Media Network and Return Predictability

Transcription

1 Media Network and Return Predictability Li Guo, Yubo Tao, and Jun Tu arxiv: v2 [q-fin.st] 4 Dec 2017 Singapore Management University August 13, 2017 Abstract Investor attention has long been noticed as an important driving force of stock returns. A large number of papers have been focusing on providing direct or indirect proxies for overall investor attention. However, we believe that additional attention is a more crucial driver than overall attention in affecting stock returns. In this paper, we propose a new class of predictors, media connection indices (MCI), using news tones of media news that mentions more than one stocks to proxy effects that induced by additional attention, i.e. sentiment spillover, sentiment co-movement and connected news coverage. In general, we show our predictors are powerful and outperform other predictors, such as sentiment indices and economic predictors, in terms of both in-sample and out-of-sample predictability. In-depth analysis show that the predictability of MCI mainly comes from negative news tones which is consistent with Tetlock et al. (2008). Our cross-sectional return predictability based on portfolio sortings confirms the existence of sentiment co-movement channel. JEL Classification: G11, G12, G41. Keywords: Return Predictability; Media Network; News Sentiment; Additional Attention; Sentiment Co-movement. All co-authors are from Singapore Management University. Send correspondence to Jun Tu, Lee Kong Chian School of Business, Singapore Management University, Singapore ; Telephone: (+65) tujun@smu.edu.sg. Jun Tu acknowledges that the study was funded through a research grant from Sim Kee Boon Institute for Financial Economics. The usual disclaimer applies. A previous version of this paper has been circulated under the title, Media Network and Return Predictability. 1

2 In recent years, numerous studies have applied media coverage or google search frequency to investigated how investor limited attention affects investors trading decisions and portfolio performances (see Fang and Peress, 2009; Fang et al., 2014; Da et al., 2011, etc.). Indeed, media coverage and google search frequency are natural tools for measuring overall investor attention, but it may not be the object that we really want to study. Since the investors who are holding the stock will pay attention to the news of this stock anyway, what makes a difference is additional investor attention drawn from investors who have not paid attention to this stock before (who are more likely to be marginal investors). This motivates us to study the additional attention (or co-search attention) by analysing the news that mentions more than one stocks (connected news hereafter). For a piece of news that mentions two stocks at the same time, it not only just conveys knowledge of media coverage, but also informs us there is a media linkage between these two stocks. Such media linkage enables us to weave a tangled web of stocks and lead the investors that focus on a specific stock to pay additional attention to all the attached stocks. In addition, since this news also conveys news tones (positive, negative or neutral) for both stocks, we thus have more information available for studying how additional news sentiment and co-movment of news sentiment affect stock returns. In this paper, we first construct a media network using connected news and compose a novel class of predictors based on the media news connection. Then, we show our new predictors that proxy sentiment spillover, sentiment co-movement and connected news coverage can generate negative return predictability significantly for both in-sample and out-of-sample. The intuition behind these results is that, due to short-sale constraint, the investors who pay additional attention to stocks with bad news cannot adjust their positions while those who pay additional attention to stocks with good news can buy shares. This asymmetry finally drives an overpricing of stocks with bad news, and hence leads to a negative future return. Finally, we confirm the predictability of our measure comes from aforementioned co-movement channels by predicting cross-sectional portfolios and rule out the possible explanations from investor sentiment and information co-movement by showing no predictability of those proxies. In our empirical tests, the media-connected co-movement index predicts a negative return with a higher in-sample and out-of-sample performance than sentiment indices of Baker and Wurgler (2006) and Huang et al. (2014). In fact, we have documented 2.55% and 2.70% monthly in-sample and out-of-sample R 2 s in OLS predictive regressions respectively. In addition, our findings are statistically as well as economically significant even though we control for different economic predictors used in Goyal and Welch (2008) and sentiment indices of Baker and Wurgler (2006) and Huang et al. (2014). We then test the 2

3 performance of our index in predicting returns during the recession and expansion periods documented by NBER, and it shows that our measure obtains larger and positive R 2 s in recession periods comparing to sentiment indices and combined economic predictors. Further, we find the predictability of our measure mainly comes from the negative new tones which is consistent with Tetlock et al. (2008) in which he find negative tones are more informative. Lastly, we test the portfolio implications of media connection index across different investors with various risk aversion levels. The empirical results show that the portfolios constructed based on media connection index obtain larger Sharpe ratios and certainty equivalent return gains than sentiment indices and combined economic predictors for mean-variance investors. Our paper has shed new light upon a different aspect of media s role in return predictability. In the past decades, the literature that investigates the media s role in financial markets mainly examines how the pessimism sentiment revealed from the content is associated with stock prices. Tetlock (2007) presents that the linguistic tone, especially negative tones, can predict market excess returns. Tetlock et al. (2008) further explore the cross-section predictability of returns by processing firm-specific news. Similarly, Zhang et al. (2016) document a sector specific reaction based on their distilled sentiment measure. Jegadeesh and Wu (2013) further improves Tetlock (2007) by using a term weighting method of content analysis based on OLS and Naïve Bayes, and they also find significant return predictability of news articles. Unlike these literature that focuses on extracting investors sentiment between the lines, our indices take into account the connected news and thus extract investor s sentiment spillovers and sentiment co-movement effects. These effects are shown to have powerful in-sample and out-of-sample predictabilities on market returns. We also contribute to the literature that studies investor attention. In Peng and Xiong (2006), they documented that investors tend to process more market information than firm-specific information due to limited attention, and thus generates important features in return co-movement. A follow-up work Peng et al. (2007) show that combining with limited attention and attention shifts, people can explain time-varying asset comovement. In terms of media attention, Odean (1999) and Barber and Odean (2007) found that individual investors are more likely to trade the stocks that have grabbed their attentions due to limited attention in searching what to trade, especially for buying stocks. Fang and Peress (2009) and Fang et al. (2014) further examined the cross-sectional return predictability and mutual funds trading and performances using media coverage as proxy of attention-grabbing events, and they also find evidence that both individual and institutional investors subject to limited attention. Different from these papers, we 3

4 implicitly impose limited attention by assuming first-degree structure of media network. Moreover, we also find strong evidence of sentiment co-movement which is induces by investor additional attention or attention distraction. These results complement to the investor attention literatures and open up the potential for new empirical designs. Lastly, we contribute to the literature on application of network analysis in financial studies. Cohen and Frazzini (2008) and Menzly and Ozbas (2010) find that economic links among certain individual firms and industries contribute to cross-firm and cross-industry return predictability. They interpret their results as evidence of gradual information diffusion across economically connected firms, in line with the theoretical model of Hong et al. (2007). Rapach et al. (2015) investigate the predictability of industry returns based on a wide array of industry interdependencies. Most recent, Härdle et al. (2016) propose a new method, tail-event driven network risk, to detect risk network. Based on this measure, they provide direct evidence on tail event interdependencies of financial institutions. Some follow-up empirical studies include Chen et al. (2017), Chen et al. (2016) and Xu et al. (2016). Most related, Scherbina and Schlusche (2015) have provide empirical evidence that media news reveals additional linkages among individual stocks. They find the lagged return of stocks in the linked group according to media news can predict subsequent return of other stocks within the same group. Different from above literature, we are the first paper to construct the market-wide media network and provide direct evidence on its market return predictability. The rest of the paper is organized as follows. In section 1, we review the literature exploring media connections and media network in financial markets and make some essential assumptions for subsequent analysis. In section 2, we show how to compose an aggregate measure of media connection which can overcome the deficit of other measures and describe our data sources. Then, we conduct some empirical tests and present our results in section 3. In section 4, we verify the sentiment co-movement channel and other possible channels of return predictability. Lastly, we conclude in section 5. 1 Media Connection and Media Network In this section, we review the literatures that study the impact of the media connections and media networks on financial and economic matters, and introduce several reasonable assumptions for constructing the new predictors of stock returns. Media connection, by definition, is an inter-relationship that is built via news stories which may through explicit mentions or implicit affections. The explicit mentions, also known as media co-occurrence, is the most natural way of formulating the connectivity 4

5 of two entities. Özgür et al. (2008) first studied the social network inferred from the co-occurrence network of Reuters news. They show that the network exhibits small-world features with power law degree distribution and it provides a better prediction of the ranking on importance of people involved in the news comparing to other algorithms. Scherbina and Schlusche (2015) studied the cross-predictability of stock returns by identifying the economic linkage from co-mentions in the news story. They constructed a linkage signal using the weighted average of the connected stock returns and they find that the linked stocks cross-predict one anothers returns in the future significantly, and the predictability increases with the number of the connected news. Apart from the explicit mentions, the connection may also be built through implicit affections. One of the most popular channels is the industrial chain. As shown in Cohen and Frazzini (2008), economic links among certain individual firms and industries contribute significantly to cross-firm and cross-industry return predictability. Rapach et al. (2015) extends the perspective of Cohen and Frazzini (2008) by defining a connection between industries with the predictability of returns. Through these industrial interdependencies, the news that conveys information on one industry will also percolates into the other industries. Further, due to the competitive relation of stocks within the industry, the good (bad) news to one stock will be bad (good) news to its competitors. In addition, business interaction is another important channel that transfers news information from one firm to another. Based on media connections, we can formulate a media network by taking the whole picture of the connected stocks as a connected graph with news tones or connectivity tagged on each stocks. In network analysis context, all these information can be captured by the adjacency matrix or weighted adjacency matrix. Apart from adjacency matrix, we also need to make some essential and reasonable assumptions on news arrival and network structures in advance to simplify our analysis. The assumptions are as follows: Assumption 1 (Random News Arrival). Connected news arrives randomly and investors have no prior information on the distribution of news arrival. Assumption 2 (First-degree Network). The information that the connected news conveys only affects the directly connected stocks. Assumption 3 (Majority Opinion). The aggregated news tones reflect the majorities opinions on future prices of both connected stocks. The first assumption is quite common and reasonable as investors face two tiers of randomness. The first tier randomness comes from the arrival of firm-specific news event and the second tier comes from the news connections. In reality, a news event is always 5

6 unpredictable, and even though investors realize a news event will occur, the stocks that the news will mention are still mysterious to the investors. As a result, the investors who rely on connected news will only make short-term investment decisions in order to avoid exposure in uncertainties. The second assumption comes from two facts. The first fact is that investors attentions are always limited which restrict them to spend time on collecting information on second-degree connection. Another fact is that second-degree information is marginal to first-degree information as vast quantity of news guarantees that most of second-degree connections form by two piece of news will be summarized as first-degree connections in some news. This simplifies our analysis greatly as we only need to consider the static network graph and analyse the first-degree adjacency matrixin each period. The last assumption just ensures our indices constructed by aggregation will not be dominated by some extreme opinions. 2 Data and Methodology In this section, we first introduce the data sources and explain the methodology for constructing the media connection indices. By making a comparison with conventional measures, such as correlation coefficients and Euclidean distances, we show that our measure will not only provide us the media coverage and news sentiment, but also the potential of stock co-movement and sentiment interaction between investors. 2.1 Media-connection-based Predictors The data we use for identifying media connection is the firm-specific news from the Thomson Reuters News Archive dataset ranging from Jan-1996 to Dec The data contains various types of news, e.g. reviews, stories, analysis and reports etc., about markets, industries and corporations. In this paper, we identify the news that has mentioned at least two stocks as connected news and the others as self-connected news. This dichotomy allows us to isolate the effect of the media connection by calculating the connectivity measure with one stock as the centre of a news network, and the aggregation of connectivity measures over the whole portfolio will provide information on the whole news network. The media connection is identified though the connected news where at least two stocks are mentioned in the text. Beyond just connectivity, the Thomson Reuters news data also provide us the news tones (positive, negative and neutral) for each mentioned stocks within each article. With both connectivity and news tones available, we can compute 3 types of daily pairwise connection scores of the news to each stock mentioned 6

7 with different combinations of them, that is, CS k (1),i,j,t = T one k i,t Cnct k j,t, (2.1) CS k (2),i,j,t = T one k i,t T one k j,t, (2.2) CS k (3),i,j,t = Cnct k i,t Cnct k j,t, (2.3) with i, j = 1, 2,, N and k = 1, 2, K t, where N is the total number of stocks in the sample, the superscript k denotes the kth news in day t and K t is the total number of news of day t which may vary everyday, Cnct is the dummy variable indicating news mentions that comes from connection information matrix news 1 news 2 news Kt stock 1 Cnct 1 1,t Cnct 2 1,t Cnct Kt 1,t stock 2 Cnct 1 2,t Cnct 2 2,t Cnct Kt 2,t stock N Cnct 1 N,t Cnct 2 N,t Cnct Kt N,t and T one is news tone that comes from the media tones information matrix, news 1 news 2 news Kt stock 1 T one 1 1,t T one 2 1,t T one Kt 1,t stock 2 T one 1 2,t T one 2 2,t T one Kt 2,t stock N T one 1 N,t T one 2 N,t T one Kt N,t By construction, the connection scores have distinct economic meanings with each other. CS (1),i,j, accounts for the spillover effect of stock i s investors attention on stock j if they are connected. CS (2),i,j, captures the news-sentiment-induced co-movement effect between stock i and j. It is positive when the sign of news tones of stocks are the same which indicates a possible co-movement of the stock prices, while it is negative when the sign of news tones of stocks are opposite which means price co-movement is less likely to happen. Further, the magnitude of CS (1) (CS (2) ) implies the strength of the connection which ranges from -1 to 1. With a higher magnitude, the connection of the two stocks will be tighter and a stronger sentiment spillover (a more synchronized movement) will occur. CS (3),i,j, only takes 0-1 value according to media-connectedness between stocks. Therefore, the aggragation of CS (3) will only capture the media coverage of stocks. To see the advantage of our co-movement measure over other interrelation measures,. 7

8 we consider the following example: suppose there are two connected news for two stocks, and the news tones for each stock is given by: [ news 1 news 2 stock stock ]. where a positive number indicates a positive news tone for this stock. By cross-sectional correlation, we will have the correlation coefficient between the stocks is -1. However, the news tones for these stocks are all positive in both news, which may indicates a positive co-movement relationship. Therefore, correlation coefficient is not a proper measure for describing the news-induced stock co-movement based on this simple example. As simple correlation fails in describing media connections, people may come up with Euclidean distance. However, distance is incapable of capturing the sentiment information in news tones, and the following example explains our point. [ news 1 news 2 stock stock ], [ news 3 news 4 stock stock ] For above cases, the Euclidean distance informs us that the distance between these two pairs of stocks are the same. However, we can deduce that the prices of stock 3 and 4 are likely to be co-moving while the prices of stock 1 and 2 may be affected oppositely. As a result, a proper connection score that can correctly reflect the relationship between connected stocks based on the news tones is needed, and we will show that our construction of connection scores will be able to retain the information given in the news completely in the next section. With these basic elements available, we construct the Media Connectivity Matrices on daily basis K t k=1 K t C (p),t = k=1 K t CS k (p),1,1,t CS k (p),2,1,t. CS(p),N,1,t k k=1 K t k=1 K t CS k (p),1,2,t CS(p),2,2,t k k=1 K t k=1.... CS k (p),n,2,t K t CS(p),1,N,t k k=1 K t CS(p),2,N,t k k=1. K t CS(p),N,N,t k k=1, p {1, 2, 3}. (2.4) By construction, the media connectivity matrices provide us with sentiment spillovers, co- 8

9 movement possibility and news coverage. Based on these Media Connectivity Matrices, we finally aggregate the network information to compose Media Connection Indices (MCI) on daily basis, MCI (p),t = N N K t CS(p),i,j,t k i=1 j i k=1 N N K t CS(p),i,j,t k i=1 j=1 k=1 N N K t CS(p),i,j,t 1 k i=1 j i k=1 N N K t CS(p),i,j,t 1 k i=1 j=1 k=1, p {1, 2, 3}. (2.5) The indices are formulated by taking differences of fractions between the sum of offdiagonal elements and the sum of every element. This formulation is helpful in isolating the information of media coverage from other information sets and eliminate the potential persistency in index series. Therefore, by construction, we now name MCI (1),t Mediaconnected Sentiment Index, MCI Media-connected Co-movement Index, and MCI (3),t Media-connected Coverage Index. [Insert Figure 1 here.] In Figure 1, we visualize the time series of media connection indices to gain a direct impression of their properties. Interestingly, we can observe that the media-connected sentiment index peaks at the Subprime crisis period while media-connected co-movement and coverage indices peak at the dot-com crisis period. Besides, these indices show complementary characteristics in variation, that is, when MCI (1) is not volatile, the other two indices will become more volatile and vice versa. These features indicate that mediaconnection sentiment index contains different information content with the other two types of media connection indices. Also, we may observe that media-connected co-movement index and media-connected coverage index seem to behave alike. This is because the co-movement only appears when stocks are media-connected. Therefore, MCI (2) can be viewed as a weighted MCI (3) with weights T one T one on each connected news, and thus we would expect MCI (2) will be a more informative measure than MCI (3). 2.2 Alternative Predictors Apart from the media news data, we calculate the investor sentiment index in Baker and Wurgler (2006) and the investor sentiment aligned index in Huang et al. (2014) for comparing with the sentiment content of the media connection indices. Moreover, we also calculate the average correlation index in Pollet and Wilson (2010) for comparing with the information content of the media connection indices. 9

10 For illustration purposes, we plot media-connected sentiment index and investor sentiment indices by Baker and Wurgler (2006) and Huang et al. (2014) in Figure 2. [Insert Figure 2 here.] From the figure we can observe that media-connected sentiment index behaves differently to other sentiment indices in general. Particularly, it is more volatile than sentiment indices in Subprime crisis but it fails to capture the high sentiment in dot-com crisis. This indicates that even though media-connected sentiment index captures some sentiment varitions, it mainly contains the news sentiment spillovers. Therefore we would expect our media-connected sentiment index provide some marginal information that not involved in sentiment indices. In addition, we also collect 14 economic predictors that are linked directly to economic fundamentals used in Goyal and Welch (2008) from Amit Goyal s website. Specifically, they are the log dividend-price ratio (D/P), log dividend yield (D/Y), log earnings-price ratio (E/P), log dividend payout ratio (D/E), stock return variance (SVAR), book-tomarket ratio (B/M), net equity expansion (NTIS), treasury bill rate (TBL), long-term bond yield (LTY), long-term bond return (LTR), term spread (TMS), default yield spread (DFY), default return spread (DFR) and inflation rates (INFL). The basic summary statistics of these predictors are reported in Table 1. [Insert Table 1 here.] From the summary statistics we can observe that the monthly excess market return has a mean of 0.82% and a standard deviation of 5.07%, implying a monthly Sharpe ratio of While the excess market return has little autocorrelation, most of economic predictors are highly persistent. The summary statistics are generally consistent with the literature. 3 Predicting Stock Market Returns with Media Connection In this section, we provide a number of empirical results. Section 3.1 examines the predictability of media network index and average correlation index on the aggregate market. Section 3.2 compares the media network index with economic predictors. Section 3.3 analyzes the out-of-sample predictability, and Section 3.4 assesses the economic value of predictability via asset allocation. 10

11 3.1 Forecasting the Market Consider the standard predictive regression model, R m t+1 = α + βmci t + ɛ t+1, (3.1) where R m t+1 is the excess market return, i.e., the monthly log return on the S&P500 index in excess of the risk-free rate. For comparison, we also run the same in-sample predictive regression with BW sentiment index, Sent BW t, and PLS sentiment index, Sent P t LS. Specifically, we test the null hypothesis H 0 : β = 0, which means media connections has no predictability of stock returns, against the alternative H 1 : β 0. Under the null hypothesis, (3.1) reduces to the constant expected return model, R m t+1 = α + ɛ t+1. [Insert Table 2 here.] Table 2 reports the results of in-sample predictive regressions. Panel A, B, and C provide the estimation results for the media connection indices. In all cases, the MCI calculated by optimism scores can predict negative returns significantly, and this predictability mostly come from the negative tones of media news. Moreover, the media-connected coverage index also predicts a negative return significantly. This result is consistent with Fang and Peress (2009) that intense media coverage induces a lower return. Panel D reports the in-sample predictability of investor sentiment indices. Consistent with Baker and Wurgler (2006) and Huang et al. (2014), both sentiment indices predict a negative return whereas they are not statistically significant unless we apply a one-sided test critical value. The last three columns report the overall R 2 and R 2 s in expansion and recession periods recorded by NBER. The results show that MCIs provide larger in-sample R 2 s than sentiment indices. Economically, the OLS coefficient suggests that a one-standard deviation increase in MCI opt is associated with an approximate 0.82% decrease in expected excess market return for the next month. On the one hand, recall that the average monthly excess market return during our sample period is 0.82%, thus the slope of 0.82% implies that the expected excess market return based on MCI opt varies by the same magnitude of its average level, which indicates a strong economic impact. On the other hand, if we annualize the 0.82% decrease in one month by the multiplication of 12, the annualized level of 9.84% is somewhat large. In this case, one may interpret this as the model implied expected change that may not be identical to the reasonable expected change of the investors in the market. Empirically, this level is comparable with conventional macroeconomic predictors. For example, a one-standard-deviation increase in the D/P 11

12 ratio, the CAY and the net payout ratio tends to increase the risk premium by 3.60%, 7.39%, and 10.2% per annum, respectively (see, e.g. Lettau and Ludvigson (2001) and Boudoukh et al. (2007)). Meanwhile, the R 2 s of MCI opt (1),t, MCIopt, MCI (3),t with OLS forecast are 1.44%, 2.55%, and 2.12% respectively, which are all substantially greater than 0.87% of Sent BW and 1.12% of Sent P LS. This implies that if this level of predictability can be sustained out-of-sample, it will be of substantial economic significance (Kandel and Stambaugh (1996)). Indeed, Campbell and Thompson (2008) show that, given the large unpredictable component inherent in the monthly market returns, a monthly out-of-sample R 2 of 0.5% can generate significant economic value and our findings in section 3.3 are consistent with this argument. To figure out the driving force of the predictability, we also calculate media connection indices based on negative and positive tones and re-estimate model (3.1) respectively. We find that the main prediction power of media connection indices are from negative tones whose t-statistics are and with R 2 s being 1.71% and 1.79% while the positive tone has no significant prediction power on future returns. This result is consistent with Tetlock (2007) who assert the negative tone is more informative than the positive tone, and our findings complete their argument in the media network perspective. Apart from just analyse the predictability over the whole period, it is also important to analyse the predictability during business-cycles to gain a better understanding about the fundamental driving forces. Following Rapach et al. (2010), we compute the R 2 statistics separately for economic expansions (R 2 up) and recessions (R 2 down ), R 2 c = 1 T t=1 1 {t T c} ɛ 2 t T t=1 1 {t T c} (Rt m R, c {up, down}, (3.2) m ) 2 where 1 {t Tup} (1 {t Tup}) is an indicator that takes a value of one when month t is in an NBER expansion (recession) period, i.e. T up (T down ), and zero otherwise; ɛ t is the fitted residual based on the in-sample estimates of the predictive regression model in (3.1); R m is the full-sample mean of R m t ; and T is the number of observations for the full sample. Note that, unlike the full-sample R 2 statistic, the Rup 2 (Rdown 2 ) have no sign restrictions. Columns 4 and 5 of Table 2 report the Rup 2 and Rdown 2 statistics. It is shown that MCIs gain a slightly higher return predictability over the expansions and a significantly higher one over recession periods than Sent BW, and they only gain a higher return predictability over the expansions while Sent P LS gains the higher over the recessions. These are supportive results to our theory as during the recessions, bad news exists more often and investors subjected to short-sale constraint fail to react to them which in the end leads to a negative 12

13 return in the following period. Summarizing Table 2, the media connection indices exhibit significant in-sample predictability for the monthly excess market return both statistically and economically, which is much stronger than the sentiment indices. In addition, media connection indices perform much better in the both expansion and recession periods, especially over the recessions. This suggests that media connection indeed reflect investors additional attention on connected stocks and news content has asymmetric effect on stock returns accordingly. 3.2 Comparison with Economic Predictors In this section, we compare the forecasting power of media connection indices with economic predictors and examine whether its forecasting power is driven by omitted economic variables related to business cycle fundamentals or investor sentiment. Specifically, we examine whether the forecasting power of MCIs remains significant after controlling for economic predictors and investor sentiment. To analyse the marginal forecasting power of MCIs, we conduct the following bivariate predictive regressions based on MCIs and Z t, R m t+1 = α + βmci q (p),t + φz t + ɛ t+1, p {1, 2, 3}, (3.3) where q {opt, neg, pos} and Z t is one of alternative predictors described in section 2.2, and our main interest is the coefficient β, and to test H 0 : β = 0 against H 1 : β 0. [Insert Table 3 here.] Table 3 shows that the estimates of β in (3.3) are negative and stable in magnitude, in line with the results of predictive regression (3.1) reported in Table 2. More importantly, β remains statistically significant when augmented by the economic predictors. These results demonstrate that MCIs contain sizeable complementary forecasting information beyond what is contained in the economic predictors and investor sentiment. Meanwhile, the coefficients of investor sentiment indices remain to be insignificant and of same magnitudes after controlling for media connection indices, suggesting that the information content of media-connection based predictors are not overlapping with investor sentiment predictors. 3.3 Out-of-sample Forecasts Despite that the in-sample analysis provides more efficient parameter estimates and thus more precise return forecasts by utilizing all available data, Goyal and Welch (2008), 13

14 among others, argue that out-of-sample tests seem more relevant for assessing genuine return predictability in real time and avoid the over-fitting issue. In addition, out-ofsample tests are much less affected by finite sample biases such as the Stambaugh bias (Busetti and Marcucci (2013)). predictive performance of media connection indices. Hence, it is essential to investigate the out-of-sample For out-of-sample forecasts at time t, we only use information available up to t to forecast stock returns at t + 1. Following Goyal and Welch (2008), Kelly and Pruitt (2013), and many others, we run the out-of-sample analysis by estimating the predictive regression model recursively based on different types of media connection indices, ˆR t+1 m = ˆα t + ˆβ t MCI (q) (p),1:t;t, p {1, 2, 3}, (3.4) where q {opt, pos, neg}, ˆα t and ˆβ t are the OLS estimates from regressing {R m r+1} t 1 r=1 with model (3.1) recursively. Like our in-sample analogues in Table 2, we consider different types of media connection indices based on optimism, positive and negative news tones respectively. For comparison purposes, we also carry out out-of-sample test with Sent BW t and Sent P LS t, and the results are reported in Panel B of Table 4. To evaluate the out-of-sample forecasting performance, we apply the widely used Campbell and Thompson (2008) ROS 2 statistics based on unconstrained forecast and truncated forecast that imposing non-negative equity premium constraint. The unconstrained R 2 OS statistic measures the proportional reduction in mean squared forecast error (MSFE) for the predictive regression forecast relative to the historical average benchmark. Goyal and Welch (2008) show that the historical average is a very stringent out-of-sample benchmark, and individual economic variables typically fail to outperform the historical average. To compute ROS 2, let r be a fixed number chosen for the initial sample training, so that the future expected return can be estimated at time t = r + 1, r + 2,..., T. Then, we compute s = T r out-of-sample forecasts: { ˆR t+1} m T t=r 1. More specifically, we use the data over 1996:01 to 2003:12 as the initial estimation period so that the forecast evaluation period spans over 2004:01 to 2014:12. ˆR 2 OS = 1 T 1 t=r (Rm t+1 ˆR m t+1) 2 T 1 t=r (Rm t+1 R m t+1) 2, (3.5) where R m t+1 denotes the historical average benchmark corresponding to the constant ex- 14

15 pected return model (R m t+1 = α + ɛ t+1 ), i.e. R m t+1 = 1 t t Rs m. (3.6) s=1 By construction, the ROS 2 statistic lies in the range (, 1]. If R2 OS > 0, it means that the forecast ˆR t+1 m outperforms the historical average Rt+1 m in terms of MSFE. The out-of-sample R 2 based on truncated forecast, R OS 2, is calculated by simply truncate the unconstrained OLS forecast by imposing non-negative equity premium assumption, i.e. R m t+1 = max{0, ˆR m t+1}, (3.7) where ˆR m t+1 is the out-of-sample forecast by unconstrained OLS, and the out-of-sample R 2 now becomes R 2 OS = 1 T 1 t=r (Rm t+1 R m t+1) 2 T 1 t=r (Rm t+1 R m t+1) 2. (3.8) By imposing non-negative equity premium constaint, Campbell and Thompson (2008) show that the out-of-sample forecast will impove and become more statistically significant. The statistical significance of the out-of-sample R 2 s we report is based on MSFEadjusted statistic of Clark and West (2007) (CW-test hereafter). It tests the null hypothesis that the historical average MSFE is not greater than the predictive regression forecast MSFE against the one-sided (right-tail) alternative hypothesis that the historical average MSFE is greater than the predictive regression forecast MSFE, corresponding to H 0 : ROS 2 0 against H 1 : ROS 2 > 0. Clark and West (2007) show that the test has a standard normal limiting distribution when comparing forecasts from the nested models. Intuitively, under the null hypothesis that the constant expected return model generates the data, the predictive regression model produces a noisier forecast than the historical average benchmark as it estimates slope parameters with zero population values. We thus expect the benchmark models MSFE to be smaller than the predictive regression model s MSFE under the null. The MSFE-adjusted statistic accounts for the negative expected difference between the historical average MSFE and predictive regression MSFE under the null, so that it can reject the null even if the R 2 OS statistic is negative. [Insert Table 4 here.] Panel A, B and C of Table 4 show that MCI indices in general generate positive and significant ROS 2 statistics and thus delivers a lower MSFE than the historical average. Further, consistent with Campbell and Thompson (2008), the out-of-sample predictability 15

16 of MCIs improves by truncated regression. Thus, it is safe to conclude that MCIs have strong out-of-sample predictive ability for market returns, which confirms our conjectures in previous in-sample results (Table 2). Comparing with MCIs, Sent BW exhibits much weaker out-of-sample predictive ability for market excess returns as shown in Panel D. Its ROS 2 is negative and insignificant in general with exception in expansion periods. Interestingly, the PLS sentiment presents very good out-of-sample return predictability in all cases. This result once again show that the sentiment aligned approach indeed isolates the true factors from the noises for predicting market as explained in Huang et al. (2014). Despite Sent P LS showing strong predicting power, our media connection indices (MCI opt and MCI (3),t) still outperform it in general. It proves that our media connection indices are indeed powerful predictors for market returns. In addition, the last two columns of Table 4 show that, even though the predictability of media connection indices are mainly concentrated in recessions, MCI neg (1),t presents strong out-of-sample forecasting ability during both expansions and recessions, which also support Tetlock et al. (2008) that negative tone is more informative than positive tone of media news. Besides, MCI (3),t also perform well in both expansions and recessions, which confirms media coverage in being an important channel for additional attention to take effect on stock returns. [Insert Figure 3 here.] Since MCI is constructed with tones of connected-news, its predictability may partially come from the investors sentiment. To understand differences in forecasting power between sentiment indices and media-connected co-movement index (MCI ), Figure 3 depicts the predicted returns based on Sent BW t, Sent P t LS and MCI opt for the 2004: :12 out-of-sample period. It is clear that the MCI opt predicted returns are much more volatile than the forecasts of sentiment indices. As the actual realized excess returns (plotted in the figure as 6-month moving average for better visibility) are even more volatile than the MCI opt predicted returns. This explains why the connectednews-based index does a better job than the hard-information-based sentiment measures in capturing the expected variation in the market return. [Insert Figure 4 here.] Following Goyal and Welch (2008) and Rapach et al. (2010), Figure 4 presents the time-series plots of the differences between cumulative squared forecast error (CSFE) for the historical average benchmark forecasts and the CSFE for predictive regression fore- 16

17 casts based on MCI opt and sentiment indices over 2004: :12. This time-series plot is an informative graphical device on the consistency of out-of-sample forecasting performance over time. When the difference in CSFE increases, the model forecast outperforms the historical average, while the opposite holds when the curve decreases. The solid blue line in Figure 4 shows that our media-connected co-movement index, MCI opt consistently outperforms the historical average except for the first few periods. The curve has slopes that are predominantly positive, indicating that the good out-of-sample performance of MCI opt steps from the whole sample period rather than some special episodes. The figure also graphically illustrates the performances over the NBER dated business cycles, complementary to Table 4. For comparison, we also plot the differences in CSFE of investor sentiment indices in dashed lines. The dashed red line shows that Sent BW fails to consistently outperform the historical average. As a consequence, it does a poor job in terms of monthly out-of-sample forecasts. The Sent P LS, which is depicted by dashed yellow line, however is shown to perform better than Sent BW, it is still not as good as media-connected co-movement index. These results suggest that our media connection indices contain some useful information in predicting market returns that investor sentiment indices are fail to capture. Lastly, we compare the out-of-sample performance of media connection indices with the combined economic predictors proposed in Rapach et al. (2010). From Panel E of Table 4 we can conclude that the out-of-sample predictability of the combined economic predictors during our sample period is very poor in general except for the expansion periods. This result implies that the out-of-sample predictability of our media connection indices does not come from the hard information either. In summary, out-of-sample analysis shows that media connection indices are powerful and reliable predictors for the excess market returns, and consistently outperforms investor sentiment indices and combined economic predictors across different sample periods which is consistent with our previous in-sample results (Tables 2 and 3). 3.4 Asset Allocation Implications Now we examine the economic value of stock market forecasts based on the media connection indices. Following Kandel and Stambaugh (1996), Campbell and Thompson (2008) and Ferreira and Santa-Clara (2011), among others, we compute the certainty equivalent return (CER) gain and Sharpe ratio for a mean-variance investor who optimally allocates across equities and the risk-free asset using the out-of-sample predictive regression forecasts. 17

18 At the end of period t, the investor optimally allocates w t = 1 γ ˆR t+1 ms ˆσ t+1 2, (3.9) of the portfolio to equities during period t + 1, where γ is the relative risk aversion ms coefficient of CRRA utility, ˆR t+1 is the out-of-sample forecast of the simple excess market return, and ˆσ t+1 2 is the variance forecast. The investor then allocates 1 w t of the portfolio to risk-free assets, and the t + 1 realized portfolio return is R p t+1 = w t R ms t+1 + R f t+1, (3.10) where R f t+1 is the gross risk-free return. Following Campbell and Thompson (2008), we assume that the investor uses a eight-year moving window of past monthly returns to estimate the variance of the excess market return and constrains w t to lie in between 0 and 1.5 to exclude short sales and to allow for at most 50% leverage. The CER of the portfolio is CER = û p 0.5γˆσ 2 p, (3.11) where û n and ˆσ 2 n are the sample mean and variance for the investors portfolio over the q forecasting evaluation periods respectively. The CER gain is the difference between the CER for the investor who uses a predictive regression forecast of market return generated by (3.4) and the CER for an investor who uses the historical average forecast. We multiply this difference by 12 so that it can be interpreted as the annual portfolio management fee that an investor would be willing to pay to have access to the predictive regression forecast instead of the historical average forecast. To examine the effect of risk aversion, we consider portfolio rules based on risk aversion coefficients of 1, 3 and 5, respectively. In addition, the transaction costs considered here is 50bps which is generally considered as a relatively high number. For assessing the statistical significance, we follow DeMiguel et al. (2009) by testing whether the CER gain is indistinguishable from zero by applying the standard asymptotic theory as in their paper. In addition, we also calculate the monthly Sharpe ratio of the portfolio, which is the mean portfolio return in excess of the risk-free rate divided by the standard deviation of the excess portfolio return. Following again DeMiguel et al. (2009), we use the approach of Jobson and Korkie (1981) corrected by Memmel (2003) to test whether the Sharpe ratio of the portfolio strategy based on predictive regression is statistically indifferent from that of the portfolio strategy based on historical average. 18

19 [Insert Table 5 here.] Table 5 shows that, of all media connection indices, MCI opt stands out in term of the economic value, which is well above all the other predictors, and is of economic significance. The net-of-transaction-cost CER gains for MCI opt across the risk aversions are consistently positive and economically large, ranging from 1.13% to 3.59%. More specifically, an investor with a risk aversion of 1, 3, or 5 would be willing to pay an annual portfolio management fee up to 3.59%, 2.38%, and 1.13%, respectively, to have access to the predictive regression forecast based on MCI opt instead of using the historical average forecast. The Sharpe ratio of portfolios formed based on MCI opt is , which is the highest among all the portfolios based on other predictors and 25% more than the market Sharpe ratio, 0.16, with a buy-and-hold strategy. The alternative predictors that can match the performance of media connection indices are PLS sentiment index whose CER gains range from 0.70% to 4.01% and the median combined economic predictor whose CER gains vary from 0.89% to 3.33%. However their CER gains and Sharpe ratio are close to media connection indices, their results are not as statistically significant as media connection indices given investors high risk aversion. By Figure 1 and Figure 2, we can find both MCI (2) and Sent P LS peak at the dotcom crisis period while other predictors are less responsive to this period. This explains the outperformance of the portfolios based on these two predictors as investors have better predictability during the crisis period and thus have a better chance of avoiding unexpected loss in wealth. Overall, Table 5 demonstrates that the media connection indices, especially MCI opt, can generate sizable economic value for a mean-variance investor with high risk aversion, while sentiment indices and combined economic predictors can only generate significant CER gains under low risk aversion. specifications and the same level of transaction cost. The results are robust to common risk aversion 4 Source of Return Predictability According to the construction of media connection indices, several types of information are automatically involved, e.g. soft information and investor sentiment reflected by new tones, which enlightens us to provide three potential sources for predictability. Firstly, high sentiment co-movement may generate a negative return because of asymmetric sentiment effect on investors who are subjected to short-sale constraint. Secondly, the negative forecast may purely driven by investor sentiment. Lastly, the negative predicted returns is 19

20 a result of delayed information diffusion caused by media network. All these conjectures need to be proved or disproved through empirical tests. In this section, we will explore through which channel do media connection indices take effects in predicting market returns. For this purpose, we first check the predictability of MCIs on cross-sectional portfolios. Then, we test if MCIs can predict sentiment indices or average correlation index in Pollet and Wilson (2010) to see if their predictabilities are sentiment driven or explained by delayed information diffusion. 4.1 Sentiment Co-movement Channel In terms of sentiment co-movement, we proxy it by news tone co-movement which is modelled by MCI opt. By classic behaviour theory, investors receiving positive news on a stock would like to increase their positions in this stock and vice versa. In our case, the investors whose stock holdings are connected by media news will also react to the tones of connected stock as the connection grabs the attention of the connected investors. However, due to short-sale constraint, the connected investors can only react to good news by increasing their positions in connected stock while cannot short-sale the connected stocks when bad news occurs. This asymmetry caused by short-sale will end up with an overpricing in connected stocks and thus lead to a negative future return. On the top of this logic, we therefore sort the stocks on different levels of media connection indices respectively. In general, we expect low media connection deliver high portfolio returns. Figure 5 presents the returns of sorted portfolios with respect to different MCIs. [Insert Figure 5 here.] As expected, all three media connection indices exhibit good cross-sectional return predictability in general, especially the media-connected co-movement index which has highest returns. In Figure 5a, we can see that even though low media-connected sentiment predict high return significantly, the median and high group are convoluted. This reveals that the pure sentiment spillovers does not contribute to the return predictability significantly. In contrast, Figure 5b and 5c show that sentiment co-movement channel truly exists as MCI opt sorted portfolios gain highest returns and MCI (3),t sorted portfolios obtain largest low-minus-high return spread. These results confirm intensed connected-news coverage help enhance attention distraction and thus consolidate sentiment co-movement channel. 20