How Brexit affects financial markets

Размещено на http://www.allbest.ru/

Размещено на http://www.allbest.ru/

How Brexit Affects Financial Markets

Pilyugina Polina Sergeevna

Москва 2018

Table of Contents

• 1. Introduction
• 2. Literature review
• 2.1 Analysis of one particular media source
• 2.2 Media coverage analysis
• 2.3 Reversion effect
• 3. Model construction
• 4. Data gathering
• 4.1 Media coverage
• 4.2 Sentiment analysis
• 4.3 Stock market measures
• 5. Data analysis
• 5.1 Simple models for media coverage
• 5.2 Models with industry specification
• 5.3 Models with source type specification
• 5.4 Robustness
• 6. Limitations
• Conclusions
• 8. Bibliography
• Appendix
• Abstract
• This work examines the influence of Brexit on the financial markets through the media coverage. The novelty of this work is the application of the social media analysis to such unique event. The methodology used in this work is only emerging and no standardised approach exist, while this work combines the features of the most cited works on this topic. The main research question of this work is whether the media coverage of Brexit has significant effect on daily returns of FTSE100 and DJIA. The findings of this work allow to state the significance of this effect and even more, to more deeply understand the influence of particular industries.
• Keywords: Brexit, financial markets, media coverage, media effect.
• Introduction
• Globalization as an economic phenomenon has begun several centuries ago with the emergence of advanced technologies in transportation and communication. Mainly this term refers to the development of the international network of integrated economic systems, although this process results not only in economic, but also in political, social and cultural movements. International trade requires transport routes between countries and trading platforms, which would allow the movement of goods, services and financial instruments. From economical perspective, most political leaders have seen the process of globalization as beneficial for their economies, that is why they applied a lot of effort to contribute to this process. In the recent years several political and trade unions emerged in order to simplify the formal procedure of international trade, such as World Trade Organization, G20, European Union, BRICS etc.
• All these organizations were intended to facilitate the process of globalization and to provide comfortable and transparent environment for the movement of not only goods and services, but also people and ideas. From the first sight, this process seems to be beneficial for all participants, however there are several pitfalls, because countries, even as closely situated as in Europe, have large political, economical, cultural and religious differences and such integration may weaken their nation sovereignty; thus their objectives from participation in such unions may not coincide. Such goal incongruence resulted in growing concern among politicians about whether the benefits of participation in such unions overweight possible losses.
• From the economic perspective, some countries initially had less developed economies than others, and thus they benefit from EU, but at the cost of highly developed economies: as every country is interconnected to another, richer countries have no choice other than to support those participants who have a tough time Consider the Greek government-debt crisis as the most recent example.. That is why in recent years this concern escalated in historically unparalleled event of future withdrawal of United Kingdom from European Union, known as Brexit. The uniqueness of this event and the fact that UK is one of the biggest players on the political landscape of Europe makes Brexit exceptionally fascinating to examine and this work is dedicated to its effect on financial markets.
• From the beginning of UK participation in the EEC The European Economic Community is one of three European Communities, which were then incorporated in European Union in 1993. the question about whether it is reasonable for the UK was at play. Consequently, in 1993 a Eurosceptic political party called the UK Independence Party (UKIP) was formed and they insisted that EU ruins the national sovereignty of UK. Moreover, members of this party considered EU as being responsible for huge numbers of immigrants in UK and criticized this institution for corruption and undemocratic decisions. UKIP has begun growing popularity among UK citizens in 2004, until in 2014 the UKIP achieved the first place in UK European elections. This victory is considered to be the point of departure for the process of Brexit negotiation.
• The process of Brexit itself started with referendum held on 23 of June in 2016, which resulted in 51.9% of votes in favour of UK's withdrawal from EU. Notably, this is the first time ever the member state leaves EU. On 29^th of March 2017 the UK government triggered Article 50 and thus started the legal and political procedure of leaving the EU, which will end on 29^th of March 2019. From that point there were several severe changes in UK's government, lots of negotiations and bureaucratic procedures. Although the process is still in progress, everyone agrees that it would definitely strongly affect UK and EU members, and this effect is not that obvious by now. The long-run effects of Brexit on the global economic situation will become clear only some years after the official end of procedure, but this process has already affected financial markets. This effect is composed from current events and expectations about future, which are also affected by negotiations, happening right now, and how they appear to public. Moreover, even though there are several points of view on the underlying reasons of Brexit, the one common view is that it was greatly facilitated by social media and the Internet, spreading the ideas, opinions and gossips among people. That was the reason why this work focuses on social media factors when analysing the effects of Brexit.
• The effect of information on financial markets has been always of the great interest for financial analysists. It is not always obvious, how investors would react on the particular event (dividend payments, M&A etc.), and also how long would it take for their reaction to affect financial markets. As Eugene Fama (1970) suggested, the market is efficient when stock prices fully reflect all available information and thus no investor can consistently gain abnormal returns using trading strategies, based on available information. He outlined several forms of informational efficiency: weak form, when prices reflect all past information, semi-strong, when prices additionally reflect all public information, and strong, when prices also reflect private information.
• There is still no common answer about whether markets are efficient, because all studies are subject to joint hypothesis problem JHP refers to untestability of EMH, due to simultaneous testing the chosen model for stock returns and the market efficiency. Type I and type II errors may arise, thus making any conclusion uncertain., but there is some evidence in favour of all efficiency forms, as well as against of them, while semi-strong efficiency hypothesis has found the most evidence. In order to test semi-strong form of efficiency event studies are widely used, but they are particularly useful when the event takes some short time, and there is only one announcement. However, in the recent years increasing amount of studies examining news media effects emerged and their authors found the evidence, contradicting EMH, that media news and especially financial news had influenced stock markets, while there were no changes in fundamentals.
• First of all, nowadays people find out about events almost solely through social media and internet. Some governments even use control over media to promulgate pro-government ideas. That is why nowadays event itself sometimes does not matter, what really matters are how and when social media illustrates it, thus making people dependent on opinions of those, who create media. Moreover, this effect also persists in financial markets, when prices react only after event is reflected in social media. One particularly curious example is examined by Huberman Regev (2001): they examined the shock on financial markets made by an article in New York Times about potential breakthrough in cancer-curing drugs. The point is that the effect was huge: it affected not only the company, which was described in article, but some other biotechnological stocks, while it was not the first article devoted to that issue. In that case financial markets and society in general ignored the event itself, although it was described in several popular magazines, but overreacted five months later with permanent increase in company's stock price, when no new events really occurred. It is still not clear why only that particular article in NYT resulted in such a huge shock, but the phenomenon suggests that the effect of media should be more carefully examined in analysing the behaviour of stock prices after events.
• Consequently, analysing such unique event as Brexit and its effect on financial markets becomes very tricky procedure. First of all, this event is continuing right now and will be finished only in 2019, while yet having great effect on financial markets. This makes event studies not particularly useful, because the particular date of event (announcement) is unclear and there are still negotiations in place and with the end of each phase new facts are revealed, which can be seen as a distinct event. Secondly, in the world where social media and the internet play such a huge role, it is improper to ignore their effects. There is growing concern among economists and financial analysist about social media effects: it is now absolutely undoubtable, that this effect is huge, but its mechanisms still need examination.
• This study intends to analyse the effect of Brexit through the social media publications, instead of using simple timeline. The data on articles was obtained from factiva.org and was analysed through sentiment analysis package. In the beginning, literature on the issue is observed for model construction, after that data on publications and stock returns is obtained. The main part of this work is composed of econometric analysis of daily returns of FTSE and DJA with number of publications as one of exogenous variables. Finally, the results obtained through mathematical work are analysed in order to make some conclusions.
• 1. Literature review
• In the recent years growing amount of economic papers is dedicated to social media effect, mainly because with the development of the internet it becomes possible to obtain and aggregate all sorts of data about articles, since most of magazines are available online. Moreover, online articles have time advantage over printed versions and most investors rely on internet magazines, RSS feeds and other online resources, rather than printed newspapers, which technically are not able to deliver information in time. In addition, textual analysis and other analytical tools are possible to use only with electronic data, that is why most of the recent works are based on digital resources and so does this work. Nevertheless, for the purpose of the model construction studies based on printed newspapers would also be examined.
• Most of the papers tend to analyse the effect of news and media reports not related to some particular events, but rather the general correlation. There are several common trends in this works, although no general approach exists to quantification of social media effects because of the variety of data available. General approach is to include some “media factors” in regression analysis of stock returns, but the choice of particular factors is extremely different. Some of the studies include natural language analysis, some account for the type of articles (general news or experts' opinions), some include more complex factors like sentiment or stress strength of an article in combination with all previous. For the sake of model construction several papers were analysed on this topic.

1.1 Analysis of one particular media source

One of the first influential studies on social media effects was provided by Tetlock (2007). This study is focused on the column “Abreast of the Market” in Wall Street Journal and its influence on the stock markets. This column, as follows from the name, is focused on the market analysis and author of this column comments on the recent financial activities. This column was chosen because WSJ and its electronic distributor Dow Jones Newswires had the largest number of readers and for their long history both established good reputation among investors. The data was obtained for the period from 1984 to 1999.

In the study content of the “Abreast of the Market” was analysed using the program General Inquirer, which classified words in 77 categories and for each day number of the words in each category was obtained. After that the pessimism factor was formed by the factor analysis of principal components of these categories, which include examination of covariance matrix for the most important categories. As the result the evidence was found that changes in pessimism factor indeed has proven to significantly influence changes in daily stock returns. For the control variables he used five lags (defined as L5 operator) of detrended squared residuals of DJIA returns to proxy volatility, dummies to control for January and day-of-the-week effects, as well as dummy for the 19^th of October, 1987 shock; all these variables were included in . For the main data analysis vector autoregressive (VAR) model was used, in which the endogenous variables were returns (, volume () (as suggested by Campbell, Grossman, and Wang (1993)) and the media factor (. The final model was:

where was assumed to be heteroskedastic error term. The assumption about independence of error in the and of disturbances allowed author to use OLS regressions for estimations, with VAR estimates being similar to the Granger causality tests, as suggested by the author. The results of estimation showed the pessimism factor lags to be significant in the regression. After that the regression with as an exogenous variable was constructed in order to investigate, whether daily returns of DJIA and its trading volumes can influence the pessimism in the column. Coefficient of was significant, which implies that fluctuations in DJIA indeed affected the contents of the column, so the two variables are interlinked. The author overall estimated 3 OLS regressions which showed that pessimism factor affects both daily returns of DJIA and its trading volumes, as well as it is affected by changes in the returns. These results prove that media indeed has influence on the market and its content matters for the investors, although this effect seems to disappear in the long-run. In particular, increase in the pessimism causes temporary decline in stock returns. However, there are several drawbacks in this work. Firstly, the time period, chosen in the study, is seen as very outdated: the significance of this column for investors now and then is drastically different, because of the emergence of the internet. Secondly, the sentimental analysis focuses on pessimism factor and negative words, which may cause biased results.

Another study based on the column “Abreast of the Market” was conducted by Dougal et al. (2012) and it was also focused on the sentiment of the column, but their analysis was more deeply concerned with the personalities of the authors. Particularly, this work is intended to find out whether the content differences between journalists, who write for “Abreast of the Market”, have significant effect on investors. This column is particularly convenient for such study, because the information about authors is easily extractable and can be analysed. This study also included sentimental analysis of news, by automatically counting “negative” and “positive” words, as specified by financial dictionaries. Media factor consisted of 25 journalist indicators (one for each person), which reflected particular characteristics of their writing style, and it has proven to be significant factor in prediction of the DJIA returns. In particular, after the journalists were simply indicated for the regression, it's predicting power increased by 30-40%. Different specifications of the model resulted in conceptually the same conclusions.

Back to the model specification, it was based on the model from Tetlock (2007), in order to make these specifications comparable, but with several important improvements. First of all, the vector of DJIA returns was adjusted to dividend yields and stock splits, and it was excess return, so the Treasury bill rate was subtracted from each observation. Secondly, the vector of control variables accounted for the same sources of stock market predictability. It consisted of dummies for day-of-the-week and January effects, time factors (year fixed effects), detrended log of volume and detrended squared residuals of DJIA to account for volatility. Thirdly, standard errors were calculated using Newey and West (1987) and White to account for heteroscedasticity and autocorrelation in residuals. The most interesting part of this study was media factor construction. Unlike Tetlock (2007), this study was more focused on each particular journalist. The database included articles, published in “Abreast of the Market” for 9592 trading days, and almost every article was attributed to one particular journalist. The study focuses on the 25 authors, who wrote at least 50 articles for this column. Main difference of this analysis was that this study intended to compare the days when one particular author wrote an article to those when he\she did not, for this purpose the number of rotations was calculated. Then average excess returns were calculated for the day when the article was published, before and at the other days. The following regression was estimated:

where is excess daily return of DJIA, is a dummy variable, which equals one when journalist has written an article in period and as was specified previously includes all control variables. The results of estimation allowed the authors to conclude that individual writers can indeed somehow affect investor behaviour, and adding media factor significantly increases the explanatory power of the regression. However, the particular combination of influential journalists' characteristics is still unclear and requires more detailed analysis. Moreover, this work was focused on the macro-level influence, while individual investor behaviour and preferences were ignored.

Another work dedicated to one particular media was conducted by Uhl (2014) and it was focused on the Reuters news agency. It is based on the grounds of several previous researches, which put in doubt the EMH, suggested by Fama, because of significant overreaction of investors to new information.

The difference of this work from previous is not only in the choice of media source, but mostly in the sentimental analysis. The sample of articles was taken from January 2003 to December 2010 and it included more advanced sentimental analyses, made by Reuters algorithm, which assigned values -1, 1 and 0 to each article, depending on the content. Based on this analysis two independent time series of positive and negative articles were constructed. Another notable innovation of this work is that the authors used monthly returns on DJIA, not daily returns, in order to account for the long-run effects, and also, to account for macroeconomic factors, they used the time series of the Conference Board Leading Economic Indicators (LEI) index (the combined index consists of macroeconomic factors and is easier to operate in such model). To account for non-stationarity differenced logs were taken of DJIA, volumes and LEI. VAR model with three exogenous lags was used. Overall, the results suggest that both negative and positive sentiment in news significantly influence the returns in stocks and this influence is stronger, than the influence of macroeconomic index. However, this may occur because changes in LEI are incorporated in the stock prices faster, then in the month. Although the most important contribution of this study is that the media effect persists in monthly data.

1.2 Media coverage analysis

Another particularly interesting work, which examined the particular event was already mentioned in this paper: Huberman Regev (2001). As it was already mentioned, they examined the phenomenon of market overreaction on stale news. The results of this work were already mentioned, while in this part the choice of particular model will be examined. First thing is that this article is different from all others in the sense that it is dedicated to the particular event, not the general media effect, and moreover, the event itself allows to distinguish between the “media effect” and the “event effect”, as there was 5-month time span between them. The study was focused on the stock returns of seven biotechnological companies. As the result only May 4th (the day of publishing the influential article) had some effect on all biotech companies, while the event itself did not cause any disturbances. That implies that stock prices reacted not to the change of fundamentals, but rather to nonfinancial information.

The study of Engelberg and Parsons (2011) was in many ways different from what was examined before. If all other papers examined the media effect on market as whole, this study is focused on differences in investor behaviour with different media coverage. The sample of 19 non-intersecting trading markets, associated with 19 big cities in the Unitied States of America, was taken for this study. For each city the most influential newspaper was assigned, which allowed to examined the media effects, because of the difference in media coverage and in the contents of each source. The most important finding of this study is that the reaction of each market on announcements depends on whether this announcement was outlined by each local newspaper or not. However, there are different possible reasons why local publishers might ignore some announcements.

This decision might be related to the geographical location of the firm, which has made an announcement: local newspapers might be more likely to focus their interest on local firms, because local investors might be more interested in particular local firms more than in some others. If this effect is presented and is significant, this study might be biased because of the omitted variable bias. In order to account for such possibility, the model includes several controls for determinants of the local coverage and local trading. The inclusion of this variables, however, did not change the significance of local media coverage. Another set of controls included fixed effects for pairs of firm and city, firm and announcement date, city and earnings date. Also, the fixed effects reflecting differences in particular cities were included in the model. The notable result is that even after the inclusion of all three sets of controls the local media coverage remained the significant factor in the model. Overall, this study also allows to insist that local media coverage affects the investors behaviour. However, this study strongly relied on printed newspapers (the data was found for the time period between 1991 and 1996 years), and does not take into account, that with the development of the internet the borders of each region of coverage are most likely to become wider.

One of the most recent studies on the media influence was conduced by Joanna Strycharz, Strauss, Trilling (2018). They focused their analysis on three companies listed on the Amsterdam exchange index: Phillips, Shell and ING (which were chosen because they are among the largest and belong to multiple industrial sectors). For companies it is essential that media not only communicates information about them to the public, but also how they appear in media: that is why big companies even have special departments for media communication. This study looks on the general media coverage of companies to explain particularly influential media factors: not only how media coverage affects prices, but also how sentiment affects and whether it is significant, as well as how topics of articles influence the result. Reversion effect problem is also addressed in this study, that is why the VAR model is used.

Stock market data was obtained in the form of absolute difference in daily closing prices. Taking the absolute value seems strange, but it allows to account for the fact that what seems positive for one investor, might appear negative to another. The time period was from 1 January 2014 to 31 December 2015. Articles was parsed using Python scripts from RSS feeds, which allowed to download articles and retrieve the text of an article. Only articles which were dedicated to one of three companies were retrieved and then automatically and manually checked for relevance. Also SentiStrength was used to analyse sentiment of each article. Another algorithm was used to assign topics for each article and for each company different number of topics was assigned: 12 for Phillips, 14 for Shell and 11 for ING. Overall there were three time series of media factors: one with overall number of article, another two with sentiment and topic specification.

All data was assigned to trading days with weekend articles treated as if they were published on the first consequent trading day. Then econometric analysis was applied to account for non-stationarity and to find the optimal number of lags for VAR model. Although the residuals of sock prices time series showed signs of heteroscedasticity, overall tests indicated the robustness of the model. Tests for robustness included test of normality for all variables, test for outliers, tests for multicollinearity, heteroscedasticity and omitted variables. Overall 50 VAR models were constructed. The number of articles per day appeared to be significant, while explaining only the small proportion of fluctuations. Sentiment has proven to have significant positive effect on the share prices of Shell and Phillips, as well as specification of topics for an article. Overall the model has proven the significance of media in explaining the stock market behaviour, while the reversion effect was not found significant. Nevertheless, the model is limited to the particular companies, while it does not give any insights about why Shell and Phillips were affected by media to the far more extent than ING.

1.3 Reversion effect

One of the most serious possible drawbacks of all media studies is the reversion effect. The question is whether the media influences the market, or it is shaped by market events. All papers try to address this problem using VAR models and controls, but the study by Scheufele, Haas, and Brosius (2011) is solely dedicated to this problem. The authors also pointed the problem of choice between micro-, meso- and macro-level of the analysis: it is impossible to construct one empirical analysis on all three levels, because changes in the stock market are caused by all investors and it is impossible to distinguish between them. Moreover, investors behaviour is affected by characteristics, such as degree of risk aversion, which are not observed. Thus the micro-level analysis does not provide information about the market as whole, while macro-level analysis relies on aggregated effects. The main question of this work is whether media mirror or shape stock market changes.

Their analysis is focused on eight German companies and their media coverage. For the market factors they have chosen to inspect the relative change in closing prices (from which they subtracted the relative change in Deutscher Aktien Index) and trading volume in Germany for the period from July to August 2005. The formula for adjusted closing price change was as follows:

As for the media factor, the data was found about printed, television and online news, reflecting some information about selected companies, for each of 44 trading days in sample. The contents of each paper were examined in order to find whether each of them contains some advice on whether investors should buy, sell or hold a security. Articles, that were published on weekends were attributed to the next closest trading day. Finally, three factors were calculated: number of mentions of company name or its products, the valence of media (difference between positive and negative) and the valence of stock values.

The data analysis consisted of several steps and included both descriptive analysis and econometric analysis of time series. On the first step for each time series ARIMA models were estimated and residuals were used in analysis as media factors. The residual in general should not be auto-correlated, and thus should follow the White Noise. Secondly, correlation matrix was calculated for each adjusted time series of media factors and adjusted closing prices: both with lags ±5 days and with no lag. Positive lags represent stock market influence on media, while negative represent the effect of media on the market. They have used prewhitening modelling technique, instead of VAR model, as it allowed for strong test because of the assumption about widespread relation between media and stock market and the assumption about macro-level difficulty in accounting for the micro-level characteristics. Correlation matrix showed much stronger correlation of negative lags than of positive lags (by construction, correlations at positive lags imply media to influence the market, while at negative lags correlation means reverse effect), which can be considered as an evidence in favour of “mirroring” media. Although the results of the study generally support the hypothesis that media mirrors the market, rather than shapes it, the study showed some correlation existed for positive lags also, while not that strong. Also, this study relies on several rather strong assumptions about macro-level effects and the model used is unusual for this type of studies. Moreover, the sample period is very small to analyse long-term strategies and the result may be driven by this particular small sample only, and also in 2005 the spread of the internet was not so wide.

To sum up, the models for such studies vary widely, because no standardised way had emerged yet. Moreover, most studies are concerned on some particular companies or some particular media source, but not about some special events, that is why this study is one of the first to attempt modelling the effects of Brexit through its media coverage and to compare the effects of the same media coverage on two different markets. This attempt may be formulated by several research questions, which this study seeks to answer:

RQ1. Whether the inclusion of the Brexit media factor increases the explanatory power of the regression and appears to be significant?

RQ2. Whether this effect varies for articles from different industries\sources?

RQ3. Whether the media effect of Brexit is different for FTSE and DJIA?

The hypothesis is that Brexit is not purely negative event, and its consequences differ in different industries. Also we expect the effect of Brexit on FTSE100 to be stronger, that its effect on DJIA.

2. Model construction

The choice of the model somehow relies on the previous works on this topic, while reflecting the crucial differences in underlying assumption. First of all, most studies used VAR models and combination of negative and positive lags in order to account for the reversion effects, while this study is focused on media coverage of one particular event. Although Brexit affects economy and financial markets, the procedure of Brexit is not affected by financial market fluctuations. Moreover, the articles about economic consequences of Brexit are mostly focused on possible macroeconomic changes, which will occur after the UK finally withdraws from EU, or the legal procedures of Brexit. That all combined allows us to make one particularly useful assumption of Brexit articles being exogenous variable and not consistently reversely affected by stock market fluctuations. That allows to use OLS model with some extensions.

Stationarity of the time series was of the great concern, so the time series of FTSE and DJIA closing prices was examined for stationarity and the final choice was to include daily returns of each index, calculated as follows:

where is closing price for the day . Both time series of closing prices were found non-stationary by Augmented Dickey Fuller test, while returns, calculated as above, were proven to be stationary by ADF test, as can be seen in Figures 5-12 in Appendix section. The H₀ of ADF test is that the process has unit root (is not stationary), and it is not rejected for closing price time series (see Figures 15-22 in Appendix) for any significance level more than 10%, while it is rejected for both series of returns, as p-values of the tests are equal to zero.

The media coverage factor was chosen to be the daily number of articles, published about Brexit, sorted by industries and by sources. Another media factor was sentiment, calculated for each day. More detailed explanation of this variables will be presented in the next section. To control for the know sources of stock market predictability the number of variables was included: control dummies for day-of-the-week, January effect, but also lagged values of squared residuals and log of daily trading volumes. The generalized model is as follows:

(1)

where is set of variables, equal to the number of papers (in different specifications it would be number of articles in each industry/source), is the average sentiment, is set of control variables. To account for heteroscedasticity White standard errors were used, as suggested by Dougal, Engelberg, Garcнa, Parsons (2012).

brexit financial market media

3. Data gathering

Firstly, the dummy for the timeline of Brexit was constructed. It was based on the official timeline, published on the web site of UK parliament (written by Walker) and it consisted of dates, on which steps of Brexit were made. The data was retrieved in form of dummy variable, which takes value 1 if there was some event listed in that document, and 0 if there were no events.

3.1 Media coverage

In order to obtain data on articles the web site factiva.org was used. Overall this service provides access to more than 30 thousand different sources, including the most popular and influential newspapers (like Financial Times or The Wall Street Journal etc.), journals (The Economist, Forbes etc.), news feeds and business web sites. The aim of this work was to analyse the wide variety of articles and Factiva was the most suitable source. This service itself provides several search instruments for data categorization which make it easier to create a database.

First of all, Factiva allows for complex search requests using regular expressions, count of words and other specifications. Although there is extremely wide variety of them, for the purpose of this work the following specifications were used: time period, industry and source type. In order to exclude irrelevant articles, the search request required at least 3 mentions of the word “Brexit” in the text. Whereas the use of regular expressions was possible, there were no need in them for several reasons. First of all, Brexit is a unique term and it is used only to identify one particular event, so it cannot be confused with anything else. Secondly, Factiva automatically selects articles with punctuation signs after the keywords. Moreover, service allows to automatically exclude duplicates, necrologies and another irrelevant posts from the search results.

However, the main feature of this web site is its ability to categorize articles in different dimensions. Each article is attributed to the particular industry based on textual analysis and these categories are not intersecting, so there is no need for manual assignment of questionable articles. The total amount of industries is over 100, but they are clustered in 14 different general industries: Agriculture, Energy, Retail, Media, Health Care, Automotive, Resources, Business, Consumer Goods, Construction, Financial Services, Industrial Goods, Technology, Logistics. These industries indicate to which industry the particular article applies, so they can be considered as topics. Second categorization which was used in this work is based on the source type. There are six general source categories: “Dow Jones Newswires”, “Major News and Business Sources”, “Press Release Wires”, “Reuters Newswires”, “The Wall Street Journal - All sources” and “Others”. All these categories are assigned to article by Factiva algorithms and they do not intersect (category “Others” was constructed as not any of previously mentioned). However, it is not possible to get this information from particular search result in Factiva, so the categorization can be obtained only using separate search requests.

Although Factiva provides very deep and high-quality data base, it is impossible to retrieve big amounts of data from this service, which notably complicated the work and required some creative method of data gathering. Therefore, the process of data retrieving consisted of several steps. First, search requests were constructed as follows: at least 3 mentions of the word “Brexit”, industry type and source type. All search requests included time period parameter which was set to 2 years from June 2018, as in June 2016 referendum was held and the Brexit started and the term appeared. Additionally, the language was specified to be English, as this work intends to analyse the effect on FTSE and DJIA, not to say that number of articles in other languages is significantly lower. That is the final search request used in Python script was

la=en and atleast3 brexit and(in={0} and rts={1})

where instead of {0} each type of industry was substituted, and the type of source instead of {1}. Data period and duplicate exclusion were set manually. On this step, Python script constructs all 84 search requests and then they were manually entered and Factiva search result pages were manually saved in the form of html tables. Then they were parsed in the form of csv file with one article being a unit of database, and each observation contained the industry and source specification, title, date and abstract. The diagrams, illustrating the distribution of articles by industries and sources can be found below.

Figure 1

Figure 2

The graph of the total number of articles dynamics shows that the popularity of this topic has declined during the 2018 and it was the largest during the June 2016, when the referendum was held.

Figure 3

3.2 Sentiment analysis

Before proceeding to data base formation, Python script was written, which parsed raw html data into excel table, but with one additional step: TextBlob semantic analysis was used to assign polarity parameter to each article. TextBlob is a library for Python based on Natural Language Toolkit, which uses complex textual analysis and natural language processing technics. This library was chosen because it is based on NLTK, which is widely used in sentiment analysis. The polarity parameter itself is in range of -1 to 1 for each article, and it was calculated for each article separately. After that when the final database was formed using excel and python library Pandas, for each date the number of articles for each pair of industry and source was calculated, and polarity parameters of this articles were summed up and divided by the number of articles for each day.

The resulting database consisted of two tables: one with number of articles per day for each of 14 industries, another with number of articles per day for each type of source. Both tables included the average polarity of all articles per day.

3.3 Stock market measures

The next step was to merge all the data on articles with financial data. As it was mentioned before, analysis was based on two market indexes: Financial Times Stock Exchange Index (FTSE) 100 and Dow Jones Industrial Averages (DJIA). As suggested by Scheufele, Haas, and Brosius (2011), articles published on weekends were treated as if they were published on Mondays, and the same procedure was made with sentiment parameter. Because of the difference in distribution of holidays in UK and US, two slightly different data sets were constructed for FTSE100 and DJIA models. The data on closing prices and volumes was retrieved from Yahoo! Finance.

The last step was to include the set of control variables. The choice of control variables was based on existing works on the financial modelling. The set of controls included dummies for day-of-the-week effects (dummy for Mondays and dummy for Fridays), dummy for the month of January, squared residuals of the returns and differenced logarithms of volumes. The residuals were obtained as suggested in Tetlock (2007): 60-day moving average was subtracted from the values and taken as a residual.



Figure 4a. Summary of the data for FTSE100 (Estimated by Stata)

Figure 4b. Summary of the data for DJIA (Estimated by Stata)

4. Data analysis

In order to identify the effect of news about Brexit on financial markets the general regression (1) was estimated in Stata. Three specifications were estimated overall for each market index: one with coverage measured as simply total number of articles (“sum” variable), one with number of articles distributed in different industries, and one with number of articles distributed in different sources. Following the suggested in Tetlock (2007) framework, the OLS models with White standard errors were estimated. However, the data specificity requires slightly different specifications.

For the first step correlation matrix was calculated for the explanatory variables. The correlation matrix provides high positive correlation, which is consistent with the specificity of data: most likely the increase in number of articles in different industries is caused by same exogenous shock. Although there were almost no extreme correlation values, this is important to check multicollinearity bias in the model.

Figure 5. Correlation matrix of articles by industries (Estimated by Stata)

The correlation matrix for articles by sources shows very high correlations between almost all variables, which seems reasonable. Surprisingly, number of articles in Press Release Wires has the least correlation with other variables. This correlation matrix show that the results of regression analysis should be taken with caution, as the



Figure 6. Correlation matrix of articles by sources (Estimated by Stata)

4.1 Simple models for media coverage

To proceed with the analysis, different regression specifications were estimated in Stata. Firstly, we estimate the model using FTSE100 index, as it is of the main concern for our model. First specification that was estimated was:

(2)

where is daily returns, calculated by the formula in Section 3, are lagged logarithms of trading volume, are lagged squared residuals and is the vector of control variables. This regression (and all that follow, until specified) was estimated by Stata using White standard errors. It was then compared to the model, including the media factor:

+

(3)

where stands for the lagged number of variables. The inclusion of the variable increases estimation power of the regression by 25% (R² increases from 0.0542 to 0.0674). The coefficient of variable was found to be significant at 10% significance level and it suggests that 1 unit increase in number of articles on Brexit, published on previous date, increases the daily return of FTSE100 by 0.00121%. Moreover, comparison of the Akaike information criteria allows to state, that the model (3) slightly outperforms the model (3). Although the explanatory power of the regression is small, it is common situation for the works, trying to predict the market indexes. The third model estimated was:

+

(4)

where is the lagged value of average sentiment. Although the coefficient of is not significant at 10% significance level, it increases the predicting power of the model even more (R² = 0.0691) and also AIC is less, then for (3). The summary of statistics for three models can be seen in the next figure. These regressions allow to answer the RQ1, since inclusion of Brexit media coverage indeed increases the explanatory power of the regression and its effect is statistically significant. Although the coefficient is rather small, it is positive and its sign is statistically significant also.

Figure 7. Summary of three model specifications estimation for FTSE

Variable	(2)	(3)	(4)
L1.return	.10232798	.08172446	.07958521
L1.res	10.990008	9.1534424	9.2804971
L1.logvlm	-.00213765	-.00307435	-.00319623
DoW	.00001653	-.00010531	-.0001521
Jan	.0001693	.00023108	.00022601
L1.sum		.00001214	.00001229
L1.avsent			.00896186
_cons	.04369995	.06206991	.06432557
AIC	-2094.042	-2096.185	-2094.714
R-squared	0.0542	0.0674	0.0691

This table contains coefficients of specified variables, with bolded values corresponding to variables, significant at 10% s.l. Dow and Jan variables are dummy controls for day-of-the-week and January.

The same regressions were that ran on DJIA data. Overall the inclusion of increases the power of the regression, and in case of DJIA this effect is stronger, than it was for FTSE100 (R² increases from 0.0105 to 0.0215), while its value in this case is much lower. However, the coefficient of is significant at 5% s.l in this model, while its value is slightly less, than in FTSE100 specification. The coefficient of is still insignificant, while it increases the explanatory power of the regression. AIC suggests that the best model among estimated is the one corresponding to the regression (2)

Figure 8. Summary of three model specifications estimation for DJIA

Variable	(2)	(3)	(4)
L1.return	.02293399	-.00092043	.00014956
L1.res	1.80245	.73069648	.70310237
L1.logvlm	-.00018997	.0005281	.00052982
DoW	-.000548	-.00065193	-.00062644
Jan	.00255683	.00238991	.00239137
L1.sum		.00001176	.00001164
L1.avsent			-.00545244
_cons	.00418132	-.01051434	-.01038544
AIC	-2040.745	-2041.992	-2040.184
R-squared	0.0105	0.0215	0.0222

This table contains coefficients of specified variables, with bolded values corresponding to variables, significant at 10% s.l. Dow and Jan variables are dummy controls for day-of-the-week and January.

This estimation allows to make some conclusions about the RQ3. Even though the explanatory power of the regression is less for DJIA, the effect of media coverage about Brexit is significant for both FTSE100 and DJIA, being surprisingly even more significant for DJIA. Moreover, inclusion of Brexit effect increased the explanatory power for both specifications. From these facts we can draw the conclusion, that media coverage of Brexit has influenced both FTSE100 and DJIA daily returns. In addition, the timeline dummy was found to be insignificant and its inclusion did not improve the model, so it was not included.

4.2 Models with industry specification

In this section augmented models will be considered, in which the industries will be specified. We expect the “Financial Services” industry to be the most significant among the industry-specific variables. However, in investigation of this model we should bear in mind the correlation matrix and possible multicollinearity problems. Firstly, the augmented regression is estimated, which includes all 14 industries:

(5)

where is the set of 14 variables, which correspond to the number of articles in each of 14 industries. The inclusion of this 14 variables increases explanatory power of the regression significantly, compared to set of previous models. Although only four industries were statistically significant (Agriculture, Media, Financial Services and Industrial Goods), it is still interesting to investigate the signs of effects for other industries. There were overall seven industries with negative coefficients estimated. Among significant variables only Media industry has negative effect on daily returns, which is surprising. All other significant variables have small positive effect. The second model estimated included only those variables, which were found significantly at 10% s.l. in previous model:

(6)

where the name of each new explanatory variable corresponds to the number of articles in particular industry. Variables are still significant in this regression and the change in coefficients is small. Although the R² has decreased in the new model, the AIC for the second model is less, than for the first, which suggests that the second model outperforms the first one. Moreover, in decreased even compared to the (2) regression, which means that specification of industries is significant in predicting the daily returns of FTSE100. This suggests, that the effect of Brexit on industries is different, and the news about Brexit progress can negatively affect the market.

Figure 9. Summary of four model specifications estimation

	FTSE100	DJIA
Variable	(5)	(6)	(5)	(6')
L1.return	.10252964	.0999496	.00073449	.01293075
	0.2216	0.2157	0.9924	0.8687
L1.res	8.3798972	10.126655	3.7906534	4.5662015
	0.2065	0.1108	0.4261	0.3240
L1.logvlm	-.00151737	-.00182304	-.00013571	-.00016587
	0.4474	0.3697	0.8849	0.8528
DoW	.00047351	.00040704	-.00071799	-.00066652
	0.5931	0.6114	0.4796	0.4545
Jan	.00019308	.00029599	.00258539	.00249638
	0.8446	0.7659	0.0310	0.0329
Agriculture	.00032549	.00029745	.00021379
	0.0242	0.0283	0.1615
Energy	.00034754		.0001966
	0.1934		0.4247 ...

Подобные документы

• The Global Money Markets and Money Management

  The General Economic Conditions for the Use of Money. Money and Money Substitutes. The Global Money Markets. US Money Market. Money Management. Cash Management for Finance Managers. The activity of financial institutions in the money market involves.
  
  реферат [20,9 K], добавлен 01.12.2006
  

• To investigate a credit channel of monetary policy transmission

  Study credit channel using clustering and test the difference in mean portfolio returns. The calculated debt-to-capital, interest coverage, current ratio, payables turnover ratio. Analysis of stock market behavior. Comparison of portfolios’ performances.
  
  курсовая работа [1,5 M], добавлен 23.10.2016
  

• Financial institutions

  The concept, types and regulation of financial institutions. Their main functions: providing insurance and loans, asset swaps market participants. Activities and basic operations of credit unions, brokerage firms, investment funds and mutual funds.
  
  реферат [14,0 K], добавлен 01.12.2010
  

• Specificity exchange market

  Types and functions exchange. Conjuncture of exchange market in theory. The concept of the exchange. Types of Exchanges and Exchange operations. The concept of market conditions, goals, and methods of analysis. Stages of market research product markets.
  
  курсовая работа [43,3 K], добавлен 08.02.2014
  

• Off-budget funds - their place in a financial system

  Economic essence of off-budget funds, the reasons of their occurrence. Pension and insurance funds. National fund of the Republic of Kazakhstan. The analysis of directions and results of activity of off-budget funds. Off-budget funds of local controls.
  
  курсовая работа [29,4 K], добавлен 21.10.2013
  

• A balance sheet

  Тhe balance sheet company's financial condition is divided into 2 kinds: personal and corporate. Each of these species has some characteristics and detail information about the assets, liabilities and provided shareholders' equity of the company.
  
  реферат [409,2 K], добавлен 25.12.2008
  

• Capital market

  Capital Structure Definition. Trade-off theory explanation to determine the capital structure. Common factors having most impact on firm’s capital structure in retail sector. Analysis the influence they have on the listed firm’s debt-equity ratio.
  
  курсовая работа [144,4 K], добавлен 16.07.2016
  

• Accumulation Pension Funds: condition and development

  Theoretical aspects of accumulation pension system. Analysis of current status and development of accumulative pension system in Kazakhstan. Ways to improve the pension system and enhancing its social significance accumulative pension fund provision.
  
  курсовая работа [1,1 M], добавлен 06.11.2013
  

• Вероятность получить государственную поддержку и конкурентоспособность – диагностирование самоотбора

  Дивестиции как процесс продажи части подконтрольных компании активов и изъятие капиталовложений, ее разновидности в зависимости от условий. Методология формирования выборки и Event Study. Анализ и результаты Event Study для российских металлургов.
  
  курсовая работа [294,3 K], добавлен 08.02.2017
  

• Theory of Finance

  Fisher Separation Theorem. Consumption Vs. Investment. Utility Analysis. Indifference Curves. The satisfaction levels. Indifference Curves and Trade Off between Present and Future Consumptions. Marginal Rate of Substitution. Capital Market Line.
  
  презентация [1,5 M], добавлен 22.06.2015
  

• Лондонская фондовая биржа

  Понятие фондовой биржи, ее задачи и функции. Главные обязанности брокеров, дилеров, маклеров, спекулянтов. Фондовые индексы и их назначение. Лондонская фондовая биржа: краткая история развития, сегодняшний день, перспективы. Динамика индекса FTSE 100.
  
  реферат [115,2 K], добавлен 30.11.2016
  

• Анализ влияния изменения рейтингов на ценообразование ценных бумаг

  Исследование влияния деятельности рейтинговых агентств на доходность еврооблигаций российских эмитентов, обращающихся на зарубежных торговых площадках. Анализ показателей доходности ценных бумаг в определенные временные периоды методом Event Study.
  
  дипломная работа [244,5 K], добавлен 31.08.2016
  

• The Tax System of the United States

  Federalism and the Tax System. Federal Taxes and Intergovernmental Revenues. Tax Reform. The Progressivity of the Tax System. Political Influences on the Tax System. Main principles of US tax system. The importance of Social Security taxes.
  
  реферат [15,9 K], добавлен 01.12.2006
  

• Влияние объявлений о дивидендных выплатах на цены акций российских компаний

  Обоснования направления влияния дивидендных выплат. Политика выплаты российских компаний. Метод event study. Расчет нормальной доходности акции, влияние объявлений о дивидендных выплатах на цены. Усредненная избыточная доходность по типам новостей.
  
  курсовая работа [454,5 K], добавлен 13.10.2016
  

• Financial markets and trading strategies

  The behavior of traders on financial markets. Rules used by traders to determine their trading policies. A computer model of the stock exchange. The basic idea and key definitions. A program realization of that model. Current and expected results.
  
  реферат [36,7 K], добавлен 14.02.2016
  

• Making the media in Internet

  Theoretical basics of Internet advertising. The analysis of the media planning process. The establishing media objectives through developing media strategies and tactics. The effectiveness of the media planning in Internet. The example of the media plan.
  
  курсовая работа [64,2 K], добавлен 25.03.2014
  

• The mobile application HeadyUp

  Social network theory and network effect. Six degrees of separation. Three degrees of influence. Habit-forming mobile products. Geo-targeting trend technology. Concept of the financial bubble. Quantitative research method, qualitative research.
  
  дипломная работа [3,0 M], добавлен 30.12.2015
  

• Media in China

  Chinese media and government. Xinhua (the China News Agency) and People's Daily, the two most important print media. Internet censorship in China. Central Television, talk Radio, cable TV and satellites. The role of "internal" media. Market competition.
  
  курсовая работа [404,3 K], добавлен 09.12.2010
  

• Trading algorithms in financial markets

  Description of exchange stocks as financial point-of-sale platforms. Description of point-of-sale algorithm of broker trade at the financial market. Parameters of price gaps on financial auctions and optimization of currency point-of-sale algorithms.
  
  контрольная работа [1011,9 K], добавлен 14.02.2016
  

• Study aspects of the impact of modern media on the British

  Consideration of the mass media as an instrument of influence on human consciousness. The study of the positive and negative aspects of the radio, television, press, magazines, Internet. Advantages and disadvantages of the media in the Great Britain.
  
  дипломная работа [2,3 M], добавлен 14.10.2014
  

• главная
• рубрики
• по алфавиту
• вернуться в начало страницы
• вернуться к началу текста
• вернуться к подобным работам