Long-Term Evidence on the Effect of Aggregate Earnings on Prices

Transcription

1 Long-Term Evidence on the Effect of Aggregate Earnings on Prices Yunhao Chen, Xiaoquan Jiang, and Bong-Soo Lee We examine the time-series properties and determinants of the relation between aggregate earnings information and stock prices (aggregate earnings response coefficient or AERC) employing return decompositions with data since We confirm that AERC is negative even though firms respond positively to individual firm earnings information, but we also find that AERC is time varying. Furthermore, we show that AERC components based on expected earnings, cash flows, and discount rates are also time varying and differ in relative importance. The relation between stock prices and reported earnings has attracted extensive attention in accounting and finance research. Since Ball and Brown (1968), the literature has shown that positive earnings changes are associated with positive stock price reactions for individual firms (a positive earnings response coefficient or ERC). However, recent studies find that stock prices react negatively to aggregate earnings news (a negative aggregate earnings response coefficient, or AERC) as documented in Kothari, Lewellen, and Warner (2006), Ball, Sadka, and Sadka, (2009), and Cready and Gurun (2010). For example, using the quarterly S&P Index from 1970 to 2000, Kothari et al. (2006) and Sadka and Sadka (2009) find that price responses are unrelated to past aggregate earnings. Instead, they are negatively correlated with current aggregate earnings changes. Furthermore, by using a sample of 28 countries and running time-series regressions of market returns on aggregate earnings changes for each country, He and Hu (2013) find that only four countries have negative coefficients for aggregate earnings changes and none of these negative coefficients are statistically significant. To explain the puzzling contrast between firm price reactions to firm earnings and aggregate earnings, we examine whether and how the AERC varies over time and what determines the price responses to aggregate earnings changes. Furthermore, while recent studies on AERC have furthered our understanding of the relationship between aggregate earnings and firm values, they primarily focus on the period from the 1970s to the 2000s. In this study, we attempt to provide a comprehensive picture of AERC over a longer period from 1871 to We find the previously documented negative AERC is time-specific and AERC is time varying (see Figure 1). We confirm that the time-varying property of AERC is due to two major factors. One factor is that the stock market regulations affect the information environment of the stock market, which, in turn, affects the investors perception of the value of firm disclosure. For example, in 2000, the passage of Regulation Fair Disclosure (Reg FD) has brought more timely and frequent communication between companies and investors thereby enhancing information We thank two anonymous referees and Marc Lipson (Editor) for many insightful comments. Also we thank the participants at 2012 Asian Finance Association Conference in Taipei, 2012 American Accounting Association Conference in Washington DC, and 2012 Financial Management Association Conference in Atlanta. Yunhao Chen is a visiting Assistant Professor in the Accounting Department in the School of Business Administration at the University of Miami in Coral Gables, FL. Xiaoquan Jiang is an Associate Professor in the Department of Finance in the College of Business at Florida International University in Miami, FL. Bong-Soo Lee is a Professor in the Department of Finance in the College of Business at Florida State University in Tallahassee, FL. Financial Management Summer 2015 pages

2 324 Financial Management Summer 2015 Figure 1. Rolling Regressions Betas of Returns, Expected Returns, Cash Flow News, and Discount Rate News on de/p with S&P Data This figure illustrates the betas from the 30-quarter rolling regressions of aggregate returns, expected returns, cash flow news, and discount rate news on aggregate earnings using S&P data from 1871:Q1 to 2011:Q4. Rm is the S&P 500 Index return, de is the quarterly differenced earnings of the S&P 500 Index, and P is the market value of the S&P 500 Index. ER, Ncf, andndr represent expected return, cash flow news, and the discount rate news components of stock returns that are generated from VAR, respectively. The relation between the return and its components is as described in Equation (1): r t = E t 1 r t + N CFt N DRt. BETA_RET

3 Chen, Jiang, & Lee Long-Term Evidence on the Effect of Aggregate Earnings on Prices 325 Figure 1. Continued CF_ERC Panel C. Rolling Regression Betas of SP CF_ERC on de/p JAN1870 JAN1880 JAN1890 JAN1900 JAN1910 JAN1920 JAN1930 JAN1940 JAN1950 JAN1960 JAN1970 JAN1980 JAN1990 JAN2000 JAN2010 JAN2020 DR_ERC 15 Panel D: Rolling Regression Betas of SP DR_ERC on de/p JAN1870 JAN1880 JAN1890 JAN1900 JAN1910 JAN1920 JAN1930 JAN1940 JAN1950 JAN1960 JAN1970 JAN1980 JAN1990 JAN2000 JAN2010 JAN2020 transparency (Bailey et al., 2003; Heflin, Subramanyam, and Zhang, 2003; Eleswarapu, Thompson, and Venkataraman, 2004). 1 The enactment of the Sarbanes-Oxley (SOX) Act in 2002 has also been successful in improving the quality of firm disclosure and restoring confidence in the capital 1 Regulation Fair Disclosure was promulgated by the Securities and Exchange Commission (SEC) in August 2000, which requires all publicly traded companies to disclose material information to all investors.

4 326 Financial Management Summer 2015 markets (Cohen, Dey, and Lys, 2008; Koh, Matsumoto, and Rajgopal, 2008; Kolev, Marquardt, and McVay, 2008). Anilowski, Feng, and Skinner (2007) find that aggregate earnings guidance is associated with aggregate earnings news and, to some extent, market returns. With the enhanced information environment and disclosure quality, stock prices reflect available news in a more efficient fashion allowing investors to better draw inferences for next period earnings changes. The other factor is that macroeconomic conditions and business cycles affect aggregate uncertainty and preferences (Chen, 1991; Menzly, Santos, and Veronesi, 2004; Henkel, Martin, and Nardari, 2011). Aggregate earnings are widely recognized as having implications for the growth potential of stocks and, therefore, as a plausible measure of aggregate cash flows (i.e., the cash flow effect) (Campbell and Shiller, 1988a, 1988b; Fama and French, 1989; Hecht and Vuolteenaho, 2006; Kothari et al., 2006). The literature regarding firm-level ERC suggests that earnings also contain information on firm risk (Easton and Zmijewski, 1989; Ball, Kothari, and Watts, 1993). At the aggregate level, Kothari et al. (2006) provide evidence that aggregate earnings and discount rate news move together (i.e., discount rate effect). Furthermore, Shivakumar (2007, 2010) proposes that aggregate earnings surprises could signal unexpected inflation that can be directly linked to discount rate movements (Fisher, 1907) and other macroeconomic variables. We argue that earnings convey information on both growth (cash flows) and risk (discount rates) and that the relative importance of each effect varies over time. If aggregate earnings are dominated by cash flow effects, we would observe a positive earnings-price relation. However, if aggregate earnings are dominated by discount rate effects, we would find a negative earnings-price relation. The sign of AERC depends upon the relative importance of cash flow news, discount rate news, and their time-varying correlations. 2 Kothari et al. (2006) suggest that investors may negatively react to positive aggregate earnings news as investors are more risk-averse, indicating discount rate news is dominant. In contrast, Sadka and Sadka (2009) and Choi, Kalay, and Sadka (2011) argue that earnings changes are predictable (i.e., price leads earnings). 3 We use these ideas to explain why aggregate ERC is time varying (i.e., why it is positive in some subsample periods and negative in others). We estimate the three components of market returns (expected returns, discount rate news, and cash flow news) by applying Campbell s (1991) return decomposition and examine the determinants of the relation between three market return components and aggregate earnings. 4 We find that the relative importance of return components determines the variation of aggregate ERC over time. Our study contributes to the literature by providing new insights into price response to aggregate earnings. First, we explore the puzzling finding that the ERC can be negative at the aggregate level, by providing evidence that aggregate ERC is time varying. Specifically, it was generally positive prior to the 1960s, and then it becomes negative in the subsample periods from 1970 to 2000 before it becomes positive again after 2000, which is the post-sox period. The results are consistent with Kothari et al. (2006) and Sadka and Sadka (2009) when we analyze the data covering the time period specified in their paper. In addition, we contribute to the understanding of the determinants of aggregate ERC by examining the correlation between earnings changes 2 In addition, the present value model suggests that stock prices reflect growth and risk information. Since both earnings and stock prices contain common growth and risk information that is time varying, the relation (comovement) between aggregate earnings and market returns would be time varying as well. 3 Similar points are made in Collins, Kothari, and Rayburn (1987) and He and Hu (2013). 4 Campbell s (1991) VAR-based approach faces the challenge that the results may depend upon model specification. Therefore, in addition to estimating the return components using VAR decomposition, we also use a direct approach as in Campbell, Polk, and Vuolteenaho (2010) to estimate the three components of market returns and verify the robustness of the results that use VAR decomposition. See Section III.D for discussions on VAR methodology.

5 Chen, Jiang, & Lee Long-Term Evidence on the Effect of Aggregate Earnings on Prices 327 and expected returns, cash flow news, and discount rate news. In the prior literature, Kothari et al. (2006) focus on the news, while Sadka and Sadka (2009) focus on the expected return component. We follow the prior literature and estimate all three return components using vector autoregression (VAR) decomposition and examine the relation between earnings changes and return components. We provide evidence that component ERCs are time varying, suggesting that the relative importance of each component varies over time. Specifically, we find that both cash flow news and discount rate news play an important role in determining the aggregate ERC. However, we also determine that discount rate news is the primary driver of aggregate ERCs from 1970 to 2000, while the expected returns and cash flow news play a major role in the aggregate ERCs after The remainder of the paper proceeds as follows. In Section I, we discuss our research design including the return decompositions, VAR estimates, and the earnings return regressions. Section II describes the sample and data used in this study. Section III describes the empirical analysis and presents the empirical results, while Section IV provides our conclusions. I. Research Design In this section, we discuss our research design. We follow Campbell (1991) to decompose the realized returns into one-period-ahead expected returns, revision in the expected future cash flows, and revision in the expected discount rates. A. Return Decomposition In this study, we are interested in what determines stock returns since the main purpose of our study is to examine the relationship between stock returns and earnings changes. Therefore, we apply Campbell s (1991) return decomposition. Campbell (1991) decomposes the log return on equities into following three components: r t = E t 1 r t + (E t E t 1 ) ρ j d t+ j (E t E t 1 ) ρ j r t+ j = E t 1 r t + N CF,t N DR,t. (1) j=0 j=1 This equation states that the log return is associated with the one-period-ahead expected return, the revision in the expectations of future cash flows (cash flow news), and the revision in the expectations of future discount rates (discount rate news). As noted by Campbell (1991), Equation (1) is a dynamic accounting identity relating the current return innovation to revisions in expectations. 5 5 Hecht and Vuolteenaho (2006) apply this decomposition to measure the relative importance of these three effects in regressions of returns on cash flow proxies. Based on Equation (1), the explanatory power of cash flow proxies may arise from the correlation of cash flow proxies (predictors) with one-period expected returns, cash flow news, and expected return news. They argue that, If expected return variation is responsible for the high explanatory power of the aggregate regressions, these R 2 s should not be interpreted as evidence of cash-flow news driving the returns. Similarly, if expected-return news is highly variable and positively correlated with cash-flow news, the low R 2 s in regressions of firm-level returns on earnings do not necessarily imply that earnings are a noisy or delayed measure of the cash-flowgenerating ability of the firm. Even if earnings are a clean signal of cash-flow news, expected-return effects can garble the earnings-returns relation. (p. 160)

6 328 Financial Management Summer 2015 B. VAR Estimates and Regressions of Returns on Earning Shocks Following Campbell (1991) and Campbell and Vuolteenaho (2004), we use a VAR model to estimate the one-period-ahead expected returns, cash flow news, and the discount rate news series. We assume that the data are generated by a first-order VAR model: Z t = A 0 + AZ t 1 + u t, (2) where Z t is a vector of log market returns (r), the cross-sectional equity risk premium (λ), the log book-to-market ratio (bp), and the log return on equity (roe) describing the economy at time t. A 0 and A are a vector and a matrix of constant parameters, respectively, and u t is a vector of shocks. The choice of state variables in VAR is consistent with Vuolteenaho (2002). With VAR expressed in this form, the components of the identity in Equation (1) can be obtained by E t 1 r t = e1 (A 0 + AZ t 1 ), (3a) N CF,t = [e1 + e1 ρ A(I ρ A) 1 ]u t, (3b) N DR,t = e1 ρ A(I ρ A) 1 u t, (3c) where ei = [ ], picking the ith state variable in the state vector, and I is an identity matrix. With these components available, we are able to examine aggregate ERC and the return component ERCs in detail. C. Contemporaneous Returns-Earnings Relation We examine the correlation between aggregate earnings and stock returns following the conventional procedure in the prior literature. In addition, we apply Campbell s (1991) decomposition to estimate and test the contemporaneous relation between stock return components and earnings shocks. Consider a typical regression of returns on aggregate earnings changes: r t =α+β Y t +e t, (4) where Y is a difference of earnings. Using the decomposition of Campbell (1991), we can rewrite the Regression (4) as follows: E t 1 r t =α 1 +β ER Y t +e ER,t, (5a)

7 Chen, Jiang, & Lee Long-Term Evidence on the Effect of Aggregate Earnings on Prices 329 N CF,t = α 2 + β CF Y t + e CF,t, (5b) N DR,t = α 3 + β DR Y t + e DR,t. (5c) Since the independent variables in the Regressions (5a), (5b), and (5c) remain the same and the sum of the expected return and cash flow news with negative discount rate news is the realized return (Equation (1), Regressions (5a), (5b), and (5c) can also be combined as (Hecht and Vuolteenaho, 2006): r t = α + (β ER + β CF β DR ) Y t + (e ER,t + e CF,t e DRt ), (6) where α = α 1 + α 2 α 3. Regressions (5a), (5b), (5c), and (6) demonstrate that there are three sources driving the relation between stock returns and earnings: 1) the one-period expected return, 2) cash flow news, and 3) the discount rate news. A small magnitude (negative) of ERC may be due to the offsetting values of the betas in the return components. II. Sample and Data Description In our analysis of the relationship between stock returns and aggregate earnings, the main dependent variable is the return from the S&P 500 Composite Index and the independent variable is the aggregate earnings of the S&P 500 Composite Index. We obtain the historical earnings, price, and dividend information on the index from Shiller s Website. 6 The raw data are of monthly frequency. We construct quarterly data from the first quarter of 1871 to the fourth quarter of 2011 by compounding monthly returns and aggregating monthly earnings and dividends. To avoid the effect of outliers, we winsorize the market returns and changes in earnings at 5% and 95%. 7 For VAR, we select the following variables to be included in the state vector. First, the log return on the market (LogRm) is the log gross returns on the S&P 500 Index (For robustness, we also use the value-weighted and equal-weighted Center for Research in Security Prices [CRSP] index). In addition, LogBP is the log book-to-market ratio on the S&P 500 Index. Finally, LogROE is the log return on equity on the S&P 500 Index. 6 Shiller s Web site is shiller/data.htm. S&P s 500 Index is a capitalization-weighted index of 500 stocks. The index is designed to measure the performance of the broad domestic economy through changes in the aggregate market value of 500 stocks representing all major industries. Prior to 1957, the primary daily stock market index was the S&P 90, a value-weighted index based on 90 stocks starting in The index is extended back to 1871 by using the data in Cowles (1939). From 1926 to 2011, the earnings series is the S&P earnings per share adjusted to the index total for the year. The nominal dividend prior to 1926 was also taken from Cowles (1939). The nominal dividend series for is dividends per share adjusted to the index, four quarter total, for the S&P Composite Index. Using annual data that has consumption per capital data back to 1889, we calculate the correlation between the real S&P Index earnings growth and the real consumption per capital growth. It is for the full sample, from 1889 to 2000, from 1889 to 1969, from 1970 to 2000, and from 2001 to This indicates that aggregate earnings growth is a reasonable proxy for aggregate economic fluctuation, which is also time varying. We use the consumption per capital growth rate as a proxy for the macroeconomic growth rate due to the unavailability of the quarterly gross domestic product (GDP) growth data prior to We noticed some large shocks during the sample period. For example, the 1930 observation can be explained by the Great Depression of 1929, which is an unusual economic event. The observations in 1970s may be related to two oil crisis shocks in 1973 and The observations in the late 1980s may be related to the Savings and Loan Crisis, which led over half the savings and loans to fail. To circumvent the effect of extreme observations, we winsorized the market returns and changes of earnings at 5% and 95%. The results using data winsorized at 1% at both ends are consistent with the tabulated results.

8 330 Financial Management Summer 2015 Table I. Descriptive Statistics and Pearson Correlations This table reports the descriptive statistics and Pearson correlations among the main variables of interest. Rm is the S&P 500 Index return, BP is the S&P 500 Index book-to-market ratio, de is the quarterly differenced earnings of the S&P 500 Index, and P is the market value of the S&P 500 Index. ER, Ncf,andNdr represent expected return, cash flow news, and discount rate news components of stock return that are generated from VAR, respectively. The sample period is from 1871:Q1 to 2011:Q4. Variables LogRm LogBP LogROE de/p ER Ncf Ndr Panel A. Descriptive Statistics Mean Std Min Q Median Q Max LogRm LogBP LogROE de/p ER Ncf Ndr Panel B. Pearson Correlations LogRm <.001 <.001 <.001 LogBP < < LogROE <.001 < de/p <.001 <.001 <.001 ER Ncf <.001 To explore whether the aggregate earnings and returns relation varies with macroeconomic variables, we regress quarterly stock returns and each of their three components separately on quarterly differenced earnings conditional upon several macroeconomic activities and risk measures. We obtain most of the macroeconomic data from the Federal Reserve Bank at St. Louis GDP and consumer price index (CPI) are from 1947Q1. The recession probability index (LRI) is the logarithm of smoothed recession probabilities for the United States from 1967:Q2 obtained from a dynamic factor Markov-switching model applied to four monthly coincident variables: 1) nonfarm payroll employment, 2) the index of industrial production, 3) real personal income excluding transfer payments, and 4) real manufacturing and trade sales. 8 Dreg is a dummy variable that is equal to one if the data are after 2000, and zero otherwise. In Table I, we report the descriptive statistics of earnings changes, book-to-market ratios, market returns, and their components (Panel A), and the Pearson correlations among the variables of 8 For additional details, including an analysis of the performance of this model for dating business cycles in real time, see Chauvet and Piger (2008).

9 Chen, Jiang, & Lee Long-Term Evidence on the Effect of Aggregate Earnings on Prices 331 interest (Panel B). While the historical time-series behavior of market returns and earnings shocks are well known, the historical behavior of the market return components are less clear. The means of each component are zero since we use a demeaned series in VAR. Note that the volatility of cash flow news and discount rate news are much higher than the volatility of expected returns. The market return is positively correlated with expected returns and cash flow news, and negatively correlated with discount rate news. By construction, expected return is orthogonal to news (cash flow news and discount rate news). We also calculate the descriptive statistics for each subsample period without reporting the tables to save space. The results indicate that all of the variables exhibit the time-varying property, particularly the book-to-market ratio and expected returns. III. Empirical Analysis We examine aggregate ERC in two dimensions: 1) the relation between earnings changes and returns and (2) the correlation between earnings changes and return components, namely, expected return, cash flow news, and discount rate news. We provide evidence on both the contemporaneous and dynamic (causal) relations. We measure earnings changes as the first differenced earnings for quarter t scaled by the market value of the S&P 500 Index at t 1 (de/p). We also conduct robustness tests using alternative measures of aggregate earnings, such as earnings changes scaled by the book value of equity (de/b), and earnings surprise as a residual from an autoregressive regression where earnings are regressed on lagged earnings and lagged returns. A. VAR Estimation In order to decompose returns into expected returns, cash flow news, and discount rate news, we estimate a parsimonious VAR with one lag. Table II reports the VAR model parameter estimates. Each row of the table corresponds to a different equation of the VAR. The columns report coefficients on the three explanatory variables: 1) LogRm, 2) LogBP, and 3) LogROE. The first row in Panel A of Table II reports the quarterly VAR state variables. LogBP and LogROE are significant in predicting annual excess returns on the aggregate stock market, consistent with the results in Vuolteenaho (2002). Both LogBP and LogROE are highly persistent. Both LogRm and LogBP contain future profitability information. Using annual frequency data, Panel B reports similar results although the significance is a little weaker. B. Contemporaneous Relation between Earnings Changes and Returns (or Return Components) To provide an overall picture of the correlation between aggregate earnings changes and returns (or return components), we run 30 quarter rolling regressions. Figure 1 illustrates the time-varying relationship between aggregate earnings changes and returns (or return components). Panel A depicts the aggregate ERC over time. Specifically, the aggregate ERC in Panel A is primarily positive or close to zero prior to the 1960s except for some years. It then hovers around zero until the mid-1960s (slightly negative between the 1940s and the late 1950s and slightly positive for about a decade after that). From the mid-1960s to mid-1990s, the aggregate ERC is largely negative except for a few years. This result is consistent with the evidence in Kothari et al. (2006)

10 332 Financial Management Summer 2015 Table II. Vector Autoregressive Regression Parameter Estimates The table reports the OLS parameter estimates for a first-order VAR model including the demeaned log excess market return (LogRm), the demeaned log book-to-market (LogBP), and the demeaned log return on equity (LogROE). Each set of three rows corresponds to a different dependent variable. We report parameters, t-statistics and adjusted R-squares. Intercept LogRm t LogBP t LogROE t R 2 Panel A. Quarterly Data from 1871:Q1 to 2011:Q4 LogRm t+1 PARMS t-stat LogBP t+1 PARMS t-stat LogROE t+1 PARMS t-stat Panel B. Annual Data from 1871 to 2011 LogRm t+1 PARMS t-stat LogBP t+1 PARMS t-stat LogROE t+1 PARMS t-stat that the aggregate ERC is negative for their sample period from 1970 to However, Figure 1 provides additional contrasting evidence that after mid-1990, aggregate ERC becomes positive again. We conjecture that the switch in the pattern around 2000 may be due to the improvement in the information environment after Reg FD and the enactment of the SOX Act in 2002 (see further discussions below). Overall, this figure broadens our view and provides evidence that aggregate ERC is time varying. Panels B to D depict components of the aggregate ERC over time: expected return ERC (Panel B), cash flow news ERC (Panel C), and discount rate news ERC (Panel D). The component ERCs are generally less volatile than the aggregate ERC. Expected return ERCs mostly hover near zero. Cash flow news ERCs are largely positive with some dips in the late 1960s and late 1980s. Discount rate news ERCs are primarily positive with some dips in the mid-1880s and mid-1930s. In sum, these figures indicate that both aggregate ERC and the components of the aggregate ERC are time varying. We examine the correlation between aggregate earnings changes and returns (aggregate ERC) by regressing the market returns on current earnings changes. We allow the VAR model to vary in each subsample period and estimate the VAR parameters and generate return components for each subsample period accordingly. Then, we examine the relationship between earnings changes and return components by regressing the return components on earnings changes. Table III reports the aggregate ERC and component ERCs in the full sample and subsamples. The left panel reports the results using Earnings Change, while the right panel reports the results using 9 Kothari et al. (2006) also repeat their tests using earnings on the S&P 500 going back to From 1936 to 1969, among the four variations of earnings and returns measures, they determine that one combination is negative and significant, while the other three combinations are not significantly different from zero.

11 Chen, Jiang, & Lee Long-Term Evidence on the Effect of Aggregate Earnings on Prices 333 Table III. Aggregate Quarterly Return and Quarterly Earnings Shock The table reports the slope estimate, t-statistic, and adjusted R-square (R 2 ) when quarterly stock returns are regressed on quarterly earnings shock: Rm t+1 = α + β de/p t +e t+1,whererm is the S&P 500 Index return, de/p is quarterly differenced earnings of S&P 500 Index scaled by the lagged S&P 500 Index price (P) as Earnings Change. Earnings Surprise is the residual when de/p is regressed on lagged de/p and lagged returns.er, Ncf, and Ndr represent expected return, cash flow news, and discount rate news components of stock return that are generated from VAR, respectively. The relation between the return and its components is as described in Equation (1): r t = E t 1 r t + N CFt N DRt.AdjR 2 reflects the joint explanatory power of both variables. Earnings Changes Earnings Surprise α β AdjR 2 α β AdjR 2 Panel A. 1871:Q1-2011:Q4 Rm PARMS t-stat ER PARMS t-stat Ncf PARMS t-stat Ndr PARMS t-stat Panel B. 1871:Q1-2000:Q4 Rm PARMS t-stat ER PARMS t-stat Ncf PARMS t-stat Ndr PARMS t-stat Panel C. 1871:Q1-1969:Q4 Rm PARMS t-stat ER PARMS t-stat Ncf PARMS t-stat Ndr PARMS t-stat Panel D. 1970:Q1-2000:Q4 Rm PARMS t-stat ER PARMS t-stat (Continued)

12 334 Financial Management Summer 2015 Table III. Aggregate Quarterly Return and Quarterly Earnings Shock (Continued) Earnings Changes Earnings Surprise α β AdjR 2 α β AdjR 2 Panel D. 1970:Q1-2000:Q4 Ncf PARMS t-stat Ndr PARMS t-stat Panel E. 2001:Q1-2011:Q4 Rm PARMS t-stat ER PARMS t-stat Ncf PARMS t-stat Ndr PARMS t-stat Earnings Surprise. Earnings Change is the actual value of de/p, and Earnings Surprise is the forecast error when de/p is regressed on lagged de/p and lagged market returns. 10 We discuss the results using earnings change first. Panel A reports the aggregate ERC using all of the observations for the full sample period. The results indicate that the overall aggregate ERC is positive and significant (1.806 with t = 2.612). This is consistent with the ERC at the firm level. To verify that the overall aggregate ERC is not driven by the unusual events in the last decade, we also examine the relation excluding the years from The results in Panel B indicate that the aggregate ERC (1.435 with t = 1.996) is also positive and significant when recent years since 2001 are excluded. The results demonstrate that the overall positive and significant aggregate ERC is not driven by the recent turbulence in the economy. In order to examine whether the relationship between aggregate earnings and market returns is affected by regulation regime changes and the economic environment and to make the analysis comparable to prior studies, we partition the whole sample into three subsample periods: 1) from 1871 to 1969, 2) from 1970 to 2000, and 3) from 2001 to In the subsample period from 10 In the earnings surprise case, both the fitted value and the forecast error from the forecasting regression are included in the second-stage return regression in the same way as in Kothari et al. (2006). For simplicity and comparability, we report the coefficient on the earnings surprise only. The R 2 reflects the joint explanatory power of both variables. 11 Aboody, Hughes, and Ozel (2014) find that the role of fair value had a significant impact on the correlation between earnings and bond returns during the recent financial crisis around In addition, firms have significant goodwill write-offs from 2001 to We partition the sample into three subsamples based on a few considerations. One is to match the sample period in prior studies, such as Kothari et al. (2006) and Sadka and Sadka (2009), and the other is to match the periods where significant stock market regulations are in effect. For example, the period from 1970 to 2000 is to match the sample period in Kothari et al. (2006), while the period from 2001 to 2011 is to match the post-regulation FD and post-sox periods. We lump together the years from 1871 to 1969 as one period due to the similarities in the data quality. For example, the stock market data quality prior to 1926 is generally considered to be not as reliable as the data in later years. The years from 1926 to 1950s include the WWI, WWII, and the Korean war periods.

13 Chen, Jiang, & Lee Long-Term Evidence on the Effect of Aggregate Earnings on Prices to 1969, the aggregate ERC (2.367 with t = 2.999) is positive and significant, not different from that of the full sample period. However, in the subsample period from 1970 to 2000, the aggregate ERC is significantly negative ( with t = 2.402). This finding confirms the results in Kothari et al. (2006) that the aggregate ERC is negative during this sample period. Using Compustat data, Kothari et al. (2006) find the aggregate ERC is with t = In the subsample period from 2001 to 2011, aggregate ERC (5.621 with t = 2.316) becomes positive and significant again. This sample period is an extension of Kothari et al. (2006), and we find that the aggregate ERC has taken a dramatic temporal shift since Taken together, we find that the relation between aggregate earnings and returns is time varying. The VAR analysis of Campbell (1991) allows us to separately measure the three return components: 1) expected returns, 2) cash flow news (both with positive effects), and 3) discount rate news (with a negative effect). We denote the coefficient of expected return on earnings changes as E_ERC, the coefficient of cash flow news on earnings changes as CF_ERC, and the coefficient of discount rate news on earnings changes as DR_ERC. We then examine the relationship between earnings changes and return components by regressing each return component on earnings changes and obtaining three individual component ERCs (E_ERC, CF_ERC, and DR_ERC) for each subsample period. The regression results in Table III indicate that the component ERCs are time varying as well. Specifically, in the full sample period, all component ERCs, namely, E_ERC (0.534 with t = 4.338), CF_ERC (5.239 with t = 8.804), and DR_ERC (3.981 with t = ), are all significantly positive with CF_ERC larger than DR_ERC in magnitude, suggesting that the positive effects of E_ERC and CF_ERC dominate the negative effect of DR_ERC, resulting in overall positive ERC. The results of component ERCs are similar when the recent financial crisis period is excluded. In the subsample period from 1871 to 1969, E_ERC (0.502 with t = 3.521), CF_ERC (5.470 witht = 8.421), and DR_ERC (3.686 with t = ) are all significantly positive suggesting that the positive effects of E_ERC and CF_ERC dominate the negative effect of DR_ERC, resulting in positive ERC in the subsample period. In the subsample from 1970 to 2000, both E_ERC and CF_ERC are insignificantly different from zero. However, the DR_ERC (3.630 with t = 2.724) is significantly positive suggesting that the negative effect of DR_ERC is driving the aggregate negative ERC. This evidence provides empirical support for the conjecture in Kothari et al. (2006) that discount rate shocks explain a significant fraction of aggregate stock returns. In addition, cash flow news tends to react less to earnings shocks, partly because individual firms earnings shocks tend to be diverse and cancel out to some extent due to the considerable cross-sectional variation as shown in Bali, Ozgur, and Tehranian (2008), Vuolteenaho (2002), and Patatoukas and Yan (2010). However, it should be noted that this relation holds only in this subsample period of In fact, in the recent subsample period from 2001 to 2011, both E_ERC (1.316 with t = 4.301) and CF_ERC (4.853 with t = 3.404) are significantly positive, and DR_ERC (0.264 with t = 0.211) is not significantly different from zero. These results suggest that in this recent sample period, the aggregate ERC is mainly driven by positive E_ERC and CF_ERC. This result is consistent with Ball et al. (2009) and Sadka and Sadka (2009), indicating that the expected returns explain a large portion of the relation in this period. This time period constitutes the period around and after the enactment of the SOX Act of 2002, whose objective was to restore the integrity of financial reporting by curtailing earnings management (Cohen et al., 2008). As a result, recent research provides evidence suggesting these new regulations have been successful in improving the quality of firm disclosure and restoring confidence in the capital market (Cohen et al., 2008; Koh et al., 2008; Kolev et al., 2008). It is not surprising that with the improved quality

14 336 Financial Management Summer 2015 of firm disclosure, investors are better able to extract information to form earnings expectations in the last decade or so. To test the robustness of the results, we also conduct an array of additional tests. The right panel of Table III reports the regression results using earnings surprises. The magnitude of the coefficients and significance levels are similar to those in the regressions using earnings changes except for E_ERC. E_ERC is not significant as, by construction, expected return is orthogonal to earnings surprises. However, estimates of CF_ERC and DR_ERC are very similar to those of earnings changes. In addition to using de/p, which scales earnings changes with market price, we also conducted another set of tests using de/b that scales earnings changes with the book value of equity. The untabulated results are similar to those in Table III in terms of both the magnitude of the coefficients and the significance levels. Therefore, our results remain robust whether we use earnings changes or earnings surprises, and whether the earnings changes are scaled by market price or book value. These findings suggest that both aggregate ERC and component ERCs are time varying and the aggregate ERC is determined by the relative importance of component ERCs in different sample periods. C. Contemporaneous Relation between Earnings and Return Components (Multivariate Analysis) In the following set of tests, we examine how the relationship between earnings and all return components changes over time by regressing earnings changes on three return components together for each subsample period: de/p = a + β 1 ER + β 2 Ncf + β 3 Ndr, (7) where de/p, ER, Ncf, and Ndr are as defined above. Although this may not be the conventional way to examine ERC, we are mainly interested in the contemporaneous relation between earnings and return components. In this multivariate regression, we are able to compare the relative importance among three return components in explaining earnings changes. Table IV reports the multivariate regression results of Equation (7). In the full sample period (results shown in Panel A), β 1 (0.061 with t = 5.107), β 2 (0.019 with t = 7.779), and β 3 (0.044 with t = ) are all significantly positive, suggesting all three return components are correlated with earnings changes. When the observations from the recent financial crisis period are excluded (Panel B), it does not change the results qualitatively. In the subsample period from 1871 to 1969 (Panel C), β 1 (0.061 with t = 4.311), β 2 (0.025 with t = 8.501), and β 3 (0.051with t = ) are all significantly positive. In the subsample period from 1970 to 2000 (Panel D), the β 2 ( with t = 7.030) is significantly negative, β 3 (0.089 with t = 7.725) is significantly positive, and β 1 (0.002 with t = 0.060) is not significantly different from zero. This result not only is consistent with Kothari et al. (2006), who find that discount rate news drives the aggregate ERC in this sample period, but also reveals that cash flow news also plays a role in this period with a negative effect. It is interesting to note that in the recent subsample period from 2001 to 2011 (Panel E), β 1 (0.203 with t = 4.509), β 2 (0.057 with t = 4.667), and β 3 (0.048 with t = 3.095) are all significantly positive. However, β 1 in this sample period is much larger in magnitude than those in other sample periods suggesting that the aggregate ERC in this period is dominated by the expected return. Cash flow news effect and discount rate news effect may offset each other in this period. This is consistent with Sadka and Sadka (2009), who suggest that price leads earnings and the expected earnings explain the overall

15 Chen, Jiang, & Lee Long-Term Evidence on the Effect of Aggregate Earnings on Prices 337 Table IV. Regression of Earnings on Stock Return Components The table reports the slope estimates, t-statistics, and adjusted R-square (R 2 ) when quarterly differenced earnings are regressed on stock return components: de/p t = α + β 1 ER t + β 2 Ncf t + β 3 Ndr t,wherede is the quarterly differenced earnings of the S&P 500 Index and P is the market value of the S&P 500 Index. ER, Ncf, and Ndr represent expected returns, cash flow news, and the discount rate news components of stock returns that are generated from VAR, respectively. α β 1 β 2 β 3 R 2 Panel A. 1871:Q1-2011:Q4 PARMS t-stat Panel B. 1871:Q1-2000:Q4 PARMS t-stat Panel C. 1871:Q1-1969:Q4 PARMS t-stat Panel D. 1970:Q1-2000:Q4 PARMS t-stat Panel E. 2001:Q1-2011:Q4 PARMS t-stat aggregate ERC. It also reflects the fact that investors have better information from improved firm disclosure after Reg FD and the enactment of the SOX Act. Overall, the findings from the multivariate analysis are consistent with those in Table III and imply that the relationship between earnings and return components is time varying. The timevarying ERCs are determined by the relative importance of the expected returns, cash flow news, and discount rate news. It is worth noting that the coefficient of cash flow news is always positive and significant except for the subsample period of in which it is negative, consistent with the results in Table III. After 2000, expected return dominates both cash flow news and discount rate news. This may be due to the improvement in the quality of firm disclosure and investor confidence after the enactment of the SOX Act. Our evidence is compatible with that provided by Kothari et al. (2006) and Sadka and Sadka (2009). D. Discussions on VAR Methodology Although the VAR methodology in Campbell (1991) enables us to estimate the return components separately and identify the role of each component in understanding aggregate ERCs, the

16 338 Financial Management Summer 2015 methodology is contingent on the VAR model employed for the analysis. In particular, the proxy for cash flow news based on the indirect approach relies on VAR specification. In addition, the inclusion of earnings in the state variables in the VAR model may cause a potential mechanical relation between cash flow news and earnings change. Chen and Zhao (2009) also suggest that estimating cash flow directly can yield the opposite conclusions (Ball et al., 2009). To address these concerns, we run a set of robustness tests using different measures of cash flow news. We use the changes in the future return on equity ( ROE) as our first proxy for cash flow news. We report the correlation between cash flow news and earnings shock in Table V. Panel A of Table V delivers two main messages. First, the magnitude of cash flow news ERC and its explanatory power for market returns are sensitive to the measures of cash flow news. In the original indirect approach based on the VAR model, the cash flow news ERC (adjusted R-square) ranges from ( 0.008) to (0.115). Using the direct approach based on the changes in future ROE (Panel A), the cash flow news ERC (adjusted R-square) ranges from (0.142) to (0.105). However, the different measure of cash flow news provides consistent results in the sense that cash flow news ERCs are indeed time varying. Particularly, the cash flow news ERC tends to be significantly positive in the early subsample periods, while it tends to be weak or insignificant in later subsample periods. When we replace earnings changes with earnings surprise as a proxy for earnings shock, the results are qualitatively unchanged. In addition, to mitigate the potential mechanical relation in this context, we adopt the changes in the one-period-ahead dividend consumption ratio DC as a proxy for cash flow news and report the results in Panel B of Table V. 13 We find a similar pattern in that the cash flow news ERC is time varying and tends to be significantly positive in the full sample period and early subsample periods, while it is insignificant or weak in late subsample periods. Moreover, since one period change in future ROE ( DC) is subject to short-term fluctuations, Campbell et al. (2010) propose a present value approach to lengthen the horizon to emphasize longer term trends that correspond more closely to the revisions in the infinite horizon expectations that are relevant for stock prices. Following Campbell et al. (2010), we adopt the present value of changes of future three years (12 quarters ) ROE ( DC) as a proxy for cash flow news. Again, we find a similar pattern, as shown in Panels C and D of Table V, in that the cash flow news ERC is time varying and tends to be significantly positive in early subsample periods, while it is insignificant or weak in later subsample periods. In addition, following the third approach above, as a robustness check, we also consider horizons of two years, four years, and five years, and the untabulated results show similar patterns. E. Dynamic Relation between Earnings and Return/Return Components Previous studies in aggregate ERC tend to focus on contemporaneous relations between aggregate earnings changes and market returns. However, dynamic relations between aggregate earnings changes and market returns provide further insight into the aggregate ERC. For example, time-varying ERC and price leads earnings reflect dynamic relations in nature. Dynamic relations also help us understand the time-varying aggregate ERC in relation to business cycles (Chen, 1991). Therefore, in addition to the contemporaneous relationship between earnings changes and returns, we also explore dynamic causal relations between earnings changes and returns (or return components). We conduct Granger-causality tests to examine whether price leads earnings, earnings lead price, or both. 13 Menzly et al. (2004) and Lettau and Ludvigson (2005) find that the dividend consumption ratio is able to predict future dividend growth.

17 Chen, Jiang, & Lee Long-Term Evidence on the Effect of Aggregate Earnings on Prices 339 Table V. Aggregate Quarterly Cash Flow News and Quarterly Earnings Shock The table reports the slope estimates, t-statistics, and adjusted R-square (R 2 ) when different quarterly cash flow news proxies are regressed on quarterly earnings shocks: N CF t+1 = α + β de/p t + e t+1,wherede/p t is the quarterly differenced earnings of the S&P 500 Index scaled by the lagged S&P 500 Index price (P)as Earnings Change. Earnings Surprise is the residual when de/p is regressed on lagged de/p and lagged returns. We use four different measures for the proxy of cash flow news (N cf ): 1) the change of future ROE in Panel A, 2) the change of the future dividend consumption ratio in Panel B, 3) the present value of future ROE in Panel C, and 4) the present value of the future dividend consumption ratio in Panel D. R 2 reflects the joint explanatory power of both variables. Earnings Changes Earnings Surprise Sample Period α β R 2 α β R 2 Panel A. Ncf = Change of Future ROE 1871:Q1-2011:Q4 PARMS t-stat :Q1-2000:Q4 PARMS t-stat :Q1-1969:Q4 PARMS t-stat :Q1-2000:Q4 PARMS t-stat :Q1-2011:Q4 PARMS t-stat Panel B. Change of Future Dividend Consumption Ratio 1871:Q1-2011:Q4 PARMS t-stat :Q1-2000:Q4 PARMS t-stat :Q1-1969:Q4 PARMS t-stat :Q1-2000:Q4 PARMS t-stat :Q1-2011:Q4 PARMS t-stat Panel C. Present Value of Change of Future ROE 1871:Q1-2011:Q4 PARMS t-stat :Q1-2000:Q4 PARMS t-stat :Q1-1969:Q4 PARMS t-stat :Q1-2000:Q4 PARMS t-stat :Q1-2011:Q4 PARMS t-stat (Continued)