Thesis Laurens Mittertreiner
6013B0326Y 10607803
Economie & Bedrijfskunde
Finance & Organization
Magdalena Rola-Janicka
31st January 2017
The Media Effect Revisited: Changes in the Relation Between Media and Stock Returns due to the Rise of the Internet
Table of contents
Introduction Page 3
1. Literature Review Page 4
2. Data Description Page 6
3. Methodology Page 8
4. Results and Empirical Analysis Page 11
5. Discussion and Conclusion Page 15
Reference List Page 16
Appendix Page 17
Statement of Originality
This document is written by Laurens Mittertreiner who declares to take full responsibility for the contents of this document.
I declare that the text and work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Introduction
Fang and Peress (2009) found that stocks with no media coverage earn a higher return than stocks with high media coverage, adjusted for well-known risk factors. This is called the media effect. They did research on this effect using a sample ending in 2002. The proxy for media coverage they used was printed media only, since the coverage of web sources did not play a significant role yet in that time. Nowadays the world is digitalizing quickly and the role printed media plays in news disseminating decreases. Due to the rise of the internet more news sources have come available. Not only the amount of news sources has grown but also the ease with which news can reach investors. In the past, news about stocks could reach the investor mainly through printed media and analyst reports. Nowadays, investors continually have access to firm specific news by getting online on news sites, social media, etcetera. Furthermore, it has become easier for anyone to publish news on anything.
The interest in the relation between media and the stock market is on the rise. This thesis contributes to this topic by examining whether the rise of the internet has had an impact on the media effect. Therefore, the cross-sectional relations between media coverage and stock returns are researched over two time periods, one before and one after the rise of the internet. Portfolios varying by media coverage are formed and raw return differences are tested in univariate analysis. Furthermore, the relation between media coverage and stock returns is tested in two regression settings. A Chow test is performed to test for significantly different coefficients of media coverage in the two time periods. The results show that the media effect is stronger in the second sample. Also, the Chow test result is found to be significant. The results could suggest that due to the rise of the internet the media effect has become stronger. An interesting implication of this would be that firms can affect their cost of capital by actively generating investor interest through online information dissemination. However, there are several limitations to this research and in order to draw the above conclusion with certainty it is essential to include web coverage data in further research.
The remainder of this thesis contains the following sections. In section I the important literature related to this thesis is reviewed. Section II involves the data gathering process and gives descriptive statistics on the data. Section III describes the methodology to answer the research question. Hypotheses following this research question are also formulated in this
section. In section IV results will be presented and an empirical analysis will be made. Section V concludes and discusses.
1. Literature Review
This thesis relates to the literature that studies the relation between media and the stock market. Several recent papers did research on the link between media and stock liquidity and found a positive relation. For example, Antweiler and Frank (2004) find that stock news predicts market volatility. The effect of news on stock returns is statistically significant but economically small. Meschke (2004) documents a strong run-up and reversal on stocks for 11 days after CEO interviews, broadcasted on CNBC. Grullon, Kanatas et al. (2004) report that higher advertising expenses, as a proxy for the extent a firm is known by investors, cause higher common stock liquidity and a higher amount of individual and institutional investors.
Whereas the above papers focus on the relation between media and stock liquidity, there are also several papers about the relation between media and stock returns. Klibanoff et al.
(1998), for instance, find that dramatic country-specific front page news appearing in the New York Times affects the pricing of closed-end country funds. They argue that prices normally underreact to changes in fundamentals, but with accompanying front page news prices move more towards fundamentals. Their findings confirm the expectation that some investors react more quickly due to news dissemination. Tetlock (2007) in his papers measures the linguistic content of news articles. He documents that media pessimism predicts downward pressure on stocks and subsequent reversal. In his paper, he quantifies the level of pessimism in a popular column in the Wall Street Journal and finds that stock prices undergo downward pressure and afterwards a reversal to fundamentals caused by a high level of media pessimism. Tetlock et al. (2008) find that the portion of negative words in news stories predicts individual firms’ stock returns and accounting earnings. Chan (2003) finds a time-series relation of return continuations and reversals. He reports that stocks that have low returns in times they have head line news (Chan names them “news losers”) experience negative return drift for over a year whereas stocks that have low returns accompanied by no news (these stocks are called “no-news losers”) do not experience this effect and their returns reverse. Closely related but distinct from Chan (2003) in various ways is the paper of Fang and Peress (2009). Fang and Peress are the first to find a cross-section relation between media coverage and stock returns. They find that stocks with no media coverage earn higher returns than stocks with high media
coverage. This is called the media effect. Because of certain publication time of printed media, the news covered is not likely to be the most actual news, so they state that the extent of information dissemination affects stock returns. In Fang and Peress (2009) possible explanations of the media effect are researched.
The first possible explanation is the return continuations and reversals documented by Chan (2003). Firms with low returns accompanied by news headlines experience negative return drift for over 12 months whereas firms with low returns accompanied by no news see their returns reverse. This effect could generate the difference in returns between high and low media covered stocks. The distinction between the two researches however, is that Chan uses news headlines and Fang and Peress use news articles as a proxy for media coverage. As it turns out, many stocks classified as no coverage stocks by Fang and Peress, have had news headlines. This means that stocks classified as news-losers by Chan would have been no coverage stocks in the research on the media effect. Thus, equalizing no-news losers and news losers as no coverage stocks and coverage stocks is not accurate. Fang and Peress conclude after testing whether the media effect is caused by negative return drift of high coverage losers or reversal of negative returns of no-coverage losers that the findings of Chan don’t explain the media effect.
Another possible explanation is called the “impediments-to-trade” hypothesis, also known as the “illiquidity” hypothesis. The media effect is a return differential between stocks adjusted for well-known risk factors. This return difference is an arbitrage opportunity that is not ruled out by the market. Therefore, there must be impediments causing rational agents not to trade on those no coverage stocks. An important impediment would be illiquidity. According to the liquidity preference theory, investors, all other factors being equal, prefer cash or other highly liquid holdings. Stocks that are more liquid are easier to sell fast for full value. So, investors must be offered a higher return, as a compensation for higher risk, to be willing to sacrifice more liquidity. This hypothesis suggests that the media effect should be the strongest among the most illiquid stocks. After testing for this, Fang and Peress cannot conclude with certainty that the illiquidity hypothesis explains the media effect.
The third and last explanation is the so-called “investor recognition hypothesis”. Higher media coverage of stocks can cause higher investor recognition by disseminating news to a broad audience. According to Merton (1987), in informationally incomplete markets stocks
that are poorly recognized by investors need to offer a higher return to compensate their investors for being imperfectly diversified. The market for stocks that are poorly recognized will be much smaller. This causes an increase in firm-specific risk that must be compensated by a higher offered return. Part of the idiosyncratic risk of the firm must be priced because of imperfect diversification caused by a lack of investor recognition. Thus, the return difference caused by the media effect is not a mispricing but a compensation for higher risks not
included in standard pricing models. According to this hypothesis the media effect should be the strongest among stocks that are poorly recognized. Poorly recognized stocks are identified by Fang and Peress as stocks with low analyst forecast and high individual ownership. Those two variables reflect the extent of information incompleteness. After testing for it, Fang and Peress find strong evidence for the investor recognition hypothesis given the results that the media effect is greater for those stocks.
In conclusion, Fang and Peress state that the investor recognition hypothesis causes the media effect and the impediments-to-trade hypothesis maintains the effect.
2. Data Description
The two samples that are used contain 15 companies listed on the AMX or AScX. The Amsterdam Midkap Index is the second biggest Amsterdam stock index and contains stocks 26 up to and including 50, according to size. The Amsterdam Small Cap Index is the third biggest Amsterdam stock index. Stocks 51 up to and including 75 are listed in this index. Since the data on media coverage contains articles in Dutch national newspapers the sample companies should preferably be Dutch, preferably have mainly Dutch investors and should not be listed on the biggest Amsterdam index, the AEX. Companies listed on the AEX are more internationally active with more international investors and therefore Dutch printed media data would not be a good proxy for media coverage. The sample periods are January 1st 1997 till December 31st 2002 and January 1st 2010 till December 31st 2015. The sample companies are the same for both samples and thus automatically existent and listed on a stock exchange during both sample periods. Fang and Peress (2009) end their sample period in 2002 when internet media start spreading, but they do not yet play a substantial role in news dissemination. In the second sample period it can be said with certainty that the news
dissemination through the internet does play a substantial role and the internet is available to almost everyone.
The data on media coverage are retrieved from the LexisNexis academic database. I search for the number of articles per company per month in the three main Dutch national
newspapers (NRC Handelsblad, Telegraaf and Volkskrant) and the main Dutch financial newspaper (Financieele Dagblad). The number of articles related to the companies serve as a proxy for media coverage. In the LexisNexis database articles related to companies are identified by searching for the company name in the Hlead section. This sections contains the headline, highlight and lead sections of articles. Whether in doubt of relevance of an article for the sample specific company I make the decision with a manual check of the content.
In the tables below descriptive statistics of media coverage data are presented. Table I reports the proportion of all three media coverage categories. We can see that a little bit less than half of all observations contains zero articles. Low media coverage has a lower proportion than no media coverage but higher than the lowest proportion, that of high media coverage. Table II reports the descriptive statistics of the variable articles. The sample companies have on average approximately 2 articles written about them per month with a maximum number of 28 articles per month.
Table I Proportion of Media Coverage Categories
Observations: 2038 Proportion
No Media Coverage 0.4661432
Low Media Coverage 0.3312071
High Media Coverage 0.2026497
Table II Descriptives of Variable Articles
Variable Obs Mean Std. Dev. Min Max
Articles 2.038 1.776251 2.956896 0 28
Monthly stock prices are obtained from Datastream, as well as monthly data on shares outstanding and yearly data on earnings before interest and taxes (EBIT).
3. Methodology
The raw return differences are examined in univariate analysis. For both samples as well as for the whole data set the media effect is computed in raw returns using the same following method. Each month the sample companies are divided into three categories: no coverage, low coverage and high coverage. The amount of coverage equals the number of articles written about the specific company. No coverage corresponds to 0 articles. The division between low coverage and high coverage is made by means of the median. Companies with a number of articles equal or lower to the median are put in the low media coverage portfolio and companies with more articles than the median form the high media coverage portfolio. For each month, the equal-weighted average return of the portfolios is computed. Finally, the average monthly return of each portfolio is computed. The statistical significance of the difference in returns between the no coverage portfolio and the coverage portfolios is tested by an independent samples t-test.
The relation between media coverage and stock returns is also analyzed in a regression setting. The regression equation is the following:
Y = β0 + β1X1 + β2X2 + β3X3 + β4X4 + ε (1)
Where
Y is the dependent variable stock return β0 is the constant
X1 is the independent dummy variable low media coverage X2 is the independent dummy variable high media coverage X3 is the control variable past month return
X4 is the control variable firm size
This is an OLS regression with the dependent variable stock return and two dummy variables, low media coverage and high media coverage respectively, as independent variables. The values for the dummy variables are from the division into the three categories stated above. Past month return and firm size are the most important control variables in this time series regression. Fang and Peress (2009) did a regression on the determinants of media coverage in
a regression setting. Firm size was found to be a significant determinant of media coverage, whereas past month return was not found to be significant. However, both variables could be correlated with media coverage and both variables could be determinants of returns. For those two reasons these control variables are included in the regression.
The regression is run for each of the samples separately in order to test whether the
coefficients of the dummy variables are significant. If found significant, there is proven to be a media effect in the concerning time period. The statistical hypotheses for these tests are the following:
H0: β1 = 0 H0: β2 = 0
H1: β1 < 0 H1: β2 < 0
Where
β1 is the coefficient of the independent dummy variable low media coverage β2 is the coefficient of the independent dummy variable high media coverage
In a second regression setting media coverage expressed in the number of articles is regressed on stock returns. The same control variables are included in the regression. This is also an OLS regression, with the following equation:
Y = β0 + β1X1 + β2X2 + β3X3 + ε (2)
Where
Y is the dependent variable stock return β0 is the constant
X1 is the independent variable articles X2 is the control variable past month return X3 is the control variable firm size
The regression is run for each of the samples separately in order to test whether the independent variable articles is significant. If found significant, there is a linear relation
between articles and return in the concerning time period and thus a media effect. The hypothesis is again that more coverage leads to lower returns. In statistical forms this would be:
H0: β1 = 0
H1: β1 < 0
Where
β1 is the coefficient of the independent variable articles
The results of the above two regressions as well as the results of a Chow test, described below, will be analyzed to answer the research question whether the media effect in the two samples differs due to the rise of the internet. The hypothesis on the difference of the media effect is not very straightforward. There are two plausible expectations. It could be that companies, getting no or low coverage by the printed media, will get more coverage due to web sources. It has become easier to publish news on anything, so companies ignored by the traditional printed media can get covered by online sources, for example by their social media accounts. This would mean that the media effect will become weaker in the second sample, since the explanations of the media effect, illiquidity and investor recognition, weaken. The alternative and more likely expectation is that web coverage is correlated with printed
coverage. Looking at data on webnews on news sites of the printed news sources support this correlation. Lots of articles covered on the web are similar to the ones covered in printed media. Because of the internet, there is better access to information and this information is quicker and more broadly disseminated. This would mean that the media effect is stronger in the second sample due to an increasing amount of coverage of already covered firms.
To check whether there is a difference between the media effect of the two samples, a Chow test is conducted on the second regression involving the independent variable articles. The Chow test determines whether the articles coefficient in the two linear regressions of both samples significantly differ.
The hypothesis of this thesis will be that the media effect in the second sample will be stronger than in the first sample. The statistical hypotheses are the following:
H0: β1(sample 1) = β1(sample 2)
H1: β1(sample 1) > β1(sample 2)
Where
β1 is the coefficient of the independent variable articles
4. Results and Empirical Analysis
The raw returns and the raw return differences of the two samples and the total sample, computed in univariate analysis, are documented in Table I below.
Table III Univariate Comparisons Media Coverage and Stock Returns
Average Monthly Return Difference Return statistics t- Difference Return statistics t-Media
Coverage No Low High No – Low No – Low No – High No – High Sample 1 -0.021561 0.63997 -1.05799 -0.66153 -0.5621 1.03643 0.8050 Sample 2 1.10129 1.10287 0.86339 -0.00157 -0.0019 0.23739 0.2008
Both
Samples 0.53987 0.87142 -0.09704 -0.33155 -0.4575 0.63691 0.7271 1Return values are given in percentages.
An unpaired two-sample t-test with unequal variances on a 5%-significance level is performed to test for significance of the average return differences between the coverage portfolios and the no coverage portfolio. None of the t-statistics are significant, as presented in the table above. To test with more observations included the same test is conducted on the average return difference of the no media coverage returns and the high and low media coverage returns instead of on the portfolio returns. These results give somewhat lower p-values, but are not statistically significant as well.
In Table IV below results can be found of the regression with the independent dummy variables low media coverage and high media coverage.
Table IV Dummy Variables Media Coverage Regressed on Stock Returns
Dependent Variable: Stock Return
___________________________________________________________________________
Sample 1 Sample 2 Both Samples
Low Media Coverage 0.0092344 -0.0011507 0.0046306
(1.18) (-0.18) (0.93)
High Media Coverage -0.0058915 -0.0031075 -0.0027599
(-0.62) (-0.43) (-0.47)
Past Month Return -0.0092097 0.0102542 0.0010745
(-0.29) (0.34) (0.05)
Firm Size1 -7.06e-06 -0.0000105 -5.71e-06
(-0.86) (-1.77)* (-1.20) Constant -0.0008379 0.0159994 0.0054795 (-0.15) (2.84)* (1.39) Observations 973 1065 2038 R2 0.0035 0.0032 0.0015 1in millions
t-statistics are reported in parentheses. *, ** and *** represent statistical significance on the 10%, 5% and 1%, respectively.
The coefficients of the dummy variables of the two samples as well as for the whole data set are not significant. The coefficients of determination of all three models are also very poor. The insignificance of the coefficients could have several reasons. There could be a bias in the form of an omitted variable bias. Two important control variables, however, are included in the regression. The regressions are also performed with robust standard errors, but the dummy variables remain insignificant. The only difference with robust standard errors is the p-value of the variable firm size, that is slightly lower in the regressions. To allow for some more variation in the independent variable and therefore probably some more explanatory power in
the regression, stock returns are also regressed on the number of articles instead of on the two dummies only. Results of that regression setting are presented below in Table V.
Table V Number of Articles Regressed on Stock Returns
Dependent Variable: Stock Return
Sample 1 Sample 2 Both Samples
Articles -0.0016307 -0.0028033 -0.0021358
(-1.53) (-2.71)*** (-2.88)***
Past Month Return -0.0094407 0.0064751 -0.0005282
(-0.30) (0.21) (-0.02)
Firm Size1 -6.72e-06 -0.0000104 -5.32e-06
(-0.82) (-1.75)* (-1.12) Constant 0.0036705 0.0198638 0.0100554 (-0.82) (4.00)*** (-1.12)*** Observations 973 1065 2038 R2 0.0033 0.0099 0.0048 1in millions
t-statistics are reported in parentheses. *, ** and *** represent statistical significance on the 10%, 5% and 1%, respectively.
The coefficient of articles is not significant in sample 1 according to the results presented above in Table V. However, this is based on a two-sided t-test. If we look at a one-sided t-test as stated in the statistical hypothesis the result is significant on a 10%-significance level. In sample 2 as well as in the total data set, the coefficient of articles is significant at a 1% significance level. So, the media effect is stronger in the second sample according to these results. This may suggest that the correlation between printed media and internet coverage is affecting the results. Internet media coverage could be captured by the articles coefficient, introducing a bias, and cause stronger results.
In both regressions, there is a possible problem with endogeneity. Endogeneity can arise through various causes, but in this regression simultaneous causality could play a role. Media coverage is a determinant of stock returns but stock returns in their turn can be a determinant of media coverage. Unusual high or low returns of a firm may be an inducement for a
newspaper article. In gathering media coverage data this possible problem has found support because part of the articles covering the sample firms were based on stock performance. The possible problem that arises is media coverage reflecting stock returns instead of causing stock returns. The simultaneous causality leads to correlation between the independent variable articles and the error term. This in turn leads to bias and inconsistency of the OLS estimator. Engelberg and Parsons (2011) also encountered this possible endogeneity problem in their research on the influence of media coverage of financial events on investor behavior. The possible simultaneous causality problem was that local media reflected trading patterns rather than causing them. To mitigate the problem they collected a number of control
variables that measured the preexisting interest in certain stocks. To address the simultaneous causality problem of this thesis I could add the control variable earnings (EBIT). Earnings is according to theory strongly related with stock returns. Data on earnings of my sample companies, however, are only available on a yearly basis. So, there would be too few observations to run a new regression.
The F-statistic and corresponding p-value of the Chow test are presented in table VI below. Table VI Results Chow test
F (2, 2032) 4.07**
Prob > F 0.0172
*, ** and *** represent statistical significance on the 10%, 5% and 1%, respectively.
As can be seen in the table above, the result of this test is significant on a 5% significance level. Therefore, it can be concluded that the coefficient of the independent variable articles differs between the two sample periods.
5. Discussion and Conclusion
The univariate analysis of the raw return differences as well as the regression of the media dummy variables on stock returns did not give significant results. A possible limitation would be the number of observations. Fang and Peress found significant results using all companies listed on the NYSE and 500 companies listed on the NASDAQ. Extending the number of sample companies in further research would be recommended. The insignificance of the two dummies in the first regression setting could also be a lack of explanatory power of the independent variables. Therefore, in the second regression setting the number of articles are regressed on stock returns. These results are found to be significant. Moreover, the effect in the second sample is stronger than in the first sample. The Chow test concluded that the coefficients of the both samples indeed differ. The hypothesis that the media effect will be stronger in the second sample due to the correlation between the printed media coverage and web coverage seems supported by the results. However, to draw strong conclusions and relate the rise of the internet to a stronger media effect, it is essential for further research to include data on web coverage. By including data on web coverage conclusions can be drawn which of the two competing hypotheses is found to be true. To tackle the problem with endogeneity in this research a solution for further research could be to also look at the content of the
coverage.
Policy implications of the media effect, already documented by Fang and Peress, are that firms can affect their cost of capital through their media relations activities. The explanation for the media effect is that firms should offer their investors higher returns due to the firms’ lack of investor recognition. In the time of the research of Fang and Peress firms were suffering from reduced analyst coverage. Firms’ media relations activities could generate investor interest to create more investor recognition and thus reduce their cost of capital. Due to the rise of the internet generating investor interest has become much easier. In the past, information dissemination had to go through traditional channels such as printed media and stock analyst reports. Firms are now able to actively spread news online through social media for example. If the rise of the internet is found to be related with the media effect firms can affect their cost of capital by actively disseminating corporate information online.
Reference List
Antweiler, Werner, & Murray Z. Frank (2004). Is all that talk just noise? The information content of internet stock message boards. The Journal of Finance (59), 1259-1293.
Barber, Brad, & Terry Odean (2008). All That Glitters: The Effect of Attention and News on the Buying Behavior of Individual and Institutional Investors. Review of Financial Studies (21), 785-818.
Chan, W. (2003). Stock Price Reaction to News and No-News: Drift and Reversal after Headlines. The Journal of Financial Economics, 70, 223-260.
Engelberg, J. & Parsons, C.A. (2011) The causal impact of media in financial markets. The Journal of Finance (66), (67-97)
Fang, L., & Peress, J. (2009). Media Coverage and the Cross-section of Stock Returns. The Journal of Finance, 64(5), 2023-2052.
Grullon, Gustavo, George Kanatas, & James P. Weston (2004). Advertising, breadth of ownership, and liquidity. Review of Financial Studies (17), 439-461.
Klibanoff, P., Lamont, O., & Wizman, T.A. (1998). Investor Reaction to Salient News in Closed-End Country Funds. The Journal of Finance (53), 673-699.
Merton, R.C. (1987). A simple model of capital market equilibrium with incomplete information. The Journal of Finance (42), 483-510.
Meshke, F.J. (2004). CEO interviews on CNBC, Working paper, Arizona State University. Stock, J.H., & Watson, M.W. (2015). Introduction to Econometrics. Pearson.
Tetlock, P.C. (2007). Giving Content to Investor Sentiment: The Role of Media in the Stock Market. The Journal of Finance (62), 1139-1168.
Tetlock, P.C., Saar-Tsechansky, M., Macskassy, S. (2008). More than Words: Quantifying Language to Measure Firm’s Fundamentals. The Journal of Finance (63), 1437-1467.
Appendix
List of sample companies:
1. Koninklijke BAM Groep nv 2. Vastned Retail N.V.
3. Wereldhave N.V. 4. Sligro Food Group N.V.
5. Eurocommercial Properties N.V. 6. TKH Group NV 7. BE Semiconductor Industries N.V. 8. Heijmans NV 9. Brunel International NV 10. Beter Bed NV 11. Accell Group N.V. 12. Kendrion N.V. 13. Nedap N.V. 14. Wessanen N.V. 15. Ordina NV
