Social Ties and Predictable Returns*

Lin Peng Muhammed Yönaç Dexin Zhou

This Draft: February 22, 2021

Abstract

We propose a new measure of cross-firm linkages that are based on the social connect-edness between regions of firms’ locations and show that the linkages are positivelyassociated with comovements in firm value. Returns of socially-connected industrypeers strongly predict the future returns of focal firms, with a long-short portfolio gen-erating a monthly value-weighted alpha of 71 basis points. The result is stronger forsmall firms and firms with low institutional ownership or analyst coverage. Further,those peer firms’ returns strongly predict focal firms’ future fundamentals, earningssurprises, and analysts’ forecast errors. The findings suggest that social connection-based linkages contain important information that is not fully incorporated into pricesor explained by existing factors.

JEL Codes: G14, D85Keywords: social networks, slow information diffusion, return predictability

Preliminary, please do not circulate

*Lin Peng, Zicklin School of Business, Baruch College, One Bernard Baruch Way, Box 10-225, NewYork, NY 10010. Phone: (646) 312-3491, Email: lin.peng@baruch.cuny.edu. Muhammed Yönaç, Zick-lin School of Business, Baruch College, One Bernard Baruch Way, Box 10-225, New York, NY 10010.Email: muhammed.yonac@baruch.cuny.edu. Dexin Zhou, Zicklin School of Business, Baruch College, OneBernard Baruch Way, Box 10-225, New York, NY 10010. Email: Dexin.zhou@baruch.cuny.edu. We thankseminar participants at Baruch College, CUNY. Lin Peng acknowledges the Krell research grant and theWassarman research grant for financial support. All errors remain our own.

1 IntroductionThere has been extensive theoretical and empirical literature studying how investor atten-tion affects information processing, suggesting that limited attention can lead to marketunderreactions to value-relevant news.1 The underreaction-based mechanism explainspredictable returns in firms connected by shared economic links such as industry, loca-tion, and analyst coverage (e.g., Moskowitz and Grinblatt, 1999; DellaVigna and Pollet,2007; Cohen and Frazzini, 2008; Ali and Hirshleifer, 2020). More recently, a growing liter-ature highlights the interconnected nature of firms in the economy (e.g., Acemoglu et al.,2012; Bailey et al., 2018a) and shows that economic exchanges between regions are pos-itively associated with the intensity of social connections (e.g., Breschi and Lenzi, 2016;Cohen et al., 2017; Bailey et al., 2020).

Motivated by the recent studies, in this paper, we explore a new dimension of cross-firm linkages that are based on social connections between regions in which the firms arelocated. We posit that the fundamentals of firms located in regions with strong socialconnections also tend to comove more strongly. If investors are inattentive to this fun-damental linkage, then asset prices would not be able to incorporate such information ina timely fashion. As information diffuses slowly across the population of investors, thereturns of firms located in socially-connected regions can help predict the future stockreturns and performances of related firms.

To test this hypothesis, we construct a novel measure of socially-connected peer firmreturns (social-peer firm return, for short) using the Facebook-based Social ConnectednessIndex (SCI) (Bailey et al., 2018a). We empirically study whether a focal firm’s return iscorrelated with industry peers located in regions that share strong social connections withthe firm’s locations and the extent to which this information is incorporated into prices byexamining the ability of social-peer returns in predicting the focal firm’s future returns.

Our measure of social ties (SCI measure) is based on friendship links on Facebook, theworld’s largest online social networking service. More specifically, the SCI between a pairof counties is the number of friendship links between users from the two counties relativeto the product of the counties’ populations and captures social connectedness between

1See, for example, Hong and Stein (1999), Huberman and Regev (2001), Hirshleifer and Teoh (2003), Houand Moskowitz (2005), Peng and Xiong (2006), Hirshleifer et al. (2009), DellaVigna and Pollet (2009).

1

the counties. Given Facebook’s scale (with over 243 million active users in the U.S. as of2018) and the relative representativeness of its user body, SCI provides a comprehensivemeasure of the geographic structure of the U.S. social networks.2 Since firms that belongto the same industry are more likely to share common fundamentals, we define the social-peer return of a focal firm as the SCI-weighted returns of all firms that belong to the focalfirm’s industry. By construction, this measure maximizes the information from industrypeers from socially connected regions.

Using this social-peer firm return measure, we first show that a focal firm’s returnis more highly correlated with industry peers located in counties that have a high socialconnectedness with the firm’s locations. Controlling for industry returns, a one standarddeviation of increase in the returns of social peer firms is associated with 65 basis pointsincrease in focal firm returns. This suggests that social connectedness is useful in cap-turing the fundamental economic linkages across firms independent of overall industryperformances.

We next explore whether the social connections-based firm linkages are fully incorpo-rated into prices or if returns of connected firms can be explored to gain superior perfor-mances. Toward this goal, we construct a portfolio by taking a long position in firms withhigh social-peer returns and short position in firms with low social-peer returns. We showthat this long-short portfolio generates a Fama and French (2015) five-factor-adjusted re-turns of 155 basis points (equal-weighted) and 71 basis points (value-weighted) in thefollowing month.

This result is robust when we examine two-way sorted portfolios and conduct Famaand MacBeth (1973) regression analysis controlling for firm characteristics that are knownto predict returns.3 In particular, we show that our results are not driven by industrymomentum (e.g., Moskowitz and Grinblatt, 1999; Hoberg and Phillips, 2018), geographiclead lag effect (e.g., Parsons et al., 2020), customer-supplier lead lag effect (e.g., Cohenand Frazzini, 2008), and shared analyst coverage (Ali and Hirshleifer, 2020). Even afteraccounting for these lead-lag relationship, an increase in social-peer firm returns from the10% to the 90% level predicts a return increase of 42 basis points for the focal firms in the

2See, Bailey et al. (2018a), Kuchler et al. (2020a), Kuchler et al. (2020b) for further discussion on these points.3In all of our cross-sectional regression specifications, we include lag one-month return, size, book-to-market, momentum, illiquidity, idiosyncratic volatility, skewness, and co-skewness.

2

next month.To gain insight into the potential channels that may explain our findings, we explore

how firm characteristics influence the strength of the predictability. If the documentedreturn predictability is due to investor inattention to value-relevant information fromsocially connected peer firm returns, we should expect a stronger predictability for fo-cal firms that are less visible to the investors. Consistent with this, we find that returnpredictability results are stronger among smaller firms, firms with low institutional own-ership, and firms with low analyst coverage, precisely those firms that receive little in-vestors attention.

We also investigate whether the return predictability may be attributable to investors’overreaction to social-peer firm returns by investigating long-term return predictability.We find no evidence of return reversal following the most recent month. Instead, there isa moderate return continuation for up to five years. The lack of return reversal is inconsis-tent with the overreaction explanation and further confirms that the return predictabilityis due to investors’ delayed incorporation of value-relevant information from social-peerfirms.

We further analyze the degree that our results are driven by the ability of social-peerfirm returns to predict the future fundamentals of the focal firm. Using panel regres-sions, we find that high social-peer firm returns are positively associated with cumula-tive measures of firm performances in the long run. Specifically, all else equal, an increasein social-peer firm returns from the 10% to the 90% values increases focal firm cumula-tive gross profitability over assets by 82.50%, EBITDA over assets by 10.65%, and assetturnover by 179.81% in the next five years. Economically, the increases are equivalent to44%, 14% and 29% of the mean of the corresponding variables.

Finally, we show that social-peer firm returns positively predicts focal firms’ futureearnings surprises. Economically, a 10%-to-90% change in social-peer firm returns is as-sociated with a 1.91% increase in the cumulative earnings surprises of the focal firms forthe next five years, or 13% of the standard deviation of the cumulative earnings surprises.Similarly, the same increase in social-peer firm returns predicts that sell-side analysts sub-stantially underestimate the focal firm earnings by 48 basis points in the next year, or 25%of the standard deviation of the one-year cumulative forecast error. Therefore, the resultsuggests that financial analysts are sluggish in adjusting their forecasts to incorporate

3

relevant information from socially connected firms.Taken together, we present a new and robust lead-lag relationship based on social

ties between firms’ headquarters locations. These findings suggest that social-peer firms’returns contain important information regarding the fundamentals of focal firms in a waythat is not fully incorporated into prices or explained by existing factors. We thereforeconclude that social connectedness of firms’ locations provides a new way to identifyimportant economic linkages between firms.

Our paper is closely related to recent studies suggesting that firms share similar loca-tions or business footprints tend to have correlated performances or returns (e.g., Pirin-sky and Wang, 2006; Parsons et al., 2020; Jin and Li, 2020). Other papers have shown thatfirms’ locations are associated with the firms’ fundamentals such as growths and prof-itability (Dougal et al., 2015; Huang, 2015; Dougal et al., 2018; Smajlbegovic, 2019), stockreturns (Korniotis and Kumar, 2013; Tuzel and Zhang, 2017), and borrowing cost (Hasanet al., 2017). Also using Facebook data, Kuchler et al. (2020a) find that an area’s socialproximity to capital is positively related to the valuation and stock liquidity of firms lo-cated in that area. Our unique insight is that social connectedness across regions capturesa substantial component of economic ties of firms that are not explained by the local so-cioeconomic conditions or physical proximity.

We also contribute to the understanding of how networks shape economic relation-ships. In the social finance literature, the evidence often indicates that the primary roleof social networks are to increase recognition or transmission of information (e.g., Cohenet al., 2008; Hwang and Kim, 2009; Hochberg et al., 2007; Pool et al., 2012, 2015; Heimer,2016; Kuchler et al., 2020a). While previous research documents that firms can be con-nected through customer-supplier relationship or co-location (e.g., Cohen and Frazzini,2008; Pirinsky and Wang, 2006; Parsons et al., 2020), our results suggest that the the effectof social connections can also be important, in a way that is neither localized nor limitedto customer-supplier relations.4

Finally, this study adds to the extensive literature that studies the return predictabilityin financial markets (see McLean and Pontiff, 2016; Harvey, 2017, for recent meta stud-ies). More specifically, several papers document lead-lag patterns that are difficult to

4Another strand of literature studies how shocks are propagated through within-firm network (e.g., Giroudand Mueller, 2019).

4

explain using efficient market hypotheses (e.g., Jegadeesh and Titman, 1993; Moskowitzand Grinblatt, 1999; Hou, 2007; Hoberg and Phillips, 2018; Lee et al., 2019; Parsons et al.,2020; Jin and Li, 2020). Consistent with these papers, our results suggest that investorsand financial analysts fail to fully identify the economic links between firms.

The rest of the paper is organized as follows. In section 2, we provide a detailed de-scription of the data used in this study. In section 3, we discuss the return predictabilityresults. In section 4, we analyze whether social-peer firm returns capture fundamentalinformation about the focal firm and whether sell-side analysts fully incorporate this in-formation. In section 5, we conclude our paper.

2 Data

2.1 Social Connectedness

We measure social connections between regions with the Social Connectedness Index, whichis first proposed in Bailey et al. (2018a) and is based on Facebook friendship links betweenany two given U.S. counties as of April 2016. Facebook had more than 2.1 billion monthlyactive users globally and 239 million active users in the U.S. and Canada as of 2017 andcovers 68% of the adult population and 79% of online adults in the U.S (Duggan et al.,2016). The same survey also reveals that Facebook usage rates among U.S.-based onlineadults were relatively constant across various demographics and locations. Since individ-uals in the U.S. to connect with real-world friends, SCI based on Facebook links is likelyto be a relatively precise proxy for the social relationships across different regions in theU.S.5

More specifically, the SCI between two U.S. counties, is defined as:

Social Connectedness Index =Friendship Linksi,j

Populationi × Populationj,

where Friendship Links is the number of friendship links between county i and j on Face-book and Population denotes the county-level population in headquarters counties. Thismeasure represents a relative friendship probability between these two counties. The SCI

5Bailey et al. (2018a,b, 2019a,b,c, 2020); Kuchler et al. (2020b,a), and Rehbein et al. (2020) provide evidencethat friendships observed on Facebook are a good proxy for real-world connections.

5

variable is measured as of 2016. Bailey et al. (2018a) find that social connections acrossregions are persistent and do not change dramatically over time. Thus, information basedon the 2016 measure is a reasonable proxy for historical social connectedness.

2.2 Firm-Level Data

Our sample includes all domestic (U.S.) common stocks (i.e., share code 10 or 11) in CRSPuniverse and covers the period between January 1962 to December 2019. The accountingvariables, including earnings, are obtained from the CRSP-Compustat merged database.Analyst earnings forecasts and institutional ownership data are from I/B/E/S and Thom-son Reuters institutional (13F) holdings database respectively. Customer-supplier firmlinks are obtained through the Linking Suite by WRDS. Firm headquarters data are gath-ered from the augmented 10-X header data file on the Notre Dame Software Repositoryfor Accounting and Finance (SRAF) maintained by Bill McDonald; the file covers the pe-riod between 1994 and 2018. We further augment this data by parsing the 2019 10-Q and10-K filings on SEC’s Edgar database.6 The Economic Area-ZIP code link file is obtainedfrom Riccardo Sabbatucci’s website.7 Data on Fama-French five factors and Fama-French48-industry classification are obtained from Kenneth French’s data library.8

Unless otherwise stated, all variables are measured as of the end of the portfolio for-mation month (i.e., month t). We require a minimum of 24 monthly observations forvariables created using monthly data and 15 daily observations for those created usingdaily data. We exclude stocks with a price lower than $5 and higher that $1000 at the endof month t from our analyses to ensure that our results are not affected by illiquidity.

2.3 Social-Peer Firm Return

Our main variable of interest is the social peer firm return (SPFRET), defined as the social-connectedness-weighted average return of stocks that are from the same industry as the

6For firms that exists in our sample prior to 1994, we assume a headquarter location that is identical to thefirst available headquarter location in SEC filings if the firm also exists after 1994. This back-filling proce-dure is consistent with Parsons et al. (2020). They have argued that although this potentially introduces ameasurement error into our analyses through mismatching firms and headquarters prior to 1994, it wouldonly create noise and dampen the effects that we aim to estimate and thus work against us. To further en-sure the robustness of our results, we check the robustness of our results by excluding observations before1996, when Edgar filing becomes mandatory. We find that our results are robust.

7See https://sites.google.com/site/riccardosabbatucci/Research?authuser=08See https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html

6

focal stock. We only include firms from a different state to mitigate the effect of geo-graphic proximity on our measures.

Formally, for the focal stock i at the end of month t, SPFRET is given by

SPFRETi,t ≡∑j∈Ii,j/∈Si

SCIci,cjRETj,t

∑j∈Ii,j/∈SiSCIci,cj

,

where Ii and Si are the industry and state of stock i respectively, while ci and cj are therespective counties of stock i and stock j.9 SCIci,cj is the social connectedness index (SCI)between counties ci and cj. The compound SPFRET between t− 11 and t− 1 is denotedby SPFMOM. We consider both SPFRET and SPFMOM in our return prediction tests.

We control for a number of additional predictive variables in our regressions. Theirdefinitions are as follows: INDRET is the month-t equal-weighted average return ofstocks with the same Fama-French 48 (FF48) industry classification as the focal stock, IN-DMOM is obtained by compounding INDRET between month t− 11 and t− 1, GEORETis the month t equal-weighted average return of all stocks from the same Economic Area(EA) as the focal stock but from a different FF48 industry (Parsons et al., 2020).10 GEO-MOM is obtained by compounding GEORET from month t − 11 to t − 1. CFRET is themonth t weighted average return of stocks that share at least one analyst with the focalstock over previous 12 months, where weights are the number of shared analysts betweenstocks; CFMOM is computed the same way, except that month-t stock return is replacedby the compound stock return between t− 11 and t− 1 (Ali and Hirshleifer, 2020). CRETis the equal-weighted average stock return of the main customers of the focal firm, wherea 6-month gap is required between the fiscal year-end of the supplier and stock returns toensure that the firm-customer links are known before the returns they are used to explain(Cohen and Frazzini, 2008).

Besides the predictors described above, we also use a battery of other focal-stock-levelcontrol variables shown to predict returns in literature, which are measured as of the endof month t unless otherwise stated. Their definitions are as follows: RET is the monthly

9Although we allow the industry and headquarters location of a stock to change in time, time subscript thas been omitted from I, S and c for notational clarity.

10EAs are defined by the Bureau of Economic Analysis. They are intended to capture local nodes of eco-nomic activity and typically involve a main metropolitan area, along with smaller nearby regions fromwhich workers commute Parsons et al. (2020).

7

stock return. Following Shumway (1997), we adjust stock returns for delisting in order toavoid survivorship bias. 11 SIZE is the logarithm of the market capitalization (in milliondollars) as measured at the end of previous June. BMKT is the CAPM beta computedusing a 60-month window with a minimum window of 24 months. BM is the book-to-market ratio, whose value between the end of June of year T and May of year T + 1 iscomputed as the book value of stockholdersâ equity, plus deferred taxes and investmenttax credit (if available), minus the book value of the preferred stock for the last fiscalyear ending in T − 1, scaled by the market value of equity at end of December of T − 1.Depending on availability, the redemption, liquidation, or par value (in that order) is usedto estimate the book value of the preferred stock. MOM is obtained by compounding RETbetween month t− 11 and t− 1. IVOL is the monthly idiosyncratic volatility, computedas the standard deviation of the daily residuals obtained by regressing the daily excessstock returns on the daily market excess return, small-minus-big (SMB) and high-minus-low (HML) factors over the previous month. ILLIQ is the Amihud’s illiquidity (Amihud,2002), defined as the average daily ratio of the absolute stock return to the dollar tradingvolume within the previous month. MAX is the maximum daily stock return realizedover the previous month. SKEW is the sample skewness of the daily stock returns fromthe previous month. COSKEW is the stock’s monthly co-skewness, defined as

COSKEWi,t =E[εi,tR2

m,t]√

E[ε2

i,t

]E[

R2m,t

] ,

where Ri,t and Rm,t are the excess returns of stock i and market at the end of month trespectively, and εi,t is the residual from the regression of the excess stock return on theexcess market return using the monthly return observations over the prior 60 months(Harvey and Siddique, 2000).

11Specifically, we use the delisting return from CRSP when a stock is delisted. If a delisting return is notavailable, we assume that it is 100%, unless the delisting code is 500 (reason unavailable), 520 (went toOTC), 551-573, 580 (various reasons), 574 (bankruptcy), or 584 (does not meet exchange financial guide-lines). For these observations, we assume that the delisting return is −30%.

8

2.4 Summary Statistics

Table 1 presents the time-series averages of some cross-sectional statistics and correla-tions. All variables except RETt are measured at the end of month t. Values in Panel Aare in percentages12 The newly introduced variable SPFRET⊥ the abnormal return thatis obtained by regressing SPFRET on INDRET using a panel regression with month fixedeffects. In other words, it is the component of SPFRET that is orthogonal to INDRET.Panel A shows why this orthogonalization is warranted: the average cross-sectional cor-relation between SPFRET and INDRET is 0.70. This is not surprising given that SPFRETis formed by using stocks from the same industry as the focal stock. We also see thatthe average correlation between SPFMOM and INDMOM is 0.71. The results that we re-port in the main text are mostly based on the raw (non-orthogonal) versions of SPFRETand SPFMOM. However, because of these high correlations with the industry variables,we prepare another set of results using the orthogonalized versions as a robustness checkand report it in the appendix. Both the raw and orthogonalized versions give very similarresults.

[Insert Table 1 here]

3 The Return Predictability of Social-Peer Firm ReturnsA growing literature shows that social connectedness fosters economic interactions (e.g.,Cohen et al., 2017; Bailey et al., 2018b). Thus, social connectedness can potentially serveas a proxy for the economic links between firms across different regions. As a result,performances of firms located in socially connected counties are likely positively relatedto the focal firm’s performance. A number of prior studies show that investors tend tounderreact to the information contained in economically relevant firms (e.g., Cohen andFrazzini, 2008; Lee et al., 2019; Parsons et al., 2020; Jin and Li, 2020). In this section, weexamine whether returns of social-peer firms can predict focal firms’ returns.

3.1 Univariate-Sorted Portfolios

We first examine the return predictability of social peer returns by constructing portfoliossorted based on SPFRET. As described in the previous section, SPFRET is constructed

12All ratios in the paper are reported in percentages.

9

using returns of all firms in the focal firms’ industry. To maximize the impact of thefirms with high social connectedness, we weight these firms by the SCI between the head-quarters counties of the peer firm and the focal firm. We also consider a second sortingvariable, SPFRET⊥, which is the residual of SPFRET regressed on industry returns. Theorthongalized measure captures information from social peer firms that is unrelated toaggregate industry performances.

Table 2 reports the results of the univariate portfolio sorts. Specifically, we sort oursample firms into deciles based on either SPFRET (Panel A) or SPFRET⊥ (Panel B). Then,we construct equal-weighted return for each portfolio, along with the return of a high-minus-low portfolio obtained through going long in the decile-10 portfolio and short inthe decile-1 portfolio are calculated over the following month. The first row in each panelshows the raw returns. To ensure that our results are not driven by risk factors or anoma-lies, we additionally report characteristics-based alphas computed by using DGTW ad-justments (see Daniel et al., 1997; Wermers, 2003) in the second row, and the Fama-French5-factor (FF5) alphas in the third row. The corresponding t-statistics based on Newey andWest (1994) standard errors are reported in parentheses below the alphas.

Panel B shows that, even after controlling for INDRET, the high-minus-low portfoliohas a monthly characteristics-based (CB) alpha of 0.51% per month, and a monthly FF5alpha of 0.68%, with t-statistics of 6.15 and 6.30 respectively. Thus, the predictabilityis unlikely to be driven by industry momentum (e.g., Moskowitz and Grinblatt, 1999).In Table A1, we report the value-weighted portfolio returns as well. For value-weightedportfolios, the DGTW and FF5 alphas are 0.30% and 0.66% with t-statistics of 3.71 and 4.64respectively. Overall, this set of results indicate that month-t SPFRET has a substantialpredictive power for the focal stock return in month t + 1 and the predictability is largelyindependent of the industry momentum effects or risk adjustments.

[Insert Table 2 here]

3.2 Controlling for Industry Effects

Next, we conduct additional tests to contrast the effect of social connectedness with re-turns of firms in the same industry in order to further ensure that our results are notdriven by industry momentum as documented in Moskowitz and Grinblatt (1999). InTable 3, we report the results of bivariate portfolio sort based on SPFRET and INDRET,

10

where INDRET is the portfolio return of the firms in the same Fama-French 48 as the focalfirm, which captures the effect of industry momentum. In Panel A, we conduct sequentialsort, where stocks are first sorted into quintiles based on their INDRET and then, withineach INDRET quintile and then further sorted into quintiles using SPFRET breakpointsfor that INDRET quintile only.

In Panel B, second sorting is done by using SPFRET breakpoints for all stocks fromthat specific month (independent sorting). After sorting, equal-weighted return for eachof the 25 resultant portfolios are calculated over the following month, along with thereturns of five high-minus-low portfolios that are long in stocks in the SPFRET quintile5 and short in the SPFRET quintile 1 stocks for each INDRET quintile. This proceduregenerates 30 time-series of portfolio returns. We calculate alphas using Fama and French(2015) models for all 30 portfolios. The final rows in both panels report the average alphasfor the SPFRET quintiles. Essentially, these average alphas can be viewed as informationderived from SPFRET, after controlling for INDRET. The high-minus-low average alphain Panel A is 0.48% per month with a t-statistics of 8.42, whereas the average high-minus-low alpha in Panel B is 0.70% per month with a t-statistics of 6.17. Hence, the bivariatesort results confirm the large economic and statistical predictability of focal stock returnsby SPFRET.

[Insert Table 3 here]

Table A2 reports the same set of results using SPFRET⊥. Our results are fully con-sistent with those based on SPFRET. The main takeaway from these results is that whileSPFRET and INDRET are highly related, SPFRET contains useful information about focalfirms that is not captured by the average industry returns.

Another way to make this contrast is to explicitly split industry peer firms into highand low social connectedness groups. Empirically, we sort stocks from the same industryas the focal stock into two portfolios based on the SCI between their county and the focalcounty while using the median SCI as the breakpoint. We report the result from this splitin in Table 4. IND_HIGH denotes the equal-weighted return of the high-SCI portfolio.In columns 1 to 3, all stocks except the focal stock are used while forming the portfolios,while only the stocks of firms headquartered out of the focal state are used in columns 4through 6. In Panel A, the dependent variable is the contemporaneous (i.e., month t) fo-

11

cal stock return. In Panel B, it is the following month’s focal stock return. In both panels,covariates include the basic control variables described in Section 2.3 (RET, SIZE, BMKT,BM, IVOL, ILLIQ, MAX, SKEW and COSKEW). All independent variables are standard-ized cross-sectionally and winsorized at 1% and 99%. t-statistics with standard errorsbased on Newey and West (1994) adjustments are reported in parentheses.

Panel A shows that SPFRET has substantial explanatory power for the contemporane-ous focal stock returns even after controlling for aggregate industry returns. Specifically,a one standard deviation increase in SPFRET while holding INDRET constant is asso-ciated with a 0.65% in the contemporaneous focal stock return if all stocks are used toform the SPFRET, and by 0.51% if only the out-of-state stocks are used. Comparing thesewith the coefficient estimates for INDRET, we find that co-movement between RETt andthe idiosyncratic component of SPFRET is almost as high as the co-movement betweenRETt and the industry return. These results suggest that socially connected stocks con-tain incremental information about the focal firm above and beyond the overall industryreturns.

Panel B shows that SPFRET has substantial predictive power for next month’s focalstock return, even after controlling for the industry return. A one standard deviationincrease in SPFRET while holding INDRET constant increases next month’s focal stockreturn by 0.23% if all stocks are used to form the SPFRET, and by 0.19% if only the out-of-state stocks are used. Again, these numbers are almost as large at the coefficient estimatesfor INDRET, indicating that return predictability of social-peer returns is not a result ofindustry momentum.

[Insert Table 4 here]

3.3 Fama-MacBeth Regression Analysis

We next examine whether our results are driven by other known predictive variables byinclude additional controls in our cross-sectional regression. Table 5 reports the resultsof Fama-MacBeth regressions where the focal stock return in next month is regressedon (SPFRET), along with short-term industry momentum (INDRET), long-term industrymomentum (INDMOM), geographic lead-lag effect (GEORET), shared analyst coveragelead-lag effect (CFRET), and customer-supplier lead-lag effect (CRET). We also include abattery of well known cross-sectional predictive variables, including RET, SIZE, BMKT,

12

BM, MOM, IVOL, ILLIQ, MAX, SKEW, COSKEW.13 For brevity, we do not report theircoefficients.

In column 1, we start with SPFRET along with basic controls as our covariates andstart adding other predictors that capture economic linkages between firms as we movealong to the other columns. Coefficient estimate for SPFRET remains significant at 5%level even after controlling for all predictors as column 6 shows, and it is significant at the1% level in the rest of the columns.

Note that the number of observations drop substantially once CRET, the average re-turn of the customer portfolio, is added to the regressors. We can see the reason in Panel Aof Table 1: the average number of monthly observations for CRET is 503 while the numberof observations for the other variables are all well over 2000. Thus, adding CRET to thecovariates reduces the number of observations significantly, and this in turn reduces thestatistical significance for all covariates. Still, even in this restricted sample, the economicmagnitude of the effect of SPFRET on RETt+1 is significant: An increase in SPFRET fromthe 10% to the 90% of its distribution leads to a 42 basis points increase in next month’sfocal stock returns. This magnitude is substantial, given that CFRET has been shown tosubsume the effects of other lead-lag return predictors (see Ali and Hirshleifer, 2020).

[Insert Table 5 here]

Table A3 repeats the analyses above by replacing SPFRET with its orthogonalizedversions, SPFRET⊥. In each column, orthogonalization of SPFRET is performed by run-ning a panel regression of SPFRET on RET, MOM, the covariates that are displayed forthat column and month fixed effects, using all available stock-month observations priorto month t. We find that our results continue to be highly significant. In sum, this set ofresults lend further credence to our predictability results. In particular, they show that thereturn predictability resulted from social-peer firms’ returns is a new phenomenon thathas not been documented in the existing literature.

13These additional control represents short-term reversals, size, beta, book-to-market, momentum, idiosyn-cratic volatility, Amihud illiquidity, max daily return in the previous month, realized skewness, and co-skewness.

13

3.4 Heterogeneity Analysis

The return predictability documented in the previous subsection is consistent with slowinformation diffusion (e.g., Hong and Stein, 1999). Additionally, there are also additionalstudies that argue prices may respond to information slowly due to limited cognitiveresources of investors (e.g., Hirshleifer and Teoh, 2003; Peng and Xiong, 2006). If ourresults are, in part, driven by investors’ inattention, we should expect the predictabilityresults to be stronger among firms with low visibility and institutional ownership.

Empirically, we test this set of hypotheses in Table 6 by conducting Fama and Mac-Beth (1973) regressions of RETt+1 on SPFRET, INDRET, INDMOM, GEORET, CFRET andthe other basic controls for various sub-samples of our data 14. In particular, at the endof month t we sort all common stocks into two sub-samples on three different dimen-sions: size, institutional ownership and analyst coverage. For size sorting, we use NYSEbreakpoints.

Columns 1 and 2 report the results based on size split. We find that the return pre-dictability is twice as large in small firms comparing to large firms. Columns 3 and 4 re-port results based on institutional ownership split. We find that while return predictabil-ity exists among stocks in both groups, the economic magnitude of the coefficient forlow ownership group is 60% higher than the high ownership group. Similarly, we findthat return predictability only exists among low analyst coverage group. We repeat theanalyses above by replacing SPFRET with SPFRET⊥ in Table A4 and our results remainrobust. Taken together, these results indicate are consistent with patterns of slow infor-mation diffusion resulted from investor inattention (e.g., Hirshleifer and Teoh, 2003; Pengand Xiong, 2006).

[Insert Table 6 here]

3.5 Return Predictability in the Long Run

Having established that returns of socially connected firms can predict focal firms’ re-turns, we next analyze the horizon of this return predictability. This analysis helps usidentify the mechanism for the return predictability documented in the previous subsec-tion. For example, our return predictability can be a result of investors’ underreaction to

14We omit CRET from the regressors because of the sample reduction issue that we saw in Table 5.

14

fundamental information(e.g., Hirshleifer and Teoh, 2003; Peng and Xiong, 2006). How-ever, the same set of result can also be consistent with investors overreacting to the pos-itive information from socially connected firms (e.g., De Bondt and Thaler, 1985; Danielet al., 1998). We can empirically examine these two hypotheses. If the predictability isa result of investors’ underreaction, we should not observe a reversal of the return pre-dictability in the long-run. Otherwise, we expect to find that returns of socially connectedfirms should be negatively associated with long-run stock returns after the immediatefuture.

Table 7 reports the long-horizon return predictability results for our main variables.Returns are measured using characteristics-based CARs. In Panel A, we sort all commonstocks into deciles based on their SPFRET (columns 1 to 3) and SPFMOM (columns 4 to 6).Next, we calculate the equal-weighted 12-month, 24-month and 60-month CARs for eachportfolio. We also compute the same CARs for the high-minus-low portfolio that is longin the decile-10 portfolio and short in the decile-1 portfolio. This gives us nine monthlytime series of CARs. For each time series, we then calculate and report the average CARas the characteristic-based alpha for the corresponding horizon. The results show noapparent return reversal.15 Panel B reports the results of the Fama-MacBeth regressionswhere characteristics-based CARs for 12-month, 24-month and 60-month horizons arerun against SPFRET, SPFMOM and various other controls. We apply the same price anddata filters as in Panel A. The coefficient estimates for both SPFRET and SPFMOM arestatistically and economically significant at all three horizons. Once again, there is noevidence of return reversal. In fact, the coefficient estimates are monotonically increasingin return horizon.

[Insert Table 7 here]

Table A5 repeats the same analyses by replacing SPFRET and SPFMOM with theirorthogonalized versions: SPFRET⊥ and SPFMOM⊥. SPFRET⊥ for stock i at the end ofmonth t is defined as SPFRETi,t − β̂

′tXi,t, with β̂t being the vector of coefficient esti-

mates for the panel regression FNDRETi,s = α + β′tXi,s + θs + εs for all s ≤ t, where

15Although, the alpha for CAR24 is slightly smaller than that for CAR12 for both SPFRET and SPFMOM,the null hypothesis that the two are equal cannot be rejected at 10% level. For SPFRET, the null hypothesisthat alphas for CAR60 and CAR24 are equal is rejected at 5% level, whereas it is rejected at 1% level forSPFMOM.

15

θs is the fixed effect for month s. Elements of Xi,s, the vector of covariates for stock ias measured at the end of month t, include stock’s own return, equal-weighted indus-try return, 12- to 60-month stock momenta and 12- to 60-month equal-weighted industrymomenta.16 Month t is excluded from momentum calculations. For SPFMOM, orthog-onalization method is identical to that for SPFRET, except variables that are measuredat the end of month t are excluded from the covariate vector to prevent the look-aheadbias. The results are in line with those in Table 7; both the portfolio sort alphas andFama-MacBeth coefficient estimates are monotonically increasing with the return hori-zon, which implies that not only there is no return reversal in the long-run, but the returnpredictability may also exist at the five-year horizon. Thus, our evidence is inconsistentwith the overreaction explanation.

4 Predicting Firm Performances and Future EarningsSo far, our analyses reveal that returns of socially connected firms can predict return inthe following month and may have incremental predictability in the following five years.We have argued that this return predictability is resulted from more intense economicexchange between regions. However, we have not presented direct evidence consistentwith this view. Thus, in this subsection, we intend to test if returns of socially connectedfirms contain information about focal firms’ fundamental performances in the future.

4.1 Long-Run Firm Fundamentals

We first test whether social-peer firm returns predict focal firms’ long-term fundamen-tals. If social peer firms have strong fundamental connections to focal firms, we expectthat positive news from social peer firms should be associated with better focal firm per-formance in the future.

In Table 8, we examine the long-horizon predictability in firm fundamentals due toSPFRET and SPFMOM using panel regressions, where we regress one-, two- and five-year cumulative value of the fundamental on SPFRET, SPFMOM and the same controlvariables used in Table 7, along with month fixed effects. We choose three financial ratios

16Orthogonalization with respect to long-run industry and stock momenta allows us to control for long-run reversals documented in De Bondt and Thaler (1985), which could otherwise affect our long-horizonresults.

16

as our dependent variables: asset turnover (AT), earnings before interest, taxes, depreci-ation and amortization (EBITDA) and gross profitability (GP). AT is defined as the totalsales scaled by the beginning-of-year total assets. EBITDA is calculated by subtractingthe cost of goods sold (COGS) and selling, general and administrative expenses (SGAX)from the total revenue. GP is calculated by subtracting the COGS from the total revenue.AT measures asset management efficiency, while EBITDA and GP are profitability mea-sures. We scale both EBITDA and GP by the beginning-of-year total assets before usingthem in the regressions. At the end of month t, for each common stock, we calculate thefundamentals over the next financial year. This avoids the overlap between the predic-tors and the variables being predicted. We further require each stock to have five years ofuninterrupted fundamental observations.

The results show that both SPFRET and SPFMOM have significant predictive powerfor all three fundamental variables, and the predictability increases monotonically withthe horizon. Table A6 reports the same results for SPFRET⊥ and SPFMOM⊥, orthogonal-ized the same way as in Table A5. The two sets of results are very similar to each other interms of the monotonicity of return predictability. This set of results indicate social-peerfirm returns and social-peer firm indeed contain information about focal firms’ futurefundamental performances.

[Insert Table 8 here]

4.2 Earnings Surprises and Analyst Forecast Errors

We next analyze whether market participants fully incorporate returns of social-peerfirms in their expectations. Following Cohen and Frazzini (2008); Hong et al. (2000), weproxy investor expectations using sell-side analyst forecast of earnings. If market par-ticipants do not fully incorporate social peer firms’ information to their expectations, weexpect social-peer returns to capture the difference between realized earnings and con-sensus analyst expectations.

Table 9 presents the panel regression results for earnings predictability by SPFRETand SPFMOM. In columns 1 to 3, we regress the standardized unexpected earnings (SUE)calculated by using a random-walk model on SPFRET, SPFMOM following Bernard andThomas (1990); Mendenhall (1991). We include the same control variables as those used

17

in Table 7 and Table 8, along with month fixed effects. Specifically, we calculate the un-expected earnings (UE) of stock i in calender quarter q of an earnings announcement asUEi,q = EPSi,q − EPSi,q−4, where EPSi,q and EPSi,q−4 are the stock’s basic earnings pershare (EPS) excluding extraordinary items in quarters q and q− 4 respectively. EPS is ad-justed for stock splits and reverse splits through division by the Compustat item AJEXQ,the quarterly cumulative adjustment factor. Standardized unexpected earnings in quarterq (SUEi,q) is defined as UEi,q scaled by its standard deviation over the past eight quarters,with a minimum of four UE observations available. The n-year cumulative SUE for astock at the end of month t is then the arithmetic sum of its SUEs over the next 4n cal-ender quarters, excluding the quarter that includes month t. This prevents the overlapbetween the predictors and the predicted variables. Results show that both SPFRET andSPFMOM strongly predicts the SUE up to five years into the future.

Following DellaVigna and Pollet (2009), in columns 4 through 6, we redefine the SUEfor calender quarter q as the difference between the announced earnings for that quarterand analyst consensus forecast (i.e., the median forecast), scaled by the stock price 5 trad-ing days before the earnings announcement date.17 Our results show that SPFRET canpredict analyst-forecast-based SUEs up to two years into the future, while SPFMOM hassignificant predictive power over the following year. These results suggest that sell-sideanalysts are sluggish in incorporate the information social-peer firms into their earningsforecasts, with a delay of up to two years. In table A7 we repeat the same analyses usingSPFRET⊥ and SPFMOM⊥, orthogonalized the same way as in Table A5 and Table A6.The results turn out to be very similar to those in Table 9.

[Insert Table 9 here]

These tests are consistent with the idea that firms from social connectedness regioncontain fundamental information about focal firm. However, market participants, includ-ing sell-side analysts and investors do not timely digest such information. As a result, weobserve significant return predictability using return information of firms located in so-cially connected regions.17We adjust actuals, forecasts and prices for stock splits and reverse splits through scaling them by the item

CFACSHR, the cumulative adjustment factor, from the daily CRSP file. We only use the forecasts madefor the next quarter and within 90 days of the earnings announcement date. If an analyst makes multipleforecasts within this time window, we use the latest forecast. We also require at least 5 distinct analystsmake an earnings forecast for the firm for the next quarter.

18

5 ConclusionThis paper explores a new dimension of cross-firm linkages that are based on social con-nections between regions in which the firms are located. Using Facebook-based data, weanalyze whether the intensity of social connectedness across regions is related to correla-tions in firms’ fundamentals and their stock returns.

We construct a novel measure of socially-connected peer firm returns and find thatthese returns are highly correlated with the returns of focal firms. Controlling for in-dustry returns, a one standard deviation of increase in the returns of social peer firms isassociated with a 71 basis points increase in focal firm returns for the following month.This predictability is not subsumed by industry momentum, other variables that capturelead-lag relationship, or alternative cross-sectional predictive variables. Additionally, thispredictability lasts for up to five years does not reverse in the long run.

Consistent with investor inattention generates sluggish price adjustments, our resultsare stronger among firms with low visibility (measured by market capitalization, institu-tional ownership, or low analyst coverage). Further, we show that returns of social-peerfirms help predict focal firms’ long-term fundamentals for up to five years. In addition,social-peer firms’ returns predict the future revisions of earnings forecasts by financialanalysts, suggesting that the analysts fail to account for such information.

Our findings suggest that social ties across regions offer a new way to capture impor-tant fundamental economic linkages across firms. The social-connection based linkagemeasure can be useful for future studies that try to uncover the way in which shockspropagate across the network of firms.

19

ReferencesAcemoglu, Daron, Vasco M Carvalho, Asuman Ozdaglar, and Alireza Tahbaz-Salehi,

2012, The network origins of aggregate fluctuations, Econometrica 80, 1977–2016.

Ali, Usman, and David Hirshleifer, 2020, Shared analyst coverage: Unifying momentumspillover effects, Journal of Financial Economics 136, 649–675.

Amihud, Yakov, 2002, Illiquidity and stock returns: cross-section and time-series effects,Journal of financial markets 5, 31–56.

Bailey, Michael, Rachel Cao, Theresa Kuchler, Johannes Stroebel, and Arlene Wong, 2018a,Social connectedness: Measurements, determinants, and effects, Journal of Economic Per-spectives 32, 259–80.

Bailey, Michael, Ruiqing Cao, Theresa Kuchler, and Johannes Stroebel, 2018b, The eco-nomic effects of social networks: Evidence from the housing market, Journal of PoliticalEconomy 126, 2224–2276.

Bailey, Michael, Eduardo Dávila, Theresa Kuchler, and Johannes Stroebel, 2019a, Houseprice beliefs and mortgage leverage choice, The Review of Economic Studies 86, 2403–2452.

Bailey, Michael, Patrick Farrell, Theresa Kuchler, and Johannes Stroebel, 2019b, Socialconnectedness in urban areas, NBER Working Paper .

Bailey, Michael, Abhinav Gupta, Sebastian Hillenbrand, Theresa Kuchler, Robert Rich-mond, and Johannes Stroebel, 2020, International trade and social connectedness, Work-ing Paper .

Bailey, Michael, Drew M Johnston, Theresa Kuchler, Johannes Stroebel, and Arlene Wong,2019c, Peer effects in product adoption, Technical report, National Bureau of EconomicResearch.

Bernard, Victor L, and Jacob K Thomas, 1990, Evidence that stock prices do not fullyreflect the implications of current earnings for future earnings, Journal of accounting andeconomics 13, 305–340.

20

Breschi, Stefano, and Camilla Lenzi, 2016, Co-invention networks and inventive produc-tivity in us cities, Journal of Urban Economics 92, 66–75.

Cohen, Lauren, and Andrea Frazzini, 2008, Economic links and predictable returns, TheJournal of Finance 63, 1977–2011.

Cohen, Lauren, Andrea Frazzini, and Christopher Malloy, 2008, The small world of in-vesting: Board connections and mutual fund returns, Journal of Political Economy 116,951–979.

Cohen, Lauren, Umit G Gurun, and Christopher Malloy, 2017, Resident networks andcorporate connections: Evidence from world war ii internment camps, The Journal ofFinance 72, 207–248.

Daniel, Kent, Mark Grinblatt, Sheridan Titman, and Russ Wermers, 1997, Measuring mu-tual fund performancewith characteristic-based benchmarks, Journal of Financial Eco-nomics 52, 1035–1058.

Daniel, Kent, David Hirshleifer, and Avanidhar Subrahmanyam, 1998, Investor psychol-ogy and security market under-and overreactions, the Journal of Finance 53, 1839–1885.

De Bondt, Werner FM, and Richard Thaler, 1985, Does the stock market overreact?, TheJournal of finance 40, 793–805.

DellaVigna, Stefano, and Joshua Pollet, 2007, Demographics and industry returns, Ameri-can Economic Review 97, 1667–1702.

DellaVigna, Stefano, and Joshua Pollet, 2009, Investor inattention and Friday earningsannouncements, Journal of Finance 64, 709–749.

Dougal, Casey, Christopher A Parsons, and Sheridan Titman, 2015, Urban vibrancy andcorporate growth, The Journal of Finance 70, 163–210.

Dougal, Casey, Christopher A Parsons, and Sheridan Titman, 2018, Urban vibrancy andfirm value creation, Working Paper .

Duggan, Maeve, Shannon Greenwood, and Andrew Perrin, 2016, Social media update2016. Pew Research Center.

21

Fama, Eugene F, and Kenneth R French, 2015, A five-factor asset pricing model, Journal offinancial economics 116, 1–22.

Fama, Eugene F, and James D MacBeth, 1973, Risk, return, and equilibrium: Empiricaltests, Journal of political economy 81, 607–636.

Giroud, Xavier, and Holger M Mueller, 2019, Firms’ internal networks and local economicshocks, American Economic Review 109, 3617–49.

Harvey, Campbell R, 2017, Presidential address: The scientific outlook in financial eco-nomics, The Journal of Finance 72, 1399–1440.

Harvey, Campbell R., and Akhtar Siddique, 2000, Conditional skewness in asset pricingtests, Journal of Finance 55, 1263–1295.

Hasan, Iftekhar, Chun Keung Hoi, Qiang Wu, and Hao Zhang, 2017, Social capital anddebt contracting: Evidence from bank loans and public bonds, Journal of Financial andQuantitative Analysis 52, 1017–1047.

Heimer, Rawley Z, 2016, Peer pressure: Social interaction and the disposition effect, TheReview of Financial Studies 29, 3177–3209.

Hirshleifer, David, Sonya S. Lim, and Siew Hong Teoh, 2009, Driven to distraction: Extra-neous events and underreaction to earnings news, Journal of Finance 64, 2289–2325.

Hirshleifer, David, and Siew Hong Teoh, 2003, Limited attention, information disclosure,and financial reporting, Journal of accounting and economics 36, 337–386.

Hoberg, Gerard, and Gordon M Phillips, 2018, Text-based industry momentum, Journal ofFinancial and Quantitative Analysis 53, 2355–2388.

Hochberg, Yael V, Alexander Ljungqvist, and Yang Lu, 2007, Whom you know matters:Venture capital networks and investment performance, The Journal of Finance 62, 251–301.

Hong, Harrison, Terence Lim, and Jeremy C Stein, 2000, Bad news travels slowly: Size,analyst coverage, and the profitability of momentum strategies, The Journal of Finance55, 265–295.

22

Hong, Harrison, and Jeremy C Stein, 1999, A unified theory of underreaction, momentumtrading, and overreaction in asset markets, The Journal of finance 54, 2143–2184.

Hou, Kewei, 2007, Industry information diffusion and the lead-lag effect in stock returns,The Review of Financial Studies 20, 1113–1138.

Hou, Kewei, and Tobias Moskowitz, 2005, Market frictions, price delay, and the cross-section of expected returns, Review of Financial Studies 18, 981–1020.

Huang, Xing, 2015, Thinking outside the borders: Investors’ underreaction to foreignoperations information, The Review of Financial Studies 28, 3109–3152.

Huberman, Gur, and Tomer Regev, 2001, Contagious speculation and a cure for cancer,Journal of Finance 56, 387–396.

Hwang, Byoung-Hyoun, and Seoyoung Kim, 2009, It pays to have friends, Journal of fi-nancial economics 93, 138–158.

Jegadeesh, Narasimhan, and Sheridan Titman, 1993, Returns to buying winners and sell-ing losers: Implications for stock market efficiency, The Journal of finance 48, 65–91.

Jin, Zuben, and Frank Weikai Li, 2020, Geographic links and predictable returns, Availableat SSRN 3617417 .

Korniotis, George M, and Alok Kumar, 2013, State-level business cycles and local returnpredictability, The Journal of Finance 68, 1037–1096.

Kuchler, Theresa, Yan Li, Lin Peng, Johannes Stroebel, and Dexin Zhou, 2020a, Socialproximity to capital: Implications for investors and firms, Technical report, NationalBureau of Economic Research.

Kuchler, Theresa, Dominic Russel, and Johannes Stroebel, 2020b, The geographic spreadof covid-19 correlates with structure of social networks as measured by facebook, Tech-nical report, National Bureau of Economic Research.

Lee, Charles MC, Stephen Teng Sun, Rongfei Wang, and Ran Zhang, 2019, Technologicallinks and predictable returns, Journal of Financial Economics 132, 76–96.

23

McLean, R David, and Jeffrey Pontiff, 2016, Does academic research destroy stock returnpredictability?, The Journal of Finance 71, 5–32.

Mendenhall, Richard R, 1991, Evidence on the possible underweighting of earnings-related information, Journal of Accounting Research 29, 170–179.

Moskowitz, Tobias J, and Mark Grinblatt, 1999, Do industries explain momentum?, TheJournal of finance 54, 1249–1290.

Newey, Whitney K, and Kenneth D West, 1994, Automatic lag selection in covariancematrix estimation, The Review of Economic Studies 61, 631–653.

Parsons, Christopher A, Riccardo Sabbatucci, and Sheridan Titman, 2020, Geographiclead-lag effects, The Review of Financial Studies 33, 4721–4770.

Peng, Lin, and Wei Xiong, 2006, Investor attention, overconfidence and category learning,Journal of Financial Economics 80, 563–602.

Pirinsky, Christo, and Qinghai Wang, 2006, Does corporate headquarters location matterfor stock returns?, The Journal of Finance 61, 1991–2015.

Pool, Veronika K, Noah Stoffman, and Scott E Yonker, 2012, No place like home: Fa-miliarity in mutual fund manager portfolio choice, The Review of Financial Studies 25,2563–2599.

Pool, Veronika K, Noah Stoffman, and Scott E Yonker, 2015, The people in your neighbor-hood: Social interactions and mutual fund portfolios, The Journal of Finance 70, 2679–2732.

Rehbein, Oliver, Simon Rother, et al., 2020, Distance in bank lending: The role of socialnetworks, Technical report, University of Bonn and University of Mannheim, Germany.

Shumway, Tyler, 1997, The delisting bias in crsp data, Journal of Finance 52, 327–340.

Smajlbegovic, Esad, 2019, Regional economic activity and stock returns, Journal of Finan-cial and Quantitative Analysis 54, 1051–1082.

24

Tuzel, Selale, and Miao Ben Zhang, 2017, Local risk, local factors, and asset prices, TheJournal of Finance 72, 325–370.

Wermers, Russ, 2003, Is money really “smart”? new evidence on the relation betweenmutual fund flows, manager behavior, and performance persistence, Working Paper .

25

Table 1: Summary Statistics and Correlations

The table reports the summary statistics and correlation matrix for the main variables used in thepaper. All variables except for RETt+1 are measured at the of portfolio formation month t. RETt+1is the monthly focal stock return following the portfolio formation month. SPFRET (social-peerfirm return) is the SCI-weighted average return of stocks that are in the same FF48 industry but notin the same state as the focal stock. SPFMOM is the compound SPFRET between t− 11 and t− 1.INDRET is the equal-weighted average return of all stocks that are from the same FF48 industryas the focal stock. INDMOM is obtained by compounding INDRET over the months t-11 to t-1.GEORET is the equal-weighted average return of stocks that are in the same economic area butnot in the same FF48 industry as the focal stock. GEOMOM is obtained by compounding GEORETover the months t-11 to t-1. CFRET is the weighted average return of stocks that share at least oneanalyst with the focal stock over previous 12 months, where weights are the number of sharedanalysts between stocks. CFMOM is obtained by compounding CFRET from month t− 11 to t− 1.CRET is the equal-weighted average stock return of the main customers of the focal firm. PanelA reports the time-series averages of the number of monthly observations and cross-sectionalmean, standard deviation, minimum, cutoff at the 25 and 75 percentile, min and maximum foreach variable. The values are reported in percentages. Panel B exhibits the time-series averagesof cross-sectional correlations between the variables. The additional variable SPFRET⊥ is SPFRETorthogonalized to INDRET using a panel regression and all available stock-month observationsup to and including month t. All values in Panel A are in percentages.

Panel A: Summary statistics.

N Mean Std. Dev. Min. 25% 75% Max.RETt+1 2562 0.845 11.724 -76.005 -4.776 6.102 102.843

SPFRET 2472 1.484 3.817 -18.936 -0.822 3.610 27.520SPFMOM 2315 18.552 17.626 -40.726 7.117 27.804 134.343INDRET 2478 0.744 3.178 -9.329 -1.293 2.633 12.727

INDMOM 2511 9.740 14.486 -26.552 0.178 18.162 62.890GEORET 2149 1.054 3.558 -28.161 -0.295 2.298 40.651

GEOMOM 2148 13.040 13.470 -57.905 7.247 17.699 154.069CFRET 2683 1.112 4.145 -23.613 -1.232 3.370 30.594

CFMOM 2683 17.895 18.963 -40.110 6.756 26.530 229.910CRET 503 0.999 8.282 -47.720 -3.085 4.982 46.901

26

Panel B: Correlation matrix.

RETt+1 SPFRET SPFMOM SPFRET⊥ INDRET INDMOM GEORET GEOMOM CFRET CFMOM CRETRETt+1 1.000

SPFRET 0.023 1.000SPFMOM 0.010 0.078 1.000SPFRET⊥ 0.012 0.694 0.014 1.000INDRET 0.027 0.700 0.096 0.008 1.000

INDMOM 0.012 0.044 0.710 -0.041 0.101 1.000GEORET -0.002 0.044 0.018 0.010 0.047 0.018 1.000

GEOMOM 0.008 -0.002 0.058 -0.012 0.013 0.073 0.035 1.000CFRET 0.033 0.392 0.029 0.130 0.414 0.013 0.097 0.004 1.000

CFMOM 0.021 0.048 0.425 -0.000 0.067 0.437 0.024 0.110 0.043 1.000CRET 0.014 0.140 0.009 0.052 0.139 0.010 0.044 -0.002 0.206 0.011 1.000

27

Table 2: Socially-Peer Firm Returns and Focal Firm Returns: Univariate Sorting

The table reports the results of the univariate portfolio sorts based on SPFRET and SPFRET⊥. SPFRET is the SCI-weighted average return of stocks from the same FF48 industry as the focal stock but from a different state. SPFRET⊥is the abnormal return generated from a panel regression of SPFRET on the equal-weighted industry return, using allavailable stock-month observations prior to the portfolio formation month (i.e., month t) and month fixed effects. Atthe end of each month from January 1962 to December 2019, all common stocks are sorted into deciles based on eitherSPFRET (Panel A) or SPFRET⊥ (Panel B). Then, equal-weighted return for each portfolio, along with the return of ahigh-minus-low portfolio obtained through going long in the decile-10 portfolio and short in the decile-1 portfolioare calculated over the following month. The first row in each panel shows the raw returns, second row shows thecharacteristics-based alphas computed by using DGTW adjustments (see Daniel et al., 1997; Wermers, 2003), and thethird row reports the Fama-French 5-factor (FF5) alphas. All returns are in percentages. The corresponding t-statisticsbased on Newey and West (1994) standard errors are reported in parentheses below the alphas. *, **, and *** indicatestatistical significance at the 10%, 5%, and 1% levels, respectively.

Panel A: Contemporaneous Relation.

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (10-1)

Raw Return 0.072 0.266 0.521∗∗ 0.803∗∗∗ 0.893∗∗∗ 0.879∗∗∗ 0.940∗∗∗ 1.161∗∗∗ 1.345∗∗∗ 1.538∗∗∗ 1.466∗∗∗

(0.307) (1.180) (2.393) (3.705) (4.168) (4.019) (4.472) (5.302) (5.990) (6.374) (8.518)

DGTW Alpha −0.494∗∗∗ −0.341∗∗∗ −0.147∗∗ 0.056 0.130∗∗ 0.127∗∗∗ 0.127∗∗ 0.363∗∗∗ 0.444∗∗∗ 0.631∗∗∗ 1.125∗∗∗

(−6.246) (−3.891) (−2.136) (0.991) (2.223) (2.645) (2.533) (6.221) (5.913) (6.528) (7.541)

FF5 Alpha −1.033∗∗∗ −0.767∗∗∗ −0.547∗∗∗ −0.274∗∗∗ −0.227∗∗∗ −0.248∗∗∗ −0.132∗∗ 0.086 0.330∗∗∗ 0.520∗∗∗ 1.552∗∗∗

(−8.283) (−5.819) (−5.688) (−3.127) (−2.605) (−3.044) (−2.255) (1.007) (3.004) (3.880) (6.944)

Panel B: Predictive relation.

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (10-1)

Raw Return 0.541∗∗ 0.607∗∗∗ 0.631∗∗∗ 0.716∗∗∗ 0.769∗∗∗ 0.918∗∗∗ 0.974∗∗∗ 1.001∗∗∗ 1.054∗∗∗ 1.204∗∗∗ 0.663∗∗∗

(2.561) (2.700) (3.070) (3.561) (3.723) (4.303) (4.617) (4.713) (4.789) (5.349) (6.814)

DGTW Alpha −0.148∗∗ −0.085 −0.088 −0.020 0.025 0.147∗∗∗ 0.185∗∗∗ 0.233∗∗∗ 0.286∗∗∗ 0.358∗∗∗ 0.506∗∗∗

(−2.069) (−1.480) (−1.537) (−0.364) (0.435) (2.714) (3.459) (4.621) (4.930) (5.638) (6.150)

FF5 Alpha −0.580∗∗∗ −0.458∗∗∗ −0.438∗∗∗ −0.332∗∗∗ −0.305∗∗∗ −0.131∗ −0.076 −0.060 −0.016 0.101 0.681∗∗∗

(−5.270) (−4.989) (−6.505) (−3.905) (−4.479) (−1.674) (−1.173) (−0.854) (−0.270) (1.640) (6.304)

28

Table 3: Return Predictability of Social-Peer Firm Returns: BivariateSortingThe table reports the results of bivariate portfolio sort based on SPFRET and INDRET, whereINDRET is the portfolio return of the firms in the same FF48 industry as the focal firm. InPanel A, we conduct sequential sort, where stocks are first sorted into quintiles based on theirINDRET and then, within each INDRET quintile and then further sorted into quintiles usingSPFRET breakpoints for that INDRET quintile only. In Panel B, second sorting is done by usingSPFRET breakpoints for all stocks from that specific month (independent sorting). After sorting,equal-weighted return for each of the 25 resultant portfolios are calculated over the followingmonth, along with the returns of five high-minus-low portfolios that are long in stocks in theFNDRET quintile 5 and short in the SPFRET quintile 1 stocks for each INDRET quintile. Thisprocedure generates 30 time-series of portfolio returns. Using each time-series, FF5 alphas for all30 portfolios are calculated. Final rows in both panels report the average alphas for the SPFRETquintiles. All returns are in percentages. Newey-West t-statistics are reported in parentheses. *,**, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.

Panel A: Dependent sort.

SPFRET

(1) (2) (3) (4) (5) (5-1)

INDRET1 −1.265∗∗∗ −1.209∗∗∗ −0.923∗∗∗ −0.750∗∗∗ −0.594∗∗∗ 0.671∗∗∗

(−8.034) (−8.597) (−6.446) (−4.742) (−3.691) (5.146)

INDRET2 −0.664∗∗∗ −0.527∗∗∗ −0.366∗∗∗ −0.202 −0.320∗∗∗ 0.344∗∗∗

(−4.724) (−4.304) (−3.197) (−1.545) (−3.594) (2.727)

INDRET3 −0.416∗∗∗ −0.172∗ −0.220∗ −0.086 −0.152∗ 0.264∗∗∗

(−4.326) (−1.836) (−1.950) (−1.093) (−1.718) (2.721)

INDRET4 −0.355∗∗∗ −0.070 −0.050 0.109 0.162 0.517∗∗∗

(−3.877) (−0.695) (−0.496) (1.068) (1.471) (4.080)

INDRET5 0.231∗∗ 0.275∗∗ 0.443∗∗∗ 0.559∗∗∗ 0.834∗∗∗ 0.603∗∗∗

(1.985) (2.038) (3.095) (3.579) (4.651) (4.036)

Average -0.494∗∗∗ -0.341∗∗∗ -0.223∗∗∗ -0.074 -0.014 0.480∗∗∗

(-8.532) (-6.027) (-3.992) (-1.326) (-0.239) (8.418)

29

Panel B: Independent sort.

SPFRET

(1) (2) (3) (4) (5) (5-1)

INDRET1 −1.158∗∗∗ −0.731∗∗∗ −0.420∗∗ −0.517∗∗ 0.275 1.390∗∗∗

(−8.403) (−5.342) (−1.976) (−2.316) (1.049) (5.197)

INDRET2 −0.684∗∗∗ −0.380∗∗∗ −0.367∗∗∗ −0.300∗∗ −0.126 0.544∗∗∗

(−4.967) (−3.305) (−3.382) (−2.113) (−0.644) (2.697)

INDRET3 −0.440∗∗∗ −0.295∗∗∗ −0.243∗∗∗ −0.085 0.007 0.432∗∗

(−3.018) (−2.716) (−3.119) (−0.896) (0.045) (2.217)

INDRET4 −0.168 −0.370∗∗∗ −0.151 −0.010 0.152 0.330(−0.832) (−2.887) (−1.397) (−0.126) (1.353) (1.521)

INDRET5 −0.189 −0.045 0.125 0.288∗∗ 0.621∗∗∗ 0.797∗∗

(−0.607) (−0.218) (0.853) (2.275) (4.187) (2.113)

Average -0.527∗∗∗ -0.364∗∗∗ -0.211∗∗∗ -0.125∗ 0.186∗∗ 0.699∗∗∗

(-6.016) (-5.493) (-3.331) (-1.935) (2.264) (6.170)

30

Table 4: Social Ties and Industry Momentum

We investigate whether social-peer firm returns contain incremental information about focalfirms relative to industry momentum effects. At the end of month t, we sort stocks from the sameindustry as the focal stock into two portfolios based on the SCI between their county and the focalcounty while using the median SCI as the breakpoint. IND_HIGH denotes the equal-weightedreturn of the high-SCI portfolio. In columns 1 to 3, all stocks except the focal stock are used whileforming the portfolios, while only the stocks of firms headquartered out of the focal state areused in columns 4 through 6. In Panel A, the dependent variable is the contemporaneous (i.e.,month t) focal stock return. In Panel B, it is the following month’s focal stock return. In bothpanels, covariates include the basic control variables described in Section 2.3 (RET, SIZE, BMKT,BM, IVOL, ILLIQ, MAX, SKEW and COSKEW). All values are in percentages. All independentvariables are standardized cross-sectionally and winsorized at 1% and 99%. t-statistics withstandard errors based on Newey and West (1994) adjustments are reported in parentheses. *, **,and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.

Panel A: Return co-movement.

RETtAll Stocks Out-of-State Stocks

(1) (2) (3) (4) (5) (6)

INDRET 1.207∗∗∗ 0.729∗∗∗ 1.122∗∗∗ 0.717∗∗∗

(21.954) (22.293) (20.124) (19.393)

IND_HIGH 1.185∗∗∗ 0.649∗∗∗ 1.085∗∗∗ 0.511∗∗∗

(18.510) (11.848) (17.832) (9.337)

Controls Yes Yes Yes Yes Yes Yes# Periods 673 673 673 673 673 673# Stocks 1,963 1,951 1,951 1,963 1,950 1,950R2 0.353 0.353 0.356 0.352 0.351 0.354

31

Panel B: Return predictability.

RETt+1All Stocks Out-of-State Stocks

(1) (2) (3) (4) (5) (6)

INDRET 0.432∗∗∗ 0.285∗∗∗ 0.406∗∗∗ 0.266∗∗∗

(9.457) (6.867) (9.740) (6.404)

IND_HIGH 0.428∗∗∗ 0.225∗∗∗ 0.406∗∗∗ 0.194∗∗∗

(9.793) (6.860) (9.459) (5.928)

Controls Yes Yes Yes Yes Yes Yes# Periods 672 672 672 672 672 672# Stocks 1,961 1,950 1,950 1,961 1,948 1,948R2 0.083 0.083 0.086 0.083 0.083 0.085

32

Table 5: Return Predictability of Social-Peer Firm Returns: Cross-Section Regressions

The table reports the results of Fama-MacBeth regressions where the focal stock return in nextmonth is regressed on SPFRET, along with short-term industry momentum (INDRET), long-termindustry momentum (INDMOM), geographic lead-lag effect (GEORET), shared analyst coveragelead-lag effect (CFRET), and customer-supplier lead-lag effect (CRET). We also include a batteryof well known cross-sectional predictive variables, including RET, SIZE, BMKT, BM, MOM,IVOL, ILLIQ, MAX, SKEW, and COSKEW (see 2.3 for detailed descriptions). For brevity, wedo not report their coefficients. All values are in percentages. All independent variables arestandardized cross-sectionally and winsorized at 1% and 99%. t-statistics with standard errorsbased on Newey and West (1994) adjustments are reported in parentheses. *, **, and *** indicatestatistical significance at the 10%, 5%, and 1% levels, respectively.

RETt+1

(1) (2) (3) (4) (5) (6)

SPFRET 0.420∗∗∗ 0.212∗∗∗ 0.213∗∗∗ 0.143∗∗∗ 0.195∗∗∗ 0.128∗∗

(9.467) (7.244) (6.972) (3.789) (2.924) (2.057)

INDRET 0.295∗∗∗ 0.278∗∗∗ 0.082∗ 0.217∗∗∗ 0.082(7.364) (6.739) (1.805) (2.578) (1.081)

INDMOM 0.093∗∗ 0.100∗∗ 0.088∗∗ 0.089 0.093(2.508) (2.550) (2.061) (1.467) (1.404)

GEORET 0.035∗∗ 0.018 −0.017 −0.032(2.098) (0.834) (−0.342) (−0.595)

CFRET 0.501∗∗∗ 0.416∗∗∗

(7.727) (4.612)

CRET 0.240∗∗∗ 0.190∗∗∗

(5.415) (3.715)

Controls Yes Yes Yes Yes Yes Yes# Periods 672 672 672 453 492 445# Stocks 1,951 1,951 1,689 1,942 368 353R2 0.082 0.090 0.093 0.074 0.133 0.129

33

Table 6: Predictability Hetereogeneity

The table reports Fama-MacBeth regression results for subsamples of data with different focalstock characteristics. At the end of month t, we sort all common stocks into two sub-samples onthree different dimensions: size, institutional ownership and analyst coverage. For size sorting,we use NYSE breakpoints. For each subsample, next month’s focal stock return is regressed onSPFRET, along with short-term industry momentum (INDRET), long-term industry momentum(INDMOM), geographic lead-lag effect (GEORET), common analyst coverage lead-lag effect(CFRET), and customer-supplier lead-lag effect (CRET). We also include a battery of well knowncross-sectional predictive variables, including RET, SIZE, BMKT, BM, MOM, IVOL, ILLIQ,MAX, SKEW, COSKEW (see Section 2.3 for detailed descriptions). All values are in percentages.All independent variables are winsorized at 1% and 99% and standardized cross-sectionally.Newey-West t-statistics are reported in parentheses. *, **, and *** indicate statistical significanceat the 10%, 5%, and 1% levels, respectively.

Size Inst. Own. Analyst Cov.

Small Large Low High Low High(1) (2) (3) (4) (5) (6)

SPFRET 0.193∗∗∗ 0.081 0.159∗∗∗ 0.096∗∗ 0.240∗∗∗ 0.056(3.545) (1.572) (3.086) (2.525) (4.671) (1.139)

INDRET 0.129∗∗ −0.040 0.187∗∗∗ −0.032 0.006 −0.009(2.360) (−0.735) (2.704) (−0.720) (0.108) (−0.148)

INDMOM 0.143∗∗ 0.066 0.103∗ 0.017 0.083∗∗ 0.054(2.375) (1.335) (1.899) (0.361) (2.033) (0.964)

GEORET −0.0002 0.067∗∗∗ 0.023 0.044∗ 0.037 0.018(−0.006) (2.639) (0.968) (1.691) (1.070) (0.629)

CFRET 0.545∗∗∗ 0.385∗∗∗ 0.485∗∗∗ 0.470∗∗∗ 0.529∗∗∗ 0.506∗∗∗

(7.834) (6.300) (7.027) (7.356) (7.375) (6.756)

Controls Yes Yes Yes Yes Yes Yes# Periods 453 453 453 453 441 441# Stocks 1,355 587 962 962 700 699R2 0.069 0.130 0.086 0.104 0.088 0.127

34

Table 7: Long-Run Return Predictability

The table reports the long-run returns of univariate portfolios sorted on social-peer firm returns.In Columns 1 to 3 of Panel A, all focal stocks are sorted into deciles at the end of month t basedon their SPFRET, the SCI-weighted average month-t return of stocks that are from the same FF48industry as the focal stock, but from a different state. Future long-horizon characteristics-basedcumulative abnormal returns (CAR) of decile 1 portfolio (Low), decile 10 portfolio (High) and theportfolio that is long decile 10 and short decile 1 (High− Low) are then calculated. CAR12 CAR24and CAR60 are the CARs calculated over 12-, 24 and 60-month horizons respectively. CAR overhorizon h is defined as the sum of monthly abnormal returns from month t + 1 to t + h, where theabnormal stock return for month t is defined as the difference between the stock return and thereturn of the corresponding DGTW portfolio in t (Daniel et al., 1997; Wermers, 2003). In Columns4 to 6, sorting is done with respect to SPFMOM, the compound SPFRET between t− 11 and t− 1.Panel B reports the results of Fama-MacBeth regressions where future long-horizon CARs areregressed on SPFRET and SPFMOM along with various other predictors. INDRET is the month-tequal-weighted average return of stocks from the same industry as the focal stock, INDMOM isthe compound INDRET between t− 11 and t− 1, GEORET is the month-t equal-weighted averagereturn of all stocks from the same economic area as the focal stock but from a different industry,GEOMOM is the compound GEORET from t− 11 to t− 1 and CFRET and CFMOM are the month-t weighted average return and momentum of stocks that share at least one analyst with the focalstock over previous 12 months, where weights are the number of shared analysts between stocks.Other control variables include RET, SIZE, BMKT, BM, MOM, IVOL, ILLIQ, MAX, SKEW andCOSKEW (see Section 2.3 for detailed descriptions). All values are in percentages. All independentvariables are winsorized at 1% and 99% and standardized cross-sectionally in Panel B. Newey-West heteroskedasticity and autocorrelation-corrected t-statistics are reported in parentheses. *,**, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.

35

Panel A: Portfolio sort on SPFRET and SPFMOM.

SPFRET SPFMOM

Low High High-Low Low High High-Low(1) (2) (3) (4) (5) (6)

CAR12 2.270∗∗∗ 5.324∗∗∗ 3.054∗∗∗ 2.422∗∗∗ 5.622∗∗∗ 3.201∗∗∗

(4.004) (7.157) (5.269) (3.402) (5.327) (2.727)

CAR24 6.877∗∗∗ 9.558∗∗∗ 2.681∗∗∗ 6.778∗∗∗ 9.532∗∗∗ 2.754∗

(6.129) (7.408) (3.947) (5.676) (5.838) (1.654)

CAR60 20.580∗∗∗ 24.834∗∗∗ 4.254∗∗∗ 19.326∗∗∗ 26.304∗∗∗ 6.977∗∗

(8.666) (8.411) (2.917) (8.842) (6.903) (1.985)

36

Panel B: Fama-MacBeth regression results.

CAR12 CAR24 CAR60

(1) (2) (3)

SPFRET 0.406∗∗ 0.586∗∗ 0.904∗∗∗

(2.466) (2.418) (2.990)

SPFMOM 1.079∗∗∗ 1.292∗ 2.358∗∗∗

(2.632) (1.899) (2.874)

INDRET 0.171 −0.379 −0.155(0.991) (−1.343) (−0.361)

INDMOM −0.970∗∗ −1.471∗∗ −0.600(−2.246) (−1.961) (−0.565)

GEORET 0.196∗ 0.114 0.162(1.899) (0.742) (0.748)

GEOMOM −0.024 −0.319 0.230(−0.103) (−0.949) (0.459)

CFRET 1.132∗∗∗ 1.109∗∗∗ 1.361∗∗∗

(6.053) (5.321) (4.282)

CFMOM 1.077∗∗ 0.859∗ 1.127(2.350) (1.735) (1.541)

Controls Yes Yes Yes# Periods 394 394 394# Stocks 1,495 1,495 1,495R2 0.069 0.072 0.080

37

Table 8: Social-Peer Firm Returns and Focal Firm Fundamentals

The table reports the panel regression results for long-horizon predictability in firm fundamen-tals due to SPFRET and SPFMOM. One-, two- and five-year cumulative value of three differentfinancial ratios are regressed on SPFRET, SPFMOM and the same control variables used in Ta-ble 7 (see Section 2.3 for a detailed description of the control variables), along with month fixedeffects. The financial ratios examined are: asset turnover (AT), earnings before interest, taxes, de-preciation and amortization (EBITDA) and gross profit (GP). AT is defined as the total sales scaledby the beginning-of-year total assets. EBITDA is calculated by subtracting the cost of goods sold(COGS) and selling, general and administrative expenses (SGAX) from the total revenue (TR). GPis calculated by subtracting the COGS from the total revenue. EBITDA and GP are scaled by thebeginning-of-year total assets. At the end of each month t, n-year cumulative value of funda-mentals for a stock is calculated as the arithmetic sum of the values of the fundamentals over thenext n financial years of the corresponding firm. All values are in percentages. All independentvariables are winsorized at 1% and 99% and standardized cross-sectionally. Standard errors aredouble-clustered on month and stock dimensions. t-statistics are reported in parentheses. *, **,and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.

38

Asset Turnover EBITDA Gross Profitability

1 Year 2 Years 5 Years 1 Year 2 Years 5 Years 1 Year 2 Years 5 Years(1) (2) (3) (4) (5) (6) (7) (8) (9)

SPFRET 4.293∗∗∗ 7.040∗∗∗ 14.967∗∗∗ 0.338∗∗∗ 0.518∗∗∗ 1.024∗∗∗ 1.840∗∗∗ 3.142∗∗∗ 6.983∗∗∗

(4.154) (4.199) (4.171) (3.386) (3.177) (3.065) (5.944) (6.328) (6.683)

SPFMOM 15.589∗∗∗ 25.841∗∗∗ 54.652∗∗∗ 1.017∗∗∗ 1.561∗∗∗ 3.236∗∗∗ 6.472∗∗∗ 11.244∗∗∗ 25.075∗∗∗

(8.963) (9.022) (8.792) (5.466) (4.950) (4.834) (11.810) (12.229) (12.505)

INDRET −4.315∗∗∗ −6.693∗∗∗ −13.187∗∗∗ −0.272∗∗ −0.401∗∗ −0.671∗ −1.638∗∗∗ −2.674∗∗∗ −5.574∗∗∗

(−3.618) (−3.443) (−3.152) (−2.439) (−2.172) (−1.753) (−4.403) (−4.416) (−4.311)

INDMOM −15.037∗∗∗ −24.217∗∗∗ −50.044∗∗∗ −1.098∗∗∗ −1.638∗∗∗ −3.162∗∗∗ −6.091∗∗∗ −10.289∗∗∗ −22.272∗∗∗

(−7.641) (−7.533) (−7.325) (−5.738) (−5.099) (−4.687) (−10.146) (−10.409) (−10.550)

GEORET 0.464 0.788 1.643 0.051 0.069 0.094 0.170 0.236 0.432(1.176) (1.172) (1.081) (0.920) (0.721) (0.469) (1.460) (1.249) (1.073)

GEOMOM 0.883 1.430 3.359 0.029 −0.056 −0.057 0.163 0.142 0.518(1.350) (1.278) (1.369) (0.335) (−0.378) (−0.178) (0.794) (0.421) (0.711)

CFRET 0.705 0.955 1.284 −0.0005 −0.036 −0.286 0.260 0.368 0.432(0.893) (0.728) (0.449) (−0.004) (−0.200) (−0.760) (0.903) (0.783) (0.433)

CFMOM 2.659∗∗∗ 3.261∗∗ 4.880 0.243 0.239 0.006 1.370∗∗∗ 1.972∗∗∗ 3.411∗∗∗

(2.747) (1.992) (1.393) (1.607) (0.946) (0.013) (4.117) (3.575) (2.929)

Controls Yes Yes Yes Yes Yes Yes Yes Yes YesMonth FE Yes Yes Yes Yes Yes Yes Yes Yes YesObservations 507,451 507,451 507,451 507,451 507,451 507,451 507,451 507,451 507,451R2 0.083 0.077 0.072 0.116 0.110 0.096 0.142 0.147 0.143

39

Table 9: Social-Peer Firm Returns, Focal Firms’ Earnings Surprises, andAnalyst Forecast Errors

The table presents the panel regression results of predicting earnings surprises and forecast errorsby SPFRET and SPFMOM. In columns 1 to 3, cumulative standardised unexpected earnings(SUE), calculated by using a random-walk model, is regressed on SPFRET, SPFMOM along withthe same control variables used in Table 8 and month fixed effects. In columns 4 to 6, cumulativeanalyst forecast error is used as the dependent variable. Details of how the dependent variablesare calculated are given in Section 4. The length of the horizon over which the cumulation is doneis reported over each column. All values are in percentages. All independent variables are win-sorized at 1% and 99% and standardized cross-sectionally. Standard errors are double-clusteredon month and stock dimensions. t-statistics are reported in parentheses. *, **, and *** indicatestatistical significance at the 10%, 5%, and 1% levels, respectively.

Cumulative SUE Cumulative Analyst Forecast Errors

1 Year 2 Years 5 Years 1 Year 2 Years 5 Years(1) (2) (3) (4) (5) (6)

SPFRET 0.029∗ 0.075∗∗∗ 0.146∗∗ 0.043∗∗ 0.074∗∗ 0.012(1.802) (2.653) (2.572) (2.074) (2.132) (0.202)

SPFMOM 0.132∗∗∗ 0.275∗∗∗ 0.582∗∗∗ 0.146∗∗∗ 0.116 0.057(3.593) (3.980) (3.955) (2.886) (1.163) (0.323)

INDRET 0.014 −0.038 −0.084 −0.030 −0.060 −0.025(0.771) (−1.182) (−1.275) (−1.103) (−1.587) (−0.441)

INDMOM −0.117∗∗∗ −0.308∗∗∗ −0.559∗∗∗ −0.036 −0.031 0.108(−3.202) (−4.574) (−4.009) (−0.797) (−0.375) (0.769)

GEORET 0.011 0.002 −0.007 0.013 0.023 0.062∗

(1.153) (0.106) (−0.267) (0.958) (1.021) (1.876)

GEOMOM −0.011 −0.044 −0.045 0.014 0.008 0.062(−0.551) (−1.200) (−0.674) (0.477) (0.180) (0.706)

CFRET 0.061∗∗∗ 0.088∗∗∗ 0.059 0.018 0.023 0.035(3.898) (3.186) (1.169) (0.897) (0.882) (0.893)

CFMOM 0.241∗∗∗ 0.308∗∗∗ 0.304∗∗∗ −0.047 −0.016 −0.165(8.299) (6.409) (3.523) (−1.246) (−0.282) (−1.594)

Controls Yes Yes Yes Yes Yes YesMonth FE Yes Yes Yes Yes Yes YesObservations 560,855 560,855 560,855 43,936 43,936 43,936R2 0.097 0.087 0.069 0.066 0.071 0.079

40

Appendix

1

Table A1: Univariate Sorting (Value-Weighted Portfolios)

The table reports the results of the univariate portfolio sorts based on SPFRET and SPFRET⊥, where portfolios are value-weighted.SPFRET (social-peer firm return) is the SCI-weighted average return of stocks that are from the same FF48 industry as the focal stock,but from a different state. SPFRET⊥ is the abnormal return generated from a panel regression of SPFRET on the equal-weighted returnof stocks from the same FF48 industry as the focal stock, using all available stock-month observations prior to the portfolio formationmonth (i.e., month t) and month fixed effects. At the end of each month from January 1962 to December 2019, all common stocks aresorted into deciles based on either SPFRET (Panel A) or SPFRET⊥ (Panel B). Then, value-weighted return for each portfolio, along withthe return of a high-minus-low portfolio obtained through going long in the decile-10 portfolio and short in the decile-1 portfolio arecalculated over the following month. The first row in each panel shows the raw returns, second row shows the characteristics-basedalphas computed by using DGTW adjustments (see Daniel et al., 1997; Wermers, 2003), and the third row reports the Fama-French5-factor (FF5) alphas. All returns are in percentages. The corresponding t-statistics based on Newey and West (1994) standard er-rors are reported in parentheses below the alphas. *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.

Panel A: Sorting by SPFRET.

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (10-1)

Raw Return 0.397∗ 0.508∗∗∗ 0.693∗∗∗ 0.939∗∗∗ 0.873∗∗∗ 0.880∗∗∗ 0.896∗∗∗ 1.071∗∗∗ 1.113∗∗∗ 1.062∗∗∗ 0.665∗∗∗

(1.943) (2.625) (3.597) (5.110) (4.640) (4.708) (5.051) (5.628) (5.770) (5.380) (4.480)

DGTW Alpha −0.225∗∗∗ −0.185∗∗ −0.053 0.119∗∗ 0.052 0.056 0.042 0.235∗∗∗ 0.224∗∗∗ 0.218∗∗∗ 0.443∗∗∗

(−3.421) (−2.510) (−0.924) (2.120) (0.805) (1.029) (0.853) (3.985) (3.024) (3.154) (4.215)

FF5 Alpha −0.575∗∗∗ −0.432∗∗∗ −0.303∗∗∗ −0.013 −0.134 −0.132 −0.082 0.115 0.227∗ 0.134 0.709∗∗∗

(−5.208) (−3.913) (−3.544) (−0.120) (−1.518) (−1.378) (−1.204) (1.207) (1.940) (1.181) (4.394)

Panel B: Sorting by SPFRET⊥.

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (10-1)

Raw Return 0.490∗∗ 0.651∗∗∗ 0.641∗∗∗ 0.785∗∗∗ 0.697∗∗∗ 0.934∗∗∗ 1.077∗∗∗ 1.057∗∗∗ 0.989∗∗∗ 1.030∗∗∗ 0.541∗∗∗

(2.541) (3.356) (3.463) (4.420) (3.663) (4.994) (5.669) (5.593) (5.397) (6.135) (4.531)

DGTW Alpha −0.172∗∗∗ −0.026 −0.056 −0.026 −0.063 0.121∗ 0.209∗∗∗ 0.191∗∗∗ 0.143∗∗∗ 0.125∗∗ 0.298∗∗∗

(−2.752) (−0.490) (−0.959) (−0.656) (−1.127) (1.956) (3.262) (3.536) (2.666) (2.377) (3.712)

FF5 Alpha −0.495∗∗∗ −0.315∗∗∗ −0.337∗∗∗ −0.158∗∗ −0.282∗∗∗ −0.014 0.107 0.095 0.058 0.164∗ 0.658∗∗∗

(−5.005) (−3.381) (−3.723) (−2.393) (−2.780) (−0.166) (1.215) (1.136) (0.665) (1.934) (4.644)

2

Table A2: Bivariate Sorting / Orthogonalized FNDRETThe table reports the results of bivariate portfolio sort based on SPFRET⊥, as defined in Table A1,and INDRET, the equal-weighted return of stocks in the same FF48 industry as the focal firm. InPanel A, we conduct sequential sort, where stocks are first sorted into quintiles based on theirINDRET and then, within each INDRET quintile and then further sorted into quintiles usingSPFRET breakpoints for that INDRET quintile only. In Panel B, second sorting is done by usingSPFRET breakpoints for all stocks from that specific month (independent sorting). After sorting,equal-weighted return for each of the 25 resultant portfolios are calculated over the followingmonth, along with the returns of five high-minus-low portfolios that are long in stocks in theSPFRET quintile 5 and short in the SPFRET quintile 1 stocks for each INDRET quintile. Thisprocedure generates 30 time-series of portfolio returns. Using each time-series, FF5 alphas for all30 portfolios are calculated. Final rows in both panels report the average alphas for the SPFRETquintiles. All returns are in percentages. Newey-West t-statistics are reported in parentheses. *,**, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.

Panel A: Dependent sort.

SPFRET⊥(1) (2) (3) (4) (5) (5-1)

INDRET1 −1.180∗∗∗ −1.163∗∗∗ −0.984∗∗∗ −0.783∗∗∗ −0.631∗∗∗ 0.549∗∗∗

(−7.724) (−7.998) (−5.871) (−5.078) (−3.919) (4.092)

INDRET2 −0.645∗∗∗ −0.472∗∗∗ −0.400∗∗∗ −0.243∗∗ −0.321∗∗∗ 0.324∗∗

(−4.565) (−3.831) (−3.123) (−2.149) (−3.196) (2.508)

INDRET3 −0.408∗∗∗ −0.207∗∗ −0.174∗∗ −0.109 −0.147 0.261∗∗∗

(−3.610) (−2.372) (−1.969) (−1.299) (−1.640) (2.678)

INDRET4 −0.301∗∗∗ −0.095 −0.074 0.174∗ 0.089 0.390∗∗∗

(−3.034) (−0.914) (−0.711) (1.750) (0.861) (3.387)

INDRET5 0.218∗ 0.402∗∗ 0.385∗∗∗ 0.597∗∗∗ 0.740∗∗∗ 0.522∗∗∗

(1.664) (2.441) (2.723) (3.833) (5.080) (4.274)

Average -0.455∗∗∗ -0.313∗∗∗ -0.234∗∗∗ -0.154∗∗∗ -0.016 0.445∗∗∗

(-7.497) (-5.604) (-4.277) (-2.751) (-0.281) (7.587)

3

Panel B: Independent sort.

SPFRET⊥(1) (2) (3) (4) (5) (5-1)

INDRET1 −1.077∗∗∗ −1.091∗∗∗ −0.974∗∗∗ −0.698∗∗∗ −0.484∗∗∗ 0.612∗∗∗

(−6.888) (−7.587) (−6.977) (−5.256) (−3.125) (4.255)

INDRET2 −0.688∗∗∗ −0.485∗∗∗ −0.452∗∗∗ −0.266∗∗ −0.297∗∗∗ 0.387∗∗∗

(−5.176) (−4.025) (−3.637) (−2.242) (−2.665) (2.852)

INDRET3 −0.378∗∗∗ −0.321∗∗∗ −0.091 −0.236∗∗∗ −0.140 0.240∗∗

(−3.773) (−3.447) (−1.028) (−2.900) (−1.388) (1.996)

INDRET4 −0.313∗∗∗ −0.036 −0.042 −0.072 0.093 0.419∗∗∗

(−3.555) (−0.350) (−0.439) (−0.837) (0.601) (2.913)

INDRET5 0.181 0.363∗∗ 0.392∗∗∗ 0.503∗∗∗ 0.747∗∗∗ 0.566∗∗∗

(1.418) (2.422) (2.957) (3.132) (4.937) (4.277)

Average -0.463∗∗∗ -0.307∗∗∗ -0.249∗∗∗ -0.073 -0.054 0.409∗∗∗

(-8.039) (-5.36) (-4.326) (-1.305) (-0.979) (7.606)

4

Table A3: Return Predictability of Peer Firms’ Returns: Regression Anal-ysis / Orthogonalized FNDRET

The table reports the results of Fama-MacBeth regressions where the focal stock return in nextmonth is regressed on SPFRET⊥ (i.e.,the orthogonalized SPRFRET), along with short-termindustry momentum (INDRET), long-term industry momentum (INDMOM), geographic lead-lageffect (GEORET), common analyst coverage lead-lag effect (CFRET), and customer-supplierlead-lag effect (CRET). We also include a battery of well known cross-sectional predictivevariables, including RET, SIZE, BMKT, BM, MOM, IVOL, ILLIQ, MAX, SKEW, and COSKEW (see2.3 for detailed descriptions). For brevity, we do not report their coefficients. In each column,orthogonalization of SPFRET is performed by running a panel regression of SPFRET on RET,MOM, the covariates that are displayed for that column and month fixed effects, using allavailable stock-month observations prior to month t. All independent variables are standardizedcross-sectionally and winsorized at 1% and 99%. t-statistics with standard errors based onNewey and West (1994) adjustments are reported in parentheses. *, **, and *** indicate statisticalsignificance at the 10%, 5%, and 1% levels, respectively.

RETt+1

(1) (2) (3) (4) (5) (6)

SPFRET⊥ 0.423∗∗∗ 0.157∗∗∗ 0.161∗∗∗ 0.098∗∗∗ 0.152∗∗∗ 0.085∗

(9.932) (7.416) (7.314) (3.730) (3.366) (1.880)

INDRET 0.437∗∗∗ 0.417∗∗∗ 0.143∗∗∗ 0.355∗∗∗ 0.157∗∗

(9.923) (9.656) (4.088) (4.657) (2.484)

INDMOM 0.100∗∗∗ 0.109∗∗∗ 0.093∗∗ 0.099 0.064(2.653) (2.814) (2.145) (1.599) (1.030)

GEORET 0.033∗ 0.021 −0.011 −0.042(1.960) (0.958) (−0.225) (−0.787)

CFRET 0.521∗∗∗ 0.406∗∗∗

(7.543) (4.259)

CRET 0.253∗∗∗ 0.167∗∗∗

(5.634) (3.555)

Controls Yes Yes Yes Yes Yes Yes# Periods 661 661 661 432 481 424# Stocks 1,981 1,981 1,716 2,005 375 366R2 0.082 0.088 0.089 0.069 0.124 0.121

5

Table A4: Predictability Hetereogeneity / Orthogonalized FNDRET

The table reports Fama-MacBeth regression results with orthogonalized SPFRET for subsamplesof data with different focal stock characteristics. At the end of month t, all common stocks aresorted into two sub-samples on three different dimensions: size, institutional ownership andanalyst coverage. For size sorting, NYSE breakpoints are used. For each subsample, next month’sfocal stock return is regressed on SPFRET⊥, as defined in Table A2, along with short-termindustry momentum (INDRET), long-term industry momentum (INDMOM), geographic lead-lageffect (GEORET), common analyst coverage lead-lag effect (CFRET), and customer-supplierlead-lag effect (CRET). We also include a battery of well known cross-sectional predictivevariables, including RET, SIZE, BMKT, BM, MOM, IVOL, ILLIQ, MAX, SKEW, COSKEW (seeSection 2.3 for detailed descriptions). All values are in percentages. All independent variablesare winsorized at 1% and 99% and standardized cross-sectionally. Newey-West t-statistics arereported in parentheses. *, **, and *** indicate statistical significance at the 10%, 5%, and 1%levels, respectively.

Size Inst. Own. Analyst Cov.

Small Large Low High Low High(1) (2) (3) (4) (5) (6)

SPFRET⊥ 0.139∗∗∗ 0.063∗ 0.123∗∗∗ 0.068∗∗∗ 0.176∗∗∗ 0.037(4.300) (1.943) (3.588) (2.579) (5.153) (1.100)

INDRET 0.258∗∗∗ 0.012 0.294∗∗∗ 0.030 0.163∗∗∗ 0.025(5.894) (0.268) (5.888) (0.802) (4.326) (0.517)

INDMOM 0.143∗∗ 0.068 0.103∗ 0.015 0.082∗∗ 0.056(2.365) (1.352) (1.895) (0.325) (2.011) (0.992)

GEORET 0.0002 0.067∗∗∗ 0.023 0.044∗ 0.037 0.018(0.006) (2.675) (0.971) (1.679) (1.056) (0.643)

CFRET 0.545∗∗∗ 0.385∗∗∗ 0.485∗∗∗ 0.471∗∗∗ 0.529∗∗∗ 0.505∗∗∗

(7.813) (6.290) (7.015) (7.339) (7.369) (6.744)

Controls Yes Yes Yes Yes Yes Yes# Periods 453 453 453 453 441 441# Stocks 1,355 587 962 962 700 699R2 0.069 0.130 0.086 0.104 0.088 0.127

6

Table A5: Long-Run Return Predictability / Orthogonalized SPFRETand SPFMOM

The table reports the univariate portfolio sort results on the long-run return predictability due tosocial connectedness using orthogonalized SPFRET and SPFMOM. In Columns 1 to 3 of Panel A,all focal stocks are sorted into deciles at the end of month t based on the orthogonalized SPFRET inmonth t, denoted by SPFRET⊥, which, for stock i at month t, is defined as SPFRETi,t − β̂

′tXi,t, β̂t

being the vector of coefficient estimates for the panel regression SPFRETi,s = α + β′tXi,s + θs + εs

for all s ≤ t, where θs is the fixed effect for month s. Elements of Xi,s, the vector of covariates forstock i as measured at the end of month t, include stock’s own return, equal-weighted industryreturn, 12- to 60-month stock momenta and 12- to 60-month equal-weighted industry momenta.Month t is excluded from momentum calculations. Future long-horizon characteristics-based cu-mulative abnormal returns (CAR) of decile 1 porfolio (Low), decile 10 portfolio (High) and theportfolio that is long decile 10 and short decile 1 (High− Low) are then calculated. CAR12 throughCAR60 are the CARs calculated over horizons from 12 months to 60 months. CAR over horizon his defined as the sum of monthly abnormal returns from month t + 1 to t + h, where the abnormalstock return for month t is defined as the difference between the stock return and the return ofthe corresponding characteristics-matched portfolio in t (Daniel et al., 1997; Wermers, 2003).Incolumns 4 to 6, sorting is done with respect to the orthogonalized, SPFMOM, the compoundSPFRET between t− 11 and t− 1. Orthogonalization method is identical to that for SPFRET, ex-cept that variables that are measured at the end of month t are excluded from the covariate vectorto prevent look-ahead bias. Panel B reports the results of Fama-MacBeth regressions where futurelong-horizon CARs are regressed on SPFRET⊥ and SPFMOM⊥ along with various other predic-tors. INDRET is the month-t equal-weighted average return of stocks from the same FF48 industryas the focal stock, INDMOM is the compound INDRET between t− 11 and t− 1, GEORET is themonth-t equal-weighted average return of all stocks from the same economic area as the focalstock but from a different industry, GEOMOM is the compound GEORET from t− 11 to t− 1 andCFRET and CFMOM are the month-t weighted average return and momentum of stocks that shareat least one analyst with the focal stock over previous 12 months, where weights are the numberof shared analysts between stocks. Other control variables include RET, SIZE, BMKT, BM, MOM,IVOL, ILLIQ, MAX, SKEW and COSKEW (see Section 2.3 for detailed descriptions). All valuesare in percentages. Stocks that do not have 60 months of return data post portfolio formation arealso excluded. All independent variables are winsorized at 1% and 99% and standardized cross-sectionally in Panel B, Newey-West heteroskedasticity and autocorrelation-corrected t-statisticsare reported in parentheses. *, **, and *** indicate statistical significance at the 10%, 5%, and 1%levels, respectively.

7

Panel A: Portfolio sort on SPFRET⊥ and SPFMOM⊥.

SPFRET⊥ SPFMOM⊥Low High High-Low Low High High-Low(1) (2) (3) (4) (5) (6)

CAR12 5.758∗∗∗ 6.507∗∗∗ 0.749∗ 4.846∗∗∗ 7.377∗∗∗ 2.530∗∗

(8.147) (8.425) (1.914) (6.056) (6.967) (2.068)

CAR24 11.446∗∗∗ 13.170∗∗∗ 1.724∗∗ 10.884∗∗∗ 14.179∗∗∗ 3.295(9.007) (8.524) (2.136) (8.096) (6.588) (1.351)

CAR60 28.598∗∗∗ 32.050∗∗∗ 3.452∗∗∗ 26.532∗∗∗ 35.141∗∗∗ 8.609∗∗∗

(12.165) (10.868) (3.711) (13.728) (10.572) (3.713)

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

8

Panel B: Fama-MacBeth regression results.

CAR12 CAR24 CAR60

(1) (2) (3)

SPFRET⊥ 0.222∗∗ 0.470∗∗∗ 0.689∗∗∗

(2.424) (2.669) (3.221)

SPFMOM⊥ 0.896∗∗∗ 1.121∗∗ 2.031∗∗∗

(2.916) (1.982) (3.886)

INDRET 0.488∗∗∗ 0.101 0.488(3.167) (0.523) (1.484)

INDMOM −0.125 −0.578 1.199(−0.479) (−1.233) (1.297)

GEORET 0.196∗∗ 0.138 0.079(2.204) (1.051) (0.381)

GEOMOM −0.012 −0.299 0.040(−0.055) (−0.941) (0.091)

CFRET 1.131∗∗∗ 1.184∗∗∗ 1.383∗∗∗

(6.297) (5.897) (4.651)

CFMOM 1.089∗∗ 0.964∗ 1.066(2.493) (1.953) (1.249)

Controls Yes Yes Yes# Periods 373 373 373# Stocks 1,317 1,317 1,317R2 0.070 0.071 0.081

9

Table A6: The Prediction of Long-Run Firm Fundamentals / Orthogonal-ized FNDRET and FNDMOM

The table reports the panel regression results for long-horizon predictability in firm fundamentalsdue to SPFRET⊥ and SPFMOM⊥, as defined in Table A5. One-, two- and five-year cumulativevalue of three different financial ratios are regressed on SPFRET⊥, SPFRET⊥ and the same controlvariables used in Table A5 (see Section 2.3 for a detailed description of the control variables),along with month fixed effects. The financial ratios examined are: asset turnover (AT), earningsbefore interest, taxes, depreciation and amortization (EBITDA) and gross profitability (GP). ATis defined as the total sales scaled by the beginning-of-year total assets. EBITDA is calculatedby subtracting the cost of goods sold (COGS) and selling, general and administrative expenses(SGAX) from the total revenue (TR). GP is calculated by subtracting the COGS from the totalrevenue. EBITDA and GP are scaled by the beginning-of-year total assets. At the end of eachmonth t, n-year cumulative value of fundamentals for a stock is calculated as the arithmetic sum ofthe values of the fundamentals over the next n financial years of the corresponding firm. All valuesare in percentages. All independent variables are winsorized at 1% and 99% and standardizedcross-sectionally. Standard errors are double-clustered on month and stock dimensions. t-statisticsare reported in parentheses. *, **, and *** indicate statistical significance at the 10%, 5%, and 1%levels, respectively.

10

Asset Turnover EBITDA Gross Profitability

1 Year 2 Years 5 Years 1 Year 2 Years 5 Years 1 Year 2 Years 5 Years(1) (2) (3) (4) (5) (6) (7) (8) (9)

SPFRET⊥ 3.055∗∗∗ 5.127∗∗∗ 11.008∗∗∗ 0.256∗∗∗ 0.415∗∗∗ 0.859∗∗∗ 1.364∗∗∗ 2.376∗∗∗ 5.315∗∗∗

(3.856) (3.985) (3.964) (3.574) (3.577) (3.605) (5.755) (6.243) (6.556)

SPFMOM⊥ 9.227∗∗∗ 15.702∗∗∗ 33.720∗∗∗ 0.672∗∗∗ 1.074∗∗∗ 2.324∗∗∗ 3.934∗∗∗ 6.999∗∗∗ 15.947∗∗∗

(7.566) (7.679) (7.465) (5.003) (4.639) (4.623) (9.724) (10.083) (10.347)

INDRET −1.667 −2.452 −4.326 −0.014 −0.014 0.053 −0.442 −0.675 −1.218(−1.625) (−1.451) (−1.185) (−0.157) (−0.097) (0.176) (−1.392) (−1.301) (−1.100)

INDMOM −4.511∗∗∗ −6.650∗∗∗ −12.969∗∗∗ −0.272∗∗ −0.352∗ −0.582 −1.487∗∗∗ −2.283∗∗∗ −4.547∗∗∗

(−3.486) (−3.159) (−2.924) (−2.203) (−1.698) (−1.346) (−3.719) (−3.518) (−3.319)

GEORET 0.529 0.882 1.705 0.073 0.094 0.151 0.231∗∗ 0.316∗ 0.533(1.388) (1.367) (1.167) (1.605) (1.210) (0.959) (2.035) (1.736) (1.354)

GEOMOM 1.429∗∗ 2.420∗∗ 4.824∗ 0.122 0.120 0.315 0.343 0.436 0.940(2.180) (2.140) (1.922) (1.528) (0.864) (1.056) (1.621) (1.252) (1.248)

CFRET 1.519∗ 2.303 4.110 0.100 0.132 0.085 0.586∗ 0.923∗ 1.669(1.737) (1.604) (1.319) (0.936) (0.726) (0.228) (1.876) (1.830) (1.569)

CFMOM 4.745∗∗∗ 6.480∗∗∗ 11.235∗∗∗ 0.355∗∗ 0.417 0.396 1.951∗∗∗ 2.919∗∗∗ 5.563∗∗∗

(4.602) (3.761) (3.053) (2.305) (1.612) (0.741) (5.490) (4.981) (4.496)

Controls Yes Yes Yes Yes Yes Yes Yes Yes YesMonth FE Yes Yes Yes Yes Yes Yes Yes Yes YesObservations 430,207 430,207 430,207 430,207 430,207 430,207 430,207 430,207 430,207R2 0.082 0.077 0.072 0.128 0.122 0.107 0.143 0.150 0.147

11

Table A7: The Prediction of Future Earnings Surprises / OrthogonalizedFNDRET and FNDMOM

The table presents the panel regression results for earnings predictability by SPFRET⊥ andSPFMOM⊥, as defined in Table A5. In columns 1 to 3, cumulative standardised unexpectedearnings (SUE), calculated by using a random-walk model, is regressed on SPFRET⊥, SPFRET⊥along with the same control variables used in Table A5 and month fixed effects. In columns 4to 6, cumulative analyst forecast error is used as the dependent variable. Details of how thedependent variables are calculated are given in Section 4. The length of the horizon over whichthe cumulation is done is reported over each column. All values are in percentages. Stocks thatdo not have uninterrupted five years of earnings observations post month t are also excluded.All independent variables are winsorized at 1% and 99% and standardized cross-sectionally.Standard errors are double-clustered on month and stock dimensions. t-statistics are reportedin parentheses. *, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels,respectively.

Cumulative SUE Cumulative Analyst Forecast Errors

1 Year 2 Years 5 Years 1 Year 2 Years 5 Years(1) (2) (3) (4) (5) (6)

SPFRET⊥ 0.004 0.035∗ 0.080∗ 0.040∗∗∗ 0.060∗∗ 0.021(0.331) (1.676) (1.893) (2.641) (2.369) (0.530)

SPFMOM⊥ 0.066∗∗ 0.144∗∗∗ 0.320∗∗∗ 0.107∗∗∗ 0.088 0.006(2.442) (2.808) (2.941) (2.774) (1.283) (0.054)

INDRET 0.037∗∗ 0.011 0.002 −0.014 −0.026 −0.043(2.293) (0.410) (0.039) (−0.680) (−0.908) (−1.241)

INDMOM −0.030 −0.131∗∗∗ −0.199∗∗ 0.063∗ 0.041 0.112(−1.103) (−2.831) (−2.190) (1.728) (0.790) (1.086)

GEORET 0.009 0.008 0.002 0.013 0.019 0.055∗

(0.994) (0.516) (0.062) (1.032) (0.922) (1.753)

GEOMOM −0.003 −0.014 −0.009 0.017 0.011 0.040(−0.132) (−0.352) (−0.125) (0.540) (0.240) (0.475)

CFRET 0.073∗∗∗ 0.112∗∗∗ 0.103∗ 0.036∗ 0.044 0.050(4.540) (3.890) (1.911) (1.740) (1.602) (1.209)

CFMOM 0.258∗∗∗ 0.354∗∗∗ 0.380∗∗∗ −0.013 0.020 −0.115(8.452) (6.854) (4.033) (−0.346) (0.347) (−1.093)

Controls Yes Yes Yes Yes Yes YesMonth FE Yes Yes Yes Yes Yes YesObservations 468,976 468,976 468,976 40,733 40,733 40,733R2 0.101 0.091 0.071 0.069 0.074 0.079

12