Aggregate Earnings Surprises and Inflation Forecasts *

Transcription

1 Aggregate Earnings Surprises and Inflation Forecasts * S.P. Kothari MIT Sloan School of Management kothari@mit.edu Lakshmanan Shivakumar London Business School lshivakumar@london.edu Oktay Urcan London Business School ourcan@london.edu April 24, 2013 Abstract We show that aggregate earnings surprises contain information about future inflation, but macroeconomic forecasters do not fully utilize this information in generating their forecasts. Earnings news, aggregated across firms releasing earnings in a three-month period, predicts forecast errors in Producer Price Index (PPI) released in the subsequent two months. These aggregate earnings surprises do not predict forecast errors for Consumer Price Index (CPI). The results are robust to alternative specifications, and they are driven by a broad cross-section of firms rather than limited to isolated industries like the financial services industry or the retail industry. With respect to the capital markets, the bond market s reaction to PPI news is predictable based on previously released aggregate earnings news. Collectively, the results indicate that neither macroeconomic forecasters nor bond market investors fully incorporate information in aggregate earnings surprises for future PPI. * We appreciate helpful comments from Bill Cready, Yaniv Konchitchki, Peter Pope and seminar participants at Harvard Business School, Cass Business School and the 11 th London Business School Accounting Symposium. Oktay Urcan acknowledges the financial support from London Business School Research and Materials Development (RAMD) Fund.

2 1. Introduction Accurate forecasts of future inflation are crucial for almost all economic agents. Inflation expectations affect monetary and fiscal policies of central banks and governments, which, in turn, affect business decisions. Inflation expectations also directly influence a variety of business decisions, including the pricing of services, wages in labor contracts, corporate investment and financing decisions (as inflation expectations directly affect the cost of capital) and firms hedging decisions. Inflation expectations also affect banks lending decisions through their effect on interest rates. Not surprisingly, a vast academic literature focuses on the forecasting of inflation and on the evaluation of inflation forecasts. Two important unanswered questions in this literature are whether (i) corporate earnings are useful in predicting inflation; and (ii) macroeconomic forecasters and capital market participants efficiently incorporate inflation information in earnings into their forecasts and thus in setting security prices. We address these two questions in this study. Corporations constitute a large segment of the macroeconomy. Public companies collectively represent a substantial fraction of the macroeconomy as employers and as producers of goods and providers of services in the economy. Corporations also make up a large fraction of the private sector investment in an economy. We therefore expect corporations activities, in aggregate, to significantly affect the macroeconomy. Consistent with this expectation, macroeconomic variables like the GDP, interest rates, and inflation are correlated with aggregate corporate output (revenues) and profits (e.g., Brown and Ball, 1967; Higson, Holly and Kattuman, 2002; Bernstein and Arnott, 2003; Shivakumar 2007). Of particular relevance to our study is the evidence that aggregate earnings news positively correlates with changes in interest rates and future inflation (see Kothari, Lewellen 2

3 and Warner, 2006; Shivakumar, 2007; Cready and Gurun, 2010). The Bureau of Economic Analysis (BEA) releases statistics on aggregate corporate earnings every quarter. BEA typically makes these announcements following individual firms preliminary earnings announcements. In fact, the BEA relies in part on the financial information corporations file with the Securities and Exchange Commission in its estimate of aggregate corporate earnings. But the earnings releases of individual firms precede the BEA releases. We examine whether the resulting aggregate earnings information anticipates inflation news in subsequent periods. Unexpected changes in corporate profitability are likely to affect future inflation in an economy for at least three reasons. First, current profitability, through the persistence of earnings, is indicative of the profitability of future investments in property, plant and equipment, research and development, and human resources. Corporate profitability thus directly affects the perceived attractiveness of investment opportunities, which in turn influences managers investment decisions. Corporate earnings also indirectly affect firms capital investments in an economy. Greater profits facilitate greater investments because financing frictions such as the cost of raising equity capital and the costs of debt overhang are lowered with the increased availability of internal funds (see Hennessy, Levy and Whited, 2007; Lewellen and Lewellen, 2012). The investment decisions influence corporate demand for raw materials and labor, which in turn raises the prices of these inputs, assuming the supply of raw materials and labor is less than perfectly elastic. Second, over the long run, corporate profits, dividends, and share prices all move in sync. Increased share prices indirectly and dividend payouts directly 3

4 contribute to consumer spending, which induces additional demand for goods and services in an economy. 1 Finally, corporate profitability and bank lending are also positively interrelated. Earnings growth typically results from revenue growth, which is associated with increased working capital needs of corporations and their suppliers. Banks are ubiquitous in providing working capital financing to businesses. Improved profitability also lowers the credit risk (or bankruptcy risk) of a business. Thus, banks perceive lower risk in lending as corporate profitability rises, which too contributes to increased bank lending. The increased availability of credit that flows into corporate investments leads to greater demand for products and services, which in turn, increases inflation. Consistent with this effect of bank lending, Bassett et al. (2010) document a significant increase in real GDP and inflation in the quarters immediately following a shock that relaxes bank lending standards. Overall, through their direct effect on aggregate demand for capital goods, through their influence on consumer spending, and through their effect on bank lending, we expect corporate earnings to positively affect future inflation. 2 Preceding discussion outlines economic reasons why we might expect corporate earnings to presage future inflation. However, the literature has not examined whether macroeconomic forecasters fully exploit this link. Forecasters might believe that listed firms earnings are too granular to be useful in forecasting at 1 Davis and Palumbo (2001) estimate that for each dollar of unanticipated stock price increase, consumer spending increases between 4 and 7 cents. 2 An association between corporate profits earned in a quarter and inflation in the contemporaneous quarter can be purely mechanical. A firm s accounting earnings are calculated as the difference between revenues and expenses. Revenues are the product of current selling prices and the quantity sold. Expenses, in contrast, are historical in that the inventory sold in a period is carried on the firm s books at historical costs, which would not fully capture the inflation affecting the selling prices of the finished goods (See Ball, Kothari and Watts, 1993). In contrast to this contemporaneous relation, the focus of this study is on evaluating whether earnings announced in a given period predicts inflation announced in subsequent periods. 4

5 the economy-wide level, or they might believe that corporate earnings do not contain information incremental to scheduled, periodic releases of macroeconomic data. We evaluate (i) whether aggregating individual firms earnings predicts future inflation, and if so, (ii) whether macroeconomic forecasters and capital market participants incorporate such information efficiently into their forecasts. Our findings suggest that macroeconomic forecasters and capital market participants could improve their current forecasts of inflation by incorporating the information in aggregate corporate earnings into their forecasts. Recent studies evaluating the role of aggregate earnings in capital markets provide evidence consistent with aggregate earnings containing inflation news. Kothari, Lewellen and Warner (2006) and Cready and Gurun (2010) show that aggregate earnings surprises are negatively correlated with contemporaneous stock market returns and bond price changes, and are positively correlated with contemporaneous changes in T-bill rates. These findings suggest aggregate earnings convey inflation news. Shivakumar (2007) evaluates whether aggregate corporate earnings changes predict a variety of future macroeconomic activities, including industrial production growth, real GDP, and inflation, and finds that aggregate earnings are related only to future inflation. Preceding discussion summarizes theoretical and empirical reasons to expect aggregate earnings news to provide information about inflation. We examine whether macro-forecasters incorporate the inflation information in aggregate earnings surprises in their forecasts. We investigate this question empirically employing widely used forecasts of Consumer Price Index (CPI) and Producer Price Index (PPI) from Money Market Services International (MMS) survey. 5

6 We begin by documenting that the aggregate of all earnings surprises announced publicly over a three-month window is positively correlated with future actual PPI, but not CPI, after controlling for lagged inflation realizations. We conjecture that these differences in the results for PPI and CPI arise because the CPI primarily captures movements in consumption or retail prices, and includes the effects of taxes and of imports, whereas the PPI primarily captures changes in prices of goods and services, which are more closely linked to producers earnings. We next evaluate whether aggregate earnings surprises measured in a particular month predict future inflation forecast errors. Aggregate earnings surprises are found to predict macroeconomists forecast errors in PPI released in the subsequent two months. That is, inflation forecasters do not fully incorporate the information about PPI contained in aggregate earnings surprises. No such evidence is found for CPI forecast errors, which is consistent with our earlier finding that aggregate earnings contain little information about future CPI. The predictive ability of aggregate earnings for PPI forecast errors is robust to alternative methodologies and to controls for other potential biases in the inflation forecasts. We also find that this relation is driven mainly by the aggregate earnings surprises of non-financial firms, suggesting that subsequent inflation shocks are not just the result of greater bank lending at a time when their profits are unexpectedly high. Finally, we analyze capital market reaction to inflation news. Since our earlier analysis reveals that the PPI news is predictable based on aggregate earnings surprises, we examine whether this predictive ability carries on to the bond and equity market reactions as well. Aggregate earnings surprises are positively related to 6

7 Treasury-bill yield changes at subsequent PPI announcements. We do not find similar evidence for capital market reactions to CPI announcements, which is consistent with our earlier results from the analysis of CPI forecast errors. We also do not find such evidence in the equity market, consistent with equity market prices not reacting significantly to the inflation news in aggregate earnings surprises. Equity prices seem to react to the information in earnings surprises at the time of individual firms earnings announcements. The rest of the paper is organized as follows. In the following section, we discuss prior literature on the rationality of survey forecasts and capital market reactions to inflation news. Section 3 explains our sample selection and empirical methodology, and presents the descriptive data, and Section 4 presents the main results from the analysis. Section 5 concludes. 2. Prior Literature There is little previous research on inflation information in aggregate earnings. However, there is evidence of an association between aggregate earnings and security prices, and between security prices and inflation. Below we discuss these studies briefly as a prelude to our empirical analysis. We also summarize the literature on the macroeconomic survey data on inflation, the properties of the survey data, and the pricing of inflation expectations as evidenced in the intra-day and longer-term movements in stock price indices. 2.1 Stock-market returns and inflation The literature on stock market returns and inflation can be classified into three parts: (i) The relation between inflation, nominal interest rates, and stock prices; (ii) The effect of news about interest rates or inflation on stock prices; (iii) Whether 7

8 capital market participants efficiently process information about inflation in setting stock prices. We briefly summarize this literature as it bears on the questions examined in our study. Based on Irving Fisher s proposition that nominal interest rates contain market assessments of expected inflation, Fama and Schwert (1977) evaluate whether stock returns are positively correlated with expected inflation and find that this is not the case. They surprisingly find that stock returns are negatively correlated with both the expected and unexpected components of inflation. Fama (1981) explains this puzzle as being driven by a positive correlation between stock returns and real macroeconomic variables and a negative relation between real activity and inflation. French, Ruback and Schwert (1983) extend this literature to the cross-section by testing the nominal contracting hypothesis, which states that stocks of firms with more nominal contracts should be more sensitive to unanticipated inflation, where the Consumer Price Index (CPI) is used to proxy for inflation. Evaluating the effect of unanticipated inflation on debt contracts and on depreciation tax shields, they reject the nominal contract hypothesis. Bernard (1986) extends the scope of nominal contracts examined and documents cross-sectional variation in the relation between inflation (CPI) shocks and stock returns. Modigliani and Cohn (1979) represent the genesis of the literature that evaluates whether capital market participants efficiently incorporate inflation information in security prices and in their earnings forecasts. Modigliani and Cohn (1979) conjecture that, when valuing stocks, investors fail to incorporate the effect of inflation on nominal earnings, causing stock-market yields to be depressed in periods of high inflation and to be excessive in periods of deflation. Campbell and 8

9 Vuolteenaho (2004) provide empirical support for this hypothesis and show that inflation illusion explains almost 80% of the time-series variation in mispricing of the S&P 500. Numerous other studies also demonstrate that stock prices and market participants do not fully incorporate information about inflation. For example, Chordia and Shivakumar (2005) find that lagged inflation predicts stock prices, and that this predictability varies systematically across firms, based on a firm s earnings exposure to inflation. Basu, Markov and Shivakumar (2010) suggest that the source of the predictability might be because analysts do not incorporate expected inflation information efficiently in their earnings forecasts. Konchitchki (2011) attributes the predictability in part to financial statements not recognizing inflation gains and losses in a timely manner. Based on stock price analysis, he concludes that investors do not seem to fully distinguish monetary and non-monetary assets, which is fundamental to understanding the economic effects of inflation for a firm. In contrast to this literature, which is focused on firm-level forecasts, our study investigates macro-level forecasting by testing whether macroeconomic forecasters efficiently incorporate inflation information contained in lagged aggregate earnings surprises. 2.2 Aggregate earnings and stock-market returns Recent literature demonstrates a negative relation between aggregate earnings growth and stock returns (see Kothari et al., 2006), which is potentially rooted in news about inflation in aggregate earnings. The likelihood of news about inflation in aggregate earnings serves the basis for our examination of whether macroeconomic 9

10 forecasters and market participants efficiently incorporate the inflation information in aggregate earnings. Kothari et al., (2006) conjecture that the surprising finding of a negative relation between aggregate earnings and stock returns might be because aggregate earnings growth contains news about changes in discount rates. Consistent with this conjecture, they find that aggregate earnings growth is strongly correlated with changes to several discount rate proxies, such as T-bill rates, slope of term structure and default spread. Research following Kothari et al. (2006) offers alternative explanations as well as reinforces the conjecture about inflation news in aggregate earnings. Specifically, Ball, Sadka and Sadka (2009) suggest that aggregate earnings growth is related to expected discount rates (which include expected inflation), rather than to discount rate news. Cready and Gurun (2010) more directly evaluate the proposition of Kothari et al. (2006) that aggregate earnings changes captures shocks to discount rates by using short-window market reactions to earnings announcements as a proxy for earnings news. They show that earnings announcement returns aggregated across the stocks announcing earnings in a day are negatively related to returns on the broader market indices. They further show that the earnings announcement returns are associated with contemporaneous changes in inflation expectations. Their evidence supports the view that aggregated earnings news contains information about discount rates. 2.3 Macroeconomic survey data on inflation and stock prices Survey data from Money Market Services (MMS) has been widely used in the macroeconomics and finance literature to isolate the unanticipated information component in scheduled macroeconomic announcements, and to study whether capital 10

11 markets react to macroeconomic announcements. 3 Consistent with macroeconomic news measured from these survey forecasts having significant capital market effects, Urich and Wachtel (1984) document that the unanticipated component of PPI, but not of CPI, has an immediate positive effect on short-term interest rates, and that there is no evidence of delayed market reaction to the inflation news. Pearce and Roley (1985) find that daily stock prices react to monetary news, but not to news about CPI, where the news is measured relative to expectations provided by MMS survey data. McQueen and Roley (1993) show that the relation between PPI and stock returns is particularly strong during periods of high economic growth. Hardouvelis (1987) finds that Treasury bond markets, but not stock markets, react to both PPI and CPI news. In addition, a vast literature focuses on the intra-day impact of macroeconomic news releases on capital markets. This research is mostly motivated by finance microstructure issues such as market s efficiency in processing new information (see Lieberman, 2011). Because macroeconomic survey data generate economically important capital market reactions, research has also investigated survey forecasts for rationality and biasedness. These studies generally yield mixed results. Pearce and Roley (1985) find that unbiasedness of inflation forecasts from MMS surveys cannot be rejected. Aggarwal, Mohanty and Song (1995) find that survey forecasts of several macroeconomic series, including CPI, are consistent with the rational expectations hypothesis, but that forecasts of a few other macroeconomic series, including PPI, are not consistent with it. McQueen and Roley (1993) also report mixed results. They find that the survey data, although not always unbiased or efficient, are generally 3 Studies using MMS data to compute macroeconomic news include Urich and Wachtel (1984), McQueen and Roley (1993), Elton (1999), Balduzzi, Elton and Green (2001), Flannery and Protopapadakis (2002), Andersen et al. (2003), Bernanke and Kuttner (2005), Gürkaynak, Sack and Swanson (2005) and Andersen et al. (2007). 11

12 superior to auto-regressive time-series models of macroeconomic data in that they have smaller root-mean-square errors. To summarize, inflation expectations rank among the critical variables in an economy, and have important ramifications for resource allocation and for product and financial markets (Thomas, 1999). Firms employ inflation forecasts in a variety of their decisions, including pricing of products, setting wages, planning investment, and capital budgeting. It is important to understand whether these expectations reflect all available public information. Although prior studies focus on whether forecasts are unbiased and efficient, they do not identify potential sources of bias. We extend the literature by evaluating whether macroeconomic forecasters rationally incorporate inflation information in aggregate earnings surprises. 3. Sample Selection and Data Description 3.1 Sample selection and variable definitions Our initial sample consists of all NYSE, AMEX and NASDAQ firms in the merged CRSP/Compustat database, with data available on earnings announcement dates. From this sample, we include only the firms available on IBES. For each quarterly earnings announcement, we use the most recent IBES consensus analyst forecast prior to the announcement to calculate firm-level earnings surprises. 4 Specifically, a firm-level earnings surprise is calculated as IBES-reported quarterly earnings minus the most recent consensus analyst earnings forecast from IBES, divided by the absolute value of the most recent consensus analyst earnings forecast. We calculate aggregate earnings surprises by taking value-weighted and equal- 4 We do not impose any requirement on the minimum number of analysts following a firm. However, in unreported analysis, we find that our conclusions are even more strongly supported if we analyze only firms that are followed by at least 5 analysts. Our results are also robust to computing the consensus forecasts from IBES Detail History file and including only analysts forecasts issued in the 60 days prior to an earnings announcement. 12

13 weighted averages of firm-level earnings surprises over a quarter (three months). In particular, VIBES (EIBES) is the quarterly value-weighted (equal-weighted) average of IBES earnings surprises. In calculating aggregate earnings surprises, we exclude all firms with share price less than $1, and earnings surprises in the top and bottom 0.5% of the empirical distribution. We obtain data on monthly actual inflation from the Bureau of Labor Statistics over the period from October 1984 to December The starting date of this sample period is governed by the availability of IBES data to construct aggregate earnings surprises. 5 Our analyses focus on both PPI and CPI measures of inflation. These have significant differences in the way they are constructed, which could potentially cause differences in the results for these two inflation proxies. First, whereas PPI is constructed from the overall market output of US producers, CPI is based on goods and services purchased by urban US households. Thus PPI excludes imports, but CPI includes them. Second, PPI is based on the revenue received by producers, but CPI is based on out-of-pocket expenditure by consumers. Therefore CPI includes sales and excise taxes but PPI does not. Thus, to the extent that the US listed companies reflect outputs by US producers rather than consumption by consumers, we would expect aggregate earnings surprise to be more strongly related to PPI than to CPI. 6 5 As we discuss in our robustness tests, we also use seasonal random-walk-based measures of earnings surprises from Compustat as an alternative measure of aggregate earnings surprises. Although this proxy does not require analyst data, for comparability we restrict the sample using this measure to the same period as those relying on IBES-based earnings surprise measures. Nonetheless, we have confirmed that our results are robust to extending the data for the random-walk-based measures to start from January 1980, which is when inflation forecast data are available from MMS surveys. 6 Aggregate earnings of US listed firms include earnings from products sold outside the US. The effect of such foreign earnings on the PPI depends on whether the product is manufactured in the US or not. Our discussions with officials at Bureau of Labor Statistics reveal that, if a product is manufactured in the US, then the prices of such products are considered for PPI, but not otherwise. In our analysis, we considered excluding foreign earnings from the definition of aggregate earnings. Given that some of the foreign earnings are relevant for PPI, a priori it is not obvious that foreign earnings should be excluded from aggregate earnings. In any case, we could not conduct our analysis on domestic 13

14 PPI and CPI numbers are generally announced at 8:30 a.m. EST, before stock markets open. During our sample period, for 94% of the time, the PPI announcements preceded the CPI announcements, with an average difference of five days. We utilize MMS surveys to construct our proxies for market participants inflation expectations. The dataset includes median polled forecast values for a variety of macroeconomic series, including PPI and CPI. The MMS surveys are conducted every Friday morning among senior economists and bond traders with major commercial banks, brokerage houses, and some consulting firms, mostly in the greater New York, Chicago, and San Francisco areas. MMS surveys provide a timely source of inflation expectation to market. For CPI and PPI surveys specifically, we find that the median numbers of days between the surveys and the actual release of the inflation index are five and seven, respectively. We use the difference between actual and expected inflation as our proxy for inflation forecast errors. For each month t, we compute the forecast errors for CPI (ECPI t ) as the percentage monthly change in actual CPI released that month minus the median forecasted percentage change in CPI. EPPI t, which is the forecast errors for PPI released in month t, is similarly computed. Although we use the original inflation figures reported every month for the actuals, our results are qualitatively unaffected when we replace the original figures by the final revised figures. 3.2 Research design and summary statistics To investigate whether macroeconomic forecasters efficiently incorporate inflation-related information in aggregate earnings surprises, we estimate the following pooled regression: earnings, as segment data on domestic earnings is available only on an annual basis in COMPUSTAT, whereas our analysis requires data at a monthly or at most quarterly frequency. Also, analysts forecasts for domestic earnings are unavailable to compute surprises on domestic earnings. 14

15 = +, , +, +, +, +, + (1) where Inflation Surprise t+j is either ECPI t+j or EPPI t+j for inflation released in month t+j (j = 0, 1 or 2) and Aggregate Earnings Surprise t-1,t-3 is either VIBES t-1,t-3 or EIBES t- 1,t-3, computed using earnings announced in months t-1 to t-3. The control variables are defined below. 7 To allay concerns that any observed relation between aggregate earnings surprise and future inflation is confounded by omission of correlated macroeconomic information, Equation (1) includes control variables for two macro-variables a real output variable and a consumption measure and three interest rate variables. Specifically, the following five control variables are included in the regression model: (i) Change in default spreads, calculated as the difference in interest rates between AAA bonds and BAA bonds from month t-4 to t-1 ( DEFAULT t-1,t-3 ), (ii) Change in yield spreads, calculated as the difference in interest rates between federal funds rate and risk free rate from month t-4 to month t-1 ( YIELD t-1,t-3 ) (iii) Change in term spreads, calculated as the difference in interest rates between 10-year government bonds and risk free rate from month t-4 to month t-1 ( TERM t-1,t-3 ) (iv) Percentage change in industrial production index from month t-4 to month t-1 (PROD t-1,t-3 ) and 7 Throughout, we use subscripts to refer to the month in which we measure the variable. For instance, ACPI t refers to CPI released in month t. If a variable is aggregated across more than one month or is computed as a change over more than a month, then the first subscript refers to the starting month of the measurement period and the second subscript refers to the ending month. Thus, VIBES t-1,t-3 refers to the aggregate value-weighted earnings surprises that are based on earnings released in months t-3 to t- 1, while CONS t-1,t-3 refers to the percentage change in personal consumption expenditures that occurred in months t-3 to t-1 (which is calculated as percentage change from month t-4 to month t-1). 15

16 (v) Percentage change in personal consumption expenditures from month t-4 to month t-1 (CONS t-1,t-3 ). 8 We obtain data for macro-economic controls from Federal Reserve Economic Data (FRED) website of St. Louis Fed. 9 Additionally, to control for potential serial correlation, lagged inflation surprises in month t-1 to t-3 are included in the set of controls. 10 We regress inflation surprise announced in month t (or, alternatively, in month t+1 or t+2) on aggregate earnings surprise calculated from earnings announced over months t 3 to t 1. We aggregate quarterly earnings announced over a rolling threemonth period to ensure that our aggregate measures encapsulates the earnings information of all listed companies. 11 We obtain qualitatively similar results, however, when we replace Aggregate Earnings Surprise t 1,t 3 by the aggregate earnings surprise measured in month t 1 alone. The regression Equation (1) is estimated using 300 monthly observations from January 1985 to December Table 1 reports summary statistics for the variables used in our tests. During the sample period, the average CPI and PPI inflation rates were 0.24% and 0.18% per month. Moreover, macroeconomic forecasts were unbiased predictors of inflation. The median inflation forecast errors are zero for both the CPI and PPI measures (ECPI and EPPI), and the mean forecast errors are 0.013% for CPI and 0.033% for PPI. For both CPI and PPI, the median and mean forecast errors are statistically indistinguishable from zero at the 5% level or better. 8 Our results are qualitatively similar when we replace the changes in financial variables with their levels at end of month t-1. 9 Additionally, including equity market returns, defined as the value-weighted CRSP index measured over months t-3 to t-1, leaves our results unchanged. The coefficient on the equity market returns is insignificant in the regressions. 10 For both CPI and PPI, aggregating forecast errors over months t 3 to t 1 rather than including each month s inflation forecast errors separately in the regressions has little effect on our conclusions. 11 We aggregate earnings surprises over months t 3 to t 1 rather than include each month s aggregate earnings surprise separately to avoid multicollinearity issues, as aggregate earnings surprises within a quarter are autocorrelated at the monthly frequency, with autocorrelation coefficients varying from 0.17 to 0.62 for the various aggregate earnings surprise measures. 16

17 We do not find any evidence that CPI and PPI forecast errors are first-order serially correlated at conventional statistical significance levels. The mean aggregate earnings surprise based on analysts earnings forecasts is 6.7% for the value-weighted measure (VIBES t-1,t-3 ) and 14.4% for the equalweighted measure (EIBES t-1,t-3 ). These averages are significantly different from zero, as are the median aggregate earnings surprises. This indicates that analysts tend to be optimistic in their earnings expectations, and is consistent with the prior literature (e.g., Brown, Foster and Noreen, 1985). 12 Not surprisingly, aggregate earnings surprises are highly serially correlated as overlapping firm-level earnings are used to compute the monthly aggregate earnings. Among the control variables, the mean and median values for DEFAULT t-1,t- 3, YIELD t-1,t-3, and TERM t-1,t-3 are insignificantly different from zero. This suggests that there was no trend in the interest rate variables during the sample period. The average production index increased by 0.467%, and average personal consumption increased by 1.372% during our sample period. These indicate that the sample period is characterized by economic growth. Table 2 reports Pearson correlations (above the diagonal) and Spearman correlations (below the diagonal) among the variables of interest. The actual (as well as forecasted) PPI and CPI measures of inflations are highly correlated with each other. The correlation coefficients across these alternative measures of inflation exceed 0.6, suggesting a large overlap in the price increases measured using PPI and CPI. We find that ECPI t and EPPI t, forecast error proxies for CPI and PPI, are significantly positively correlated with each other (Spearman correlation coefficient is 12 The extant literature on analysts forecasts documents a variety of biases and inefficiencies in analysts earnings forecasts. To the extent that biases in analysts forecasts induce noise in our earnings surprise measures for the aggregate market, our analyses would be conservatively biased. Nonetheless, we check the robustness of our results to time-series based measures of earnings surprises. The results from the robustness analysis are discussed in Section

18 0.183). Both the PPI and CPI forecast errors tend to be insignificantly correlated with aggregate earnings surprises in this univariate analysis. No clear pattern of statistical significance is observed in the correlations between forecast errors and macroeconomic control variables. But, we find consistently positive and significant correlations between aggregate earnings surprises and change in yield spreads as well as industrial production index. 4. Multivariate Evidence This section begins with a discussion of the analysis evaluating inflation information in aggregate earnings news. We then analyze whether macro-forecasters and capital market participants efficiently incorporate inflation information in their forecasts. 4.1 The relation between aggregate earnings surprises and inflation realizations We estimate a vector auto-regressive (VAR) model to examine whether aggregate earnings surprises have information about future innovations to inflation. In particular, we estimate the following model: = (2) where z t is a vector. The variables included in the vector are: (i) actual inflation released in month t (ACPI t or APPI t ), (ii) aggregate earnings surprise (VIBES t,t-2 or EIBES t,t-2 ), (iii) default spread in month t (DEFAULT t ), (iii) yield spread in month t (YIELD t ), (iv) term spread in month t (TERM t ), (v) percentage monthly change in industrial production index (PROD t ) released in month t, and (vi) percentage monthly change in personal consumption expenditures (CONS t ) released in month t Although this analysis is similar in spirit to that in Shivakumar (2007), who estimates OLS regression of future inflation and other macroeconomic series on lagged aggregate earnings, there are a 18

19 All the variables in the VAR system appear to be stationary based on the Phillips-Perron tests for unit root. 14 Our use of 4 lags in the system is identified based on Akaike s information criterion and on the Final Prediction Error criterion. Consistent with prior macroeconomic literature on time-series prediction of level of inflation (e.g., Balduzzi, 1995 and Estrella and Mishkin, 1997), the VAR analysis includes macroeconomic variables in percentage changes and financial variables in percentage rates. Table 3 reports results from estimating Equation (2). Although the VAR model includes a variety of variables and many lags of each variable, to conserve space, we report only the coefficient estimates for the first lag of the variables from regressions using our main variable of interest, viz. actual inflation, as the dependent variable. The key finding is that, irrespective of the aggregate earnings surprise proxy, lagged aggregate earnings surprises predict innovations to actual inflation only for PPI. The coefficients on VIBES t,t-2 and EIBES t,t-2 in regression of actual PPI inflation are and These coefficients are statistically significant at the 5% level or better. More importantly, these coefficients imply that a one-standard-deviation increase in VIBES t,t-2 (EIBES t,t-2 ) predicts PPI to rise by 0.205% (0.301%), which is economically large compared to the mean (median) value of APPI t of 0.181% (0.200%) in our sample. In untabulated results, we find that the coefficient on other lags of aggregate earnings (lags 2, 3, and 4) are generally statistically insignificant. Focusing on the coefficients of lagged inflation, we find that the coefficients on 1 st lag of actual inflation are always positive and significant, indicating persistence in inflation. In undocumented results, we consistently observe that the coefficients on couple of key differences between the two studies. First, the focus in Shivakumar (2007) is on macroeconomic levels rather than on innovations to macroeconomic activity. Second, Shivakumar (2007) considers each macroeconomic series independently rather than as part of a system of equations. 14 Our findings are consistent with Halunga et al. (2009), who confirm that inflation, measured as change in CPI, is integrated of order zero since the early 1980s. 19

20 2 nd lag of inflation are negative in CPI regressions. Coefficients on all other lags of inflation variables are statistically insignificant. Collectively, the results show that aggregate earnings surprises forecast future inflation, even after taking into account potential persistence in actual inflation and controlling for the effects of a broad set of macro-economic and financial control variables. This finding holds only for PPI. It supports the view that unexpected corporate profits anticipate future inflation, as conjectured earlier. 4.2 Predictive ability of aggregate earnings surprises for inflation forecast errors If macroeconomic forecasters were to incorporate information in aggregate earnings surprises efficiently, then such surprises should be unrelated to forecast errors revealed at subsequent inflation announcements. This is similar in spirit to the rationale underlying the tests of the efficiency of analysts earnings forecasts. We test the above prediction by estimating regression Equation (1). The regression results are reported in Table 4. Panel A presents the results of regressing inflation surprises on monthly value-weighted IBES-based earnings surprise measures. Panel B presents the results from regressions using monthly equalweighted IBES-based earnings surprise measures. In all regressions of PPI-inflation forecast errors, we consistently observe a significant coefficient on aggregate earnings surprise proxies. In both Panel A and Panel B, coefficients on aggregate earnings surprises are significantly positive for the one-month-ahead and two-months-ahead PPI forecast errors, indicating that macroeconomic forecasters do not fully incorporate the information in aggregate earnings surprises for future PPI inflation. The coefficient on VIBES t-1,t-3 (EIBES t-1,t-3 ) is (0.465) for the one-month-ahead PPI-inflation forecast error when macro- 20

21 economic control variables are included in the models. This coefficients decline monotonically for subsequent months forecast errors. The (0.465) coefficient implies that a one-standard-deviation increase in VIBES t-1,t-3 (EIBES t-1,t-3 ) increases forecast errors by 0.081% (0.080%), which is economically large, about 40%, compared to the average monthly change in PPI of 0.181%. 15 When we focus on regressions of CPI-inflation forecast errors, the coefficient on aggregate earnings surprise proxies is always insignificant. This result holds across both panels A and B, indicating that irrespective of how aggregate earnings are measured and irrespective of whether we control for macro variables, aggregate earnings surprises have little predictive ability for CPI-inflation forecast errors. In addition to failing to fully incorporate PPI-inflation information in earnings, forecasters also appear to neglect some of the information in lagged inflation itself. This is evident from the significance of lagged inflation forecast errors included as control variables in Equation 1. The coefficients are significantly negative in the second lag when CPI forecast error is the dependent variable and in the third lag when PPI forecast error is the dependent variable. The coefficients on the remaining macro control variables are generally statistically insignificant. Overall, the results from Table 4 suggest that macroeconomic forecasters do not fully incorporate the inflation-related information in aggregate earnings news in their forecasts for future PPI. We do not find any such inefficiency in the CPI forecast. This is consistent with our results in Table 3, where we find that aggregate 15 To investigate whether the predictive ability of aggregate earnings surprise for PPI differs across macroeconomic states, in untabulated tests, we interacted VIBES t-1,t-3 (EIBES t-1,t-3 ) with a dummy for economic recessions defined by National Bureau of Economic Research (NBER). The coefficient on this interaction term is statistically insignificant, while that on aggregate earnings surprise is very similar to those reported here, indicating that the predictive ability of aggregate earnings on inflation forecast errors does not differ across macroeconomic expansions and contractions. 21

22 earnings surprises predict future innovations in PPI-inflation, but not the innovations in CPI-inflation. The dichotomy in the results between PPI and CPI inflation is in line with the view that listed firms revenues are dominated by transactions with other businesses rather than with individuals, and so their earnings are more informative about inflation in producers prices, i.e., PPI, rather than about retail price inflation reflected in CPI. In addition, since PPI figures are released on average 5 days before the CPI announcements, forecasters have more information available to forecast CPI. 4.3 Robustness tests Time-series measures of earnings surprise Since prior studies show that individual analyst forecasts tend to be optimistically biased, we test the sensitivity of our analysis to an alternative timeseries-based measure of aggregate earnings surprise. This analysis also allows us to relax the assumption that a firm needs to have analyst following in IBES before its earnings surprises are included in aggregate earnings. This alternative measure of earnings surprise is computed as seasonally differenced quarterly reported earnings per share from Compustat, divided by the absolute value of the quarterly reported earnings per share four quarters ago. Earnings are measured before extraordinary items and discontinued operations. VCOMP t-1,t-3 (ECOMP t-1,t-3 ) is the quarterly valueweighted (equal-weighted) average of earnings surprises calculated using Compustat earnings surprises. In unreported results, we find that the mean Compustat-based measures of earnings surprises, VCOMP t-1,t-3 and ECOMP t-1,t-3, are insignificantly different from zero, although the median values are significantly positive. We report results of estimating Equation (1) when we use Compustat based aggregate earnings surprises in Table 5. Panel A (B) reports results from analyses using value-weighted (equal-weighted) Compustat aggregate earnings surprises. 22

23 Consistent with Table 4, we find that Compustat-based aggregate earnings surprises significantly predict future PPI forecast errors up to two months but not CPI forecast errors. We obtain these results irrespective of whether we use value-weighted or equal-weighted aggregate earnings surprises and irrespective of whether we include macro-economic control variables Alternative definition of aggregate earnings surprise We check the robustness of our results to the approach used to aggregate firmlevel earnings surprises by calculating aggregate earnings surprise as the ratio of two aggregate numbers. Specifically, we first compute a firm-level unscaled earnings surprise as the difference between reported earnings and the most recent analysts consensus forecasts and then for each month t, aggregate this unscaled earnings surprise across all firms announcing earnings in months t-2 to t. We then divide the aggregate unscaled earnings surprise by the aggregate of the absolute value of IBES consensus earnings forecasts over the months t-2 to t to obtain our alternative measure of aggregate earnings surprise (AGGIBES t,t-2 ). Consistent with our earlier results, Panel C of Table 5 documents that the coefficients on AGGIBES t-1,t-3 are significantly positive for up to three months when PPI forecast error is the dependent variable. When CPI forecast error is the dependent variable, the coefficients tend to be negative and marginally statistically significant. These coefficients are, however, much smaller in magnitude than those in PPI forecast error regressions and have low economic significance Quarterly analysis The precise timing of the impact of aggregate earnings surprises on future inflation is unknown. Although the analysis in the previous section indicates that aggregate earnings surprises affect PPI forecasts for up to two-months ahead, we 23

24 aggregate inflation forecast errors over a quarter, and re-estimate Equation (1) at a quarterly frequency to increase the power of the tests. For quarterly regressions, we regress average inflation forecast errors announced in each calendar quarter q+1 (i.e., average of forecast errors for inflation figures released in months t, t+1, and t+2) on aggregate earnings news calculated in calendar quarter q (i.e., based on earnings released in months t 1, t 2, and t 3). The quarterly regressions also control for lagged inflation forecast errors, measured in quarter q and include the set of macro control variables as before. We have 100 quarterly observations for these regressions. Results of this quarterly analysis are reported in Panel D of Table 5. The results show that the coefficient on aggregate earnings surprise is consistently positive and significant for regressions of PPI-inflation forecast errors, but not for CPIinflation forecast errors. The coefficient estimates are and in the PPI forecast error regressions, with t-statistics of 2.97 and 2.76, depending on the specific proxy employed for aggregate earnings surprise. The aggregate earnings surprise also has reasonable explanatory power for PPI forecast errors, as seen by the adjusted R 2 of 8% and 7% in the quarterly regressions. The coefficients on lagged forecast errors are significantly negative. Finally macro-economic control variables are generally statistically insignificant except that changes in yield spreads are weakly and negatively associated with quarterly PPI forecast errors Relationship of aggregate earnings news with PPI versus CPI forecast errors Our finding of a relation of aggregate earnings news with PPI forecast errors, but not with CPI forecast errors appears to contradict the evidence in prior studies that document a positive relation between aggregate earnings news and future CPI (e.g., 16 We obtain qualitatively similar results when we use Compustat based aggregate earnings surprises in the quarterly analysis. 24

25 Shivakumar, 2007). To better understand this difference, we first replicate the evidence in Shivakumar (2007). Every month, we regress the actual CPI-inflation in that month on aggregate earnings changes measured over prior three months (VCOMP t-1,t-3 ). 17 The results from this regression are reported in Panel A of Table 6. Consistent with Shivakumar (2007), we find a significantly positive coefficient on VCOMP t-1,t-3. However, when lagged inflation and other control variables are included in the regression, the coefficient on VCOMP t-1,t-3 turns insignificant. These results indicate that the CPI information contained in lagged aggregate earnings changes documented by Shivakumar (2007) is not incremental to information in other lagged financial and macroeconomic variable. However, when we replace the CPI-inflation with PPIinflation, the coefficient on VCOMP t-1,t-3 is significantly positive, consistent with the results from the VAR analysis reported in Table 3. As a further robustness check for whether the lack of a relation between CPIinflation and aggregate earnings news is unique to the MMS forecast that we use, we re-estimate Equation (1) after replacing ECPI with forecast errors derived from Survey of Professional Forecasters (SPF_ECPI) maintained by the Federal Reserve of Philadelphia. 18 Unlike MMS forecasts, forecasts from the Survey of Professional Forecasters (SPF) are available only on a quarterly basis and are available for CPI, but not for PPI. 19 So, the regressions using SPF data are estimated only on a quarterly basis using CPI forecast errors and mirror the quarterly analysis in Section We focus on VCOMP t-1,t-3 in these regressions to be consistent with Shivakumar (2007). However, we obtain similar conclusions when use ECOMP t-1,t-3, VIBES t-1,t-3 or EIBES t-1,t We use median CPI forecasts reported by Survey of Professional Forecasters and convert annualized CPI forecasts to quarterly rates. 19 The timing of the SPF surveys prior to 1990Q2 is not known with certainty. From 1990 Q2, for each quarter, the surveys are sent out at the end of the first month of that quarter and the responses are due in either the 2 nd or 3 rd week of the middle month. Thus, some uncertainty exists in this database about what exactly the forecasters know at the time the survey is taken. 25