Do Aggregate Analyst Recommendations Predict Future Aggregate Discount Rates? Bruce K. Billings Florida State University

Transcription

1 Do Aggregate Analyst Recommendations Predict Future Aggregate Discount Rates? Bruce K. Billings Florida State University Sami Keskek Florida State University Spencer Pierce Florida State University July 2017 ABSTRACT We document evidence suggesting that analysts incorrectly process current aggregate information about future aggregate discount rates. Current aggregate recommendation changes are negatively related to future excess returns in periods when discount rate changes have a pronounced effect in future returns. Consistent with this, aggregate recommendation changes are positively related to several measures of future discount rate changes. Lastly, we provide evidence that the negative relation between future excess returns and current aggregate recommendation changes is fully explained by the relation between recommendation changes and future discount rate change proxies. To the best of our knowledge, our study is the first to examine the relation between analysts aggregate recommendations and aggregate risk. JEL codes: A10, E10, E44, E52, G11, G14, M40, M41 Keywords: Aggregate recommendation changes, risk, aggregate discount rates, aggregate earnings growth We thank Landon Mauler, Rick Morton, Beverly Walther, and workshop participants at Florida State University for helpful comments and suggestions.

2 1. Introduction While firm-level analyst recommendations have been extensively examined, only recently have studies considered whether aggregate recommendations inform investors about higher-level news (i.e., market- and industry-level). 1 Howe, Unlu, and Yan (2009) report that aggregate recommendation changes predict aggregate earnings growth and future excess returns, consistent with analysts incorporating information about aggregate cash flows in their recommendation changes that investors appear slow to process. This interpretation is based on their evidence that aggregate earnings growth and excess returns are positively related. Using Campbell s (1991) return decomposition, this indicates that cash flow news dominates discount rate news in that relation during their sample period. This contrasts with prior evidence that discount rate news has historically dominated cash flow news in the aggregate growth-returns relation (Kothari, Lewellen, and Warner, 2006; Sadka and Sadka, 2009). Two questions, then, naturally arise: do analysts recommendation changes predict future aggregate discount rates and do investors similarly process this information with delay? Kothari, So, and Verdi (2016) highlight that little evidence linking analysts outputs and expected returns (i.e., discount rate news) exists. We provide evidence that helps to resolves these questions. 2 There are several reasons why aggregate recommendation changes should predict future aggregate risk. First and foremost, recommendation changes should be positively related to revisions in analysts expectations about future returns, which are inversely related to future discount rates (Campbell, 1991). Analysts industry expertise should facilitate their ability to 1 The predictive value of analyst recommendations for future returns at the firm level has been well established in studies that include Stickel (1995), Womack (1996), Barber, Lehavy, McNichols, and Trueman (2001), Jegadeesh and Kim (2006), Green (2006), and Barber, Lehavy, and Trueman (2010). 2 Our focus on recommendation changes at the aggregate level is motivated by the fact that discount rate changes are likely to be more important than cash flow changes at this level. Kothari et al. (2006) argue that cash flow changes are largely idiosyncratic, and so subject to the effects of diversification at the aggregate level. Discount rate changes are likely strongly correlated across stocks and so will make up a larger portion of aggregate returns. 1

3 predict aggregate risk (Gilson, Healy, Noe, and Palepu 2001). Analysts have been shown to incorporate macroeconomic news in their forecasts (Hugon, Kumar, and Lin 2016), albeit inefficiently, and macroeconomic news is predictive of future aggregate risk (Gallo et al. 2016). Analysts engage in risk assessments, and one might reasonably expect they would incorporate those risk assessments in their recommendations. 3 Further, the slow response of investors to cash flow news (Howe et al. 2009) and robust evidence of post-recommendation drift (Womack 1996) suggest that investors may similarly respond slowly to discount rate information in recommendation changes. 4 If aggregate recommendation changes include discount rate news that investors are slow to process, the positive relation between aggregate recommendation changes and future excess returns should persist even in periods when discount rate news dominates cash flows news. However, there are reasons to doubt that analysts recommendations correctly predict future aggregate risk. Firm-level empirical evidence indicates that analyst recommendations rely less on valuation estimates and more on heuristics (Bradshaw, 2004; Block, 1999), raising fundamental concerns about whether and how risk estimates (i.e., discount rates) are used to arrive at recommendations. Further, their lack of macroeconomic expertise (Hugon et al., 2016) may inhibit analysts from correctly processing macroeconomic news as it relates to future aggregate risk. Even more uncertain is whether analysts risk assessments correctly predict discount rates, particularly given evidence that discount rates are difficult to predict (Fama and French, 1997). Kothari et al. (2006) fail to observe that aggregate earnings growth is predictive of future returns when discount rate news dominates, leading them to conclude that investors 3 For example, FINRA Rule 2241 requires that member analysts disclose a fair presentation of risks that may impede the achievement of their recommendations. 4 Cready and Gurun (2010) suggest that investors do not fully process discount rate information at the earnings announcement date. Further supporting the contention that investors may be slow to respond to risk-related information, Grullon and Michaely (2004) observe stock price drift following changes in risk related to repurchase announcements. 2

4 neither under- nor over-react to discount rate news in aggregate earnings growth. Whether aggregate recommendation changes predict future discount rates and whether investors process this information slowly remain empirical questions that we address. Kothari et al. (2006) posit that the sign of the aggregate earnings growth-returns relation is an indicator of when discount rate news dominates cash flow news in returns. Relying on this general categorization, we employ three different approaches to identify periods when discount rate news is more likely to dominate the growth-returns relation: (1) the period , based on evidence that the dominance of cash flow news in the growth-returns relation during the early 2000s was temporary (Sadka and Sadka, 2009; Jorgenson, Li, and Sadka, 2009), (2) a portfolio approach based on the sign of the current growth-current returns relation, and (3) a second portfolio approach based on the sign of the future growth-future returns relation. We begin by estimating the three relations described in Howe et al. (2009) that characterize the information content of analysts recommendation changes: aggregate recommendation changes and future aggregate earnings growth (recommendation-future growth), aggregate earnings growth and excess returns (growth-returns), and aggregate recommendation changes and future excess returns (recommendation-future returns). We confirm that all the three relations are positive during their sample period. We then estimate the three relations for the period. Here, we continue to observe a positive recommendations-future growth relation, indicating that the sign of that relation appears timeinvariant. However, we fail to find a significant earnings growth-returns relation at either market- or industry-level aggregation, consistent with cash flow news and discount rate news offsetting during this period (Patatoukas, 2014). More surprisingly, the recommendations-future returns relation is negative. Given Howe et al. s (2009) and our initial evidence that analysts 3

5 correctly process cash flow news, we suspect that the negative recommendations-future returns relation may be related to analysts expectations regarding discount rate changes. We investigate this further using two other approaches to better isolate periods where information about future discount rates dominates. Our second approach uses the sign of the current growth-returns relation as an indicator of the probability of future discount rates changes. A positive current growth-returns relation indicates that cash flow shocks dominate the current period news, which increases the probability of an increase in future discount rate (Kothari et al., 2006; Shivakumar, 2007; Gallo et al., 2016). 5 In contrast, a negative current growth-returns relation indicates that discount rate shocks dominate the current period news, which we argue decreases the probability of further discount rate changes in the near future. Consistent with this argument, we find that quarters dominated by cash flow (discount rate) shocks are more likely to be followed by quarters dominated by discount rate (cash flow) shocks. We form portfolios by quarter based on the sign of the current growth-returns relation and estimate the recommendation-future returns relation within each portfolio. Results confirm that when the current growth-returns relation is negative (i.e., cash flows are likely to dominate in the next period), we observe a positive recommendations-future returns relation. This is consistent with recommendation changes correctly predicting the positive effects of current shocks for future cash flows, however investors are slow to process this information. However, when the current earnings growth-returns relation is positive (i.e. discount rate changes are more 5 Kothari et al. (2006) report a positive relation between aggregate earnings growth and changes in T-bill rates, a proxy for discount rate changes. Similarly, Shivakumar (2007) documents that aggregate earnings growth is positively related to future inflation. As we describe later, we also observe positive correlations between current earnings growth and our measures of future discount rate changes and negative serial correlation in quarterly earnings growth-returns coefficients, supporting this claim. Gallo et al. (2016) state that the Federal Open Market Committee, the Federal Reserve s primary monetary policymaking body, is likely to increase the Federal funds target rate following periods when economic performance is stronger than expected. 4

6 likely in the next period), the recommendation-future returns relation is negative. This result is consistent with analysts and investors incorrectly predicting the effects of period t cash flow shocks on future discount rate changes, and investors recognizing those discount rate effects in period t+1 returns. 6 Our third approach is similar to the second except we use the sign of the future earnings growth-returns relation as an ex post but more direct (rather than a likely) indicator that cash flow or discount rate changes have a more prominent effect on excess returns in period t+1. While it is a direct indicator, we acknowledge that this ex post measure cannot be fully predicted using information available to analysts in period t. However, if period t aggregate recommendation changes predict some portion of those future discount rate changes that investors fail to incorporate in returns until period t+1, we should observe a positive recommendations-future returns relation. We form portfolios based on the sign of the future growth-future returns relation by quarter using industry-level aggregation. Within each portfolio, we estimate the recommendation-future returns relation. 7 Results indicate that when future returns are primarily affected by cash flows, the recommendation-future returns relation is positive, consistent with Howe et al. (2009). However, when future returns are primarily affected by discount rate changes, the recommendation-future returns relation is negative. This negative relation is consistent with our prior evidence, further supporting our conclusion that analysts and investors incorrectly predict the effects of period t cash flow shocks on future discount rate changes, and investors begin to recognize those discount rate effects in period t+1 returns. 6 An expected increase (decrease) in discount rates should result in a decrease (increase) in both stock returns and analysts recommendations. 7 Partitioning the sample by the sign of the future growth-returns relation and regressing future excess returns on current recommendation changes within each partition does not truncate the sample by the sign of the dependent variable, future returns. We observe trivial correlations of and between the future growth-returns coefficients and future excess returns for market-level and industry-level aggregation, respectively. 5

7 These findings prompt us to more directly examine whether aggregate recommendation changes are predictive of future discount rate changes. We estimate four discount rate proxies based on prior research (e.g., Fama and French, 1989; Kothari et al., 2006; Gallo et al., 2016): future T-bill rate changes, future changes in default spread on corporate debt, future Federal funds target rate changes, and future effective Federal funds rate surprises. 8 If analysts correctly predict some part of future discount rate changes, recommendation changes should be negatively related to future discount rate changes. However, we observe positive relations for all four. Finally, we more directly tie our evidence of analysts failure to correctly predict the future discount rate changes accounts to our prior returns evidence by re-estimating the recommendation-future returns relation after including these future discount rate proxies, focusing on the contexts where we observe the negative recommendation-future returns relation. We find that the negative recommendation-future returns relation becomes insignificant after including future discount rate change proxies, with future change in default spread on corporate debt producing the strongest effect. In particular, the disappearance of the negative recommendation-returns relation in periods identified as high discount rate risk using ex ante current growth-returns relation provides compelling evidence that aggregate analyst recommendation changes fail to correctly predict future discount rate changes. 9 Our results indicate that aggregate recommendation changes fail to predict that greater aggregate earnings shocks lead to a greater probability of a future discount rate changes. When investors subsequently adjust stock prices for future discount rate changes, a negative relation 8 The effective Federal funds rate is the weighted average interest rate banks charge each other for overnight loans of reserve balances. It is determined by the market but is largely influenced by the Federal Reserve, who use open market operations to communicate their funds rate target. Prior studies document a negative association between stock returns and changes in target rates (Jensen, Mercer, and Johnson, 1996; Thorbecke, 1997; Jensen and Mercer, 2002). 9 These results support our conclusion of analyst inefficiency over the alternative that the negative recommendations-future returns relation results from analysts correctly incorporating discount rate information that only investors incorrectly incorporate in period t but recognize in period t+1. 6

8 between the correlated error in current recommendations and subsequent stock return adjustment is revealed. One explanation for this behavior may be that analysts incentives to provide accurate forecasts result in a focus on predicting future cash flows to the exclusion of future discount rates. Another contributing factor may be analysts lack of macroeconomic expertise (Hugon et al., 2016). Evidence of investor inefficiency regarding expectations of future discount rates is consistent with evidence in Gallo et al. (2016). We make several contributions to the literature. First, we contribute to studies examining the relation between analyst outputs and returns. We provide empirical evidence that aggregate recommendation changes not only fail to correctly predict the effects of future discount rate changes on future excess returns, but they relate inversely. We interpret this to indicate that analysts focus more on predicting future aggregate cash flows than on aggregate discount rate changes when evaluating investment opportunities. We are unaware of any study investigating whether analysts provide information about aggregate risk. Our study addresses a gap in the literature noted by Kothari et al. (2016) who state that while the implications of analysts forecasts for cash flows is clear, and the empirical evidence is vast, the links between analysts forecasts and expected returns are less established, presenting a promising opportunity for future research. Second, we clarify and expand results in Howe et al. (2009) in a way that reconciles their findings with the developing literature examining the aggregate earnings-aggregate return relation. Specifically, their results are limited to sample periods when cash flow news dominates the future growth-returns relation, and their conclusions do not generalize to periods when future discount rate changes are expected. Third, we contribute to evidence in Hugon et al. (2016) that analysts lack macroeconomic expertise in predicting future earnings. Our results suggest that 7

9 their lack of expertise is more pronounced with respect to predicting future discount rates than future cash flows. We proceed as follows. In section 2, we discuss background literature. We follow this in section 3 with a description of our sample selection procedures, variable measurement, and initial statistics. Section 4 includes our research design and results. We conclude in section Background literature Our study relates to the large collection of studies that investigate the information content of accounting earnings and analyst recommendations. 10 The vast majority of these studies focus on firm-level earnings. Beginning with Ball and Brown (1968), evidence confirms a consistently positive association between earnings changes and stock returns at the firm level. Similarly, Stickel (1995), Womack (1996), and Barber et al. (2001) report a positive association between analyst recommendation changes and stock returns at the firm level. Further, these studies often show that recommendation changes predict future stock returns, which continue to move in the direction of recommendation changes in the months following their release. Campbell (1991) provides a theoretical framework for interpreting the relations between stock returns and both earnings changes and analyst recommendation changes. He decomposes excess stock returns into expected future cash flow revisions (i.e., cash flow news), with a positive sign, and expected future discount rate revisions (i.e., discount rate news), with a negative sign. Using his framework, the evidence in these firm-level studies is consistent with 10 See Kothari (2001) for a review of studies examining the information content of accounting earnings. See Ramnath, Rock, and Shane (2008) for a review of studies examining the information content of analyst recommendation. Somewhat related, studies like Anilowski, Feng, and Skinner (2007) examine the relations between management earnings guidance, aggregate earnings, and market returns. They report that management earnings guidance, particularly downward guidance, captures aggregate earnings news and is weakly related to market returns, consistent with guidance potentially having market-wide effects. 8

10 earnings changes and recommendation changes conveying cash flow news to investors, which investors appear slow to process. More recent studies have begun to investigate the information content of aggregate (i.e., market- and industry-level) earnings changes and analyst recommendation changes. Kothari et al. (2006) surprisingly observe a negative association between aggregate earnings changes and contemporaneous excess returns over the period, as well as various sub-periods. They conclude that while firm-level earnings changes are dominated by cash flow news, aggregate earnings changes are dominated by discount rate news. 11 Subsequent studies provide supporting evidence. Cready and Gurun (2010) show that the negative relation between aggregate earnings growth and excess returns exists at the earnings announcement date. Unlike Kothari et al. (2006), Cready and Gurun (2010) also document that investors are slow to process discount rate information. Patatoukas (2014) observes that about 28% of the time-series variation in aggregate earnings changes is explained by proxies for discount rate news and another 48% is explained by proxies for cash flow news. Further, the sign of the overall relation between aggregate earnings changes and excess returns depends on whether cash flow or discount rate news dominates. Gallo et al. (2016) demonstrate that aggregate earnings convey information about future changes in Federal funds rates (i.e., discount rate news), and this relation at least partially drives the negative relation between aggregate earnings changes and excess returns. They also show that investors are slow to process future fund rate change information in aggregate earnings, resulting in drift at the aggregate level. Relatedly, Shivakumar (2007) documents that aggregate earnings changes are positively related to future inflation. 11 Alternatively, evidence in Sadka and Sadka (2009) and Choi, Kalay, and Sadka (2016) suggests that the negative relation between aggregate earnings changes and excess returns results from a negative relation between expected aggregate earnings and expected aggregate returns. This interpretation is based on the claim that aggregate earnings changes provide no new information to investors but rather merely confirm their expectations. However, Patatoukas (2014) provides compelling evidence that aggregate earnings growth provides both cash flow news and discount rate news to investors. 9

11 There is limited evidence regarding the information content of aggregate analyst recommendation changes. Howe et al. (2009) show that aggregate recommendation changes predict both quarter-ahead aggregate earnings growth and quarter-ahead excess returns. They suggest these two results are related because analyst recommendations provide information about future aggregate earnings growth which investors are slow to process. Under the framework in Campbell (1991), their evidence is consistent with the conclusion that analyst recommendation changes conveying market- and industry-level cash flow news that investors are slow to process. We are unaware of any study that assesses the predictive ability of aggregate recommendation changes with respect to aggregate risk (i.e., discount rate news). The positive recommendation-future growth relation (Howe et al., 2009) and evidence that aggregate earnings growth reflects discount rate news (Kothari et al., 2006; Patatoukas, 2014) suggest that aggregate recommendation changes should convey discount rate news at least via their ability to predict future aggregate earnings growth. Analysts engage in risk assessments, and they should incorporate those risk assessments in their recommendations given the negative effects of discount rate news on stock returns. FINRA Rule 2241, in fact, requires that member analysts disclose a fair presentation of the risks that may impede the achievement of their recommendations. Gilson, Healy, Noe, and Palepu (2001) show that analysts possess industry expertise, which contributes to more accurate forecasting. 12 Industry expertise should enhance analysts ability to recognize threats and opportunities at the industry level, which, in turn should contribute to their ability to predict changes in aggregate risk. Additionally, Hugon et al. (2016) report that analysts earnings forecast revisions incorporate macroeconomic news about gross domestic product (GDP), albeit inefficiently, and macroeconomic news is arguably an input for 12 Also see Clement (1999) and Jacob, Lyz, and Neale (1999) for additional evidence that industry expertise contributes to better analyst forecast accuracy. 10

12 predicting aggregate risk (Gallo et al., 2016). 13 Under this scenario, aggregate recommendation changes will be positively related to future excess returns. That is, analysts upgrades (downgrades) should reflect an expected decrease (increase) in future discount rates and corresponding increase (decrease) in excess returns. However, analyst recommendations may not correctly incorporate aggregate risk. Bradshaw (2004) and Block (1999) provide evidence that analyst recommendations rely less on valuation estimates and more on heuristics, raising fundamental concerns about whether and how risk estimates (i.e., discount rates) are used to arrive at recommendations. Further, evidence in Hugon et al. (2016) suggests that lack of macroeconomic expertise may adversely affect analysts ability to correctly incorporate in their recommendations the implications of macroeconomic news for future aggregate risk. Even more uncertain is whether analysts risk assessments are predictive of aggregate risk (i.e., discount rates) in future excess returns, particularly given evidence that discount rates are difficult to predict (Fama and French (1997)). Failing to observe any correlation between current aggregate earnings growth and future excess stock returns, Kothari et al. (2006) conclude that investors neither over- nor under-react to discount rate news in aggregate earnings growth. So, whether aggregate recommendation changes are predictive of future discount rate changes, and whether investors are slow to process discount rate that information leading to predictable future returns, remain empirical questions that we address. 3. Sample selection, variable measurement, and descriptive statistics 3.1 Sample selection 13 Hutton, Lee, and Shu (2014) report that analysts earnings forecasts are more accurate than managers for firms whose fortunes are more highly influenced by macroeconomic factors such as GDP and energy costs. They argue that analysts may be more objective in evaluating the implications of economy-wide events for future performance. Further, they illustrate the fact that analysts reports include discussion of macroeconomic factors. 11

13 Our sample selection procedures closely follow those in Howe et al. (2009). We first obtain analyst recommendations for the U.S. firms in the I/B/E/S Analyst Recommendation Detail History file for the period from January 1994 to December We require the recommendations to satisfy the following criteria: (1) the recommendation must be associated with a CUSIP number and have a recommendation date; (2) the recommendation must be made by an analyst with a non-missing analyst code; (3) the firm must be in the Center for Research in Security Prices (CRSP) database during the month of recommendation; and (4) the firm must have a share code of 10 or 11 and a non-missing SIC code on CRSP. We use the eight-digit CUSIP numbers to match I/B/E/S recommendation data to the CRSP database to obtain value-weighted market returns and selected firm characteristics including share price, shares outstanding, and SIC codes. We then use the SIC codes to group firms into the 48 industries in Fama and French (1997). For our industry-level analyses, we use the monthly value-weighted Fama-French industry returns from Kenneth French s website. 14 We measure aggregate market-level and industry-level earnings growth using quarterly earnings numbers from Compustat. As in Howe et al. (2009), we control for several macroeconomic indicators including the risk free rate, dividend yields, default spread, and term spread. We measure the risk free rate as the three-month Treasury bill rates and calculate the term spread as the difference between 10- year Treasury bond yields and three-month Treasury bill rates. The default spread is calculated as the difference between Moody s Baa corporate bond yield and the 10-year Treasury bond

14 yield. 15 Finally, we obtain monthly dividend yields for the S&P 500 index from Amit Goyal s website Variable measurement and descriptive statistics We measure market-level and industry-level changes in recommendations as in Howe et al. (2009). Specifically, I/B/E/S assigns a score of 1 to strong buy recommendations and a score of 5 to sell recommendations. We first reverse the ordering so that more favorable recommendations have larger numbers. We then calculate the average of the most recent outstanding recommendations from all analysts for all stocks in each month. 17 We estimate the aggregate market-level recommendation change (CHG_REC) as the monthly change in average recommendations for each month between January 1994 and December CCC_RRR t = 1 N t N t n=1 RRRRRR n,t 1 N t 1 RRRRRR N m=1 m,t 1 t = 1,,252, (1) t 1 where N t and N t 1 are the total number of recommendations across analysts and stocks in month t and t-1, respectively. Similarly, we find the average recommendation at the industry-level and calculate the monthly change for each industry-month 1994:01 and 2014: Next, we calculate aggregate earnings growth (EG) in each quarter. We first obtain seasonally differenced quarterly income before extraordinary items for each firm (de). Following Kothari et al. (2006), we require firms to have a March, June, September, or December fiscal year end to ensure that fiscal years are aligned. We then calculate the aggregate earnings growth as the cross-sectional sum of de scaled by the cross-sectional sum of market capitalization four 15 We obtain the three-month Treasury bill rates, 10-year Treasury bond yields, and Moody s Baa corporate bond yields from the Federal Reserve Bank of St. Louis As in Howe et al. (2009), we restrict the sample to recommendations issued during the past 12 months to eliminate stale recommendations. 18 We require analyst recommendations for at least 10 firms in an industry-month for the industry to be included in our analyses in the month. 13

15 quarters ago, where market capitalization (MCAP) is calculated as shares outstanding times share price. N q N q EE = de i=1 i,q / MMMM i=1 i,q 4 (2) See the Appendix for variable definitions. Table 1, Panel A reports the descriptive statistics for variables in our analyses. Specifically, we find that the mean and the median CHG_REC is and 0.001, respectively. These are nearly identical to the mean (-0.001) and median (0.002) reported in Howe et al. (2009). CHG_REC has a standard deviation of 0.018, which suggests substantial variation over our sample period. Figure 1 plots the market-level CHG_REC. Consistent with Howe et al. (2009), we observe a downward spike in September The downward spike occurs simultaneously with the implementation of NASD Rule 2711, which requires brokerage firms to define their ratings and requires that definition to be consistent with its meaning. Following Howe et al. (2009), we eliminate the observations around the implementation of NASD Rule 2711 to mitigate its effect on the results. In Panel B, we report the Pearson correlations among our key variables. Specifically, we find that CHG_REC is positively correlated with earnings growth in both current and next quarters, and with future 3-month excess market returns. We also find a positive correlation between future excess returns and both current and next quarter earnings growth. As in Howe et al. (2009), we use the average CHG_REC over the past three months in our multivariate analyses. 4. Research design and results 4.1 Regression models 14

16 Three relations form the foundation for our analysis: the relations between current aggregate recommendation changes and future earnings growth (recommendation-future growth), earnings growth and excess returns (future growth-future returns), and current aggregate recommendation changes and future excess returns (recommendations-future returns). Each is expressed in the following regression models. Recommendation-future growth: FEG t+1 = a + δ * CHG_REC t + υ (3) Future growth-future returns: FRET3m t+1 = a + β * FEG t+1 + u (4) Recommendation-future returns: FRET3m t+1 = α 0 + α 1 * CHG_REC t + α 2 * CRET3m t + ε (5) As in Howe et al. (2009) and Jagadeesh, Kim, Krische, and Lee (2004), our analysis in equations (3) and (5) focuses on predicting future growth and future returns, respectively, in the next quarter. FRET3m is future excess market returns, measured as the value-weighted market returns minus the risk free rate for three-month horizon beginning the month following the CHG_REC date. In equation (5), we include current excess market returns (CRET3m) for the three-month horizon ending in the month of the CHG_REC date to control for possible autocorrelation in market returns. As in Howe et al. (2009), we assess the incremental predictive ability of CHG_REC for future excess returns by estimating equation (5) after including macroeconomic variables linked in prior studies to market expected returns (e.g., Fama and Schwert, 1977; Keim and Stambaugh, 1986; Campbell, 1987; Fama and French, 1988). In particular, we include dividend yield (DIVIDEND) of the S&P500 index, the three-month Treasury bill rate (TB3M), default yield (DEFAULT) calculated as the difference between 15

17 Moody s Baa corporate bond yield and ten-year Treasury bond yield, and the term premium (TERM) calculated as the difference between long-term government bond yield and three-month Treasury bill rate. To address potential serial correlation and conditional heteroscedasticity in residuals, we use Newey West (1987) standard errors when estimating all equations, consistent with Howe et al. (2009). 4.2 Time-period results We begin by estimating the three fundamental relations represented by equations (3), (4), and (5) for Howe et al. s (2009) sample period. An important relation for interpreting the information content of aggregate recommendation changes is the relation between aggregate earnings growth and excess returns. Howe et al. (2009) observe a positive future growth-future returns relation during their sample period, consistent with cash flow news dominating this relation. However, Jorgenson et al. (2009) argue that negative shocks to earnings in 2001 and reversals in 2003 related to initial implementation of goodwill reforms are the primary contributor to the positive relation. When data from 2001 and 2003 are excluded from their sample period, Jorgenson et al. (2009) report that the sign of the growth-returns relation switches from positive to negative. Based on this evidence, our initial effort to identify a period when discount rate news dominates the future growth-returns relation is to estimate the three fundamental relations for the period subsequent to the effects of goodwill reform, Results from estimating the recommendations-future growth relation (equation (3)) for our sub-periods is presented in Table 2. Panel A and Panel B report results for market-level and 19 Identifying periods where the sign of the future growth-returns relation is negative represents an ex post approach to more directly identify periods where there is a higher probability of substantial future discount rate changes. This should enhance our ability to determine whether aggregate recommendation changes are predictive of these future discount rate changes. 16

18 industry-level aggregation, respectively. In both panels, column (2) results for Howe et al. s (2009) period confirm their evidence that aggregate recommendations predict quarter-ahead aggregate earnings growth. Further, positive coefficients are observed for CHG_REC in columns (1), the full sample period, and (3), , while statistical significance at conventional levels for occurs only for industry-level aggregation. Overall, however, results generally indicate that the sign of the recommendation-future growth relation is positive and time invariant. We next report results from estimating the future growth-future returns relation (equation (4)) in Table 3. Since earnings and returns data are readily available for , we begin by reporting those results in column (1) to serve as a comparison to Kothari et al. (2006). The 4.77 coefficient (p-value of 0.08) for market-level aggregation, comparable to the 2.35 coefficient (p-value < 0.10) reported in Kothari et al. (2006), illustrates that discount rate news generally dominated the growth-returns relation prior to In more recent years, evidence of a positive coefficient for FEG in column (3) for at both market-level (Panel A) and industrylevel (Panel B) aggregation confirms results in Howe et al. (2009). However, in column (4) we observe statistically insignificant coefficients for FEG in Based on evidence in Patatoukas (2014), we interpret the lack of statistical significance as the result of a relatively equal balance of discount rate news and cash flow news that offset in the growth-returns relation. 20 This supports evidence in Jorgenson et al. (2009) that the dominance of cash flow news in the growth-returns relation during Howe et al. s (2009) sample period is time-period specific. 20 Patatoukas (2014) presents evidence of both positive cash flow news and negative discount rate news affecting the relation between aggregate earnings growth and stock returns, even though the correlation between stock returns and aggregate earnings growth is insignificant. This illustrates the offsetting effects of cash flow news and discount rate news in the growth-returns relation. 17

19 Table 4 reports results from estimating the recommendations-future returns relation (equation (5)). We also include various macroeconomic variables that have been linked to expected aggregate returns in prior studies. Positive coefficient estimates for CHG_REC in column (2) of both Panel A and Panel B are consistent with those in Howe et al. (2009), indicating that market- and industry-level aggregate recommendation changes, respectively, predict cash flow news in quarter-ahead excess returns. Results for control variables are also consistent with results in Howe et al. (2009). 21 In contrast, however, we observe negative coefficients for CHG_REC that are statistically significant in for both market- (Panel A) and industry- (Panel B) level aggregation. Our initial evidence suggests that analysts correctly process cash flow news, so we suspect that the negative recommendations-future returns relation for may be related to analysts expectations regarding discount rate changes. We investigate this further using alternative approaches that better isolate periods when future discount rate changes are more likely. 4.3 Portfolio analysis results Our first portfolio analysis approach is based on Gallo et al. (2016) and Shivakumar (2007), who note that periods with positive macroeconomic shocks are likely to lead to periods of higher expected risk. Gallo et al. (2016) state that the Federal Open Market Committee of the Federal Reserve is likely to increase the interest rate banks charge each other for overnight loans of reserve balances (i.e., the Federal funds rate) following periods when macroeconomic performance (i.e., GDP) is stronger than expected. In support, they provide evidence that positive (negative) aggregate earnings news is predictive of increases (decreases) in the Federal funds 21 Where we can reasonably make comparison, our adjusted R 2 values compare favorably with those reported in Howe et al. (2009). For example, the adjusted R 2 of 14.7% we report in column (2) of Table 4, Panel A for their sample period is similar to the 16.9% reported in model (5) of their Table 4. 18

20 rate. 22 Shivakumar (2007) documents a positive relation between aggregate earnings changes and future inflation, consistent with earnings surprises, which are positively correlated with cash flow surprises, being an indicator of a higher probability of future discount rate changes. This suggests that the probability of a future discount rate change increases in periods following those where the dominant news is about cash flows. In contrast, we argue that in periods where the dominant news is about discount rate changes, the probability of an additional discount rate change in the near future is lower. Consistent with this argument, we find that quarters dominated by cash flow (discount rate) shocks are more likely to be followed by quarters dominated by discount rate (cash flow) shocks. 23 We form portfolios based on the sign of the current growth-returns relation estimated for each quarter of the sample period using the following first-stage model. CRET3m = a + γ * CEG + u (6) Current quarters with a positive coefficient (γ > 0, indicating cash flow news dominates) identify periods where the probability of a future discount rate change increases, while current quarters with a negative coefficient (γ < 0, indicating discount rate news dominates) are less likely to be followed by a change in future discount rate. Descriptive statistics from estimating equation (6) are reported in Table 5 Panel A. Cash flow news dominates the current quarter more frequently (59.6%) than discount rate news (40.4%) over our sample period. Within each portfolio, we estimate the recommendation-future returns relation (equation (5)). Industry-level results reported in Panel B indicate that where discount rate news dominates the current quarter (columns (1) (3)), the recommendation-future returns relation is positive, 22 Similarly, we observe positive correlations between current earnings growth and our proxies for future discount rate changes, which we describe later. 23 In support, we observe negative serial correlation (-0.17 significant at <0.10) in the coefficient relating aggregate earnings growth to excess returns, consistent with a higher likelihood that significant discount rate changes will follow periods when cash flow news dominates. 19

21 consistent with industry-level aggregate analyst recommendation changes being predictive of future aggregate cash flows, and investors are slow to respond. 24 However, when cash flow news dominates the current quarter (columns (4) (6)), the recommendation-future returns relation is negative. One interpretation for this negative relation is that analysts and investors correctly predict the implications of current cash flow shocks for future cash flows but fail to recognize that positive shocks also increase the likelihood of future discount rate increases. When investors recognize the increase in discount rates in the subsequent period by lowering stock prices, the error in current aggregate recommendation changes correlates negatively with the subsequent stock price adjustment. To assess the robustness of the negative recommendation-future returns relation, we employ a second approach which more directly identifies periods when future stock returns incorporate substantial discount rate changes as indicated by a negative future earnings growthreturns relation. While this ex post approach enables us to directly identify periods in which discount rate changes are a more significant component of future returns, we do not presume that analysts have perfect foreknowledge. Rather, if some predictable portion of these future discount rate changes is incorporate in current recommendation changes, and investors respond to this information slowly, we expect a positive recommendation-future returns relation. Alternatively, our prior evidence causes us to expect a negative relation. As a first stage, we estimate equation (4) for each quarter of the sample period. Consistent with Campbell (1991) and Kothari et al. (2006), we interpret quarters with a negative (positive) future growth- returns relation as an indication that discount rate changes (cash flows) dominate that period s returns. 24 At market-level aggregation, we also observe coefficient signs for CHG_REC consistent with those reported in Panel B for industry-level aggregation, however, results are generally insignificant. Low power from our small sample size is a substantial contributor. 20

22 Results reported in Table 6 Panel A indicate that cash flow news dominates more frequently (58.4%) than discount rate news (41.6%) during our sample period. We form portfolios based on the sign of β from our quarterly estimates of equation (4). The portfolio of periods with β < 0 (β > 0) indicates that discount rate (cash flow) changes have a more significant effect in future returns. As a second stage, we estimate the recommendationfuture returns relation (equation (5)) within each portfolio. 25 At industry-level aggregation (Panel B), we observe consistently negative coefficients relating CHG_REC to FRET3m when discount rate changes dominate the future growth- returns portfolio (β < 0). We also observe consistently positive coefficients relating CHG_REC to FRET3m when cash flows dominate the future growth- returns portfolio (β > 0). 26 At the firm level, LaPorta (1996) and Dechow and Sloan (1997) show that analysts forecasts of earnings growth are overly optimistic. Though our prior evidence shows an overall positive relation between recommendations and future earnings growth at the aggregate level, we also consider whether analysts excessive optimism at the aggregate level regarding future growth rather than failure to predict discount rate changes may account for the negative recommendations-future returns relation. If so, we would expect a negative recommendationsfuture growth relation for these periods when the expected higher growth fails to materialize. Within the portfolio of periods with β < 0 (i.e., those with a negative future growth-returns relation), we re-estimate the recommendations-future growth relation. Results (untabulated) confirm our earlier evidence. The coefficient linking CHG_REC to FEG at both market- and 25 We emphasize that partitioning the sample by the sign of the future growth-returns relation and regressing future excess returns on current recommendation changes within each partition does not truncate the sample by the sign of the dependent variable, future returns. Trivial correlations of and between the future growth-returns coefficients and future excess returns for market-level and industry-level aggregation, respectively, confirm this assertion. 26 In untabulated results, for market-level aggregation we also observe coefficient signs for CHG_REC consistent with those reported in Panel B for industry-level aggregation, however, results are generally insignificant. Low power from our small sample size (N = 33 for β < 0 and 47 for β > 0) is a substantial contributor. 21

23 industry-level aggregation is positive (0.129 and 0.085, respectively) and significant (p-value < 0.01 for both). Further, results presented in Tables 5 and 6 indicate that the negative recommendations-future returns relation is robust to including either current aggregate earnings growth (CEG) or quarter-ahead aggregate earnings growth (FEG). 27 These results further support our conclusions. To summarize our portfolio results, while recommendation changes predict future aggregate cash flows, they inversely predict future discount rate changes. We conclude that analysts correctly anticipate the positive relations between aggregate earnings shocks (i.e., cash flows) and future cash flows, while investors respond to this information slowly. However, aggregate recommendation changes fail to predict that a greater aggregate earnings shock leads to a greater probability of a future discount rate change. When investors subsequently adjust stock prices for the discount rate change, a negative relation between the correlated error in current recommendations and subsequent stock return adjustment is revealed. 28 A likely explanation is that analysts incentives to produce accurate earnings forecasts cause them to focus on the implications of current cash flow news for future cash flows to the exclusion of their effects on future discount rates. 29 Evidence in Hugon et al. (2016) suggests analysts lack of macroeconomic expertise may also contribute to this failure. Our results suggest the lack of 27 The incremental predictive value of aggregate recommendation changes for future excess returns after controlling for either current or future earnings growth is not surprising. Analysts may consider a variety of macroeconomic indicators (i.e., GDP growth, inflation, employment) that are relevant in predicting future industry- and market-level excess returns. We do not expect aggregate earnings growth to capture all of them. 28 Bernanke and Kuttner (2005) document that an unanticipated cut in the discount rate is associated with a stock price increase. 29 Incentives to focus on earnings forecast accuracy include the ex post verifiability of the forecast. Analysts who are identified as All-Americans by an annual poll conducted by the magazine Institutional Investor are highly rewarded. One of the most important criteria for high ranking in the poll is the analyst s perceived expertise in forecasting earnings. Further, Hong and Kubik (2003) report that more accurate forecasters are more (less) likely to move up (down) the hierarchy of brokerage houses. 22