- Email: [email protected]

Contents lists available at ScienceDirect

Journal of Monetary Economics journal homepage: www.elsevier.com/locate/jme

Credit conditions and stock return predictability Sudheer Chava a, Michael Gallmeyer b, Heungju Park c,n a

Scheller College of Business, Georgia Institute of Technology, United States McIntire School of Commerce, University of Virginia, United States c HSBC Business School, Peking University, University Town, Shenzhen 518055, China b

a r t i c l e i n f o

abstract

Article history: Received 10 June 2014 Received in revised form 6 June 2015 Accepted 10 June 2015 Available online 2 July 2015

U.S. stock return predictability is analyzed using a measure of credit standards (Standards) derived from the Federal Reserve Board's Senior Loan Officer Opinion Survey on Bank Lending Practices. Standards is a strong predictor of stock returns at a business cycle frequency, especially in the post-1990 data period. Empirically, a tightening of Standards predicts lower future stock returns. Standards performs well both in-sample and out-ofsample and is robust to a host of consistency checks. Standards captures stock return predictability at a business cycle frequency and is driven primarily by the ability of Standards to predict cash flow news. & 2015 Elsevier B.V. All rights reserved.

Keywords: Stock predictability Credit supply Macroeconomics Survey data

1. Introduction A large literature in finance and economics has documented stock return predictability using variables such as market valuation ratios, short- and long-term interest rates, firm financing patterns, the consumption-to-wealth ratio, and many other economic variables.1 But recently, an active debate has arisen over whether any of these economic variables predict future excess stock returns better than historical average excess returns. Goyal and Welch (2008) argue that many predictive variables used in the literature perform poorly both in-sample and out-of-sample, especially over the last 30 years. In contrast, Campbell and Thompson (2008) show that many predictive regressions beat the historical average return once weak restrictions are imposed on the signs of coefficients and return forecasts. This paper contributes to this literature by providing evidence that an economically motivated predictive variable that measures credit conditions from a survey of bank officers has robust in-sample and out-of-sample predictive power in forecasting excess stock returns. Further, the predictive power is strongest in the post-1990 time period and is quantitatively significant. Our work is motivated by several papers that study how supply-based measures of credit could impact the economy. Some of this work was prompted by papers that have studied the impact of the Federal Reserve's monetary policy on stock returns (Patelis (1997), Thorbecke (1997), and Bernanke and Kuttner (2005) among others) as well as the behavior of business condition proxies such as term premia, default premia, and dividend yields (Jensen et al., 1996). A possible explanation of the predictive power of monetary indicators relates to the credit channel of monetary policy transmission (Bernanke and Gertler, 1995). In particular, tighter monetary policy leads to a reduced and costlier bank loan supply that in turn impacts future stock returns. However, past work has not considered the direct influence of bank loan supply changes

n

Corresponding author. Tel.: þ86 755 26033191. E-mail addresses: [email protected] (S. Chava), [email protected] (M. Gallmeyer), [email protected] (H. Park). 1 See Koijen and van Nieuwerburgh (2013) for an extensive review of the predictability literature.

http://dx.doi.org/10.1016/j.jmoneco.2015.06.004 0304-3932/& 2015 Elsevier B.V. All rights reserved.

118

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

on stock returns. In particular, it is unclear whether the credit channel either through a monetary policy transmission mechanism or some other economic channel has predictive power for stock returns. In this paper, this issue is addressed by examining whether shocks to the aggregate supply of bank loans predict stock returns. Besides the credit channel transmission mechanism, fluctuations in the supply of bank loans can be caused by frictions in the credit creation process through bank views of future market conditions.2 In particular, if agency costs time-vary as in the financial propagation mechanism described in Fazzari et al. (1988) and Bernanke and Gertler (1989), banks can change their supply of credit based on their views of borrowers' balance sheets. In a speech at the Federal Reserve Bank of Atlanta, Bernanke (2007) argues that the supply of bank loans is tightly linked to the credit channel of monetary policy. Bank lending standards, or the terms in which loans are offered, have been used as a bank loan supply measure in several papers to study whether banks change their loan supply systematically over the business cycle and if there is an important loan supply effect in macroeconomic fluctuations. Asea and Blomberg (1998) examine the relationship between the cyclical component of aggregate unemployment and bank lending standards. Using a bank-level panel data set constructed from the terms of individual loan contracts, they find that cycles in bank lending standards are important in explaining aggregate economic activity. Our work uses survey data on bank lending standards obtained from the Federal Reserve's Senior Loan Officer Opinion Survey (SLOS). An earlier study using this data is Lown and Morgan (2006) who find that shocks to lending standards are significantly correlated with innovations in commercial loans and in real output. In particular, they find that “bank lending standards are far more informative about future lending than are loan rates.” Building from this work, Bassett et al. (2014) use bank-level responses on changes in bank lending standards with an econometric model to control for the effect of loan demand. They find that tightening shocks to their credit supply indicator is significantly related to a decline in output and a widening of corporate credit spreads. Gorton and He (2008) show that the relative performance of commercial and industrial loans leads to endogenous credit cycles and is an autonomous source of macroeconomic fluctuations. Despite this linkage of bank lending to macroeconomic variables, limited evidence exists whether changes in bank loan supply predict stock returns, which is our contribution. Keim and Stambaugh (1986), Campbell (1987), Fama and French (1988, 1989), Schwert (1990), and Cooper and Priestley (2009) provide evidence that business condition proxies such as aggregate dividend yield, default spreads, term spreads, the level of short-term interest rates, and a measure of the output gap can predict stock returns. Given some of these variables are also driven by market prices, it is difficult to discern if their predictive power is driven by rational time-varying opportunity sets or simply mispricing. Our work examines whether bank lending standards, a variable that captures aggregate supply-side credit conditions, that is not a direct function of equity market prices, serves as a leading indicator of future stock returns. In contemporaneous work to our own, Adrian et al. (2010) analyze the stock return predictability of several financial intermediary balance sheet variables and find that the annual growth rate of security broker–dealer leverage predicts future stock returns. In a supplementary appendix, our bank lending standards measure, meant to capture the supply of credit from commercial banks, still forecasts stock returns once controlling for the Adrian et al. (2010) broker–dealer leverage measure. Our work joins a growing literature that uses survey data to explain stock returns and macroeconomic variables. Campbell and Diebold (2009) find that expected business cycle conditions obtained from the Livingston survey data has forecasting ability for stock returns. Ang et al. (2007) use the Livingston survey, the Survey of Professional Forecasters, and the Michigan survey to build inflation expectations. They show that the survey-based measures of inflation outperform other forecasting methods out-of-sample. For predictions of various macro variables, Engel et al. (2007), Engel and Rogers (2006, 2009) and Ghysels and Wright (2009) use the Consensus Forecasts survey data. Lown and Morgan (2006) document the predictive power of the SLOS on loan growth, GDP growth, and various other measures of business activity. The SLOS is used to provide direct evidence of the relationship between credit conditions through a bank loan supply measure and future excess stock returns. Overall, our measure of credit conditions derived from the Federal Reserve Board's SLOS is a strong predictor of U.S. stock returns at frequencies up to and including a year. This measure contains information beyond the variables shown to have predictive power from the past predictability literature. Given this measure has been shown to predict macroeconomic variables in Lown and Morgan (2006), our work provides a direct link to stock return predictability and an aggregate macroeconomic supply variable. This credit condition measure performs well both in-sample and out-of-sample. It is also robust to a small sample bias analysis and a host of consistency checks that are summarized in a supplemental appendix. Economically, it is important to understand the source of the predictability of bank lending standards. Our evidence, both from a Campbell and Shiller (1988) decomposition and from a vector autoregression approach advocated in Cochrane (2008, 2011) is consistent with bank lending standards predicting stock returns through a cash flow channel. In particular, bank lending standards strongly predicts future cash flow growth directly. Tightening credit standards predict lower expected future cash flows, hence lower future stock returns. These results are in line with agency or information imperfections at both the firm and bank level impacting real economic activity (Bernanke and Blinder, 1988; Bernanke and Gertler, 1989; Holmstrom and Tirole, 1997; Stein, 1998). Banks inability or unwillingness to extend credit to firms causes firms to reduce investment in some positive NPV investment projects leading to a drop in firm value (Chava and Purnanandam, 2011). Our work joins a small, but growing, literature emphasizing the importance of cash flows in understanding stock market predictability. Recent examples include Lettau and Ludvigson (2005), Larrain and Yogo (2008), Bansal and Yaron (2011),

2

See Berlin (2009) for a survey of models that explain bank lending cycles.

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

119

Garrett and Priestley (2012), Bollerslev et al. (2015), and Maio and Santa-Clara (2015). Given our work presents evidence that Standards predicts future returns through cash flows and the market does not immediately impound this information, our results might seem more consistent with a behavioral interpretation in the market efficiency/behavioral finance debate suggesting that the predictability might disappear in the future. Alternatively, if the concept of rational expectations is expanded as in Condie and Ganguli (2011) to incorporate ambiguity aversion, rational but ambiguity-averse investors may not fully react to news surprises. Rational but ambiguity-averse investors exhibit information inertia even though they face no trading frictions Caskey (2009), Illeditsch (2011), and Condie et al. (2012). The paper is organized as follows. Section 2 describes the data used including the SLOS. The empirical methodology is explained in Section 3. Section 4 presents evidence on return predictability. In Section 5, the source of predictability is explored. Section 6 concludes. 2. Data Our measure of credit conditions is now described along with the stock return data and the additional predictor variables used in our predictability tests. Descriptive statistics of the stock returns and predictor variables are also reported. 2.1. Senior loan officer survey data Our measure of aggregate supply-side credit conditions, through bank lending standards, is derived from a quarterly survey of bank senior loan officers published by the Federal Reserve Board. The survey, titled the Senior Loan Officer Opinion Survey on Bank Lending Practices (SLOS), polls major U.S. banks around the country about credit conditions. The survey was first publicly available starting in the first quarter of 1967 with approximately 120 banks participating. As of the fourth quarter of 2013, 72 banks participated, capturing the general trend of the number of U.S. banks shrinking over time.3 The participating banks capture a sizeable portion of lending by U.S. banks. From Lown and Morgan (2006), survey banks account for “about 60% of all loans by U.S. banks and about 70% of all U.S. bank business loans.” The survey's questions can be classified as measures of supply and demand for commercial and industrial loans, commercial real estate loans, residential mortgage loans, and consumer loans. Our focus is on credit standards for approving commercial and industrial (C&I) loans. The question in the survey pertaining to C&I loans is currently (as of the fourth quarter 2013 survey) asked as follows: For applications for C&I loans or credit lines – other than those to be used to finance mergers and acquisitions – from large and middle-market firms [annual sales of $50 million or more]4 that your bank currently is willing to approve, how have the terms of those loans changed over the past three months? (1) tightened considerably, (2) tightened somewhat, (3) remained basically unchanged, (4) eased somewhat, (5) eased considerably. To convert the survey data into a quantifiable time-series variable, Lown et al. (2000) and Lown and Morgan (2002, 2006) are followed by creating a credit standards index (Standards) as a net percentage of banks tightening credit. Specifically, Standards is computed as the number of banks reporting tightening standards less the number of banks reporting easing standards divided by the total number reporting. The quarterly data is constructed by using the surveys conducted in January (Q1), April (Q2), July (Q3), and October (Q4) of each year from 1967 to 2013. The Federal Reserve makes the results of these surveys public in the month following when the survey was taken. For example in 2013, the Q1 through Q4 surveys were released on February 4, May 6, August 5, and November 4 respectively. Hence, the Standards number pertaining to a specific quarter is publicly known well before the end of that quarter.5 Lown and Morgan (2006) find that changes in Standards are strongly correlated with real output and bank loan changes. In particular, they show that Standards strongly dominates loan interest rates in explaining variation in the supply of business loans and aggregate output. They also show that Standards remains significant when proxies for loan demand are included, which suggests Standards can be used as a proxy for loan supply as done in our work. Other recent studies also employ Standards as a measure of bank loan supply. Gorton and He (2008) analyze the relationship between their Performance Difference Index (PDI) and Standards to explain the time-series behavior of the Credit Standard Survey responses. Leary (2009) uses Standards as an alternate proxy for changes in bank loan supply to show the role of credit supply in capital structure choice. Bassett et al. (2014) use bank-level responses on changes in Standards to analyze the relationship between tightening shocks to bank credit supply and economic activity. Our work differs in that Standards is used as a measure of aggregate supply-based credit conditions to examine the link between stock return predictability and supply-side credit conditions. 3 Recent survey results are available at http://www.federalreserve.gov/boarddocs/surveys. The survey is carried out for both large banks and small banks. Controlling for the impact of small banks, the predictability of Standards is examined with only the large banks from Q1:1997 to Q4:2013 using the survey reports from the Federal Reserve Board website. The predictability of Standards is still robust. 4 The survey also has a question of credit standards for small firms. The predictability of a small firm Standards measure is also analyzed. These results are similar to our main findings. 5 The Standards series, including updates for the most recent survey, is available at Donald Morgan's web site: http://www.newyorkfed.org/research/ economists/morgan/index.html.

120

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

The Standards data series from Q2:1990 to Q4:2013 is used. Though the SLOS was made public starting in 1967, the data from the commercial and industrial (C&I) loan standards question pre-1990 faces several issues.6 First, the wording of the C&I loan standards question was not consistent across the pre-1990 time period. From 1978 through 1983, the C&I loan standards question was split into two separate questions. The first question asked how standards changed for prime rate loans, while the second question asked how standards changed for above prime rate loans. However, as documented in Brady (1985), the link between market loan rates and the prime rate weakened during this time. Banks largely began pricing loans to large borrowers at market rates. Prime-based rate loans were largely reserved for smaller and low credit quality borrowers. Hence, the C&I loans standards question was no longer a reflection of changing credit supply for large borrowers. Second, the C&I loan standards question was even suspended for a time, as it was not asked from Q1:1984 until Q2:1990. Finally, Schreft and Owens (1991) note that from 1967 through 1983 survey respondents almost never reported a net easing of standards on business loans, suggesting a possible bias in the early years of the survey. They hypothesize that the incentive to always report tightening standards might exist “if respondent banks perceive a risk of closer regulatory scrutiny if they admit to having eased standards.” Given these issues with the pre-1990 C&I loans standards question, the focus of our study is on the post-1990 data.7 Fig. 1 plots the Standards measure across time with the shaded regions representing the NBER recession periods. In our main analysis period, Q2:1990–Q4:2013, there are three NBER-dated recessions. In all cases, it appears that Standards has tightened entering a recession. Equally important, banks appear to relax lending standards exiting a recession. From the figure, it appears that Standards is a leading indicator of a business cycle. At least at a univariate level, it seems plausible that Standards is a contender for predicting stock returns.

2.2. Stock returns and other predictor variables used To study predictability, stock returns on the CRSP value-weighted index (CRSP-VW) and the S&P 500 index are analyzed. All stock returns are expressed as continuously compounded returns with dividends included. To calculate excess stock returns, the continuously-compounded 30-day T-bill rate is used as the risk-free rate. To compare the forecasting power of Standards in the predictability regressions, some of the standard price-based predictor variables used in the predictability literature are also considered: the dividend–price ratio (dp), the 30-day T-bill rate (RF), the term spread (TERM), and the default yield spread (DEF). The dividend–price ratio, dp, is the difference between the log of dividends and the log of the CRSP-VW index price. The dividends are 12 month moving sums of dividends paid on the CRSP-VW index. TERM is computed as the difference between the yield on a 10-year and a 1-year government bond. DEF is computed as the difference between the BAA-rated and AAA-rated corporate bond yield. Data on bond yields are collected from the FRED database at the Federal Reserve Bank of St. Louis. The forecasting power of Standards is also compared to the aggregate consumption–wealth ratio measure, cay, from Lettau and Ludvigson (2001); a measure of corporate issuing activity ntis from Goyal and Welch (2008); and a measure of the output gap from Cooper and Priestley (2009). As a measure of the aggregate consumption–wealth ratio, Lettau and Ludvigson (2001) estimate ct ¼ α þ βa yt þ

k X i ¼ k

ba;i Δat i þ

k X

by;i Δyt i þϵt ;

ð1Þ

i ¼ k

where t ¼ k þ 1; …; T k, c is aggregate consumption, a is aggregate wealth, y is aggregate income, and ϵ is an error term. dt ¼ ct β^a at β^ y yt , t ¼ 1; …; T. Goyal and Welch Using an estimated coefficient from the above equation provides cay cay (2008) also estimate a cay measure that excludes advance information from the estimation equation. The Goyal–Welch measure of cay is used for our predictability test.8 Goyal and Welch (2008) use Net Equity Expansion (ntis) as a measure of corporate issuing activity. The variable ntis is computed as the ratio of the 12-month moving sum of net issues by NYSE listed stocks divided by their total end-of-year market capitalization. This dollar amount of net equity issuing activity (IPOs, SEOs, stock repurchases, less dividends) for NYSE listed stocks is computed from CRSP data as NetIssuet ¼ Mcapt Mcapt 1 ð1 þvwretxt Þ;

ð2Þ

where Mcap is the total market capitalization and vwret is the value-weighted return (excluding dividends) on the NYSE index. Goyal and Welch document that ntis is closely related to a payout variable proposed in Boudoukh et al. (2007). 6

See Schreft and Owens (1991) for a discussion of how the SLOS evolved pre-1990. As a robustness check presented in our supplementary appendix, a Standards series from Q1:1967 to Q4:2013 is constructed by splicing together the Q1:1967 to Q4:1983 data with the Q2:1990 to Q4:2013 data. To fill in the missing data from Q1:1984 to Q1:1990, the one question that has remained relative constant through the entire lifetime of the SLOS is used – a question concerning a bank's willingness to make consumer installment loans. Using a similarly constructed variable for this consumer willingness question, the Standards variable from Q1:1967 to Q4:1983 is regressed on it. The missing Standards data from Q1:1984 to Q1:1990 is then constructed from this fitted regression model that uses the consumer willingness variable. 8 The Goyal–Welch measure of cay is available by Q4:2012 at Amit Goyal's web site: http://www.hec.unil.ch/agoyal/. 7

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

121

Fig. 1. Change in standards from the senior loan officer survey 1967–2013. Note: This figure plots the changes in the Standards variable constructed from the Senior Loan Officer Survey. The sample period is Q1:1967 to Q4:2013. The shaded regions represent NBER-dated recessions. The Standards variable is not available from Q1:1983 to Q1:1990 because the survey dropped the C&I loans standards question during this time.

To predict stock returns, Cooper and Priestley (2009) construct a measure of the output gap, gap, which is measured as the deviation of the log of industrial production from a trend that incorporates both a linear and a quadratic component: pt ¼ a þ b t þ c t 2 þ ϵt ;

ð3Þ

where p is the log of industrial production, t is a time trend, and ϵ is an error term. The gap variable used in our work is estimated the same way using our sample period data. 2.3. Descriptive statistics Descriptive statistics (number of observations, mean, min, max, standard deviation, and autocorrelation) of the various predictor variables and stock returns are presented in Panel A of Table 1. The descriptive statistics of the standard predictor variables, as well as the stock returns, are in line with the results reported in previous work (e.g. Goyal and Welch, 2008), so the discussion of these results is skipped in the interest of space. The key variable of interest in the analysis, Standards, has an autocorrelation of 0.875 at a quarterly frequency. This autocorrelation while high, is the second lowest of all the predictor variables considered in the analysis (only DEF has a lower autocorrelation coefficient than Standards). Panel B in Table 1 presents the correlations across various predictor variables. Standards is the highly positively correlated variable with the default spread, DEF, (60%). Not surprisingly, Standards is negatively correlated with net stock issuances ntis ( 43%). The correlations across other predictor variables are consistent with earlier literature; in particular, the absolute value of the correlations between the two term structure variables, TERM and RF, and the output gap variable gap are all above 56%. Additionally, the aggregate consumption–wealth ratio variable, cay, is most highly correlated with the dividend– price ratio dp at 51%. 3. Empirical methods Following much of the existing predictability literature, the in-sample predictive ability of Standards for excess stock returns is first assessed. The following univariate regression is estimated: r t ¼ αþ β Standardst 1 þ ϵt ;

ð4Þ

122

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

Table 1 Descriptive statistics. Panel A: Descriptive Statistics of Stock Return Predictive Variables Variable Ret Excess Ret Standards DEF TERM RF dp cay ntis gap

Obs.

Mean 95 95 95 95 95 95 95 91 91 95

0.024 0.016 0.07 0.01 0.015 0.003 3.937 0.002 0.009 0.000

Stdev. 0.087 0.087 0.239 0.004 0.011 0.002 0.287 0.025 0.021 0.043

Min

Max

0.272 0.193 0.273 0.183 0.241 0.836 0.005 0.034 0.004 0.033 0.000 0.006 4.513 3.235 0.043 0.045 0.053 0.046 0.139 0.088

Autocorr. 0.052 0.051 0.875 0.799 0.917 0.935 0.935 0.939 0.907 0.904

Panel B: Correlations of Stock Return Predictive Variables Standards Standards DEF TERM RF dp cay ntis gap

1.000 0.599 0.043 0.065 0.083 0.065 0.425 0.005

DEF

1.000 0.298 0.495 0.338 0.220 0.550 0.374

TERM

RF

dp

cay

1.000 0.711 1.000 0.366 0.055 1.000 0.162 0.417 0.509 1.000 0.225 0.188 0.021 0.561 0.700 0.559 0.434 0.208

ntis

1.000 0.141

gap

1.000

Note: This table reports descriptive statistics and correlations for the excess stock return predictive variables. Ret is the log return and Excess Ret is the log excess return on the CRSP-VW index. Standards is the tightening standards measure. DEF is the BAA bond yield minus the AAA bond yield. TERM is the difference between the 10 year Treasury yield and the 1 year Treasury yield. RF is the 1 month T-bill rate. The log dividend–price ratio is denoted dp. The variable cay is the Lettau and Ludvigson (2001) consumption–wealth ratio variable. The variable ntis is the ratio of the 12 month moving sum of net issues by NYSE listed stocks divided by the total end-of-year market capitalization. The variable gap is the deviation of the log of industrial production from a trend that incorporates both a linear and a quadratic component. The sample period is Q2:1990 to Q4:2013.

where r t is the excess stock return, Standards is the net percent tightening of the C&I loan supply, and ϵ is an error term. The in-sample predictive ability of Standards is assessed via the t-statistic of the β estimate and the adjusted R2 from the excess return regression. Under the null hypothesis that Standards does not predict excess returns, β ¼ 0. Newey and West (1987) standard errors that correct for serial correlation and heteroscedasticity are reported. For robustness tests of the predictability of stock returns using Standards, the following predictor variables: DEF, TERM, RF, dp, cay, ntis, and gap, are also considered. These variables are collected in the vector Z and added to the regression that includes Standards to estimate: r t ¼ α þβ Standardst 1 þγ Z t 1 þϵt ;

ð5Þ

where γ is a vector of coefficient estimates on the variables Z t 1 , and ϵ is an error term. After controlling for these predictor variables, the in-sample predictive ability of Standards to explain future stock returns is assessed. To generate out-of-sample predictions, four test statistics designed to determine whether the Standards forecasting model has superior forecasting performance relative to a model of historical average returns are computed. The out-ofsample R2 ðR2oos Þ), which follows Fama and French (1989), is computed as R2oos ¼ 1

MSEA ; MSEN

ð6Þ

where MSEA is the mean-squared error from the forecasting model using Standards, and MSEN is the mean-squared error from the historical mean model. If the R2oos is positive, then the predictive regression has a lower average mean-squared prediction error than the historical mean return model. The second out-of-sample test statistic (RMSE) computed is the difference between the root-mean-squared prediction error using the historical average return model and the root-mean-squared prediction error using the predictive regression model: pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ΔRMSE ¼ MSEN MSEA : ð7Þ The third test statistic is an out-of-sample MSE-F test developed by McCracken (2007). It tests whether the historical mean model has a mean-squared forecasting error that is equal to that of the Standards forecasting model: MSE F ¼ ðT h þ 1Þ

MSEN MSEA ; MSEA

where T is the number of observations and h is the degree of overlap (h ¼ 1 for no overlap).

ð8Þ

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

123

The last out-of sample test is the ENC-NEW test proposed by Clark and McCracken (2001). The ENC-NEW test is used to examine whether the forecasts from the historical mean model encompass those from the Standards forecasting model: T P

ENC NEW ¼

ðϵ2 ϵt et Þ T hþ 1 t ¼ 1 t ; T MSEA

ð9Þ

where ϵt is the vector of out-of-sample errors from the historical mean model and et is the vector of out-of-sample errors from the Standards forecasting model. For both the MSE-F and ENC-NEW tests, the methodology in Clark and McCracken (2005) is followed which provides bootstrapped critical values for these tests. For the out-of-sample tests, an initial estimation window of 12 years (48 quarters) is used. This corresponds to half of our sample period. The out-of-sample tests are conducted in two ways: a recursive regression and a rolling regression. The recursive approach assumes the model is estimated with more data as the forecasting date moves forward in time. The rolling approach assumes the model is estimated with a moving window of the most recent 48 observations as the forecasting moves forward in time. 4. Stock return predictability The ability of Standards to predict stock returns is now explored. The in-sample evidence is explored first, followed by the out-of-sample evidence. Lastly, several robustness checks of the stock return predictability regressions are considered including a small-sample analysis. 4.1. In-sample evidence Table 2 reports in-sample forecasting regressions with Standards for the quarterly log excess returns on the CRSP-VW index and the S&P 500 index. In all of the regressions in Table 2, the t-statistics are reported using a Newey and West (1987) correction to account for serial correlation in the residuals.9 Panel A of Table 2 reports results from the regressions of the CRSP-VW quarterly excess returns on one lag of the Standards variable. Panel B of Table 2 reports results using the log excess returns on the S&P 500 index and the results are very similar to those on the CRSP-VW index. The excess CRSP-VW returns are strongly predictable with negative coefficients on the Standards variable at traditional significance levels.10 Also, the adjusted R2 coefficient in the univariate regression is 7%, which is very similar to the R2 's in the multivariate regressions. The negative sign implies that a tightening loan supply results in a subsequent drop in stock returns. As demonstrated in Section 5, this negative sign is consistent with Standards providing information about cash flows. Additionally, Panel A reports estimates from predictability regressions that include a variety of variables used in past predictability studies. Shiller (1981), Campbell and Shiller (1988), and Fama and French (1988) find that the dividend–price ratio has predictive power for excess stock returns. Bekaert and Hodrick (1992) find that the T-bill rate predicts returns, while Fama and French (1989) study the forecasting power of the term and the default spreads. Henkel et al. (2011) present evidence that the dividend yield and term structure variables are effective predictors almost exclusively during recessions only. These financial market variables (DEF, TRM, RF, and dp) are considered in our predictive regressions on the excess CRSP-VW return. Note that Standards still retains its forecasting power with roughly the same coefficient size and same level of statistical significance when controlling for the financial market-based variables. Lettau and Ludvigson (2001) find that the ratio of consumption to wealth, cay, predicts stock returns at a quarterly frequency which is replicated for our sample period. In the third column of Panel A, both cay and Standards jointly lead to an R2 coefficient of 8%. Both coefficients are statistically significant. Recent studies find evidence that corporate issuing activity forecasts returns. Boudoukh et al. (2007), Larrain and Yogo (2008), Robertson and Wright (2006), and Bansal and Yaron (2011) document that payout yields (as opposed to simple dividend yields) derived from dividends, repurchases, and issuances are robust predictors of excess stock returns. Moreover, Goyal and Welch (2008) find that ntis, which measures equity issuing and repurchasing (plus dividends) relative to the price level, has good in-sample performance, but a negative out-of-sample adjusted R2 . The variable ntis is added to our predictability regression to determine its in-sample performance relative to Standards. The fourth column of Panel A reports a regression of returns on both ntis and Standards. It shows that ntis is not statistically significant, but Standards is. More recently, Cooper and Priestley (2009) show that the output gap, gap, as measured by the deviation of the log of industrial production from a trend, predicts excess returns on stock indices and Treasury bonds. While the results of Cooper and Priestley (2009) are replicable over their sample period, gap does not seem to have forecasting power for excess stock 9 The log excess return results reported are nearly identical to log actual returns, raw actual returns, and raw excess returns. In addition to Newey and West (1987) standard errors, Hansen and Hodrick (1980) and Hodrick (1992) standard errors are computed. These results are similar and are available from the authors. 10 To analyze the economic significance, the in-sample tests using standardized variables are analyzed. Standards has the highest economic significance.

124

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

Table 2 Forecasting quarterly excess stock returns. Panel A: Excess returns on CRSP Standards 0.10 ( 2.45) DEF TERM RF dp

0.18 ( 2.45) 6.44 (1.48) 0.58 (0.39) 9.92 (0.86) 0.03 (0.72)

cay

0.10 ( 2.45)

0.09 ( 2.58)

0.57 (2.07)

ntis

0.13 (0.26)

0.05 (0.21) [0.09]

0.02 (3.19) [0.08]

0.02 (2.26) [0.05]

0.29 ( 1.30) 0.02 (3.34) [0.08]

0.17 ( 2.44) 4.88 (1.19) 0.19 (0.13) 7.52 (0.69) 0.03 (0.83)

0.11 ( 2.70)

0.10 ( 2.76)

0.11 ( 2.78)

gap Constant R

2

0.02 (3.59) [0.07]

Panel B: Excess returns on S&P 500 Standards 0.11 ( 2.68) DEF TERM RF dp cay

0.56 (2.12)

ntis

0.16 (0.31)

gap Constant R

2

0.11 ( 2.58)

0.02 (2.71) [0.08]

0.08 (0.36) [0.10]

0.02 (2.37) [0.09]

0.01 (1.63) [0.07]

0.24 ( 1.12) 0.02 (2.54) [0.09]

0.19 ( 2.58) 7.47 (1.39) 0.91 ( 0.44) 2.31 (0.16) 0.01 ( 0.25) 0.97 (1.56) 0.12 ( 0.15) 0.21 ( 0.53) 0.09 ( 0.31) [0.08]

0.17 ( 2.44) 5.80 (1.13) 1.44 ( 0.70) 0.43 ( 0.03) 0.01 ( 0.16) 0.93 (1.60) 0.05 ( 0.07) 0.24 ( 0.62) 0.05 ( 0.16) [0.10]

Note: This table reports estimates of OLS regressions of stock returns on one-quarter lagged predictive variables: rt ¼ α þ β Standardst 1 þ γ Z t 1 þ ϵt , where r t is the log excess return on the CRSP-VW index and the S&P 500 index. The predictive variables are all defined in Table 1. Newey–West corrected tstatistics appear in parentheses below the coefficient estimates and adjusted R2 statistics appear in square brackets. The sample period of variables is Q2:1990 to Q4:2013 except for that of cay and ntis(Q2:1990–Q4:2012).

returns over our sample period. After controlling for gap in our regression, Standards still has a significant negative coefficient. The last column of Panel A presents the in-sample forecasting regression with all the variables included. Interestingly, only Standards has a significant coefficient among all the predictor variables and the adjusted R2 is very similar to that in the univariate regression with Standards. This suggests that Standards is capturing future excess stock returns at a quarterly frequency, while other predictor variables have little predictive power of excess stock returns at this horizon. Though our sample is limited to the period after the 1990s, the in-sample predictability of Standards is noteworthy in view of the findings in Goyal and Welch (2008), which show that most predictor variables lose their in-sample forecasting power after the oil price crisis in the 1970s.

4.2. Out-of-sample evidence Two recent papers, Goyal and Welch (2008) and Campbell and Thompson (2008), examine the out-of-sample forecasting ability of predictor variables that can predict in-sample. Goyal and Welch (2008) find little evidence that most predictor variables can predict out-of-sample better than a constant, while Campbell and Thompson (2008) find that the predictors

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

125

have out-of-sample predictive power with sensible restrictions on the forecasting models. The forecasting ability of Standards is now examined in out-of-sample tests and compared to other predictor variables.11 Table 3 compares forecasts based on the historic mean model to those based on each predictor variable using the CRSPVW excess returns. The four out-of-sample tests (adjusted R2 , ΔRMSE, MSE-F, and ENC-NEW) are conducted in recursive and rolling regressions. For the tests, the initial estimation period is Q2:1990–Q1:2002. The first row of Table 3 shows that the forecasting model with Standards has superior forecasting performance relative to the historic mean model in both the recursive and the rolling regressions. The out-of-sample R2 is 5.4% in the recursive regression and 4.7% in the rolling regression. The ΔRMSE is 0.003 in the recursive regression and 0.002 in the rolling regression, which implies that the forecast errors with Standards are lower than those with the historic average return. The MSE-F test rejects the null hypothesis that the MSEs from the forecasts that use Standards is equal to those based on the historical average return. The ENC-NEW test also rejects the null hypothesis that the forecasts from the historical mean model encompass those from the Standards forecasting model. These results suggest that Standards plays a strong role as a predictor of excess stock returns since the 1990s. These results contrast with Goyal and Welch (2008) who find that variables typically used in predictability regressions have been unsuccessful out-of-sample over the last few decades. Interestingly, no economic restrictions on the forecasting model are imposed as in Campbell and Thompson (2008). The remaining rows of Table 3 report the out-of-sample test results with the other predictor variables. The variables dp and ntis show better forecasting ability than the historical average return in both the recursive and the rolling regressions. However, the adjusted R2 and ΔRMSE for Standards are over twice as large as that of dp and ntis, implying Standards has a higher forecasting power than dp and ntis.

4.3. Long-horizon forecasts Much of the existing predictability literature finds that some of the predictor variables, such as dp and cay, forecast excess stock returns in sample at long horizons better than at short horizons. With the exception of gap, most of these variables seem to predict stock returns at horizons longer than the length of a typical recession.12 In this section, it is investigated whether Standards tracks longer-term tendencies in stock markets in addition to providing shorter-term forecasts. Table 4 reports long-horizon forecasting regressions of quarterly excess returns on the CRSP-VW index. The dependent variable is the H-quarter log excess return on the CRSP-VW index, equal to r t þ 1 þ… þ r t þ H . Horizons of H ¼ 1, 2, 4, 8, and 12 quarters are used. From the top panel of Table 4, the forecasting power of Standards at horizons ranging from 1 to 12 quarters is reported. From the univariate regressions, the coefficient for Standards is hump-shaped and peaks around 4 quarters in the sample. At a 4 quarter horizon, the coefficient estimate for Standards is insignificant and the adjusted R2 is approximately 5%, so the predictive power decreases at horizons greater than 4 quarters. Here, Standards seems to better forecast future excess stock returns at a business cycle frequency as the informational content of Standards decreases at longer horizons. After including the price-based variables (DEF, TRM, RF, and dp), Standards still exhibits a hump-shaped forecasting pattern. The forecasting significance peaks at 4 quarters, thus declines at longer horizons. The addition of the price-based variables leads to statistical significance at longer horizons for Standards. Regarding the adjusted R2 coefficient, it increases with the horizon and is not hump-shaped. This is driven by the increased predictive power of the dividend–price rate, dp, with the horizon and is consistent with the predictability literature summarized for example in Koijen and van Nieuwerburgh (2013). In the bottom panel of Table 4, cay, ntis, and gap are added to the middle panel regressions. The hump-shaped forecasting pattern of Standards is robust, while the predictive power of cay and the adjusted R2 increase with the horizon, which supports the findings of Lettau and Ludvigson (2001). Here the predictive power of Standards occurs at a shorter horizon than most of the predictive variables explored in the literature.

4.4. Small sample robustness of stock return predictability The robustness of Standards as a stock return predictor is examined in a small sample analysis.13 Many predictability studies find that regression coefficients and standard errors obtained from predictive regressions with a highly persistent predictor, exhibit small sample biases (Mankiw and Shapiro, 1986; Nelson and Kim, 1993; Elliott and Stock, 1994, and Stambaugh, 1999). These biases have the potential to be severe, especially when the predictor variables are scaled by price. 11 Out-of-sample forecasting tests with the other predictor variables are also analyzed. However, results for cay and gap are not reported as the estimation periods for both variables in the out-of-sample test are relatively short. From our unreported results, cay shows better forecasting ability than the historical average return in the recursive regression. 12 Cooper and Priestley (2009) document “the average length of NBER contractions in the 1945–2001 period is 10 months.” 13 Several other robustness checks are available in our supplementary appendix including using other SLOS variables, analyzing the predictability of Standards with other credit condition variables, extending the Standards data series to use earlier data, and studying stock return predictability in the Canadian stock market. Throughout, Standards still retains its predictive power.

126

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

Table 3 Forecasting quarterly excess stock returns out-of-sample. Variable

Recursive approach 2

R oos Standards DEF TERM RF dp ntis

0.054 0.140 0.020 0.050 0.023 0.006

Rolling approach

ΔRMSE

MSE–F

ENC–NEW

R oos

0.003 0.006 0.001 0.002 0.001 0.000

2.667nn 5.759 0.931 2.237 1.086n 0.246

1.968nn 0.836 0.214 0.394 0.342 0.235

0.047 0.210 0.027 0.012 0.097 0.013

2

ΔRMSE

MSE–F

ENC–NEW

0.002 0.010 0.001 0.001 0.005 0.001

2.302nn 8.162 1.241 0.550 4.160 0.555

2.146nn 1.122 0.256 0.304 0.540 0.651

Note: This table reports the results of an out-of-sample forecast comparison of the log excess return on the CRSP-VW index. The comparisons are forecasts of excess stock returns based on a constant (unconditional forecast), and forecasts based on a constant and a 1-quarter lagged predictive variable (conditional forecast). The predictive variables are all defined in Table 1. We conduct the out-of-sample test in two ways. First, the recursive approach assumes the model is estimated with more data as the forecasting date moves forward in time. Second, the rolling approach assumes the model is 2 estimated with a moving window of the most recent 40 observations as the forecasting moves forward in time. The column R oos is the out-of-sample R2 and is defined in Eq. (6). ΔRMSE is the RMSE difference between the unconditional forecast and the conditional forecast and is defined in Eq. (7). MSE–F gives the F-test of McCracken (2007), which tests for an equal MSE of the unconditional forecast, and the conditional forecast and is defined in Eq. (8). ENC– NEW provides the Clark and McCracken (2001) encompassing test statistic and is defined in Eq. (9). Significance levels of MSE–F and ENC–NEW at 90%, 95%, and 99% levels are denoted by one, two, and three stars, respectively. The sample period is Q2:1990 to Q4:2013.

Table 4 Long horizon regression: quarterly excess stock returns. Regressors

Standards R

2

Standards DEF TERM RF dp R

2

Standards DEF TERM RF dp cay ntis gap R

2

Forecast horizon H 1

2

4

8

12

0.10 ( 2.45) [0.07]

0.17 ( 1.98) [0.09]

0.19 ( 1.41) [0.05]

0.19 ( 0.97) [0.02]

0.06 ( 0.29) [ 0.01]

0.18 ( 2.45) 6.44 (1.48) 0.58 (0.39) 9.92 (0.86) 0.03 (0.72) [0.09]

0.37 ( 2.58) 17.25 (2.25) 1.50 (0.56) 26.77 (1.22) 0.06 (0.79) [0.23]

0.46 ( 3.50) 22.51 (3.00) 2.16 (0.65) 29.47 (1.15) 0.14 (1.59) [0.26]

0.44 ( 2.54) 17.63 (1.44) 6.46 (0.91) 13.40 (0.31) 0.32 (2.29) [0.37]

0.06 ( 0.25) 9.32 ( 0.51) 8.70 (0.96) 22.20 ( 0.32) 0.54 (2.60) [0.45]

0.19 ( 2.58) 7.47 (1.39) 0.91 ( 0.44) 2.31 (0.16) 0.01 ( 0.25) 0.97 (1.56) 0.12 ( 0.15) 0.21 ( 0.53) [0.08]

0.35 ( 3.33) 19.91 (2.75) 3.31 ( 1.11) 1.42 ( 0.07) 0.01 ( 0.13) 2.06 (2.21) 0.87 (0.64) 0.28 ( 0.46) [0.31]

0.33 ( 2.57) 23.61 (2.95) 10.12 ( 2.00) 45.44 ( 1.39) 0.08 (0.65) 3.61 (3.22) 2.82 (1.21) 0.61 ( 0.87) [0.47]

0.59 ( 4.10) 26.57 (3.22) 7.55 ( 1.74) 97.69 ( 2.61) 0.06 ( 0.36) 10.90 (5.41) 2.25 ( 1.40) 0.48 ( 0.66) [0.69]

0.38 ( 2.37) 3.70 (0.39) 2.15 ( 0.45) 94.81 ( 1.94) 0.04 ( 0.22) 12.82 (6.55) 6.15 ( 3.16) 1.58 ( 1.66) [0.75]

Note: This table reports results from long-horizon regressions of quarterly log returns on lagged variables: r t þ 1 þ … þ r t þ H ¼ α þ β :Standardst þ γ Z t þ ϵt þ 1 , where H denotes the return horizon in quarters and the dependent variable is the sum of H log returns on the CRSP Value-weighted stock market index, r t þ 1 þ … þ r t þ H . The regressors are all defined in Table 1. Newey–West corrected t-statistics appear in parentheses below the coefficient estimate and adjusted R2 statistics appear in square brackets. The sample period is Q2:1990 to Q4:2013.

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

127

Table 5 Robustness: test of small sample bias. Panel A: Campbell and Yogo (2006) test Variable

Standards DEF TERM RF dp cay ntis gap

t-stat(β^ )

β^

t-test

Q-test

0.125 0.016 0.007 0.014 0.055 0.047 0.063 0.028

[ 0.204, 0.053] [ 0.131,0.075] [ 0.050,0.078] [ 0.059,0.033] [ 0.036,0.089] [ 0.036,0.085] [ 0.010,0.133] [ 0.075,0.026]

[ 0.207, 0.058] [ 0.127,0.107] [ 0.057,0.072] [ 0.058,0.034] [ 0.044,0.066] [ 0.032,0.087] [ 0.009,0.134] [ 0.073,0.027]

95% CI

99% CI

( 2.27 2.30) ( 2.30 2.28)

( 3.09 3.13) ( 3.14 3.12)

2.793 0.263 0.194 0.518 1.752 1.467 1.442 0.946

Panel B: Bootstrap stock return test Variable t-stat(β^ ) 2.45 2.68

CRSP S&P

90% CI: β

2

95% CI

99% CI

0.07 0.08

( 0.01 0.04) ( 0.01 0.04)

( 0.01 0.07) ( 0.01 0.07)

R

Note: This table reports tests of small sample bias. Panel A shows OLS estimates along with 90% Bonferroni confidence intervals following Campbell and Yogo (2006). The second and third columns report the t-statistics and the point estimate β^ from regressions of the log excess CRSP-VW return on a constant and on a one-quarter lagged predictive variable. The predictive variables are all defined in Table 1. The next two columns report the 90% Bonferroni confidence intervals for β using the t-test and Q-test, respectively. Panel B reports confidence intervals from a bootstrap procedure. 100,000 artificial time series of the size of our data set are generated under the null hypothesis of no predictability. The data generating process is r t ¼ γ þ et , Standardst ¼ μ þ ϕ Standardst 1 þ νt where r t is the log excess return on the CRSP-VW index and the S&P 500 index. The parameters in the data-generating process are set to the sample estimates for the bootstrap. OLS regressions with a Newey–West standard error correction: r t ¼ α þ β Standardst 1 þ ϵt are estimated to 2 compute the empirical distributions of the t-statistic of β^ and the R coefficient. We draw from the residuals of the system estimated under the null hypothesis. The sample period is from Q2:1990 to Q4:2013.

Though Standards is a persistent variable, its degree of persistence is not as strong as measures such as the dividend price ratio (see Table 1). Additionally, it is not a priced-based variable. However, given the length of the Standards data series, it is explored whether the in-sample results of Standards could be driven by small sample biases. To address these small sample bias problems, two robustness checks are performed. First, the small-sample tests of Campbell and Yogo (2006) are computed. Campbell and Yogo employ local-to-unity asymptotics to achieve a better approximation of the finite sample distribution when the predictor variable is persistent. Their construction of the confidence interval uses the Bonferroni method which combines a confidence interval for the largest autoregressive root of the predictor variable with confidence intervals for the predictive coefficient conditional on the largest autoregressive root. These results are presented in Panel A of Table 5. Following Campbell and Yogo, the confidence interval for β~ ¼ ðσ e =σ u Þβ instead of β are reported.14 In the fourth (fifth) column of the table, the 90% Bonferroni confidence intervals for β using the ttest (Q-test) are reported, whose the null hypothesis is β ¼ 0. Both the Bonferroni t-test and the Q-test reject the null of no predictability for only Standards. Our second method for addressing small sample bias problems is to use both a bootstrap and a Monte Carlo simulation of the predictive regressions. The data for both simulations are generated under the null hypothesis of no predictability: r t ¼ γ þet ;

ð10Þ

where γ is a constant. Also, an AR(1) specification for Standards is used: Standardst ¼ μ þ ϕStandardst 1 þ νt ;

ð11Þ

where μ and ϕ are those estimated from the Standards data. Then, artificial sequences of excess stock returns and Standards are generated by drawing randomly from the sample residuals for the bootstrap procedure or a normal distribution for the Monte Carlo simulation under the no predictability null. One hundred thousand samples equal to the length of the Standards data series are drawn. Using these samples created under either a bootstrap or Monte Carlo simulation, Eq. (4) is estimated yielding a distribution of our test statistics. Panel B of Table 5 reports the results of the bootstrap procedure for the Newey–West t-statistics and adjusted R2 coefficients of the predictive regression with Standards.15 For both the CRSP-VW and S&P 500 excess stock returns, the estimated t-statistics of Standards lies outside of the 95% confidence interval based on the empirical distribution from the bootstrap procedure. Thus, the hypothesis that Standards has no predictive power for excess stock returns can be rejected. In 14 The standard deviations σ e and σ u are computed from the residuals of the following regression model: r t ¼ α þ βxt 1 þ ut , xt ¼ γ þ ρxt 1 þ et where r t denotes the excess stock return in period t and xt denotes the predictor variable in period t. 15 The results of Monte Carlo simulation are nearly identical to those of the bootstrap procedure.

128

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

addition, the results show that the estimated adjusted R2 coefficient is outside of the 99% confidence intervals for the bootstrap adjusted R2 coefficients. Therefore, the predictability of Standards is robust to small sample biases.

5. Channel of predictability The results so far show that excess stock returns are strongly predictable with a negative coefficient on Standards and the predictive power decreasing with the time horizon. The negative coefficient suggests that tightening Standards predicts a subsequent drop in stock returns. Asset pricing models with time-varying discount rates, such as the Campbell and Cochrane (1999) model, would typically imply investors require higher expected returns in bad times. However, a cash flow component could be important too as captured for example in the Bansal and Yaron (2004) model and empirically studied in Bansal and Yaron (2011) and Larrain and Yogo (2008) using measures of total cash flows. The source of the predictability of Standards is now explored to better understand the negative coefficient. As a first step, the predictability of Standards is examined controlling for other cash flow expectation variables.16 From the Survey of Professional Forecasters, the average expected growth rate of GDP over the next four quarters and the average expected CPI inflation rate over the next four quarters are used. Table 6 reports the in-sample predictive regressions. Standards is statistically significant with a negative coefficient in all specifications. This evidence is consistent with either Standards accounting for cash flow expectation information not captured by the other variables or Standards being related to discount rate news. Most existing attempts to disentangle variation in future cash flows versus variation in expected discount rates start with the Campbell and Shiller (1988) decomposition of dividend yields dpt

1 X

ρj 1 r t þ j

j¼1

1 X

ρj 1 Δdt þ j ;

ð12Þ

j¼1

where dp is log dividend–price ratio, ρ 0:961=4 is a constant of approximation, r is stock returns, and Δd is growth rates of dividends. Simple forecasting regressions with returns, dividend growth, dividend yield, or Standards as dependent variables and lagged dividend yield and Standards as independent variables are used. There is a growing literature that extends the predictability regressions implied by inverting the identity in Eq. (12) with additional regressors such as Cochrane (2011), Bollerslev et al. (2015), and Maio and Santa-Clara (2015). As Cochrane (2008, 2011) advocates, this is a particularly clean way to understand if additional regressors, Standards in our case, provides information about cash flows or discount rate variation. The models are estimated using OLS and robust covariance matrix estimators. To summarize the impact of dividend yields and Standards on returns and dividend growth, the first-order VAR and multivariate regression from Cochrane (2011) are used: 3 " # " # 2 dp ðiÞ ðiÞ εt þ 1 dpt þ 1 dpt ¼Φ þ 4 Standards 5 Standardst þ 1 εt þ 1 Standardst 2 4

Þ r ðtiþ 1 ðiÞ

▵dt þ 1

3 5¼B

"

ðiÞ

dpt

Standardst

#

" þ

εrt þ 1

#

;

ð13Þ

ε▵d t þ1

where Φ and B are 2 2 coefficient matrices. The long-horizon return forecasts are then 3 " lr # 2 P1 " # j1 ðiÞ rt þ j rt j¼1ρ dpt 1 4 5 ¼ BðI ρΦÞ P1 ; lr j1 ▵dt þ j ▵dt Standardst j¼1ρ

ð14Þ

for ρ ¼ 0:961=4 and with the long-run regression coefficients defined as BðI ρΦÞ 1 . Table 7 presents the forecasting results. To match with our previous results, the quarterly log excess returns on the CRSPVW index as the stock return and the quarterly growth rates of real dividends of the S&P 500 index from Robert Schiller's website as the dividend growth rates are used. At a quarterly frequency, Standards predicts both returns and dividend growth rates with respective t-statistics of 2.26 and 5.48. From the point estimates, a one standard deviation increase in Standards implies a 2.6% drop in returns and a 1.3% drop in dividend growth. This result in particular highlights that Standards is useful in predicting cash flows. This result joins a growing literature emphasizing the importance of cash flow predictability. See for example Lettau and Ludvigson (2005), Garrett and Priestley (2012), Bollerslev et al. (2015), and Maio and Santa-Clara (2015). Inference in long-horizon regressions with relatively short sample periods is problematic. 16 The quarterly growth rate of analyst earnings forecasts over the next year for the S&P 500 index from I/B/E/S is also used as a cash flow expectation variable. I/B/E/S provides the earnings forecasts for the S&P 500 index by Q4:2009. Controlling for the quarterly growth rate of earnings forecasts for the S&P 500 index, Standards is the best predictor of future excess stock returns.

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

129

Table 6 Forecasting quarterly excess stock returns with cash flow expectation variables. Excess stock returns on the CRSP-VW index Standards GDP4Qavg CPI4Qavg

0.16 ( 3.65) 0.05 ( 2.64) 0.03 (1.62)

0.18 ( 2.34) 0.04 ( 2.18) 0.03 ( 0.64) 1.52 (0.30) 1.85 (1.24) 24.36 (1.81) 0.08 (1.18)

DEF TERM RF dp

0.16 ( 3.72) 0.04 ( 2.36) 0.00 ( 0.08)

cay

0.15 ( 3.94) 0.05 ( 2.71) 0.03 (1.37)

0.16 ( 3.59) 0.05 ( 2.74) 0.03 (1.68)

1.07 (1.79)

ntis

0.39 (0.73) 0.24 ( 1.56) 0.17 (3.36) [0.14]

gap Constant R

2

0.18 (2.97) [0.13]

0.50 (1.31) [0.14]

0.24 (3.71) [0.15]

0.19 (2.95) [0.12]

0.18 ( 2.17) 0.02 ( 1.00) 0.09 ( 1.17) 2.10 (0.32) 0.54 (0.28) 20.97 (1.62) 0.10 (1.07) 1.44 (1.60) 0.03 ( 0.05) 0.21 (0.61) 0.67 (1.23) [0.13]

Note: This table reports estimates of OLS regressions of excess stock returns on one-quarter lagged predictive variables: r t ¼ α þ β Standardst 1 þ γ Z t 1 þ ϵt , where r t is the log excess stock return on the CRSP-VW index. The variable GDP4Qavg is the average expected growth rate of GDP over the next four quarters from the Survey of Professional Forecasters. The variable CPI4Qavg is the average expected CPI inflation rate over the next four quarters from the Survey of Professional Forecasters. Newey–West corrected t-statistics appear in parentheses below the coefficient estimates and adjusted R2 statistics are in square brackets. The sample period is Q2:1990–Q4:2013.

Table 7 Forecasting quarterly excess stock returns and dividend growth rates. Regressors

rt þ 1 Δdt þ 1 dpt þ 1 Standardst þ 1 j1 rt þ j r lrt ¼ Σ 1 j ¼ 1ρ lr

j1 Δdt þ j Δdt ¼ Σ 1 j ¼ 1ρ

Coefficients

Other statistics

dpt

Standardst

R2

σ½Et ðyt þ 1 Þ%

σ½Et ðyt þ 1 Þ ½Et ðyt þ 1 Þ

0.017 (1.88) 0.002 ( 1.06) 0.932 (28.67) 0.074 ( 1.58) 0.320

0.026 ( 2.26) 0.013 ( 5.48) 0.041 (0.98) 0.884 (16.67) 0.100

0.12

3.01

1.91

0.34

1.34

2.33

0.050

0.087

0.91 0.81 0.331 0.099

Note: This table reports results of a multi-variable VAR for quarterly excess returns and dividend growth rates. r is the log excess return on the CRSP-VW index and Δd is the quarterly growth rates of real dividends of the S&P 500 index from Robert Schiller's website. dp is the log dividend–price ratio and Standards is the tightening standards measure. Each regression includes a constant. Both dp and Standards are rescaled (σðdpÞ ¼ 1 and σðStandardsÞ ¼ 1). The t-statistics appears in parentheses below the coefficient estimate. σ½Et ðyt þ 1 Þ% is the volatility of the fitted value from the forecast. The sample period is Q2:1990–Q4:2013.

Nonetheless, the point estimates of the long-run regression coefficients at the bottom of Table 7 do continue to show that Standards continues to provide information about future cash flows. The importance of Standards for quarterly return and cash flow predictability based on the point estimates from the predictive regression and the VAR are shown in Fig. 2. The left panel provides results for the expected return, while the right panel provides results for expected dividend growth. Overall, the figure demonstrates that Standards is useful for forecasting at a quarterly horizon as it provides additional information beyond the dividend price ratio. Standards in particular is quite useful in capturing the variation in dividend growth that is not well captured by the slow-moving dividend price ratio.17

17

Additional cash flow predictability evidence is summarized in our supplementary appendix.

130

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

20

8

15

6

10 4

Percentage

Percentage

5 0 −5 −10

2 0 −2

−15 −4 −20 −25

dp only dp and Standards rt+1

−30 1990

1995

−6

2000

2005

Year

2010

2015

−8 1990

dp only dp and Standards Δ d t+1

1995

2000

2005

2010

2015

Year

Fig. 2. Forecast of quarterly returns and dividend growth using standards. Note: The forecasts are fitted values of regressions of returns and dividend growth on dividend yield and Standards: r is the log excess return on the CRSP-VW index and Δd is the quarterly growth rates of real dividends of the S& P 500 index from Robert Schiller's website. dp is the log dividend–price ratio and Standards is the tightening standards measure. (a) Stock returns. (b) Dividend growth.

These results provide evidence that the predictability of Standards for excess stock returns is mainly through the cash flow news channel. This is consistent with the view that the unwillingness of banks to lend funds can cause firms to forego some positive NPV projects, leading to a decrease in firm value (Chava and Purnanandam, 2011; Chava and Hsu, 2014). Given the market does not immediately impound this information into the stock price, our results might be interpreted as siding with a behavioral interpretation in the market efficiency/behavioral finance debate. Alternatively, if the concept of rational expectations is expanded as in Condie and Ganguli (2011) to incorporate ambiguity aversion, rational investors who exhibit ambiguity aversion may not fully react to news surprises. This leads to information inertia, even though investors face no trading frictions as discussed in Caskey (2009), Illeditsch (2011), and Condie et al. (2012) for example. Empirically testing this ambiguity-based information inertia hypothesis could be especially fruitful in future work. 6. Conclusion Evidence is provided that Standards, a measure of aggregate supply-based credit conditions, is a strong predictor of U.S. stock returns. Given that Standards has been shown to predict aggregate macroeconomic variables, our results provide a direct link between a macroeconomic supply variable and the predictability of asset returns. Additionally, Standards is not derived from financial market prices mitigating concerns that the source of its predictive power is from capturing mispricing in financial markets. Standards captures predictability at a business cycle frequency. Our evidence suggests that Standards' predictability is primarily driven by a cash flow news channel.

Acknowledgments We thank Urban Jermann (editor), Joao Gomes (associate editor), an anonymous referee, Dong-Hyun Ahn, Greg Bauer, Frederico Belo, David Chapman, Zhanhui Chen, Robert Faff, Burton Hollifield, Shane Johnson, Nishad Kapadia, Hwagyun Kim, Inmoo Lee, J. Spencer Martin, Bumjean Sohn and seminar participants at Texas A&M University, Peking University, the 2010 European Finance Association Meeting, the McGill Risk Management Conference, the 2010 UBC Summer Finance Conference, and the 2011 University of Melbourne Finance Down Under Conference for helpful input. We thank Kimberley Berdy for editorial assistance. All errors are our own. Appendix A. Supplementary data Supplementary data associated with this article can be found in the online version at http://dx.doi.org/10.1016/j.jmoneco. 2015.06.004. References Adrian, T., Moench, E., Shin, H.S., 2010. Financial intermediation, asset prices, and macroeconomic dynamics. Fed. Reserv. Bank N. Y. Staff Rep. 422, 1–40. Ang, A., Bekaert, G., Wei, M., 2007. Do macro variables, asset markets, or surveys forecast inflation better? J. Monet. Econ. 54 (May (4)), 1163–1212.

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

131

Asea, P.K., Blomberg, B., 1998. Lending cycles. J. Econom. 83 (12), 89–128. Bansal, R., Yaron, A., 2004. Risks for the long run: a potential resolution of asset pricing puzzles. J. Financ. 59 (August (4)), 1481–1509. Bansal, R., Yaron, A., 2011. The asset pricing-macro nexus and return-cash flow predictability. Duke University (Unpublished manuscript). Bassett, W.F., Chosak, M.B., Driscoll, J.C., Zakrajšek, E., 2014. Changes in bank lending standards and the macroeconomy. J. Monet. Econ. 62, 23–40. Bekaert, G., Hodrick, R.J., 1992. Characterizing predictable components in excess returns on equity and foreign exchange markets. J. Financ. 47 (June (2)), 467–509. Berlin, M., 2009. Bank credit standards. Bus. Rev. Q2, 1–10. Bernanke, B.S., 2007. The financial accelerator and the credit channel. Speech at The Credit Channel of Monetary Policy in the Twenty-First Century Conference, Federal Reserve Bank of Atlanta, June. Bernanke, B.S., Blinder, A.S., 1988. Credit, money, and aggregate demand. Am. Econ. Rev. 78 (May (2)), 435–439. Bernanke, B.S., Gertler, M., 1989. Agency costs, net worth, and business fluctuations. Am. Econ. Rev. 79 (March (1)), 14–31. Bernanke, B.S., Gertler, M., 1995. Inside the black box: the credit channel of monetary policy transmission. J. Econ. Perspect. 9 (4), 27–48. Bernanke, B.S., Kuttner, K.N., 2005. What explains the stock market's reaction to federal reserve policy? J. Financ. 60 (3), 1221–1257. Bollerslev, T., Xu, L., Zhou, H., 2015. Stock return and cash flow predictability: the role of volatility risk. J. Econom 187 (2), 458–471. Boudoukh, J., Michaely, R., Richardson, M., Roberts, M.R., 2007. On the importance of measuring payout yield: implications for empirical asset pricing. J. Financ. 62 (April (39)), 877–915. Brady, T.F., 1985. The role of the prime rate in pricing of business loans by commercial banks, 1977–84. Staff Studies 146, Board of Governors of the U.S. Federal Reserve System. Campbell, J.Y., 1987. Stock returns and the term structure. J. Financ. Econ. 18 (June (2)), 373–399. Campbell, J.Y., Cochrane, J.H., 1999. By force of habit: a consumption-based explanation of aggregate stock market behavior. J. Polit. Econ. 107, 205–251. Campbell, J.Y., Shiller, R.J., 1988. The dividend–price ratio and expectations of future dividends and discount factors. Rev. Financ. Stud. 1 (3), 195–228. Campbell, J.Y., Thompson, S.B., 2008. Predicting excess stock returns out of sample: can anything beat the historical average? Rev. Financ. Stud. 21 (July (4)), 1509–1531. Campbell, J.Y., Yogo, M., July 2006. Efficient tests of stock return predictability. J. Financ. Econ. 81 (1), 27–60. Campbell, S.D., Diebold, F.X., 2009. Stock returns and expected business conditions: half a century of direct evidence. J. Bus. Econ. Stat. 27 (2), 266–278. Caskey, J.A., 2009. Information in equity markets with ambiguity-averse investors. Rev. Financ. Stud. 22 (9), 3595–3627. Chava, S., Hsu, A., 2014. Financial Constraints, Monetary Policy Shocks, and the Cross-Section of Equity Returns. Georgia Institute of Technology (Unpublished manuscript). Chava, S., Purnanandam, A., 2011. The effect of banking crisis on bank-dependent borrowers. J. Financ. Econ. 99, 116–135. Clark, T.E., McCracken, M.W., 2001. Tests of equal forecast accuracy and encompassing for nested models. J. Econom. 105 (November (1)), 85–110. Clark, T.E., McCracken, M.W., 2005. Evaluating direct multistep forecasts. Econom. Rev. 24, 369–404. Cochrane, J.H., 2008. The dog that did not bark: a defense of return predictability. Rev. Financ. Stud. 21 (July (4)), 1533–1575. Cochrane, J.H., 2011. Presidential address: discount rates. J. Financ. 66 (August (4)), 1047–1108. Condie, S., Ganguli, J., Illeditsch, P.K., 2012. Information Inertia. University of Pennsylvania (Unpublished manuscript). Condie, S., Ganguli, J.V., 2011. Ambiguity and rational expectations equilibria. Rev. Econ. Stud. 78 (3), 821–845. Cooper, I., Priestley, R., July 2009. Time-varying risk premiums and the output gap. Rev. Financ. Stud. 22 (7), 2601–2633. Elliott, G., Stock, J.H., 1994. Inference in time series regression when the order of integration of a regressor is unknown. Econom. Theory August (10) (3-4), 672–700. Engel, C., Mark, N.C., West, K.D., 2007. Exchange rate models are not as bad as you think, NBER Working Papers. Engel, C., Rogers, J.H., 2006. The U.S. current account deficit and the expected share of world output. J. Monet. Econ. 53 (July (5)), 1063–1093. Engel, C., Rogers, J.H., 2009. Expected consumption growth from cross-country surveys: implications for assessing international capital markets. IMF Staff Pap. 56 (August (3)), 543–573. Fama, E.F., French, K.R., 1988. Dividend yields and expected stock returns. J. Financ. Econ. 22 (October (1)), 3–25. Fama, E.F., French, K.R., 1989. Business conditions and expected returns on stocks and bonds. J. Financ. Econ. 25 (November (1)), 23–49. Fazzari, S.M., Hubbard, R.G., Petersen, B.C., 1988. Financing constraints and corporate investment. Brook. Pap. Econ. Activity 19 (1), 141–206. Garrett, I., Priestley, R., 2012. Dividend growth, cash flow, and discount rate news. J. Financ. Quant. Anal. 47, 1003–1028. Ghysels, E., Wright, J.H., 2009. Forecasting professional forecasters. J. Bus. Econ. Stat. 27, 504–516. Gorton, G., He, P., 2008. Bank credit cycles. Rev. Econ. Stud. 75 (October (4)), 1181–1214. Goyal, A., Welch, I., 2008. A comprehensive look at the empirical performance of equity premium prediction. Rev. Financ. Stud. 21 (4), 1455–1508. Hansen, L.P., Hodrick, R.J., 1980. Forward exchange rates as optimal predictors of future spot rates: an econometric analysis. J. Polit. Econ. 88 (October (5)), 829–853. Henkel, S.J., Martin, J.S., Nardari, F., 2011. Time-varying short-horizon predictability. J. Financ. Econ. 99, 560–580. Hodrick, R.J., 1992. Dividend yields and expected stock returns: alternative procedures for inference and measurement. Rev. Financ. Stud. 5 (3), 357–386. Holmstrom, B., Tirole, J., 1997. Financial intermediation, loanable funds, and the real sector. Q. J. Econ. 112 (August (3)), 663–691. Illeditsch, P.K., 2011. Ambiguous information, portfolio inertia, and excess volatility. J. Financ. 66 (6), 2213–2247. Jensen, G.R., Johnson, R.R., Mercer, J.M., 1996. Business conditions, monetary policy, and expected security returns. J. Financ. Econ. 40 (February (2)), 213–237. Keim, D.B., Stambaugh, R.F., 1986. Predicting returns in the stock and bond markets. J. Financ. Econ. 17 (December (2)), 357–390. Koijen, R.S., van Nieuwerburgh, S., 2013. Predictability of returns and cash flows. Annu. Rev. Financ. Econ. 3 (December), 467–491. Larrain, B., Yogo, M., 2008. Does firm value move too much to be justified by subsequent changes in cash flow. J. Financ. Econ. 87 (January (1)), 200–226. Leary, M.T., 2009. Bank loan supply, lender choice, and corporate capital structure. J. Financ. 64 (June), 1143–1185. Lettau, M., Ludvigson, S., 2001. Consumption, aggregate wealth, and expected stock returns. J. Financ. 56 (June (3)), 815–849. Lettau, M., Ludvigson, S.C., 2005. Expected returns and expected dividend growth. J. Financ. Econ. 76, 583–626. Lown, C.S., Morgan, D.P., 2002. Credit effects in the monetary mechanism. Fed. Reserv. Bank N. Y. Econ. Policy Rev. (May), 217–235. Lown, C.S., Morgan, D.P., 2006. The credit cycle and the business cycle: new findings using the loan officer opinion survey. J. Money Credit Bank. 38 (September (6)), 1575–1597. Lown, C.S., Morgan, D.P., Rohatgi, S., 2000. Listening to loan officers: the impact of commercial credit standards on lending and output. Fed. Reserv. Bank N. Y. Econ. Policy Rev. (July), 1–16. Maio, P., Santa-Clara, P., 2015. Dividend yields, dividend growth, and return predictability in the cross-section of stocks. J. Financ. Quant. Anal. 50 (1-2), 33–60. Mankiw, N.G., Shapiro, M.D., 1986. Do we reject too often?: small sample properties of tests of rational expectations models Econ. Lett. 20 (2), 139–145. McCracken, M.W., 2007. Asymptotics for out of sample tests of granger causality. J. Econom. 140 (October (2)), 719–752. Nelson, C.R., Kim, M.J., 1993. Predictable stock returns: the role of small sample bias. J. Financ. 48, 641–661. Newey, W.K., West, K.D., May 1987. A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix. Econometrica 55 (3), 703–708. Patelis, A.D., 1997. Stock return predictability and the role of monetary policy. J. Financ. 52, 1951–1972. Robertson, D., Wright, S., 2006. Dividends, total cash flow to shareholders, and predictive return regressions. Rev. Econ. Stat. 88 (March (1)), 91–99. Schreft, S.L., Owens, R.E., 1991. Survey evidence of tighter credit conditions: what does it mean? Fed. Reserv. Bank Richmond Econ. Rev., March/April 29–34. Schwert, G.W., 1990. Stock returns and real activity: a century of evidence. J. Financ. 45, 1237–1257. Shiller, R.J., 1981. Do stock prices move too much to be justified by subsequent changes in dividends? Am. Econ. Rev. 71 (June (3)), 421–436.

132

S. Chava et al. / Journal of Monetary Economics 74 (2015) 117–132

Stambaugh, R.F., 1999. Predictive regressions. J. Financ. Econ. 54 (December (3)), 375–421. Stein, J.C., 1998. An adverse-selection model of bank asset and liability management with implications for the transmission of monetary policy. Rand J. Econ. 29 (Autumn (3)), 466–486. Thorbecke, W., 1997. On stock market returns and monetary policy. J. Financ. 52, 635–654.