Determining the impact of news events on
return and volatility of European indices
Author: C.V. Seinen (6050980) Supervisor: Prof. Dr. H.P. Boswijk 2nd Supervisor: Dr. N.P.A. van Giersbergen Date: 29 – 05 – 2015
University of Amsterdam Faculty of Economics and Business Master Financial Econometrics Master Thesis
Table of contents
1. Introduction ... 2 2. Literature... 4 2.1 Structure of information ... 4 2.2 Political and Social instability ... 5 2.3 Terrorism ... 7 2.4 International Cohesion ... 7 2.5 Regular event study models ... 8 2.6 Distributional Event Response Model (DERM) ... 9 3. Models and Methods ... 11 3.1 Distributional Nearest Event Response Model (DNERM) ... 11 3.2 Selecting Influential Event Types ... 13 3.3 Regression model and GARCH model ... 14 3.4 Procedure ... 15 4. Simulations ... 17 4.1 General Set‐Up ... 17 4.2 Small Sample: 100 observations ... 18 4.3 Large Sample: 1000 observations ... 19 4.4 Fit to the true model ... 20 4.5 Different Event Types ... 21 5. Data ... 22 5.1 Market Data ... 22 5.2 GDELT ... 22 5.3 Weekends and Holidays ... 25 6. Results ... 26 6.1 Instructions ... 26 6.2 Germany ... 27 6.3 France ... 29 6.4 Great Britain ... 31 6.5 The Netherlands ... 34 6.6 Summary ... 36 7. Conclusion ... 37 7.1 Conclusion ... 37 7.2 Limitations ... 38 8. References ... 40 Appendix………431. Introduction
What information becomes public and what is disregarded is a mysterious process and differs widely across different eras. While political or social sentiment play an important role, often there is a lack of availability and accessibility to multiple sources of news. In most European countries the availability and accessibility to news has greatly improved over the past decades, virtually anyone with an internet connection has access to vast amounts of information. However the ever increasing information quantity makes it harder to qualify the importance of each piece of information. In financial markets information is the primary source of influence to the market, but to distinguish useful information from useless information is not straightforward. A great amount of studies have been published trying to assess the impact of information events on the movement of the market or its assets. Examples are Bialkowski et al. (2008) who research the effect of elections on stock market behavior, or Karolyi and Martell (2010) who measure the impact of terrorist attacks against companies, on the value of the company stock value. The common denominator of most of these event studies is that they explicitly choose which events they want to examine. For various reasons this is completely sensible as the studies are dependent on data availability, accessibility, but of course mainly on what the researcher is trying to explain. In many cases this arbitrary choice in variables is a restriction on the available data and perhaps inaccurate. As news availability and accessibility becomes ever greater, it therefore becomes possible to follow a more unconstrained approach to select which variables to choose in determining the impact of certain news events. Many variables have some minor or major effect on the market, but how this is reflected into the market is generally an unknown process. Events could have a direct effect on returns or indirectly by influencing the volatility. Also the structure of the market can be of influence on what types of news are influential, so that there are differences across countries. Central research question in this thesis is therefore: are expected returns and volatilities affected by different types of news, and are these differences consistent across countries? Information comes available at a certain point in time, but the effect of this information on the market is likely to persist for some time or maybe already present in the market. Event studies mainly consider these effects exclusively for expected returns or volatility, though some examples exist that consider them jointly like Arin et al. (2008). They investigate the influence of terrorism using a Generalized Autoregressive Regressive Conditional Heteroskedasticity (GARCH) model (Bollerslev, 1986) in which they specify event variables in both mean and variance equation, to assess their influence on both returns and volatility. They use this model structure for several countries to evaluate the performance under different circumstances. This thesis follows this example and tries to generalize this model to the use of multiple events.To determine how the response to events is incorporated into the market, the Distributional Nearest Event Model (DNERM) is proposed, which is a modification of the Distributional Event Response Model (DERM) proposed by Rucker et al (2005). The model imposes a time specific response structure to the event types using non‐linear methods. The dummy structure of regular event variables is thereby transformed to a dynamical structure. The transformed event variables are then used in two types of models that are regularly used in event studies (Eckbo, 2008), namely a linear regression model and a GARCH model. The focus of modelling the data is split in explaining variations in expected returns and variations in volatility. The hypothesis is that event types can be influential on both returns and volatility, but also that event types can be exclusively influential on one of the two; so that they should be considered separately. Empirical data on news events is used of four European countries: Germany, France, United Kingdom and the Netherlands. The particular choice for the four European countries is mainly based on availability of market and event data. The indices of these countries are used to gauge the effects of news events. The daily returns, based on market indices, are used to determine the effect of the events on expected returns and the daily differences of the implied volatility index are used for the volatility. The implied volatility index is used for volatility, because volatility is not directly observable from standard daily returns, therefore the implied volatility index can be used as a proxy (Tsay, 2005). The data on events comes from the Global Database of Events, Language and Tone (GDELT) project. This database contains news events from print, broadcast and web formats, from January 1979 through present day. This database contains extensive amounts of data including many types of events and characteristics. Due to the amount of data, selections are made to assure that the modelled events have a non‐trivial connection to market returns and volatility, and also to account for computational power. Rather than making an arbitrary choice, the Least Absolute Shrinkage and Selection Operator (LASSO) is applied to make an automated selection on all the available variables. The rest of the paper is organized as follows. The second section gives a review of findings in studies that are regularly used. The third section discusses the models and methods applied and gives the step by step procedure followed in this thesis, to derive results. The fourth section contains simulations and robustness checks on the DNERM. The fifth section discusses the processing and cleansing of the used data. The sixth section is concerned with the results of the empirical data. The seventh section contains a summary of results with its implications and gives directions for future research.
2. Literature
News has always been highly influential on stock markets. Facts and sentiments are the ingredients of news, but how the mixture is chosen is important on how the news is perceived. The literature connecting stock markets to news is extensive. In the first paragraphs an overview of findings in previous empirical work will be given. The last two paragraphs contain background on regularly applied models and a first insight into the improvement of the DERM. 2.1 Structure of information Information itself is very heterogeneous of origin as its effect on asset markets can by very specific. Andersen et al. (2003) and Rucker et al. (2005) found that the reaction of exchange rates to important news events is asymmetric. The impact of certain news events is the biggest around the release and then quickly dies out, but the rate at which the impact decreases is not similar. Andersen et al. (2007) shows in addition that the asymmetric pattern is not affected by the business cycle of a given firm. They state that generally there exists distinct correlation between news response and asset market returns. Akhtar et al. (2012) links the response pattern to sentiment of the news. They confirm the asymmetric behavior of the response pattern, but in addition they conclude that the pattern of positive news is different from negative news. Tetlock et al. (2008) also find that tone used in news stories has a significant effect. Specifically negative words used in news stories forecast lower future returns. Moreover they find that the tone of the news can capture specific aspects of a firm that are otherwise difficult to quantify. They also find that negative news is incorporated into the stock prices with a delay compared to positive news. This time aspect to the tone of the news has also been investigated by Zhang (2006). He diversifies between low and high uncertainty stocks and finds that the market reaction to new information differs. Low‐ uncertainty stocks are practically unchanged by any news, while with high‐uncertainty stocks positive news predicts higher returns and negative news predicts lower future returns. Chan (2003) also finds evidence that negative news has a delayed effect in stock prices. He finds that the behavior of stock prices in the form of drift from its mean, is mostly due to the negative news events and can last up to one year. When bad news is released again it takes some time for it to be reflected into the market price, but when the investors react they overreact causing excess trading volume and higher volatility. From this Chan determines that increased volume in trades leads to increased volatility. As well as there are different types of certainty in asset markets, there is also some degree of certainty in news, that is, not all news comes as a surprise. Governments and companies makeregular statements about their progress in scheduled announcements and are considered macro news. Rangel (2011) studies the difference between scheduled macro news and unscheduled macro news on the volatility of assets. He finds that unscheduled macro news has a prolonged impact on volatility with respect to scheduled macro news. The arrival of macro news itself also has a direct impact, following traditional pricing models, especially on currency markets. Evan and Lyons (2008) show that variance of daily prices is determined by the arrival of macro news for more than 30%. Again this effect will not be incorporated into the prices immediately, but will experience some sort of delay. They also pose that the situation in currency markets is best described by a model in which dealers follow a wait and see strategy, as dealers first want to witness the actions of traders before they set prices. In stock markets not all stocks receive the same amount of news coverage. The research of Chan (2003) uses headlines of big newspapers and links them to stock market movement. As mentioned he finds that public news can lead to abnormal returns of stocks, but he also finds that some stocks have extreme returns without being mentioned in headlines of newspapers. Fang and Peress (2009) show that stocks with no news coverage outperform stocks with high news coverage frequently. Their result even holds after accounting for risk factors as size, book‐to‐market, momentum and liquidity of various stocks. Both papers clearly indicate (company) news cannot account for all movements in the stock market. 2.2 Political and Social instability News itself is a much aggregated term as it contains dozens of various topics. News topics chosen to describe the behavior of the market or stocks are often chosen to be primarily financial; like in previous mentioned articles. But of course many different types of topics can influence the stock market. One such topic is described by Beaulieu et al. (2005). They focus on the influence of political news on the volatility of stock returns. They find that in firms with high exposure to political risk, the degree of favorability, of the news, influences the volatility, with unfavorable news increasing the volatility and favorable news decreasing the volatility. These firms are characterized by having purely domestic assets or heavily relying on international operations in politically unstable region. Again negative political news has a more robust effect than positive news. The political climate can be variant of nature in some countries, which has its effect on the associated domestic stock markets. Cuadra and Sapriza (2008) show that emerging markets tend to experience larger political uncertainty which leads to higher volatility. Diamonte et al. (1996) researches the discrepancy between emerging and developed markets and finds that emerging markets are more prone to political risk, which causes negative impact on stock returns. They also determine that when political risk increases, the response of the market is more intense in emerging
markets. The impact of the change of political risk is found to be a statistically significant determinant of stock returns in emerging and developed markets, but not in developed markets. In addition Aggarwal et al. (1999) shows that in emerging markets periods of high volatility are predominantly caused by domestic news events, rather than global events. Political risk has a governmental and a civil component. Governmental actions like elections, labor disputes, corruption and reform are found to be influential on the business environment by research conducted by Asteriou and Siripoulus (2000). They find that policy uncertainty has negative effects on the development of equity markets. In addition they find that this effect is more present when the policy uncertainty is concerning productive economic decisions, which is also concluded by Asteriou and Price (2001). Politically instable countries, in which elections are imminent, implying that there is uncertainty about future policies, therefore drive away risk‐averse economic agents and thus hindering the economy. Elections themselves are not frequent events and can be rather constitutional or emanate from forceful overthrowing. Bialkowski et al. (2008) research the effect of elections on stock markets by analyzing several industrialized countries. They find that the country specific component of volatility at least doubles during elections. Much of this risk emanates from balloting periods prior to the elections and can cause heavy fluctuating returns. Durnev (2010) shows that corporate investments respond significantly to political risk surrounding elections. In countries with high corruption and governmental censorship, this impact is even more sustained. Outcomes of elections are, at least in a democracy, dependent on the public opinion. This civil component of political risk is a hard to measure variable as it is ever changing. Although polls are used to measure the public opinion, no such continuous measure exists. However not specifically intended for elections, Bollen et al. (2011) tries to overcome this gap and uses social media to quantify the public opinion/mood to link it to the stock market. They state that public mood or sentiment is likely to be equivalently as important to stock market as the news. This statement about the influence of public mood is supported by articles of Melvin and Tan (1996) and King and Soule (2007). Melvin and Tan (1996) show that civil unrest increases volatility on exchange rates of countries directly after such an event and indirectly by adverse reaction of investors on the increased political risk the country faces. Though civil unrest can spark at any moment, not all such events are purely random as they can be incited by social movements that try to raise public awareness about certain matters. This is the conclusion of research done by King and Soule (2007). They show that activists have indirect influence on asset markets through the use of media. They emphasize that the media is twofaced in the sense that media mediates the news that they receive from activists but also functions as their
spokesperson. For that matter successful activists can have a firm grip on the news. Besides rather mild civil/political unrest such as strikes, protests and demonstrations, more severe types exist like bombings, coup attempts and assassinations. Crowley and Loviscek (2002) research the effect of such severe civil unrest in Latin American countries. Their result indicate that such an occurrence, systematically impacts currency markets and can last for up to three months. 2.3 Terrorism In the area of political unrest, terrorism is a class of its own. Though not an invention of modern times, its impact on asset markets seems to be ever more existent, since the attacks of 2001 in the United States, caused an international financial crisis. Terrorism seems therefore to be more dangerous than before as terrorists are more likely to resort to weapons or tactics that inflict mass casualties. Tucker (2001) mainly contradicts this notion by pointing out that, besides heavier weaponry, essentially nothing has changed. He states that the impact of a terrorist attack depends on its strategic planning. Abasie and Gardeazabal (2003) show in a case study about Basque country (Spain) associated stocks, that in the period of truce following intense terrorist activity, that Basque stocks start to outperform non‐Basque stocks. After the ceasefire this quickly turned around again, implying the history associated threat of terror being a heavy burden on these stocks. Chen and Siems (2004) find evidence that investors flee the market after terrorist attacks. Because of the speed of today’s worldwide news dissemination, they emphasize that the fleeing can get enormous in a short amount of time causing serious devaluation problems on the stock market. Arin et al. (2008) finds that this response to terrorism varies across countries, but is statistically significant in both mean and variance for all countries that they have examined. Karolyi and Martell (2010) zoom in on terrorism that affects companies and find that terrorist shocks to companies share prices are mostly affected by human capital losses, such as kidnappings. Physical losses, such as destroyed buildings, have far less impact. Again these shocks have a significant country specific component, but in addition they find that attacks on companies in countries that are considered safe, are associated with larger negative price reactions. 2.4 International Cohesion Conflicts are not always contained within the borders of one country so there is high propensity for its influence on financial markets to be internationally. Guidolin and La Ferrara (2010) find that conflicts in Asia and the Middle East have a strong effect on global financial markets. This impact predominantly comes forth from international conflicts in these regions. The social structure in these countries is a key predictor as almost all international conflicts taking place in highly polarized
countries have a negative impact on global markets. While this is intuitively comforting Schneider and Troeger (2006) show that not all markets follow this prediction. In particular they find that following some conflictive events in the Gulf, the NYSE responded positive. This means there is evidence that despite an increasingly integrated world economy, not all political events have the same effect. That stock markets co‐move is well known, but what the precise mechanism of this interaction is, is ill understood. Albuquerque and Vega (2009) show, among many others, that world markets are predominantly steered by macroeconomic news from large economies, like from the United States. In their case study of Portugal, they find that returns on Portuguese markets respond to US macroeconomic news, but US markets do not respond to Portuguese macro‐economic news. This implies that large economies have global importance to small economies. Besides the United States other highly influencing markets exist like Germany and United Kingdom. Berben and Jansen (2005) find a significant increase in stock market co‐movement among the previous mentioned markets. For the same markets and the same period Morana and Beltrati (2008) conclude slightly different as they can find no evidence of increasing correlation between the markets, however they do find evidence of consistent global influence on volatility across the markets. Rua and Nunes (2009) add to this that co‐movement of stock returns is highly dependent on frequencies. In general they find that markets co‐move strongly at low frequencies, which means their co‐movement is observed over longer periods of time and not so much on short terms. 2.5 Regular event study models Classical event studies use dummy variables, in linear regression models, to explain differences in the abnormal returns of an asset in a market (Eckbo, 2008). Abnormal returns are returns corrected for the overall market performance. The rationale of this technique is that a correction needs to be made on the standard return for information that affects the return of the whole market. Therefore the movement of the whole market often makes it hard to interpret the coefficient of the asset itself. The abnormal returns reflect the performance of the asset with exclusion of general market movements. The adjustment for the overall market is usually done by using the Capital Asset Pricing Model (CAPM). First the systematic risk is estimated 1 and then the abnormal returns are obtained by subtracting the adjusted market returns from the actual returns 2 . In these equations is the expected portfolio return for a chosen time interval (hour, day, week,
month, etc.), the risk‐free rate, the expected market return, the systematic risk, the return at time and the abnormal return at time .
∙ 1 2 Response to news varies over time, therefore a constant coefficient model will be heteroskedastic (Fama and French, 1989). To allow for heteroskedasticity in the data a GARCH( , ) model (Bollerslev, 1989) is applied 3 . This model is chosen above other models for its unique properties. The advantage of a GARCH model is that it treats heteroskedasticity as a variance to be modelled, instead of trying to resolve the problem of heteroskedasticity. Moreover, it is of particular use in event studies, because for each error term a variance can be computed. This enables the possibility to assess the effects of events directly on volatility.
Equation 3 shows the GARCH model, where a constant term, ε the error term at time , is the amount of ARCH terms, the amount of GARCH terms, the amount of event variables, and
the coefficients of the event types in respectively the mean and variance equation, the event variables, and is an independent and identically distributed (IID) process with mean zero and unit variance. ′ ε where ε and ~ 0,1 ′ β 3 2.6 Distributional Event Response Model (DERM) Usually to model an event, a dummy variable is used to represent the point in time on which the event took place. Then to allow for delayed effects various lags are added to the model. This approach however arbitrarily restricts the response pattern of an event. Proposed by Rucker et al. (2005) the DERM allows the response to be less static, by fitting a distribution on the time interval around the event. This inherently means that an event type will no longer exhibit influence exclusively on the event date, but also after the event and possibly before
the event. The influence after the event date can be interpreted as regular lags, implying a prolonged or delayed affect. The influence before an event could arise if there is common knowledge about the arrival of that particular event, which is then the result of speculations. The DERM model is used to fit a response curve to determine the time specific structure and bandwidth of different event types. The framework of the distribution has to be chosen in advance, but the specifics of the distribution are found by minimizing the sum of squared errors. Equation 4 shows the structure of the DERM model for event types, where are some market covariates, the accompanying coefficients, the number of events of one specific event type, indicates the difference in days between the observation at time and event of event type , the coefficient that serves as an impact factor for each event of event type , … ; is a chosen density function with parameter , and an error term. Each
; at every observation, gets assigned a specific value depending on what density and parameter is chosen. ; , ~ 0, 4 The drawback of this model is that it requires all events to be used in the model as separate events, while applying the same density function. The rational for this framework is that in this way the separate events can get a coefficient assigned to measure the distinct impact of the separate occurrences. However when using large amount of data containing many events, this leads to a great increase in variables in the model. Most of these variables are bound to have insignificant coefficients as these single events will have little importance on the whole time scale. Rucker et al. recognize this and test for joint significance, but find that event types that have significant separate events, are insignificant when considered jointly. This makes interpretation of the estimated distribution of that particular event type rather incomprehensive. To challenge these problems this thesis proposes innovations to this model, which are described in detail in section 3.
3. Models and Methods
This section gives an overview of methods and models applied. The first paragraph explains the features of the Distributional Nearest Event Model (DNERM) and how it is applied. The second paragraph explains the model used to select the event types. The third paragraph briefly describes the use of the DNERM with respect to the linear regression model and the GARCH model. The last paragraph is an overview of the procedure followed to derive results. 3.1 Distributional Nearest Event Response Model (DNERM) An event is often represented by a dummy variable, but when an event has a delayed effect then lags need to be added. While this is common practice in classical event studies, this cherry picking of an order of lags only roughly estimates the time specific structure of the response to that event. As discussed in section 2, Rucker et al. (2005) propose the Distributional Event Response Model (DERM) that allows to fit a time specific response function around each event. Every event type is modelled with an own characteristic distribution, but all events are estimated separately in the model. This is to estimate an impact factor for all events so that they can be distinguished from each other. Applying the DERM to data containing lots of events, therefore, can quickly lead to a very complex model. This thesis proposes a modification to this model by looking at the nearest event at every given time instead of each event separately, hence DNERM 5 . The idea of the modification is to identify the true response pattern of an event type, without adding all events to the model thereby keeping the model parsimonious. The scope of event types is also chosen narrower to be able to make a clear inference on the coefficients. To retain the different magnitudes for events of the same event type, a text mining algorithm proposed by Shook et al. (2012) is used. For each event this algorithm returns a value representing the ‘tone’ of that particular event. This value consists of a score between ‐100 (extremely negative) and +100 (extremely positive). The tone of each event is the percentage difference of words found in an article of a positive lexicon and in a negative lexicon. Equation 5 shows the DNERM model, where |represent the tone value of event type conditional on which event is most near and is the error term which is assumed to be normally distributed. The difference indicator | shows the difference in days between the observation at time and the nearest event of event type conditional on which event is the closest. | | ; , ~ 0, 5
Rucker et al. (2005) use normal and uniform distributions to fit the event types, but results of other studies (Andersen et al. 2003, Andersen et al. 2007, Ahktar et al. 2012) indicate that incorporation of news into the market is asymmetric, so that distributions that allow for asymmetry are explored as well in this thesis. The family of extreme value distributions: Gumbel, Fréchet and Weibull are tried as possible improvements. Respectively the distributions allows for an exponential tail, light tail with finite upper bound and heavy tail with finite lower bound. Except for the Gumbel distribution these class of distributions allows for varying skewness and kurtosis. The Fréchet and Weibull distribution in addition constrain respectively the upper and lower bound of the support, so that tails can be controlled. Also The Exponentially modified Gaussian (EMG) (Foley and Dorsey, 1984), shown in equation 6 , is considered as possible improvement. In this equation is the variable, the mean of the Gaussian component, the variance of the Gaussian component and the rate of exponential component. This distribution allows for a positive skewed version of the normal distribution, but is not necessarily restricted to it. It has, like the normal distribution, infinite support for the difference indicator, but can impose an exponential component to the distribution; positively skewing it in the process. This type of model is commonly used in the field of psychology to study of response times (Palmer et al, 2011). ; , , 2e erfc √2 , 6 erfc is the complementary error function: erfc 1 erf 2 √ e To arrive at the optimal specification of the event types the DNERM is estimated continuously with a search algorithm for varying parameter inputs, until the sum of squared errors is minimized. The search algorithm starts from an initial point and is then adjusted until the sum of squared errors cannot be lowered any further. For accuracy this algorithm is run 1000 times for each of the distributions, each time starting from a different initial point. The parameter values that result in the model with the smallest sum of squared errors, is chosen as the most optimal model. The sum of squared errors is chosen as a comparison device, because the structure of the model does not change by varying and therefore these values are comparable across the different models.
3.2 Selecting Influential Event Types Due to the amount of data, selections are made to assure that the modelled events have a non‐trivial connection to market returns and volatility, and also to account for computational power. So although all event types could possibly be influential, only the event types with the strongest effects are of interest. Rather than making an arbitrary choice in variables, the Least Absolute Shrinkage and Selection Operator (LASSO) is used. This is an automated selection method proposed by Tibshirani (1996). All available event variables and some additional variables are fed to the LASSO. By construction of the event variables, see section 5, the variables contain on average more information on Monday, because events that happen over the weekend are attributed to the first open trading day; which is often Monday. So to correct for possible structural breaks a dummy variable is added which is one on Mondays and zero otherwise. Also, following Fang and Peress (2009), an indicator variable is added which takes value one when there is news and zero when there is no news. In addition to allow for delayed effects of event types, the LASSO is ran multiple times including various lag lengths for all event variables. Event variables that are selected at least once by the LASSO in these iterations, are used as explanatory variables in the linear regression model and GARCH model.
Equation 7 shows the penalized likelihood problem, where is the number of observations, the amount of explanatory variables, the explanatory variables, the estimated LASSO coefficients, the linear regression estimates for event type and ∈ 0, ∞ is the tuning parameter. The LASSO minimizes a penalized sum of squares. The penalty term controls the amount of penalization. Increasing this penalty term will cause some of the coefficients to be shrunk and others to be set exactly to zero (Hastie et al., 2009). argmin 1 2 | | 7 At every point of the parameter , a different model is estimated. To choose the optimal parameter value an ‐fold cross validation is used. For a number the data is split into parts, or folds. ‐fold cross validation then considers training on all but the th part of the data and subsequently
validating this set on the th part, iterating over 1, … , ; i.e. training on , , ∉ and validating on , , ∈ . For each value of the tuning parameter ∈ , … , , for an arbitrary number of evaluating points, the estimate is computed on the training set and the total error of the validation set is recorded 8 .
8
∈
Next the cross‐validation error curve can be fitted for every value of , by taking the average error for each parameter value over all folds 9 and plotting the parameter values. The parameter value that minimizes this curve is then chosen as most optimal point (Hastie et al., 2009). At the most optimal point the LASSO returns a set of explanatory variables, possibly containing the original events, lagged events and/or the other covariates. CV 1 9 All the used event variables are of in a tone adjusted dummy structure form, which means that the variables contain the tone value of the event at the time it occurred and zero otherwise. This is done to allow for varying influence over time of the same event type. These variables are then used for optimizing in the DNERM. One drawback from the use of LASSO is that it deals poorly with collinear variables (Tibshirani, 1996). When faced with two highly collinear variables it tends to drop one of them arbitrarily, therefore a number of replications are performed to ensure the results are consistent. 3.3 Regression model and GARCH model The abnormal returns technique 2 is used to model the influences of the events on expected returns using the daily closing prices of the market indices. Equation 2 uses a market return and the return of an asset. While such a setting in not applicable to this research, the principle of this technique is maintained by taking a higher order market aggregation in which the indices of countries can be seen as assets. This higher order market aggregation is represented by the EURO STOXX50 index. For the implied volatility, the changes with respect to daily closing prices of the implied volatility indices are used. While the dependent variable is different, the structure of both the regression models for the expected return and the volatility are the same. This structure is shown in equation 10 , where is the number of selected explanatory variables by LASSO, the
coefficient of event type e and the DNERM adjusted event type variable at time .
10 For the GARCH model the events of the returns and volatility are combined into one model. The GARCH model has the structure of equation 3 , in which the events specified in the mean equation are the variables found by LASSO for the abnormal log returns of the market’s index, and the
variables used in the variance equation the variables selected by LASSO for the log differences of the market’s implied volatility index. The structure for the estimations is the same as described in 3 , where the event variables are represented by the dummy variables or the DNERM adjusted variables. The dependent variable for each of these versions of the GARCH models is the abnormal return series from the market’s index. 3.4 Procedure The empirical analysis of the event data on European countries starts off by determining which news events to select from all the possible events. The selection of events from the entire set, see section 5 for details, consists of about 150 different event types. Although all types could be possibly influential, only the events with the strongest effects are of interest. To make the choice in variables as arbitrary free as possible, the LASSO is used. For varying lag lengths, all possible event variables are fed to the LASSO. From all these results, the variables that are selected at least once by one of the LASSO’s runs, are chosen to be influential. This means that in some cases multiple different lags of the same event type are selected. Variables are selected separately for their influence on the returns and or volatility of the market. This means that also for both entities the LASSO is applied separately. To find explanatory variables for returns, the country’s market index is used. From this index the log returns are calculated from the daily closing prices and adjusted for the overall market return, by using the abnormal returns approach. The abnormal returns therefore are the dependent variable in the LASSO model. The market return is based on the daily closing prices of the EURO STOXX50 index. From daily closing prices the volatility is however not directly observable. To compensate for this, the daily log changes of the market’s implied volatility index are used; which are seen as an appropriate substitute (Tsay, 2005). The implied volatility index is used as an indicator of response rather than an indicator of performance towards the overall market, therefore just the unadjusted daily log changes are used as dependent variable. This separate approach enables to distinguish between variables that influence returns and volatility exclusively or jointly. The original event types do not contain information on the impact of the separate events, because they are of a dummy structure. These impact factors cannot be estimated in the DNERM. Therefore a text mining algorithm, proposed by Shook et al. (2012) is used to determine the impact of these separate events. This is regarded to as the ‘tone’ of a news event and consists of a score between ‐100 (extremely negative) and +100 (extremely positive). It is used as proxy for impact measurement as the score reflects the emotional density of the particular news event. While not as accurate as a human reader, automated tonal scoring can achieve useful approximations at scale (Leetaru and Schrodt, 2013). This measurement can be used as a method of filtering the context of
events as a subtle measure of the importance of an event and as a proxy for the impact of that event. Using this tool the structure of the event types can be refitted by applying distributions to the dummy structure, conditional on the tone value of each event. In total six different distributions are applied in the DNERM to the selected event types: the Uniform, Normal, Weibull, Gumbel, Fréchet and Exponentially Modified Gaussian (EMG) distribution. These distributions are assumed to resemble specific characteristics that could approximate the true response behavior of news events affecting the market. These DNERMs are estimated in a linear regression model and a GARCH model. For comparison the unadjusted dummy variables are also estimated in these models.
4. Simulations
This section contains various controlled settings to test how the DNERM performs under different circumstances. This is done by varying the sample size, the amount of variables used and the size of the added noise, all independently from each other. The results of single simulations from the different set‐ups are shown in Appendix A1.1 to A1.8. In this section the different DNERM specifications are referred to by their underlying distributions. 4.1 General Set‐Up A set‐up consists of 100 or 1000 observations, with three variables or twelve variables. These numbers are chosen arbitrarily to create different circumstances. For each set up a set of three event variables is evaluated. These events possess specific characteristics, namely an event that: frequently occurs (E1), less frequently occurs (E2) and rarely occurs (E3). These events are all in the tone adjusted dummy form. The event variables are then transformed under six distribution specifications of the DNERM, in which the parameters of the distributions are chosen in advance. For each of these specifications a return series is produced which contains a normally distributed small or large noise (ten times bigger than small noise) with mean zero and unit variance. These six created returns series are then all estimated with the six different specifications of the DNERM, where the parameters of the distributions for each of the generated return series remain constant. Intuitively, the distribution used to produce the return series is of course estimated most accurately with the variant of the DNERM with the same distribution. But as it is generally unknown what the distribution of the true variable is, the main interest of these simulations is how well the versions of the DNERM perform when applied to the other constructed return series. The comparison of the DNERM variants is done by accuracy of estimated coefficients, minimization of sum of squared errors and a theoretical fit to the true distribution. Specifically on the accuracy ranking of the coefficients, models gain points for significant accurate or semi‐accurate coefficients, see Appendix 1.1‐1.8, but are more heavily penalized for significant coefficients that are of the opposite sign of the true coefficient. In the following paragraphs the results of multiple simulations are discussed in general; for reference representative results of single simulations are shown in Appendix 1.1‐1.8. Aggregated results are shown in Table 1 for small sample simulations (100 observations) and in Table 2 for large sample simulations (1000 observations). For these two tables, for example, the rank of the Uniform DNERM variant is based on its performance in all six different set‐ups displayed in the appendices 1 and 2. Table 3 shows the average fit of each specification of the DNERM under each set up. The underlying results of Table 3 are shown in Appendix 2.1‐2.8.4.2 Small Sample: 100 observations Table 1 shows the results for the different set ups. The ranking of models is primarily reliant on the amount of noise present in the time series. For the model with three variables the results are jittery, as no model is consistently selected as most optimal when the noise is varied. For the model with twelve variables the results are more stable, with the Normal distribution being the most optimal followed by the Gumbel distribution. Table 1: Ranking of variants of DNERM for different set ups, using 100 observations
# Dist. Small noise Large noise
Betas SSE Overall Betas SSE Overall
3 Event Types Uniform 1 5 4 2 2 1 Normal 2 4 2 1 3 3 Weibull 6 7 7 2 4 4 Gumbel 3 3 3 4 4 5 Frechet 7 2 5 7 1 2 EMG 4 1 1 5 6 6 Dummy 5 6 6 6 7 7 12 Event Types Uniform 2 5 2 3 7 7 Normal 1 1 1 6 1 1 Weibull 7 6 6 7 2 6 Gumbel 5 2 4 2 4 2 Frechet 3 3 3 5 3 3 EMG 4 4 5 4 5 5 Dummy 6 7 7 1 6 4 The distributions are ranked for their performance opposed to each other on the accuracy of estimated coefficients and minimization of sum of squared errors (SSE), under different set ups. With 1 being the best and 7 being the least. See Appendix 1.1, 1.2, 1.5, 1.6. The Fréchet distribution estimation performs particularly well based on SSE under all four conditions, but based on accuracy of coefficients it performs poorly. Conversely the Uniform distribution performs better on coefficient accuracy for the three variable model and the Dummy variant for the twelve variable model, but both perform least on SSE minimization. Overall the Normal, Gumbel and Fréchet distribution rank the highest on average. When looking at the performance of the individual distributions, models that perform better on coefficient accuracy simultaneously perform worse is goodness of fit. This mainly due to the fact that models that perform better on accuracy have on average more estimated coefficients that are semi‐significant or insignificant (see Appendix 1.1, 1.2, 1.5, 1.6). This means that in the ranking for accuracy these models gain some points but don’t lose points. Conversely models that minimize the SSE better automatically have more significant coefficients. In this
case however, the coefficients are often wrong, resulting in strong penalization. In general for selecting the optimal model in small sample simulations, the optimal choice is dependent on what is what is more important: accuracy or fit.
Table 2: Ranking of variants of DNERM for different set ups, using 1000 observations
# Dist. Small noise Large noise
Betas SSE Overall Betas SSE Overall
3 Event Types Uniform 2 6 3 2 2 2 Normal 7 3 5 1 1 1 Weibull 6 5 6 6 6 7 Gumbel 5 4 4 7 4 6 Frechet 3 2 2 3 5 3 EMG 1 1 1 5 3 4 Dummy 4 7 7 4 7 5 12 Event Types Uniform 1 6 6 1 6 5 Normal 6 2 3 6 2 3 Weibull 3 5 5 2 5 4 Gumbel 4 2 2 3 2 2 Frechet 2 4 4 5 4 6 EMG 4 1 1 4 1 1 Dummy 7 7 7 7 7 7 The distributions are ranked for their performance opposed to each other on the accuracy of estimated coefficients and minimization of sum of squared errors (SSE), under different set ups. With 1 being the best and 7 being the least. See Appendix 1.3, 1.4, 1.7, 1.8. 4.3 Large Sample: 1000 observations Table 2 shows the results for the different set ups. Alternating the EMG, Gumbel and the Normal distribution rank highest. Opposite to the small sample simulations the same models score well on accuracy and SSE minimization at the same time. For this sample size it is clear that the dummy variant and the uniform distribution are outperformed in the twelve events models. Special attention needs to be paid to the Weibull and Fréchet distribution. These distributions do not allow events to have influence prior to the event date. As a consequence simulated events that exhibit influence prior to the event date, are estimated poorly by these two distributions. Both distributions also have the power of assigning the most value to the actual event date, more than other distributions can. This means that if there is reason to believe that an event type can only have influence after its event date or that the response to an event type is predominantly on the event date, then preferably one of these two distributions used for modelling. In this case however the parameters for the distributions were chosen random in each replication, so that the performance of these distributions in particular is possibly negatively affected.
4.4 Fit to the true model The individual results shown in the Appendices 2.1‐2.8 surprisingly show that not always the input distribution gets estimated the best with the same estimation distribution. Some distributions systematically estimate the given distribution more precisely. Table 3 shows these results aggregated by averaging over all fits to true distributions where the estimated DNERM has significant coefficients. Under all different set ups, on average the EMG performs best in capturing the time specific structure of all event types. Depending on the set up, the Gumbel, Normal and Uniform distribution perform better or equivalently well. The top three selected distributions all range around the 50%, which implicates that about half of the time specific structure of the event types is described by the estimated distribution. The static form of the Uniform distribution is in many set‐ups remarkably close to non‐ symmetric distributions, EMG and Gumbel. This is mainly due to the fact that the Uniform distribution less frequently finds significant coefficients (non‐significant values are discarded), but when it does apparently the fit is on average good. The opposite is true for the Frechet distribution. This distribution often finds (ill estimated) significant coefficients, while the distribution has very little resemblance to the true distribution. Especially for the Frechet distribution as input, none of the used versions of the DNERM gives a good approximation. Table 3 : Ranking of DNERM variants with respect to fit to the true distribution Observations 100 1000
Noise Small Large Small Large
# Distribution 3 Event Types Uniform 57% 44% 51% 41% Normal 66% 64% 37% 41% Weibull 35% 22% 29% 35% Gumbel 70% 50% 30% 42% Frechet 29% 17% 33% 19% EMG 72% 50% 41% 36% 12 Event Types Uniform 40% 37% 53% 54% Normal 58% 28% 52% 49% Weibull 39% 32% 32% 40% Gumbel 66% 16% 51% 48% Frechet 52% 29% 47% 34% EMG 53% 56% 54% 51% The distributions are ranked for their performance opposed to each other on the theoretical fit of the estimated distribution to the original distribution, with green being the best, dark yellow the 2nd and light yellow the 3rd best. For the ranking the average is taken from all results for each distribution within each set‐up, see Appendix 2.1‐2.8, missing results are excluded in derivation of the average.
4.5 Different Event Types The appendices 1.1‐1.8 and 2.1‐2.8 show the performance of the DNERM for the different event types. Type 1 (E1), frequently occurring event, is often found to be significant and is consistently estimated most accurate in nearly every specification of the DNERM. The time specific response structure is also estimated best for this type of event. The other two event types are alternately overestimated or ill estimated. This implies that event types with many events are estimated with more accuracy and the fit to the true distribution is approximated better. When the noise is increased both the accuracy and the time specific structure of all the event types are considerably lesser estimated.
5. Data
For the four countries, Germany, France, the United Kingdom and the Netherlands, two types of empirical data used in this thesis, namely market data and event data. In this section the choices made in the structuring, cleansing and coupling of data are discussed. 5.1 Market Data The market data is obtained from finance.yahoo.com, www.stoxx.com and www.google.com/finance. Table 4 and 5 show some descriptive statistics for this data, which consist of the market index and the implied volatility index for each of the countries. In addition also data is obtained for a more general index of Europe. After accounting for closed trading days, no observations are missing within each set. The sets all contain only the daily closing prices. Table 4: Description used market variablesCountry Ticker Min Max Standard
Deviation # Observations Origin
Netherlands AEX ‐0.044 0.042 0.006 2324 finance.yahoo
France CAC40 ‐0.046 0.041 0.007 2324 finance.yahoo
Germany DAX ‐0.047 0.032 0.006 2318 finance.yahoo
United Kingdom FTSE100 ‐0.041 0.040 0.005 2331 finance.yahoo
Europe EURO STOXX50 ‐0.044 0.039 0.006 2362 stoxx.com
Netherlands VAEX ‐0.462 0.454 0.037 1213 google.com/finance
France VCAC ‐0.143 0.158 0.031 1210 google.com/finance
Germany VDAX ‐0.123 0.117 0.025 2299 google.com/finance
United Kingdom VFTSE ‐0.128 0.158 0.031 1211 google.com/finance
‐ The Min, Max and Standard Deviation are based on the daily log differences of the indices, based on closing prices ‐ AEX, CAC40, DAX, FTSE100, EURO STOXX50 and VDAX use observations between 02‐01‐2006 and 30‐01‐2015 ‐ VAEX, VCAC and VFTSE use observations between 27‐10‐2010 and 30‐01‐2015 5.2 GDELT The event data is obtained from the “Global Data on Events, Location and Tone” (GDELT) project. Making use of Google Big Query, the data is acquired from the database. The primary sources that this project uses to identify news events, include all international news agencies like Africa News, Agence France Presse, Associated Press Online, Associated Press World stream, BBC Monitoring, Christian Science Monitor, Facts on File, Foreign Broadcast Information Service, United Press International, and the Washington Post. The processing of events is based upon a ‘Conflict and Mediation Event Observations’ (CAMEO) framework, which is a system of coding specially designed for news events. Events are therefore stored as a code rather than the whole original news story. This event code that is consists of a three layers a primary, secondary and tertiary number, indicating what type of
event it represents. Every event code always starts with one of the twenty primary categories: Make Public Statement (01), Appeal (02), Express Intent to Cooperate (03), Consult (04), Engage in Diplomatic Cooperation (05), Engage in Material Cooperation (06), Provide Aid (07), Yield (08), Investigate (09), Demand (10), Disapprove (11), Reject (12), Threaten (13), Protest (14), Exhibit Military Posture (15), Reduce Relations (16), Coerce (17), Assault (18), Fight (19) and Engage in Unconventional Mass Violence (20). The secondary and tertiary categories are added if a further specification can be given of the event. Also for every event the actors are identified, along with many other characteristics. All the fields used from the GDELT database are described in Appendix 3. Table 5: Correlation between log returns of market variables
Index* AEX CAC40 DAX FTSE100 Euro STOXX50
AEX 1 0.94 0.89 0.91 0.94 CAC40 0.94 1 0.93 0.91 0.95 DAX 0.89 0.93 1 0.87 0.91 FTSE100 0.91 0.91 0.87 1 0.95 Euro STOXX50 0.94 0.95 0.91 0.95 1 Implied
Volatility* VAEX VCAC VDAX VFTSE VSTOX
VAEX 1 0.64 0.62 0.64 0.64 VCAC 0.64 1 0.80 0.87 0.86 VDAX 0.62 0.80 1 0.80 0.87 VFTSE 0.64 0.87 0.80 1 0.81 VSTOX 0.64 0.86 0.87 0.81 1 *The maximum amount of observations available were used to derive these correlations. For the regular indices 2306 observations are used and for the implied volatility indices 1193 The event data is not directly applicable for the purpose of this paper, as the majority of the data is not related to financial markets. This paper therefore utilizes the characteristics of every event type so that specific actors can be chosen to constrain the relationship between which entity preforms an action upon another entity. In the database these entities are described by actor1 and actor2 respectively, but for clarity this relationship will be further described as actor and receiver. A choice is thus made of which combination of events and actor‐receiver pairs are used to distil data from, from the database. Based on literature discussed in section 2, a selection of event codes is chosen and divided into five main groups, namely economically (E), governmental (G), public/people (P), terrorism (T) and media (M), related topics. A set containing all events codes relating to Germany, France, United Kingdom and the Netherland is used to determine actor‐receiver pairs that are related the most to variables used in previous event studies. These groups consist of actor‐ receiver pairs conditional on origin, which means they are in principal subdivided into three groups, namely if the event came from outside the country (Incoming), from inside the country (Internal) or
to other countries (Outgoing). After placing restrictions on which pairs to include, also restrictions are placed on which event codes are used by which groups. This is done to make the interpretations of coefficients less ambiguous. The detailed description of which actor‐receiver pairs are used and which event codes they are assigned, is shown in Appendix 4. In this appendix an overview is shown of how the selection and filtration takes place. The detailed description of each event code can be found in Appendix 5. Based on this categorization and filtration for every country there is a certain ammount of events availiable which can be used in the models. Due to availibilty of news events, market data and scope of this paper a selection of the following countries was made: Germany(DEU), France(FRA), United Kingdom(GBR) and the Netherlands(NLD). The first three are strongest markets in the European Union and therefore are covered by media extensively. The Netherlands is added as a market to show the difference between small and large economies. As can been seen in Figure 1, the amount of suitable events increases significantly after 2006. This is mainly due to the substantial changes over the past two decades in both international news and the availibilty of news on the internet (Leetaru and Schrodt, 2013). For this reason the analysis in this paper will primarily rely on GDELT data after 2006. In Table 6 for each country a summary is given of the amount of events that are available in the timespan 02‐ 01‐2006 to 30‐01‐2015. 0 500 1000 1500 2000 2500 3000 2002 2004 2006 2008 2010 2012 2014 Number of Even ts Years Figure 1: Ammount of Suitable Events per Country per Year DEU FRA GBR NLD
A major disadvantage of the GDELT data is that there is a high probability of duplicate events being coded as different events. This could arrise when the events have different sources and/or are written slightly different, so that they are treated as separate events. Also often stories about news events evolve over time, so that the characteristics of the separate events could correspond to the same underlying event. To control for these problems a two features of the dataset are utilized. The first is that only root events are used. These are events occurring in the lead paragraph of a document, which tend to be more important (Leetaru and Schrodt, 2013). The second is that all events occuring on the same day are aggregated to one event, keeping the characteristics of the most important event. The most important event is chosen based on the highest value of average tone, highest number of mentions and highest number of sources. Table 6: Available amount of events per category, origin and country from jan‐2006 to jan‐2015
Country Origin E G M P T Total Events All Events
DEU Incoming 1842 664 804 275 138 3723 7132 Internal 474 534 311 186 10 1515 Outgoing 1300 567 24 0 3 1894 FRA Incoming 1958 1420 1112 399 318 5207 9569 Internal 551 1010 386 998 14 2959 Outgoing 692 632 70 0 9 1403 GBR Incoming 3575 1661 1845 712 408 8201 15709 Internal 1372 1636 316 1186 24 4534 Outgoing 2333 562 67 0 12 2974 NLD Incoming 875 375 235 45 41 1571 2608 Internal 188 143 32 35 0 398 Outgoing 452 175 12 0 0 639 Per country the news events are categorized in news groups (E, G, M, P, T) and origin. E = Economically related news, G = Governmental related news, M = Media related news, P = Public/Civilian related news, T = Terrorism associated news. The origin of the news is subdivided into: Incoming = all news wherein the country is only the receiver, Internal = all news where the country is the actor as well as the receiver, Outgoing = all news wherein the country is only the actor. 5.3 Weekends and Holidays Financial markets are closed on weekends and holidays, but news events can occur every day of the year. Therefore to merge the market data with the event data, all events that occur on a closed trading day are added to the first open trading day following the event. The assumption is made that events that occur on closed trading days cannot have influence on the market until it opens again.
