Geographic Diffusion of Information and Stock Returns Jawad M. Addoum * University of Miami Alok Kumar University of Miami Kelvin Law Tilburg University October 21, 2013 Abstract This study shows that value-relevant information about publicly traded U.S. firms is geographically distributed and the market is slow in aggregating this information. Specifically, we demonstrate that the earnings and cash flows of firms can be predicted using the past performance of other firms in economically relevant geographical regions. However, sell-side equity analysts and institutional investors do not fully incorporate geographically dispersed information in their earnings forecasts and trades, respectively. Consequently, firms exhibit stronger post-earnings-announcement drift and stronger momentum in returns when geographic information is more dispersed and difficult to aggregate. A Long-Short trading strategy that exploits the slow diffusion of geographic information earns an annual, abnormal risk-adjusted return of about 9%. Keywords: Geographical connections, earnings predictability, return predictability, information diffusion, mispricing, sell-side analysts. * Please address all correspondence to Jawad M. Addoum, Department of Finance, School of Business Administration, 514 Jenkins Building, University of Miami, Coral Gables, FL 33124; Phone: 305-284-8286; Email: jawad.addoum@miami.edu. Alok Kumar can be reached at 305-284-1882 or akumar@miami.edu. Kelvin Law can be reached at (+31) 13-466-2219 or k.f.law@tilburguniversity.edu. We thank Major Coleman, George Korniotis and Chris Malloy and seminar participants at University of Miami, Tilburg University, and Syracuse University for helpful comments and valuable suggestions. We are responsible for all remaining errors and omissions.
Geographic Diffusion of Information and Stock Returns Abstract This study shows that value-relevant information about publicly traded U.S. firms is geographically distributed and the market is slow in aggregating this information. Specifically, we demonstrate that the earnings and cash flows of firms can be predicted using the past performance of other firms in economically relevant geographical regions. However, sell-side equity analysts and institutional investors do not fully incorporate geographically dispersed information in their earnings forecasts and trades, respectively. Consequently, firms exhibit stronger post-earnings-announcement drift and stronger momentum in returns when geographic information is more dispersed and difficult to aggregate. A Long-Short trading strategy that exploits the slow diffusion of geographic information earns an annual, abnormal risk-adjusted return of about 9%. Keywords: Geographical connections, earnings predictability, return predictability, information diffusion, mispricing, sell-side analysts.
1. Introduction A growing literature in finance demonstrates that value-relevant firm-specific information is distributed geographically (García and Norli (2012), Giroud (2013), Bernile, Kumar, and Sulaeman (2013)). In particular, this literature finds that the typical U.S. firm has geographical presence in six U.S. states and its business activities often occur in locations away from its corporate headquarters. Each of these economically-relevant non-headquarters locations may contain bits and pieces of information about the firm that may not be easily accessible and may be difficult to aggregate. Even firm insiders may have some difficulty accessing and aggregating this geographically dispersed information. In a recent study, Giroud (2013) shows that firm managers are able to improve their monitoring and information gathering activities when the plant locations become more accessible following the introduction of airline routes. If firm-specific value-relevant information is dispersed geographically, it is natural to ask how quickly this information gets aggregated into stock prices. Given the prior evidence in the literature, it is unlikely that market participants are able to aggregate the geographically dispersed information immediately. The frictions generated by physical distance between corporate headquarters and centers of economic activities relevant to the firm are likely to slow down the information aggregation process. This delay in the information aggregation process could subsequently generate predictable patterns in stock returns. To better understand our core idea, consider the following example of Walmart. While headquartered in Arkansas, Walmart has over 4,000 retail locations across the U.S. and it also has 158 distribution centers serving as distribution hubs for daily operations. Thus, Walmart s day-to-day economic activities and operations are often located in states other than its corporate headquarters. Walmart s customers, suppliers, research and development facilities, and other economic activities 1
are geographically dispersed. Since firms located in the same geographical regions are likely to experience common local economic shocks, Walmart s future performance would be correlated with the performance of firms located near its corporate headquarters (HQ) and economically connected (EC) states. If value-relevant information from non-hq economically relevant geographical locations is aggregated and incorporated into stock prices with delay, Walmart s performance and stock returns may be predictable. In this paper, we generalize this idea and test the potential asset-pricing implications of slow diffusion and delayed aggregation of geographically distributed firm-specific information. First, we demonstrate that the geographical connections of firms contain information that can be used to predict their future fundamentals. In particular, we find that the one-quarter-ahead earnings and cash flows of a firm are predictable based on the performance of other firms located around corporate headquarters and in regions that are economically relevant for the firm. To measure the economic interest of a firm in a given U.S. state, we follow García and Norli (2012) and conduct textual analysis of firms annual 10-K filings. Specifically, we capture the relative importance of a state to a firm by measuring the citation share of each U.S. state, which is defined as the number of times a U.S. state is mentioned in the 10-K filings divided by the total number of times all U.S. states are mentioned in those filings. A state with firm-level citation share of one (zero) indicates that the firm has all (none) of its economic activities located in that state. Using these firm-state-year citation share estimates of economically connected states, we measure the geographical dispersion in a firm s economic interests and estimate a series of Fama- MacBeth (1973) type predictive regressions. Specifically, we construct EC Earnings (EC Cash Flows) measures, defined as the citation-share weighted average earnings (cash flows) of firms located in economically relevant states, excluding the headquarters state. We also measure the earnings and cash flows of other firms located around firm headquarters. 2
Consistent with our conjecture, we find that the performance of other firms located in both HQ and EC states contain value-relevant information for the future earnings and cash flow of a given firm, even after we account for the lagged performance and fundamentals. Further, the evidence of predictability using information in EC states is stronger than predictive power of HQbased measures. We also examine the predictive power of geographical information over longer horizons, and find that our geography-based measures have predictive ability for up to eight quarters. Our evidence of cash flow predictability is incremental over the predictive ability of firmspecific lagged earnings measures (e.g., Fama and French (2000), Vuolteenaho (2002)). We also demonstrate that our results are not driven by state-level business cycles or firm characteristics, as all regressions include controls for state-level economic variables and firm fundamentals. In economic terms, a one standard deviation increase in the earnings (cash flows) of other firms in the HQ state in the current quarter predicts a 0.075% (0.197%) increase in the firm s earnings (cash flows) in the next quarter. 1 Relative to the average quarterly earnings (cash flows) of 0.9% (1.7%), these magnitudes are economically meaningful. Further, a one standard deviation increase in the earnings (cash flow) of firms headquartered in EC states in the current quarter predicts a 0.156% (0.468%) increase in their earnings (cash flows) next quarter. Again, compared to the average quarterly earnings (cash flows), these magnitudes are economically meaningful and stronger than the effects of HQ-based measures. In the next set of tests, we examine how quickly geographically-dispersed information is aggregated by market participants and subsequently reflected in stock prices. We first examine whether sell-side analysts help incorporate firms geographically dispersed information through their earnings forecasts. We find that equity analysts do not fully incorporate firms geographically 1 The economic magnitude is calculated as the coefficient estimate multiplied by its standard deviation: 0.050 0.015 = 0.075% or 0.048 0.041 = 0.197%. 3
dispersed earnings-relevant information into their forecasts. Instead of using dispersed geographical information, they form their forecasts primarily based on lagged firm earnings. As a result, their earnings forecasts do not fully account for the earnings information of other firms headquartered in economically connected states. Specifically, the lagged EC-based earnings measure significantly predicts analysts future forecast errors. A one standard deviation increase in lagged EC-Earnings corresponds to a 5.251% increase in analysts consensus forecast error in the next quarter. Further, we find that the explanatory power of EC-based earnings is higher than the lagged earnings or HQbased earnings measures. Since equity analysts are an important part of the price formation process and geographically dispersed information is not fully incorporated into their earnings forecasts, it is likely that geographic information diffuses slowly and generates predictable patterns in stock returns. We examine the effects of slower diffusion of information in two economic settings: post-earnings announcement drift (PEAD) and momentum in stock returns. Our choice is motivated by the evidence in previous studies, which suggest that slow diffusion of information may be a key mechanism that generates predictable patterns in stock returns in these two settings. We find that both post-earning-announcement drift and momentum returns are more pronounced among firms that have more geographically dispersed information that is difficult to aggregate. In both instances, we find that the slow diffusion of geographically dispersed information generates predictable patterns in stock returns. In economic terms, our estimates indicate that firms with above-median geographic dispersion have monthly momentum returns that are about 0.35% higher (t-statistic = 2.78) than firms with below-median dispersion. Similarly, we find that a one standard deviation increase in geographic dispersion is associated with 0.20 to 0.35% (t-statistics are between 2.37 to 2.72) higher post-earnings-announcement returns. This evidence is consistent with 4
our broad conjecture and shows that market participants do not fully incorporate geographically dispersed information. To quantify the economic magnitudes of these predictable return patterns, we construct outof-sample geography-based trading strategies. We use look-ahead bias-free forecasts of firm-level earnings-per-share (EPS) and sort stocks based on their Expected Earnings Surprise, which is defined as the forecasted EPS minus the analyst consensus. A Long-Short trading strategy that takes a Long (Short) position in firms with high (low) expected earnings surprise generates a monthly alpha of 75 basis points or an annual premium of about 9%. The performance of our trading strategy is robust during the sample period, and cannot be explained by standard asset pricing factors such as the market, size, value, momentum, short- and long-term reversal, and liquidity factors. We also analyze the sensitivity of the trading performance estimates when there is a delay between portfolio formation date and the start of the portfolio performance measurement period. We find that the performance of our trading strategy deteriorates as the gap between the portfolio formation date and the start of the portfolio performance measurement period increases. Consistent with the gradual information diffusion model of Hong, Lim, and Stein (2000), we find that it takes approximately three months for geographically dispersed information to be eventually incorporated into stock prices. Collectively, these results make several contributions to a growing literature in finance that exploits the geographical links among publicly-traded firms in the U.S. (García and Norli (2012), Giroud (2013), Bernile, Kumar, and Sulaeman (2013)). Our first main contribution is to show that earnings of geographically related firms have the ability to predict the future performance of a given firm. To our knowledge, the evidence of earnings and cash flow predictability using geographically dispersed information is new to this literature. 5
Second, we present novel evidence of predictable patterns in stock returns. Specifically, stock market participants incorporate geographically dispersed information about firms with a delay and this slow information diffusion generates predictable patterns in stock returns. Even relatively more sophisticated market participants such as equity analysts and institutional investors do not successfully incorporate firms geographical information in their forecasts and trades, respectively. As a result, earnings surprises of firms are predictable. In addition, firms exhibit stronger postearnings-announcement drift and stronger momentum in returns when geographic information is more dispersed and more difficult to aggregate. Our study also relates to prior studies, which demonstrate that local investors and equity analysts exhibit local bias and may possess an informational advantage. 2 Our findings suggest that a significant part of local information still remains unexploited. Last, our paper is related to recent studies that examine economic links among firms that emerge through customer-supplier networks (e.g., Cohen and Frazzini (2008), Menzly and Ozbas (2010)). Our evidence suggests that common exposures to geographical regions generate additional economic links among firms. The rest of the paper is organized as follows. Section 2 provides a summary of various data sources used in the empirical analysis. Section 3 presents evidence of earnings and cash flow predictability using geographic connections. Sections 4 and 5 examine whether market participants are able to quickly aggregate this geographically dispersed information into stock prices. We conclude in Section 6 with a brief discussion. 2. Data and Methods In this section, we provide details about our data sources and present summary statistics for the main variables used in the empirical analysis. 2 For example, see Coval and Moskowitz (1999, 2001), Huberman (2001), Hau (2001), Grinblatt and Keloharju (2001), Malloy (2005), Ivković, Sialm, and Weisbenner (2008), Teo (2009), van Nieuwerburgh and Veldkamp (2009). 6
2.1 Measures of Economic Connections We identify the headquarters of all U.S. publicly listed firms by retrieving their annual Form 10-K filings from Electronic Data Gathering, Analysis, and Retrieval (EDGAR) system. Our sample period is from 1994 to 2012. To comply with listing disclosure requirements, all U.S.-based firms are required to file Form 10-K with the U.S. Securities and Exchange Commission (SEC) annually. Each 10-K has a prescribed reporting structure and disclosure items, containing a comprehensive overview of a firm s business operations, summary of financial conditions, and audited financial statements. Apart from containing the operational details relating to a firm s business activities (e.g., organizational structure, executive compensation, industry competition, auditor s reports, and regulatory issues etc.), it also lists the detailed information regarding physical locations of a firm s assets (e.g., factories, warehouses, distribution centers, and sales offices, etc.), segment reporting, and management discussions and analysis of financial condition or results of operations. 3 We capture the economic exposure of a firm to each U.S. state by conducting a textual analysis of its annual 10-K filings (García and Norli (2012), Bernile, Kumar, and Sulaeman (2013)). For each firm-year, we parse the 10-K filings and count the number of times references are made to each of the 50 U.S. states and Washington D.C. Specifically, we focus on the occurrence of geographic references in the following four specific sections: (a) Item 1: Business, (b) Item 2: Properties, (c) Item 6: Consolidated Financial Data, and (d) Item 7: Management s Discussion and Analysis. As these four sections summarizes the locality of a firm s main business operations including a firm s plant and equipment, major physical assets, store locations, office locations, and 3 See http://www.sec.gov/answers/form10k.htm for additional details about the information reported in Form 10-K. 7
acquisition activities the resulting citation count broadly captures the economic ties between a firm and its geographical distribution of economic interests. While our textual analysis based measure of economic connections has measurement noise, it has several advantages. First, instead of relying on disclosure in firm s filings (e.g., Exhibit 21 or geographical segment reporting), our method greatly expands to a broader sample. Moreover, as some of the voluntary disclosures may be subject to management s discretion or interpretation (e.g., use of geographical regions in segment reporting, defining materiality in Exhibit 21), content analysis is less prone to this potential measurement bias. Specifically, firm managers have a great discretion in deciding a reportable geographical segment (under accounting standards SFAS 131 or IFRS 8), they have incentive to pool geographical segments to avoid disclosing commercially valuable information to competitors that is unavailable elsewhere (Harris (1998)). After obtaining the total count of each state s mentions in 10-K filings, we compute a citation share (CS) for each state in a firm s 10-K filing, defined as the number of times a U.S. state is cited divided by the total number of citations of all U.S. states in a firm s 10-K filing in year t. The maximum (minimum) of citation share is zero (one), where a higher (lower) citation share implies that a given firm s economic activities are more (less) concentrated in a given state. 4 2.2 Other Datasets We use the Central Index Key (CIK) to merge firms 10-K filings data with firm fundamentals from Compustat. CIK is assigned by the SEC to uniquely identify registered firms for meeting disclosure requirements. After merging, we retrieve firms headquarter location from the CRSP-Compustat Merged (CCM) file. The fundamentals of firms with at least $10 million of average total assets, two years of data, and $1 closing price are obtained from Compustat. All firm-year 4 See Bernile, Kumar, and Sulaeman (2013) for additional details about the citation share measure. 8
observations with SIC codes in the ranges of 4900-4949 (utility firms) and 6000-6999 (financial institutions) are excluded. Corresponding price and return data for trading strategies and factor model tests are then obtained from the Center for Research on Security Prices (CRSP). For analyst forecasts, we obtain split-adjusted earnings per share (EPS) forecasts from Thomson-Reuters IBES unadjusted detail file. To identity analysts geographical location, we use the names and brokerage house information in the Broker Translation File to hand-match records in Nelsons Directory of Investment Research (Malloy (2005)). Monthly data on the risk-free rate, market excess return (MKT), size factor (SMB), value factor (HML), and momentum factor (UMD), short-term reversal (STR), and long-term reversal (LTR) factors are obtained from Ken French s website. The quarterly measures of the economic activity indices of U.S. states are obtained following the Korniotis and Kumar (2013) method. 3. Earnings and Cash Flow Predictability We begin our empirical analysis by estimating a series of predictive regressions. Specifically, we investigate whether the earnings of a firm can be predicted using the earnings information of other firms headquartered in the same state where the firm is headquartered or has an economic presence. Our conjecture is that the earnings information of other firms located in the same geographical area would contain information about a firm s future earnings. Thus, if a firm s operations are geographically dispersed, then the lagged performance of firms with economic presence in those economically connected states may contain valuable information about the future performance of the firm. We test this conjecture using data on both earnings and cash flows. Specifically, we investigate whether the earnings and cash flows of other firms headquartered in economically connected (EC) states have any incremental power to predict a given firm s future earnings and cash 9
flows. We also compare the predictive power of the EC based measures and earnings measures obtained using firms around corporate headquarters. 3.1 Predictability Using Information Around Corporate Headquarters Our first test examines whether a firm s earnings and cash flows can be predicted by the financial performance of other firms headquartered in the same state. Specifically, we estimate the following predictive Fama-MacBeth (1973) style regression:,,,,, (1) At the end of each quarter, we regress firm j s Earnings in quarter t+1 on an intercept, lagged Earnings in quarter t, lagged HQ Earnings in quarter t, and a vector of control variables (X), which include firm-specific characteristics and measures of the HQ state s economic environment in quarter t. Earnings is defined as operating income after depreciation divided by average total assets (Richardson, Sloan, Soliman, and Tuna (2005)). The lagged Earnings is included in the specification as prior studies show that firm s earnings is persistent (Fama and French (2000), Dichev and Tang (2009), Frankel and Litov (2009)). HQ Earnings is the value-weighted Earnings of firms headquartered in the same state. β 2 and β 3 jointly measure the one-quarter-ahead predictability of firm earnings, where the main coefficient of interest is β 2. Our main conjecture is that HQ Earnings in quarter t would contain incremental explanatory power beyond the predictive ability of firm-specific lagged Earnings in quarter t. A positive coefficient estimate would indicate that the earnings information of other firms headquartered in the same state can be used to predict a given firm s earnings in quarter t+1. As the error terms are likely to be auto-correlated in the Fama-MacBeth (1973) regressions, all standard errors are adjusted using 10
the Newey-West (1987) method with 4-year lag (i.e., 16 quarters). To minimize the noise in our measures, we require at least two other firms when we compute the HQ-based earnings measures. X is a vector of control variables, which includes the following firm attributes: 1) Firm Size, 2) Leverage, 3) Loss, 4) Market-to-Book, 5) Dividend Yield, 6) No-Dividend Indicator, 7) Dividend-Price, and 8) Economic Activities Index. Firm Size is the natural logarithm of total assets. Leverage is the sum of shortterm and long-term debts, divided by total assets. Loss (No-Dividend Indicator) is a dummy that takes a value of one when operating income (dividend) is negative (zero), and zero otherwise. Market-to-Book is the sum of market equity, short-term debt, and long-term debt, divided by total assets. Dividend Yield is the dividends divided by shareholders equity. Together, these control variables should account for differences in firm size, growth opportunities, operations, and profitability, which could affect firm earnings. Additional details about all variables are available in Appendix A. Beyond these firm attributes, we include two variables that capture the differences in statelevel business cycles across headquarter states. The first control variable is the Dividend-Price, defined as the value-weighted average of the log of one plus the dividend-price (D/P) ratio of firms headquartered in the same state. Dividend D is the sum of the past four quarterly dividends, whereas P is the end-of-month stock price. The monthly stock prices are obtained from CRSP, and the quarterly stock-level dividends are obtained from Compustat. We also include the Economic Activities Index in the specification, which is a summary index of state-level macroeconomic conditions in a firm s headquarters state. We consider three state-level economic indicators that are likely to capture business cycle variation at the local state-level: 1) income growth, 2) unemployment rate, and 3) housing collateral ratio. All these state-level macroeconomic data are available at the quarterly frequency. Following Korniotis and Kumar (2013), the Economic Activities Index is defined as the sum of standardized values of state-level income growth 11
and state-level housing collateral ratio, minus the standardized value of the relative state-level unemployment ratio, divided by three. In this definition, the state-level income growth is the log difference between state income in a given quarter and state income in the same quarter in the previous year. This measure is motivated by prior studies that interpret this measure as a proxy for the return to human capital (e.g., Jagannathan and Wang (1996), Campbell (1996)). A high-level of income growth reflects positive macroeconomic conditions in a state. State-level unemployment rate is the ratio of the current unemployment rate to the moving average of the unemployment rates from the previous 16 quarters. It is a recession indicator for the state economy. The moving average serves as a proxy for the projected or expected level of unemployment and a positive (negative) deviation from this projected or expected level signals a good (bad) signal for the local economy. The unemployment rates are obtained from the Bureau of Labor Statistics (BLS). State-level housing collateral ratio is the log ratio of state-level housing equity to state labor income following Lustig and van Nieuwerburgh (2005, 2009). It is computed using the Lustig and van Nieuwerburgh (2005) method, where the data are obtained from Stijn van Nieuwerburgh s website. This measure captures the tightness of borrowing constraints and the degree of risk sharing at the local state level. A high housing collateral ratio predicts high consumption growth at the statelevel, as the consumption of individuals is better insulated against adverse labor income shocks. For our second set of baseline tests, we use a similar specification as equation (1) to examine the predictability of one-quarter-ahead cash flows. Specifically, we replace the dependent variable Earnings with Cash Flow, and estimate the following specification:,,,,, (2) 12
In this equation, Cash Flow is the cash flows from operating activities divided by average total assets. Similar to the Earnings specification, the main coefficient of interest is β 2. A positive coefficient would indicate that the cash flows of firms situated in the same state predicts a given firm s future cash flows. Our choice of firm fundamentals Earnings and Cash Flow are motivated by prior studies that widely study these metrics as proxies of financial performance of firms (e.g., Sloan (1996), Fama and French (2000)). The primary difference between these two measures is that the former captures the accrual component of a firm s earnings, which typically includes future or deferred cash flow, depreciation, and allowances. In contrast, the latter measures the actual cash flow component of earnings, and depends on the actual timing of the earnings realization. We present the summary statistics of all variables in Panel A of Table I. The mean Earnings (Cash Flow) is 0.009 (0.017), which indicates that the average firm earnings (cash flows) is 0.9% (1.7%) of total assets each quarter. This finding is consistent with prior studies that find that cash flows from operations are on average higher than earnings (e.g., Dichev and Tang (2009)). We also find that firm-specific earnings (Earnings) are more volatile than aggregate state-level earnings (HQ Earnings), as the standard deviation of Earnings is about 3.5 times the standard deviation of HQ Earnings. The same pattern is observed with the Cash Flow measure, as the aggregate state-level cash flow (HQ Cash Flow) is less volatile than firm-specific cash flows (Cash Flow). Panel B of Table I reports the correlations among these measures. We find that the aggregate state-level earnings/cash flows (i.e., HQ Earnings/HQ Cash Flow) is strongly and positively correlated with firms earnings/cash flows (i.e., Earnings/Cash Flow) at the 1% level. Table II reports the estimates from the earnings and cash flow regression specifications. Specifically, Columns 1 and 4 report the univariate regression results, Columns 2 and 5 report the results regressing firm s Earnings/Cash Flow on lagged Earnings/Cash Flow and lagged HQ 13
Earnings/Cash Flow, while Column 3 and 6 reports the results with the full specification that includes all control variables. First, consistent with prior evidence, we find that firms earnings and cash flows are persistent, as the estimated coefficients on lagged firm-specific Earnings and Cash Flow are positive and statistically significant. The average R 2 is 0.680, confirming the findings in prior literature that the firm-specific earnings process is fairly persistent. The economic magnitudes of lagged Earnings and Cash Flow are also similar to the findings in prior studies. 6 Second, we find that controlling for firms lagged performance, the earnings and cash information of other firms headquartered in the same state predicts the firm s future financial performance. Specifically, the estimated coefficients on HQ Earnings and HQ Cash Flow are positive and statistically significant, ranging from 0.048 (t-statistic = 2.87) to 0.124 (t-statistic = 3.58). Examining the economic magnitudes of these estimates, we find that a one standard deviation change in HQ Earnings (HQ Cash Flow) leads to a 0.075% (0.197%) change in Earnings (Cash Flow) in quarter t+1. Relative to average quarterly earnings (cash flow) of 0.9% (1.7%), these magnitudes are economically meaningful. Overall, the estimates from earnings and cash flow regression specifications indicate that the information contained in the financial performance of other firms headquartered in the same state predicts a given firm s earnings and cash flows. 3.2 Predictability using Information From Economically Relevant Regions Next, we examine whether the fundamental information in a firm s EC states has incremental ability to predict its future performance, incremental over the ability of firm fundamentals of other firms located in the same state. To test this conjecture, we re-estimate our 6 For instance, see Richard, Sloan, Soliman, and Tuna (2005), Dichev and Tang (2009), and Frankel and Litov (2009). 14
baseline regressions with three additional independent variables using the following regression specification:,,,,,, (3) Similar to HQ Earnings, EC Earnings is the citation-share weighted Earnings of firms located in economically connected states, excluding the firms in the HQ state. As before, all EC-based measures require a minimum of two firms. The following example illustrates the intuition behind our identification strategy: consider a technology firm that is headquartered in Texas but has 2/3 of its operations in California and 1/3 of its operations in Florida. EC Earnings is the sum of a) 2/3 of the average earnings of firms that are headquartered in California or have economic presence in that state, and b) 1/3 of the average earnings of all firms with corporate headquarters or economic presence in Florida. If these geographical locations contain value-relevant information about the technology firm, our measure should predict the firm s future performance, as captured by the coefficient of interest β 3. A positive coefficient would indicate that the earnings information of firms located in a firm s economically connected states is able to predict the firm s earnings in the next quarter. Beyond the standard firm-level control variables, two additional control variables EC Dividend-Price and EC Economic Activities Index are included to capture the macroeconomic conditions in EC states. These control variables are important, as EC Earnings could potentially capture the overall economic environment of firms economically-relevant states. Thus, including these additional variables helps mitigate this concern. Specifically, EC Dividend-Price is constructed following HQ Dividend-Price, defined as the citation-share weighted dividend-price index (Dividend- Price) of all firms located in economically connected states, excluding the HQ state. Similarly, EC Economic Activities Index is the citation-share weighted Economic Activities Index of all firms in 15
economically connected states, with HQ state excluded. We estimate the following predictive cash flow regression specification:,,,,,, (4) Here, EC Cash Flow is the citation-share weighted Cash Flow of firms located in economically connected states, excluding the HQ state. A positive coefficient of β 3 would indicate that the cash flow information of firms in a firm s economically connected states can predict its future cash flows. The summary statistics in Panel A of Table I show that EC Earnings/EC Cash Flow exhibit high and significant positive correlations with firm-specific Earnings/Cash Flow. EC-measures also have the lowest volatilities, as on average they have less than a quarter of the standard deviation of Earnings/Cash Flow. This evidence is not surprising. Since firms typically have economic presence in multiple states, a diversified geographical presence lowers the volatility of these measures. The estimation results for EC based earnings and cash flow regressions are reported in Table III. Similar to the previous results, Columns 1 and 4 report the results regressing firm s Earnings/Cash Flow on 1) lagged Earnings/Cash Flow, 2) lagged HQ Earnings/Cash Flow, lagged EC Earnings/Cash Flow, and control variables, whereas Columns 3 and 6 report the results for the full specification that includes all control variables. We find that EC Earnings contain information about the future earnings of a firm. This effect is incremental over the ability of HQ Earnings to predict a firm s future earnings. Specifically, the estimated coefficients on EC Earnings are all positive and statistically significant, ranging from 0.275 (t-statistic = 12.95) to 0.441 (t-statistic = 24.13). In economic terms, a one standard deviation increase in EC Earnings corresponds to a 0.156% increase in firms Earnings in the next quarter, which is economically meaningful. This economic magnitude is greater than the predictive power of HQ Earnings in economic terms. In this expanded regression specification, we also find that the 16
estimated coefficients on HQ Earnings remain positive and statistically significant across all specifications. However, when EC Earnings is included in the specification, the magnitudes of HQ Earnings weaken in comparison to the corresponding estimates reported in Table II, and ranges from 0.039 (t-statistic = 6.42) to 0.080 (t-statistic = 6.50). A similar pattern is present in the Cash Flow regression results. We find that EC Cash Flow exhibits a strong ability to predict firms cash flows in the next quarter. Specifically, the estimated coefficients on EC Cash Flow are positive and statistically significant across Columns 4-6, ranging from 0.152 (t-statistic = 2.35) to 0.382 (t-statistic = 4.47). Similar to our earlier evidence, EC Cash Flow has stronger predictive ability than HQ Cash Flow, where a one standard deviation increase in EC Cash Flow predicts a 0.468% increase in firms Cash Flow in the next quarter. Since all regressions specifications include the EC Economic Activities Index, these results are unlikely to reflect the effects of the overall macroeconomic environment in economically-relevant states. Moreover, since all EC measures exclude firms HQ states, our estimates only reflect the average performance of firms in economically-relevant states. Collectively, these predictive regression results show that the fundamental information of firms located in economically connected states contain valuable incremental information about the future performance of firms. 3.3 Longer-Horizon Earnings and Cash Flow Predictability In this section, we examine the predictive power of EC based measures over longer horizons. The results from longer-horizon predictability are reported in Table IV. Specifically, Columns 1, 2, 3, and 4 present the estimates using 2-, 3-, 4-, and 8-quarter-ahead earnings measures. All estimated coefficients on EC Earnings remain positive and statistically significant across all specifications, ranging from 0.294 (t-statistic = 5.49) to 0.391 (t-statistic = 14.80). Based on the estimated coefficients on EC Earnings, the predictability gradually weakens as the prediction horizon increases 17
from t+1 (one quarter ahead) to t+8 (two years ahead). Specifications focusing on firms cash flows from operations exhibit a similar trend, with all coefficients on EC Cash Flow remaining positive and statistically significant. Moreover, the average R 2 s for Earnings (Cash Flow) gradually decrease from 0.593 in t+2 to 0.469 in t+8 (0.397 in t+2 to 0.355 in t+8). Overall, the evidence from long-horizon predictability regressions shows that our geography-based earnings and cash flow measures have predictive power even over longer horizons. 3.4 Earnings Predictability: Robustness Tests In the last set of earnings predictability tests, we examine whether the baseline results are robust to an alternative definition of economically connected states. In the first test, instead of using all cited states, we alternatively compute our earnings and cash flow measures only using the top three states with the highest citation-share (and HQ state excluded). We re-estimate our baseline regression specifications using this alternative definition to examine the robustness of our earlier results using citation shares unconditionally. The estimation results are summarized in Table V. First, the EC information continues to exhibit predictive power even under this alternative definition. Specifically, the coefficients on EC Earnings remain positive and statistically significant, ranging from 0.209 (Column 3) to 0.320 (Column 1), and statistically significant at the 1% level. The same pattern is observed in the Cash Flow regression estimates. The estimated coefficients on EC Cash Flow are positive and statistically significant at the 1% level, ranging from 0.140 (in Column 6) to 0.296 (in Column 4). Further, we find that EC information continues to exhibit stronger predictive power relative to that of HQ information across all specifications. For instance, given that the mean and standard deviation of EC Earnings (HQ Earnings) is 0.010 and 0.005 (0.003 and 0.015), respectively, a one 18
standard deviation change in EC Earnings (HQ Earnings) leads to 0.120% (0.062%) increase in firms Earnings in the next quarter, which is economically significant. In the next test, we examine whether our results are sensitive to firms incorporated in Delaware. A potential concern with our baseline results may be that our results are simply driven by firms incorporated in Delaware. A significant portion of firms are incorporated in Delaware and firms may cite Delaware often in their 10-K reports even if they do not have a meaningful economic presence in the state. To address this potential concern, we exclude all sample firms that are incorporated in Delaware and re-estimate our baseline regressions as in Table III. Although, the total number of observations drops significantly, our main results remain very similar. Specifically, the coefficient estimates of EC Earnings range from 0.233 (t-statistic = 5.60) to 0.281 (t-statistic = 6.23), whereas the estimated coefficient estimates of EC Cash Flow range from 0.221 (t-statistic = 3.96) to 0.285 (t-statistic = 7.78). This evidence indicates that our baseline results reported in Table III are unlikely to be driven by Delaware firms. For additional robustness, we also examine whether our results are similar when we use data with annual frequency. Although the number of cross-sections drops substantially from 64 quarters to 16 years, the strong predictive power of EC Earnings and EC Cash Flow remains intact. This evidence suggests that our earnings and cash flow predictability regressions are robust. 4. Aggregation of Geographically Dispersed Information Our results so far show that information about earnings and cash flows of firms are geographically dispersed. Specifically, the aggregated fundamental information of firms located in economically relevant states of a firm predicts its future firm-specific earnings and cash flows. A logical question is how this geographically-dispersed information is aggregated by market participants and subsequently reflected in stock prices. 19
There is mounting evidence that financial intermediaries such as equity analysts play an important role in processing and disseminating firm-specific information. In this section, we examine whether those financial intermediaries speed up the information aggregation process through timely updates of their earnings forecasts. We also investigate whether institutional investors around HQ and EC locations speed up the information aggregation process. This analysis is motivated by the prior evidence in the literature, which demonstrates that local investors at both HQ and EC locations have a local informational advantage. In the last part of the section, we focus on the asset pricing implications of potential inefficiencies in the information aggregation process. 4.1 Geographic Dispersion of Information and Analyst Behavior In this section, we examine whether equity analysts incorporate geographically dispersed information about a firm in their earnings forecasts. Before presenting the main results, we reestimate our baseline earnings predictability regressions using the subsample of firms that have analyst coverage. This test is motivated by prior studies, which demonstrate that stocks without analyst coverage are small stocks with relatively low liquidity (Hong, Lim, and Stein (2000)). We want to ensure that our evidence of earnings predictability is not restricted to firms with no analyst coverage. The estimates in Column 1 of Table VI show that the predictability patterns are strong even among stocks with analyst coverage. Specifically, the estimated coefficient on EC Earnings is 0.304 (tstatistic = 9.36), which is larger than the estimate of 0.275 (t-statistic = 12.95) reported previously in Column 3 of Table III where we consider the full sample of firms. Since prior evidence shows that analyst coverage is skewed toward stocks with large market capitalization (Bhushan (1989), Hong, Lim, and Stein (2000)), this evidence suggests that our predictability results are not concentrated among small stocks with low market capitalization or poor liquidity. 20
Next, we examine whether equity analysts incorporate firms geographical information into their forecasts. Given the prior evidence on the information aggregation role of equity analysts, we expect analysts to incorporate geographically dispersed earnings information into their forecasts. We test this conjecture using two additional dependent variables: Analyst Consensus Forecast and Forecast Error. Analyst Mean (Median) Consensus Forecast is the mean (median) consensus based on all outstanding forecasts issued within the 90 days period prior to the earnings announcement. If analysts issue multiple forecasts, only their last forecasts are used to ensure that analysts with frequent forecasts are not over-weighted when computing the consensus forecast. Forecast Error is defined as the absolute value of the actual earnings minus analyst consensus scaled by lagged stock price. We also include Number of Analysts as an additional control variable, which is defined as the natural logarithm of the number of analysts following a stock in a given quarter. This additional control variable is included in the specifications because prior studies show that the number of analysts is strongly correlated with a firm s information environment (e.g., Hong, Lim, and Stein (2000)). To reduce the noise generated by extremely small denominators when computing these two measures, we follow Livnat and Mendenhall (2006) and include only firms with at least $1 per share at the end of each quarter, non-missing market (book) value of equity, and at least $5 million market capitalization at the end of the prior quarter. These filters screen out stocks with low liquidity or low market capitalization from the sample. In our tests, we use the consensus forecast and forecast error measures to examine whether analysts use publicly available information from firms economically relevant locations. If analysts incorporate firms geographically dispersed information into their forecasts, this geographic information would predict the levels of their consensus forecasts but not the errors in their forecasts. 21
In particular, if equity analysts are more effective in incorporating firms fundamentals from economically connected states, their consensus forecasts should be more informative. As their aggregated forecasts deviate less from the actual earnings, the forecast error would be lower. In contrast, if analysts do not fully adjust their earnings forecasts to reflect firms economic activities in economically relevant states, their consensus should not be related to the earnings information aggregated from firms economically relevant states, while consensus-based forecast error would be positively related to the earnings information in those economically relevant states. The results from Analyst Consensus Forecast regressions are reported in Columns 2 and 3 of Table VI. We find that analysts do not fully incorporate firms geographically dispersed information into their forecasts, as the estimated coefficients on EC Earnings are marginal in statistical significance (estimate = 0.038; t-statistic = 1.97). This result is unlikely to be driven by extreme values, as the results using mean and median consensus forecasts are nearly identical. However, analysts form their forecasts using firms historical earnings, as the estimated coefficients on Earnings are significantly positive. This evidence is consistent with findings from prior research (e.g., Fama and French (2000)). Those coefficient estimates have the highest statistical significance across all independent variables, followed by the estimated coefficients on HQ Earnings, which are also significantly positive (estimate = 0.015; t-statistic = 6.21). These results suggest that, when forming their earnings forecasts, analysts incorporate more information from a given firm s historical earnings and the average performance of firms located in that firm s headquarters state. While this evidence confirms prior findings that analysts use past earnings information when they form their earnings forecasts (e.g., Mendenhall (1991), Abarbanell and Bernard (1992)), it also indicates that analysts have difficulty aggregating value-relevant geographically dispersed information. 22
The results from Forecast Error regressions portray a similar picture and further illustrate that analysts do not fully incorporate geographically distributed information into their earnings forecasts. Specifically, the estimated coefficients on EC Earnings are positive and statistically significant (estimates vary between 0.156 and 0.157; t-statistics vary between 3.50 and 3.56). Comparing the magnitudes of the coefficient estimates, we find that estimates of EC Earnings are larger than those for HQ Earnings and Earnings. A one standard deviation change in EC Earnings leads to a 5.251% change in Forecast Error (vs. 3.276% for HQ Earnings and 0.977% for Earnings). Moreover, the estimated coefficients on lagged Earnings are statistically insignificant, which provides further evidence suggesting that analysts forecasts tend to fixate on firms lagged fundamentals. In additional tests, we also examine whether local analysts who are located close to corporate headquarters or economically relevant states are more likely to incorporate information about local firms more efficiently. The evidence reported in Appendix B shows that these local analysts are unable to incorporate geographically dispersed information either. We find this evidence both before and after the implementation of Regulation Fair Disclosure (Reg FD). Overall, these regression estimates show that, while value-relevant information about a firm is geographically dispersed, equity analysts do not incorporate this information effectively into their earnings forecasts. They are unable to facilitate the aggregation of this dispersed information. 4.2 Geographic Dispersion of Information and Market Reaction to Earnings If equity analysts are unable to effectively account for geographical information in their earnings forecasts, a natural question is how this information is incorporated into asset prices. While earnings surprises are usually incorporated into stock prices quickly, an extensive literature documents and analyzes the drift in prices that occurs in the weeks and months following earnings announcements (Mendenhall (1991), Abarbanell and Bernard (1992)). If market participants have 23