, J., and C. Osler (2000): “Rational Speculators and Exchange Rate
Volatility,” European Economic Review, 44, 231–253.
Chang, P., and C. Osler (1999): “Methodical madness: technical analysis
and the irrationality of exchange-rate forecasts,” The Economic Journal, 109,
636–661.
Cheung, Y.-W., and C. Y.-P. Wong (1999): “Foreign exchange traders in
Hong Kong, Tokyo and Singapore : a survey study,” Advances in Pacific Basin
Financial Markets, 5, 111–134.
BIBLIOGRAPHY
127
Chordia, T., R. Roll, and A. Subrahmanyam (2005): “Evidence on the
speed of Convergence to Market Efficiency,” Journal of Financial Economics,
76, 271–292.
Dacorogna, M., R. Gençay, U. Müller, R. Olsen, and O. Pictet
(2001): An introduction to high frequency finance. Academic Press.
Dacorogna, M., R. Gençay, U. Müller, and O. Pictet (2001): “Effective return, risk aversion and drawdowns,” Physica A, 289, 229–248.
Dacorogna, M., U. Müller, R. Nagler, R. Olsen, and O. Pictet
(1993): “A geographical model for the daily and weekly seasonal volatility in
the foreign exchange market,” Journal of International Money and Finance,
12, 413–438.
Danielsson, J., and R. Payne (2002): “Real trading patterns and prices in the
spot foreign exchange markets,” Journal of International Money and Finance,
21, 203–222.
De Long, J., A. Shleifer, L. Summers, and R. Waldmann (1990): “Noise
trading risk in financial markets,” Journal of Political Economy, 98, 703–738.
Degennaro, R., and R. Shrieves (1997): “Public information releases, private information arrival and volatility in the foreign exchange market,” Journal
of Empirical Finance, 4, 295–315.
Dempster, M., and C. Jones (1998a): “Can technical pattern trading be
profitably automated ? 1. Channel,” Centre for Financial Research, Judge
Institute of Management Studies, University of Cambridge, working paper.
(1998b): “Can technical pattern trading be profitably automated ?
2. The head & shoulders,” Centre for Financial Research, Judge Institute of
Management Studies, University of Cambridge, working paper.
Denuit, M., and P. Lambert (2002): “Constraints on Concordance measures
in bivariate discrete data,” Discussion Paper 02-12, Institut de Statistique,
Université Catholique de Louvain, Louvain-la-Neuve, Belgium.
128
BIBLIOGRAPHY
Detry, P. (2001): “Other evidences of the predictive power of technical analysis: the moving averages rules on European indexes,” EFMA 2001 Lugano
Meetings.
Diebold, F. X., T. A. Gunther, and A. S. Tay (1998): “Evaluating Density
Forecasts with Applications to Financial Risk Management,” International
Economic Review, 39(4), 863–883.
Dooley, M., and J. Shafer (1984): “Analysis of short-run exchange rate
behavior : March 1973 to 1981,” Floating Exchange Rates and The State of
World Trade and Payments, pp. 43–70.
Eddelbüttel, D., and T. McCurdy (1998): “The impact of news on foreign exchange rates : evidence from high frequency data,” Working paper,
University of Toronto.
Efron, B. (1986): “Double Exponential Families and Their Use in Generalized
Linear Regression,” Journal of the American Statistical Association, 81(395),
709–721.
Evans, M. (2002): “FX Trading and Exchange Rate Dynamics,” Journal of
Finance, 57(6), 2405–2447.
Evans, M., and R. Lyons (1999): “Order flow and exchange rate dynamics,”
Berkeley Working paper.
(2002): “Order Flow and Exchange Rate Dynamics,” Journal of Political Economy, 110, 170–180.
(2003): “How is Macro News Transmitted to Exchange Rates?,” NBER
Paper 9433.
Fama, E. (1970): “On the Efficient Capital Markets: A Review of Theory and
Empirical Work,” Journal of Finance, 25, 383–417.
Fiess, N., and R. MacDonald (2002): “Towards the fundamentals of technical analysis: analysing the information content of High, Low and Close prices,”
Economic Modeling, 19, 353–374.
Fisher, R. A. (1932): Statistical methods for reasearch workers.
BIBLIOGRAPHY
129
Frankel, J., and K. Froot (1986): “Understanding the U.S. Dollar in the
Eighties: The Expectations of Chartists and Fundamentalists,” Economic
Record.
(1990): “Chartists, fundamentalists, and trading in the foreign exchange
market,” American Economic Review.
Friedman, M. (1953): Essays in Positive Economics, Chicago. university of
Chicago Press.
Gençay, R. (1998): “The predictability of security returns with simple technical
trading rules,” Journal of Empirical Finance, 5, 347–359.
Gençay, R., G. Ballocchi, M. Dacorogna, R. Olsen, and O. Pictet
(2002): “Real-time trading models and the statistical properties of foreign
exchange rates,” International Economic Review, 43, 463–491.
Gençay, R., M. Dacorogna, R. Olsen, and O. Pictet (2003): “Foreign
exchange trading models and market behavior,” Journal of Economic Dynamics & Control, 27, 909–935.
Gençay, R., and T. Stengos (1997): “Technical trading rules and the size
of the risk premium in security returns,” Studies in Nonlinear Dynamics and
Econometrics, 2, 23–34.
Glosten, L., and P. Milgrom (1985): “Bid ask and transaction prices in a
specialist market with heterogeneously informed traders,” Journal of Financial
Economics, 13, 71–100.
Goodhart, C., and A. Demos (1991): “Reuters’ screen images of the foreign
exchange markets: The Yen/Dollar and the Sterling/Dollar spot market,” The
Journal of International Securities Markets, 5.
Goodhart, C., and L. Figliuli (1991): “Every minute counts in financial
markets,” Journal of International Money and Finance, 10, 23–52.
Heinen, A., and E. Rengifo (2003): “Multivariate Modeling of Time Series
Count Data: An Autoregressive Conditional Poisson Model,” CORE Discussion Paper 2003/25.
130
BIBLIOGRAPHY
Jeanne, O., and A. Rose (2002): “Noise trading and exchange rate regimes,”
The Quarterly Journal of Economics.
Jensen, M. (1970): “Random walks and technical theories: some additional
evidence,” Journal of Finance, 25(2), 469–482.
Joe, H. (1997): Multivariate Models and Dependence Concepts. Chapman and
Hall, London.
LeBaron, B. (1999): “Technical trading rule profitability and foreign exchange
intervention,” Journal of International Economics, 49, 125–143.
Levich, R., and L. Thomas (1993): “The significance of the trading-rule
profits in the foreign exchange market : a bootstrap approach,” Journal of
International Money and Finance, 12, 1705–1765.
Levy, R. (1971): “The predictive significance of five-point patterns,” Journal
of Business, 41, 316–323.
Lo, A., and A. MacKinlay (1990): “An econometric analysis of nonsynchronous trading,” Journal of Econometrics, 45, 181–211.
Lo, A., H. Mamaysky, and J. Wang (2000): “Foundations of technical analysis: computational algorithms, statistical inference, and empirical implementation,” Journal of Finance, 56, 1705–1765.
Lucas, R. (1978): “sset Prices in an Exchange Economy,” Econometrica, 46,
1429–1445.
Lui, Y., and D. Mole (1998a): “The use of fundamental and technical analyses
by foreign exchange dealers : Hong-Kong evidence,” Journal of International
Money and Finance, 17, 535–545.
(1998b): “The Use of Fundamental and Technical Analyses by Foreign
Exchange Dealers: Hong Kong Evidence,” Journal of International money and
Finance, 17, 535–545.
Lyons, R. (1991): “Private beliefs and information externalities in the foreign
exchange market,” NBER Working Paper 3889.
BIBLIOGRAPHY
131
(1995): “Tests of microstructural hypotheses in the foreign exchange
market,” Journal of Financial Economics, 39, 321–351.
(1997): “A simultaneous trade model of the foreign exchange hot
potato,” Journal of International Economics, 45, 181–211.
(2001): The microstructure approach to exchange rates. MIT Press.
Melvin, M., and X. Yin (2000): “Public information arrival, exchange rate
volatility and quote frequency,” The Economic Journal, 110, 644–661.
Menkhoff, L. (1998): “The noise trading approach - questionnaire evidence
from foreign exchange,” Journal of International Money and Finance, 17, 547–
564.
Murphy, J. (1986): Technical Analysis of the Futures Markets: a Comprehensive Guide to Trading methods and Applications. New York Institute of
Finance.
(1999): Technical analysis of the financial markets. New York Institute
of Finance.
Neely, C. (1997): “Technical analysis in the foreign exchange market: a layman’s guide,” Federal Reserve Bank of Saint Louis Working Paper.
Neely, C., and P. Weller (2003): “Intraday technical trading in the foreign
exchange market,” Journal of International Money and Finance, 22(2), 223–
237.
Neely, C., P. Weller, and R. Dittmar (1997): “Is technical analysis in
the foreign exchange market profitable ? A Genetic Programming Approach,”
Journal of Financial and Quantitative Analysis, 32(4), 405–426.
Nelsen, R. B. (1999): An introduction to Copulas. Springer, New York.
Nelson, D. (1990): “Conditional heteroskedasticity in asset returns: a new
approach,” Econometrica, 59, 347–370.
O’Hara, M. (1995): Market microstructure. Blackwell business.
132
BIBLIOGRAPHY
Osler, C. (1998): “Identifying noise traders: the head-and-shoulders pattern
in U.S. equities,” Federal Reserve Bank of New York Working Paper.
Osler, C. (2000): “Support for resistance: technical analysis and intraday
exchange rates,” Federal Reserve Bank of New York Working Paper.
(2002): “Stop-loss orders and price cascades in currency markets,”
Mimeo, Federal Reserve Bank of New York.
(2003): “Currency Orders and Exchange Rate Dynamics: An Explanation for the Predictive Success of Technical Analysis,” The Journal of Finance,
5, 1791–1819.
Patton, A. (2002): “Skewness, Asymmetric Dependence, and Portfolios,”
mimeo.
Peiers, B. (1997): “Informed traders, intervention and price leadership: a
deeper view of the microstructure of the foreign exchange market,” The Journal of Finance, 4, 1589–1614.
Pring, M. (1985): Technical analysis Explained: The Successful Investor’s
Guide to Spotting Investment Trends and Turning points. McGraw-Hill.
Prost, A., and R. Prechter (1985): Eliott waves principle. New classics
library.
Ready, M. (1997): “Profits from technical trading rules,” Working Paper, University of Wisconsin-Madison.
Rime, D. (2000): “Private or public information in foreign exchange market?
an empirical analysis,” University of Oslo, Memorandum 14.
Sapp, S. (2002): “Price leadership in the Spot Foreign Exchange Market,” Journal of Financial and Quantitative Analysis, 37(3), 425–448.
Shiller, R. (2003): “From Efficient Markets Theory to Behavioral Finance,”
Journal of Economic Perspectives, 17, 83–104.
Silverman, B. (1986): Density estimation for statistics and data analysis.
Chapman and hall.
BIBLIOGRAPHY
133
Sklar, A. (1959): “Fonctions de répartitions à n dimensions et leurs marges,”
Public Institute of Statistics of the Univeristy of Paris, 8, 229–231.
Sweeney, R. (1986): “Beating the foreign exchange market,” Journal of Finance, 41, 304–314.
Wang, J.-X. (2001): “Quote Revision and Information flow among Foreign
Exchange Dealers,” Journal of International Financial Markets, Institutions
and Money, 11, 115–136.
Index
activity, 4–9, 11, 18–22, 27–29, 34, 35, equidispersed, 53
39, 41–48, 50, 54–57, 59, 62, 65, estimation, 17, 21, 30, 36, 38, 39, 44,
105, 110, 114, 117–119, 121
55, 58, 107, 119, 122
agenda, 24, 116
estimatior, 80
arbitrage, 73
forecast, 4, 10, 16, 25, 33, 55, 57, 64,
ARCH, 6
74, 75, 85, 86, 107, 110, 111
autocorrelation, 22, 30, 35, 55, 57, 60,
foreign exchange, 1, 3, 6, 9–12, 15, 41,
117
42, 48, 54, 59, 65, 74–78, 89, 93,
bank for international settlements, 3
94, 106, 107, 110, 111, 123
basis point, 2, 3, 91–93
behavior, 12, 73, 74, 76, 88, 89, 105, GARCH, 20
107, 109, 114, 115, 121, 123
heterogeneous, 8, 45, 46, 65, 114, 121
bias, 22, 80, 115
high frequency, 7, 16, 20, 39, 74, 78, 79
bid-ask, 1, 42, 74, 92
homogeneous, 12, 114, 115, 121, 123
breakout, 11, 106
hypothesis, 8, 30–32, 34, 38, 58, 59, 62,
Brownian motion, 75, 88
64, 65, 74, 76, 88, 118, 120
chartism, 6, 76, 107
interdealer, 2–4
couterparty, 3
interest rate, 19, 24, 26–28, 33, 34, 38,
cycles, 20
44, 51, 59, 62, 106, 110, 122
direction, 85, 86, 112
interpretation, 9, 16, 39, 43, 45–47, 65
dynamics, 6, 7, 9, 11, 16, 18, 55, 77, intraday, 7, 17, 20
106, 107, 111, 114, 117, 123
kernel, 80, 95–98, 110
EBS, 1, 2, 41, 42
macroeconomic, 17–19, 24–29, 33–35,
ECB, 24
38, 44, 45, 50, 51, 57, 59, 62
economic, 6, 17–19, 24–26, 51, 110, 116
EGARCH, 7, 17, 30, 32, 34, 117–119, microstructure, 15, 17, 73
122
mimic, 109, 114
135
136
INDEX
news announcement, 5–9, 11, 12, 15– scheduled news, 7, 11, 17, 34, 39, 43,
20, 22, 24–29, 32, 34, 35, 38, 39,
57, 118
42–47, 50, 51, 53, 54, 56, 57, 62, scheduled speeches, 24
65, 106, 115, 122, 123
noise, 5, 8–10, 12, 44, 79, 80, 95, 105,
107, 108
noise trader, 5, 73, 78, 105, 107
noise trading, 9, 107
normal distribution, 53, 67
overdispersed, 53, 57
position, 7, 16–18, 20, 22, 28, 29, 33,
78, 79, 86, 110, 114, 122
seasonality, 7, 17, 20, 28, 29, 37, 50, 54,
56, 58, 59
spread, 2, 19, 20, 74, 92, 95
support level, 11, 12, 77, 106, 107, 109,
111, 112, 114, 115, 121–123
technical analysis, 9, 11, 73, 74, 76, 77,
81, 105, 107, 111
Telerate, 1, 41
time interval, 9, 19, 32, 35, 36, 38, 75,
79, 80, 84, 85, 93, 95, 97, 98,
price, 1
100, 117
random walk, 10, 73, 110
regular, 7, 16, 27, 95, 98, 114
resistance level, 11, 12, 77, 106, 107,
109, 111, 112, 114, 115, 121–123
return, 15, 17, 20–22, 24, 28, 30, 36, 37,
39, 73, 74, 77, 78, 88, 110, 117,
123
underdispersed, 53, 57
unscheduled, 7, 15, 16, 18, 33, 34, 39,
44
unscheduled announcements, 24, 29
unscheduled event, 18
unscheduled news, 11, 39, 43, 59
Reuters, 1, 2, 7, 17, 24–26, 39, 41, 42, volatility decrease, 28, 123
51, 116
Risk, 4, 10, 28, 29, 75, 80, 87, 92, 93,
106, 110
Risk aversion, 62, 86, 87, 92, 122
Risk management, 4
Risk premium, 88
scheduled, 7, 11, 15–17, 20, 24, 25, 27,
29, 33, 34, 38, 39, 44, 50, 57,
106
scheduled announcements, 18, 27
scheduled events, 38, 115
volatility increase, 5, 7, 11, 12, 16, 17,
19, 28, 29, 32, 33, 39, 106, 115,
122, 123