Credit Cycles, Expectations, and Corporate Investment

Abstract

We provide a systematic empirical assessment of the Minsky hypothesis that business fluctuations stem from irrational swings in expectations. Using predictable firm-level forecast errors, we build an aggregate index of irrational expectations and use it to provide three sets of results. First, we show that our index predicts aggregate credit cycles. Next, we show that these predictable credit cycles drive cycles in firm-level debt issuance and investment and similar cycles between financially constrained and unconstrained firms, as Minsky predicts. Finally, we show more pronounced cycles in firm-level financing and investment for firms with ex ante more optimistic expectations. (JEL G31, G32, G40, E32, E44)

In the aftermath of the credit boom of 2006–2007 and subsequent financial crisis and Great Recession, empirical work documents a systematic link between cycles in financial markets and the real economy. Credit expansions predict lower gross domestic product (GDP) growth (Schularick and Taylor 2012; López-Salido, Stein, and Zakrajšek 2017) and lower returns to bank stocks (Baron and Xiong 2017) and corporate bonds (Greenwood and Hanson 2013). Minsky (1977) hypothesizes that boom-bust patterns in credit and output growth reflect swings in nonrational expectations by economic agents. During bull markets, runs of good news regarding firm fundamentals result in overly optimistic investors and managers. This excessive optimism drives an excessive decrease in the cost of capital, inducing managers to borrow and invest too much and increasing the overall fragility of the economic system. Subsequently, systematic disappointment triggers abrupt reversals in credit and investment. Bordalo, Gennaioli, and Shleifer (2018) propose a formal model of this mechanism under diagnostic expectations and present supportive evidence that the expectations of credit analysts about future credit conditions are extrapolative. To assess the role of expectations in business cycles, however, much remains to be done.

In this paper, we make progress on two fronts. First, we construct an aggregate index of irrational expectations about long-term earnings growth using predictable firm-level forecast errors. We show that our index has strong explanatory power for aggregate cycles in credit quality, investment, and debt issuance. These findings provide support for Minsky’s hypothesis that overextrapolation of shocks to firm-level fundamentals is a driver of credit cycles. Second, we analyze whether irrational expectations affect capital markets through supply of capital by investors, demand for capital by firms, or both. Understanding the mechanism through which expectations affect capital markets is important both from a positive and a normative standpoint. To separate the role of supply and demand, we compare financially constrained and unconstrained firms and find that our measure of irrational expectations predicts similar cycles in investment and debt issuance between these firms. These results are consistent with both supply and demand effects; that is, our results are consistent with cycles in investment and financing driven by both irrational markets and irrational managers, as Minsky (1977) requires. Our final results show heterogeneity in investment and credit cycles across firms; firms with greater firm-level overextrapolation experience a greater initial increase and a greater subsequent reversal in investment and debt issuance. To test Minsky (1977), it is crucial to combine aggregate data on expectations and credit markets with micro data on firm-level expectations, forecast errors, and financial constraints, as we do in this paper.

In the first part of the paper, we construct a measure of firm-level biased expectations of future fundamentals. Minsky (1977) describes how the interplay of firm-level debt issuance, cash flows, and long-term investment projects, when combined with overreaction to shocks to firm fundamentals, creates fragility within the economic system. We measure expectations of fundamentals using long-term earnings expectations data from IBES, in line with Bordalo et al. (2019) and consistent with the importance of cash flows to the theory Minsky (1977) presents. In particular, we capture the expectations of firms’ future fundamentals using the forecasts of earnings 3 to 5 years ahead. Revisions of these forecasts and forecast errors measure the extent to which these expectations are biased and whether these expectations over- or underreact to news.

The creation of our index has two steps. First, to capture the predictable component of firm-level forecast errors, we calculate predicted forecast error from a regression of forecast error on the previous two forecast revisions, as in Bordalo et al. (2019). Second, to aggregate our index across the economy, we weight this firm-level measure of biased expectations by the contemporaneous firm-level debt issuance. Intuitively, our index is higher when firms for which analysts are more optimistic access credit markets. If Minsky (1977) is correct, such biased expectations of fundamentals should predict large investment and debt issuance; in turn, large investment and debt issuance by these firms during booms should sow the seeds of subsequent downturn and, thus, should predict subsequent declines in investment and debt issuance.

We show that an increase in our aggregate measure of biased expectations in year t strongly predicts both short-term credit booms in year t + 1 and long-term credit reversals in years t + 3 and t + 4. Aggregate cycles in irrational expectations formation and revision explain a large portion of the time-series variation of credit market sentiment (CMS); the correlation between our index and CMS is 0.75.¹ To the best of our knowledge, this is the first evidence directly showing that credit market sentiment reflects biased expectations about the fundamentals of firms accessing credit markets.

In the second part of the paper, we show that our index drives cycles in both firm-level corporate investment and debt issuance. We measure corporate investment using the methodology of Peters and Taylor (2017), allowing us to measure investment in both physical capital and intangible capital. We measure debt issuance as change in long-term debt. We estimate the long-run effects of our index on corporate investment and debt issuance by estimating panel impulse response functions using the local projection method from Jordá (2005). Our results are strong for both investment and debt issuance. In the short run, a one-standard-deviation increase in our index is associated with a 4.6% increase (relative to its mean) in total investment the following year, which represents the weighted average of a 3.6% increase in investment in physical capital and a 4.7% increase in investment in intangible capital. In the longer run, the same one-standard-deviation increase in our index in year t is associated with a 2.7% and 2.5% decrease in total investment in years t + 4 and t + 5, which is driven by a large and persistent decrease in physical investment. We find stronger cycles in debt issuance.

The fact that both corporate financing and investment respond to optimism-driven credit cycles indicates that supply effects are clearly at play in the data. To ask whether there are also demand-side effects, we allow for possible differential effects of our index on the investment and debt issuance of financially constrained and unconstrained firms. If excess optimism influences only the supply of capital by credit market investors, and managers are rational, then one would expect higher optimism to reduce credit spreads and increase debt issuance for all firms. In this case, however, only financially constrained firms would increase their investment. Unconstrained firms run by rational managers would be already at the optimum level of investment, so they would just use the proceeds of debt issuance to repurchase shares. In doing so, they would act as cross-market arbitrageurs (Ma 2019), ultimately reducing financial market instability. Alternatively, if excess optimism influences only the demand for capital by corporate managers and investors are rational, then both constrained and unconstrained firms would want to invest more, but capital supply would remain stable. As a result, one would expect higher optimism to come with higher debt issuance and higher investment by all firms. In this case, however, rational capital supply would imply an increase in credit spreads, thereby eventually disciplining the excess demand of irrational managers.

If excess optimism influences both the demand and the supply of capital, then debt issuance and investment should increase across the board for both constrained and unconstrained firms. Moreover, if the supply-side effect is sufficiently strong, we expect to see a decrease in credit spreads and a deterioration in credit quality. In this case, irrational markets do not discipline the excess demand of irrational managers, thereby planting the seeds of future predictable reversals in investment, financing, and economic growth, as in Minsky (1977).

To compare investment and credit cycles between financially constrained and unconstrained firms, we measure financial constraints using three definitions: the Hadlock and Pierce (2010) index, the Whited and Wu (2006) index, and the Kaplan and Zingales (1997) index. Using all three measures, our estimates of the boom-bust cycles in corporate investment and debt issuance are virtually identical for financially constrained and unconstrained firms. These findings imply that our estimates using our index do isolate the real effects of swings in irrational expectations on corporate investment and financing. Put differently, financial constraints do not have incremental explanatory power on corporate investment and financing cycles once irrational credit market cycles are considered. The fact that optimism-driven investment and financing cycles are very similar for constrained and unconstrained firms indicates that demand effects are also at play in the data.² We rationalize our results in a parsimonious dynamic Q-theory model of investment in which both managers and investors hold diagnostic expectations.

We demonstrate that the strong effect of our index on corporate investment and financing is robust to controlling for a large set of aggregate proxies for first- and second-moment shocks to the economy. One advantage of our panel analysis is that, unlike purely aggregate analyses, in addition to economy-wide indicators, we can also directly control for many firm-level determinants of investment and financing activity. This allows us to show that the strong effect of predictable firm-level forecast error on corporate investment and financing is not due to time-invariant firm-level heterogeneity, time-varying firm-level default risk, or other firm-level proxies for investment opportunities and balance sheet strength. We also consider whether our results are related to the cross-market arbitrage effects in Ma (2019) by examining whether our results depend on firm size. We find that only firms in the top size decile rebalance their capital structure, consistent with the results in Ma (2019). However, we show that even firms in the top size decile experience significant boom-bust cycles in corporate investment.

We contribute to the literature studying cycles in credit markets and economic activity (Greenwood and Hanson 2013; López-Salido, Stein, and Zakrajšek 2017; Bordalo, Gennaioli, and Shleifer 2018) by emphasizing the role of irrational expectations about fundamentals in generating and amplifying both credit cycles and business cycles. Greenwood and Hanson (2013) examine aggregate cycles and show countercyclical spreads, which suggest that irrational capital supply is at play, but they do not examine data on expectations and do not consider demand effects. López-Salido, Stein, and Zakrajšek (2017) find that frothy credit market conditions predict low GDP growth and look at the investment of firms for different categories of credit ratings, again suggesting that mechanisms of irrational capital supply are at play, but they do not examine data on expectations and thus cannot isolate the effects of irrational swings in expectations from the rational updating of beliefs about future credit conditions.

Conversely, Gennaioli, Ma, and Shleifer (2016) examine data on managerial expectations and corporate investment in the cross-section but do not examine their dynamics over the cycle or study how either expectations or investment vary with credit market sentiment. Bordalo et al. (2020b) build a Real Business Cycle (RBC) model with heterogeneous firms and financial frictions to show how firm-level cycles based on irrational expectations aggregate into credit and macroeconomic cycles. They show how a calibrated RBC model with diagnostic expectations can produce large business fluctuations even with small TFP shocks. Barrero (2022) shows using a dynamic general equilibrium model that managerial overextrapolation reduces firm value. In Barrero (2022), managers overreact to transitory profitability shocks, generate excess volatility, and overspend on adjustment costs. Relative to these papers, we investigate empirically the connection between expectations of fundamentals and aggregate credit and investment cycles, emphasizing the importance of the distinction between financially constrained and unconstrained firms for assessing the mechanism.

Other papers investigate different mechanisms. Mian, Sufi, and Verner (2017) study household borrowing and, consistent with our results, find that the rise in aggregate household debt predicts negative errors in GDP growth forecasts by IMF and OECD, but do not examine bond spreads, firm-level investment, or debt issuance, nor do they systematically isolate the effect of biased expectations from those of rational updates of information in credit cycles. Fahlenbrach, Prilmeier, and Stulz (2018) study intermediaries’ balance sheets and, consistent with our results, find that there is excess optimism for banks expanding credit more rapidly, but they do not examine credit spreads, firm-level investment, or debt issuance. Real investment is a key element of business cycles, and the bond market is a key financing source for corporate investment in the United States, which implies that our results on the corporate borrowing channel complement those of Mian, Sufi, and Verner (2017) on the household borrowing channel. Krishnamurthy and Muir (2017) show that changes in credit spreads and the extent of credit growth predict the severity of crises, suggesting expectations play a role, but they do not examine data on expectations. Greenwood and Hanson (2015) study the dry bulk shipping industry and document cycles in earnings and investment. Specifically, they find that high current ship earnings are associated with high used ship prices and heightened industry investment in new ships, and forecast low future returns, but they do not examine data on expectations. Simsek (2013) provides theoretical analyses showing how belief disagreement between lenders and borrowers can affect aggregate leverage and asset prices asymmetrically in good states relative to bad ones. A different literature, starting with Bernanke and Gertler (1989) and Kiyotaki and Moore (1997) emphasizes financial frictions as a transmission mechanism of productivity shocks to economic activity. These papers help organize our analysis about the economic mechanisms at play in our data, and we will discuss them in more detail in Section 5.

1 Data

This section describes our data. In Section 1.1, we describe data on expectations; in Section 1.2, we describe firm-level data; in Section 1.3, we describe our measure of credit market sentiment; and in Section 1.4, we describe macro-level data.

1.1 Expectations data

To examine overextrapolation of shocks to firm fundamentals, we use data on both predicted and realized firm earnings. Our data on analysts’ expectations is from IBES; we use analyst-level forecasts for earnings per share (EPS) from the IBES Unadjusted Detail History File. We focus on analysts’ forecasts of firms’ long-run EPS growth rate, or “LTG,” which is defined by IBES as “|$ \ldots $|expected annual increase in operating earnings over the company’s next full business cycle,” or, approximately, the next 3 to 5 years. This length of time is consistent with the timing of boom-bust cycles Minsky (1977) describes. To calculate consensus LTG forecasts, we use medians at the firm-month level. Our LTG data run from 1983 through December 2017. Realized EPS figures are obtained from the IBES Unadjusted Detail Actuals file. To adjust IBES EPS figures, we use CRSP daily data on stock splits (cfacshr).

In Section 2, we will discuss how we use these expectations data to obtain forecast errors and forecast revisions. For now, we point out that, because our CMS data and our firm-level balance sheet data are at an annual frequency, we convert IBES data from monthly to annual frequency. We do so by taking a simple average over the estimates in each calendar year, so we use annualized consensus forecasts to calculate both forecast errors and forecast revisions at a series of horizons.

1.2 Firm-level data

We study a large, unbalanced panel of Compustat firms at an annual frequency that spans 1990 through 2017.³ The panel excludes financial firms (ie, firms with a one-digit SIC code of six), utilities (ie, firms with two-digit SIC code of 49), firms not incorporated in the United States, and firm-years with negative assets, sales, or book equity. Otherwise, our sample includes all observations with data on investment, financing, and other firm-level controls. Further details on our empirical setup can be found in Section 3.2. All variables are winsorized at the 1^st and 99^th percentiles. Table D.2 in the Internet Appendix presents summary statistics of firm-level variables. We have over 70,000 firm-year observations and over 8,286 unique firms, with a median of 2,543 firms per year.⁴

1.3 Measuring credit-market sentiment

Throughout the paper, we measure credit market sentiment (CMS) using the Issuer Quality Index (ISS) developed by Greenwood and Hanson (2013). This index is designed to capture average issuer quality in the economy. ISS is calculated as the difference between the average of default probabilities of firms with the highest net debt issuance in a given year and the average of default probabilities of firms with the lowest net debt issuance in that year. Default probabilities at the firm level are estimated as in Bharath and Shumway (2008) and can be thought of as statistically equivalent to a credit rating, with the added benefit that they can be computed for a large set of firms as early as 1963. For this calculation, net debt issuance is defined by Greenwood and Hanson (2013) as the change in total assets minus the change in book equity, scaled by lagged total assets. Firms are categorized as high (low) net debt issuance if they are in the top (bottom) NYSE net debt issuance quintile. In the remainder of the paper, when we refer to credit market sentiment (CMS), we mean the ISS index of Greenwood and Hanson (2013). CMS tends to be high at or several quarters before the beginning of each NBER recession. Consistent with these trends, Greenwood and Hanson (2013) show that CMS significantly negatively predicts excess corporate bond returns in the following 2 years. When credit market sentiment is high (ie, in a credit market boom), according to the forecasting model of Greenwood and Hanson (2013), the expected return to bearing credit risk is low. In our Internet Appendix, we also present robustness tests using the high-yield-share index (HYS), also proposed by Greenwood and Hanson (2013), that provide the same conclusions as tests using ISS.⁵ HYS is calculated as the share of nonfinancial corporate bond issuance each year with a high-yield rating from Moody’s.

1.4 Macro-level data

In our empirical tests, we control for a series of macroeconomic variables: the Leading Economic Index from the Conference Board, the Jurado, Ludvigson, and Ng (2015) index of macro uncertainty, the Baker and Wurgler (2006) equity sentiment index, the default spread, the risk-free rate, and the Economic Policy Uncertainty (EPU) index of Baker, Bloom, and Davis (2016). Macroeconomic variables are measured at the end of the firm’s current fiscal year. If macroeconomic variables are reported more frequently than annually, we average their value over each year to provide an annual measure.

Table D.3 in the Internet Appendix presents correlations between CMS and our macroeconomic controls.⁶ Correlations are generally low between CMS and: the risk-free rate (0.34), the Baker and Wurgler (2006) equity sentiment index (0.22), the default spread (–0.27), and the macro uncertainty index (–0.29). The correlation between CMS and the Leading Economic Indicator (LEI) is also quite low, 0.10. A large inverse correlation, –0.62, exists between CMS and the EPU index of Baker, Bloom, and Davis (2016).

We ensure that the inclusion (or exclusion) of any of these macroeconomic controls does not drive our findings. All our findings are robust to excluding any of the above controls, one or two at a time. In addition, in unreported results, we find that our results are also robust to controlling for several alternative measures of first-moment shocks (ie, the Michigan Consumer Confidence Index, the Chicago FED National Activity Index, and expected GDP growth from SPF), second-moment shocks (ie, the VIX index from the Chicago Board Options Exchange, the standard deviation of GDP forecasts from SPF, and the cross-sectional standard deviation of firm-level profit growth), or sentiment in the equity market (ie Shiller’s Cyclically Adjusted PE Ratio).

2 Expectations and Credit Cycles

In this section, we study the connection, in the aggregate, between analysts’ expectations and credit cycles. Section 2.1 constructs a firm-level measure of biased expectations of fundamentals as the portion of forecast errors that is predictable by past forecast revisions. Section 2.2 aggregates our measure of biased expectations to the macro-level and documents its time-series correlation with CMS. In Section 2.3, we demonstrate that our measure of biased expectations can predict a number of alternative measures of credit market sentiment and discuss our measure relative to these variables.

2.1 Extracting the predictable component of forecast errors

Testing the Minsky (1957, 1977) hypothesis requires a measure of overextrapolation of shocks to firm fundamentals. In Minsky’s theory, firms overextend themselves through overborrowing and overinvesting and are increasingly exposed to any devaluation of their assets, at which point their cash flows are insufficient to cover their interest and principal payments. To capture overextrapolation of shocks to firm fundamentals relevant to the theory in Minsky (1957, 1977), we use equity analysts’ LTG forecasts, which measure the expected annual increase in operating earnings over the next 3 to 5 years (or approximately one business cycle), as well as revisions to those forecasts and errors (differences between actual and forecasted growth). While Minsky (1977) describes (operating) cash flows to meet principal and interest repayments, we use earnings per share (EPS) forecasts and subsequent revisions to provide a measure of overenthusiasm regarding firm cash flows.⁷ Moreover, several studies analyzing the information content of earnings forecasts, including from equity analyst reports and earnings conference calls, demonstrate the relevance of earnings forecasts to credit markets (Shivakumar et al. 2011; Donovan et al. 2021).

To extract the predictable component of forecast errors, we need both forecast revisions and forecast errors. To calculate forecast revisions (FR) for firm i in year t, we calculate:

$$\displaystyle FR_{i,t-1\rightarrow t}=E_{t}[\mathit{LTG}]-E_{t-1}[\mathit{LTG}].$$

(1)

|$ FR_{i,t-1\rightarrow t}$| is the difference between LTG forecasted in year t (⁠|$ E_{t}[\mathit{LTG}]$|⁠) and year t—1 (⁠|$ E_{t-1}[\mathit{LTG}]$|⁠). |$ E_{t}[\mathit{LTG}]$| is the average over year t of median monthly forecasts. We use medians rather than means to avoid the effect of potential outliers.

To calculate forecast error (FE), for firm i in year t, we calculate:

$$\displaystyle FE_{i,t\rightarrow t+k}=(EPS_{t+k}/EPS_{t})^{1/k}-E_{t}[\mathit{LTG}],$$

(2)

where |$ k=1,\ldots,5$|⁠, or, the actual growth in EPS from time t to t + k minus the projected LTG at time t. Taking the k^th root is necessary to annualize realized EPS growth so that it is comparable with LTG forecasts, which are annualized in IBES.

To extract the predictable component of forecast error, we follow Bordalo et al. (2020a) and Coibion and Gorodnichenko (2015) and estimate:

$$\displaystyle FE_{i,t\rightarrow t+k}=\alpha_{t+k}+\beta_{1,t}FR_{i,t-2\rightarrow t-1}+\beta_{2,t}FR_{i,t-1\rightarrow t}+\epsilon_{i,t+k},$$

(3)

where |$ k=1,\ldots,5$|⁠. There are time subscripts on the coefficients |$ \beta_{1,t}$| and |$ \beta_{2,t}$| because they are estimated on a rolling basis. For any given year t, we estimate Equation (3) only with data up to t, so that each regression does not include any forward-looking information.

Table 1 presents results from regressions of Equation (3) for 1-year to 5-year forecast error for the longest time series possible in our data (t = 2017). The results show that forecast revisions in years t—1 and t—2 are significantly negatively correlated with the 2- to 5-year forecast errors (⁠|$ FE_{t+2}$| to |$ FE_{t+5}$|⁠). Even the 1-year forecast errors in column 1 (⁠|$ FE_{i,t\rightarrow t+1}$|⁠) are significantly negatively related to the forecast revision 1 year prior (⁠|$ FR_{i,t-2\rightarrow t-1}$|⁠). At the 1-year horizon, we do not find a significant statistical relation between forecast errors (⁠|$ FE_{i,t\rightarrow t+1}$|⁠) and forecast revisions (⁠|$ FR_{i,t-1\rightarrow t}$|⁠).⁸ However, outside of that coefficient, all coefficients in Table 1 are strongly statistically significant.

Table 1

Predicting forecast errors with past forecast revisions

	|$ FE_{t\rightarrow t+1}$|	|$ FE_{t\rightarrow t+2}$|	|$ FE_{t\rightarrow t+3}$|	|$ FE_{t\rightarrow t+4}$|	|$ FE_{t\rightarrow t+5}$|
|$ FR_{t-1\rightarrow t}$|	–0.191	–0.567***	–0.623***	–0.644***	–0.638***
	(–1.32)	(–5.22)	(–8.78)	(–11.01)	(–14.74)
|$ FR_{t-2\rightarrow t-1}$|	–0.465***	–0.528***	–0.500***	–0.475***	–0.480***
	(–3.89)	(–4.89)	(–6.92)	(–8.28)	(–13.18)
Observations	37,856	32,345	29,563	26,960	24,697
R  2	.001	.014	.029	.045	.060

This table presents regressions of |$ FE_{i,t\rightarrow t+k}$| on two forecast revisions. Each column corresponds to a different |$ k=1,\ldots,5$|⁠. Forecast errors and revisions are calculated using forecasts of long-term EPS growth (LTG) from IBES. See Section 2 for details on their construction. Standard errors are corrected for autocorrelation using the Newey and West (1987) procedure. t-statistics are reported in parentheses.

*

p<0.1;

**

p<0.05;

***

p< 0.01.

The results in Table 1 show strong correlations between forecast revisions and subsequent forecast errors. As Coibion and Gorodnichenko (2015) discuss, under rational expectations, the correlation between forecast errors and forecast revisions should be zero. Consistent with results in Bordalo et al. (2020a), we find negative and significant coefficients in a regression of forecast errors on previous forecast revisions, which suggests that forecasts are extrapolative (Coibion and Gorodnichenko 2015; Bordalo et al. 2019, 2020a).

Recent literature questions whether Coibion and Gorodnichenko (2015) style regressions should be estimated with micro-level data, as in Bordalo et al. (2020a). Kohlhas and Walther (2021) show that Coibion and Gorodnichenko (2015) style regressions in micro-data are sensitive to outliers and that the negative coefficient in Bordalo et al. (2020a) is driven by 1% of the data. Juodis and Kučinskas (2023) and Kučinskas and Peters (2022) conclude that Coibion and Gorodnichenko (2015) regressions in micro-data primarily recover a relation between forecast errors and idiosyncratic noise in expectations. Because we aggregate our forecasts to the firm-year, so that our regressions are not estimated at the individual forecaster level, the critiques in Juodis and Kučinskas (2023) and Kučinskas and Peters (2022) are unlikely to be relevant to our setting. To ensure that our results are robust to the concerns in Kohlhas and Walther (2021), Juodis and Kučinskas (2023), and Kučinskas and Peters (2022), we perform two robustness tests. First, as Kohlhas and Walther (2021) recommend, we truncate each variable at the top and bottom 1% and reestimate our results in Table 1. Second, as Kučinskas and Peters (2022) recommend, we winsorize each variable at the top and bottom 5% and reestimate our results in Table 1. Results from these samples are in Internet Appendix  Table D.5 and are very similar to those in Table 1, so our results are robust to the concerns that Kohlhas and Walther (2021), Juodis and Kučinskas (2023), and Kučinskas and Peters (2022) raise.

For each firm i and year t, we estimate firm-level predicted forecast error (firmPFE) as fitted firm-level forecast errors estimated via Equation (3) with a 4-year time horizon (ie, k = 4),⁹ or:

$$\displaystyle \mathit{firmPF}E_{i,t}=\hat{\alpha}_{t}+\hat{\beta}_{1,t}FR_{i,t-2\rightarrow t-1}+\hat{\beta}_{2,t}FR_{i,t-1\rightarrow t}.$$

(4)

Intuitively, Equation (4) measures the component of forecast error (FE) that is predictable using the previous 2 consecutive forecast revisions (FR). Estimated firmPFE provides firm-level measures of overextrapolation of shocks to firm fundamentals necessary to test the theory in Minsky (1957, 1977).

2.2 Aggregating forecast errors to predict CMS

Next, we aggregate firmPFE at the macro-level to determine its correlation with credit market sentiment. Minsky (1977) suggests there should be a positive correlation between credit market sentiment and our firmPFE measure in the near term and an inverse correlation in the long term.

A crucial component of the Minsky (1977) hypothesis is that overextrapolation of shocks to firm fundamentals drives overborrowing and overinvestment by those same firms. So, to create a macro-level variable reflecting both firmPFE and the credit market activity of these firms, we weight firmPFE by firm-level debt issuance:

$$\displaystyle \mathit{macroPF}E_{t}=\sum \limits_{i=1}^{N}w_{i,t}\left(-\mathit{firmPF}E_{i,t}\right),$$

(5)

where

$$\displaystyle w_{i,t}=\frac{\Delta \mathit{Deb}t_{i,t-1\rightarrow t}}{\mathit{Asset}s_{i,t-1}},$$

(6)

and debt issuance (⁠|$ \Delta \mathit{Deb}t_{i,t-1\rightarrow t}$|⁠) is the change in long-term debt (dltt + dlc in Compustat) scaled by one lag of total assets. The minus sign in front of |$ \mathit{firmPF}E_{i,t}$| adjusts for the fact that periods of higher sentiment are periods in which predicted forecast errors should be more negative. Because we multiply |$ \mathit{firmPF}E_{i,t}$| by –1, we expect our macroPFE index to be positively correlated with contemporaneous CMS. In panel A of Figure 1, we plot macroPFE (green) and CMS (blue) over time. This figure shows a very strong correlation between the two time series, 0.75.¹⁰ This correlation suggests that we are able to closely capture the information contained in the Greenwood and Hanson (2013) CMS measure using our aggregation of predicted forecast errors.

Fig. 1

CMS and predicted forecast error

This figure plots the credit market sentiment (CMS) index of Greenwood and Hanson (2013) in blue alongside the debt-weighted firm-level predicted forecast error (macroPFE, top panel, green) and average firm-level predicted forecast error (⁠|$ \overline{\mathit{firmPFE}}$|⁠, bottom panel, orange). All variables are standardized to have a standard deviation of 1 and mean of 0.

In our estimation of Equation (4), allowing the β coefficients to vary over time (by reestimating the regression every year) has an important implication for the relation between our index and CMS. The magnitude of these coefficients captures the strength of overextrapolation tendencies present in the market at any point in time. Hence, our macroPFE_t index captures two important sources of time-series variation: variation in the extent to which the market overextrapolates (ie, |$ \beta_{1,t}$| and |$ \beta_{2,t}$| in Equation (4)), and variation in the debt-issuance weighted average of forecast revisions. For example, our index is particularly high when the market is in a high overextrapolation state (very negative |$ \beta_{1,t}$| and |$ \beta_{2,t}$|⁠) and firms that have received particularly good news (very positive forecast revisions) have issued a lot of debt. The debt weighting, therefore, helps align macroPFE_t with sentiment in the credit market in particular as opposed to sentiment in financial markets as a whole.

For comparison with our debt-weighted macroPFE, we calculate an index wherein firms are equally weighted. For this scenario, in Equation (5), we set |$ w_{i,t}=1/N$|⁠, and calculate an annual average of firmPFE, which we denote as |$ \overline{\mathit{firmPFE}}$|⁠. Panel B of Figure 1 plots |$ \overline{\mathit{firmPFE}}$| (orange) and CMS (blue) over time. A comparison of panels A and B in Figure 1 highlights the importance of debt-weighting to capturing credit market sentiment with firmPFE. The correlation between |$ \overline{\mathit{firmPFE}}$| and CMS is 0.29, less than half the correlation between macroPFE and CMS. Thus, the average level of firm-level overextrapolation, generally, does not capture CMS as closely as our debt-weighted firm-level measure of overextrapolation.

To test whether macroPFE predicts boom-bust cycles in credit market sentiment, we regress CMS on lags of our aggregate firmPFE measure. In panel A of Table 2, we report results from this regression:

$$\displaystyle CMS_{t+k}=\alpha +\beta \mathit{macroPF}E_{t}+\epsilon_{t},$$

(7)

where |$ k=0,\ldots,5$|⁠.

Table 2

Predicting CMS with predictable firm-level forecast error

	CMSt	|$ CMS_{t+1}$|	|$ CMS_{t+2}$|	|$ CMS_{t+3}$|	|$ CMS_{t+4}$|	|$ CMS_{t+5}$|
A. macroPFE is firmPFE weighted by total debt issuance (change in long-term debt)
macroPFE	0.750***	0.484***	0.035	–0.367**	–0.558***	–0.434***
	(7.942)	(4.583)	(0.256)	(–2.471)	(–7.507)	(–3.737)
Constant	–0.000	0.017	0.085	0.106	0.097	0.077
	(–0.000)	(0.089)	(0.363)	(0.415)	(0.362)	(0.306)
Observations	28	27	26	25	24	23
R  2	.563	.237	.001	.150	.341	.208
						B. |$ \overline{\text{firmPFE}}$|  is a simple average of firmPFE
|$ \overline{\mathit{firmPFE}}$|	0.287	0.299*	0.166*	0.063	–0.098	–0.067
	(1.338)	(1.907)	(1.803)	(0.237)	(–0.256)	(–0.237)
Constant	0.000	0.024	0.083	0.098	0.094	0.069
	(0.000)	(0.095)	(0.342)	(0.387)	(0.356)	(0.262)
Observations	28	27	26	25	24	23
R  2	.082	.091	.031	.004	.011	.005

This table presents regression of |$ CMS_{t+k}$| on an aggregation of firmPFE, which is calculated differently in each panel. In panel A, macroPFE is calculated by weighting firmPFE by change in total debt in the last year (dltt_t + dlc_t—dltt|$_{t-1}$| – dlc|$_{t-1}$|⁠, scaled by total assets at |$ t-1$|⁠). In panel B, |$ \overline{\mathit{firmPFE}}$| is a simple average annually of firmPFE. The predicted forecast errors are fitted values from panel regressions of forecast errors on forecast revisions. See Section 2 for details. Each column corresponds to a different |$ k=0,\ldots,5$|⁠. All variables are standardized so the economic magnitude of the coefficients is in terms of number of standard deviations from the mean. Standard errors are corrected for autocorrelation for up to six lags using the Newey and West (1987) procedure. t-statistics are reported in parentheses.

*

p<0.1;

**

p<0.05;

***

p< 0.01.

The results in panel A of Table 2 show that macroPFE predicts a boom-bust cycle in credit market sentiment. The coefficient for macroPFE in the first column represents a regression of contemporaneous CMS on macroPFE. This coefficient is positive and statistically significant. Both macroPFE and CMS are scaled to have a mean of zero and a standard deviation of one, so the coefficient estimate of 0.750 shows that, for a one-standard-deviation increase in macroPFE, CMS increases by 0.750 standard deviations. There is also a positive, albeit smaller in magnitude, change in CMS in year t + 1 for a one-standard-deviation increase in macroPFE in year t. There is an inverse relation between CMS in years t + 3, t + 4, and t + 5 and macroPFE in year t. A one-standard-deviation increase in macroPFE in year t is associated with a 0.367-, 0.558-, and 0.434-standard-deviation decrease in CMS in years t + 3, t + 4, and t + 5, respectively.

To the best of our knowledge, the results in panel A of Table 2 comprise the first evidence directly showing that credit market sentiment reflects biased expectations of fundamentals of firms accessing credit markets. Minsky (1977) hypothesizes that, in credit market booms, increases in analysts’ optimism drive down the cost of capital. As a result, credit expands via a disproportionate issuance of debt by firms with higher default probabilities. Then, credit predictably dries out in years t + 3 and t + 4. The lack of positive news supporting the credit expansion generates systematic disappointment, which drives up the cost of capital and triggers a credit contraction. In our results in Table 2, we show that our debt-weighted firm-level measure of overoptimism (macroPFE) significantly predicts credit market sentiment. We find a positive relation between contemporaneous macroPFE and CMS and a strong inverse relation between macroPFE in year t and CMS in years t + 3, t + 4, and t + 5, consistent with the reversal that Minsky (1977) describes.

In contrast, we see a weaker cycle in CMS related to average firm-level overoptimism in panel B of Table 2. For the results in panel B of Table 2, we estimate a series of regressions in which we swap out macroPFE with a simple average of firmPFE (⁠|$ \overline{\mathit{firmPFE}}$|⁠). The results show a positive correlation between |$ \overline{\mathit{firmPFE}}$| and CMS in the following 2 years, but the coefficients for |$ \overline{\mathit{firmPFE}}$| for CMS in years t and t + 1 are smaller in magnitude than those in panel A. The statistical significance of the positive relation between aggregated firmPFE and CMS is greatly attenuated by equal weighting firmPFE, rather than weighting firmPFE by firm-level debt issuance. Although there is an inverse correlation between |$ \overline{\mathit{firmPFE}}$| in t and CMS in years t + 4 and t + 5, these coefficients are not statistically significant. Additionally, the magnitude of these coefficients is very small, relative to the coefficients for macroPFE in panel A. Taken together, this evidence supports the univariate evidence in Figure 1, suggesting that weighting firmPFE by debt is important to capturing information relevant to predicting CMS. Relatively high levels of firm-level overextrapolation is insufficient to drive cycles in credit market conditions; firms for which overextrapolation of shocks to fundamentals exists must also issue contemporaneous debt for a strong correlation between aggregated firmPFE and CMS to exist. A key contribution of this paper is demonstrating the strong correlation between both contemporaneous and leading CMS and firm-level predicted forecast error weighted by debt issuance.

In Table D.6 in our Internet Appendix, we present regression results, including broad measures of macroeconomic uncertainty using the Jurado, Ludvigson, and Ng (2015) index, as well as for the first moment of expected economic conditions using the Leading Economic Index from the Conference Board, and with CMS regressed only on macro-level controls. The results in Table D.6 show that our results hold (and, if anything, become stronger) when we control for these measures of first and second moments of macroeconomic conditions.

Next, we provide three validation exercises of the CMS measures that we use in this paper. First, in Figure 2, we examine aggregate debt issuance, as well as firm leverage scaled by assets and equity, and CMS. Our goal is to capture overreaction to shocks to firm fundamentals relevant to credit markets, so here we examine both debt issuance and the value of equity together to see whether and how the market value of equity changes as debt issuance rises. If equity and credit market sentiment rise together, then we would expect a rise in market leverage when CMS is high, as firms issue “cheap” equity, despite a lack of increase in equity’s value. Panel A of Figure 2 shows that, as expected, change in total net debt issuance, measured as change in long-term debt (dltt + dlc in Compustat), rises and falls with CMS. Panel B shows that long-term debt scaled by assets rises as CMS rises until approximately 1999 and then falls and is relatively flat over the period. In panel C, as CMS rises from 1992 to 1996, aggregate leverage scaled by equity is relatively flat. Leverage scaled by equity jumps after 1998 only as CMS begins to fall. Leverage scaled by equity dips as CMS rises and remains high in the mid-2000s and remains flat during the CMS spike peaking in 2014. This relation between CMS market value of leverage (long-term debt scaled by market value of equity) suggests that, even as debt issuance rises (panel A), market value of equity is also rising, so that market leverage is level or falls when CMS is high. These results suggest that CMS captures credit market sentiment distinct from equity market sentiment. Otherwise, market leverage would rise as debt issuance rises when CMS is high.

Fig. 2

CMS and aggregate corporate credit market measures

This figure plots the credit market sentiment (CMS) index of Greenwood and Hanson (2013) in blue alongside three measures of corporate debt activity aggregated annually across the economy: (a) total net debt issuance (dltt_t + dlc_t—dltt|$_{t-1}$| – dlc|$_{t-1}$|⁠) scaled by total assets at |$ t-1$| in panel A, (b) total long-term debt (dltt_t + dlc_t) scaled by total assets at |$ t-1$| in panel B, and (c) total long-term debt (dltt_t + dlc_t), scaled by the market value of equity at |$ t-1$| in panel C. All variables are standardized to have a standard deviation of 1 and mean of 0.

Second, we examine the correlation between our measure of firm-level overextrapolation of shocks to firm fundamentals (firmPFE) and firm-level excess bond premium (EBP) from Gilchrist and Zakrajšek (2012).¹¹ EBP is the residual from a regression of bond-level credit spreads on firm-level expected default rates and bond characteristics.¹² Since overoptimism should be associated with lower (more negative) forecast errors, as well as lower excess bond premia (opposite signs for overpessimism), we should observe a positive relation between firmPFE and EBP.¹³

In panel A of Table 3, we regress our EBP measure on firmPFE. Since EBP is calculated at the bond-month level, and firmPFE is calculated at the firm-year level, we run our regressions at three different aggregation levels and report the results across the columns. In the first column, to estimate regressions at the bond-month level, we simply duplicate the firm-level values of firmPFE and debt issuance for all the bonds of the firm, in all months of the corresponding year. The result in column 1 shows a strong positive relation between EBP and firmPFE that is statistically significant at the 1% level, suggesting that our firmPFE measure does indeed capture expectation errors in the credit market. In column 2 of Table 3, we aggregate bond-level EBPs to the firm-level using bond-level amount outstanding as weights. Our results are similar if we instead use equal-weighting. We then estimate firm-month-level regressions, where annual firmPFE is duplicated for all the months in the corresponding year. This result is in column 2 and again shows a strong positive relation between firmPFE and EBP. Finally, in the third column, we aggregate the firm-month EBP values used in column 2 to the annual level (by taking a simple average over the months in each year) and find a qualitatively similar result to that in column 2. By showing a significant correlation between our firmPFE and the EBP from Gilchrist and Zakrajšek (2012), the results in panel A of Table 3 show that firmPFE captures firm-level overextrapolation of shocks to firm fundamentals.

Table 3

Validation of firmPFE and macroPFE

A. Regressions of excess bond premium on predictable forecast errors
	Bond-month-level	Firm-month-level	Firm-year-level
firmPFE	0.082***	0.059***	0.058***
	(5.41)	(4.57)	(2.91)
Firm fixed effects	Yes	Yes	Yes
Observations	335,391	86,569	7,967
R  2	.158	.239	.317

B. Predicting BAA forecast errors with macroPFE
	|$ \mathit{Ferro}r_{t\rightarrow t+12}$|	|$ \mathit{Ferro}r_{t\rightarrow t+15}$|	|$ \mathit{Ferro}r_{t\rightarrow t+18}$|
macroPFE	0.686***	0.691***	0.668***
	(5.65)	(4.91)	(3.87)
Observations	18	18	18
R  2	.471	.477	.446

Panel A reports coefficient estimates from regressing excess bond premia (EBP) on firmPFE. The dependent variable EBP is measured at the bond-month level following Gilchrist and Zakrajsek (2012), and firmPFE is measured at the firm-year level. In the first column of panel A, we run regressions at the bond-month level, where annual firmPFE is simply duplicated for all bonds of the firm, in all months of the corresponding year. In the second column of panel A, we aggregate the dependent variable at the month level, using a weighted average of the EBP values of the firm’s bonds that month. The weights are given by the bonds’ amount outstanding that month. We then run firm-month-level regressions with this weighted average EBP as the dependent variable and annual firmPFE values duplicated for all months of the corresponding year. Finally, in the third column of panel A, we run regressions at the firm-year level, where the dependent variable is a simple average of the monthly EBP averages used in the regressions with results reported in panel A, column 2. All regressions in panel A include firm fixed effects and cluster standard errors by firm and time. In panel B, we present coefficient estimates from regressions of 12-, 15-, and 18-months-ahead forecast errors of the BAA yield on macroPFE. Forecast errors are calculated each month t as the realized future BAA yield (12, 15, and 18 months into the future) and the consensus (median) forecast for it as of time t. Monthly forecast errors are averaged across each calendar year to match the frequency of our macroPFE index. Standard errors are corrected for autocorrelation using the Newey and West (1987) procedure. t-statistics are reported in parentheses.

*

p<0.1;

**

p<0.05;

***

p< 0.01.

Finally, in panel B of Table 3, we show that our macroPFE index predicts forecast revisions in Blue Chip Financial Forecasts, which directly capture credit market expectations. We calculate forecast errors in Blue Chip BAA forecasts using 12-, 15-, and 18-months-ahead forecasts and regress these errors on contemporaneous macroPFE. BAA forecast errors are calculated as the realized future BAA yield minus the current consensus (median) forecast for that BAA yield. These forecast errors are calculated monthly, so we take an average over the monthly values in each year to match the annual frequency of the macroPFE index. Both the dependent and independent variables are standardized, so that the coefficient estimates can be interpreted as the correlation between the two variables. The horizon for the forecast errors is limited by the fact that Blue Chip only contains forecasts of the BAA yield for up to six quarters in the future. Additionally, our sample size for these tests is limited to 18 observations because Blue Chip data starts in 2000, and our macroPFE runs through 2017. However, despite the limited sample, we find a strong correlation, 66 to 69%, between macroPFE and Blue Chip BAA forecast errors, depending on the horizon. The positive sign of the coefficients is as expected since overoptimism in the credit market (ie, higher macroPFE) should be associated with excessively low expectations for bond yields (ie, higher forecast errors).

2.3 Alternative measures of market sentiment

While our main measure of credit market sentiment is the ISS index from Greenwood and Hanson (2013), we also consider whether our macroPFE measure predicts a boom-bust pattern in other measures of sentiment. In Table D.7 in our Internet Appendix, we find similar results to those in Table 2 using the high-yield share of new issues (HYS) index also from Greenwood and Hanson (2013). The contemporaneous correlation between macroPFE and HYS is positive, and an increase in macroPFE in year t predicts a significant decline in HYS in year t + 3. In panel B of Table D.7, we show a significant relation between macroPFE and the default spread, defined as the difference between Moody’s Baa corporate bond yield and the Aaa yield. The contemporaneous correlation between macroPFE and the default spread is negative. An increase in macroPFE in year t predicts a significant decline in the default spread in year t + 3.

Table D.8 shows similar results to panel B of Table D.7 for measures of credit spread from Gilchrist and Zakrajšek (2012). Table D.8 shows that, at year t, macroPFE is inversely correlated with the GZ spread, the excess bond premium, and the estimated probability of default (see panels A, B, and C, respectively). Additionally, an increase in macroPFE in year t predicts a significant increase in the GZ spread in years t + 3 and t + 4 and in both components of the GZ spread in years t + 2, t + 3, and t + 4. The excess bond premium measures credit spreads net of the estimated probability of default, so the similar pattern in results between panels B of Tables D.7 and D.8 are as expected.

In addition to measures of credit spread, some studies use balance sheet measures to approximate credit market conditions. For example, Mian, Sufi, and Verner (2017) study aggregated household debt-to-GDP and Schularick and Taylor (2012) and Jordá, Schularick, and Taylor (2013) examine various measures of aggregated bank loans. In the Internet Appendix  Table D.9, we test whether macroPFE can predict boom-bust cycles in measures from Schularick and Taylor (2012) in data available through 2008.¹⁴ The measures from Schularick and Taylor (2012) we examine are: total bank loans scaled by CPI (panel A), total bank loans scaled by total assets (panel B), change in total bank loans scaled by CPI (panel C), and change in total bank loans scaled by total assets (panel D).

The results in the Internet Appendix  Table D.9 suggest that our macroPFE measure predicts a boom-bust cycle in the measures from Schularick and Taylor (2012). In panel A, a boom-bust pattern is apparent in the coefficients for macroPFE_t predicting bank loans scaled by CPI in years t to t + 5; however, none of these coefficients is statistically different from zero. Results in panel B show that macroPFE in year t predicts a significant increase in bank loans scaled by bank assets in years t + 1 and t + 2 and a significant decline in bank loans scaled by bank assets in years t + 4 and t + 5. Results in panel C show that a one-standard-deviation increase in macroPFE in year t decreases change in bank loans scaled by CPI by 1%, 1.5%, and 1.2% in years t + 2, t + 3, and t + 4, respectively. Finally, in panel D, results show that a one-standard-deviation increase in macroPFE in year t predicts change in bank loans scaled by assets increases by 2% in year t and 1.5% in year t + 1 but decrease by 0.9%, 1.2%, and 0.4% in years t + 3, t + 4, and t + 5, respectively.

Results in Table D.9 are generally weaker when bank loans are scaled by CPI rather than bank assets. Consistent with the relative strength of results in Table D.9, the contemporaneous correlation between macroPFE and bank loans scaled by CPI is only 0.12 (Internet Appendix  Table D.4), whereas the correlation between macroPFE and bank loans scaled by bank assets is 0.56. However, the results in Table D.9 are consistent with macroPFE predicting boom-bust cycles in bank loans. Taken together, results across Tables D.7, D.8, and D.9 imply that macroPFE captures the time-series dynamics of the credit market rather broadly.

Table 4

Predictable firm-level forecast error drives cycles in corporate investment

	Year 1	Year 2	Year 3	Year 4	Year 5
A. Dependent variable: Total investment
macroPFE	0.046***	0.044***	0.011	–0.027***	–0.025***
	(5.357)	(5.349)	(1.259)	(–3.092)	(–3.010)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	57,597	52,648	47,150	42,139	37,643
R  2	.379	.314	.293	.288	.281
B. Dependent variable: Physical investment
macroPFE	0.036***	0.012	–0.030***	–0.079***	–0.073***
	(4.003)	(1.398)	(–3.547)	(–9.472)	(–8.294)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	57,597	52,658	47,712	42,664	38,160
R  2	.397	.351	.346	.335	.337
C. Dependent variable: Intangible investment
macroPFE	0.047***	0.063***	0.036***	0.009	0.010
	(4.447)	(6.069)	(3.380)	(0.825)	(0.828)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	57,709	52,747	47,245	42,223	37,721
R  2	.366	.309	.285	.282	.275

This table presents coefficient estimates from regressing total investment (panel A), investment in physical capital (panel B) and investment in intangible capital (panel C) up to 5 years in the future on aggregated predictable firm-level forecast error (macroPFE) and controls. Predictable firm-level forecast error (see Section 2 for details) is weighted by change in total debt (as in defined in panel A of Table 2). Column headings Year “k” (k = 1|$ \ldots $|5) mean the dependent variable is measured at time t + k, while all independent variables are measured at time t. The investment and Tobin’s q variables are measured as in Peters and Taylor (2017). Specifically, total capital is gross PPE (ie, physical capital) plus the sum of goodwill, capitalized R&D, and capitalized SG&A (ie, intangible capital). Total investment is the percentage change in total capital, investment in physical capital is the change in physical capital divided by lagged total capital, and investment in intangible capital is the change in intangible capital divided by lagged total capital. Tobin’s q is the market value of equity plus book value of debt divided by total capital. All specifications include firm fixed effects, firm-level controls (Tobin’s q, Cash flow to assets, log total assets, Cash to assets, Book leverage, Sales growth, ROA, default probability) and macro-level controls (the Leading Economic Index from the Conference Board, the Jurado, Ludvigson, and Ng [2015] index of macro uncertainty, the Baker and Wurgler [2006] sentiment index, the default spread, the risk-free rate, and the policy uncertainty index of Baker, Bloom, and Davis [2016]), as well as the current and lagged level of the dependent variable. Standard errors are bootstrapped according to Abadie et al. (2023). t-statistics are reported in parentheses.

*

p<0.1;

**

p<0.05;

***

p< 0.01.

Considering alternative measures of market sentiment also allows us to highlight a key distinction in the creation of our measure from the literature. Whereas other studies use aggregated levels of debt, for example, household debt or bank loans (Mian, Sufi, and Verner 2017; Schularick and Taylor 2012; Jordá, Schularick, and Taylor 2013) to capture information about credit market conditions, our focus is on corporate debt issuance. We calculate macroPFE as an aggregation of the product of predictable firm-level forecast errors and firm-level debt issuance because Minsky (1957, 1977) suggest that debt issuance by firms for which there is overextrapolation of shocks to firm fundamentals should also experience eventual reversals in investment and debt issuance. Of course, firm leverage and long-term debt issuance are correlated. However, as detailed in the remainder of this section, we find that corporate debt issuance, not leverage, is the key to predicting CMS.

In Table D.10, we provide results for variations on results in Table 2. In panels A and B, we aggregate firmPFE to macroPFE by weighting by firm-level leverage (long-term debt scaled by total assets), rather than by net debt issuance. In panels C and D, we aggregate firmPFE to macroPFE using net debt issuance, as in panel A of Table 2; results in panel C repeat those in panel A of Table 2 for comparison with the results in panel D. In panels B and D, we omit from our calculation of macroPFE firms in the highest 5% of net debt issuance annually. A comparison of results across all four panels helps demonstrate whether firm-level leverage or issuance is most relevant to predicting CMS.

The results in Table D.10 show that corporate debt issuance, not leverage, is key to predicting CMS. Results in panel A of Table D.10 show that an increase in macroPFE in year t calculated with leverage rather than issuance predicts a significant increase in CMS in years t through t + 3. However, there is a weaker reversal in panel A of Table D.10 than in our main results in panel A of Table 2 (which we replicate in panel C of Table D.10). If we weight macroPFE by firm leverage, there is only a statistically significant decline in CMS in year t + 4 related to macroPFE_t; this estimate is only significant at the 10% level. In panel B, when we remove the top 5% of issuers from our sample, the reversal further weakens and loses all statistical significance. In comparison, in panel D, when we remove the top 5% of issuers from our sample in which we weight macroPFE by issuance, rather than leverage, we continue to see a significant boom-bust cycle in CMS predicted by macroPFE. Results in panel D are smaller in magnitude than those in panel C, in which we leave all issuers in the calculation of macroPFE. However, even omitting top issuers, macroPFE weighted by issuance predicts a significant increase in CMS in years t and t + 1 and a decline in CMS in years t + 4 and t + 5. These results demonstrate that measuring macroPFE with debt issuance is key to testing the theory in Minsky (1957, 1977).

3 Credit Cycles and Corporate Investment and Financing

In this section, we study the extent to which credit market sentiment predicts corporate investment and financing at the firm level. In Section 3.1, we describe our approach. Section 3.2 studies how macroPFE predicts corporate investment and financing, in both the short and long run. Section 3.3 compares how macroPFE predicts corporate investment and financing for financially constrained versus unconstrained firms.

3.1 Baseline specification

Our baseline regressions generally take the following form:

$$\displaystyle Y_{i,t+k}=\alpha_{i}+\beta_{k}m\mathit{acroPF}E_{t}+\gamma F_{i,t}+\delta M_{t}+Y_{i,t}+Y_{i,t-1}+\varepsilon_{i,t+k}$$

(8)

where k indexes the year in which the dependent variable is measured relative to year t in which the independent variables are measured. |$ Y_{i,t+k}$| is either (a) corporate investment in tangible capital, intangibles, or the sum of tangible capital and intangibles, or (b) a measure of debt financing, including long-term debt issuance (change in total long-term debt), change in long-term debt due within the next year, and change in long-term debt due in more than 1 year. macroPFE_t is our index of aggregated predictable firm-level forecast errors described in the previous section; |$ F_{i,t}$| is a vector of firm-level controls; M_t is a vector of macro-level controls; and α_i is a set of firm fixed effects. We detail these controls when we report our results. In addition, consistent with the Jordá (2005) method, we control for the contemporaneous level of the dependent variable, as well as its lag (⁠|$ Y_{i,t}$| and |$ Y_{i,t-1}$|⁠, respectively).¹⁵

Estimating Equation (8) for |$ k=1,2,3,4,5$| traces out the Jordá (2005) local projection impulse response function, β_k. Note that, in our context, the error terms for two firms, i and j, in a given year t are likely to be correlated, since macroPFE does not vary in the cross-section. Similarly, given the persistence in both macroPFE and corporate investment, the error terms of firm i at time t and firm j at time t + 1 are also potentially correlated. To adjust standard errors for both types of correlation, we bootstrap standard errors for all estimations of Equation (8) according to the recommendations in Abadie et al. (2023). Finally, to help with the interpretation of economic magnitudes, we normalize all dependent variables by their mean and all the independent variables by their standard deviation. Hence, each coefficient can be interpreted as the marginal effect of a one-standard-deviation change in the independent variable, expressed as a percentage of the mean of the dependent variable.

3.2 Using macroPFE to predict investment and debt issuance

We begin by estimating Equation (8) with measures of firm investment as dependent variables. To construct these variables, we follow Peters and Taylor (2017), who argue that intangible capital (ie, the sum of goodwill, capitalized R&D, and capitalized SG&A) has become an increasingly important factor of production and should therefore be included alongside physical capital (ie, gross PPE) in any analysis of corporate investment activity. Our investment measures are change in total capital, change in physical capital, and change in intangible capital, each expressed as a percent of lagged total capital.

Results in Table 4 show a boom-bust pattern in corporate investment related to macroPFE that is consistent with the Minsky hypothesis. Column 1 presents results for a regression with a dependent variable measured in year t + 1; column 5 reports results for a dependent variable measured in year t + 5. Results for investment in total capital are presented in panel A. For a one-standard-deviation (one-unit) increase in macroPFE in year t, investment changes relative to its mean by: 4.6% in year t + 1, 4.4% in year t + 2, 1.1% in year t + 3, –2.7% in year t + 4, and –2.5% in year t + 5.

Results in panels B and C of Table 4 show the effect of macroPFE on investment in physical capital and intangibles, respectively, and support the results in panel A. In panel B, the results show that a one-standard-deviation increase in macroPFE in year t corresponds to a 3.6% increase (relative to its mean) in physical investment in year t + 1 and a significant decline in investment in year t + 3 to year t + 5. The decline in investment in physical capital is economically significant, 7.9% and 7.3% in years t + 4 and t + 5, respectively. In panel C, a one-standard-deviation increase in macroPFE in year t corresponds to a more prolonged period of higher investment in intangibles in years t + 1, t + 2, and t + 3. The increase in intangible investment peaks in t + 2 at 6.3%, declines to 3.6% in t + 3, and attenuates to zero in years t + 4 and t + 5. The results in panel A reflect a weighted average of the effects seen in panels B and C, as total investment (panel A) is the sum of investment in tangible (panel B) and intangible (panel C) capital.

In all of our tests, we control for a large set of firm level variables. At the firm-level, we control for Tobin’s q (the market value of equity plus book value of debt, divided by total capital), the ratio of cash flow to assets, and several controls for the strength of the balance sheet: the logarithm of total assets to proxy for firm size, the ratio of cash to assets and book leverage to proxy for corporate liquidity, and sales growth and ROA to proxy for the firm’s operating performance. The estimated coefficients for the covariates, though not reported to preserve space, have expected signs; firms with higher investment opportunities, higher liquidity, and better performance invest more. None of the covariates affects our baseline result.

In all tests, we also control for a broad set of macroeconomic variables, which is crucial because our variable of interest, macroPFE, varies only in the time series. In terms of potentially confounding macroeconomic conditions, we control for: (a) the risk-free rate, (b) aggregate investment opportunities (Leading Economic Indicator Index from the Conference Board), (c) the macroeconomic uncertainty index from Jurado, Ludvigson, and Ng (2015), (d) the equity market sentiment index from Baker and Wurgler (2006), (e) the EPU index¹⁶ from Baker, Bloom, and Davis (2016), and (f) the aggregate valuation of debt (the default spread).

Furthermore, we ensure that our documented effect of credit market sentiment on corporate investment does not operate through firm-level credit risk. Firm-level credit risk can affect investment in two ways. First, if a boom in credit market sentiment increases credit risk at the firm level, then we should observe an increase in both firm-level default probability and firm-level investment through an asset-substitution-type of mechanism, as argued, for example, by Gomes, Grotteria, and Wachter (2019). Alternatively, higher credit risk may come with poor investment opportunities and, hence, lower subsequent investment. To address these possibilities, we control in all our panel regressions for firm-level default probability, as measured by the Bharath and Shumway (2008) index.¹⁷ Our results indicate that higher firm-level default probability is negatively associated with subsequent investment, but our results on the predictive power of macroPFE are unaffected.

Next, we estimate Equation (8) with measures of firm debt issuance as dependent variables. Table 5 reports these results. The dependent variable for panel A is total long-term debt issuance, measured as change in long-term debt (dltt + dlc) divided by lagged total assets. In panels B and C, the dependent variable is change in the portion of long-term debt due to mature in more than 1 year (dltt) and less than 1 year (dlc), respectively. By breaking down change in long-term debt into these components, we ensure that our results are not driven by changes in debt due to expire in the near-term.

Table 5

Predictable firm-level forecast error drives cycles in debt issuance

	Year 1	Year 2	Year 3	Year 4	Year 5
A. Dependent variable: Total debt issuance (change in long-term debt)
macroPFE	0.240***	0.134***	–0.020	–0.123***	–0.103***
	(10.449)	(5.990)	(–0.946)	(–5.267)	(–3.976)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	56,282	51,506	46,763	41,809	37,425
R  2	.258	.216	.191	.177	.173
B. Dependent variable: Change in long-term debt expiring in >1 year
macroPFE	0.234***	0.127***	0.020	–0.067**	–0.112***
	(9.060)	(4.982)	(0.702)	(–2.579)	(–3.966)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	56,017	51,256	46,556	41,654	37,300
R  2	.224	.197	.182	.170	.168
C. Dependent variable: Change in long-term debt expiring in <1 year
macroPFE	0.380***	0.125***	–0.120***	–0.215***	–0.130***
	(7.844)	(2.779)	(–2.816)	(–5.021)	(–2.917)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	56,910	52,042	47,270	42,301	37,872
R  2	.229	.144	.132	.127	.122

This table presents coefficient estimates from regressing total net debt issuance (panel A), change in long-term debt due to expire in more than one year (panel B) and change in long-term debt due to expire in less than one year (panel C) up to 5 years in the future, on aggregated predictable firm-level forecast error (macroPFE) and controls. Predictable firm-level forecast error (see Section 2 for details) is weighted by change in total debt (as in defined in panel A of Table 2). Column headings Year “k” (k = 1|$ \ldots $|5) mean the dependent variable is measured at time t + k, while all independent variables are measured at time t. Total net debt issuance is the change in long-term debt (“dltt”+“dlc” in Compustat) scaled by lagged total assets. Change in long-term debt due to expire in more than one year is “dltt”in Compustat scaled by lagged total assets. Change in long-term debt due to expire in less than one year is “dlc” in Compustat scaled by lagged total assets. All specifications include firm fixed effects, firm-level controls (Tobin’s q, Cash flow to assets, log total assets, Cash to assets, Book leverage, Sales growth, ROA, default probability) and macro-level controls (the Leading Economic Index from the Conference Board, the Jurado, Ludvigson, and Ng [2015] index of macro uncertainty, the Baker and Wurgler [2006] sentiment index, the default spread, the risk-free rate, and the policy uncertainty index of Baker, Bloom, and Davis [2016]), as well as the current and lagged level of the dependent variable. Standard errors are bootstrapped according to Abadie et al. (2023). t-statistics are reported in parentheses.

*

p<0.1;

**

p<0.05;

***

p< 0.01.

Consistent with the Minsky hypothesis and our investment results in Table 4, results in Table 5 show a boom-bust pattern in debt issuance related to macroPFE. In panel A, for a one-standard-deviation increase in macroPFE in year t, the results show 24.0% and 13.4% increases in long-term debt issuance as a percent of total assets in years t + 1 and t + 2, respectively, and declines of 12.3% and 10.3% in years t + 4 and t + 5, respectively. The results in panels B and C of Table 5 show that the cycle in debt issuance is not driven by either component of long-term debt and, thus, broadly reflects changes to firm-level debt issuance.

In the Internet Appendix, we present results demonstrating the robustness of the results reported in Tables 4 and 5. First, we examine whether results are sensitive to the exclusion of the financial crisis of 2008–2009. In Tables D.11 and D.12 we report in panel A results in the pre-2008 subsample and, in panel B, we report results excluding the years 2008–2009 of the financial crisis. Our results are very similar to the baseline results in Tables 4 and 5.

Next, we explore the idea that a portion of the proceeds raised by issuing debt in response to credit market sentiment might be used to repurchase shares (Ma 2019), rather than to invest. Tables D.13, D.14, and D.15 in the Internet Appendix examine the effect of macroPFE on net debt issuance, net equity repurchases, and total external financing (net debt issuance minus net equity repurchases), as Ma (2019) defines. We report results using all firms in our sample (Table D.13), only firms in the top size decile (Table D.14), and only firms in the bottom nine size deciles (Table D.15).

Table D.13 shows that credit market sentiment is not associated with any kind of cycle in equity repurchases for the average firm (panel B). We see an inverse relation between macroPFE and net equity repurchases across years t + 1 to t + 5, signifying that the increase in debt issuance that occurs in years t + 1 and t + 2 associated with macroPFE does not correspond with an increase in equity repurchasing. In Tables D.14 and D.15, we see that firms in the top size decile experience both higher net debt issuances and higher repurchases in years t + 1 and t + 2, consistent with Ma (2019). These results provide some evidence that firms act as cross-market arbitrageurs. That is, when credit is cheap, firms issue debt and use the proceeds to repurchase shares, consistent with Ma (2019). Yet, this evidence is confined to firms in the top size decile in our data. In Internet Appendix Tables D.16, D.17, and D.18, we reestimate our investment results in Table 4 in the top and bottom nine deciles by firm size. We find significant total investment booms in both samples in years t + 1 and t + 2, and a flattening of investment in large firms and a significant reversal in investment in all other firms in years t + 4 and t + 5. Additionally, we find a significant boom in years t + 1 and t + 2 and a reversal in years t + 4 and t + 5 in physical investment for both samples. Hence, we continue to find broad evidence of a cyclical pattern in corporate investment related to credit market sentiment, even when we separate out large firms from the rest of the sample.

We also examine heterogeneity across industries in the boom-and-bust patterns in investment and financing following irrational swings in credit market conditions. These results are important to understand the transmission mechanism driving the investment reversal that we document in the data. On one hand, this reversal could be due to an investment in durables that are abandoned and allowed to depreciate, as in the investment hangover channel of Rognlie, Shleifer, and Simsek (2018). In Rognlie, Shleifer, and Simsek (2018), overbuilding of durable capital, such as housing can induce a demand-driven recession with limited reallocation and low output when monetary policy is constrained. In this channel, the downturn in housing is essentially an overinvestment that gets corrected ex post. On the other hand, the investment reversal we document could also stem from an ex ante investment in non-durable capital that later fully depreciates and is not renewed when the time comes because credit has dried out. These channels are not necessarily mutually exclusive because irrational swings in credit conditions could affect investment in both durable and nondurable capital.

To attempt to shed light on these channels, we reestimate the tests in Tables 4 and 5 separately on the 10 different sectors of the economy, as classified by Fama and French (1997).¹⁸  Table D.19 in the Internet Appendix shows that the positive effects of a credit market shock in year t on investment in year t + 1 occurs in 9 of 10 sectors. (The exception is the healthcare, medical equipment, and drugs sector.) Conversely, the reversals in investment in years t + 4 and t + 5 following a credit boom in year t occur in all sectors but healthcare and nondurables (food, tobacco, etc). The sectors with the most pronounced cycles include consumer durables; oil, gas, and coal extraction and products; and the wholesale retail sector. Table D.20 in the Internet Appendix presents similar results for total net debt issuance. The results in Tables D.19 and D.20 show strong effects across very different industrial sectors, so both channels for investment reversals described above are likely at play in our data.¹⁹

3.3 Investment and financing cycles among financially constrained and unconstrained firms

Thus far, we have established a strong relationship between aggregated predictable firm-level forecast errors and subsequent corporate investment and debt issuance. These results show a supply effect in cycles of firm-level debt issuance and investment in tangible and intangible capital. In this section, we examine whether demand effects are also at play. To do so, we allow for firm-level heterogeneity in cycles of investment and debt issuance between financially constrained and unconstrained firms.

Differences in investment and debt-issuance cycles between financially constrained and unconstrained firms are informative. If macroPFE affects demand, managers irrationally increase investment and financing in response to irrational credit booms. In this case, we expect that the boom-and-bust pattern of corporate investment and financing that we document in Tables 4 and 5 should be equally present in both financially constrained and unconstrained firms (see, e.g., Bordalo, Gennaioli, and Shleifer (2018)). However, if macroPFE does not increase demand for credit (ie, if managers are rational), we expect that only financially constrained firms invest more in response to irrational credit booms. Whereas constrained firms could take advantage of improved credit conditions and reduce their underinvestment, financially unconstrained firms would already be investing at the optimal level. A similar explanation exists in the behavioral models in Stein (1996) and Baker, Stein, and Wurgler (2003), where rational managers optimally respond to irrational capital markets, and in the financial frictions literature (e.g., Bernanke and Gertler (1989) and Kiyotaki and Moore (1997)). However, in the financial frictions literature, shocks to capital supply are not due to irrational expectations, as in our data. In our framework, examining how constrained versus unconstrained firms react to macroPFE is crucial because at the core of the Minsky (1977) hypothesis rests irrationality, not only of investors supplying capital but also of firms demanding capital for their investment activity. If constrained and unconstrained firms react similarly to high levels of macroPFE, we have evidence of the irrationality on both the supply and demand side that plants the seeds of subsequent downturns, as Minsky (1977) describes.

We examine whether our results in Tables 4 and 5 differ for financially constrained firms relative to unconstrained firms using three measures of financial constraints: the Hadlock and Pierce (2010) index, the Whited and Wu (2006) index, and the Kaplan and Zingales (1997) index. In panel A of Table 6, we present results from regressions similar to Equation (8) using total investment as the dependent variable, to which we add an interaction of macroPFE and a financial constraints dummy that equals one for firms in the top quintile of the Hadlock and Pierce (2010) index and zero for all other firms.²⁰ In panels B and C, we define financially constrained firms as those in the top quintile of the Whited and Wu (2006) and Kaplan and Zingales (1997) indexes, respectively.

Table 6

Predictable cycles in corporate investment: Conditioning on financial constraints

	Dependent variable: Total investment
	Year 1	Year 2	Year 3	Year 4	Year 5
A. Financial constraints measure is top quintile of Hadlock and Pierce (2010) index
macroPFE	0.050***	0.044***	0.011	–0.024**	–0.024***
	(5.583)	(5.055)	(1.202)	(–2.632)	(–3.042)
Financially constrained	–0.088	0.040	–0.093	0.336	–0.197
	(–0.510)	(0.249)	(–0.415)	(1.213)	(–0.990)
|$ \mathit{macroPFE}\times $|Financially constrained	–0.008	–0.0002	–0.010	–0.029	0.006
	(–0.262)	(0.027)	(–0.237)	(–0.823)	(0.160)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	57,597	52,648	47,150	42,139	37,643
R  2	.381	.315	.294	.289	.282
B. Financial constraints measure is top quintile of Whited and Wu (2006) index
macroPFE	0.053***	0.046***	0.010	–0.037***	–0.029***
	(5.650)	(4.992)	(0.959)	(–3.971)	(–3.214)
Financially constrained	0.167	–0.145	–0.334*	–0.158	–0.154
	(1.012)	(–0.889)	(–1.827)	(–0.861)	(–0.764)
|$ \mathit{macroPFE}\times $|Financially constrained	–0.035	0.013	0.039	0.056*	0.055
	(–1.365)	(0.463)	(1.504)	(1.821)	(1.631)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	52,458	48,015	42,964	38,336	34,169
R  2	.384	.321	.301	.297	.286
C. Financial constraints measure is top quintile of Kaplan and Zingales (1997) index
macroPFE	0.049***	0.043***	0.007	–0.031***	–0.019**
	(5.614)	(5.166)	(0.660)	(–3.614)	(–2.184)
Financially constrained	–0.169	–0.265*	–0.364**	–0.483***	–0.075
	(–1.185)	(–1.838)	(–2.335)	(–2.822)	(–0.541)
|$ \mathit{macroPFE}\times $|Financially constrained	–0.008	0.005	0.022	0.026	–0.036
	(–0.445)	(0.202)	(0.927)	(1.052)	(–1.296)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	57,567	52,624	47,130	42,122	37,628
R  2	.381	.315	.294	.289	.281

This table presents coefficient estimates from regressing total investment up to 5 years in the future, on macroPFE (measured as in Section 2) and measures of financial constraints. Aggregated predictable firm-level forecast error (macroPFE) is calculated as predictable firm-level forecast error (see Section 2 for details), weighted by change in total debt (as in defined in panel A of Table 2). We also add interactions between all controls and a dummy variable that equals one for firms that are financially constrained. We categorize a firm as financially constrained if it falls in the top quintile for the financial constraints index of Hadlock and Pierce (2010) in panel A, the Whited and Wu (2006) measure in panel B, and the Kaplan and Zingales (1997) measure in panel C. Column headings Year “k” (k = 1|$ \ldots $|5) mean the dependent variable is measured at time t + k, while all independent variables are measured at time t. All specifications include firm fixed effects, firm-level controls (Tobin’s q, Cash flow to assets, log total assets, Cash to assets, Book leverage, Sales growth, ROA, default probability) and macro-level controls (the Leading Economic Index from the Conference Board, the Jurado, Ludvigson, and Ng [2015] index of macro uncertainty, the Baker and Wurgler [2006] sentiment index, the default spread, the risk-free rate, and the policy uncertainty index of Baker, Bloom, and Davis [2016]), as well as the current and lagged level of the dependent variable. Standard errors are bootstrapped according to Abadie et al. (2023). t-statistics are reported in parentheses.

*

p<0.1;

**

p<0.05;

***

p< 0.01.

Across all panels of Table 6, we find that there is virtually no difference in investment behavior in response to a shock in macroPFE for financially constrained versus unconstrained firms. If there was a stronger cycle for financially constrained firms, relative to financially unconstrained firms, we would expect coefficients for the interaction between macroPFE and our measure of financial constraints to reflect a cycle similar to the cycle we see in the coefficients for macroPFE: positive for years t + 1 and t + 2 and then negative for years t + 4 and t + 5. We see no apparent cycle in the coefficients for this interaction in any panel. The only significant coefficient (at the 10% level) is in panel B, for tests using the Whited and Wu (2006) index to measure financial constraints, in year t + 4. However, this coefficient is positive, or the reverse of what would be expected in year t + 4 if the cycle in investment were stronger for financially constrained firms. Thus, the results in Table 6 show no evidence that financially constrained firms have a more pronounced cycle in investment related to macroPFE than financially unconstrained firms. Moreover, because investment cycles in Table 4 are strongest in physical investment, we provide additional tests for physical investment in Table D.21 in our Internet Appendix. These results show that there are no differences in cycles in physical investment between financially constrained and unconstrained firms.

Consistent with our investment results in Table 6, results in Table 7 show no difference between cycles of debt issuance for financially constrained and unconstrained firms. In Table 7, we present results from regressions similar to Equation (8) using total long-term debt issuance as the dependent variable (as in panel A of Table 5), to which we add an interaction of macroPFE and a financial constraints dummy. Again, we see no evidence of a boom-bust cycle in the coefficients for the macroPFE and financial constraints interaction across all three panels. The only significant coefficients for this interaction come in panel C, using the Kaplan and Zingales (1997) index to measure financial constraints. However, the coefficients for the financially constrained indicator variable and the interaction between this variable and macroPFE together suggest an inverse correlation between the financial constraints measure and debt issuance for all years (t + 1 to t + 5), rather than any evidence of a cycle. Moreover, the coefficients for macroPFE itself remain relatively constant for each year across the panels and between panel A of Table 5 and each panel of Table 7. This consistency suggests that the inclusion of the financial constraints measure and interaction do not affect our baseline boom-bust results in debt issuance. The lack of a difference in cycle between financially constrained and unconstrained firms suggests that demand effects are at play in the data and provide further evidence of irrationality in support of the Minsky (1977) hypothesis.

Table 7

Predictable cycles in debt issuance: Conditioning on financial constraints

	Dependent variable: Total debt issuance
	Year 1	Year 2	Year 3	Year 4	Year 5
A. Financial constraints measure is top quintile of Hadlock and Pierce (2010) index
macroPFE	0.242***	0.143***	–0.019	–0.112***	–0.090***
	(10.137)	(6.017)	(–0.852)	(–4.546)	(–3.547)
Financially constrained	–0.375	0.079	–0.145	0.750	0.251
	(–0.859)	(0.244)	(–0.318)	(1.164)	(0.402)
|$ \mathit{macroPFE}\times $|Financially constrained	0.017	–0.071	0.008	–0.067	–0.081
	(0.122)	(–0.974)	(0.122)	(–0.772)	(–0.868)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	56,282	51,506	46,763	41,809	37,425
R  2	.259	.217	.191	.177	.173
B. Financial constraints measure is top quintile of Whited and Wu (2006) index
macroPFE	0.253***	0.132***	–0.031	–0.126***	–0.094***
	(9.860)	(5.550)	(–1.268)	(–4.658)	(–3.176)
Financially constrained	0.620	0.625	–0.752	0.844	–0.444
	(1.487)	(1.371)	(–1.625)	(1.610)	(–0.738)
|$ \mathit{macroPFE}\times $|Financially constrained	–0.061	–0.004	0.091	–0.010	0.021
	(–0.883)	(–0.078)	(1.186)	(–0.175)	(0.169)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	51,321	47,017	42,677	38,113	34,040
R  2	.264	.222	.200	.190	.178
C. Financial constraints measure is top quintile of Kaplan and Zingales (1997) index
macroPFE	0.252***	0.151***	0.010	–0.097***	–0.101***
	(10.602)	(6.219)	(0.353)	(–3.919)	(–3.771)
Financially constrained	–0.142	–0.882**	–0.378	–0.630	–0.502
	(–0.357)	(–2.380)	(–0.900)	(–1.289)	(–1.094)
|$ \mathit{macroPFE}\times $|Financially constrained	–0.075	–0.114*	–0.183**	–0.156**	–0.023
	(–1.294)	(–1.787)	(–2.620)	(–2.229)	(–0.317)
Firm-level controls	Yes	Yes	Yes	Yes	Yes
Macro-level controls	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes
Observations	56,239	51,471	46,733	41,783	37,402
R  2	.261	.217	.192	.178	.173

This table presents coefficient estimates from regressing total net debt issuance up to 5 years in the future, on macroPFE (measured as in Section 2) and measures of financial constraints. Aggregated predictable firm-level forecast error (macroPFE) is calculated as predictable firm-level forecast error (see Section 2 for details), weighted by change in total debt (as in defined in panel A of Table 2). We also add interactions between all controls and a dummy variable that equals 1 for firms that are financially constrained. We categorize a firm as financially constrained if it falls in the top quintile for the financial constraints index of Hadlock and Pierce (2010) in panel A, the Whited and Wu (2006) measure in panel B, and the Kaplan and Zingales (1997) measure in panel C. Column headings Year “k” (k = 1|$ \ldots $|5) mean the dependent variable is measured at time t + k, while all independent variables are measured at time t. Total net debt issuance is the change in long-term debt (“dltt”+“dlc” in Compustat) scaled by lagged total assets. All specifications include firm fixed effects, firm-level controls (Tobin’s q, Cash flow to assets, log total assets, Cash to assets, Book leverage, Sales growth, ROA, default probability) and macro-level controls (the Leading Economic Index from the Conference Board, the Jurado, Ludvigson, and Ng [2015] index of macro uncertainty, the Baker and Wurgler [2006] sentiment index, the default spread, the risk-free rate, and the policy uncertainty index of Baker, Bloom, and Davis [2016]), as well as the current and lagged level of the dependent variable. Standard errors are bootstrapped according to Abadie et al. (2023). t-statistics are reported in parentheses.

*

p<0.1;

**

p<0.05;

***

p< 0.01.

4 Biased Expectations and Investment

In this section, we further explore the interplay of supply and demand mechanisms in the cross-section. We begin by noting that the Minsky hypothesis predicts a symmetric effect of credit cycles on debt and investment, in which cycles during credit booms should be the mirror image of cycles during credit contractions. More specifically, during credit booms, firms with the most optimistic EPS revisions at time t should experience larger increases in investment and debt in year t + 1 and larger reversals at longer horizons, relative to firms with the least optimistic (ie, most pessimistic) EPS revisions. Conversely, during credit contractions, firms with the least optimistic EPS revisions at time t should experience larger decreases in investment and debt in year t + 1 and larger rebounds (ie, larger upswings in debt and investment at longer horizons), relative to firms with the most optimistic (re., least pessimistic) EPS revisions.

To investigate these channels in more detail, we examine whether firms with high or low predictable firm-level forecast errors experience more pronounced cycles in investment and debt issuance, and we distinguish between times when macroPFE is low and times when macroPFE is high. When macroPFE is high, firms with the highest analyst forecast revisions should respond more to credit market sentiment. These firms should have a larger positive coefficient for macroPFE at t + 1 and experience a larger reversal in t + 3 and t + 4. Intuitively, these are firms for which there is overextrapolation of recent good news regarding their fundamentals, which should lead to higher short-term investment and debt issuance. Eventually, good news for these firms should dry up and these firms should experience a more pronounced reversal in both investment and debt issuance relative to firms with lowest analyst forecast revisions.

Conversely, when macroPFE is low, firms with the lowest analyst forecast revisions should respond more to the credit contraction, that is, have a larger negative coefficient for macroPFE at t + 1 and experience a larger positive reversal in t + 3 and t + 4. Intuitively, these are firms for which there is overextrapolation of recent bad news regarding their fundamentals, which should lead to lower short-term investment and debt issuance. Eventually, bad news for these firms should dry up and these firms should experience a more pronounced positive reversal in both investment and debt issuance relative to firms with highest analyst forecast revisions.

To demonstrate the above channels, we estimate the following specification:

$$\displaystyle \begin{array}{c} Y_{i,t+k}=\sum \limits_{j=1}^{3}\beta_{j}H\mathit{ighMacroPF}E_{t}\cdot I\left(T_{i,t}=j\right)\\ \quad +\sum \limits_{j=1}^{3}\beta_{j}L\mathit{owMacroPF}E_{t}\cdot I\left(T_{i,t}=j\right)+\delta X_{i,t}+\epsilon_{t},\end{array}$$

(9)

where |$ k=0,\ldots,5,\,T_{i,t}$| is the tercile of forecast revisions that firm i belongs to at time t, HighMacroPFE_t is a binary variable equal to one if macroPFE is in the highest tercile, and LowMacroPFE_t is a binary variable equal to one if macroPFE is in the lowest tercile. Forecast revisions are measured as |$ \mathit{firmPF}E_{i,t}$| from Equation (4), constituting the portion of firm-level forecast error that is predictable using the previous two consecutive forecast revisions. The |$ X_{i,t}$| term is a vector of controls that includes all the controls from the baseline regressions in Section 3.

Panel A of Table 8 shows results for regressions of physical investment for a specification only including the binary variables indicating macroPFE tercile, omitting interactions with firmPFE tercile indicators, to establish a baseline for subsequent results. We focus on physical investment because, as Table 4 shows, results for total investment represent a weighted average of investment in physical and intangible capital; focusing on only one measure of investment provides cleaner inferences. The results in panel A of Table 8 show that when macroPFE in year t is high (ie, high tercile macroPFE equals one), physical investment increases by approximately 5% and 3.2% of assets in years t + 1 and t + 2, respectively, and decreases by approximately 6.4%, 7.2%, and 2.8% of assets in years t + 3 through t + 5, respectively. Conversely, when macroPFE in year t is low, physical investment decreases by approximately 2.8% in both years t + 1 and t + 2 and increases by 9.6% and 17.0% of assets in years t + 4 and t + 5, respectively. Demonstrating the opposing cycles in physical investment that are predicted when macroPFE is high versus low is helpful for interpreting the results in panel B, in which we interact the binary macroPFE tercile variables with indicators for whether a firm is in the high or low tercile of firmPFE, as in Equation (9).

Table 8 presents results for the specification in Equation (9) with physical investment as the outcome variable. These results are consistent with our expectations and show more pronounced cycles in investment for firms in the highest tercile of firmPFE when macroPFE is high and for firms in the lowest tercile of firmPFE when macroPFE is low. When macroPFE is high, we expect that firms in the high tercile of firmPFE have higher investment in the short term and lower investment in the long term than firms in the low tercile of firmPFE. Thus, we expect the coefficient for the interaction of high macroPFE and high firmPFE is initially greater than the coefficient for the interaction of high macroPFE and low firmPFE, so the difference between these two coefficients is positive. Then, we expect the relation to flip—in years t + 4 and t + 5, in years where we have documented a reversal, we expect the difference between the coefficient for the interaction of high macroPFE and high firmPFE and the coefficient for the interaction of high macroPFE and low firmPFE is negative. Consistent with this expectation, results in Table 8 show a positive difference between these coefficients in years t + 1 through t + 3 and a negative difference in years t + 4 and t + 5. When macroPFE is high in year t, firms in the highest tercile of firmPFE increase investment by 8.5% and 12.6% relative to assets in years t + 1 and t + 2, while firms in the lowest tercile of firmPFE only increase investment by 5.3% and 6.2%, respectively. In year t + 2, firms in the highest tercile of firmPFE have increases in investment that are more than double the size of the increase in investment for firms in the lowest tercile of firmPFE. Differences in effects between the high and low firmPFE terciles are economically, albeit not statistically, significant. These are lower power tests than those we estimated earlier because, to focus on firms with measurable firmPFE, we cut our sample by more than half.

Table 8 also reports results consistent with symmetry when macroPFE is low. If macroPFE is low in year t, firms in the lowest tercile of firmPFE reduce physical investment by 10.0% of assets, while firms in the highest tercile of firmPFE reduce investment by only 4.6% of assets. The difference between the coefficient for the interaction between low macroPFE and low firmPFE and the coefficient for the interaction between low macroPFE and high firmPFE begins at -5.4% in year t and narrows to -1.1% and –0.6% in years t + 2 and t + 3, respectively. Then, firms in the low firmPFE tercile experience a greater increase in physical investment in years t + 4 and t + 5 than firms in the high firmPFE tercile. The difference between the coefficient for the interaction between low macroPFE and low firmPFE and the coefficient for the interaction between low macroPFE and high firmPFE is 8.9% and 11.3% in years t + 4 and t + 5, respectively. When macroPFE is low, firms with low firmPFE initially have lower investment than firms with high firmPFE, but firms with low firmPFE experience a greater reversal, or increase in physical investment, in years t + 4 and t + 5 than firms with high firmPFE.

Table 8

Predictable firm-level forecast error drives cycles in physical investment: Conditioning on predictable firm-level forecast error tercile

	Year 1	Year 2	Year 3	Year 4	Year 5
A. High versus low macroPFE
High tercile macroPFE	0.050***	0.032*	–0.064***	–0.072***	–0.028
	(2.867)	(1.754)	(–3.67)	(–3.990)	(–1.494)
Low tercile macroPFE	–0.028	–0.028	–0.010	0.096***	0.170***
	(–1.535)	(–1.596)	(-0.500)	(4.599)	(8.158)
Observations	57,597	52,658	47,712	42,664	38,160
R  2	.397	.351	.346	.335	.337
B. High versus low macroPFE conditioning on predictable firm-level forecast error tercile
High tercile macroPFE ×	0.085***	0.126***	–0.044	–0.043	0.032
high tercile firmPFE	(2.363)	(3.471)	(–1.260)	(–1.362)	(0.777)
High tercile macroPFE ×	0.053	0.062*	–0.067*	–0.013	0.039
low tercile firmPFE	(1.297)	(1.753)	(–1.852)	(–0.387)	(1.139)
Low tercile macroPFE ×	–0.100***	–0.060	–0.016	0.137***	0.199***
low tercile firmPFE	(–2.488)	(–1.444)	(–0.310)	(3.347)	(4.650)
Low tercile macroPFE ×	–0.046	–0.048	–0.010	0.050	0.070*
high tercile firmPFE	(–1.377)	(–1.200)	(–0.192)	(1.201)	(1.813)
High macroPFE × high firmPFE –	0.030	0.060	0.022	–0.027	–0.010
high macroPFE × low firmPFE	(0.629)	(1.427)	(0.480)	(–0.601)	–0.254)
Low macroPFE × low firmPFE –	–0.054	–0.011	–0.006	0.089*	0.113***
low macroPFE × high firmPFE	(–1.436)	(–0.263)	(–0.149)	(1.933)	(2.872)
Observations	20,153	18,944	17,635	16,030	14,559
R  2	.432	.378	.368	.346	.343

This table presents coefficient estimates from regressing total physical investment up to 5 years in the future on a binary variables equal to one if aggregated predictable firm-level forecast error (macroPFE) is in the high or low tercile for the sample and controls. In panel B, we interact the binary tercile indicators for macroPFE with binary variables equal to one if the firm was in the highest or lowest tercile, respectively, for firmPFE in the year and zero otherwise. Predictable firm-level forecast error (see Section 2 for details) is weighted by change in total debt (as in defined in panel A of Table 2). Column headings Year “k” (k = 1|$ \ldots $|5) mean the dependent variable is measured at time t + k, while all independent variables are measured at time t. The investment and Tobin’s q variables are measured as in Peters and Taylor (2017). Investment in physical capital is the change in physical capital divided by lagged total capital, where total capital is gross PPE (ie, physical capital) plus the sum of goodwill, capitalized R&D, and capitalized SG&A (ie, intangible capital). Tobin’s q is the market value of equity plus book value of debt divided by total capital. All specifications include firm fixed effects, firm-level controls (Tobin’s q, Cash flow to assets, log total assets, Cash to assets, Book leverage, Sales growth, ROA, default probability) and macro-level controls (the Leading Economic Index from the Conference Board, the Jurado, Ludvigson, and Ng [2015] index of macro uncertainty, the Baker and Wurgler [2006] sentiment index, the default spread, the risk-free rate, and the policy uncertainty index of Baker, Bloom, and Davis [2016]), as well as the current and lagged level of the dependent variable. Standard errors are bootstrapped according to Abadie et al. (2023). t-statistics are reported in parentheses.

*

p<0.1;

**

p<0.05;

***

p< 0.01.

Results in Table 9 show similar symmetric firm-level responses to macroPFE for debt issuance. Panel A of Table 9 shows opposing cycles in debt issuance are predicted for high versus low macroPFE in year t. If macroPFE is high in year t, debt issuance is higher in years t + 1 and t + 2 and lower in years t + 3 through t + 5. In contrast, if macroPFE is low in year t, debt issuance is initially low and rebounds by year t + 5. In panel B of Table 9, results show that when macroPFE is high, high firmPFE firms have a more pronounced cycle of issuance activity. High firmPFE firms initially have higher debt issuance than low firmPFE firms in years t + 1 through t + 4 and then lower debt issuance than low firmPFE firms in year t + 5. When macroPFE is low, low firmPFE firms have a more pronounced cycle of issuance activity. In year t, low firmPFE firms reduce investment by 12.4% of assets more than high firmPFE firms. However, by year t + 4 and t + 5, low firmPFE firms have higher debt issuance than high firmPFE firms, 6.4% and 8.6% of assets, respectively. The symmetry in results, more pronounced cycles for high firmPFE firms when macroPFE is high and low firmPFE firms when macroPFE is low, is consistent with our investment results in Table 8.

Table 9

Predictable firm-level forecast error drives cycles in total debt issuance: Conditioning on predictable firm-level forecast error tercile

	Year 1	Year 2	Year 3	Year 4	Year 5
A. High versus low macroPFE
High tercile macroPFE	0.265***	0.250***	–0.105**	–0.147***	0.026
	(5.877)	(5.380)	(–2.221)	(–3.019)	(0.431)
Low tercile macroPFE	–0.295***	–0.251***	–0.273***	–0.036	0.364***
	(–6.479)	(–5.071)	(–4.884)	(–0.620)	(5.837)
Observations	56,282	51,506	46,763	41,809	37,425
R  2	.257	.216	.191	.176	.173
B. High versus low macroPFE conditioning on predictable firm-level forecast error tercile
High tercile macroPFE ×	0.516***	0.527***	–0.116	0.049	0.079
high tercile firmPFE	(5.131)	(5.112)	(–1.195)	(0.485)	(0.736)
High tercile macroPFE ×	0.288***	0.282***	–0.172	–0.048	0.148
low tercile firmPFE	(2.754)	(2.622)	(–1.577)	(–0.389)	(1.180)
Low tercile macroPFE ×	–0.456***	–0.439***	–0.516***	–0.043	0.367***
low tercile firmPFE	(–4.413)	(–4.322)	(–4.396)	(–0.427)	(2.914)
Low tercile macroPFE ×	–0.331***	–0.393***	–0.431***	–0.112	0.276**
high tercile firmPFE	(–3.751)	(–4.069)	(–4.129)	(–1.101)	(2.524)
high macroPFE × high firmPFE –	0.226*	0.240*	0.062	0.096	–0.062
high macroPFE × low firmPFE	(1.906)	(1.948)	(0.518)	(0.713)	(–0.472)
low macroPFE × low firmPFE –	–0.124	–0.053	–0.082	0.064	0.086
low macroPFE × high firmPFE	(–1.218)	(–0.523)	(–0.677)	(0.497)	(0.683)
Observations	19,838	18,651	17,369	15,811	14,373
R  2	.266	.226	.202	.189	.189

This table presents coefficient estimates from regressing total total debt issuance up to 5 years in the future on a binary variables equal to one if aggregated predictable firm-level forecast error (macroPFE) is in the high or low tercile for the sample and controls. In panel B, we interact the binary tercile indicators for macroPFE with binary variables equal to one if the firm was in the highest or lowest tercile, respectively, for firmPFE in the year and zero otherwise. Predictable firm-level forecast error (see Section 2 for details) is weighted by change in total debt (as in defined in panel A of Table 2). Column headings Year “k” (k = 1|$ \ldots $|5) mean the dependent variable is measured at time t + k, while all independent variables are measured at time t. Total net debt issuance is the change in long-term debt (“dltt”+“dlc” in Compustat) scaled by lagged total assets. Tobin’s q is measured as in Peters and Taylor (2017), the market value of equity plus book value of debt divided by total capital. All specifications include firm fixed effects, firm-level controls (Tobin’s q, Cash flow to assets, log total assets, Cash to assets, Book leverage, Sales growth, ROA, default probability) and macro-level controls (the Leading Economic Index from the Conference Board, the Jurado, Ludvigson, and Ng [2015] index of macro uncertainty, the Baker and Wurgler [2006] sentiment index, the default spread, the risk-free rate, and the policy uncertainty index of Baker, Bloom, and Davis [2016]), as well as the current and lagged level of the dependent variable. Standard errors are bootstrapped according to Abadie et al. (2023). t-statistics are reported in parentheses.

*

p<0.1;

**

p<0.05;

***

p< 0.01.01

5 Theories of Credit Cycles and Investment

Our results highlight a robust positive correlation between irrational swings in credit market sentiment in year t and corporate investment in both tangible and intangible capital in year t + 1. In the longer term, the effect reverses: high irrational sentiment in year t is followed by a large and significant reduction in corporate debt financing and investment in years t + 4 and t + 5. Both booms and reversals are strong across the board and not limited to specific subsets of firms or industries.

In this section, we discuss the theories that are most directly consistent with these results. We place existing theories in two broad groups: those relying on the revision of (rational) expectations and some kind of financial frictions, and those relying on the revision of (biased) expectations, with or without financial frictions. Section 5.1 discusses theories of rational expectations; Section 5.2 discusses behavioral theories based on biased expectations; and Section 5.3 formalizes the preceding discussions in the context of the neoclassical Q-theory framework.

5.1 Rational expectations and financial frictions

The large literature on the macroeconomic role of financial frictions recognizes that exogenous shocks to prices or productivity, despite causing an immediate rational revision of expectations, may not generate an immediate adjustment of corporate borrowing and investment behavior in the presence of financial frictions.²¹ The seminal contributions of Bernanke and Gertler (1989), Kiyotaki and Moore (1997), and Bernanke, Gertler, and Gilchrist (1999) highlight three main channels through which financial frictions affect the macroeconomy. First, when agents are levered, temporary shocks can have persistent effects on economic activity because they affect the agents’ net worth, which takes time to rebuild. Second, shocks are directly amplified in the presence of leverage. Bernanke and Gertler (1989) and Carlstrom and Fuerst (1997) quantify these effects by building on the idea that collateral value is costly to verify when information is asymmetric. Third, Kiyotaki and Moore (1997) show that shocks are further indirectly amplified through intertemporal feedback loops. In Kiyotaki and Moore (1997), an increase (decrease) in prices generates an increase (decrease) in the net worth of levered agents, thereby relaxing (tightening) their collateral constraints. This leads to an increase (decrease) in investment and output, which further increases (decreases) these agents’ net worth.²² Together, these insights show that even relatively small shocks can have potentially large effects on the macroeconomy.²³

In these models, collateral constraints depend on asset values and are always binding, based on the idea that financial frictions generally prevent agents from investing up to the first best level. As a result, positive shocks to prices and collateral help agents invest closer to the first best. Furthermore, these models provide a justification for ex post policy interventions because after a positive shock, agents fail to internalize that their decision to borrow and invest will affect prices and future transmissions of the shocks.²⁴

Kocherlakota (2000) argues that the quantitative degree of amplification of these models is sensitive to the model parameterization and is ultimately insufficient to explain observed fluctuations. Therefore, after the financial crisis of 2008–2009, more recent macroeconomic models of financial frictions focus on providing nonlinear dynamics. Brunnermeier and Sannikov (2014) present a model in which constraints are binding only occasionally, so that, at the steady state, firms absorb moderate shocks easily by adjusting payouts. However, after an unusually large shock, firms can no longer adjust payouts and need to deleverage (ie, sell capital to cut down their exposures).²⁵ Similarly, Bianchi (2011) and Mendoza (2010) study international macro-finance models based on occasionally binding collateral constraints and externalities of individual borrowing decisions on prices. These models also generate strong state dependency—once the economy is in a crisis regime, even small shocks are subject to amplification, leading to significant endogenous risk.²⁶

In our setting, these models can rationalize why a credit market sentiment shock in year t should be followed by increased corporate borrowing and investment in year t + 1. These models have difficulty in rationalizing in a parsimonious way why aggregate shocks at time t do not just eventually die out, but instead generate a large and predictable reversal in corporate borrowing and investment in year t + 4 across the board for all types of firms and irrespective of financial frictions.²⁷ To be sure, these reversals could reflect subsequent exogenous shocks of the opposite sign, or could be due mechanically to a strong negative moving average component in credit market sentiment. However, these explanations pose two problems. First, they are not parsimonious, as they posit that the time-series structure of exogenous shocks closely mirrors the data patterns to be explained without specifying further falsifiable predictions. Second, these explanations neglect the fact that prior evidence shows a systematic, cyclical component in credit market sentiment, since a credit market sentiment boom in year t predicts both low returns in year t + 1 (Greenwood and Hanson 2013) and low aggregate economic activity in years t + 3 and t + 4 (López-Salido, Stein, and Zakrajšek 2017).

5.2 Biased expectations

Literature starting with Minsky (1977) and Kindleberger (1978) stresses the role of biased expectations in generating and amplifying financial market cycles and economic fluctuations. More recently, a set of theories emphasizes that credit market sentiment can affect investment exclusively through revisions of biased expectations. Greenwood and Hanson (2013) show that credit booms come with a deterioration of the credit quality of the average issuer of debt, and in the aggregate predict low subsequent returns to corporate bondholders. López-Salido, Stein, and Zakrajšek (2017) show that credit booms drive the aggregate mix of external financing and, in turn, subsequent aggregate fluctuations in economic activity. This evidence implies that a supply effect is at play in credit markets, although it does not examine expectations directly. This approach emphasizes that, rather than a sequence of idiosyncratic unexpected shocks of opposite signs, financial market instability features cyclical and predictable components as well as endogenous reversals.

In general, it is difficult to reconcile these cyclical and predictable components with rational expectations. Under rational expectations, one would expect that a credit boom with low average quality of debt issuance should be followed by higher subsequent credit risk and higher expected returns, which is the opposite of what the data show. Similarly, it is difficult to reconcile our findings with behavioral models exclusively focusing on demand shocks, for example, models where consumers face a preference shock as in Laibson (1997) or Barro (1999), or where agents are generally overoptimistic about the future. Besides featuring consumption as a prominent channel (as opposed to borrowing and investment as in our data), these models predict that an irrational upward shift in credit demand should come with higher interest rates and low debt issuance, which is the opposite of what the data show.

More recently, a small but growing number of studies have presented formal analyses of how behavioral biases affect economic activity also through a credit supply channel. Bordalo, Gennaioli, and Shleifer (2018) present a model of diagnostic expectations whereby agents overweight future outcomes that become more likely in light of current data (see also Barberis et al. (2015) and Greenwood, Hanson, and Jin (2019)). Greenwood and Hanson (2015) study investment boom-and-bust cycles and returns on capital in the dry bulk shipping industry and find that high current ship earnings are associated with high used ship prices and heightened industry investment in new ships, but forecast low future returns. In their model, firms overextrapolate exogenous demand shocks and partially neglect the endogenous investment response of their competitors. Bordalo et al. (2020b) and Maxted (2024) present general equilibrium analyses.²⁸

In these models, agents overextrapolate a shock to fundamentals too far in the future. After a number of subsequent realizations turn out worse than expected, agents abruptly revise their expectations downward, generating a reversal. In these models, a single shock to fundamentals generates both positive short-term boosts and endogenous longer-term reversals in economic activity. In our setting, these models explain why shocks to fundamentals should propagate through credit supply via biased expectations, so that when fundamentals turn out worse than expected, firms redesign their investment plans, triggering long-term reversals in investment, as shown in Tables 4 and 5 and in Tables 6 and 7. In the next section, we formalize these ideas in a Q-theory framework. We then explore the overextrapolation mechanism in more detail.

5.3 A Q-theory framework for investment cycles

In this section, we summarize the previous discussion within the context of the neoclassical Q-theory framework. We begin by laying out the baseline framework with rational expectations (Section 5.3.1) and solving the model (Section 5.3.2). Then, we consider an augmented rational expectations model with financial frictions (Section 5.3.3). Next, we consider diagnostic expectations (Section 5.3.4). After that, we proceed with a quantitative evaluation of these models by presenting impulse response functions (Section 5.3.6).

5.3.1 Baseline framework with rational expectations

Consider a firm run by a risk-neutral owner with an infinite horizon, who discounts the future by a factor |$ \beta \lt 1$|⁠. The firm’s output in period 1 is obtained by combining capital, K, and labor, L, using a constant returns to scale production function, |$ A_{t}K_{t}^{\alpha}L_{t}^{1-\alpha}$|⁠, with |$ \alpha \lt 1$|⁠. At the beginning of period t, the owner hires labor L_t at wage ω_t and makes decisions about investment during the period, I_t. The firm’s optimal policy in year t maximizes the expected present value of earnings:

$$\displaystyle \max \limits_{\{I_{s},L_{s},K_{s+1}\}_{s=t}^{\infty}}\mathbb{E}_{t}\left\{\sum \limits_{s\geq t}^{\infty}\beta^{s-t}\left[A_{s}K_{s}^{\alpha}L_{s}^{1-\alpha}-\omega_{s}L_{s}-I_{s}-C(I_{s},K_{s})K_{s}\right]\right\},$$

subject to the capital accumulation equation, |$ K_{s+1}=(1-\delta)K_{s}+I_{s}$|⁠, where δ denotes depreciation. We assume the commonly used quadratic investment adjustment costs:

$$\displaystyle C\left(I_{s},K_{s}\right)=\frac{\chi}{2}\left(\frac{I_{s}}{K_{s}}-\delta \right)^{2},$$

which allow for convex adjustment costs (⁠|$ \chi \gt 0$|⁠) as long as the |$ \frac{I_{s}}{K_{s}}$| ratio differs from its steady-state value, δ, and displays constant returns to scale. In the maximization problem above, the operator |$ \mathbb{E}_{t}(.)$| denotes the owner’s expectations conditional on available information at the beginning of year t, computed according to possibly biased beliefs. We allow for departures from rational expectations but restrict the analysis to beliefs that preserve the law of iterated expectations.²⁹

5.3.2 Solution of baseline model

The Lagrangian is:

$$\displaystyle \begin{align} \mathcal{L}& =\mathbb{E}_{t}\Bigg\{\sum \limits_{s\geq t}^{\infty}\beta^{s-t}[A_{s}K_{s}^{\alpha}L_{s}^{1-\alpha}-\omega_{s}L_{s}-I_{s}-C(I_{s},K_{s})K_{s}\\ & \quad \quad \ \ \ -q_{s}(K_{s+1}-I_{s}-(1-\delta)K_{s})]\Bigg\},\end{align}$$

and the first-order conditions are:

$$\displaystyle \frac{\partial \mathcal{L}}{\partial L_{t}}=0\Leftrightarrow \quad \left(1-\alpha \right)A_{t}K_{t}^{\alpha}L_{t}^{-\alpha}=\omega_{t},$$

(10)

 

$$\displaystyle \frac{\partial \mathcal{L}}{\partial I_{t}}=0\Leftrightarrow \quad q_{t}-1-\chi \left(\frac{I_{t}}{K_{t}}-\delta \right)=0,$$

(11)

 

$$\displaystyle \begin{array}{c} \frac{\partial \mathcal{L}}{\partial K_{t+1}}=0\Leftrightarrow \quad q_{t}=\beta \mathbb{E}_{t}\left[\alpha A_{t+1}K_{t+1}^{\alpha -1}L_{t+1}^{1-\alpha}+\chi \frac{I_{t+1}}{K_{t+1}}\left(\frac{I_{t+1}}{K_{t+1}}-\delta \right)^{}\\ \quad -\frac{\chi}{2}\left(\frac{I_{t+1}}{K_{t+1}}-\delta \right)^{2}+q_{t+1}\left(1-\delta \right)\right]\\ TV\Leftrightarrow \quad \lim \limits_{T\rightarrow \infty}\beta^{T}\mathbb{E}_{t}\left[q_{t+T}K_{t+T+1}\right]=0.\end{array}$$

(12)

Then, we multiply both sides of Equation (12) by current capital stock, |$ K_{t+1}$|⁠. We use the capital accumulation equation |$ K_{t+1}=\frac{K_{t+2}-I_{t+1}}{(1-\delta)}$| to replace |$ K_{t+1}$| in front of |$ q_{t+1}$| and exploit constant returns to scale in output and investment costs. Under the standard definition of profits, |$ \Pi_{t}=A_{t}K_{t}^{\alpha}L_{t}^{1-\alpha}-\omega_{t}L_{t}-C(I_{t},K_{t})K_{t}-I_{t}$|⁠, we obtain the stochastic difference equation:

$$\displaystyle K_{t+1}q_{t}=\beta \mathbb{E}_{t}[\Pi_{t+1}+K_{t+2}q_{t+1}].$$

After iterating forward and imposing the transversality condition, we obtain the standard investment equation:

$$\displaystyle \frac{I_{t}}{K_{t}}=\delta -\frac{1}{\chi}+\frac{\beta}{\chi}\frac{\mathbb{E}_{t}[\sum \limits_{s\geq t+1}^{\infty}\beta^{s-(t+1)}\Pi_{s}]}{K_{t+1}}.$$

(13)

Equation (13) is the theoretical counterpart of Equation (8) and thus forms the basis of our empirical tests. In Internet Appendix A, we report the full solution of the baseline model with rational expectations, including the steady-state values of all variables.

5.3.3 Baseline model augmented with financial frictions

Next, we introduce financial frictions. As a starting point, we begin with Kiyotaki and Moore (1997), who introduced the concept of borrowing under collateral constraints, which is now popular in many applications. Specifically, we explicitly introduce borrowing, B_t, as an additional choice variable. Borrowing B_t generates tax advantages |$ \tau B_{t}$|⁠. As a result, absent constraints, the firm borrows and sets a capital structure with 100% debt. However, we constrain borrowing with the liquidation value of the firm’s physical assets. Specifically, we model collateral constraints by introducing a cost of borrowing, |$ C^{D}(B_{s},K_{s})$|⁠, as follows:

$$\displaystyle C^{D}\left(B_{s},K_{s}\right)=\phi_{0}e^{-{\phi_{1}}\cdot \left(\frac{\eta K_{s}}{B_{s}}-1\right)},$$

where η is the liquidation value of collateral as a fraction of its book value, K, with |$ \eta \lt (1-\delta)$|⁠. Distressed capital is thus sold at a discount, as in Shleifer and Vishny (1992).

This cost formulation (used in Croce et al. (2012), among others) convexities the occasionally nonbinding collateral constraint |$ B_{t}\leq \eta K_{t}$|⁠, which allows the firm to borrow up to the value of its collateral (ie, the liquidation value of its capital stock). In this formulation, the parameter |$ \phi_{1}$| is set (very) high to discourage the firm from borrowing more than the collateral value. Accordingly, the parameter |$ \phi_{0}$| is set (very) low so that the firm will choose |$ B_{t}=\eta K_{t}$| at the steady state. By modeling this constraint as a continuous and differentiable function, we can solve the model with standard numerical methods. We present the solution of the model with rational expectations and financial frictions in Internet Appendix B.

5.3.4 Rational versus biased expectations: An illustration

With regard to diagnostic expectations, we note that introducing diagnostic expectations in our setting generates additional cross-sectional predictions, relating forecast revisions and forecast errors to corporate investment and corporate borrowing. To see this, we begin by approximating Equation (13) with

$$\displaystyle i_{t}=b_{0}+b_{1}\mathbb{E}_{t}(\pi_{t+1}),$$

(14)

where lowercase letters indicate scaling by capital stock (ie, |$ \mathbb{E}_{t}\left(\pi_{t+1}\right)=\frac{\mathbb{E}_{t}[\Pi_{t+1}]}{K_{t+1}}$| and |$ i_{t}=\frac{I_{t}}{K_{t}}$|⁠). This approximation is reliable if expectations about the level of future earnings display significant persistence, namely, |$ \frac{\mathbb{E}_{t}[\Pi_{t+1}]}{K_{t+1}}$| is not too far from |$ \frac{\mathbb{E}_{t}[\Pi_{t+2}]}{K_{t+2}}$| and, more generally, from earnings far away in the future.

Assume now that profits follow an |$ AR(1)$| model such that |$ \pi_{t+1}=\rho \pi_{t}+\epsilon_{t+1}$| and |$ \mathbb{E}_{t}(\pi_{t+1})=\rho \pi_{t}$|⁠. By substituting these into Equation (14) and assuming rational expectations, we obtain:

$$\displaystyle i_{t}=b_{0}(1-\rho)+\rho i_{t-1}+b_{1}\rho \epsilon_{t},$$

which implies that an |$ AR(1)$| process for π_t translates into an |$ AR(1)$| process for i_t.

Now consider the diagnostic expectations formulation of Bordalo, Gennaioli, and Shleifer (2018):

$$\displaystyle \mathbb{E}_{t}^{\theta}(\pi_{t+1})=\mathbb{E}_{t}(\pi_{t+1})+\theta [\mathbb{E}_{t}(\pi_{t+1})-\mathbb{E}_{t-1}(\pi_{t+1})].$$

Under diagnostic expectations, it is then possible to show that

$$\displaystyle i_{t}=b_{0}(1-\rho)+\rho i_{t-1}+b_{1}\rho (1+\theta)\epsilon_{t}-b_{1}\theta \rho^{2}\epsilon_{t-1}.$$

Under diagnostic expectations, an |$ AR(1)$| process in π does not translate into an |$ AR(1)$| process in i_t. Diagnostic expectations introduce a moving average component, so that a positive realized shock to π, ϵ_t, translates into a positive spike in |$ i_{t+1}$| and also into a reversal in |$ i_{t+2}$|⁠. In other words, for θ = 0, we have the rational expectations case, in which an |$ AR(1)$| process in π translates into an |$ AR(1)$| process in i_t. In contrast, for |$ \theta \gt 0$|⁠, a moving average component appears (ie, the term multiplied by |$ \epsilon_{t-1}$|⁠). Consequently, for |$ \theta \gt 0$|⁠, we have both a larger investment boost in year 1 and reversal in year 2.

Following the logic of Bordalo, Gennaioli, and Shleifer (2018), we can formulate additional testable implications from this diagnostic expectations model. Define the forecast error at time t + k as |$ \mathbb{E}_{t}[\pi_{t+k}-\mathbb{E}_{t}^{\theta}(\pi_{t+k})]$| (ie, realized profits minus predicted profits, where the prediction is subject to bias θ) and the revision of expectations about profits π at time t + k, |$ 0\lt k\lt T$|⁠, as |$ \mathbb{E}_{t}[\mathbb{E}_{t+k}^{\theta}(\pi_{t+T})-\mathbb{E}_{t}^{\theta}(\pi_{t+T})]$| (i.e, forecast made at time t + k minus forecast made at time t). Then, it is possible to show that both forecast errors and forecast revisions at time t + k are predictable in light of information held at time t. In Internet Appendix C, we show that the revision of forecasts, |$ \mathbb{E}_{t}[i_{t+1}-\mathbb{E}_{t}^{\theta}(i_{t+1})]$|⁠, is such that

$$\displaystyle \mathbb{E}_{t}[i_{t+1}-\mathbb{E}_{t}^{\theta}(i_{t+1})]=-b_{1}\theta \rho^{2}\epsilon_{t}.$$

Thus, positive news about profits today make the firm invest more tomorrow and increase predicted profits tomorrow. However, the realized profits tomorrow are systematically smaller than predicted. Similarly, we can derive:

$$\displaystyle \mathbb{E}_{t}[\mathbb{E}_{t+k}^{\theta}(i_{t+T})-\mathbb{E}_{t}^{\theta}(i_{t+T})]=-b_{1}\theta \rho^{T+1}\epsilon_{t}.$$

Again, positive news today about profits today increase expected profits in the future, and these expectations systematically steer away from realized profits going forward.³⁰

5.3.5 Risky debt and supply effect

So far, we have considered diagnostic expectations on the part of managers only. In this section, we also consider diagnostic expectations of credit investors, to study both demand and supply effects, which we show are at play in our data.

To do so, we introduce risky (ie, defaultable) debt. In steady state with risk neutrality, expected cash flow growth should equal the risk-free rate, r (ie, |$ \mathbb{E}[g]=r$|⁠). In addition, for risky debt, the following holds:

$$\displaystyle \mathbb{E}[g]=r=\pi_{s}\cdot r^{B}+\pi_{f}\cdot 0,$$

where |$ \mathbb{E}[g]$| is the expected cash flow growth by analysts and managers, π_s is the probability of the “good state” or “success” (in which creditors are repaid in full), and π_f is probability of the bad state or “failure” where |$ \pi_{f}=1-\pi_{s}$|⁠, in which we assume creditors are not repaid at all.

In this simple setting, there are no costs of default, so the interest rate that creditors need to be promised in equilibrium, r^B, is such that:

$$\displaystyle r^{B}=\frac{\mathbb{E}[g]}{\pi_{s}}=\frac{r}{\pi_{s}}.$$

The proof is straightforward. A riskless investment of B yields |$ r\cdot B$|⁠, while a risky investment yields |$ B(1+r^{B})-B=r^{B}B$| with probability π_s and zero otherwise. In equilibrium without risk aversion, the two must equal each other. As a result, |$ r\cdot B=\pi_{s}\cdot [B(1+r^{B})-B]$|⁠, which yields the result.

If credit markets’ expectations are unbiased, then π_s is the true probability of the good state. If, however, as we document empirically, managers’ expectations are possibly biased, |$ \mathbb{E}[g]=\mathbb{E}^{\theta}[g]$|⁠. Assuming demand and supply to be neither perfectly elastic nor fully inelastic (ie, assuming demand is downward sloping and supply is upward sloping) implies that unexpected good news about credit markets should move demand (much) more than supply. That is, |$ \mathbb{E}^{\theta}[g]$| should increase and π_s should remain constant. Consequently, if only managers hold diagnostic expectations, then in good times interest rates should go up and borrowing should go down. This effect is the opposite of what the data show. On the other hand, if credit market expectations, |$ \pi_{s}=\pi_{s}^{\theta}$| is larger than the true probability, |$ \overline{\pi}_{s}$|⁠. In this case, when sentiment is high (ie, |$ \pi_{s}^{\theta}\gt \overline{\pi}_{s}$| and expected future default rates are then too low), then supply should also move, and interest rates can decrease in good times, as our data on debt issuance implies.

To incorporate this discussion into our framework, assume default happens if a negative shock makes productivity fall below a given threshold, |$ A^{\ast}$|⁠, so that |$ \pi_{s,t+1}=\underset{t}{\Pr}\left( \log \,A_{t+1}\gt \log \,A^{\ast}\right)$|⁠. Recall that productivity evolves according to |$ \log \,A_{t+1}=\rho \, \log \,A_{t}+\varepsilon_{t+1}$|⁠, implying that under rational expectations, |$ \mathbb{E}_{t}[ \log \,(A_{t+1})]=\rho \, \log \,(A_{t})$|⁠, and under diagnostic expectations, |$ \mathbb{E}_{t}^{\theta}[ \log \,(A_{t+1})]=\rho \, \log \,(A_{t})+\theta \rho \varepsilon_{t}$|⁠.

Under rational expectations,

$$\displaystyle \pi_{s,t+1}=1-\Phi \left[\frac{\, \log \,A^{\ast}-\mathbb{E}_{t}[ \log \,(A_{t+1})]}{\sigma}\right]=1-\Phi \left[\frac{\, \log \,A^{\ast}-\rho \, \log \,(A_{t})}{\sigma}\right],$$

while under diagnostic expectations,

$$\displaystyle \pi_{s,t+1}=1-\Phi \left[\frac{\, \log \,A^{\ast}-\mathbb{E}_{t}^{\theta}[ \log \,(A_{t+1})]}{\sigma}\right]=1-\Phi \left[\frac{\, \log \,A^{\ast}-\rho \, \log \,(A_{t})-\theta \rho \varepsilon_{t}}{\sigma}\right],$$

where |$ \Phi $| is the CDF of the standard normal distribution. To compute our impulse response functions, we pick |$ A^{\ast}$| such that the probability of success in steady state, |$ \pi^{\ast}$|⁠, is |$ \pi^{\ast}=0.95$|⁠, which yields |$ A^{\ast}= \exp \,[\sigma \Phi^{-1}(1-\pi^{\ast})]$|⁠.

5.3.6 Impulse response functions

To produce impulse response functions from the theoretical models discussed in the previous section and make them comparable to our empirical setting, we begin by abstracting from labor. Namely, we impose |$ L_{t}=\overline{L}=1$| for all t. Then, we use |$ \alpha =0.7$| (as commonly in settings with only capital and no labor), |$ \delta =0.15$|⁠, χ = 5, interest rate r = 0.04, and discount factor |$ \beta =1/(1+r)$|⁠. We assume productivity follows an |$ AR(1)$| process in logs, |$ \ln [A_{t}]=\rho \ln [A_{t-1}]+\epsilon_{t}$|⁠, with |$ \epsilon_{t}\sim N(0,\sigma^{2}),\,\rho \in [0,1]$|⁠, where we take |$ \rho =0.4$| and |$ \sigma =0.05$|⁠. Throughout this section, we produce impulse response functions using the Generalized Stochastic Simulation Algorithm (GSSA) developed by Judd, Maliar, and Maliar (2011).

Regarding the model with diagnostic expectations of Section 5.3.4, denoted by |$ a_{t}=\ln (A_{t})$|⁠, we define diagnostic expectations of productivity, |$ A_{t+1}$|⁠, consistent with Bordalo, Gennaioli, and Shleifer (2018), as:

$$\displaystyle \mathbb{E}_{t}^{\theta}[a_{t+1}]=\mathbb{E}_{t}[a_{t+1}]+\theta (\mathbb{E}_{t}[a_{t+1}]-\mathbb{E}_{t-1}[a_{t+1}]).$$

This specification has a number of convenient features. First, it nests rational expectations as a special case when θ = 0. Second, it implies overextrapolation of fundamentals when |$ \theta \gt 0$|⁠, consistent with psychological evidence. Third, it is a forward-looking formulation that preserves the law of iterated expectations. Fourth, as a result of the above it is immune to the Lucas critique. Fifth, it is a portable model of expectation formation in the sense of Rabin (2013).

Regarding the model with financial frictions of Section 5.3.3, in our calibration, we choose |$ \eta =0.33,\,\tau =0.35$|⁠, and |$ \phi_{1}=2000$|⁠, as is common in this literature (see, e.g., Croce et al. (2012)).

To provide an immediate visual mapping with our regression results, Figure 3 reports the impulse response functions of the investment-to-capital ratio, I/K, as generated by the above models in response to a shock to productivity. In the upper left quadrant we examine the baseline model with no debt. When managers hold rational expectations, the ratio I/K responds positively to a productivity shock in the first period and then decreases, turning negative after three periods. That is, the firm still invests a positive quantity, I > 0, but lower than the depreciation rate of capital, δ, which, in our model, also represents the steady-state value of I/K. As a result, |$ 0\lt I\lt \delta K$|⁠, and the firm becomes smaller. After that, the ratio I/K converges back to its steady-state level.³¹ We compare the IRF of this Q-model with rational managerial expectations with the IRF generated by the same model under diagnostic expectations, and we find that they generate a bigger fluctuation, about three-fold in our calibration.

Fig. 3

Impulse response functions

This figure plots the impulse response functions (IRFs) of neoclassical Q-theory under rational expectations (RE) and diagnostic expectations (DE). All variables are measured at time t = 0. Year “k” (k = 1|$ \ldots $|5) means the dependent variable is measured at time t + k. All IRFs are plotted as a percentage of their steady-state values. Hence, the economic magnitude (y-axis) can be interpreted as the change (expressed as % of the mean) in the dependent variable caused by a one-standard-deviation shock to TFP. The x-axis represent the steady-state value of the variable.

The upper-right quadrant examines the model with riskless debt of Section 5.3.3. Even with rational expectations, this model generates much larger fluctuations than the baseline model. However, this model carries the cross-sectional implication that the same firms facing more financial frictions should experience larger fluctuations. This is not what we see in the data—results in Tables 6 and 7 show that both financially constrained and unconstrained firms behave very similarly in response to the same shock in macroPFE. The implication is that financial frictions alone cannot explain our empirical results.

In the bottom two quadrants, we examine the model with risky debt, and we report the ratios I/K (bottom left) and B/K (bottom right). Both panels consistently show that considering diagnostic expectations of both managers and credit market investors generates a more than twofold fluctuation in investment and debt, respectively, than models considering either rational managers or rational creditors. These findings are consistent with our empirical results that show both demand and supply effects are at play in the data.

In sum, the results of this section indicate that a model with diagnostic expectations is the only deviation from standard Q-theory that can rationalize our empirical results. Our results also demonstrate that it is important to model the diagnostic expectations of both managers and creditors. In our concluding section, we draw the implications of our findings for theory and policy.

6 Conclusion

In a landmark contribution, Minsky (1977) hypothesized that capital market fluctuations are due to irrational swings in expectations. To assess this hypothesis systematically, we construct an aggregate index of irrational expectations using predictable firm-level forecast errors and use it to provide three sets of results. First, we show that irrational expectations drive aggregate credit cycles. We then find that our index of aggregate predictable firm-level forecast errors drives cycles in firm-level debt issuance and investment. Finally, we show that when credit market sentiment is high, firm-level financing and investment cycles are more pronounced for firms with ex ante more optimistic expectations. Crucially, financial constraints do not have additional explanatory power for firm-level cycles in the cross-section once irrational expectations are considered. We rationalize these results within a parsimonious dynamic Q-theory model with risky debt in which both corporate managers and credit investors hold diagnostic expectations.

Assessing the role of irrational swings in expectations is crucial to determine whether unstable capital markets have the potential to self-stabilize or whether monetary authorities need to take corrective actions. Following the stock market crash of 1929, the experience of the Great Depression of the 1930s has taught us that a tight monetary policy can greatly amplify the adverse effects of a capital market downturn (Friedman and Schwartz 1963). Building on this insight, recent policies of quantitative easing following the financial crisis of 2007–2008 have clearly avoided a repeat of the economic slump of the 1930s (Bernanke 2023).

It remains unclear, however, whether monetary policy should be similarly proactive, albeit in the opposite direction, during periods of capital market booms. Alan Greenspan and others have famously argued in favor of minimal policy interventions to allow markets to self-regulate in the face of booming capital markets. This view is clearly correct whenever an excess demand for financing by over-optimistic managers is met with an increase in the cost of capital brought about by rational capital suppliers. This view is also correct whenever excessively low credit spreads are exploited by rational managers who issue debt and use the proceeds to buy back shares, restoring stability in capital markets while doing so.

Our results imply that buoyant capital markets often go hand in hand with overextrapolation by managers who respond to low credit spreads with excessive debt issuance and excessive real investment. Thus, financial market booms plant the seeds for their own demise, bringing about subsequent reversals and further instability down the road. Overextrapolation not only blurs the traditional forces of supply and demand but also obfuscates the workings of otherwise rational arbitrage. In particular, our results suggest that the presence of financially unconstrained firms that overwhelmingly keep issuing securities and investing in the face of low spreads and unclear investment opportunities represents a clear sign of overheating in capital markets. Such overheating may warrant close scrutiny by monetary authorities, with a view of considering undertaking corrective actions.

More broadly, our results indicate the need to incorporate biased beliefs in a realistic theory of business cycles. Two facts point us in this direction, namely, the strong empirical role of irrational expectations, both in the aggregate and in the cross-section, and the fact that financial frictions do not have additional explanatory power once irrational expectations are considered. These pieces of evidence indicate that a realistic theory of belief formation should be a key ingredient of a realistic theory of business cycles.

Code Availability

The replication code for this paper is available in the Harvard Dataverse at: https://doi.org/10.7910/DVN/KPAM5B.

Acknowledgement

This paper has benefited from comments and discussions with Pedro Bordalo, Stefano Cassella, Scott Cederburg, Max Croce, Zhaojing Chen, Sergey Chernenko, Janet Gao, Nicola Gennaioli, Sam Hanson, Yeejin Jang, Bige Kahraman, Tommaso Monacelli, Beau Page, Nicola Pavoni, and James Reeder, as well as conference and seminar participants at the NBER Behavioral Finance Meetings, the American Economic Association Meetings, Bocconi, Maastricht, Rotterdam, and Tilburg. All errors are our own. Supplementary data can be found on The Review of Financial Studies web site.

Footnotes

1

Our main measure of CMS is the “ISS” index from Greenwood and Hanson (2013), calculated as the difference in average default probability between firms with the highest and lowest net debt issuance. We present, in our Internet Appendix, results using the alternative measure in Greenwood and Hanson (2013) (“HYS”, or the share of nonfinancial corporate bond issuance each year with a high-yield rating from Moody’s) and credit spread (the difference between Moody’s Baa corporate bond yield and the Aaa yield). These results are qualitatively similar to the ones in the main text.

2

In particular, we show that expectations-driven changes in credit supply generate large fluctuations in excess of those generated by expectations-driven changes in credit demand, in line with the evidence in Greenwood and Hanson (2013) and our own results. More research is needed to assess the extent to which credit mispricing is due to biased expectations of fundamentals, such as default rates, or excessive extrapolation of past market conditions.

3

The starting year of 1990 reflects the availability of expectations data needed to compute our variable macroPFE. The ending year of 2017 reflects the availability of the measures of investment developed by Peters and Taylor (2017) and of the Baker and Wurgler (2006) equity sentiment index.

4

Note that our sample size decreases to around 56,000 observations in our main firm-level tests, because our empirical design requires us to control for lagged levels of the dependent variable.

5

See the results in Table D.7.

6

Correlations between CMS and auxiliary measures of credit market sentiment that we use in robustness work are given separately in the Internet Appendix  Table D.4.

7

Projections from credit market analysts are less helpful to capture overenthusiasm as Minsky (1957, 1977) describes for three reasons: (1) credit market analyst reports are typically hand-collected and processed, requiring substantial resources and subjective defining and classifying of forecasted variables, (2) reports from credit market analysts typically cover a smaller set of firms and cluster in time around ratings events, creating a problematic source of selection bias in a study of credit market sentiment, and (3) information from credit analyst reports is harder to directly tie to the theories in Minsky (1957, 1977) than firm cash flows, including, for example, projected timing of ratings downgrades.

8

While Bouchard et al. (2019) find a strong positive relation between 1-year forecast errors and forecast revisions, it is important to note that they use short-term forecasts to construct these variables, while we use long-term forecasts.

9

Bordalo et al. (2019) examine horizons from 3 to 5 years, so we take the midpoint.

10

This correlation is even higher in subsamples. In Table D.4, we limit our sample to 1990 to 2008 and find a 0.88 correlation between macroPFE and CMS.

11

To estimate firm-level credit spreads and EBP, we use data on bond characteristics from the Mergent FISD database and data on bond prices from the Lehman/Warga database (prior to 1997), Bank of America Merrill Lynch (from 1997 to 2002), and TRACE (post-2002). When we take monthly averages of our credit-spread and EBP measures, we obtain macro-level indices that have correlations of 98% and 97% with the aggregate bond-level credit spread and EBP measures that Gilchrist and Zakrajšek (2012) make available.

12

Firm-level expected default rates are measured as in Bharath and Shumway (2008).

13

This expectation is consistent with the negative relation between aggregate EBP and our macroPFE index because firmPFE enters with a negative sign in the construction of macroPFE.

14

Replication materials for Schularick and Taylor (2012) are available here: https://www.openicpsr.org/openicpsr/project/112505/version/V1/view.

15

Controlling for lags of the dependent variable in the presence of firm fixed effects introduces a potential “Nickell bias” in our estimates (Nickell 1981). We believe the bias is small, since we have 28 years in our time-series tests using macroPFE. Nevertheless, we verify that our results are qualitatively the same if we reestimate all our specifications using the “system GMM” approach introduced by Arellano and Bover (1995) and Blundell and Bond (1998).

16

Gulen and Ion (2016) find that elevated policy uncertainty in the aggregate correlates negatively with corporate investment at the firm level.

17

In unreported results, we obtain similar results under alternative proxies for credit quality, such as the Campbell, Hilscher, and Szilagyi (2008) index, the Ohlson (1980) O-index, and the Altman (1968) Z-score.

18

Fama and French (1997) originally classify firms in 12 sectors. Our exclusion of utilities and financials is consistent with the rest of our analysis.

19

Our findings also compare with those in Mian, Sufi, and Verner (2017), who document that an increase in the household debt to GDP ratio predicts lower GDP growth and higher unemployment, consistent with a household balance sheet channel in economic fluctuations. The findings in Mian, Sufi, and Verner (2017) are consistent with our findings of a boom-bust pattern in real estate. Unlike Mian, Sufi, and Verner (2017), we explicitly link the boom-bust patterns in corporate investment in real estate (and in all other sectors) to the revision of biased expectations. In addition, we find strong boom-bust patterns in almost all other industrial sectors, consistent with a corporate balance sheet channel in business cycle fluctuations.

20

For all measures of financial constraints, results are similar using top tercile or above median to define financial constraints. In this variation on Equation (8), we include interactions between the financial constraints dummy and all independent variables, including macroPFE, although we only report coefficients for macroPFE, the financial constraints dummy, and the interaction between the two.

21

In a different but related vein, Kydland and Prescott (1982) consider the presence of lags between investment plans and their realization, which alone can generate fluctuations in investment around a growth path.

22

This channel builds on the fire-sales mechanism of Shleifer and Vishny (1992), according to which when a firm in financial distress needs to liquidate assets, the natural purchasers—firms in the same industry—are likely financially distressed, too. As a result, demand for liquidated assets will be low, and the assets will trade at a fire-sale discount relative to their fundamental value.

23

Khan and Thomas (2008, 2013) and Ottonello and Winberry (2018) explore these dynamics in models with heterogeneous firms.

24

Scheinkman and Weiss (1986) and Cooley, Marimon, and Quadrini (2004) present related models in which borrowing constraints stem from enforcement frictions.

25

He and Krishnamurthy (2012) study a related model in which the aggregate capital of the intermediary sector represents a key state variable for determining macroeconomic and asset pricing patterns during the financial crisis.

26

New Keynesian analyses emphasize a distinct but related mechanism that involves deleveraging and aggregate demand externalities. For example, Schmitt-Grohé and Uribe (2017) present a model in which, in the presence of downward nominal wage rigidity, a Taylor-type interest rate feedback rule, and a zero lower bound on nominal interest rates, a confidence shock can generate a slump in investment (see also Korinek and Simsek (2016) and Eggertsson and Krugman (2012).

27

Capital adjustment costs in models without labor can generate some reversals in borrowing and investment due to the lumpy nature of adjustment. In the presence of labor, no reversals occur to the extent that wages adjust more smoothly than capital.

28

Related to this literature, Simsek (2013) provides theoretical analyses showing how variations in lenders’ beliefs have an important impact on leverage and outcomes.

29

Gennaioli, Ma, and Shleifer (2016) consider a similar setting and use it to estimate the effect of diagnostic expectations in the cross-section. We examine the dynamics of expectations and investment over the cycle. In addition, we also consider risky debt, and we allow for diagnostic expectations not only of managers but also of credit market investors.

30

Considering both financial frictions and biased beliefs jointly can help illustrate the underlying logic of constructing of our macroPFE index as the aggregation of firmPFE using debt weights. In such a setting, firms differ ex ante due to their debt capacity, η, and cannot borrow more than the pledgeable value of their current capital, ηK. In equilibrium, firms issue as much as they can given their collateral constraint, |$ B^{\ast}=\eta K$|⁠. When factoring in biased beliefs, it is still true that |$ \mathbb{E}_{t}^{\theta}[B_{t+1}]=\mathbb{E}_{t}^{\theta}[\eta K_{t+1}]$|⁠, so both biased beliefs and debt capacity together matter for credit and investment cycles—following a large positive shock |$ \epsilon_{t+1}\gt 0$|⁠, firms with biased beliefs overreact and invest more (both I and K increase) and borrow more, up to ηK. Firms’ debt capacity is unobservable, so, when constructing our macroPFE index, we make use of the equilibrium relation of our model, |$ B^{\ast}=\eta K$|⁠. Even if η is unobservable, we can capture these differential responses by firms to the same shock |$ \epsilon_{t+1}$| by using the observed level of debt issuance, B.

31

Interestingly, the fact that the baseline Q-theory presents this pattern whereby I/K crosses the steady-state level and then converges to it from below crucially depends on not having labor in the model. Intuitively, in the absence of labor adjustment costs, labor adjusts faster than capital absorbing much of the overall firm response to the exogenous shock. As a result, I/K converges to its steady-state level, δ, from above.

Author notes

Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online.