Media connection and return comovement

https://doi.org/10.1016/j.jedc.2021.104191Get rights and content

Abstract

Media news may cover multiple firms in one article, which establishes a media connection across firms. We propose a media connection strength (MCS) measure between two given firms, which is defined as the number of news articles co-mentioning these two firms. We show that the MCS measure can significantly explain and forecast return comovement of media-connected firm-pairs. Further analyses show that our results are robust to various alternative explanations. We argue that the MCS measure can capture comprehensive and complex correlated fundamental information among media-connected firms and hence may provide a new mechanism for return comovement beyond the existing rational- and behavioral-based explanations.

Introduction

Understanding the driving forces of stock return comovement has important implications for many applications, such as risk management (Philippe (2001)), portfolio allocation (Qian et al. (2007)), asset price dynamics (Rosenberg and Schuermann (2006) and Brooks et al. (2002)), and trading strategies (Gatev et al. (2006); Papadakis and Wysocki (2007); Chen et al. (2019)). Theoretically, the stock return comovement of two firms can be driven by common shocks to their fundamentals, such as future cash flows or discount rates. Therefore, empirically explanatory variables based on the fundamental information of a firm, usually quantitative information, have been proposed to explain comovement (e.g., Shiller (1989); Fama and French (1993); Chen et al. (2019); Green and Hwang (2009)). In addition, recent studies document ǣexcess comovementǥ that cannot be explained by fundamental information and propose alternative behavioral bias-based explanations for return comovement (e.g., Karolyi and Stulz (1996); Barberis and Shleifer (2003); Barberis et al. (2005); Bekaert et al. (2005); Kumar and Lee (2006); Kallberg and Pasquariello (2008); Anton and Polk (2014); Kumar et al. (2016)).

In this study, we investigate a new potential channel for return comovement; namely, media connection, which is beyond the existing fundamental information channels and is not based on behavioral bias. In particular, we examine whether correlated fundamental information contained in multi-firm news articles could be an important driver for return comovement. Multi-firm news articles covering two or more firms may reflect journalists collective opinions on a certain complicated fundamental connectivity between the firms that are cited. In addition to well-documented fundamental links including customer-supplier links (Cohen and Frazzini (2008)), text-based links (Hoberg and Phillips (2016)), and EDGAR co-search links (Lee et al. (2015)), multi-firm news articles may be a proxy of other complex economic linkages, such as strategic partnerships, intra- and inter-sectoral competitive links, and credit, financing, banking, and subsidiary relations (Schwenkler and Zheng (2019)). Therefore, we propose a media connection strength (MCS) measure based on the news data collected by Thomson Reuters News Analytics (TRNA), which is defined as the number of news articles covering two firms.1 We expect that the MCS measure can offer incrementally better explanatory power regarding the comovement of fundamental performance and return comovement for media-connected firm pairs that goes beyond those existing variables measuring correlated fundamentals across firms.

We conduct empirical tests to verify our hypothesis. First, we examine the determinants of the MCS. We run cross-sectional regressions of our MCS measure on the firm-pair characteristics that are documented in the literature as associated with return comovement. We find that variables such as the mutual fund common ownership and common analyst coverage are significantly and positively associated with MCS. Further tests show that many fundamental similarities between co-mentioned stocks can significantly explain MCS, suggesting that fundamental similarities of stock pairs are indeed captured by our MCS.

Next, we examine whether the MCS measure can predict fundamental similarities of media-connected firm pairs. Following Livnat and Mendenhall (2006), we define the standardized unexpected earnings (SUE) as a proxy for the fundamental information of a firm and find that the correlation of SUE, when controlling for other well-known return comovement sources and other specific fundamental similarities, can be positively and significantly predicted by the MCS among the co-mentioned stock pairs. Specifically, one standard deviation of the increase in the MCS is associated with an increase of 0.6% in the correlation of SUE in quarter t and an increase of 1.0% in quarter t+1. Meanwhile, to further validate our fundamental argument for a media connection, we employ the correlation of sales growth (SALEG) as an alternative measure of fundamental comovement, and the results are also consistent with our expectation; that is, a one standard deviation increase in the MCS of quarter t is associated with an increase of 1.8% in the correlation of SALEG among firm pairs in quarter t and an increase of 1.9% in quarter t+1, when controlling for other well-known return comovement sources and other fundamental similarities. Finally, we also measure the correlated fundamentals of stock pairs as the negative value of the absolute difference in percentile ranking of SUE or SALEG for the media-connected stock pairs, and the results remain similar. This suggests that media connection contributes to an increasing fundamental news comovement for media-connected firm pairs beyond what can be explained by other stock-pair similarities documented in the literature.

Subsequently, we conduct a simple univariate analysis to examine if the predictive power of our MCS on fundamental similarities can be extended to an association between MCS and excessive return comovement. We first split all firm pairs into three groups based on the MCS and calculate the average return correlation of these three groups; we then form a high-minus-low group that captures the return correlation difference between the high-MCS group and the low-MCS group. We find that the return correlation of the high-MCS group is significantly higher than that of the low-MCS group. In addition, the average return correlation of the stock pairs with no-zero MCS is significantly higher than that of those with zero MCS, indicating that our MCS is indeed positively associated with a stronger return comovement. We then perform cross-sectional regressions to formally investigate if the contemporaneous and future excess return comovement of media-connected firm pairs can be significantly explained through the MCS measure. Consistent with our hypothesis, the MCS is positively and significantly associated with not only cross-sectional variation in the strength of fundamental similarities but also a stronger return comovement among firm pairs in the current month and next month, when controlling for other well-known return comovement sources and a list of firm-pair fundamental similarities. A one standard deviation increase in MCS is associated with an increase of 0.9% in return comovement and an increase of 1.0% in excess return comovement (under the Fama and French (2015) five-factor model (FF5)) in the current month, and the results are similar to that of the next month. Moreover, the predicted variation in FF5 residual correlation is in the average range of 5.33% to 16.75%, with a mean abnormal correlation of 7%, suggesting that the effect of MCS is also economically significant.

Kumar and Lee (2006), Kumar et al. (2013), and Kumar et al. (2016) find that stock prices tend to co-move if they share the same sentiment. Given that stocks mentioned in the same news article may share a similar sentiment, we examine whether the return comovement predictability of our MCS is independent of investor sentiment. Through defining high- and low-sentiment periods based on the sentiment index of Baker and Wurgler (2006), we find that the predictability of the MCS is robust and similar in magnitude under different sentiment regimes, suggesting that investor sentiment is unlikely to explain the results. In addition, we find that the MCS measure is positively associated with trading activity comovement, measured as the correlation of order flow or turnover ratio between paired firms. This is consistent with the hypothesis that the correlated fundamentals are captured through the MCS measure, which may induce correlated trading activities and the corresponding return comovement.

Nevertheless, investors may simply trade on media-connected stock pairs due to category trading habitat instead of the fundamental similarities between those stock pairs. Barberis and Shleifer (2003) argue that investors allocate capital at the level of asset categories rather than individual stocks. Firms are mentioned frequently in the same news article because they are viewed as being in the same asset category by market participants, including news media and investors. Alternatively, investors view of asset categories could be influenced by the frequency of stocks being mentioned together in news articles. In either case, investors trading in and out of media-connected stocks that are viewed as being in the same category could induce an excessive comovement in their returns.

However, this is less likely to explain our results, given that the coefficients of MCS on return comovement remain similar during both low- and high-sentiment periods. We further rule out this alternative hypothesis in two ways. Firstly, we replace the correlation of order flow (OFCOR) or the correlation of turnover (TOCOR) in the current month with the similarity in turnover (TO) and the correlation of TO in the past 60 months (VOLCOR) in the cross-sectional regressions of the return comovement so as to control for the correlated flow-induced trading induced by category trading habitat. The results show that the effect of our MCS is still there after the inclusion of these control variables. Secondly, as is detailed in the latter part of the paper, we conduct portfolio analysis to further show that category trading habitat is unlikely an explanation of our results.

Finally, we conduct robust checks to examine whether other fundamental similarities documented in the literature can be used to explain our results. Pirinsky and Wang (2006) find that stock prices tend to co-move if their headquarters are located in the same place. Thus, we add a dummy variable to our regression indicating whether two stocks are headquartered in the same U.S. county or not. Consistent with Pirinsky and Wang (2006), firm pairs located in the same county exhibit an excess return comovement in the next month. Nevertheless, the addition of the same-county dummy variable does not affect the return correlation predictability of our MCS measure.

Furthermore, Hoberg, Phillips, 2010, Hoberg, Phillips, 2016 construct a text-based industry classification based on the production description of 10,000 annual reports and find that this classification is more informative than traditional ones such as SIC or NAICS. We control this possible channel of similarity in the production of firm pairs for return comovement by including the text-based industry similarity score in our regression. The economic magnitude of the coefficient of our MCS measure is slightly weaker but remains statistically significant at the 1% level. Finally, Lee et al. (2015) find that stocks that investors co-search in the EDGAR system exhibit higher fundamental similarities than traditional industry classifications. As a measure of the attention paid by investors to EDGAR searches, the co-search measure could subsume our MCS in predicting the excess return comovement, so we include it in our regression model. Again, the return correlation predictability of our media connection measure remains significant. Overall, our results are robust to these alternative sources of stock return comovement.

Our paper contributes to several strands in the literature. First, we contribute to the debate on whether stock comovement is caused by information or noise. On the one hand, stock return comovement is driven by similarities in fundamentals such as size and book-to-market (Fama and French (1993), cash flows (Chen et al. (2019), price level (Green and Hwang (2009)), and analyst coverage (Chan and Hameed (2006); Muslu et al. (2014); Hameed et al. (2015)). On the other hand, some studies attribute the excess return comovement that cannot be explained by fundamental commonalities to behavioral biases such as categorical trading (Barberis and Shleifer (2003); Barberis et al. (2005)), sentiment (Kumar and Lee (2006); Kumar et al. (2013); Kumar et al. (2016)), investor attention (Peng and Xiong (2006); Huang et al. (2019)), forecast errors (Israelsen (2016)), and trading pressure from institutional investors (Anton and Polk (2014)). Our paper shows that stock return comovement for media-connected stock pairs is driven by the common fundamental information conveyed by media reports. Moreover, our results are consistent with the predictions based on the models of profit-maximizing information producers (Veldkamp (2006)). In particular, information production is non-rival with a high fixed cost of discovery and a low marginal cost of replication. Therefore, competitive producers such as news media tend to provide news that can maximize their profits from investors. Moreover, the information useful for predicting a subset of stocks could attract more investors than information that is useful for predicting only a single stock. As a consequence, given the high per-unit cost of information production, information useful for more stocks would induce news media to extend their coverage to these stocks, especially those whose fundamentals correlate more with each other.

Second, this study contributes to the literature on lead-lag effects in the returns of economically related stocks. Early studies tend to group firms according to their fundamental characteristics so as to find economically related firms, such as Lo and MacKinlay (1990), Brennan et al. (1993), Badrinath et al. (1995) and Chordia and Swaminathan (2000), while recent studies focus on specific economic links to identity firm connections, such as Cohen and Frazzini (2008), Lee et al. (2019), and Ali and Hirshleifer (2019). In this paper, we propose another way of identifying firm peers based on media connection, which can drive return lead-lag effects among firms via information spillover across media connection.

Third, we contribute to the literature that investigates the role of news media in financial markets ((Huang et al., 2021). Existing studies largely focus on two perspectives, one of which is to show that news tone or sentiment can be used to predict a firms future performance. Tetlock (2007) argues that language, especially negative language, could be used to predict excess market returns. Tetlock et al. (2008) analyze firm-specific news to explore the predictability of cross-sectional return. Another perspective examines whether media coverage can strongly affect future stock returns. For example, Fang and Peress (2009) find that stocks with no media coverage can earn higher returns than stocks with high media coverage, even after controlling for well-known risk factors. Engelberg and Parsons (2011) find that local media coverage can strongly predict local trading. However, there is limited evidence for the interactive effect induced by common media coverage on return comovement. In this paper, we construct a proxy for fundamental similarities between two media-connected firms through media connection, and our empirical results suggest that our MCS measure is significantly associated with fundamental similarities and return comovement.

The rest of the paper is organized as follows. Section 2 describes the data used in this paper and explains how the main variables are constructed. In Section 3, we present the main empirical results. Alternative explanations are examined and robustness tests are conducted in Section 4 to rule out alternative hypotheses. Section 5 concludes.

Section snippets

Data and main variables

We obtain stock return data from CRSP, accounting data from Compustat, and analyst coverage data from IBES. Institutional holding data is obtained from Thomson Reuters Institutional Holdings (13F) database and mutual fund holdings data is collected from the Thomson Reuters Mutual Fund Holdings database. Our data sample includes all common stocks listed on the NYSE, AMEX, and NASDAQ.

The media news data used in this paper is collected from TRNA over the period from January 1996 to December 2014.

Cross-Sectional determinants of media connection strength

We start our empirical analysis by examining whether stocks with similar fundamentals are more likely to be mentioned in the same news article. Therefore, we investigate what determines the MCS to understand the mechanism between common media coverage and firm-pair characteristics. Specifically, we run Fama and MacBeth (1973) regressions of MCS on firm-pair characteristics mentioned in data section. To see the improvement of R2, we gradually add independent variables from Column (1) to Column

MCS And category trading habitat

Barberis and Shleifer (2003) argue that investors allocate capital at the level of asset categories rather than individual stocks. Firms are mentioned frequently in the same news articles, which may be because they are viewed as being in the same asset category by market participants including news media and investors. Alternatively, investors view of asset categories could be influenced by the frequency of stocks being mentioned together in news articles. In either case, investors trading in

Conclusions

In this study, based on news data from Thomson Reuters, we construct a media connection strength (MCS) measure across two given firms, which is defined as the number of multi-firm news articles covering those two firms. We find that the MCS measure is correlated with fundamental similarity and can significantly explain and forecast return comovement for media-connected firm pairs. Further analyses show that our results are less likely to be driven by investors’ category trading activities or

References (51)

  • L. Peng et al.

    Investor attention, overconfidence and category learning

    J financ econ

    (2006)
  • J.V. Rosenberg et al.

    A general approach to integrated risk management with skewed, fat-tailed risks

    J financ econ

    (2006)
  • U. Ali et al.

    Shared analyst coverage: unifying momentum spillover effects

    J financ econ

    (2019)
  • M. Anton et al.

    Connected stocks

    Journal of Finance

    (2014)
  • S.G. Badrinath et al.

    Of shepherds, sheep, and the cross-autocorrelations in equity returns

    Review of Financial Studies

    (1995)
  • M. Baker et al.

    Investor sentiment and the cross-section of stock returns

    Journal of Finance

    (2006)
  • G. Bekaert et al.

    Market integration and contagion

    Journal of Business

    (2005)
  • M.J. Brennan et al.

    Investment analysis and the adjustment of stock prices to common information

    Review of Financial Studies

    (1993)
  • C. Brooks et al.

    The effect of asymmetries on optimal hedge ratios

    Journal of Business

    (2002)
  • H. Chen et al.

    Empirical investigation of an equity pairs trading strategy

    Manage Sci

    (2019)
  • T. Chordia et al.

    Trading volume and cross-autocorrelations in stock returns

    Journal of Finance

    (2000)
  • L. Cohen et al.

    Economic links and predictable returns

    Journal of Finance

    (2008)
  • K. Daniel et al.

    Evidence on the characteristics of cross sectional variation in stock returns

    Journal of Finance

    (1997)
  • J.E. Engelberg et al.

    The causal impact of media in financial markets

    Journal of Finance

    (2011)
  • E.F. Fama et al.

    Risk, return, and equilibrium: empirical tests

    Journal of Political Economy

    (1973)
  • Cited by (1)

    • Aerospace competition, investor attention, and stock return comovement

      2023, Journal of Economic Behavior and Organization

    We thank Dashan Huang, Jianfeng Hu, Roger Loh, Weikai Li, Chishen Wei, Weina Zhang, and participants at 2019 SMU Summer Camp and Conference on the Theories and Practices of Securities and Financial Markets 2019 for their helpful suggestions. We also thank Eugene Fama and Kenneth French for sharing the Fama-French factors returns, Malcolm Baker and Jeffrey Wurgler for sharing the sentiment index data, and Gerard Hoberg and Gordon M. Phillips for sharing the text-based network industry classifications (TNIC) data. Li Guo acknowledges the financial support sponsored by Shanghai Pujiang Program. Jun Tu acknowledges that the study was funded through a research grant from Sim Kee Boon Institute for Financial Economics. All errors remain our responsibility.

    View full text