The short term behavior of the EUR/USD exchange rate after the release of a U.S. macroeconomic data surprise

(1)

0

MSc Econometrics

Universiteit van Amsterdam

Supervisor UvA: Kees Jan van Garderen Amsterdam, June 2015

The short term behavior of the EUR/USD exchange rate

after the release of a U.S. macroeconomic data surprise

“”

Yannic Pieters

Student number: 6163297 MSc thesis

(2)

Statement of Originality

This document is written by Student Yannic Pieters who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

1

Abstract

There is a scheduled calendar in which the release time and date of important U.S. macroeconomic indicators are listed. From these U.S. macroeconomic indicators it is suspected that they influence the trading in forex. A lot of predictions are made about the value of these macroeconomic indicators but these are not always accurate. Often the prediction is different than the actual value of the macroeconomic indicator. This is called a data surprise and recent literature do find that for a lot of macroeconomic indicators there is a systematical relationship between a data surprise and the level of the exchange rate. After a data surprise in a U.S. macroeconomic indicator, a large price movement or jump occurs in the EUR/USD exchange rate.

Recent studies use high frequency data with fixed time intervals to examine the effect of a data surprise on the EUR/USD exchange rate. This study uses tick data with a variable time window to examine this effect, because it is up to now unknown what the duration of the first large price movement or jump actually is. A new filter has filtered these jumps from approximately 250 million price issues and the obtained dataset was used to give a clearer picture of jump returns and jump durations.

This study reports the following findings. For a lot of macroeconomic indicators, jump returns are influenced by the size of the data surprise, but this data surprise does not influence the corresponding jump durations. For some macroeconomic indicators there is an asymmetric response to data surprises and data surprises of different macroeconomic indicators do not influence each other in the effect. Furthermore, it is often the type of news that influences these durations, from which 43% is of 10 seconds or less. Another finding is that anticipation to data surprises has become significantly faster in recent years. At last, uncertainty in the forex market causes shorter durations.

(4)

2

1 Introduction

There are two main approaches that forex traders use to gain insight in the foreign exchange (forex) market to increase their predictive capabilities of price

movements. Whereas the technical analyses tries to find patterns in price movements by analyzing historical data, fundamental analysis is based on the overall state of the economy which can be determined by macroeconomic indicators. The release of the majority of these macroeconomic indicators is scheduled in a precise macroeconomic calendar. A number of researches substantiate the fact that the release of the value of these macroeconomic indicators has a large influence on trading in forex. A finding by Chueng and Chinn (2001) is that news releases of United States (U.S.) macroeconomic indicators is rapidly incorporated into the EUR/USD exchange rate. Evans and Lyons (2005) state that macroeconomic news arrivals tend to have substantial influence on trading in and the price of forex. Because these indicators do have a large influence on exchange rates, there are a lot of predictions made by companies about the value of these U.S. macroeconomic indicators, which traders can use in their trading strategy prior to the scheduled news release. The companies who publically distribute these expectations are for instance Bloomberg, DailyFX or ForexFactory. It has been shown by Chaboud, Chernenko, Howorka, Iyer, Liu and Wright (2004) that there is a

systematic relationship between the difference of the prediction and the actual value of a U.S. macroeconomic indicator (data surprise) and the subsequent change in the level of the exchange rate. This means that predictions are already incorporated in the exchange rate and they found that data surprises can lead to large price movements, or jumps, in the exchange rate immediately after the official data releases of macroeconomic indicators.

In recent years, literature has examined the relationship between the proportional size of the jump or jump return and the amount of surprise in the data. A common approach in these papers is to use, right after the data surprise, the rate of return in the exchange rate as the proportional size of the jump. This is often done by using high frequency data with time intervals ranging from one minute to one day. The rate of return is then calculated by comparing in the data the last quote before and the first quote after the data surprise. It is then not ruled out that during these time intervals,

(6)

4 other news has hit the market and disturbs the effect of the data surprise on the

exchange rate. These 1-min to 1-day intervals can be seen as fixed time windows. While recent papers focus mainly on jump returns, this paper also focuses on jump durations to give a better description of these jumps. To do this, the paper tries to answer the

following question: what is the relationship between a data surprise in a macroeconomic news arrival, the subsequent jump in the EUR/USD exchange rate and the duration of this jump?

A contribution to recent literature is that this thesis tries to capture the sole effect of the data surprises by using a variable time window instead of a fixed time window. This is implemented in an algorithm that filters the jumps from EUR/USD exchange rate tick data. This tick data is of high quality and obtained from Dukascopy Bank SA. It contains every price issued by Dukascopy Bank SA from January 2010 till March 2015. Our algorithm uses generally accepted definitions of a jump to find these jumps and then filters these jumps from the tick data. The most important contribution is that this thesis investigates the durations of these jumps. Recent papers state that news surprises are rapidly incorporated in exchange rates, but precise durations of these jumps are left out. The incorporation happens very quickly, but how quick and what influences this speed? The durations of the jumps, which are obtained using the new filter, will be modeled by an accelerated failure time model with time invariant

regressors. These regressors include the standardized data surprise, realized volatility of the EUR/USD exchange rate prior to the news release, based on short term raw tick data, and the type of news. In order to investigate the marginal impact of data surprises on jump returns, a multivariate regression model is fitted using Ordinary Least Squares (OLS). This model is extended to incorporate and test asymmetry in the effects between positive and negative data surprises. Also the interaction effect of concurrent data surprises is investigated. This all will be done separately for the most important U.S. macroeconomic indicators that affect the EUR/USD exchange rate.

The outline of the remainder of this thesis is as follows. Section 2 will review the relevant literature. Section 3 will introduce the different models and methods that are being used to answer the main question. Section 4 will present an overview of the data used and the dataset as a result of the filtration. The results of the model

(7)

5 estimations will be discussed in Section 5. Concluding remarks will be presented in Section 6.

(8)

6

2 Literature Review

Several studies examine the effect of data surprises in macroeconomic indicators on exchange rate returns. These studies vary in the technique and dataset that they use, but the main goal is the same: finding a relationship between the announcement of macroeconomic indicators and the behaviour of an exchange rate. This section will highlight the most important papers in this line of research.

Macroeconomic indicators are an important driving force behind price dynamics in the forex market. Since macroeconomic indicators are essential to compute the fundamental value of a currency pair, traders focus a lot on these indicators. Not only for fundamental analysis, but also for technical analysis, because a lot of volatility in the forex market is caused by the releases of these indicators. For example: Evans and Lyons (2008) showed that roughly 30% of the total volatility in the Deutsche

Mark/USD exchange rate from May 1 1996 up to and including August 31 1996 came from news effects.

In order to understand how traders react to fundamental data releases, Chinn et al. (2001) reported findings from a survey of U.S. forex traders. The survey contained several topics regarding trading in forex, from which one topic was about the

relationship between forex trading and news about economic fundamentals. According to the results, the exchange rate responded to news with extreme rapidity, usually within minutes and the U.S. Unemployment rate and U.S. GDP were amongst the most

important U.S. macroeconomic indicators. Consistently about one third of the

respondents indicated that full adjustment to the news takes place within 10 seconds for the most important macroeconomic indicators regarding the U.S. economy.

The finding of Chinn et al. (2001) is a good motivation of why high frequency data with an interval of one minute or more is sometimes inefficient. The adjustment could already be over before the end of an interval is reached. Ederington and Lee (1993) already confirmed the need for very high frequency data to detect announcement effect. Using tick data, they detected adjustment of the volatility of the exchange rate within the first minute after the announcement of a macroeconomic indicator. Ozdagli (2013) stated that very narrow interval data is necessary, because a too large interval of

(9)

7 for example 30 minutes could lead to a non-accurate estimation. This is because the reaction may be affected by other important news that was released earlier on the day.

Ehrmann and Fratzscher (2005) showed that exchange rates respond more strongly to news in periods of large market uncertainty. They state that amongst the most important U.S. macroeconomic indicators that have the most clear and consistent patterns after their release are the U.S. Non-Farm Employment Change, U.S. Advance GDP q/q, U.S. CPI and U.S. Unemployment rate. They showed this using intraday data and this is one of the few studies that uses this kind of data to find a relationship

between the announcement of a macroeconomic indicators and the behaviour of an exchange rate. Their motivation is that full reaction of asset prices to news may not occur immediately, but that the news is incorporated only gradually over the subsequent minutes. This is opposed to the opinion of the traders from the survey of Chinn et al. (2001) and it raises two important questions: what is the definition of this reaction and what is its duration? The first part of this question can partially be answered by the literature.

Chaboud et al. (2004) focused on the effect of a data surprise on the EUR/USD exchange rate using 1-min time interval high frequency data. Key finding of this paper is that there is a systematic relationship between the surprise component of the news announcement and the level of the exchange rate. Stronger-than-expected U.S. activity tends to systematically be followed by dollar appreciation. The effect of the data

surprise on the conditional mean of the exchange rate occurs quickly and usually within a minute. Also the announcements of macroeconomic indicators are immediately followed by higher trading volume and volatility and these factors remain elevated for a period of time. This is in line with the pattern that Fleming and Remolona (2009) describe. They state that the behaviour of an exchange rate after the announcement of a macroeconomic indicator can be divided in two stages. In the first stage there is a very sharp and short movement in the exchange rate which is effectively a jump and in the second stage volatility and trading volume picks up and stays high for several hours. Evans et al. (2005) give a more detailed description about these two stages and call it the total and average effect. Where the average effect is the direct effect of a news release, the total effect is the overall effect that the news causes on a longer term. The

(10)

8 average effect is quickly incorporated in the price and is effectively a jump, while the total effect could last for a few days. They conclude that there is indeed a total effect and that the previous literature did not appear well-suited to capture the total effect of news. The above gives an answer to the first of the earlier addressed two questions: the reaction can be called an average or a total effect depending on the chosen time window in which the returns of the exchange rate are evaluated.

In addition, Faust, Rogers, Wang and Wright (2007) used 20-min time interval high frequency data to examine the effect of data surprises on the EUR/USD exchange rate. They state that there is a clear and consistent pattern in the responses to data surprises for certain macroeconomic indicators. Stronger than expected

announcements of U.S. macroeconomic indicators consistently appreciate the value of the dollar. Again the U.S. Non-Farm Employment Change, U.S. Advance GDP q/q and U.S. CPI are among the most important macroeconomic indicators, but they differ in reaction consistency and power to what is reported in the results of Ehrmann et al. (2005). This is also the case for Chatrath, Miao, Ramchander and Villupuram (2014). Using 5-min time interval high frequency data they show that the Non-Farm

Employment Change, the Unemployment report, GDP q/q and Trade Balance are amongst the most important U.S. news variables, but this is another selection than the one reported by Faust et al. (2007). It seems that the choice of dataset influences the results as this difference can also be seen at other studies while the regression technique, namely a multivariate regression model as mentioned in Section 3, is the same. This is probably not very strange. It could very well be possible that the duration of the largest price movement after a data surprise differs per macroeconomic indicator. This means that using different types of high frequency data give other results. Also the importance of macroeconomic indicators can vary over time.

Furthermore, Andersen, Bollerslev, Diebold and Vega (2002) used 20-min time interval high frequency data to examine the effect of data surprises on the EUR/USD exchange rate. Again a relationship was found between data surprises and the level of the EUR/USD exchange rate and the authors state that announcement surprises produce conditional mean jumps. They also addressed the possibility of asymmetric responses to data surprises. They showed that the market reacts to news in an asymmetric fashion:

(11)

9 bad news has greater impact than good news. Evans and Speight (2011) reach a similar conclusion. By using 20-min time interval high frequency data they show that positive surprises in poor economic climate have strong influences on short-term returns in exchange rates.

The literature is very clear about one thing: data surprises of U.S.

macroeconomic indicators have an effect on the level of the EUR/USD exchange rate and there are two different reactions to such data surprises. The first one is the average effect and is described by the literature as effectively a jump in the exchange rate. The second one is the total effect and is described by the literature as the long term effect. This thesis focusses on the average effect or jump and the recent literature regarding this effect shows different results when it comes to comparing the different

macroeconomic indicators. This thesis defines this jump in such a way that it is possible to give a clearer picture of its return, duration and what influences these two factors.

(12)

10

3 Models and Methods

In this section the models and methods that are being used to answer the main question of this thesis are described. The method that will be used to estimate the regression models below is OLS. This section will be divided in 3 different topics. The first one will focus on the filter used, the second one will focus on the effect on jump returns, and the third one will focus on the effect on jump durations.

3.1 The filter: the models and methods

The description of the average effect or the jump varies widely between recent papers regarding the effect of data surprises on an exchange rate. For instance Chaboud et al. (2004) state that a large movement in the conditional mean of the exchange rate that is generally completed within a few minutes is effectively a jump. Chatrath et al. (2014) state that a jump is a discontinuous movement in the exchange rate. Hashimoto and Ito (2010) state that if most of the return reaction occurs within a minute and this new level continuous for at least 30 minutes, this reaction is a one-time sudden jump. These papers use fixed time window high frequency data and therefore can only state that a jump occurs, but cannot tell the exact duration of the jump. This paper uses tick data, which is discrete data where every move is discontinuous and therefore by

Chatrath et al. (2014) a jump. So looking at discontinuous movements in the data is not the way to find the short term effect of the data surprise on the exchange rate. In the relevant literature, the common definition of a jump is a sudden large movement in the exchange rate that occurs in a relative short amount of time. We use this definition, rather than a one-off large discrete change, to propose a new filter to extract the jumps from the available tick data. The key feature and contribution of this filter is that it not only informs on the size of the jump, but also the duration.

The price of the EUR/USD exchange rate is defined in this thesis as the

midpoint between the bid and ask price at a specific time. A news release is defined in this thesis as a specific point in time between January 2010 and March 2015 where the value of at least one macroeconomic indicator was released. If 𝑃_𝑘 is the price of the

(13)

11 EUR/USD exchange rate at the 𝑘_𝑡ℎ tick data point after the release of the news, 𝑃₀ is the last price of the EUR/USD exchange rate before the release of the news and 𝜏_𝑘 can be seen as the time in milliseconds between the news release and the release of 𝑃𝑘, then

𝜃_𝑘 = 𝑃𝑘−𝑃0

𝜏𝑘 =

∆𝑃𝑘

𝜏𝑘 can be seen as the average movement of the EUR/USD exchange rate

to reach 𝑃_𝑘 from the moment the news was released. It is trivial to see that when ∆𝑃_𝑘 increases, 𝜃_𝑘 increases and that when 𝜏_𝑘 decreases, 𝜃_𝑘 increases. This means that after a news release, tick data points representing a larger price movement in a shorter period of time, have larger values of 𝜃𝑘 and vice versa. In the relevant definitions, distance

covered or jump size is just as important. A tick data point with the same average movement as another tick data point, but with a larger value of ∆𝑃_𝑘, has more the common description of a jump. To implement this aspect in our algorithm, 𝜃_𝑘 will be multiplied by ∆𝑃𝑘. 𝛾𝑘=𝜃𝑘∆𝑃𝑘 can now be seen as a variable that assigns values to tick

data points after a news release that gives larger values to points that fit more the common description of the jump: a large price movement in a short period of time.

The proposed algorithm uses 𝛾𝑘 to find the end of a jump in the tick data after a

news release. It starts searching after a news release for the tick data point that has the largest value of 𝛾_𝑘 and the associated end of the jump, and saves the duration of the jump as 𝑇 in milliseconds. It does this in the first four minutes after a news release. Four minutes should be sufficient long if we bear in mind that recent relevant literature state that the first large price movement occurs within a few minutes or even seconds. The release time and data of all news releases used in this paper are fixed and known and therefore it is easy to find the points in the tick data where the news is released. 𝑇 can be defined as follows:

𝑇 = 𝑎𝑟𝑔 𝑚𝑎𝑥_𝜏_𝑘_{=1,…,240000} 𝛾_𝑘. (1)

If 𝐾 is the 𝑘 𝑡ℎ_{tick data point that was released at 𝑇 after the news release, then the}

price of the EUR/USD exchange rate after the jump can be defined as 𝑃_𝐾. When 𝑗 corresponds to the 𝑗𝑡ℎ_{news release, the jump return of the jump that corresponds to the}

𝑗𝑡ℎ_{news release can now be defined as 𝑟}

𝑗 =

𝑃_𝐾𝑗

𝑃_0𝑗 whereas its duration can be defined as

(14)

12 3.2 The effect on jump returns: the models and methods

For the measurement of standardized surprises, the research of Balduzzi, Elton and Clifton (2001) and Chaudhry, Ramchander and Simpsons (2005) is followed. We define the standardized data surprise at news release 𝑗 for U.S. macroeconomic indicator 𝑖 as 𝑆𝐴_𝑖,𝑗. 𝑆𝐴_𝑖,𝑗 is defined as

𝑆𝐴𝑖,𝑗= 𝐴𝑖,𝑗_𝜎_̂−𝐸𝑖,𝑗

𝑖 , (2)

where 𝐴_𝑖,𝑗 denotes the released value of U.S. macroeconomic indicator 𝑖 at news arrival 𝑗 and 𝐸𝑖,𝑗 denotes the predicted value of U.S. macroeconomic indicator 𝑖 at news arrival

𝑗. 𝜎̂_𝑖 is the sample standard deviation of the surprise component 𝐴_𝑖,𝑗− 𝐸_𝑖,𝑗. Expected is that a positive news surprise about the U.S. economy appreciates the U.S. dollar. In other words, expected is that a positive data surprise has a negative effect on the price of the EUR/USD exchange rate. In order to estimate the marginal effect of each data surprise on jump returns, we estimate the following multivariate regression model as described by Chatrath et al. (2014):

ln (𝑟_𝑗) = 𝑐 + ∑15 𝛽_𝑖𝑆𝐴_𝑖,𝑗

𝑖=1 + 𝜀𝑡𝑗. (3)

Hereafter, we use a one-time backward elimination to improve the model. A restricted set of regressors that represents the most influential factors (significant at the 10% level) is identified and this reduced model is estimated.

In order to test for possible asymmetry in the effects, we estimate another multivariate regression model, as described by Evans et al. (2011):

ln (𝑟𝑗) = 𝑐 + ∑15𝑖=1𝛽𝑖𝑆𝐴𝑖,𝑗++ ∑30𝑖=16𝛽𝑖𝑆𝐴𝑖,𝑗−+ 𝜀𝑡𝑗, (4) where, according to macroeconomic theory, 𝑆𝐴𝑖,𝑗+ and 𝑆𝐴𝑖,𝑗− describe a positive and

negative news surprise respectively for U.S. macroeconomic indicator 𝑖 at news arrival 𝑗. 𝑆𝐴_𝑖,𝑗+ = 0 when the news surprise of U.S. macroeconomic indicator 𝑖 at news arrival 𝑗 is negative and 𝑆𝐴_𝑖,𝑗− = 0 when the news surprise of U.S. macroeconomic indicator 𝑖 at news arrival 𝑗 is positive. Wald tests of coefficient restrictions are then performed on the coefficients to test the equality of reactions to positive and negative news surprises.

(15)

13 It often occurs that multiple macroeconomic indicators are released at the same time. At such a news release there are often concurrent data surprises. In order to investigate whether there is an interaction effect of concurrent data surprises we looked in the dataset for news releases where at least two data surprises occur at the same time. A “positive/negative surprise combination” refers to a combination of macroeconomic indicators who have had concurrent positive/negative data surprises for a minimum of two times in the dataset and no other macroeconomic indicators where released at these times. When there is a mix between concurrent positive and negative data surprises of macroeconomic indicators, these macroeconomic indicators form a “other surprise combination”. Table 1 shows a list of positive, negative and other surprise

combinations. In Table 1 there are numbers assigned to each surprise combination. Surprise combination 𝑚 refers to the combination next to the number 𝑚. In order to test if there is an interaction effect of concurrent data surprises, we estimate the following multivariate regression model:

ln(𝑟𝑗) = 𝑐 + ∑15𝑖=1𝛽𝑖𝑆𝐴𝑖,𝑗++ ∑30𝑖=16𝛽𝑖𝑆𝐴𝑖,𝑗−+ ∑6𝑚=1𝛽𝑘𝑃𝐶𝑗,𝑚+

∑14 𝛽_𝑘𝑁𝐶_𝑗,𝑚

𝑚=7 + ∑29𝑚=15𝛽𝑘𝑂𝐶𝑗,𝑚+ 𝜀𝑡𝑗, (5) , where 𝑃𝐶_𝑗,𝑚 corresponds to a multiplication of the values of the standardized surprises in surprise combination 𝑚 when all data surprises are positive at news release 𝑗 (0 otherwise), 𝑁𝐶𝑗,𝑚 corresponds to a multiplication of the values of the standardized

surprises in surprise combination 𝑚 when all data surprises are negative at news release 𝑗 (0 otherwise) and 𝑂𝐶_𝑗,𝑚 corresponds to a multiplication of the values of the

standardized surprises in surprise combination 𝑚 at news release 𝑗 when there are concurrent data surprises but they are not all positive nor negative (0 otherwise).

(16)

14

Table 1: List of combinations of macroeconomic indicators with concurrent data surprises

3.3 The effect on durations: the models and methods

The proposed filter identifies not only the size of the jump, but also the corresponding duration. In an efficient market all durations would be close to zero, because in an efficient market all new information is instantaneously incorporated in the exchange rate. This also means that previous prices incorporated all information and true data surprises are unanticipated and do not lead to arbitrage opportunities. But in practice there is a possible problem for the market efficiency. Fast absorption of

information could be influenced by liquidity constraints in the market. At the time news is released, market makers decrease their available liquidity in the market to reduce their risk because of the possible jump in the EUR/USD exchange rate. Because

liquidity is reduced, this means that some traders that are automatically anticipating on a data surprise have some latency in their ordering execution. This means that their trades are executed some time later, causing longer durations. The market maker or the

institutions that are responsible for the liquidity in the market could look at the type of news that comes out and pre-set available liquidity and this influences durations indirectly. Or they could look at the size of the data surprise and reduce liquidity right

(17)

15 when the news come out and thus influence durations indirectly. There is a further aspect. Not every trader is able to or wants to trade automatically. It is possible for some reasons that trading volume is also high right after the news release compared to right at the news release. It is clear that there are numerous aspects that could influence

durations, but recent literature does not evaluate these aspects. This thesis will evaluate four possible aspects that could influence the jump durations. `

As recent literature proof that for a number of U.S. macroeconomic indicators the amount of surprise in the data has effect on the price of the EUR/USD exchange rate, this could possibly also have an effect on the durations of these jumps. A larger move could take more time. To find the effect of a data surprise on the duration of jumps, we estimate the following accelerated failure time model:

ln (𝑇𝑗) = 𝑐 + ∑15𝑖=1𝛽𝑖𝑆𝐴𝑖,𝑗+ 𝜀𝑡𝑗. (6)

As trading has become more automated in recent years by using computer technology that automatically anticipates on news announcement, the question rises if these durations have become significantly shorter in recent years. Let 𝐷𝐴𝑌𝑗 denote the

amount of days between news arrival 𝑗 and the day corresponding to the first tick data point in the dataset that is used in this thesis. This means that 𝐷𝐴𝑌_𝑗 is linear in real time. The hypothesis of interest is that the coefficient on the linear trend (in real time) is negative and average durations reduce over time.

The type of news could also have an effect on these jump durations. Some news announcements are considered to be more important than others and Chaboud et al. (2004) show that the trading volume after a news release substantially differs between U.S. macroeconomic indicators. This means that there is more focus and trading on some U.S. macroeconomic indicators than others, which could influence jump durations.

To look if uncertainty in the market influences these durations, realized volatility of the EUR/USD exchange rate, based on short term raw tick data, has been added as a measure of this uncertainty. First the following accelerated failure time model is fitted to estimate the marginal impact of all these aspects on jump durations:

(18)

16 ln(𝑇_𝑗) = 𝑐 + ∑15 𝛽_𝑖𝐷_𝑖,𝑗

𝑖=1 + 𝛽16𝐷𝐴𝑌𝑗+ 𝛽17𝜎𝑗+ 𝜀𝑡𝑗, (7)

where 𝐷_𝑖,𝑗 is a dummy variable which takes on value 1 when on news arrival 𝑗 U.S. macroeconomic indicator 𝑖 is released (0 otherwise). In (7) 𝜎_𝑗 is the realized volatility of the last 100 prices of the EUR/USD exchange rate issued before the news release and can be defined as:

𝜎_𝑗 = √∑ ln ( 𝑃𝑗,−𝑡

𝑃𝑗,−𝑡−1)

2 100

𝑡=1 , (8)

where 𝑃_𝑗,−𝑡 is the 𝑡ℎ_{price issue of the EUR/USD exchange rate before news release 𝑗.}

Hereafter, we use a one-time backward elimination to improve the model. A restricted set of regressors that represents the most influential factors (significant at the 10% level) is identified and this reduced model is estimated.

(19)

17

4 Data

In this section the data that has been used in this study is described, as well as the dataset as a result of the filter described in Section 3.1

4.1 The tick data

The EUR/USD exchange rate tick data that has been used in this theses is obtained from Dukascopy Bank SA. Dukacopy Bank SA offers direct access to the Swiss Foreign Exchange Marketplace which provides the largest pool of electronic communication network (ECN) spot forex liquidity available for banks, hedge funds, other institutions and investors. An ECN is a type of computerized forum or network that facilitates the trading of financial products outside of traditional stock exchanges. Every price of the EUR/USD exchange rate that is issued by Dukascopy Bank SA comes directly from the Swiss Foreign Exchange Marketplace and is therefore transparent. These prices give a good view of what the global price of the EUR/USD exchange rate is at that moment. Price issues from January 2010 up to and including February 2015 have been used for evaluation. This leaves approximately 250 million bid and ask price issues which are processed by the filter.

4.2 The U.S. macroeconomic indicator data

The release time and date of every U.S. macroeconomic indicator that has been used in this thesis and was released between January 2010 and March 2015 are provided by ForexFactory.com1. This also holds for the prediction and actual values of these U.S. macroeconomic indicators. The release times and dates are fixed and every value of a macroeconomic indicator is scheduled to be released at a specific time and date. ForexFactory.com is currently the number one most viewed forex-related website as

1_{Forex Factory, Inc. is a United States corporation headquartered in Tampa, Florida. ForexFactory.com is}

(20)

18 ranked by internet statistic firm Alexa Internet2. The data releases itself come from various institutions like government departments, universities and private companies. Most of the U.S. macroeconomic indicators that are being evaluated in this thesis are the same as in recent literature. In papers and websites regarding forex trading, these indicators are noted as important and to have high impact on the EUR/USD exchange rate. An overview of these indicators can be found in Table 2. In this table and thesis quarter over quarter can be defined as q/q and month over month as m/m. Using common macroeconomic theory, one can conclude that a higher value of all these macroeconomic indicators, with an exception for the Unemployment Claims, indicate a stronger U.S. economy. Because of interpretation reasons, all standardized data

surprises of the Unemployment Claims have been multiplied by -1.

Table 2: List of evaluated macroeconomic indicators and the corresponding release institution.

4.3 The filtered dataset

The filter as described in Section 3.1 uses the data described in Section 4.1 and 4.2 to derive a unique dataset that is used to answer the main question of this thesis. It contains 886 news releases. A histogram with the distribution of the sample of 𝑙𝑛 (𝑟_𝑗) can be found in Figure 1 in the Appendix. A histogram with the distribution of the

2_{Alexa Internet, Inc. is a U.S. California-based subsidiary company of Amazon.com which provides}

commercial web traffic data. The referred ranking was published on June 15 2015. See http://www.alexa.com/siteinfo/forexfactory.com for full statistics of ForexFactory.com.

(21)

19 sample of 𝑟𝑗can be found in Figure 2 in the Appendix. These images give visual

evidence that the amounts of positive and negative returns following a news release are in this sample evenly distributed. Basic jump duration statistics extracted from this dataset can be found in Figure 3 and Figure 4 in the Appendix. Figure 3 shows clearly that the durations are concentrated on the far left side of the corresponding axis, and thus near zero. 75% of all durations derived by the filtration are of 1 minute or less and 43% of 10 seconds or less. This substantiates the previous literature that short and large price movements after news releases happen in a short period of time. The first bin in Figure 3 has been expanded in Figure 4 to give insight of how this first bin is

distributed. It is shown here that even 19% of all durations in the filtered dataset are of 1 second or less. There can be drawn an important conclusion from this histogram if there can indeed be shown that there is a relation between the jump returns and the

standardised surprises and between these durations and some of the known variables discussed in Section 3.3. If this is the caste, this histogram then shows us that a fixed time window is not efficient: the durations are not fully concentrated near a single value.

(22)

20

5 Results

This section will report the estimation results of the regression models described in the Section 3. The method that has been used to estimate these regression models is OLS. This section will be divided in two topics: one about the effect on jump returns and one about the effect on jump durations.

5.1 The effect on jump returns: the results

Using the obtained dataset from our filter, we are now able to estimate the effect of data surprises on jump returns. This is the main estimation in previous relevant literature and therefore important for comparison. In order to do this, we estimated the multivariate regression model (3). The results are reported in Table 3 and the results from the reduced model with only significant regressors are reported in Table 4. For interpretation reasons, coefficients are multiplied by 100. If we look at Table 3, the interpretation can be as follows: an increase of one unit of the standardized surprise at the release of the Core CPI m/m devaluates the EUR/USD exchange rate with 0,0418% in the first large price movement. The 𝑅2_{as a result of estimating the reduced model is}

13,28%. One can conclude that 13,28% of the first large price movement after these news releases can be explained by these data surprises. This table shows some interesting results. First of all, it strongly depends on the type of news whether a standardized data surprise has significant explanatory power. The macroeconomic indicators whose standardized data surprise has a significant influence on jump returns at the 1% significance level are the Advance GDP q/q, Non-Farm Employment Change and the Core CPI m/m. The macroeconomic indicator whose standardized data surprise has a significant influence on jump returns at the 5% significance level is the ISM Manufactruring PMI. The macroeconomic indicators whose standardized data surprise has a significant influence on jump returns at the 10% significance level are the Prelim GDP q/q, Unemployment Claims and Core Retail Sales m/m. This means that in seven of the 15 macroeconomic indicators that have been evaluated, significance is found of at least the 10% level. Furthermore, in 14 of the 15 macroeconomic indicators that have

(23)

21 been evaluated, the coefficient is negative. For these 14 indicators the coefficient is in line with the hypothesis that a positive data surprise of a U.S. macroeconomic indicator has a negative effect on the EUR/USD exchange rate in the first large price movement. This means that, given the knowledge of the value of the data surprise, for a lot of macroeconomic indicators the first large price movement after a data surprise is highly predictable. It can also be concluded that in this set of macroeconomic indicators, the Non-Farm Employment Change is has the most significant and largest effect on the jump returns. This is in line with almost every research regarding U.S. macroeconomic indicators.

It is interesting to see the difference in explanatory power and coefficient size between the Advance GDP q/q and the Prelim GDP q/q. Every quarter there are three U.S. GDP estimates released from which the Advance GDP q/q is the first and thus tends to have the most impact. This is also the case for the difference between

explanatory power between the Unemployment Claims and the Non-Farm Employment Change. The Non-Farm Employment Change is considered to be the earliest monthly data considering the job market in the U.S. and thus tends to have the most impact. These two cases are in line with this research and we can conclude that in these cases earliness of the data increases explanatory power and impact of the data surprise.

We conclude that for most positive data surprises a devaluation of the EUR/USD exchange rate follows in the first large price movement and vice versa. The question now rises if there is asymmetry in the effects. Multivariate regression model (4) has been estimated and Wald tests of coefficient restrictions are performed on the

coefficients to test the equality of reactions to positive and negative news surprises. The results are reported in Table 5. For some macroeconomic indicators there is indeed an asymmetric response. The macroeconomic indicator with by far the lowest Wald test p-value is the Non-Farm Employment Change: 0,27%. Positive news about this

macroeconomic indicator has significantly more impact than negative news about this macroeconomic indicator. This was also the indicator where its data surprises had the highest and most significant effect on jump returns. At a 10% significance level, we can conclude that there is also an asymmetric response to data surprises at the release of the Prelim GDP q/q, Unemployment claims, ISM manufacturing PMI and Pending Home

(24)

22 Sales m/m. From these four macroeconomic indicators, only at the Pending Home Sales negative news has significantly more impact than positive news. Remarkable is that in four out of five significant asymmetric response cases, the positive surprise coefficient is smaller than the negative surprise coefficient, which is not in line with Andersen et al (2002).

There is another important aspect that the estimation results of the multivariate regression model (4) show. At a 10% significance level the macroeconomic indicators Advance GDP q/q, Prelim GDP q/q, Unemployment Claims, ISM manufacturing PMI and Core Retail Sales m/m show significance at the positive surprise coefficients, but not at the negative surprise coefficients. The data surprises of these macroeconomic indicators have a significant effect on the jump return when the data surprise is positive, but not when it is negative. The other way around holds for the Trade Balance, Core CPI m/m, ADP Non-Farm Employment Change and Pending Home Sales m/m. The cause of this could be something fundamental, but the results of Table 5 do not rule out that behaviour rather than rationality plays an important role in the anticipation of these data surprises.

In order to test whether there is an interaction effect of concurrent data surprises on jump returns, multivariate regression model (5) was estimated. No significant effects of the combinations where found. This means that when multiple macroeconomic indicators cause a data surprise at exact the same time, they do not strengthen or weaken each other in the effect on the jump return.

(25)

23

Table 3: Estimation results of multivariate regression model (3). Superscripts “***”, “**” and “*” represent significance at the 1%, 5%, and 10% level, respectively.

Table 4: Estimation results of reduced multivariate regression model (3) where only the significant regressors, at the 10% level, in Table 3 are used as regressors. Superscripts “***”, “**” and “*” represent significance at the 1%, 5%, and 10% level, respectively.

Table 5: Estimation results of multivariate regression model (4). Superscripts “***” and “**” represent a significance asymmetry in the effects at the 1% and 10% level, respectively.

(26)

24 5.2 The effect on durations: the results

It is now clear how data surprises of macroeconomic indicators have an effect on jump returns. As is shown in Section 5.1, it varies between the macroeconomic

indicators whether its data surprise has an effect on the jump return. In order to test whether the size of a data surprise has effect on the durations of jumps, accelerated failure time model (6) has been fitted. The results where important: there is no

significant effect of data surprises on jump durations. This means that whatever the size of the data surprise is, it does not have a significant effect on jump durations: the anticipation speed remains the same.

As explained in Section 3, this thesis will evaluate some other aspects that could have an influence on jump durations also. Accelerated failure time model (7) is fitted to estimate the marginal impact of all these aspects on jump durations. The results are reported in Table 6 and the results from the reduced model with only significant

regressors can be found in Table 7. The results in Table 7 can be interpreted as follows: if the actual figure of the Non-Farm Employment Change is released and there is a data surprise, then the duration of the subsequent jump increases with 55,31%. Eight out of 15 dummy variables have a significant effect on jump durations. In Table 7 the dummy variables of the ADP Non-Farm Employment claims, Core Durable Goods Orders m/m, ISM Manufacturing PMI , Pending Home Sales m/m, UoM consumer sentiment and New Home Sales show significance at the 1% level. The dummy variable of the Non-Farm Employment Change shows significance at the 5% level. The dummy variable of the Prelim GDP m/m shows significance at the 10% level. For these eight

macroeconomic indicators, this result indicates that it is the type of news rather than the size of the data surprise that influences the duration of the jump. Important is that the release of the Non-Farm Employment Change has the largest significant positive impact on the duration as also did the size of its data surprise had on the jump returns. The release of the ADP Non-Farm Employment Change has the most significant negative impact on the duration.

Furthermore the coefficient of the variable 𝐷𝐴𝑌_𝑗 shows significance at the 1% level. The coefficient of 𝐷𝐴𝑌_𝑗 is negative which does not reject the hypothesis that

(27)

25 anticipation to news releases has become significantly faster in recent years. It can be interpreted as a 0,17% decrease in duration for one added day in time. The coefficient of the raw tick data volatility is also significant at the 1% level. This coefficient is

negative, meaning that more uncertainty in the market causes shorter durations.

Table 6: Estimation results of accelerated failure time model (7). Superscripts “***”, “**” and “*” represent significance at the 1%, 5%, and 10% level, respectively.

Table 7: Estimation results of reduced accelerated failure time model (7), where only the significant regressors, at the 10% level, in Table 6 are used as regressors. Superscripts “***”, “**” and “*” represent significance at the 1%, 5%, and 10% level, respectively.

Dependent variable

***

(28)

26

6 Concluding Remarks

In this thesis, the short term behaviour of the EUR/USD exchange rate after the release of a U.S. macroeconomic indicator was investigated. Recent relevant literature states that data surprises in U.S. macroeconomic indicators have an effect on the

EUR/USD exchange rate. The short term effect of these data surprises on the EUR/USD exchange rate is defined in the literature as a large price movement or jump in the EUR/USD exchange rate that occurs in a short period of time. This short period of time is defined as the jump duration and in the literature, other descriptions of this jump duration are given. Ranging from ten seconds to several minutes, these descriptions vary a lot. These studies use fixed time window high frequency data to calculate jump returns, but it is not excluded that a jump duration is shorter than the used fixed time window. The jump could already be over before the end of the fixed time window is reached. Therefore, the use of high frequency data to compute jump returns has to be discussed. Despite this possible problem, there is no known technique to identify these durations. A big contribution of this paper is that it focussed not only on identifying jump returns, but also on jump durations. Using this identification we investigated the relationship between a data surprise in a macroeconomic news arrival, the subsequent jump in the EUR/USD exchange rate and the duration of this jump.

This thesis has introduced a new filter to extract these jumps from EUR/USD

exchange rate tick data. This provides a different picture of these jumps. This filter uses generally accepted definitions of a jump to extract these jumps from a big tick data set containing approximately 250 million EUR/USD bid and ask prices from January 2010 up to and including February 2015. As a result, a unique dataset was formed containing 886 news releases with the corresponding jump returns, durations and other data. The size and uniqueness of this dataset can be seen as a big contribution to recent relevant literature. This dataset gave answer to the question of how long these durations actually are. 75% of all durations derived by the filter are of 1 minute or less, 43% of 10 seconds or less and 19% of 1 second or less. This means that it is true that the large price

movements occur in a very short period of time, but they are not concentrated near one minute but rather near zero seconds.

(29)

27 Using this dataset, multivariate regression models have been fitted in order to estimate the effect of data surprises on jump returns. It can be concluded that for most positive data surprises in a U.S. macroeconomic indicator a devaluation of the

EUR/USD exchange rate follows in the first large price movement and vice versa. Also, the data surprise in the U.S. Non-Farm Employment change can be seen as the one that has the largest and most significant impact on the EUR/USD exchange rate in the first large price movement. Furthermore, there are several macroeconomic indicators whose data surprise has asymmetry in the effect on the EUR/USD exchange rate. In most of these cases, positive news has more impact than negative news. Also, there is no interaction effect of concurrent data surprises on jump returns.

Furthermore, this thesis investigated the effect of different aspects on the size of jump durations. Four different aspects have been evaluated by fitting accelerated failure time models and some important conclusions have been drawn. First of all, there is no significant effect of different data surprises on jump durations. While for most U.S. macroeconomic indicators the data surprise has a significant effect on the jump return, it doesn’t significantly influence the duration of this jump. For most macroeconomic indicators, it is the type of the macroeconomic indicator that has a significant effect on the jump duration. This means that the average jump duration at a news release with a data surprise depends on what kind of macroeconomic indicator causes this data surprise rather than the size of this data surprise. The release of the Non-Farm Employment Change has the largest and most significant positive impact on jump durations, independent of the size of its data surprise. The release of the ADP Non-Farm Employment Change has the largest and most significant negative impact on jump durations, independent of the size of its data surprise. Also market uncertainty,

measured by the realized volatility of the EUR/USD exchange rateprior to the news

release, based on raw tick data, significantly decreases durations. And at last, this thesis shows that anticipation of traders to data surprises has become significantly faster in recent years: average durations reduce over time.

In an efficient market all jump durations would be close to zero. This thesis shows that this is not the case. The jump return of a jump following a data surprise is influenced by this data surprise, but the length of this jump is influenced by other

(30)

28 factors and is often not approximately zero. And since it is influenced by other factors, a suitable way to find the effects on jump returns is to use a variable time window rather than a fixed time window in the price data.

This technique can be used on a large variety of possible news effects. First of all, the used macroeconomic indicator dataset can be enriched with more indicators of different types of countries to estimate the effect of data surprises of these

macroeconomic indicators on the EUR/USD exchange rate or other exchange rates. But besides forex trading, this technique can also be used to estimate for instance the effect of a data surprise in the quarterly profits of a company on its stock value. An important aspect that can be examined in further research is the effect of available liquidity in the forex market at the time of a news release on jump returns and jump durations. As mentioned earlier, reduced liquidity could drive the jump durations away from zero. Nevertheless it is clear that these kinds of studies have to been done with tick data and variable time windows, since this thesis shows that the size of jump durations depends on various factors and are not fixed.

(31)

29

Bibliography

Andersen, G.T., Bollerslev, T., Diebold, F.X., and Vega, C. (2002). Micro Effects of

Macro Announcements: Real-Time Price Discovery in Foreign Exchange. National

Bureau of Economic Research Working Paper Series, 8959.

Balduzzi, P., Elton, and E.J., Clifton, T. (2001). Economic news and bond priced:

evidence from the U.S. Treasuy market . The Journal of Financial and Quantitative

Analysis, 36: 523-543.

Chaboud, A.P., Chernenko, S.V., Howorka, E., Iyer, R.S.K., Liu, D., and Wright, J.H. (2004). The High-Frequency Effects of U.S. Macroeconomic Data Releases on Prices

and Trading Activity in the Global Interdealer Foreign Exchange Market. Board of

Governors of the Federal Reserve System, International Finance Discussion Papers, 823 (2004).

Chatrath, A., Miao, H., Ramchander, S., and Villupuram, S. (2014). Currency jumps,

cojumps and the role of macro news. Journal of International Money and Finance, 40

(2014): 42-62.

Chaudhry, M., Ramchander, S., and Simpson, S. (2005). The impact of macroeconomic

surprises on spot and forward exchange markets. Journal of International Money and

Finance, 24 (2005) 693-718.

Chueng, Y.-W., and Chinn, M.D. 2001. Currency traders and exchange rate dynamics:

a survey of the U.S. market. Journal of International Money and Finance, 20 (2001):

439-471.

Ederington, L., and Lee, J.H. (1993). How markets process information: news releases

and volatility. Journal of Finance, 48(1193): 1161-1191.

Ehrmann, M. and Fratzscher, M. (2005). Exchange rates and fundamentals: new

evidence from real time-data. Journal of International Money and Finance, 24 (2005):

317-341.

Evans, D.D.M., and Lyons, R.K. (2005). Do currency markets absorb news quickly? Journal of International Money and Finance, 24 (2005): 197-217.

Evans, D.D.M., and Lyons, R.K. (2008). How is macro news transmitted to exchange

(32)

30

Evans, K.P., and Speight, A.E.H. (2011). Intraday euro exchange rates and

international macroeconomic announcements. The European Journal of Finance, 17:2:

83-110.

Faust, J., Rogers, J.H., Wang, S.-Y.B., and Wright, J.H. (2007). The high-frequency

response of exchange rates and interest rates to macroeconomic announcements.

Journal of Monetary Economic, 52 (2007): 1051-1068.

Fleming, M.J., and Remolona, E.M. (1997). What Moves the Bond Market? Federal Reserve Bank of New York Economic Policy Review, 3 (1997): 31-50.

Hashimoto, Y., and Ito, Y. (2010). Effects of Japanese macroeconomic statistic

announcements on the dollar/yen exchange rate: High-resolution picture. Journal of

The Japanese and International Economies, 24 (2010): 334-354.

Ozdagli, A.K. (2013). Not So Fast: High Frequency Data for Macroeconomic Event

(33)

31

APPENDIX

Figure 1: Distribution of the sample of 𝐥𝐧 (𝒓𝒋) following the 886 news releases in the dataset obtained by using the filter described in Section 3.1. The number under each bar represents the beginning value of the corresponding histogram bin. The sample mean of 𝐥𝐧 (𝒓𝒋) equals -7.30E-06 and the sample standard deviation of 𝐥𝐧 (𝒓𝒋) equals 0.0011.

Figure 2: Distribution of the sample of 𝒓𝒋 following the 886 news releases in the dataset obtained by using the filter described in Section 3.1. The number under each bar represents the beginning value of the corresponding histogram bin. The sample mean of 𝑟𝑗 equals 0.9999 and the sample standard deviation of 𝒓𝒋 equals 0.0011. d described in

(34)

32

section 3.1. The numasssssssssssssssssssssssssssssssssssssssssasassssssssssssssssssssssssssssssssssber under each bar

Figure 3: Distribution of the sample 𝑻𝒋 following the 886 news releases in the dataset obtained by using the filter described in Section 3.1. The number under each bar represents the beginning value of the corresponding histogram bin. The line corresponds to the cumulative percentage of the total sample size. The sample mean of 𝑻𝒋 equals 40.50 and the sample standard deviation of 𝑻𝒋 equals 60.75.

Figure 4: Distribution of the sample 𝑻𝒋 following the 886 news releases in the dataset obtained by using the filter described in Section 3.1. This histogram shows only durations of 10 seconds or less. The number under each bar represents the beginning value of the corresponding histogram bin. The line corresponds to the cumulative percentage of the total sample size. The sample mean of 𝑻𝒋equals 40.50 and the sample standard deviation of 𝑻𝒋 equals 60.75.