Retail

.00016708

.00020547

.0002426

0.1542

0.0822

0.0592

Media

-.00016393

-.00025339

1.254e-06

0.0593

0.0727

0.9892

Health Care

3.664e-06

-.00008926

0.9850

0.6057

Automotive

-.00008031

-.00014804

0.5335

0.2752

Resources

-.00031174

-.00008808

0.2524

0.7469

Consumer Goods

-.00006246

-.00007455

-.00005598

0.2058

0.0472

0.1744

Business

-.00004031

.00001261

0.5801

0.9121

Construction

.00001693

.00001953

0.8658

0.8660

Financial Services

.00001263

.00001332

0.000009116

0.000009237

0.0490

0.0432

0.1601

0.1810

Industrial Goods

.00051044

.0004036

.00038437

.00038166

0.0162

0.0408

0.0640

0.0328

Technology

-.00011986

-.00052667

-.00051065

0.4730

0.0297

0.0593

Logistics

-.00003334

.00007967

0.7865

0.4346

_cons

.03030614

.03687634

.00255294

.00358426

0.4579

0.3715

0.8877

0.8380

R-squared

0.1480

0.1159

0.0894

0.0703

AIC

-2096.856

-2105.943

-2036.838

-2048.835

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 regression (5) was estimated for the DJIA daily returns. As compared to regressions in previous sections, the explanatory power of (5) has increased, while information criteria did not improve. Overall, five industry specific variables were found significant: Retail, Consumer Goods, Industrial Goods and Technology, which are different from those, chosen for FTSE100. In order to improve the model insignificant at 10% variables were excluded, except for the Financial Services, as it seems too strange that the variable, which constitutes to 40% of overall number of articles and is of the most importance for financial markets is less significant, than Retail or Technology. This can be the sign of multicollinearity problem, however these categories are not the most correlated ones. Nevertheless, another specification was estimated:

(6')

As can be revealed from the Figure 9, although the explanatory power of the regression has decreased, AIC has improved significantly, compared to all previously estimated models. This can suggest, that the effects of Brexit information for FTSE100 and DJIA are different, while the reason of that cannot be revealed from the model: it can either follow from the industrial coverage of indexes, or from the geographical differences, or there might be some another uncovered reason. The main conclusion we can make is that the effects are somehow different for two markets.

Additionally, the inclusion of dummy for the Brexit timeline or average sentiment did not improve the results of the model and both variables were found insignificant during tests, so they were not included in the final versions of the models. And also it is interesting to note, that the dummy variable for January effect is highly significant for DJIA and not significant for FTSE100

4.3 Models with source type specification

The next model of interest is the model, which includes the source type specification. However, the construction of data has some drawbacks: there are two types of sources which constitute to the largest part of observations, and the variables are highly correlated. We estimate the following regression:

(7)

where sources variables are included in the RHS of the equation. Correspondence of variables' names to the source type can be found in the Appendix. The results of estimation suggest that the only significant source type is “Others”, which constitutes to the largest category, so it is most likely that the results are biased because of that. Same holds for the DJIA estimation.

Figure 10

Variable	FTSE100	DJIA
L1.return	.10302489	.05075264
	0.2124	0.5153
L1.res	10.350208	3.038558
	0.1175	0.5395
L1.logvlm	-.00200855	-.00028444
	0.3371	0.7794
DoW	.00002551	-.00039864
	0.9754	0.6257
Jan	-.00010084	.00244629
	0.9187	0.0376
others	.000017	.00001542
	0.0342	0.0183
WSJ	.00037067	.00017748
	0.1863	0.6287
DJW	-.00024095	-.00021981
	0.1720	0.1197
MNB	-.00001779	-7.811e-06
	0.7536	0.9027
PRW	-.00017691	.00009783
	0.3372	0.6376
RW	.00002716	-.00001301
	0.8915	0.9556
_cons	.0416971	.0058847
	0.3272	0.7678
R-squared	0.0924	0.0375
AIC	-2094.185	-2036.763

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.

Overall, the specifications estimated allow to answer all research questions. First of all, the effect of Brexit media coverage indeed affects the financial markets, since the inclusion of media coverage both increases the explanatory power of the regression and is significant. It is no surprise that this effect is small, because market indexes contain stocks of rather different companies. Secondly, the industry specification allows to increase the explanatory power even more, and industry variables are significant, which is consistent with the hypothesis that Brexit has diverse consequences for different industries. Specification of the source does not have such effect, although this is most likely caused by data specification problem.

Below are the graphs for comparison of fitted and predicted values for the regression (2) and regression (5). It can be seen, that (5) model accounts for significantly more variation. Model (5) was chosen in order to compare the results on FTSE100 and DJIA. Although the large part of variation is still not explained, the model accounts for several shocks, caused by Brexit. Still, the small R² problem exists in all such works, because of the nature of market indexes. Moreover, the data sample starts in the June 2016, when the Brexit was one of the most thrilling topics, so the harsh shock of returns in the beginning was explained by Brexit.

Figure 11. Graphs of fitted VS actual values for model (2) (on the left) and (5) (on the right), for FTSE100 daily returns.

Figure 12. Graphs of fitted VS actual values for model (2) (on the left) and (5) (on the right), for DJIA daily returns.

4.4 Robustness

However, we need to ensure that the model is robust to extreme values and that the results are not driven by outliers. There are several ways how robust checks can be done in econometric analysis. First of all, we can just exclude outliers from our sample, but this can result in other estimation problems, as we decrease the number of observations even more by adding lags to the model. Another choice is to use the robust regression estimation by Stata (“rreg” command), which accounts for the extreme values, while not excluding them. However, it does not estimate R² or informational criteria automatically, so the package called “rregfit” was installed additionally, which allows to correctly estimate both R² and AIC. Another issue about the robust regression in Stata that it does not allow to include White standard errors, so the estimates are subject to the heteroscedasticity problem.

All the regression specifications form the above analysis were run using “rreg” command, in order to compare the estimates and make some conclusions about the robustness of the models. In general, the results were the same for models (2), (3) and (4), while for both FTSE100 and DJIA the significance of lagged returns, lagged logarithm of volume and lagged residuals increased, with slight decrease in R². However, with augmented model the coefficients' significance changed.

Figure 13. Comparison of robust augmented regressions

Variable	FTSE100	DJIA
L1.return	.13742476	-.07305148
	0.0178	0.1283
L1.res	-1.4720331	3.0136296
	0.7066	0.1968
L1.logvlm	-.00365821	.00118716
	0.0539	0.0619
DoW	.00141138	-.00032602
	0.0974	0.6312
Jan	.00082383	.00193723
	0.5460	0.0936
Agriculture	.00019945	.00002815
	0.1394	0.7917
Energy	.00032238	.0002558
	0.0971	0.0729
Retail	.00008297	.00006154
	0.5041	0.5334
Media	-.00006802	.00013707
	0.3869	0.0340
Health Care	-.00009535	-.00031574
	0.5838	0.0219
Automotive	.00002888	.00001035
	0.7916	0.9036
Resources	.00007255	.00022238
	0.6745	0.1137
Consumer Goods	.00005075	.00002132
	0.4907	0.7166
Business	-.00003935	-.00002658
	0.5140	0.5720
Construction	-.00013556	-.00002984
	0.1521	0.7096
Financial Services	.00001754	6.211e-06
	0.0700	0.4178
Industrial Goods	.00034836	-.00014883
	0.0555	0.3002
Technology	-.00014405	-.00007534
	0.3127	0.4985
Logistics	-.00002578	.00006868
	0.8282	0.4587
_cons	.07362598	-.02317023
L1.return	0.0578	0.0622
R-squared	.08362789	.08200302
AIC	288.51985	433.49651

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.

In general, the model seems to be improved by introduction of the robust regression, while it does not affect the general conclusions about the influence. The specification of industries still increases the explanatory power of the regression. Moreover, other explanatory variables became more significant predictors of stock returns. Among significant industries for FTSE100 now are Energy, Financial Services and Industrial Goods, while Media is no more significant. Overall this results do not contradict the common logic about Brexit influence. As for DJIA, significant industries now are Energy, Media and Health Care, which is drastically different from the results of previous estimation. It can be driven by the heteroscedasticity problem, which can be seen even on the graphical representation of DJIA returns.

Figure 14. Graphs of fitted VS actual values for robust estimation of model (5) for FTSE100 (on the left) and for DJIA (on the right).

Overall, robust regressions did better on FTSE100 data, then on DKIA, which might be caused by DJIA data specificities, for example heteroscedasticity. In general, robust regressions support the main evidence about influence from previous models. However, the main drawback of this regressions is inability to use White standard errors. Also several works suggest using the ARIMA model to model each time series first, and than obtain residuals. Or several models even include GARCH regressions. However, the data specificity does not allow to run such regressions in Stata.

5. Limitations

This study is subject to several limitations, which can be improved in future. First of all, the sentimental analysis of articles was applied not to the whole articles, but for some abstracts, as the whole articles were not available for the analysis. Moreover, as it was stated in Loughran Mcdonald (2011), common methods of textual analysis make mistakes in revealing the true sentiment of financial texts, as most of them are based on the general dictionaries. They count the number of negative and positive words, while for financial market the markers of negativity are not the same, as in the everyday speech. So the textual analysis technic might be improved in accordance to the specificities of financial market and future works may focus more on the sentiment analysis.

Secondly, the choice of the model to account for Brexit influence is reasonable, but it does not incorporate some other channels, through which Brexit can possibly influence financial markets. Among them may be changes in macroeconomic fundamentals, but for now it is not possibly to correctly specify this effects, so it needs more investigation.

Also the common problems of regression analysis are present in this model, as well as the problem with market index predictability. In particular, some other financial indexes can be studied instead of FTSE100 and DJIA, for more precise conclusions about influence of Brexit on each particular company or industry. In addition, the algorithm of assignment of the industry type for each article is not clear and may be biased, although it seems reasonable to trust Dow Jones company, as being one of the largest in industry.

Conclusions

Overall this work is one of the first to study the influence of some event through the media coverage, and it is devoted to rather unique event, which has not yet been studied well. The literature on this topic is still emerging and this work contributes to the methodology of such studies, as the specification of industry type has proven to increase the explanatory power of the regression and this specification can be further applied to different events. Moreover, the effect of articles in different industries has significantly different influence on the daily returns of market indexes, which proves that different industries has diverse reactions on Brexit.

In general, the results are robust to extreme values, while there is some difference in influential industries in FTSE100. However, the change in significant industries for DJIA does not logically follow from previous results, which may imply heteroscedasticity problem of robust regressions.

In conclusion, the Brexit has obviously influenced the financial markets and this effect has not been carefully examined by now. Works on Brexit are only emerging and its overall influence will be seen only in several years from now, while some conclusions can be drawn already on the existing information and this work provides good insight into this topic. Also it allows future researchers to improve their methodology, using the model specifications used in this work.

Further implications

The results of this work rise several other research questions. The methodology allows to study the effect of Brexit media coverage on other market quotes, as well as to include some possible macroeconomic characteristics. Moreover, the database, gained for this work, can be used for future studies on Brexit effect on not only financial markets, but any other fields of interest.

Bibliography

1. Akhigbe A., Larson S.J., Madura J. (2002). Market Underreaction and Overreaction of Technology Stocks. Journal of Psychology and Financial Markets, 3(3), 141-151.

2. Campbell J.Y., Grossman S.J., Wang J. (1993). Trading Volume and Serial Correlation in Stock Returns. Quarterly Journal of Economics, 108(4), 905-939.

3. Chiang C.-F., Knight B. (2011). Media Bias and Influence: Evidence from Newspaper Endorsements. Review of Economic Studies, 78(3), 795-820.

4. Clarke H.D., Goodwin M., Whiteley P. (2017). Why Britain Voted for Brexit: An Individual-Level Analysis of the 2016 Referendum Vote. Parliamentary Affairs, 70(3), 439-464.

5. de Jong P., Elfayoumy S., Schnusenberg O. (2017). From Returns to Tweets and Back: An Investigation of the Stocks in the Dow Jones Industrial Average. Journal of Behavioral Finance, 18(1), 54-64.

6. Dougal C., Engelberg J., Garcнa D., Parsons C.A. (2012). Journalists and the Stock Market. Review of Financial Studies, 25(3), 639-679.

7. Engelberg J.E., Parsons C.A. (2011). The Causal Impact of Media in Financial Markets. The Journal of Finance, 66(1), 67-97.

8. Fiss P.C., Hirsch P.M. (2005). The discourse of globalization: Framing and sensemaking of an emerging concept. American Sociological Review, 70(1), 29-52.

9. Foster F.D., Warren G.J. (2016). Interviews with Institutional Investors: The How and Why of Active Investing. Journal of Behavioral Finance, 17(1), 60-84.

10. Gerber A.S., Karlan D., Bergan D. (2009). Does the Media Matter? A Field Experiment Measuring the Effect of Newspapers on Voting Behavior and Political Opinions. American Economic Journal: Applied Economics, 1(2), 35-52.

11. Goodwin M.J., Heath O. (2016). The 2016 Referendum, Brexit and the Left Behind: An Aggregate-level Analysis of the Result. The Political Quarterly, 87(3), 323-332.

12. Huberman G., Regev T. (2001). Contagious Speculation and a Cure for Cancer: A Nonevent that Made Stock Prices Soar. The Journal of Finance, 56(1), 387-396.

13. Jensen M.D., Snaith H. (2016). When politics prevails: the political economy of a Brexit. Journal of European Public Policy, 23(9), 1302-1310.

14. Joanna Strycharz, Strauss N., Trilling D. (2018). The Role of Media Coverage in Explaining Stock Market Fluctuations: Insights for Strategic Financial Communication. International Journal of Strategic Communication, 12(1), 67-85.

15. Laskin A.V. (2014). Nonfinancial Information in Investor Communications: International Journal of Business Communication.

16. Loughran T., Mcdonald B. (2011). When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks. Journal of Finance, 66(1), 35-65.

17. Lu X., White H. (2014). Robustness checks and robustness tests in applied economics. Journal of Econometrics, 178, 194-206.

18. Malkiel B.G., Fama E.F. (1970). Efficient Capital Markets: A Review of Theory and Empirical Work*. The Journal of Finance, 25(2), 383-417.

19. Poitras M. (2004). The Impact of Macroeconomic Announcements on Stock Prices: In Search of State Dependence on JSTOR. Southern Economic Journal, 70(3), 549-565.

20. Ronchetti E. (1985). Robust model selection in regression. Statistics Probability Letters, 3(1), 21-23.

21. Scheufele B., Haas A., Brosius H.-B. (2011). Mirror or Molder? A Study of Media Coverage, Stock Prices, and Trading Volumes in Germany. Journal of Communication, 61(1), 48-70.

22. StrauЯ N., Vliegenthart R., Verhoeven P. (2016). Lagging behind? Emotions in newspaper articles and stock market prices in the Netherlands. Public Relations Review, 42(4), 548-555.

23. Tetlock P.C. (2007). Giving Content to Investor Sentiment: The Role of Media in the Stock Market. Journal of Finance, 62(3), 1139-1168.

24. Tetlock P.C. (2011). All the News That's Fit to Reprint: Do Investors React to Stale Information? Review of Financial Studies, 24(5), 1481-1512.

25. Thelwall M., Buckley K., Paltoglou G., Cai D., Kappas A. (n.d.). Sentiment strength detection in short informal text. Journal of the American Society for Information Science and Technology, 61(12), 2544-2558.

26. Uhl M.W. (2014). Reuters Sentiment and Stock Returns. Journal of Behavioral Finance, 15(4), 287-298.

27. Undurraga T. (2016). Making News, Making the Economy: Technological Changes and Financial Pressures in Brazil. Cultural Sociology, 11(1), 77-96.

28. Vettehen P.H., Beentjes J., Nuijten K., Peeters A. (2010). Arousing News Characteristics in Dutch Television News 1990-2004: An Exploration of Competitive Strategies. Mass Communication and Society, 14(1), 93-112.

29. Walker N. (n.d.). Brexit timeline: events leading to the UK's exit from the European Union, 29.

30. Xing F.Z., Cambria E., Welsch R.E. (2018). Natural language based financial forecasting: a survey. Artificial Intelligence Review, 50(1), 49-73.

Appendix

Correspondence of variable names:

“DJW”: “Dow Jones Newswires”,

“MNB”: “Major News and Business Sources”,

“PRW”: “Press Release Wires”,

“RW”: “Reuters Newswires”,

“WSJ”: “The Wall Street Journal - All sources”,

“others”: “All other sources”

Figure 15. Plot of FTSE100 closing prices

Figure 16. ADF test for FTSE100 closing prices time series



Figure 17. Plot of FTSE100 returns

Figure 18. ADF test for FTSE100 returns time series

Figure 19. Plot of DJIA closing prices

Figure 20. ADF test for DJIA closing prices time series

Figure 21. Plot of DJIA returns

Figure 22. ADF test for DJIA returns time series

Размещено на Allbest.ru

...