What Drives Target Price Forecast Revisions and Their Investment Value?

What Drives Target Price Forecast Revisions and Their Investment Value? Zhi Da Department of Finance Mendoza College of Business University of Notre Dame zda@nd.edu (574) 631-0354 Keejae Hong Department of Accounting Belk College of Business University of North Carolina at Charlotte khong5@uncc.edu (704) 687-5394 Sangwoo Lee slee21fin@gmail.com (217) 722-5994 We thank Bob Chirinko, Peter Easton, David Harris, Kent Womack, and seminar participants at the University of Illinois at Chicago, Syracuse University, KAIST, University of Arkansas, and University of Montana for helpful comments and discussions. Corresponding Author

What Drives Target Price Forecast Revisions and Their Investment Value? Abstract Using a variance decomposition approach, we document that earnings forecasts and the discount rate implied in P/E ratio forecasts are both important in driving target price revisions during the sample period (1999 through 2011). Earnings forecasts are more important for companies with smaller market capitalization, higher book-to-market, slower sales growth, and lower past returns. However, a target price-based trading strategy produces a significant risk-adjusted profit when target price revisions are driven not only by revisions in earnings forecasts, but also by revisions in discount rates. This evidence suggests that equity analysts provide informative forecasts about discount rates implied in P/E ratio in addition to earnings. JEL Classification: G10, G14, G24 Keywords Target Price; Earnings Forecasts; Variance Decomposition 1

1. Introduction In addition to providing many different types of quantitative and qualitative information, sell-side equity analysts typically report three types of summary information: earnings forecasts, stock recommendations, and target price forecasts. 1 While the role of earnings forecasts and stock recommendations in price formation has been intensively studied, the literature has had less to say about the value of equity analysts target price forecasts. 2 Extant research on the investment value of target price forecasts shows that investors can earn excess returns by following analysts target price forecast signals, even after controlling for other information such as earnings revisions and stock recommendations (Brav and Lehavy, 2003; Asquith et al., 2005; Gleason et al., 2013). However, little is known about the source of target price forecasts through which analysts convey valuable information to investors. In this paper we examine the main inputs to target price forecasts in an attempt to identify where the investment value of target price forecasts comes from. While it is hard to identify which model each analyst is using in firm valuation, street wisdom holds that sell-side analysts formulate a target price by multiplying their earnings projections by a P/E ratio that s appropriate for the industry, or reasonable by the company s historical standards (Wang, 2003). Such street wisdom is consistent with evidence reported by Asquith et al. (2005) that 99% of Institutional Investor s All- 1 A target price forecast (or price target) is a firm s expected market price in the next 12 to 18 months, issued by sellside analysts. While almost all analyst reports include earnings forecasts and stock recommendations, not every analyst reports target price forecasts. According to Asquith et al. (2005), about 73% of research reports by All-America analyst team members include target price forecasts, while over 99% of them report earnings forecasts and stock recommendations in their research reports. 2 Bradshaw (2002), Brown and Bradshaw (2012), Brav and Lehavy (2003), and Gleason et al. (2013) focus on target price forecasts. 2

America analysts cite earnings multiples as a basis for price targets, whereas only 13% mention use of the discounted cash flow model or its many variations. 3 Motivated by the above street wisdom, we adopt a parsimonious valuation model of analysts target price forecasts. In particular, we take the target price (TP) to be the product of two terms: a forecast of future earnings (EF) and a forecast of the price-toearnings ratio (PE), which implies discount rate forecasts and earnings growth forecasts. 4 The decomposition of target prices into different valuation components allows us to examine several interesting empirical questions about the target price formation process. First, what determines target price forecasts -- do revisions in earnings forecasts or revisions in P/E ratio forecasts drives target price revisions? By addressing this question we shed light on whether analysts actually use both of these components in target price formation as they claim (Wang, 2003; Brav, Lehavy and Michaely, 2005). Second, how does the relative importance of earnings forecasts in target price formation depend on characteristics of the underlying stock? This question is motivated by earlier research that shows that analysts favor stocks with particular characteristics (Jegadeesh et al., 2004). Third, is the investment value of target prices driven mainly by analysts ability to forecast earnings or by their ability to forecast future P/E as well? We know that analysts add value by providing accurate near-term earnings forecasts (Chan et al., 1996; Gleason and Lee, 2003; Stickel, 1991). On the 3 Gleason et al. (2013) also document that not all target price forecasts are based on the same valuation model. For example, some analysts rely on simple heuristics (e.g., a valuation model based on the PEG ratio) while others use a more rigorous valuation model such as a residual-income-based model. 4 For a firm that pays out earnings as dividends, Gordon s (1962) constant-growth model shows that P E = 1/(r g), where r denotes a discount rate and g earnings growth rate. In a more rigorous decomposition, we also consider a present-value model where the forecast of P/E is a function of the cash flow growth rate forecast and the discount rate. We discuss the findings from the alternative model in section 6.3. 3

other hand, if they are not particularly good at forecasting P/E, we would expect analysts target prices to be more informative when the target prices are mostly driven by the earnings forecasts. Using a large sample of target price data from I/B/E/S for the 1999 to 2011 period, we document that both earnings forecast revisions and P/E ratio forecast revisions drive revisions in target prices. 5,6 For example, at a three-month revision horizon, about 39% of the variation in target price revisions is driven by revisions in earnings forecasts and 61% by revisions in P/E ratio forecasts. The relative importance of earnings forecast revisions increases with the revision horizon, consistent with earlier studies that show that in the long run, firm valuation ultimately depends on earnings (Easton et al., 1992; Vuolteenaho, 2002). At annual horizon, more than 60% of the target price variations come from revisions in earnings forecasts. In our additional analysis, we document that revisions in P/E ratio forecasts are mainly driven by revisions in discount rate, not by the revisions in growth rate. We also find that, in the cross-section, the relative importance of earnings revisions and P/E ratio revisions is related to characteristics of the underlying stocks. For example, earnings forecast revisions are more important for stocks with smaller market capitalization, higher book-to-market, higher capital expenditures, slower sales growth, and lower past returns. 5 To determine the driving force behind analysts target prices, we focus on revisions in - rather than levels of - target prices. When we instead use the level of target prices and their components, the results not only continue to hold, but are even stronger. 6 In an earlier version of the paper, we use target price forecasts collected by First Call. However, First Call stopped collecting and publishing target price forecasts in March 2005. We obtain very similar results and our inference does not change in any meaningful way when we use data from I/B/E/S, whose coverage extends to more recent years. 4

Our empirical tests above focus on how analysts form their target price forecasts. In our second set of analyses, we examine the profitability of investment strategies based on target prices. More specifically, we test the source of target price forecasts investment value. Consistent with prior studies (Brav and Lehavy, 2003; Da and Schaumburg, 2011; Gleason et al., 2013), we find that a long-short trading strategy based on the expected returns implied by analysts target prices (TPER) is highly profitable: a TPER strategy produces a substantial four-factor alpha of 0.87% per month. If the investment value of target prices comes primarily from analysts superior ability to forecast earnings, we would expect this ability to be relevant and lead to a more profitable TPER strategy for stocks whose target price revisions are driven mainly by earnings forecast revisions. Indeed, we find the TPER strategy to generate riskadjusted profit of more than 10% per year, significantly better than its performance among the remaining stocks whose target price revisions are driven mainly by P/E ratio forecast revisions. We confirm this result in cross-sectional regressions where we control for other firm characteristics in addition to stock recommendation revisions, suggesting that our results are not driven by other information provided by analysts. This paper contributes to the literature in several ways. First, the results provide insights into how analysts generate target price forecasts. While prior research explores the association between earnings forecasts and stock recommendations (Francis and Soffer, 1997; Bradshaw, 2004), the literature has given little attention to the relation between earnings forecasts and target prices even though target prices seem to be a finer summary measure about firms future prospects than stock recommendations. A few prior studies suggest that analysts earnings forecasts are important in the target 5

price formation process (Bandyopadhaya et al., 1995; Bradshaw, 2002; Gleason et al., 2013). Our paper confirms these findings by quantifying the important role of earnings forecasts. Second, our paper extends prior studies by showing that investment value in target price forecasts primarily comes from the information embedded in the earnings forecasts. Firms whose target price revisions are driven mainly by earnings forecast revisions are those where tangible assets account for a large fraction of their valuations (stocks with smaller market capitalization, higher book-to-market, higher capital expenditures, slower sales growth, and lower past returns.) This finding explains why a trading strategy based on target price is more profitable for small value stocks as documented in Da and Schaumburg (2011). Ultimately, our findings shed light on the information production process by equity analysts by simultaneously examining multiple forecasts that they issued. Loh and Mian (2006) show that accurate earnings forecasts facilitate superior stock recommendations. Kecskes et al. (2013) examine whether earnings forecasts or discount rate forecasts drive the value of analysts stock recommendations. The examination of target prices that are continuous measures could be more informative than looking at stock recommendations that take on 5 discrete values. The remainder of the paper is structured as follows. In Section 2 we discuss related literature and develop our research questions. We discuss the variance decomposition methodology used in this study in Section 3. In Section 4 we discuss our data sources and the key variables used in this study. In Section 5 we present empirical results on the drivers of target prices, and in Section 6 we report robustness tests. We analyze the 6

investment value of target prices in Section 7. We provide concluding remarks in Section 8. 2. Related Literature and Research Questions Prior literature on the investment value of equity analyst research generally agrees that analyst research provides value-relevant information to the capital market. 7 In particular, ample evidence shows that an investment strategy following earnings forecasts and stock recommendations can be profitable. 8 However, only since large target price data became available in machine readable format through First Call in the mid-1990s have researchers started paying more attention to analysts reporting of target prices and their investment value. In this section, we review related studies, and develop research questions related to target price formation, the role of firm characteristics and earnings forecasts in target price formation, and the source of the investment value of target prices. 2.1 Target Price Formation 2.1.1 Target Price Valuation Model Examining how analysts form target price expectations requires understanding the model that analysts use in forecasting a firm s future value or target price. However, no single model is universally adopted by all analysts, and it is hard to identify the model that each individual analyst actually uses in firm valuation. Notwithstanding, prior studies based on analysts self-reporting of target price forecasts provide some guidance on the 7 See Ramnath et al. (2008) for a recent review on the role and the value, of equity analysts in capital markets. 8 Earnings forecast studies include Givoly and Lakonishok (1979), Imhoff and Lobo (1984), Lys and Sohn (1990), Stickel (1991), Gleason and Lee (2003), and Womack (1996), and stock recommendation studies include Barber et al. (2001) and Jegadeesh et al. (2004). 7

valuation models used by analysts. These models can be classified into two groups: 1) models based on multiples and 2) the dividend discount model or its variations. Analysts that rely on a multiples approach first make an earnings forecast and then use an appropriate multiple (e.g., P/E) to forecast a firm s target price. Evidence based on analysts self-reporting suggests that this simple multiples-based valuation heuristic is more widely adopted by analysts than other models. For example, Asquith et al. (2005) document that 99% of Institutional Investor s All-America analysts cite earnings multiples as a basis for generating price targets, and Wang (2003) reports that to arrive at a target price for the future, sell-side analysts often take their earnings projections and multiply them by a P/E ratio that s appropriate for the industry, or reasonable by the company s historical standards. Bradshaw (2004) shows that simple heuristics such as the PEG ratio and long-term earnings growth better explain analysts valuations (i.e., stock recommendations) than a more rigorous valuation model such as a residual income model, a variation to the dividend discount model. 9 Based on the discussion above, we examine target price formation using a simple valuation model. 10 2.1.2 Target Price Formation Process Earnings forecasts are an important component of models used to forecast target prices. 11 For example, using a sample of 124 firms from 1983 to 1988, Bandyopadhyay et al.(1995) find evidence consistent with earnings forecasts being an important input in 9 Bradshaw (2004) does not use the term target price per se; however, he uses various types of valuation models to measure analyst expectations of firm value (=target price). 10 However, even if a simplistic valuation model based on multiples is popular among analysts, some analysts still use more rigorous valuation models. We also consider a present-value model where the target price is a function of earnings forecast, the cash flow growth rate and the discount rate. The use of different valuation models to leads to similar conclusions and results are available upon request. 11 A stylized example of a multiples-based target price (TP) valuation model is TP = earnings forecasts price-toearnings ratio. A more rigorous valuation model for target price estimates target price as the present value of future cash flows, where future cash flows are based on earnings forecasts. More formal models used in this study are discussed below. 8

forecasting stock prices by showing that variation in short-term (long-term) earnings forecasts explains about 30% (60%) of target price variation. Asquith et al. (2005) and Bradshaw (2002) also find evidence suggesting that earnings forecasts are a key input in the formation of analysts target prices. However, earnings forecasts are not the only input used in a valuation model. In a simple multiples-based target price model, target prices can be decomposed into the product of earnings forecasts and a P/E multiple. Thus, target price forecasts are affected not only by earnings forecasts, but also by other factors such as a P/E multiple. How this other factor (P/E multiple) affects target price forecasts, however, is an open question. Our first research question therefore investigates whether target price forecasts are based primarily on earnings forecasts and/or P/E multiples. 2.2 Firm Characteristics and the Relative Importance of Earnings Forecasts in Target Price Formation Prior research finds that sell-side equity analysts favor firms with certain characteristics. For example, using recommendations and recommendation changes, Jegadeesh et al. (2004) show that analysts tilt their opinions toward firms with characteristics that predict future returns (e.g., total accruals or book-to-market). Motivated by this evidence, our second research question investigates how firm characteristics affect the relative importance of target price components (i.e., revisions in earnings forecasts or P/E) in explaining target price revisions. In this study, we focus on several firm characteristics known to affect analysts stock valuation: accruals, book-to-market, capital expenditures, sales growth, size, and momentum. If firm characteristics are related to earnings quality or the level of future 9

earnings, revisions in earnings forecasts should be a main driver of revisions in target prices. In contrast, if firm characteristics are related to (or proxy for) the discount rate or growth opportunities, revisions in earnings forecasts should have less impact on revisions in target prices, and revisions in the P/E component of target prices should be a main driver of target price revisions. More specifically, we expect that for firms with higher sales growth, greater momentum, and lower book-to-market, the P/E component of the target price model better explains target price forecasts than the earnings forecast component because such firms are characterized by high growth potential, which is typically indicated in the P/E component. However, the role of other firm characteristics in explaining the relative importance of P/E multiples and earnings forecasts in target price valuation is not always clear. For example, if analysts correctly (do not correctly) interpret the future earnings persistence implication of the accruals component of earnings, the target price of firms with high accruals will be less (more) affected by changes in earnings forecasts. Likewise, capital expenditures can be seen as a signal of growth potential (i.e., the P/E component will be more important in explaining future firm value) or as a signal of value driven mainly by investment in assets (i.e., the earnings component will be more important in explaining future firm value). Finally, if size is viewed as a proxy for risk or growth potential, the P/E component will be more important in explaining target price revisions, while if one believes that earnings news is more difficult for smaller firms to diversify, then the earnings component of target prices will be more important. 2.3 Source of the Investment Value of Target Prices 10

Recent studies document the informativeness of analyst target prices. Using analyst target prices issued for more than 6,500 firms over the 1997 to 1999 period, Brav and Lehavy (2003) find incremental abnormal returns around target price revisions, even after controlling for stock recommendations and earnings forecast revisions. Asquith et al. (2005) confirm Brav and Lehavy (2003) by showing that target price revisions have greater impact on stock returns than earnings forecast revisions. Da and Schaumburg (2011) analyze the short-term performance of a long-short trading strategy based on TPER over the 1999 to 2004 period and find that a sector-neutral strategy of buying the highest TPER stocks in the S&P 500 and shorting the lowest stocks earns significant abnormal returns, suggesting the short-term informativeness of analyst target prices. Two recent studies also examine factors that affect the profitability of the information provided by analysts. First, Gleason et al. (2013) provide evidence that a target pricebased strategy is more profitable when the residual income model is used, as opposed to a model based on price multiples. Second, in a study on the source of the investment value of stock recommendations, Kecskes et al. (2013) split analysts stock recommendation changes into earnings-based and discount rate-based changes, and document that earnings-based recommendation changes are more informative than discount rate-based recommendation changes. The findings of Kecskes et al. (2013) imply that analysts changes in stock recommendations are more informative when such valuation changes are initiated by changes in earnings forecasts. Despite growing evidence on the investment value of analysts target prices, evidence on the source of this value is limited. Our study complements Gleason et al. (2013) and Kecskes (2013) by investigating whether the investment value of target prices is driven 11

mainly by analysts ability to forecast earnings or by their ability to also forecast future P/E ratios implied by target price forecasts. 3. Target Price Decomposition According to Asquith et al. (2005), a firm s target price (TP) is often derived as the product of two terms: a forecast of future earnings (EF) and a forecast of the P/E ratio (PE): TP t = EF t PE t. Analysts target prices (TP t ) and earnings forecasts (EF t ) are directly observable. We back out the implied forecasts of the P/E ratio using PE t = TP t / EF t. Taking the logarithm, we get: tp t = ef t + pe t. It is well known that the level of earnings forecasts can be contaminated by analyst biases. As biases are more likely to persist over short horizons, revisions in analysts forecasts are less affected by biases and hence usually more informative about changes in firms fundamentals. Revisions in (log) target prices can be decomposed into revisions in (log) earnings forecasts and revisions in the implied (log) P/E ratio forecasts: tp t = ef t + pe t. (1) The earnings forecast revisions reflect earnings news, while the P/E ratio forecast revisions reflect earnings growth rate news and discount rate news. To measure their relative importance in driving target price revisions, we use a variance decomposition approach. Equation (1) implies: 12

Var( tp t ) = Cov( tp t, ef t ) + Cov( tp t, pe t ). (2) Dividing both sides of equation (2) by Var( tp t ), we obtain: 1 = Cov( tp t, ef t ) Var( tp t ) + Cov( tp t, pe t ). (3) Var( tp t ) Each term on the right-hand side of equation (3) can be estimated by regressing ef t and pe t, respectively, on tp t. The slope coefficient of the first regression, βef, thus measures the percentage of the total variation in target price revisions that is driven by earnings forecast revisions. Likewise, the slope coefficient of the second regression, βpe, measures the relative importance of revisions in P/E ratio forecasts in driving target price revisions. By construction, βef and βpe sum to one. Empirically, βef serves as a lower bound on the relative importance of earnings news for two reasons. First, information about long-term earnings growth rates is incorporated in P/E ratios and will show up in βpe, as discussed further below. Second, since we define the difference between the target price revision and the earnings forecast revision as the revision in P/E ratio forecasts, noise associated with target price revisions (e.g., measurement error, analyst bias) will always be classified as revisions in P/E ratio forecasts. Thus, βpe overestimates the importance of P/E ratio forecasts, while βef underestimates the importance of earnings news. 4. Data 13

The target price data for this study come from I/B/E/S. 12 For each month from 1999 through 2011, we include stocks for which there is at least one (12-month-ahead) target price announcement during the month. Table 1 presents a summary of the sample. For each stock, there are on average 2.93 target prices and 5.79 earnings forecasts per month. The sample on average covers more than 89% of the CRSP stock universe in terms of market capitalization. The median market capitalization of stocks in our sample, averaged over the sampling period, is $1.35 billion much larger than that of all Nasdaq stocks ($85 million), and even slightly larger than that of all NYSE stocks ($963 million). A key variable of interest is the target-price-implied one-year-ahead expected return, TPER. TPER is defined as the consensus (split-adjusted) target price divided by the end-of-month stock price minus one (i.e., TPER t = TP t /P t 1), where the consensus target price TP t is the simple average of all target prices received during the month. We do not use analyst identities in constructing the consensus forecast, as several studies, including Bradshaw et al. (2013) and Bonini et al. (2010), find limited evidence of systematic differences in analysts target price forecasting abilities. The mean (median) TPER during our sample period is 26.0% (18.3%), substantially higher than one would expect for the market as a whole. One reason for this result is that analysts are far more likely to issue target prices when they favor a stock they cover. The mean (median) TPER is as high as 53.1% (34.0%) in 2000 during the final stages of the NASDAQ bubble. 12 In an earlier version of this paper, we used target price data from First Call, which only covers 1994 through 2004. We switched to the I/B/E/S database mainly for two reasons. First, First Call stopped collecting target price forecasts in April 2004 while the I/B/E/S database continues coverage through 2011. Second, more firms are covered by I/B/E/S than First Call (I/B/E/S s coverage is about 40% greater than First Call s coverage). However, none of our results change due to the use of a different database. 14

Following Da and Schaumburg (2011), we separate the sample into sectors according to the first two digits of Standard & Poor s GICS (Global Industry Classification Standard) codes. Using I/B/E/S data, Boni and Womack (2006) show that GICS sector and industry definitions are in accordance with the areas of expertise of most analysts as defined by the set of stocks that analysts cover. GICS codes are therefore a natural basis for sector definitions. We obtain target price, earnings forecast, and earnings growth rate data from I/B/E/S, and stock prices and returns from CRSP. We employ data from Compustat to compute various portfolio characteristics. 5. Empirical Findings 5.1 Variance Decomposition To examine how target prices are formed, we first conduct an annual variance decomposition at the firm level. The results are reported in Table 2. Several interesting findings are worth noting. First, the slope coefficient on earnings forecast revisions (βef) is significantly greater than zero for all three revision horizons. Given that βef serves as a lower bound on the relative importance of earnings news, a positive βef suggests that when we measure expectations from analysts perspective, earnings forecast revisions are important in determining expected changes in stock prices. Second, the relative importance of earnings forecast revisions in target price formation increases with the revision horizon. At the three-month revision horizon, on average about 39% of the variation in target price revisions is driven by revisions in earnings forecasts. At the semiannual (annual) horizon revisions in earnings forecasts explain about 52% (63%) of the total variation in target price revisions. This finding is consistent with the notion that although price variation in the short term can be driven 15

by sentiment or other factors unrelated to firm fundamentals, over longer horizons it is still tied to the expected change in future earnings. 13 This finding may also be due in part to non-synchronous consensus earnings forecasts and consensus target prices; that is, to earnings forecasts and target price forecasts not being issued simultaneously, which can lead to noise that biases the estimate of βef toward zero. However, this effect should weaken as the revision horizon increases. 14 Third, the slope coefficient on P/E ratio forecast revisions (βpe) is significantly positive. If analysts are simply calculating target prices using their earnings forecasts multiplied by a constant P/E ratio, then the revision in target price forecasts should be driven almost entirely by earnings forecast revisions, resulting in a close-to-zero βpe. This is not what we find. The values of βpe range from 0.61 at the three-month horizon to about 0.37 at the annual horizon, implying that analysts update their P/E ratio forecasts as much as their earnings forecasts, if not more. Later we explore whether such implied P/E ratio revisions have any investment value. Finally, the relative importance of earnings forecasts and implied P/E ratio forecasts for target prices changes over time. For instance, earnings forecasts are relatively less important in determining target prices during the 1999 to 2000 period surrounding the peak of the tech bubble. 5.2 Variance Decomposition and Stock Characteristics We next relate the relative importance of earnings forecast revisions (βef) to various characteristics of the underlying stock. We do not report the results associated with P/E 13 This result is consistent with evidence in Easton et al. (1992) that the explanatory power of earnings for returns increases monotonically from 4% to 60% as the return interval increases from one to ten years. 14 We discuss the non-synchronicity caused by stale forecasts in a robustness check in Section 6. 16

ratio revisions (βpe) because βpe = 1 βef. In addition, since all estimates are highly significant, with associated t-statistics all higher than 10 (in absolute terms), we report only point estimates from the variance decomposition. The high levels of significance are expected because the underlying structure is a mathematical identity. More specifically, we examine seven stock characteristics studied in Jegadeesh et al. (2004). Total accruals (TAC) are computed as earnings before extraordinary items minus cash flows from operating income at each quarter-end, scaled by the average total assets between t 1 and t. Sales growth (SG) is computed as the percentage change in sales from year t 1 to year t on a quarterly rolling basis. Annual total capital expenditures (CapExp) are calculated on a quarterly rolling basis scaled by the average total assets between t 1 and t. Book-to-market (BP) is the ratio of the book value of equity to market capitalization at each quarter-end. Market capitalization (MktCap) is the logarithm of market capitalization at quarter-end. Return 1 (Ret1) is the six-month return from month t 6 to month t 1. Return 2 (Ret2) is the six-month return from month t 12 to month t 7. Each month, we sort sample observations into five groups according to each characteristic, so overall we construct 35 stock portfolios each month. We then repeat the variance decomposition exercise for each portfolio. The estimates of βef and their difference between extreme portfolios are reported in Table 3 for three revision horizons. Across all three horizons, we document that the relative importance of earnings forecasts for target prices (βef) is significantly related to total accruals, sales growth, book-to-market, market capitalization, and past returns. For example, βef is higher for stocks with low total accruals, which is consistent with the intuition that 17

earnings news is more important for firms reporting higher quality earnings. The average βef is at least 0.15 higher for small stocks than large stocks. This finding is consistent with the argument that discount rate news is more difficult to diversify across projects than earnings news (Vuolteenaho, 2002), and suggests that for large firms with many projects, earnings news is diversified and accounts for a lower proportion of the total variation in stock prices. The average βef for value stocks with high book-to-market ratios is more than 0.15 higher than that for growth stocks with low book-to-market ratios. For a value stock, assets-in-place account for a large portion of the stock s value, so it is not surprising that earnings forecast revisions are more important for target price revisions. βef is lower for stocks with high sales growth because growth rate information is incorporated in implied P/E ratio forecasts. Johnson (2002) argues that higher past returns are an indication of higher future growth. Consistent with that argument, we find that βef is lower for stocks with higher past returns. 6. Robustness Tests 6.1 Negative Earnings Forecasts Thus far we do not include in our analysis firms with negative earnings forecasts (a little over 9% of stock-month observations) because earnings forecasts are logtransformed. In this section we conduct two additional tests to include firms with negative earnings forecasts in the analysis. First, we aggregate earnings forecasts at the market level, because earnings forecasts are almost always positive once they are aggregated at the market level. 18

Each month, we aggregate both target price forecasts and earnings forecasts at the market level, and then decompose revisions in aggregate market target prices into revisions in aggregate earnings forecasts and revisions in aggregate P/E ratio forecasts according to equation (1). We report these results at the aggregate market level in Panel A of Table 4. The results are similar to those of the firm-level variance decomposition reported in Table 2. Both earnings forecast revisions and P/E ratio revisions are important at all revision horizons, and earnings forecast revisions become more important while P/E ratio forecasts become less important as the revision horizon increases. Second, since (EF t EF t 1 ) EF t 1 approximates the ef t measure used earlier, we redefine ef t using the approximate expression. We report the results of the variance decomposition using this new definition of earnings forecast revisions in Panel B of Table 4. Once again the results are similar to those reported in Table 2. Our results are thus robust to the inclusion of firms with negative earnings forecasts. 6.2 Stale Forecasts When analysts issue earnings forecasts and target price forecasts at different times (e.g., delayed reporting of target price forecasts), variation in the target price may be attributed to P/E forecasts even though the changes in target price may be driven by revisions in earnings forecasts. In our sample, for the three-month revision horizon, approximately 2.2% of the observations have target price revisions that are not accompanied by earnings forecast revisions. As we increase the forecast horizon, stale forecasts seems to be less of an issue (i.e., 1.76% of observations for the six-month horizon and 1.31% for the 12-month horizon). The relative importance of earnings 19

forecast revisions in explaining target price revisions is very similar to that reported in Table 4 (full sample), suggesting that the inclusion of stable forecasts does not change our inferences in any meaningful way (results not tabulated). 6.3 Alternative Model Specifications P/E further decomposed into discount rate and growth rate According to the Gordon (1962) dividend discount model, P/E ratios incorporate information about growth rates and discount rates. In this section, we examine the relative importance of growth rates and discount rates implied in the P/E ratio in explaining target price revisions along with earnings forecasts. As our first model in equation (1) does not provide a functional specification of P/E ratios in terms of growth rates and discount rates, we develop a more general model of target price formation. This allows us to separate the three components of target prices (earnings forecasts, growth rate forecasts, and discount rate forecasts). To this end, we employ a residual income model (RIM), equity analysts other major valuation methodology for target price formation (Asquith et al., 2005). 15 The tractable finite T-horizon RIM specification is: 5 TP t = BVPS t + E t(ri t+i ) i=1 + E t(tv t+5 ) (1+R) i (1+R) 5, (4) where BVPS is equity book value, RI is residual income computed as EPS t+i r BVPS t+i 1, which denotes the expected future earnings for period t + i conditional on the information available at time t, EPS is earnings per share, and R is a discount rate. Our terminal value (TV) is computed following Clause and Thomas (2001), where RI is 15 A RIM model is also used in Gleason et al. (2013). As robustness tests, we also use the abnormal growth in earnings model (AGM) developed by Ohlson and Juettner-Nauroth (2005), which links the value of a firm to its earnings, abnormal earnings, growth in abnormal earnings, and discount rate, as well as the dividend discount model (DDM). Any inferences we draw from this study do not change if we use the AGM or DDM as the valuation model. 20

assumed to grow at the rate of a ten-year government bond less 3% after five years. 16 We compute the implied discount rate (R) for each firm following the approach adopted in earlier studies as described in detail below (e.g., Gebhardt et al., 2001; Pastor et al., 2008; Chen and Zhao, 2013). We use analysts earnings forecasts for years t + 1 and t + 2, and use long-term growth forecasts (G) to compute earnings forecasts for years t + 3 to t + 5 as follows: FE t+i = FE t+i 1 (1 + G t ) i 1, (5) where FE t+i is the earnings forecast for year t + i. 17 Using the approach described above, we can express a target price forecast as a function of the one-year-ahead earnings forecast (EF), the long-term earnings growth rate (G), and the implied discount rate (R). Since target price forecasts, one-yearearnings forecasts, and long-term growth forecasts are all available from I/B/E/S, we are able to estimate the implied discount rate, R. 18 Following Chen and Zhao (2013), who decompose changes in stock prices into changes due to cash flow news and changes due to discount rate news, we decompose revisions in target prices into revisions due to three components: earnings revisions, growth rate revisions, and discount rate revisions: tp t,t+j = ln ( TP t+j TP t ) = Δ(ef) t,t+j + Δ(g) t,t+j + (r) t,t+j, (6) 16 During our sample period, the mean (median) terminal RI growth rate is 1.18% (1.19%). 17 To denote i-year-ahead earnings forecasts, we use FE t+i instead of EF t+i to avoid confusion. Note that in our parsimonious model, we use EF t to denote analysts one-year-ahead earnings forecasts in year t (i.e., EF t FE t+1 ). 18 Results of several sensitivity tests using 1) a mean reverting to GDP growth rate assumption as in Pastor (2008) and 2) a ten-year terminal value horizon indicate that our results are robust to alternative model assumptions. 21

where: Δ(ef) t,t+j = ln[f(ef t+j, G t, R t )] ln[f(ef t, G t, R t )] Δ(g) t,t+j = ln[f(ef t+j, G t+j, R t )] ln[f(ef t+j, G t, R t )] Δ(r) t,t+j = ln[f(ef t+j, G t+j, R t+j )] ln[f(ef t+j, G t+j, R t )], f(. ) denotes the function of parameters, and ln denotes the natural logarithm. This approach enables us to estimate the revision in target price due to the revision in each component by allowing the component to vary over time while holding the other components fixed. Note that (x) t,t+j does not denote the change in variable x from t to t + j. Rather, it denotes the revision in log target prices from t to t + j that is attributed to the revision of component x over the same time horizon. The decomposition in equation (6) provides a convenient way to express the variance in target price revisions as the sum of three covariances: 19 Var( tp) = Cov[ p, (ef)] + Cov[ tp, (g)] + Cov[ tp, (r)]. (7) Dividing both sides of equation (7) by Var( tp), we obtain: 1 = Cov[ p, (ef)] Var( tp) + Cov[ tp, (g)] Var( tp) + Cov[ tp, (r)]. (8) Var( tp) Each term on the right-hand side of equation (8) can be estimated by regressing (ef), (g), and (r), respectively, on tp. The slope coefficient of each regression is labeled βef, βg, and βr, respectively. By construction, βef, βg, and βr sum to one, and each slope coefficient is interpreted as the percentage contribution of each component to the total variation in target price revisions. 19 For simplicity, we omit the time subscript t. 22

Panel C of Table 4 shows that the implied P/E ratio s explanatory power in accounting for the variation in target price revisions is driven mainly by variation in discount rates rather than variation in earnings growth rates. Across all revision horizons, growth rate revisions explain only 5% to 8% of the variation in target price revisions (βg). 20 Strikingly, revisions in discount rates explain almost half (30% to 51%) of the variation in target price revisions (βr), suggesting that analysts make discount rate forecasts as well as earnings forecasts in generating target prices. These findings are consistent with results in Kecskes et al. (2013), who document that analyst recommendation changes based on discount rate changes are informative, while recommendation changes based on growth rate changes do not have a significant price impact. 6.4 Variance Decomposition Conditional on Earnings Growth Rates According to the residual income model for target price formation, the results in Table 4, Panel C suggest that revisions in the implied P/E ratios that account for variation in target price revisions are due primarily to revisions in discount rates rather than to revisions in earnings growth rates. To confirm these findings, we limit attention to the subsample firms with zero expected earnings growth, in which case any changes expected in P/E ratios should come from changes in the discount rate rather than the growth rate. We report results of the variance decomposition of target price revisions in Panel D of Table 4 after restricting the sample to firms with no revision in growth rate 20 A low contribution of growth rate changes in explaining target price revisions is mainly due to the magnitude of growth rate changes being relatively small. For example, the median change in growth rate (in absolute terms) from year 3 to 5 is 1% for the six-month horizon and 1.5% for the 12-month horizon, and the median change in growth rate used in the terminal value calculation is only 0.4% for both the six-month and the 12-month horizons. Therefore, most variation in target price revisions seems to be explained by revisions in either earnings forecasts or the discount rate. 23

(i.e., g = 0). The βpe for this subsample is not much different from the βpe for the full sample. For example, at the six-month revision horizon, the βpe is 0.52 for the sample of firms with zero g while the average βpe for the full sample is 0.46 (= 1-0.54 from Table 4 Panel C). We observe a similar pattern for other revision horizons. These findings suggest that the explanatory power of P/E ratio forecasts in driving target price revisions is driven very little by growth rate revisions. This result is consistent with the results reported in Panel C of Table 4. Overall, the results reported in Table 4 provide evidence that the variation in target price revisions attributable to the variation in the implied P/E ratio revisions is largely due to revisions in discount rates and not revisions in earnings growth rates. 7. Investment Value of Target Price Components 7.1 Earnings Forecasts vs. Discount Rates Implied in P/E Ratio Forecasts Our analysis of variance decomposition above shows that both earnings forecasts and P/E ratio forecasts are important in determining target prices. Our results further suggest that target price revisions explained by P/E ratio revisions are driven primarily by revisions in the discount rate, with revisions in growth rates having little explanatory power. Prior research on the investment value of target prices suggests that analysts provide value-relevant information to investors through their target price forecasts (Brav and Lehavy, 2003; Da and Schaumburg, 2011). Here we examine how each component of target prices drives the investment value of target prices. Following Brav and Lehavy (2003) and Da and Schaumburg (2011), we use TPER as an investment signal, and implement a sector-neutral TPER strategy. TPER is 24

defined as the return implied by the consensus 12-month-ahead target price and the current market price (i.e., TPER = TP/P 1). For each month, we compute the TPER for each stock and sort stocks according to their TPER within each sector. We then construct an equally weighted portfolio that is long the highest TPER stocks in each sector and short the lowest TPER stocks in each sector. The portfolio is held over the next month before rebalancing. 21 To examine whether the investment value of target prices comes from analysts ability to forecast earnings or their ability to forecast discount rates, we test the crosssectional relation between the performance of the TPER strategy and the relative importance of each component of target prices. To do so, we need to construct a firmspecific measure of the relative importance of earnings forecast revisions in explaining target price revisions (βef). For every firm-month, we compute firm-specific βefs by regressing log earnings forecast revisions on log target price revisions over the previous 24 months, and we then sort stocks into two subsamples: stocks with above-median βef (high EF-sensitive sample) and stocks with below-median βef (low EF-sensitive sample). We require a minimum of eight observations for each regression. We then look at the performance of the TPER strategy within each subsample of firms. If the investment value of target prices is attributable solely to analysts ability to forecast future earnings, we would expect the TPER strategy to be profitable in the sample in which earnings forecasts are more important in explaining target prices (i.e., the subsample with above-median βef) but not in the sample in which earnings 21 As the portfolio is equally weighted, it is sector-neutral by construction, thereby isolating the relative strength of information in analysts price targets as suggested by Da and Schaumburg (2011). 25

forecasts are less important in explaining target prices (i.e., the subsample with belowmedian βef). If analysts also have superior ability to forecast P/E ratios, however, we would expect the TPER strategy to be profitable even among stocks for which the relative importance of earnings forecasts for explaining target prices is low (low EFsensitive sample) or for which the relative importance of P/E ratios is high. To account for the fact that stocks with different levels of TPER are associated with different risks, we compute risk-adjusted returns using a four-factor model that includes both the Fama and French (1993) three factors and the Carhart (1997) momentum factor. To account for the possibility that factor loadings are time-varying, we also compute characteristic-adjusted returns as in Daniel et al. (1997). Intuitively, characteristic-adjusted returns reflect the excess returns to our TPER portfolios in excess to those of a benchmark portfolio with similar characteristics in terms of size, book-to-market, and past returns. Table 5 reports the profitability of the TPER strategy for both the full sample and the subsamples by the relative importance of earnings forecasts in target price formation. Confirming the findings in Da and Schaumburg (2011), the TPER strategy is highly profitable. Panel A shows that the TPER strategy produces a substantial fourfactor alpha of 0.87% per month for the full sample. The alpha is also statistically significant with a t-value of 2.86. The characteristics-adjusted return to our TPER strategy yields a similar 0.89% per month (t-value of 3.27). At longer revision horizons (Panels B and C), the TPER strategy remains profitable, and we see similar excess returns in terms of magnitude as reported in Panel A. 26

By focusing on stocks whose target price revisions are driven mainly by earnings forecast revisions (high EF-sensitive sample), we examine whether the investment value of target price revisions comes from analysts superior ability to forecast earnings. Panel A of Table 5 reports that the TPER strategy produces a four-factor alpha of 1.03% per month in the high EF-sensitive sample (statistically significant with a t-value of 2.79). The characteristic-adjusted return is also profitable (1.06% per month with a t- value of 3.11), which is consistent with the notion that the investment value of target prices comes largely from analysts superior ability to forecast earnings. More interestingly, the performance results of the TPER strategy for the low EFsensitive sample show that equity analysts have a superior ability to forecast not only earnings but also discount rates implied in P/E ratios. Specifically, the TPER strategy produces a four-factor alpha of 0.70% per month (with a t-value of 2.39) and a characteristic-adjusted return of 0.71% per month (with a t-value of 2.75) in the low EF-sensitive sample. The findings at longer revision horizons (Panels B and C) are similar in that the TPER strategy proves to be profitable in both the low and the high EF-sensitive subsamples. When repeat the analysis using revisions in TPER rather than the level of TPER, we obtain similar results. In sum, the results in Tables 5 provide strong evidence that the investment value of target prices is driven not only by analysts ability to forecast future earnings but also by their superior ability to forecast discount rates implied in P/E ratios. 7.2 Trading Strategy after Controlling for Recommendation Revisions Analysts often issue target price forecasts and stock recommendations at the same time, so it is possible that the investment value of target prices reported in the previous 27