The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Federal Reserve Bank of Chicago Levered Returns and Capital Structure Imbalances Filippo Ippolito, Roberto Steri, Claudio Tebaldi, and Alessandro T. Villa January 8, 2022 WP 2022-42 https://doi.org/10.21033/wp-2022-42 Working papers are not edited, and all opinions and errors are the responsibility of the author(s). The views expressed do not necessarily reflect the views of the Federal Reserve Bank of Chicago or the Federal Reserve System. * Levered Returns and Capital Structure Imbalances Filippo Ippolito1 Universitat Pompeu Fabra and CEPR Roberto Steri2 University of Luxembourg Claudio Tebaldi3 Bocconi University Alessandro T. Villa4 FRB Chicago Abstract We revisit the relation between equity returns and financial leverage through the lens of a dynamic trade-off model with costly capital structure rebalancing. The model predicts that expected equity returns depend on whether a firm’s leverage is above or below its target leverage. We provide empirical evidence in support of the model predictions. Controlling for leverage, overlevered (underlevered) firms earn higher (lower) returns. A quantitative version of our model reproduces key facts about capital structure rebalancing and equity returns for U.S. corporations. Overall, our results indicate that financial flexibility crucially affects the link between leverage and equity returns. Keywords: leverage, cross section of returns, target leverage, dynamic capital structure, financial frictions. JEL Classification Numbers: G12, G32. ? We thank Thomas Geelen, Juan-Pedro Gomez, Minjoo Kim, Christian Laux, Gyongyi Loranth, Stefano Sacchetto, Lukas Schmid, Gustavo Schwenkler, Victoria Vanasco, Josef Zechner for valuable comments. We also have benefited from comments from seminar and conference participants at Universitat Pompeu Fabra, the Vienna Graduate School of Finance, the 2017 Financial Intermediation Research Society Annual Meeting, the 2017 Edinburgh Corporate Finance Conference, the 2017 4nations Cup, and the 2017 Annual Meeting of the Spanish Finance Association. Disclaimer: The views expressed in this paper do not represent the views of the Federal Reserve Bank of Chicago or the Federal Reserve System. Declaration of conflicts of interest: none. All errors are our own. 1 Barcelona Graduate School of Economics and CEPR, Universitat Pompeu Fabra. E-mail: filippo.ippolito@upf.edu. 2 Department of Finance at the University of Luxembourg. E-mail: roberto.steri@uni.lu. 3 Department of Finance at Bocconi University. E-mail: claudio.tebaldi@unibocconi.it. 4 Federal Reserve Bank of Chicago. E-mail: alessandro.villa@chi.frb.org. Preprint submitted to Journal of Monetary Economics January 8, 2022 1. Introduction The second proposition of Modigliani and Miller (1958) (MM) is one of the pillars of the theory of corporate finance. It is widely taught in corporate finance courses and employed by practitioners around the world. The proposition establishes a positive relation between financial leverage and expected equity returns that is ordinarily 5 applied to compute the cost of equity for different levels of leverage. Over the years, a number of academics have investigated the relation between equity returns and leverage, and have provided mixed results. In some cases, equity returns appear to be unrelated to leverage. In other cases, leverage is significantly and negatively related to equity returns. In this paper we formalize an intuition for why the fundamental result of MM can be at odds with the data. We 10 suggest that firms’ limited financial flexibility in adjusting their capital structure crucially affects the link between leverage and returns. We show that, in a dynamic setting, firms tend to partially adjust their capital structure toward a target leverage ratio. We offer empirical and quantitative evidence that accounting for capital structure imbalances, defined as deviations from target leverage, helps in coming to terms with the contradictory evidence on “levered returns”. 15 We begin by illustrating the intuition for our results in a simple two-period model. This stylized setup offers a “look-alike” MM equation for levered returns. In the model, firms choose their capital structure to maximize their equity value in the presence of three frictions: taxes, bankruptcy costs, and debt adjustment costs. In the absence of these three frictions, MM’s second proposition holds in its traditional form. Debt adjustment costs instead create room for capital structure imbalances. Firms have a motive to raise additional debt as long as the marginal tax 20 shield of an additional dollar of debt exceeds the marginal bankruptcy cost. The optimal financial policy of a firm is to increase its leverage as long as the margin between tax shields and bankruptcy costs is positive, and reduce its leverage otherwise. Absent adjustment costs, a firm changes leverage until the two marginal effects are perfectly equal. This happens when leverage is set equal to target leverage, which is defined as the optimum in the absence of adjustment costs. With adjustment costs, a firm generally finds it too costly to reach the target, and has to settle 25 for a leverage ratio that is either too low or too high with respect to the target. This is commonly referred to as “partial adjustment”, and it implies that in the cross section some firms are underlevered (leverage smaller than target), while others are overlevered (leverage larger than target). From an asset pricing perspective, a quasi MM result can still be obtained, once a correction factor γ is applied to the original equation. The resulting relation between market leverage and equity returns is the following:5 RE = RA + γd k − γd RA − RD γ , (1) 5 For a in-depth discussion of the use of market versus book values of debt in empirical tests of structural models, see Bretscher, Feldhütter, Kane, and Schmid (2020). 2 where k is the market value of the assets, d is the market value of debt, RA is the return on the unlevered assets, RD is the market return on debt, and RE is the levered equity return. 30 The correction factor γ reflects differences between average bankruptcy costs and average tax shields in correspondence of firms’ optimal capital structure. As adjustment costs induce capital structure imbalances, γ depends both on observed leverage and on how far a firm is from its target. The correction factor is equal to one in the frictionless benchmark, in which case we revert to the standard MM equation. In general, γ can be greater or smaller than one. More precisely, γ is smaller than one if the average (per-dollar of debt outstanding) tax shield exceeds 35 the average bankruptcy cost, and greater than one otherwise. In other words, for a firm with γ smaller than one, equity returns respond to leverage as if the firm had a debt load smaller than its market value d. In this case, the firm obtains higher average tax shields than bankruptcy costs and receives a net benefit from carrying debt. Instead, when γ is greater than one, equity returns behave as if the firm effectively had more debt than its market value d, because average bankruptcy costs are higher than average tax shields, and the firm bears a net cost of carrying its 40 outstanding debt stock.6 Observe that γ is different from one in the special case of static tradeoff theory, which does not include adjustment costs. Optimal capital structure choices only equate the marginal values of bankruptcy costs and tax shields, while their average values still differ in this case. This is because tax shields and bankruptcy costs, in general, vary non-linearly with firms’ debt. However, as all firms are be able to instantaneously respond to shocks by adjusting their capital structure to the target, capital structure imbalances play no role. 45 The key novelty of equation (1) with respect to the second proposition of MM is to show that both the amount and the unit cost (or benefit) of debt for a firm determine the effect of leverage on its equity returns. As in MM, capital structure affects how the risk of unlevered assets propagates to equity payouts through the amplification of both good and bad cash flow outcomes. We show that the amplification effect is also driven by the effective net cost (or benefit) of debt, which itself is driven by capital structure imbalances. By measuring how far a firm is from 50 its target, relative leverage is a measure of a firm’s capital structure imbalance. Relative leverage is defined as the difference between leverage and target leverage. It is negative for underlevered firms and positive for overlevered firms. γ and, consequently, equity returns depend positively on relative leverage. For a given leverage ratio, firms with higher relative leverage have higher equity returns, while firms with lower relative leverage earn lower equity returns. 55 The economic intuition for this result is the following. Take first the case of an underlevered firm and suppose that bankruptcy costs increase more than proportionally with the firm’s debt stock, while tax shields increase proportionally. As long as the marginal tax shield is greater than the marginal bankruptcy cost, every dollar of debt raised by the firm brings a net benefit because it generates a positive net margin. As the firm keeps raising more debt, the margin on an extra dollar of debt shrinks because the bankruptcy costs increase faster than the tax shields. 6 γ captures only the non-diversifiable components of bankruptcy costs and tax shields. Therefore, only the systematic component of bankruptcy costs and tax shields enters our mechanism. Thus, the effect of capital structure imbalances cannot be simply diversified away by shareholders through their portfolio allocations. 3 60 However, as long as the margin between tax shields and bankruptcy costs is positive on average, debt generates a net benefit for the firm. As a result, in comparison to the frictionless environment of MM, the amplification of leverage on returns is dampened. Vice versa, in the case of a sufficiently overlevered firm, debt brings a net loss to the firm, and the amplification effect of leverage on returns is stronger than in the original MM framework. Although this simple illustration is useful to develop intuition for our key results, it is too stylized to serve as 65 a basis for a quantitative investigation of its predictions. We then embed the key economic tradeoffs in a more general infinite-horizon dynamic environment, which we calibrate to U.S. data. The dynamic model allows for crosssectional firm heterogeneity in the form of productivity shocks. Every period, firms make endogenous investment and financing decisions. Firms are financially constrained as external equity is costly and firms can endogenously default on their debt claims. Within the model, we extend the definition of target in the two-period illustration 70 to a dynamic environment. In addition, we adopt a flexible asymmetric functional form for debt adjustment costs. As debt adjustment costs can arise from several sources, this choice allows us to keep the model tractable while capturing, albeit in reduced form, the direct and indirect costs of issuing and withdrawing debt.7 Within the calibrated model, we simulate artificial panels of firms and use them as a laboratory to test the empirical predictions. We complement our quantitative analysis with empirical evidence linking capital structure 75 imbalances and equity returns on a sample from the widely used CRSP and Compustat datasets. Our empirical evidence is based on standard reduced-form estimates of target leverage in the corporate finance literature. Taken together, our quantitative and empirical analyses provide a set of consistent empirical facts, which support the key predictions of our two-period illustration. First, firms partially adjust their capital structure toward their target leverage. Second, market leverage has very little explanatory power once standard controls, namely firm’s market 80 capitalization and book-to-market equity, are considered. Third, when relative leverage is included alongside leverage as an explanatory variable for equity returns, the coefficient of relative leverage is positive, statistically significant, and economically sizable. Instead, leverage has little predictive power. This suggests that heterogeneity in target leverage ratios confounds the relation between leverage and returns.8 Finally, in the cross section, we find that small firms adjust their capital structure faster than large firms toward their targets. Accordingly, the premium associated 85 with relative leverage is more pronounced for large firms, which appear to benefit less from adjusting toward their target capital structure. The paper is structured as follows. Section 2 develops the two-period illustration for our key results. Section 3 presents empirical evidence in support of the key model predictions. Section 4 lays out our infinite-horizon model, whose quantitative properties are assessed in Section 5. Section 6 concludes. 7 For example, Fischer, Heinkel, and Zechner (1989) and Strebulaev (2007) discuss underwriting and management fees, Acharya, Bharath, and Srinivasan (2007) study seniority issues that render large changes in the capital structure costly, and Diamond (1991) and Myers (1977) consider liquidation and agency costs associated with the use of debt. 8 When target leverage and market leverage are included in the same empirical specification, the coefficient of leverage is positive, while the coefficient of target leverage is negative. Thus, the traditional positive relation between leverage and returns is restored when one controls for target leverage. Intuitively, for an underlevered firm, the higher is the target, the larger are the net benefits of debt it internalizes at any given level of leverage. On the contrary, a higher target implies a smaller net cost of leverage for an overlevered firm. 4 90 Related Literature. This paper mainly relates to the literature that examines how financing frictions introduce deviations from the Modigliani-Miller theorem and ultimately determine cross-sectional spreads in expected equity returns. Recent papers along these lines include Livdan, Sapriza, and Zhang (2009), Gomes and Schmid (2010), Caskey, Hughes, and Liu (2012), Ozdagli (2012), Obreja (2013), Ozdagli (2015), Doshi, Jacobs, Kumar, and Rabinovitch (2019), and Friewald, Nagler, and Wagner (2021). Our model and the associated empirical findings provide 95 a number of new contributions. From a theoretical point of view, we reconcile the traditional intuition of MM’s Proposition 2 with a multi-period environment where leverage changes are costly. We also provide a rationale for the “non-result” of leverage on returns, i.e. the fact that leverage does not appear to drive equity returns (see Bhandari 1988, Fama and French 1992, Penman, Richardson, and Tuna 1992, George and Hwang 2010, Gomes and Schmid 2010). We show that once the value of debt is adjusted to account for the relevant frictions, the original insight 100 of MM’s proposition carries on through into a more complex set up than the one assumed in their original paper. Our theoretical findings rely on simple microeconomic trade-offs and have a clear intuition that fundamentally relies on the benefits and costs of debt finance. Our model shows that MM’s result can be obtained from the production side of the firm, rather than from a no-arbitrage argument.9 Finally, the model clarifies the interplay of different measures of leverage (observed leverage, target leverage, and relative leverage) and their individual relation with 105 equity returns. From an empirical point of view, this work is primarily related to Korteweg (2010), who uses equity prices to estimate the net benefits to leverage for a large cross section of firms in a dynamic tradeoff setting. Our results are consistent with Korteweg’s in that we provide additional evidence that the tradeoff between costs and benefits of leverage is quantitatively a first-order margin reflected in stock prices. We offer a measure of leverage that can be 110 consistently used to predict returns. We show that relative leverage can be computed with a two-step procedure using historical accounting data and that results are not sensitive to how relative leverage is measured. We provide empirical evidence that matches well to the predictions of the model, and we document that leverage per se is an unreliable predictor of returns. Another related empirical paper is Caskey, Hughes, and Liu (2012). Their analysis primarily relies on the “kink” proxy in Graham (2000) to define overlevered and underlevered firms depending on 115 variation in their marginal tax benefits. Remarkably, Caskey, Hughes, and Liu (2012) find that overlevered firms earn lower average cross-sectional return than underlevered firms. We entertain a different definition of overlevered and underlevered firms. Our empirical measures of target leverage, which are widely used in the empirical corporate finance, account for other dimensions than tax benefits, such as investment opportunities, innovation, and asset depreciation. In this context, we derive out testable hypotheses from a theoretical setup. Finally, we show that 120 our results are quantitatively relevant in a dynamic model, which is less susceptible to measurement error in target leverage. Our work builds on the macroeconomic literature that uses the “gap approach” to describe costly adjustments of production factors (e.g., Sargent 1978, Caballero and Engel 1993, Cooper and Willis 2004, King and Thomas 9 However, the absence of arbitrage opportunities in the market is still required. 5 2006, Bayer 2009). We rely on the “gap approach” to a generalize the definition of target leverage in the two-period 125 illustration to an infinite-horizon model. As in these studies, we consider adjustments toward a dynamic target, which they term “frictionless target”. We apply the gap approach to a firm’s debt stock, rather than to a factor of production. Reassuringly enough, our quantitative results based on the frictionless target are consistent with the empirical evidence based on reduced-form measures, which also define a dynamic target for each firm in the sample. More broadly, this paper pertains to the large and growing literature that links corporate investment and financing 130 decisions to the behavior of asset returns through the lens of production-based asset pricing models. An incomplete list includes Cochrane (1991), Jermann (1998), Gomes, Kogan, and Zhang (2003), Zhang (2005), Gomes, Yaron, and Zhang (2006), Liu, Whited, and Zhang (2009), Gomes and Schmid (2010), Belo, Lin, and Bazdresch (2014), Gomes, Jermann, and Schmid (2016), and Gomes and Schmid (2016). 2. A Simple Illustration 135 In this section we propose a stylized model to illustrate the relation between capital structure imbalances and expected equity returns. The main advantage of a stylized setup is to provide a transparent illustration of the underlying economic mechanism that drives our main empirical and quantitative results. 2.1. Economic Environment Consider an economy with three periods, t0 , t1 and t2 , populated by heterogeneous firms indexed by i. Firms have a set of physical assets ki that is assumed constant over the three periods. The assets generate after-tax profits A A equal to ri,1 (1 − τ )ki in t1 , where ri,1 is the pre-tax return on assets and τ is the statutory tax rate. Default takes A A , where r A is a given threshold. In t , the place when the return on assets is sufficiently low, such that ri,2 < ri,2 2 i,2 A assets produce ri,2 (1 − τi,2 )ki , where τi,2 0 = τ A A , if ri,2 < ri,2 A A . if ri,2 ≥ ri,2 A A The firm is currently in period t1 , and ri,1 is known with certainty. Instead, ri,2 is a random variable with conditional 140 A A distribution f (ri,2 |ri,1 ). Firms enter t1 with an outstanding amount of debt di,0 . In t1 the firm is assumed not to be in default and pays an D interest amount (1 + ρD i,1 )di,0 , where ρi,1 is a known coupon rate. The interest payments on debt in t2 are stochastic, D because they depend on whether default occurs or not. If there is no default, ri,2 (di,1 ) is equal to the coupon rate ρD i,2 (di,1 ). In the case of default, the return on debt is equal to the recovery rate D ri,2 (di,1 ) = A (1 + ri,2 (1 − τi,2 ))ki − βi,2 (di,1 ) − 1, di,1 6 (2) where the term βi,2 (di,1 ) ≥ 0 represents the bankruptcy costs per dollar of outstanding debt. Total bankruptcy costs are βi,2 (di,1 )di,1 . For analytical convenience, we assume that unit bankruptcy costs are proportional to each dollar of debt outstanding in the case of default, that is βi,2 (di,1 ) = βi,2 di,1 , where βi,2 β i = 0 A A , if ri,2 < ri,2 A A . if ri,2 ≥ ri,2 This implies that total bankruptcy costs increase more than proportionally (quadratically) with the stock of debt outstanding in case of default. We assume that adjusting the debt stock carries a cost Θi (di,0 , di,1 ) = θi 2 2 (di,1 − di,0 ) . The adjustment cost is a reduced form of accounting for the frictions associated with the issuance and repurchase of securities in an 145 environment with imperfect capital markets. The convexity of the cost function implies that large debt adjustments within one period entail larger unit costs than small adjustments, such as in Croce, Kung, Nguyen, and Schmid (2012) and Belo, Lin, and Yang (2014). Convex costs of adjusting capital structure are consistent with Myers and Majluf (1984) and Krasker (1986). Although the assumption of quadratic adjustment costs is analytically convenient, we relax it in the dynamic model of Section 4. The assumption on the existence of a cost of adjusting firms’ capital 150 structure is consistent with the empirical finding that firms do not adjust immediately to their target leverage ratio.10 Firm’s Problem. Tax shields and bankruptcy costs affect the cash flows of the firm and are thus reflected in the value of equity. In period t1 , firm i chooses di,1 to maximize the market value of the firm Vi,1 , that is:11 A D Vi,1 ≡ max ri,1 (1 − τ )ki − (1 + ri,1 (1 − τ ))di,0 + di,1 − Θi (di,0 , di,1 ) di,1 A D + E[M ( (1 + ri,2 (1 − τi,2 ))ki − (1 + ri,2 (1 − τi,2 ))di,1 − βi,2 (di,1 )di,1 )] , (3) where E[·] denotes the expectation operator with respect to the information set available at time t1 and M represents the stochastic discount factor between t1 and t2 as given by the market. To minimize notation, from now on we will refer to firm i more simply as the firm, and drop the subscript i. 155 2.2. Optimal Leverage Policy The presence of adjustment costs posits an economic tradeoff for all firms in the economy. Each firm optimally adjusts its capital structure by trading off sure cash flows at time t1 in exchange for expected cash flows at time t2 . 10 Fama and French (2002), Leary and Roberts (2005), Flannery and Rangan (2006), Huang and Ritter (2009), Faulkender, Flannery, Hankins, and Smith (2010), Flannery and Hankins (2010), and Halling, Yu, and Zechner (2015) provide examples of the estimation of partial adjustment models of capital structure. The adjustment speed toward the target is estimated to be as low as 7% (Fama and French (2002)) and as high as 30% ( Flannery and Rangan (2006)). 11 The maximization problem in (3) can be equivalently expressed in terms of change in debt, defined as ∆d i,1 ≡ di,1 − di,0 . To see this, replace di,1 with ∆di,1 + di,0 and maximize with respect to ∆di,1 . We can observe immediately that the maximization leads to the same first-order condition in (32). This observation highlights that, although we do not model debt maturity and seniority explicitly, what matters is the change in debt from one period to the next. 7 If debt claims are fairly priced in the market, that is E[M (1 + r2D (d1 ))] = 1, the first-order condition for the firm is " # ∂ r2D (d1 )d1 ∂ (β2 (d1 )d1 ) ∂Θ (d0 , d1 ) −E M = E M τ2 . ∂d1 ∂d1 ∂d1 (4) and has an intuitive interpretation. The firm finds the optimal debt level at the point where the change in adjustment costs resulting from raising one additional dollar of debt at time t1 (on the left-hand side) equalizes its expected discounted marginal net benefit at time t2 (on the right-hand side). The marginal net benefit is the difference between marginal tax shields and marginal bankruptcy costs. If there were no adjustment costs, the firm would perfectly 160 trade off the marginal tax shields with the marginal bankruptcy costs. Instead, the presence of adjustment costs implies that at the optimum, a wedge between marginal tax shields and marginal bankruptcy costs arises. The wedge can be either positive or negative. When the wedge is positive, the marginal tax shield is larger than the marginal bankruptcy cost, and the opposite holds otherwise. The optimality condition (4) can be rearranged to unfold the relation between capital structure imbalances and 165 expected equity returns. The following proposition first characterizes the firm’s optimal leverage policy. Proposition 1 (Optimal Financing Policy). Firms optimally set their leverage ratio λ1 of the gap between their initial leverage ratio d0 k and their target leverage ratio d1 d0 − = λ1 k k 2β(1−τ ) d0 d∗1 − k k d∗ 1 k d1 k to close a fraction , that is , (5) qD τ (rF +qD ) 1 F and d∗1 ≡ 2β(1−τ where λ1 ≡ θ+2β(1−τ1+r qD )qD . rF denotes the risk-free rate, with 1 + rF = E[M ] . qD ≡ ) 1+r F R r2A f (rA |rA )dr2A is the probability of default under the risk neutral measure, in which fQ (r2A |r1A ) ≡ (1+rF )f (r2A |r1A )M (r2A ) −∞ Q 2 1 is the risk-neutral probability density of the realization r2A . Proof. See Appendix. 170 Equation (5) characterizes the firm’s optimal financing policy as a partial adjustment model, consistent with the specifications ordinarily estimated in empirical studies, such as Flannery and Rangan (2006). The target d∗ 1 k is increasing with the marginal benefit of one dollar of debt τ and decreasing with its marginal bankruptcy cost β. The target is forward-looking in that firms anticipate future economic conditions at time t2 . Specifically, the numerator of d∗1 increases with the expected discounted value of the tax shields of debt. The larger the non-idiosyncratic 175 probability of default qD , the larger is the expected marginal tax shield τ (rF + qD ) in solvency states, because the tax-deductible credit spread that lenders charge over the risk-free rate is also larger.12 Instead, the denominator shows that the target decreases with the expected discounted value of marginal bankruptcy costs in default states, the per-dollar value of which is 2β and is realized with probability qD . The term 1 − τ in the denominator instead 12 As standard, the risk-neutral probability of default measures the non-diversifiable risk of bankruptcy (i.e., the part correlated with the stochastic discount factor), and thus the part priced by the market. 8 accounts for the indirect tax benefit of bankruptcy costs. Higher bankruptcy costs increase the coupon rate that 180 lenders charge to compensate for their lower recovery rate given default. The tax shield in solvency states also increases as a result of the higher required interest payment. Hence, the firm effectively bears only a fraction 1 − τ of the marginal bankruptcy cost. The target is equal to the optimum leverage that the firm would choose in the absence of adjustment costs.13 This definition of the target is consistent with the “gap” approach we use in Section 4 to define target leverage in 185 an infinite-horizon economy. Firms adjust their leverage ratio from gap d∗ 1 k − d0 k between initial leverage d0 k and target leverage d∗ 1 k d0 k to d1 k by closing a fraction λ1 ∈ (0, 1] of the . The parameter λ1 can be interpreted as the speed of adjustment of the firm toward its target capital structure. On the one hand, firms with low values of λ1 are slow in adjusting their capital structure to changes in their investment and financing conditions. On the other hand, firms with high speeds of adjustment are fast in closing the gap toward their targets.14 Observe that this illustration offers 190 a benchmark in which λ1 does not depend on k, i.e., the firm’s size does not affect the firm’s speed of adjustment. In the quantitative analysis of Section 5 we explore how firms’ adjustment toward the target are related to firms’ size. Proposition 1 shows the existence of capital structure imbalances in the economy. We define relative leverage RL1 as the difference between observed leverage and target leverage, that is: RL1 ≡ d∗ d1 − 1. k k (6) Accordingly, firms with RL1 > 0 are overlevered, i.e, their observed leverage is above the target, while firms with RL1 < 0 are underlevered. Equation (5) can be rewritten as: d1 d0 λ1 − =− RL1 . k k 1 − λ1 (7) From equation (7) we can see that if λ1 < 1 and in t1 a firm adjusts its leverage up from d0 k , the firm will be underlevered in t2 . Adjustment frictions (captured by λ < 1) will prevent the firm from reaching its target leverage d∗ 1 k in t1 . Symmetrically, if the firm deleverages in t1 , it will be overlevered in t2 . Figure 1 presents an example of the optimal leverage policy of the firm for two different values of θ that correspond to low adjustment costs (Panel A) and high adjustment costs (Panel B). The picture reports axis and d1 k d0 k on the horizontal on the vertical axis. Target leverage is at the intersection of the policy function with the 45-degree line. Firms that are already at their target do not adjust their capital structure (RL1 = 0). In the figure, the optimal 13 To see this, rewrite (4) assuming that there are no adjustment costs. This yields " # ∂ r2D (d1 )d1 ∂ (β2 (d1 )d1 ) E M τ2 =E M . ∂d1 ∂d1 Carrying out the necessary substitutions and solving for the optimal leverage ratio we obtain precisely the definition of target leverage as in Proposition 1. 14 When there are no adjustment costs, the speed of adjustment equals one. Equation (5) then states that 9 d1 k = d∗ 1 . k leverage policy is linear in d0 k , and the comparison between the two panels highlights that θ affects the slope of the optimal leverage policy. The higher θ, the steeper the policy function, and the more sluggish is the firm’s adjustment toward target leverage. To see this, consider two possible levels of d0 k , O LU 0 and L0 , located respectively below and above the target. The policy function in Panel B shows that a firm with initial leverage LU 0 leverages up less than its counterpart in Panel A and will be more underlevered in t2 . Analogously, all else being equal, a firm with initial leverage LO 0 lowers its leverage less, and is more overlevered if it faces higher adjustment costs. [Insert Figure 1 Here] 195 2.3. Levered Returns Since V1 is the cum-dividend equity value of the firm, its realized gross stock return R2E between t1 and t2 is defined as R2E = P2 + D2 D2 = , P1 V1 − D1 (8) where P1 and P2 are the (ex-dividend) stock prices at time t1 and at time t2 , and D1 and D2 are the dividend payments at time t1 and at time t2 . Notice that the (ex-dividend) stock price P2 at t2 is zero because there are no more cash flows after that time.15 The dividend payments of the firm at t1 and t2 are D1 ≡ (1 + r1A (1 − τ ))k − (1 + r1D (1 − τ ))d0 + d1 − Θ d0 d1 k , k D2 ≡ (1 + r2A (1 − τ2 ))k − (1 + r2D (d1 )(1 − τ2 ))d1 − β2 (d1 )d1 ). Exploiting the condition E[M( (1 + r2A (1 − τ2 ))] = 1 and the definition of D2 , after some manipulation, equation (8) can be expressed as16 R2E = R2A + γ(d1 )d1 k − γ(d1 )d1 R2A − R2D γ(d1 ) , (9) in which we define the after-tax gross return on physical assets as R2A ≡ 1 + r2A (1 − τ2 ), the effective cost of debt for the firm as R2D ≡ 1 + r2D (d1 )(1 − τ2 ) + β2 d1 , and the correction factor as γ(d1 ) ≡ E[MR2D ]. As E[M(1 + r2D (d1 ))] = 1, the correction factor is equal to γ(d1 ) = 1 − E[M (τ2 r2D (d1 ) − β2 d1 )]. (10) The expression in equation (10) has an intuitive interpretation. In a trade-off economy, the presence of tax shields τ2 and bankruptcy costs β2 d1 creates a discrepancy between the cost of each dollar of debt for the firm R2D and the market return on debt 1 + r2D (d1 ) earned by debt-holders. This discrepancy is the net benefit (or cost) to leverage D2 that, from (3), it follows that V1 − D1 = E[M D2 ]. Therefore, the equity return is given by R2E = E[M . D2 ] A that the condition E[M( (1 + r2 (1 − τ2 ))] = 1 can be derived under the assumption that all firms have already optimized their initial level of capital k0 , so that ∀i, ki = k0 . k0 is fixed at t = 1, 2. Note that if firms were allowed to re-optimize their capital stock at t = 1, firms would optimally choose to stay at k0 according to −1 + E[M( (1 + r2A (1 − τ2 ))] = 0. However, our two-period example simply offers an illustration based on a snapshot over a firm’s life cycle. The more general dynamic environment of Section 4 allows for cross-sectional firm heterogeneity, which translates into differences in firms’ target capital structures. 15 Observe 16 Note 10 (Korteweg, 2010) and is given by the term E[M (τ2 r2D (d1 ) − β2 d1 )], which is the present value of the average tax 200 shield τ2 r2D (d1 ) minus the average bankruptcy cost β2 d1 . As in MM, leverage propagates asset risk to equity payouts through the amplification of both good and bad cash flow outcomes. However, the amplification is not only driven by the amount of debt d1 , but also by the net benefit γ(d1 ) that shareholders internalize for each dollar of debt outstanding in correspondence with their optimal leverage and the resulting capital structure imbalance. The correction factor creates an amplification effect for the burden of debt when γ(d1 ) > 1. This happens when 205 the present value of average bankruptcy costs exceeds the present value of average tax shields. Instead, a reduction effect for debt occurs when γ(d1 ) < 1, which holds when average tax shields are greater than average bankruptcy costs. Levered equity returns in (9) are computed as if for each dollar of debt the firm effectively had γ(d1 ) dollars of debt because of its net benefits/costs to leverage. The effect of the discrepancy generated by τ2 and β2 (d1 ) is twofold. First, the correction factor γ(d1 ) enters the effective leverage ratio 210 ratio is larger than the observed leverage ratio R2D γ(d1 ) , d1 k−d1 . γ(d1 )d1 k−γ(d1 )d1 . When γ(d1 ) > 1, the effective leverage Second, the correction factor reduces the effective cost of debt because the additional leverage embedded in γ(d1 ) does not require any supplementary payment besides R2D . Importantly, γ(d1 ) captures the systematic component of tax shields and bankruptcy costs through their covariances with the pricing kernel M .17 Thus, the effect of capital structure imbalances cannot be simply diversified away by shareholders through their portfolio allocations. Intuitively, since bankruptcies are concentrated in bad times 215 while tax benefits are higher in good times, shares of overlevered firms are risky investments because their debt’ increases the present value of non-diversifiable bankruptcy costs and boosts the traditional MM amplification effect. We formally relate γ(d1 ) to capital structure imbalances in the following section. 2.4. Relative Leverage, Target Leverage, and Equity Returns. As discussed above, equation (9) offers a number of new insights into the relationship between leverage and 220 returns in the presence of frictions. The following proposition links ex-post realized equity returns and capital structure imbalances. Proposition 2 (Imbalances and Equity Returns). Realized equity returns of the firm between t1 and t2 are related to relative leverage and to target leverage through the following relationship R2E = R2A + db1 k − db1 b2D , · R2A − R bD ≡ where the effective debt stock db1 is defined as db1 ≡ γ ( ) d1 , R 2 γ d∗1 d1 , k k ≡1+ R2D γ( ) and the correction factor is defined as α d1 d∗ τ δD − α 1 + d1 2 k k k 17 Observe (11) (12) that, as in the case of MM, our predictions do not rely on a specific asset pricing model (e.g. multi-factor model), but hold for any pricing kernel M that ensures no arbitrage in the market. 11 or equivalently as α d1 τ δD d1 ≡1− + αRL1 + d1 γ RL1 , k 2 k k (13) qD 18 . with δD ≡ E[M (1 + r2A )|r2A < r2A ] and α ≡ 2β(1 − τ )k 1+r F 225 Proof. See Appendix. Taking expectations of both sides of (11) delivers the relation between ex-ante expected returns and capital structure imbalances at the heart of our analysis of levered returns, that is E[R2E ] = E[R2A ] + db1 k − db1 b2D ] . · E[R2A ] − E[R (14) It is worth reminding the reader that in both equations (11) and (14), leverage is expressed at market values, not at book values, because d1 and k, respectively, represent the market value of debt and of the assets. Therefore, the prediction of Proposition 2 is on the relation between leverage measured at market value and equity returns. 2.5. Empirical Predictions 230 On the basis of Proposition 2 and the interpretation of the correction factor, we can draw predictions on the relation between expected equity returns and our three measures of leverage (relative leverage, observed leverage, and target leverage). Consider two firms with the same observed leverage and different relative leverage. According to the definition of the correction factor provided in equation (13), the firm with higher relative leverage has a higher γ ( ). Plugging 235 equation (13) into equation (14), we then have that the firm with higher relative leverage earns higher expected returns. Instead, compare two firms with the same relative leverage and different observed leverage. Leverage enters equation (14) with a mixed sign because of several contrasting effects. First, the firm with higher dk1 has a lower γ RL1 , dk1 , as (13) shows. Second, since db1 ≡ γ RL1 , dk1 d1 in (14), it is unclear which of the two firms has a larger 240 amplification term c d 1 . c k−d 1 bD ≡ Finally, the effect of leverage on R 2 with higher leverage plausibly bears a higher cost of debt R2D , R2D γ (RL1 , is also indeterminate. Although the firm ) it has a lower correction factor as discussed before.19 d1 k Overall, the relation between leverage and expected returns is theoretically indeterminate. As a consequence, depending on which effect of leverage, if any, prevails in (14), leverage and returns could be positively, insignificantly, or negatively related in the data. 18 We use the shorthand E[f (X)|X < K] for E[f (X)1 {X<K} ], where X is random variable, f (·) is a function, 1{·} is an indicator function, and K is a constant. 19 As discussed in the online appendix, r D (d ) is increasing in d if investors are risk averse and bankruptcy states happen when M is 1 1 2 high (bad states). 12 245 Consider next two firms with the same observed leverage and different target leverage. The firm with higher target leverage must have a lower relative leverage because RL1 = d1 k − d∗ 1 k . Therefore, the conclusions obtained above for the relation between relative leverage and returns still hold with the opposite sign, as (12) shows. The firm with higher target leverage will earn lower average equity returns. Finally, compare two firms with the same target 250 and different leverage, a case that can be constructed by taking two otherwise identical firms that differ only in terms ∗ d of dk0 . Leverage enters with mixed signs both in the definition of γ k1 , dk1 and into equation (14). Therefore, the relation between leverage and expected returns is indeterminate even after controlling for target leverage. 3. Supporting Empirical Evidence This section provides empirical evidence linking capital structure imbalances and expected equity returns. This evidence is based on standard reduced-form estimates of target leverage in the corporate finance literature and to 255 support the following quantitative analysis. 3.1. Leverage Variables To compute the empirical counterparts of the key variables of the model, we define the market leverage ratio as M Li,t = Di,t , Di,t + M Ei,t (15) where Di,t denotes the stock of interest-bearing debt of firm i in period t and M Ei,t is the stock market capitalization of firm i in period t. M Li,t is the empirical counterpart of optimal market leverage in the two-period model. To see D this, observe that the market value of debt can be obtained as E[M (1 + rt+1 )]dt = dt , and similarly, that the market 260 A value of the assets is given by E[M (1 + rt+1 (1 − τt+1 ))]k = k. Defining the target leverage ratio as T Li,t , relative leverage is obtained as the difference between observed and target leverage, that is RLi,t ≡ M Li,t − T Li,t . (16) 3.2. Estimation of Target Leverage A natural implementation of the policy function obtained in equation (5) of Proposition 1 is the empirical model of Flannery and Rangan (2006) (FR), which is widely used in the corporate finance literature. In the model of FR firms partially adjust their leverage over time toward the desired level M Li,t at a speed of adjustment λ: M Li,t − M Li,t−1 = λ(T Li,t − M Li,t−1 ) + i,t , 13 (17) with T Li,t = βXi,t−1 . (18) T Li,t is modeled as a linear function of a set of firm-specific characteristics Xi,t−1 , and varies both over time and 265 across firms. Equations (17) and (18) lead to the following estimable model: M Li,t = (λβ)Xi,t−1 + (1 − λ)M Li,t−1 + i,t . (19) Differently from previous models employed in the literature (e.g. Hovakimian, Opler, and Titman 2001; Korajczyk and Levy 2003), the specification of FR allows for dynamic partial adjustment in the presence of frictions, and is thus consistent with the evidence provided by Leary and Roberts (2005) and Strebulaev (2007), according to which the existence of frictions prevents firms from instantaneously adjusting toward their desired capital structure. Excluding 270 M Li,t−1 from the right-hand-side of (19) is equivalent to assuming that a firm’s target leverage always coincides with its observed leverage, i.e., that there is full adjustment (λ = 1) in each period.20 The estimation of RLi,t requires three steps. First, we estimate equation (19) to obtain the coefficients for Xi,t−1 and M Li,t−1 , that are respectively given by λβ and (1 − λ). Second, we compute λ and β, and proceed to estimate T Li,t . Third, we compute RLi,t as the difference between observed M Li,t and predicted T Li,t . 275 For the estimation of equation (19), we use the Compustat annual database over the period 1965-2013 including all companies listed on AMEX, NYSE, and NASDAQ, and excluding firms that are not incorporated in the United States, financials (SIC codes 6000-6999) and utilities (SIC codes 4900-4999), because of their special characteristics. M Lit is defined as the book value of short-term plus long-term interest bearing debt (Compustat items DLTT+DLC) divided by the market value of assets (DLTT+DLC + PRCC F*CSHO). 280 Following FR, the set of control variables Xi,t−1 comprises the following variables. Profitability: EBIT (EBIT) over total assets (AT); Market Value over Assets: Book value of liabilities plus market value of equity (DLTT+DLC + PRCC F*CSHO) over total assets (AT); Depreciation: Depreciation (DP) over total assets (AT); Size: Logarithm of total assets (AT)21 ; Tangibility: Property, plant, and equipment (PPENT) over total assets (AT); R&D expenses: R&D expenses (XRD) over total assets (AT); R&D Dummy: Dummy equal to one for firms with missing values for 285 R&D expenses (XRD); Industry ML: Median industry ML calculated each year for two-digit SIC code industries; a dummy for each fiscal year; and a firm fixed effect. The control variables capture the importance of tax shields and bankruptcy costs, as well as possible additional determinants of target leverage identified by the empirical literature. 20 If λ = 1, equation (17) simplifies to M Li,t = T L∗i,t + i,t , that is E[M Li,t ] = E[T Li,t ]. 21 Assets are deflated by the consumer price index in 2000 dollars. 14 Importantly, as our estimates of target leverage are included in asset pricing tests, all our estimates need to be free of look-ahead bias. Thus, future information is never used in the formation of portfolios in the context of our 290 asset pricing tests. As the explanatory variables are lagged, our estimation of T Li,t begins in 1966. We require at least five observations for the estimation to be included in the sample, that means that the first estimates of T Li,t included in the sample are in 1970 for the set of firms with no gaps between 1965 and 1969. We then continue on a rolling basis each year using all the accounting information available until that year. In their review article on estimation techniques, Flannery and Hankins (2013) run a horse race between different 295 estimation procedures for a dynamic panel model like the one in equation (19).22 The evidence on which estimation technique performs better is mixed. As the FE approach is more computationally efficient in large datasets, we employ it for the estimation of equation (19).23 More precisely, we run a regression specification as in equation (19) that contains fixed effects for each individual firm. To guarantee reliable estimates of the firm fixed effects, we exclude firms for which there are less than five 300 years of data. We then compute the relative leverage as per equation (16).24 3.3. Descriptive Statistics Table 1 provides summary statistics for observed leverage (ML), target leverage (TL) and relative leverage (RL). Panel A shows that the estimated mean for the target is close to that of the observed market leverage (0.16). TL is highly persistent over time, as can be seen from the last column, which reports the autocorrelation of degree one. RL is less persistent than TL. The findings that RL has approximately a zero mean, that TL and ML have similar means, and that RL is not very persistent are all consistent with the idea that RL represents only a temporary deviation from the target. Panel B of Table 1 reports some key statistics for the whole Compustat sample for which leverage can be measured (first row) and for the subsample for which TL can be estimated. This subsample is fairly comparable to the full sample, but mildly biased toward larger firms. [Insert Table 1 Here] Table 2 displays a breakdown of the sample into quintiles of RL. We observe that RL and ML are positively related. Thus, firms that are overlevered also tend to have high leverage, and vice versa. Rather differently, RL and TL are not related monotonically: TL first decreases and then increases with RL, reaching a minimum in the third 305 quintile. This finding shows that over-leverage is not the result of having a low target, nor is under-leverage the 22 The comparison is across the following models: OLS, FE, Arellano and Bond (1991) difference GMM, Blundell and Bond (1998) system GMM, Huang and Ritter (2009) Four-Period Long Differencing, Hahn, Hausman, and Kuersteiner (2007) Longest Differencing, and Least Square Dummy Variable Correction. 23 For robustness, in unreported tests we run our estimates using Blundell and Bond (1998) system GMM and obtain qualitatively similar results. 24 For robustness, in the Online Appendix we consider three alternative measures of target leverage. 15 result of having a high target. Both the most overlevered and the most underlevered firms (fifth and first quintiles) have high targets. The fourth column reports the change in ML. Notably, overlevered firms tend to lever down in the next year, while underlevered firms tend to lever up. These findings are consistent with Figure 1 of FR, which shows that in the next year leverage drops for overlevered firms and it increases for underlevered firms. The largest 310 decrease in leverage is for the firms in the highest quintile of RL. As already noted above, due to the monotonic relation between RL and ML, these firms are also those with the highest level of ML. The fact that high leverage firms reduce leverage in subsequent years is consistent with the findings of Figure 2 of FR. Column 4 also shows that the change in leverage is faster for overlevered firms than for underlevered ones, consistent with the findings of Warr, Elliott, Koëter-Kant, and Öztekin (2012). The remaining columns report mean values of capital stock (K), 315 investment (I/K), profitability (ROA), market capitalization (SIZE), and Tobin’s Q (Q) for each of the quintiles of RL. Firms with higher relative leverage tend to have larger capital stocks and lower ROA.25 The relations between relative leverage and investment, market capitalization, and Tobin’s Q are non-monotonic. [Insert Table 2 Here] 3.4. Evidence on Returns and Capital Structure Imbalances In our asset pricing tests, we use monthly stock prices and returns of common shares of firms on NYSE, AMEX, 320 Nasdaq covered by the Center of Research in Security Prices (CRSP) from 1980 to 2013.26 De-listing returns are included in monthly returns. We drop observations if the trading status is reported to be halted or suspended. Following the standard procedure of Fama and French (1992), we match these monthly data to annual income statement and balance sheet data from the CRSP/COMPUSTAT merged database. More precisely, we match monthly prices and returns from July of calendar year t to June of calendar year t + 1 with data from each company’s 325 latest fiscal year ending in calendar year t − 1. We apply the same matching procedure to the rolling estimates of TL and RL described in Section 3. The matching procedure ensures a minimum gap of six months between fiscal year-ends and returns. The gap is meant to conservatively avoid look-ahead biases arising from using future accounting information to predict returns. For the purpose of our tests, we compute the natural logarithm of market capitalization (SIZE), and the natural 330 logarithm of BM equity. Market capitalization - defined as the product of a company’s stock price times the number of outstanding shares - is measured in June of calendar year t for the returns between July of calendar year t and June of calendar year t + 1. We measure BM equity as the ratio between a firm’s book equity and its market capitalization at the end of December of calendar year t−1. We compute book equity as the sum of shareholders’ equity and balance 25 Since RL is positively correlated with ML, this evidence is consistent with Gomes and Schmid (2010), who highlight that more levered firms are also larger and less profitable. 26 As discussed in Section 3 we implement rolling estimates of T L . Years from 1965 to 1980 are used as “burn in” period to obtain i,t sensible initial estimates. 16 sheet deferred taxes and investment tax credits if available, minus the book value of preferred stocks. Depending on 335 data availability, we estimate the book value of preferred stocks using, in this order, their redemption, liquidation and par value. Since we consider the natural logarithm of BM equity in our tests, we eliminate firms with negative book equity from our analysis. The six-month gap between predicting variables and returns Ri,t+1 from June/July of year t to June/July of year t + 1 implies that, in the model, leverage variables in year t are observed at the end of the six-month gap (June/July 340 of year t) and that Ri,t+1 spans from June/July of year t to June/July of year t + 1. 3.5. Portfolio Sorts In this section we perform a set of univariate portfolio sorts to examine how returns vary with both observed leverage and capital structure imbalances. To do so, we consider sorts across different quintiles of ML and our measure of capital structure imbalances RL. Table 3 reports equally-weighted excess returns in the left panel and 345 value-weighted excess returns in the right panel. The top row of both panels reports the realized returns (Re ) in excess of the risk-free rate for different portfolios sorted according to ML.27 In the panel with equally-weighted returns, the difference between the returns in the portfolio with the highest and the lowest ML is positive and equal to 6.09%. The Sharpe ratio of the high-minus-low strategy is 0.46. However, returns do not increase monotonically across columns. As ML increases across portfolios, returns first decrease and then increase. In the panel with value-weighted returns, 350 a similar pattern emerges. The difference between the returns of the portfolio with the highest and the lowest ML is positive, but returns do not increase monotonically. In this case, the Sharpe ratio of the high-minus-low strategy is 0.10. As for RL, across the two panels we report the portfolio sorts for RL and for equally-weighted and value-weighted excess returns. The difference in returns between the high and the low portfolio is positive, equal to 5.56% for value- 355 weighted portfolios, and 8.04% for equally-weighted portfolios. The growth in returns across portfolios is perfectly monotonic in some cases, while in others there tends to be a mild reduction in returns in the middle portfolios. The high-minus-low Sharpe ratios in the left panel are roughly double those obtained for the ML portfolios. Similarly, the high-minus-low Sharpe ratios for the value-weighted RL portfolios are approximately three times larger than those obtained for the ML sorts. [Insert Table 3 Here] 27 We obtain risk-free rate data from Kenneth French’s website. 17 360 3.6. Cross-Sectional Regressions Although sorts provide an immediate measure of the trading performance of simple long-short trading strategies involving RL, they might lead to portfolios with few or even no stocks and become unfeasible when multiple variables are considered at the same time. Since, as discussed in Section 2.5, our model offers predictions on the joint behavior of RL, TL, and ML, we use the Fama and MacBeth (FMB) methodology to provide direct estimates of their marginal 365 effects after controlling for the standard predictors of expected equity returns. Table 4 reports time-series averages of the estimated coefficients of monthly cross-sectional regressions of stock returns on size, book-to-market (BM), ML, TL and RL. We report FMB tests with a Newey-West correction with lag-length of 2 to assess which regressors have a coefficient that is significantly different from zero. In their most comprehensive specification, the FMB regressions of Table 4 take the following form: Ri,t = β0 + β1 M Li,t−1 + β2 RLi,t−1 + β3 T Li,t−1 + β4 log(Sizei,t−1 ) + β5 log(BM i,t−1 ) + i,t , (20) where Ri,t denotes realized returns, Sizei,t−1 market capitalization, and BMi,t−1 BM of equity. M Li,t−1 , RLi,t−1 and T Li,t−1 are never simultaneously included in the same regression specification, but M Li,t−1 is included together with RLi,t−1 in columns (4) and (5) and together with T Li,t−1 in column (6) and (7) to test the empirical predictions of Section 2.5.28 370 The leftmost column shows that leverage has a small positive and insignificant coefficient once we control for size, and BM. This “no result” for leverage shows that the results obtained in the univariate sorts of Table 3 are not robust to the controls. These slopes can be interpreted as the average monthly return of a self-financing portfolio with unit relative leverage, that hedges the effects of the controls in the sample period. In the FMB approach, the standard error is computed as the standard deviation of monthly returns on this portfolio, divided by the square root 375 of the number of months in the sample (408). Hence, the t-statistics for the coefficients of RL can be approximately translated into annualized Sharpe ratios of 0.6 (without controls) and 0.3 (with controls).29 Across all columns, size is negatively related to returns, while B/M is positively related to returns. It is important to remind the reader that the regressions in columns (1)-(3) are “model free”, in the sense that our empirical predictions are for regressions that contain at least two of the three leverage measures. We turn to 380 these specifications in columns (4)-(7). 28 Naturally, since RL i,t−1 = M Li,t−1 − T Li,t−1 , including all three variables at the same time would generate a multicollinearity problem. 29 The average monthly risk-free rate in our sample is approximately 39 basis points. 18 3.6.1. Relative Leverage and Leverage Columns (4) and (5) contain one of the main results of our analysis. Column (5) reports an FMB specification in which we include both ML and RL in addition to the controls. With respect to column (1), the coefficient of ML is now close to zero, and still not significant. Instead, the coefficient of RL is strongly significant, large, and positive. 385 The results of columns (4) and (5) are consistent with the prediction that RL is positively related to returns once we control for leverage. Instead, the relation between leverage and returns is undetermined because of contrasting effects. In the data, according to column (5), the positive effect of leverage on returns appears to dominate, although it remains not significant. 3.6.2. Leverage and Target Leverage 390 In columns (6) and (7) we report a set of regressions where we include ML and TL. Column (7) also includes the controls. The model predicts a negative relation between target leverage and returns, controlling for leverage. Indeed, this is what we observe in the data. The coefficient of TL is negative and statistically significant at the 1 percent level. After controlling for target leverage, observed leverage is positively and significantly related to returns. While consistent with the model predictions, the last result offers additional insights on the relative importance 395 of the contrasting effects of leverage on returns we described in Section 2.5. In conjunction with the estimates on columns (4) and (5), the positive coefficient of leverage after controlling for TL suggests that the term α d1 2 k in (12) has an important effect on the correction factor and, eventually, on returns. Indeed, a comparison of (12) and (13) highlights that all other terms in observed leverage that enter (14) are common to the specifications with RL and ML, and with TL and ML. Controlling for the target, observed leverage drives 400 c d 1 c k−d 1 up through the correction factor, and increases expected returns as a consequence. Ultimately, accounting for firm heterogeneity in firms’ target capital structure resurrects the traditional positive relationship between leverage and returns through an amplification effect a la MM.30 [Insert Table 4 Here] 30 From a purely empirical perspective, this result follows from the fact that (ignoring the controls and the error term), the specification Ri,t = β0 + β1 M Li,t−1 + β2 RLi,t−1 implies that Ri,t = β0 + (β1 + β2 ) M Li,t−1 − β2 T Li,t−1 , which is approximately what we observe in the data. 19 4. Dynamic Model The simple illustration in Section 2 provides intuition for our key results. However, the model is too stylized to 405 serve as a basis for a quantitative investigation of its predictions. In this section, we embed the key economic tradeoffs of our simple illustration in a more general dynamic environment. In the dynamic model, we allow for ex-post cross-sectional firm heterogeneity in the form of idiosyncratic productivity shocks to study the role of differences in target leverage across firms. We also introduce aggregate shocks, which affect firms’ profitability and discount rates. Every period, firms make endogenous investment decisions. This allows for a more realistic analysis of the 410 relationship between firm size, capital structure imbalances, and equity returns. Firms are financially constrained as external equity is costly and firms can endogenously default on their debt. Endogenous default creates a more realistic dependence of firm bankruptcy on the cost of borrowing. Finally, we adopt an asymmetric flexible functional form for debt adjustment costs. This allows for small adjustments to incur possibly large marginal costs while keeping the model tractable. Technology and Investment. Time is discrete. We consider the problem of a value-maximizing firm in a perfectly competitive environment. After-tax operating profits Πi,t for firm i in period t are given by α Πi,t = (1 − τ )(eAt eZi,t Ki,t − F ), (21) where τ ∈ (0, 1) is the corporate tax rate, At is an exogenous aggregate shock, Zi,t is a firm-specific shock, Ki,t is firm i’s capital stock, α ∈ (0, 1) is the capital share in production, and F > 0 is a fixed production cost. The variables At and Zi,t can be interpreted as shocks to demand, input prices, or productivity. At and Zi,t have bounded support A = [A, A] and Z = [Z, Z] , respectively. The law of motion of At is described by a Markovian transition function QA (At , At+1 ). Similarly, Zi,t is Markovian with transition function QZ (Zi,t , Zit +1 ). In our quantitative analysis, we parameterize At and Ki,t to provide a discrete approximation to continuous AR(1) processes as follows: log At = µA (1 − ρA ) + ρA log At−1 + σA εA t , log Zi,t = ρZ log Zi,t−1 + σZ εZ i,t . 415 Z As εA t and εi,t are truncated standard normal variables, both At and Zi,t are lognormal, with mean µA ∈ (−∞, ∞) and 0, persistence ρA ∈ (0, 1) and ρZ ∈ (0, 1), and volatility σA ∈ (0, ∞) and σZ ∈ (0, ∞), respectively.31 At the beginning of each period, firms can scale operations by choosing investment Ii,t . Next period’s capital stock Ki,t+1 satisfies the standard capital accumulation rule Ki,t+1 = (1 − δ)Ki,t + Ii,t , (22) where δ ∈ (0, 1) is the depreciation rate of capital. 31 As common in the literature, we set the mean of firm-specific shocks to zero (e.g., Hennessy and Whited 2007, Gomes and Schmid 2010). 20 Financing. Investment and distributions to shareholders can be financed with either the internal funds generated by operating profits or new issues. The latter which can take the form of new debt (net of repayments) or external 420 equity. We denote the firm’s debt stock as Bi,t . Outstanding debt pays a coupon ci,t per unit of time. As we detail below, the coupon is set in competitive credit markets to compensate expected bankruptcy costs in case of default, in which lenders recover a fraction ξ ∈ (0, 1) of the firm i’s capital stock. Firms are allowed to refinance their debt stock by issuing a net amount ∆Bi,t = Bi,t+1 − (1 + ci,t )Bi,t . Similar to Croce, Kung, Nguyen, and Schmid (2012) and Belo, Lin, and Yang (2014), firms incur costs Λ(∆Bi,t ) of adjusting their debt stock. In our quantitative analysis we choose a flexible parameterization for Λ(∆Bi,t ) to allow for asymmetries and possibly very large marginal costs for small adjustments. Debt adjustment costs can arise from different sources, including underwriting and management spreads (e.g., Fischer, Heinkel, and Zechner 1989, Strebulaev 2007), seniority issues that prevent firms from making large changes in the capital structure (e.g., Acharya, Bharath, and Srinivasan 2007), costs associated with the liquidation of short-term debt (Diamond 1991), agency costs associated with long-term debt (e.g., debt overhang and under-investment as in Myers 1977). Specifically, we choose a linear-exponential (LINEX) functional form (e.g., Varian 1975, Kim and Ruge-Murcia 2009, Aruoba, Bocola, and Schorfheide 2017), i.e., Λ(∆Bi,t ) = λB (eγB ∆Bi,t − γB ∆Bi,t − 1), 425 (23) with λB ∈ (0, ∞), and γB ∈ (−∞, ∞). In our context, the LINEX functional form is attractive for three main reasons. First, this flexible reduced-form functional form allows us to keep the model tractable without committing to a specific mechanism in our quantitative exercise. Second, as previous studies provide limited guidance on functional forms and economic magnitudes for debt adjustment costs, we can exploit data restrictions to calibrate them to realistic values in the context of our model. Third, the adjustment cost function is parsimonious. Varying only two parameters, the 430 “scale” λB , and the “asymmetry” γB , Λ(∆Bi,t ) spans a broad spectrum of functional forms.32 Firms can also raise external finance through seasoned equity offerings (SEOs). Let Ei,t denote the equity issuances. Following the extensive existing literature, we consider equity issuance costs Λ(Ei,t ) that include both a fixed and a proportional component, i.e., Λ(Ei,t ) = (λ0 + λ1 Ei,t )χEi,t >0 , (24) with λ0 ∈ [0, ∞) and λ1 ∈ [0, ∞). We interpret negative values for Ei,t as equity payouts. Investment financing decisions must satisfy the firm’s budget constraint, which takes the form of the following accounting identity between uses and sources of funds: Πi,t + ∆Bi,t + Ei,t + τ δKi,t + τ ci,t = Ii,t + ci,t Bi,t + Λ(∆Bi,t ), 32 The LINEX functional form nests the quadratic form as an approximation for γB → 0. 21 (25) where the terms τ δKi,t and τ ci,t reflect the tax deductibility of depreciation and interest expenses, respectively. Net distributions to shareholders, Di,t , are then defined as equity payout net of issuance costs, i.e., Di,t = −Ei,t − Λ(Ei,t ). (26) Valuation. Following several studies in cross-sectional production-based asset pricing (e.g., Berk, Green, and Naik 1999, Zhang 2005, Gomes and Schmid 2010), we parameterize the stochastic discount factor of the economy without explicitly modeling the investor’s problem. As in Zhang (2005), we assume the following stochastic process for the stochastic discount factor: logMt = logβ + γt (At − At+1 ), (27) where β ∈ (0, 1), γt = γ0 + γ1 (At − µA ), γ0 ∈ (0, ∞), and γ1 ∈ (−∞, 0). This functional form naturally links to the time-varying risk aversion in Campbell and Cochrane (1999), in which γ0 is interpreted as a “risk aversion” parameter and γ1 as an “habit formation” parameter. However, as we do not model the household problem explicitly, 435 we remain agnostic about the specific sources of time-varying risk aversion. Define as Vi,t the equity value of the firm. We assume that shareholders strategically default on their debt obligations if Vi,t < 0. Thus, interest payments ci,t are determined endogenously as follows: Bi,t+1 = Et Mt+1 {(1 + ci,t+1 ) Bi,t+1 χ{Vi,t+1 ≥0} + ξKi,t+1 χ{Vi,t+1 <0} } , (28) where ξ ∈ [0, 1) is the fraction of capital the lenders recover in case of default. Equity holders choose investment Ii,t and debt issuance ∆Bi,t to solve the following recursive problem: Vi,t = max {0, Di,t + Et [Mt+1 · Vi,t+1 ]}, Ii,t ,∆Bi,t (29) subject to (22), (25), and (28). In the recursive representation (29) the equity value Vi,t = V (Si,t ) is a function of the state variables Si,t = {Ki,t , Bi,t , ci,t , Zi,t , At }. Defining a Dynamic Target. To define a dynamic leverage target, we follow the “gap approach”, which has been extensively used to describe costly adjustments of production factors (e.g., Sargent 1978, Caballero and Engel 1993, 440 Cooper and Willis 2004, King and Thomas 2006, Bayer 2009). These studies consider adjustments toward a dynamic target, which is termed as the frictionless target. The latter corresponds to the level of a production factor to which an optimizing agent would eventually adjust to in the absence of changes in the stochastic variables. More formally, the frictionless target is constructed as the policy function in which adjustment costs are removed for a single period. ∗ ∗ In our setup, the frictionless target debt stock Bi,t = Bi,t + ∆Bi,t can be obtained from the following optimization problem: ∗ C max {0, Di,t + Et [Mt+1 · Vi,t+1 ]}, ∗ ,∆B ∗ Ii,t i,t 22 (30) ∗ where Di,t is the equity payout in which current-period debt adjustment costs are removed, i.e., Λ(∆Bi,t ) = 0. The 445 C is computed using the value function Vi,t in (29), as adjustment costs are set to zero only continuation value Vi,t+1 for the current period. We then compute the target leverage ratio as ∗ Bi,t ∗ , Ki,t ∗ ∗ where Ki,t = Ki,t + Ii,t . As in the gap approach, our dynamic target can be interpreted as the capital structure that, conditional on today’s state of the world, optimally positions the firm to deal with the uncertain funding needs it may have in the future. Three remarks are in order. First, the frictionless target encompasses the definition of target in the quadratic 450 adjustment model of Section 2 as a special case. More generally, Adda, Cooper, and Cooper (2003) show that the approximation of the frictionless target as the fixed point where the state-dependent policy function crosses the 45-degree line is exact in the case of a linear-quadratic dynamic programming problem. Second, the frictionless target generally differs from the static target defined as the debt level that would arise if there were never any costs of adjustments (e.g., Cooper and Willis 2004, Bayer 2009). The static and the frictionless target coincide in 455 the illustration in Section 2. However, due to its dynamic nature, the frictionless target more closely maps onto the reduced-form empirical specifications commonly used in the empirical corporate finance literature in which the target depends on firm-level variables (e.g., Flannery and Rangan 2006, Lemmon, Roberts, and Zender 2008).33 Third, the structure of our dynamic model does not mechanically impose convergence to the target. Thus, the calibrated model serves as a lab to quantify to what extent firms adjust their capital structure toward the target. 460 Model Solution. As the model admits no closed-form solution, we resort to numerical dynamic programming in our quantitative analysis. The model solution is computationally challenging. Equity and debt values are mutually dependent since the default condition affects the debt pricing equation. In a similar context, Gomes and Schmid (2010) reduce the dimensionality of the state space using total debt commitments as a state variable. However, their approach is not viable because of the presence of debt adjustment costs. Thus, we need to solve jointly for both 465 equity values and interest rate schedules and keep track of the dynamics of our five state variables, including the coupon. The Appendix details our numerical solution method. Discussion. We lay out a neoclassical dynamic model of investment and financing with endogenous default and a flexible functional form for debt adjustment costs. As debt adjustment costs can originate from several sources on which the literature provides limited guidance, we discipline their values by means of calibration. To study firms’ 470 adjustments toward target capital structures in a more realistic dynamic environment, we take advantage of the “gap approach” in the macroeconomic literature. The following quantitative analysis assesses the relevance of the mechanism we illustrate in Section 2, and for which we provide suggestive empirical evidence in Section 3.4, for leverage dynamics and equity returns. 33 DeAngelo, DeAngelo, and Whited (2011) consider a static long-run target to which firms would converge after receiving neutral shocks for many periods in a row. 23 5. Quantitative Analysis 475 5.1. Calibration and Model Fit Table 5 summarizes our baseline calibration. The calibration frequency is monthly. Details about the computation of the model-based and data variables are provided in the Appendix. The model features 17 parameters. The first 8 parameters in the table are on the technology side. We set the curvature of the profit function, α, to 0.4, to roughly match the capital share from the Bureau of Economic Analysis (BEA). This value is similar, for example, 480 to the ones used by Kydland and Prescott (1982), Gomes (2001), and Gomes, Jermann, and Schmid (2016). The depreciation rate δ is set to be 0.01. This is a fairly common value in the literature, as it implies an annual rate of roughly 12%. This value is in line with the empirical estimates in Cooper and Haltiwanger (2006) and comparable to those used by previous studies. For example, Gomes (2001) uses a depreciation rate of 0.145, while Hennessy and Whited (2005) estimate a value of 0.10. We choose the persistence of the aggregate productivity process, ρA , 1 485 and its volatility, σA , to be 0.95 3 and 0.007/3, respectively. These monthly values correspond to 0.95 and 0.007 at the quarterly frequency, consistent with several studies, including Cooley and Prescott (2021), Zhang (2005), and Gomes, Jermann, and Schmid (2016). We normalize the average aggregate productivity, µA , to -2. µA is purely a scaling constant that determines the long-run average scale of the economy. As in Zhang (2005), we calibrate the persistence ρZ and volatility σZ of the idiosyncratic shock process to 0.97 and 0.1, respectively. The fixed cost of 490 operation, F , is chosen to approximately match average profitability in our sample, which leads to a value of 0.18 (or 2.25% of the average capital stock). The next 3 parameters describe the dynamics of the pricing kernels. We choose β, γ0 , and γ1 to minimize mean square errors with respect to three aggregate data moments, namely the average Sharpe ratio, the average risk-free rate, and its volatility (as in Zhang 2005). This procedure yields β = 0.9928, γ0 = 52.71, and γ1 = −50.19. 495 The remaining 6 parameters are on the financing side. We pick the fixed and proportional equity flotation costs, λ0 and λ1 , to be 0.5 and 0.025. As, for example, Kuehn and Schmid (2012) and Bolton, Wang, and Yang (2021), we choose λ0 to approximately match the frequency of equity issuance in the data. For the proportional component λ1 , we pick the same value as in Gomes and Schmid (2010), who also study levered returns. This is also close to Gomes (2001), who chooses 0.028 based on regressions of flotation costs on amount issued. We choose the recovery 500 rate parameter ξ to be 0.125 of the firm’s capital stock. This implies an average debt recovery rate of 53.9%, which is close to the 51 % recovery rate for creditors when the firm defaults in Huang and Huang (2012). We calibrate the debt adjustment cost parameters λB and γB to approximately target the average debt issuance and the frequency of default in our sample, respectively. The scale parameter λB affects the marginal cost of issuing debt and, in the model, discourages the use of external debt financing. Instead, positive values of γB imply that 505 issuing debt is more costly than withdrawing debt. Thus, the lower γB (i.e., negative large values), the more likely firms default due to their inability of reducing leverage following negative shocks. These values imply an average cost of issuing debt of 0.05 % of the total amount of debt issued, which is at the lower end of the range used by Strebulaev 24 (2007), who chooses values in the range of 0.05 % to 0.35 %. The average costs of withdrawing debt are instead larger, around 0.1 % of the total amount of debt withdrawn. Overall, the magnitude of debt adjustment costs suggests that, 510 as in Fischer, Heinkel, and Zechner (1989), Goldstein, Ju, and Leland (2001), and Strebulaev (2007), relatively small adjustment costs significantly affect leverage dynamics. Finally, following Nikolov and Whited (2014), we choose the tax rate τ to be 0.20. This is as an approximation of the statutory corporate tax rate relative to personal tax rates. [Insert Table 5 Here] Table 6 summarizes overall model fit under the parameterization in Table 5. The table compares model-implied moments, which are tabulated in the first column, with their empirical counterparts, which are tabulated in the 515 second column. Overall, the model does a reasonable job at matching key variables describing the financing policies of US firms. The model matches quite closely the Sharpe ratios, the annual risk-free rate, and its volatility. The model produces a sizeable equity premium, which we do not target in the calibration. Both in the model and in the data, this moment is around 6%. The second set of moments in the table refer to firms’ real and financial policies. On the real side, the model matches fairly closely average profitability, the volatility and autocorrelation of profitability, 520 and investment ratios. The model does a reasonable job in reproducing average leverage, an untargeted quantity. Model-implied leverage is 21%, a slight overestimation of its data counterpart of 16 %. Our parameterization also produces default rates and book-to-market ratios with comparable magnitudes to the data. The third set of moments in the table describe firms’ capital structure rebalancing. Debt issuance is 19% in the model, and 22% in the data. Although we target this moment in our calibration, we report model-implied and data moments that describe the 525 relative frequency of positive and negative debt adjustments. The model-implied magnitudes of these additional non-targeted moments are also close to the data. Finally, the frequency of equity issuance in the model is 5%, fairly close to its data value of 4%. Overall, the model fit seems reasonable, both for moments that serve as targets in the calibration, and for untargeted key statistics. In the following sections we use our baseline calibration as a lab to provide inference about capital structure rebalancing and equity returns. [Insert Table 6 Here] 530 5.2. Adjustment toward Target Leverage Table 7 describes capital structure dynamics around target leverage under the baseline calibration of Table 5. Panel A breaks down the simulated data into quintiles of RL. The top row reports the average values of RL for each quintile. As RL is defined as the difference between market and target leverage, negative (positive) values refer to underlevered (overlevered) firms. The bottom row tabulates the corresponding one-period-ahead changes in leverage. 535 Overlevered firms tend to lever down, while underlevered firms tend to lever up. These findings, which are based on the frictionless target defined as in (30), are in line with the reduced-form evidence in Section 3.4. 25 Panel B reports model-implied and data estimates of adjustment speeds, both in the model and in the data. Panel B also investigates the role of firm’s size for capital structure adjustments. Data estimates are from the estimation of the model of Flannery and Rangan (2006) in (19), in which the speed of adjustment λ in (19) now has a fixed component and a component related to size (total assets AT), i.e., λ = λ0 + λ1 log(AT ). As in the model, target leverage is observable, and model-implied estimates are from the following specification: ∆M Li,t+1 = λi,t (−RLi,t ) + i,t+1 , (31) where ∆M Li,t+1 = M Li,t+1 − M Li,t and λi,t = λ0 + λ1 log(Ki,t ). Notice that the minus sign in front of RLi,t captures the fact that adjustments toward the target imply that positive values of RLi,t are associated with negative values of ∆M Li,t+1 , and vice versa. This is because overlevered firms would tend to reduce their leverage (and vice 540 versa), as the non-parametric evidence in Panel A indicates. Both in the model and in the data, the adjustment speed in Panel B for the full sample (“all firms”) is computed in correspondence of average size. Adjustment speeds for large (small) firms are instead evaluated in correspondence of one standard deviation above (below) average size. The row labeled “All Firms” shows that the average adjustment speed in the model is 0.17, close to its 0.21 data counterpart. Both in the model and in the data, small firms tend to adjust faster toward target leverage. Estimated 545 adjustment speeds in the “Small Firms” row are 0.22 (model) and 0.27 (data), versus 0.12 (model) and 0.25 (data) in the “Large Firms” row. Taken together, the results in Table 7 suggest that firms adjust their capital structure toward a dynamic target leverage ratio, consistent with the intuition in Proposition 1. In the cross section of size, this pattern is present for both large and small firms. [Insert Table 7 Here] 550 5.3. Returns and Capital Structure Imbalances Table 8 reports average annualized value-weighted stock returns under the baseline parameterization of Table 5. The table shows the results of double sorts on relative leverage (rows) and on leverage (columns). Panel A tabulates returns for all firms in the simulated economy. Across all leverage portfolios, overlevered firms earn higher average returns than underlevered firms. The spreads in average returns is more pronounced for firms with high 555 leverage. Leverage is also generally positively related to returns, with the exception of underlevered firms, for which low-leverage firms earn higher average returns than high-leverage firms. Panels B and C, instead, report returns for large and small firms, defined as those for which Ki,t is, respectively, above and below its median. Average returns are positively related to relative leverage, and the spread is more pronounced for large firms, whose adjustment speeds are lower. Instead, the relationship between leverage and returns 26 560 appears less stable. High-leverage firms earn lower average returns than low-leverage firms among underlevered large firms. In addition, spreads in returns are economically small for all small firms, for each bin of relative leverage. [Insert Table 8 Here] Table 9 reports estimates of cross-sectional regressions of stock returns on relative leverage and leverage on the same simulated data. Columns (1) to (3) tabulate estimates for the full simulated sample. The results mimic the patterns in Table 8 and are broadly in line with the non-targeted empirical evidence of Section 3.4. The coefficient 565 on relative leverage is larger in magnitude than the one on leverage. When size and book-to-market are included as controls in Column (3), the coefficient on leverage becomes negative. This confirms that the relationship between leverage and returns is ambiguous. Columns (4) to (6) and (7) to (9) refer, respectively, to firms above and below the median value of size Ki,t . The coefficient on relative leverage is positive for both groups of firms, but bigger for large firms. Thus, as before, capital structure imbalances appear to be more important for larger firms. 570 All in all, the results in Table 8 and Table 9 corroborate the key intuition of Proposition 2 in a more realistic environment. Capital structure imbalances are important drivers of the relationship between leverage and crosssectional returns. [Insert Table 9 Here] 6. Conclusions In this paper we present a model in which firms maximize the value of equity by choosing an optimal amount of 575 debt finance. There are three frictions: corporate taxes, bankruptcy costs and debt adjustment costs. We show that the optimal leverage policy for the firm requires a partial adjustment toward a target leverage. Partial adjustment occurs because there are costs in adjusting debt. We then derive the implications of this optimal policy for the relation between equity returns and leverage in the presence of capital structure imbalances. We illustrate our key ideas in a stylized setup, which provides a transparent illustration of the economic forces underlying our empirical 580 results. In this context, we derive a formula that is similar to the celebrated equation of the second proposition of Modigliani and Miller (1958). In the presence of the above frictions, we show that one cannot analyze the relation between leverage and returns without controlling for either target or relative leverage. We show that controlling for leverage, equity returns are increasing in relative leverage, which is defined as the difference between leverage and target leverage. Instead, 585 controlling for leverage, equity returns are decreasing in target leverage. We find that the sign of the relation between returns and leverage is theoretically indeterminate (controlling for either relative or target leverage), because there are different counteracting effects at work. 27 In our setup, both the amount of debt and its unit cost determine the overall effect of leverage on returns. As in the second proposition of Modigliani and Miller (1958), leverage decisions affect how cash flow risk propagates 590 to equity payouts through the amplification of both good and bad outcomes. Given leverage, capital structure imbalances influence the unit cost of debt through the non-diversifiable components of bankruptcy costs and tax shields, which are reflected in the correction factor. Overlevered firms are riskier investments than underlevered firms because they carry costly debt with high expected (non-diversifiable) bankruptcy costs. As a consequence, capital structure imbalances either boost or dampen the traditional amplification effect of Modigliani and Miller 595 (1958). We embed the key economic trade-offs in this simple illustration in a more general dynamic environment, in which firms make endogenous investment and default decisions. A quantitative version of our model reproduces the key facts about leverage, capital structure imbalances, and equity returns for U.S. corporations. We also provide reduced-form empirical evidence, based on standard proxies for target leverage from the corporate finance literature. 600 Overall, our quantitative and empirical analyses corroborate the key intuition from the two-period illustration. Our work represents a step toward a closer integration between the theory of capital structure, as developed in the field of corporate finance, and production-based asset pricing theory. Our results indicate that financial flexibility crucially affects the link between leverage and equity returns. We acknowledge two limitations of our study. First, for tractability, we are agnostic about the sources of adjustment costs in firms’ capital structure. Although 605 incorporating issues such as seniority and maturity poses conceptual and computational challenges, the literature is currently missing a quantitative analysis of the frictions that limit firms’ financial flexibility. Second, although it is reassuring that model-based and reduced-form evidence agree, the mapping between target leverage in the model and its empirical estimates is not fully understood. Albeit outside of the scope of this paper, a thorough evaluation of common empirical estimates of target leverage through the lens of structural models would be informative. We 610 leave these tasks to future research. 28 References Acharya, V. V., S. T. Bharath, and A. Srinivasan, 2007, Does industry-wide distress affect defaulted firms? evidence from creditor recoveries, Journal of financial economics, 85(3), 787–821. Adda, J., R. Cooper, and R. W. Cooper, 2003, Dynamic economics: quantitative methods and applications. MIT 615 press. Arellano, M., and S. Bond, 1991, Some tests of specification for panel data: Monte carlo evidence and an application to employment equations, The Review of Economic Studies, 58(2), 277–297. Aruoba, S. B., L. Bocola, and F. Schorfheide, 2017, Assessing dsge model nonlinearities, Journal of Economic Dynamics and Control, 83, 34–54. 620 Bates, T. W., K. M. Kahle, and R. M. Stulz, 2009, Why do us firms hold so much more cash than they used to?, The Journal of Finance, 64(5), 1985–2021. Bayer, C., 2009, A comment on the economics of labor adjustment: Mind the gap: Evidence from a monte carlo experiment, The American Economic Review, 99(5), 2258–2266. Belo, F., X. Lin, and S. Bazdresch, 2014, Labor hiring, investment, and stock return predictability in the cross 625 section, Journal of Political Economy, 122(1), 129–177. Belo, F., X. Lin, and F. Yang, 2014, External equity financing shocks, financial flows, and asset prices, working paper, National Bureau of Economic Research. Berk, J. B., R. C. Green, and V. Naik, 1999, Optimal investment, growth options, and security returns, The Journal of finance, 54(5), 1553–1607. 630 Bhandari, L. C., 1988, Debt/equity ratio and expected common stock returns: Empirical evidence, The Journal of Finance, 43(2), 507–528. Blundell, R., and S. Bond, 1998, Initial conditions and moment restrictions in dynamic panel data models, Journal of Econometrics, 87(1), 115–143. Bolton, P., N. Wang, and J. Yang, 2021, Leverage dynamics under costly equity issuance, . 635 Bretscher, L., P. Feldhütter, A. Kane, and L. Schmid, 2020, Marking to market corporate debt, Swiss Finance Institute Research Paper, (21-06). Caballero, R. J., and E. M. Engel, 1993, Microeconomic adjustment hazards and aggregate dynamics, The Quarterly Journal of Economics, 108(2), 359–383. 29 Campbell, J. Y., and J. H. Cochrane, 1999, By force of habit: A consumption-based explanation of aggregate stock 640 market behavior, Journal of political Economy, 107(2), 205–251. Caskey, J., J. Hughes, and J. Liu, 2012, Leverage, excess leverage, and future returns, Review of Accounting Studies, 17(2), 443–471. Cochrane, J. H., 1991, Production-based asset pricing and the link between stock returns and economic fluctuations, The Journal of Finance, 46(1), 209–237. 645 Cooley, T. F., and E. C. Prescott, 2021, 1. economic growth and business cycles, in Frontiers of business cycle research. Princeton University Press, pp. 1–38. Cooper, R., and J. L. Willis, 2004, A comment on the economics of labor adjustment: Mind the gap, The American Economic Review, 94(4), 1223–1237. Cooper, R. W., and J. C. Haltiwanger, 2006, On the nature of capital adjustment costs, The Review of Economic 650 Studies, 73(3), 611–633. Covas, F., and W. J. Den Haan, 2011, The cyclical behavior of debt and equity finance, American Economic Review, 101(2), 877–99. Croce, M. M., H. Kung, T. T. Nguyen, and L. Schmid, 2012, Fiscal policies and asset prices, Review of Financial Studies, 25(9), 2635–2672. 655 DeAngelo, H., L. DeAngelo, and T. M. Whited, 2011, Capital structure dynamics and transitory debt, Journal of Financial Economics, 99(2), 235–261. Diamond, D. W., 1991, Debt maturity structure and liquidity risk, the Quarterly Journal of economics, 106(3), 709–737. Doshi, H., K. Jacobs, P. Kumar, and R. Rabinovitch, 2019, Leverage and the cross-section of equity returns, The 660 Journal of Finance, 74(3), 1431–1471. Fama, E. F., and K. R. French, 1992, The cross-section of expected stock returns, The Journal of Finance, 47(2), 427–465. , 2002, Testing trade-off and pecking order predictions about dividends and debt, Review of Financial Studies, 15(1), 1–34. 665 Faulkender, M., M. J. Flannery, K. W. Hankins, and J. M. Smith, 2010, Cash flows and leverage adjustments, Working Paper. 30 Fischer, E. O., R. Heinkel, and J. Zechner, 1989, Dynamic capital structure choice: Theory and tests, The Journal of Finance, 44(1), 19–40. Flannery, M. J., and K. W. Hankins, 2010, Estimating dynamic panel models in corporate finance, Working Paper. 670 Flannery, M. J., and K. W. Hankins, 2013, Estimating dynamic panel models in corporate finance, Journal of Corporate Finance, 19, 1–19. Flannery, M. J., and K. P. Rangan, 2006, Partial adjustment toward target capital structures, Journal of Financial Economics, 79(3), 469–506. Friewald, N., F. Nagler, and C. Wagner, 2021, Debt refinancing and equity returns, Journal of Finance, Forthcoming. 675 George, T. J., and C.-Y. Hwang, 2010, A resolution of the distress risk and leverage puzzles in the cross section of stock returns, Journal of Financial Economics, 96(1), 56–79. Goldstein, R., N. Ju, and H. Leland, 2001, An ebit-based model of dynamic capital structure, The Journal of Business, 74(4), 483–512. Gomes, J., L. Kogan, and L. Zhang, 2003, Equilibrium cross section of returns, Journal of Political Economy, 111(4). 680 Gomes, J. F., 2001, Financing investment, American Economic Review, pp. 1263–1285. Gomes, J. F., U. J. Jermann, and L. Schmid, 2016, Sticky leverage, American Economic Review. Gomes, J. F., and L. Schmid, 2010, Levered returns, The Journal of Finance, 65(2), 467–494. , 2016, Equilibrium asset pricing with leverage and default, The Wharton School Research Paper, (92). Gomes, J. F., A. Yaron, and L. Zhang, 2006, Asset pricing implications of firms financing constraints, Review of 685 Financial Studies, 19(4), 1321–1356. Graham, J. R., 2000, How big are the tax benefits of debt?, The journal of finance, 55(5), 1901–1941. Hahn, J., J. Hausman, and G. Kuersteiner, 2007, Long difference instrumental variables estimation for dynamic panel models with fixed effects, Journal of Econometrics, 140(2), 574–617. Halling, M., J. Yu, and J. Zechner, 2015, Leverage dynamics over the business cycle, in AFA 2012 Chicago Meetings 690 Paper. Hennessy, C. A., and T. M. Whited, 2005, Debt dynamics, The Journal of Finance, 60(3), 1129–1165. , 2007, How costly is external financing? Evidence from a structural estimation, The Journal of Finance, 62(4), 1705–1745. 31 Hou, K., C. Xue, and L. Zhang, 2014, Digesting anomalies: An investment approach, Review of Financial Studies. 695 Hovakimian, A., T. Opler, and S. Titman, 2001, The debt-equity choice: An analysis of issuing firms, Journal of Financial and Quantitative Analysis, 36(1), 1–24. Huang, J.-Z., and M. Huang, 2012, How much of the corporate-treasury yield spread is due to credit risk?, The Review of Asset Pricing Studies, 2(2), 153–202. Huang, R., and J. R. Ritter, 2009, Testing theories of capital structure and estimating the speed of adjustment, 700 Journal of Financial and Quantitative Analysis, 44(2), 237–271. Jermann, U. J., 1998, Asset pricing in production economies, Journal of Monetary Economics, 41(2), 257–275. Kim, J., and F. J. Ruge-Murcia, 2009, How much inflation is necessary to grease the wheels?, Journal of Monetary Economics, 56(3), 365–377. King, R. G., and J. K. Thomas, 2006, Partial adjustment without apology, International Economic Review, 47(3), 705 779–809. Korajczyk, R. A., and A. Levy, 2003, Capital structure choice: Macroeconomic conditions and financial constraints, Journal of Financial Economics, 68(1), 75–109. Korteweg, A., 2010, The net benefits to leverage, The Journal of Finance, 65(6), 2137–2170. Krasker, W. S., 1986, Stock price movements in response to stock issues under asymmetric information, The Journal 710 of Finance, 41(1), 93–105. Kuehn, L.-A., and L. Schmid, 2012, Investment-based corporate bond pricing, Working Paper. Kydland, F. E., and E. C. Prescott, 1982, Time to build and aggregate fluctuations, Econometrica: Journal of the Econometric Society, pp. 1345–1370. Leary, M. T., and M. R. Roberts, 2005, Do firms rebalance their capital structures?, The Journal of Finance, 60(6), 715 2575–2619. Lemmon, M. L., M. R. Roberts, and J. F. Zender, 2008, Back to the beginning: persistence and the cross-section of corporate capital structure, The journal of finance, 63(4), 1575–1608. Liu, L. X., T. M. Whited, and L. Zhang, 2009, Investment-based expected stock returns, Journal of Political Economy, 117(6), 1105–1139. 720 Livdan, D., H. Sapriza, and L. Zhang, 2009, Financially constrained stock returns, The Journal of Finance, 64(4), 1827–1862. 32 Modigliani, F., and M. H. Miller, 1958, The cost of capital, corporation finance and the theory of investment, The American economic review, 48(3), 261–297. Myers, S. C., 1977, Determinants of corporate borrowing, Journal of Financial Economics, 5(2), 147–175. 725 Myers, S. C., and N. S. Majluf, 1984, Corporate financing and investment decisions when firms have information that investors do not have, Journal of Financial Economics, 13(2), 187–221. Nikolov, B., and T. M. Whited, 2014, Agency conflicts and cash: Estimates from a dynamic model, The Journal of Finance, 69(5), 1883–1921. Novy-Marx, R., 2013, The other side of value: The gross profitability premium, Journal of Financial Economics, 730 108(1), 1–28. Obreja, I., 2013, Book-to-market equity, financial leverage, and the cross-section of stock returns, Review of Financial Studies, 26(5), 1146–1189. Opler, T., L. Pinkowitz, R. Stulz, and R. Williamson, 1999, The determinants and implications of corporate cash holdings, Journal of Financial Economics, 52(1), 3–46. 735 Ozdagli, A. K., 2012, Financial leverage, corporate investment, and stock returns, Review of Financial Studies, 4, 1033–1069. Ozdagli, A. K., 2015, An injury to one: Financial frictions and the reaction of stock prices to monetary policy shocks, Federal Reserve Bank of Boston Working Paper. Penman, S. H., S. A. Richardson, and I. Tuna, 1992, The book-to-price effect in stock returns: Accounting for 740 leverage, Journal of Accounting Research, 45(2), 427–467. Sargent, T. J., 1978, Estimation of dynamic labor demand schedules under rational expectations, Journal of political Economy, 86(6), 1009–1044. Strebulaev, I. A., 2007, Do tests of capital structure theory mean what they say?, The Journal of Finance, 62(4), 1747–1787. 745 Varian, H. R., 1975, A bayesian approach to real estate assessment, Studies in Bayesian econometric and statistics in Honor of Leonard J. Savage, pp. 195–208. Warr, R. S., W. B. Elliott, J. Koëter-Kant, and Ö. Öztekin, 2012, Equity mispricing and leverage adjustment costs, Journal of Financial and Quantitative Analysis, 47(03), 589–616. Zhang, L., 2005, The value premium, The Journal of Finance, 60(1), 67–103. 33 Figure 1 Adjustments toward Target Leverage Figure 1 depicts the optimal leverage policy of a firm for two different values of the adjustment cost parameter θ. Panel A refers to the case of low adjustment costs (θ = 0.1), while Panel B refers to the case of high adjustment costs (θ = 0.3). Both panels report initial leverage (d0 /k) on the horizontal axis and optimal leverage (d1 /k) on the vertical axis. Target leverage is O at the intersection of the policy function with the 45-degree line. LU 0 and L0 respectively denote two possible levels of initial leverage below the target and above the target. The model is solved numerically with three possible states (H, M and L) at time t2 that can occur with probability 0.3, 0.5 and 0.2 respectively. In the low state L firms are insolvent with β2 (L) = 0.5. In the non-default states, the tax rate is τ = 0.3. The remaining parameters are as follows: M (H) = 0.8, M (M ) = 1.05, M (L) = 1.1, r2A (H) = 0.9, r2A (M ) = −0.1, R2A (L) = −0.45205, k = 1. Panel A: Low Adjustment Costs Panel B: High Adjustment Costs 34 Table 1 Leverage Decomposition: Descriptive Statistics The table provides summary statistics for our four measures of target leverage and for key firm characteristics. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013. Financial firms and utilities are excluded. ML is the market debt ratio, TL is the estimated target debt ratio, RL is relative leverage, defined as the difference between ML and RL. TL is obtained from the estimation of the model of Flannery and Rangan (2006) with firm fixed effects and at least 5 observations for each firm. The estimation of TL begins in 1970 (using data since 1965) and is done on a rolling basis as described in Section 3, each year using all the accounting information available until that year. Panel A reports means, standard deviations, and autocorrelation coefficients for market leverage and for target and relative leverage. Panel B reports average market leverage, target leverage, relative leverage, capital stock (K), investment (I/K), profitability (ROA), market capitalization (SIZE), Tobin Q (Q) and the number of observations. All nominal magnitudes are deflated by the consumer price index, to express all nominal values in 2000 dollars. All variables are winsorized at the 1 percent level and are measured as described in the appendix. Panel A: Leverage Decomposition Mean St. Dev. AC(1) ML: Observed Leverage 0.16 0.17 0.87 TL: FR with Firm FE (min 5 obs) 0.17 0.15 0.90 RL: ML - TL 0.00 0.10 0.67 Panel B: Firm Characteristics and Sample Selection LEV Leverage TL RL K Investment I/K ROA Valuation SIZE Q N 750 Full Sample 0.16 TL (min 5 obs) 0.17 0.17 0.00 2.49 0.10 0.10 2.64 2.26 81978 3.05 0.08 0.11 3.22 2.10 61683 35 Table 2 Firm Characteristics across Quintiles of Relative Leverage The table provides summary statistics for key firm characteristics across quintiles of firms sorted by estimated relative leverage. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013. Financial firms and utilities are excluded. ML is the market debt ratio, TL is the estimated target debt ratio, RL is relative leverage, defined as the difference between ML and RL and ∆ML is the rate of change of market leverage in the following year. Our measure of target leverage is obtained from the estimation of the model of Flannery and Rangan (2006) with firm fixed effects and at least 5 observations for each firm. The estimation of TL begins in 1970 (using data since 1965) and is done on a rolling basis as described in Section 3, each year using all the accounting information available until that year. The table reports average market leverage, target leverage, relative leverage, capital stock (K), investment (I/K), profitability (ROA), market capitalization (SIZE), Tobin Q (Q) and the number of observations. All nominal magnitudes are deflated by the consumer price index, to express all nominal values in 2000 dollars. All variables are winsorized at the 1 percent level and are measured as described in the appendix. RL - FR with Firm FE (min 5 obs) Group of RL ML Leverage TL RL ∆ML K Low 0.11 0.24 -0.13 0.04 1.84 0.10 0.14 2.26 2.17 2 0.12 0.16 -0.04 0.00 1.97 0.09 0.13 2.37 2.37 3 0.12 0.12 0.00 -0.02 2.02 0.08 0.11 2.10 2.34 4 0.17 0.13 0.04 -0.08 3.36 0.07 0.10 2.40 2.07 High 0.31 0.18 0.13 -0.13 6.05 0.08 0.09 2.33 1.51 36 Investment I/K ROA Valuation SIZE Q 37 6.09 2.03 0.46 8.04 4.07 0.86 Re [t] SR Re [t] SR RL H-L ML Sorting Variable 12.29 3.05 0.59 15.20 3.41 0.66 L 15.02 3.90 0.76 14.55 3.65 0.73 15.44 4.16 0.82 14.60 3.92 0.77 Equally-Weighted 3 5 15.45 4.10 0.82 15.85 4.09 0.82 7 20.33 4.35 0.94 21.29 4.10 0.91 H Excess Returns 5.56 1.97 0.35 1.73 0.51 0.10 H-L 5.06 1.28 0.24 7.80 1.99 0.35 L 8.98 2.73 0.50 9.07 3.11 0.54 9.62 3.40 0.60 7.56 2.84 0.49 Value-Weighted 3 5 9.30 3.30 0.58 11.33 3.76 0.67 7 10.62 3.22 0.57 9.53 2.30 0.44 H In the table stocks are sorted every June in deciles based on their values of market leverage (ML) and relative leverage (RL). RL is obtained from the estimation of TL from the model of Flannery and Rangan (2006) with firm fixed effects and at least 5 observations for each firm. The estimation of TL begins in 1970 (using data since 1965) and is done on a rolling basis as described in Section 3, each year using all the accounting information available until that year. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013 and covered by the Center of Research in Security Prices (CRSP). Financial firms and utilities are excluded. The breakpoints for portfolio sorts are computed on the subset of firms traded on the NYSE market. The table reports returns in excess of the risk-free rate, t-statistics and Sharpe ratios for the bottom decile (L), the top decile (H) and for the third, fifth and seventh decile. We also report the difference for the excess returns, the t-statistics and the Sharpe ratio between the top decile and the bottom decile (H-L). The left panel reports equally-weighted returns, while the right panel reports value-weighted returns. The sorting variables are matched to monthly returns as described in Section 4. Table 3 Univariate Portfolio Sorts Table 4 Levered Returns: Relative, Observed and Target Leverage For each month between July 1980 and December 2013, we estimate cross-sectional regressions of stock returns on relative leverage (RL), market leverage (ML), target leverage (TL), market capitalization (SIZE), and book-to-market equity (B/M). TL is obtained from the estimation of the model of Flannery and Rangan (2006) with firm fixed effects and at least 5 observations for each firm. The estimation of TL begins in 1970 (using data since 1965) and is done on a rolling basis as described in Section 3, each year using all the accounting information available until that year. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013 and covered by the Center of Research in Security Prices (CRSP). Financial firms and utilities are excluded. De-listing returns are included in monthly returns. The table reports Fama-MacBeth coefficient estimates. t-statistics are in parentheses. R2 and N denote the cross-sectional R-squared and the number of observations respectively. The independent variables are matched to monthly returns as described in Section 3. All variables are described in the appendix. The symbols (***), (**) and (*) denote statistical significance at the 1, 5 and 10 percent levels respectively. Observed vs Relative vs Target Leverage (1) ML (2) (3) (4) (5) (6) (7) 0.05 (0.19) 0.86∗∗∗ (2.99) 1.77∗∗∗ (4.38) 0.92∗∗∗ (2.82) 1.16∗∗∗ (3.57) 0.61∗ (1.84) 0.91∗∗∗ (3.14) -1.16∗∗∗ (-3.57) -0.87∗∗∗ (-2.99) -0.20∗∗∗ (-4.93) 0.14∗∗ (2.06) 0.01 748724 0.02 727164 0.14 (0.47) 1.69∗∗∗ (4.53) RL TL SIZE B/M R2 N -0.24∗∗∗ (-5.28) 0.17∗∗ (2.46) 0.02 931438 -0.20∗∗∗ (-4.84) 0.14∗ (1.96) 0.01 748724 -0.20∗∗∗ (-4.93) 0.14∗∗ (2.06) 0.02 727164 0.01 748724 38 0.02 727164 Table 5 Parameter Values The table reports parameter choices for the calibrated model. The frequency of calibration is monthly. Category Technology Description Capital share Depreciation Persistence of aggregate shock Standard deviation of aggregate shock Mean of aggregate shock Persistence of idiosyncratic shock Standard deviation of idiosyncratic shock Mean of idiosyncratic shock Fixed cost of operations Symbol α δ ρA σA µA ρz σz µz F Value 0.4 0.01 1 0.95 3 0.007/3 -2 0.97 .1 0 0.18 Pricing Kernel Time discount Constant “risk aversion” Time varying “risk aversion”’ β γ0 γ1 0.9928 52.71 -50.19 Financing Fixed equity flotation cost Proportional equity flotation cost Debt recovery in bankruptcy Debt adjustment cost (“scale”) Debt adjustment cost (“asymmetry”) Corporate tax rate λ0 λ1 ξ λB γB τ 0.5 0.025 0.125 .4 −0.2 0.2 39 Table 6 Moments The table reports model-implied and data moments for the baseline calibration of Table 5. The model-implied moments are calculated as averages of simulations of 10,000 firms and 2000 time periods. The data source for the Sharpe Ratio and risk-free rate moments is Zhang (2005). The average annual equity return is computed using data from Kenneth French’s data library. Default rates are taken from Covas and Den Haan (2011). The remaining data moments are computed from our sample of nonfinancial, unregulated firms from the annual Compustat dataset. All moments are annualized. Details on model and data variable definitions are provided in the Appendix. Model 0.43 0.02 0.04 0.06 Data 0.43 0.02 0.03 0.06 Average profitability Volatility of profitability Autocorrelation of profitability Average investment Average leverage Frequency of default Average book-to-market ratio 0.16 0.07 0.61 0.16 0.21 0.03 0.48 0.15 0.09 0.75 0.14 0.16 0.02 0.57 Average debt issuance Frequency of positive debt adjustments Frequency of negative debt adjustments Frequency of equity issuance 0.19 0.62 0.38 0.05 0.22 0.59 0.41 0.04 Average Sharpe Ratio Average risk-free rate Volatility of risk-free rate Average equity premium 40 Table 7 Adjustments toward Target Leverage - Model Panel A describes firms’ capital structure adjustments toward target leverage under the baseline calibration of Table 5. RL denotes relative leverage (defined as the difference between market and target leverage) and adjustment denotes the one-period-ahead change in leverage. Panel B reports model-implied and data estimates of adjustment speeds for the full sample of firms, for small firms and for large firms. Data figures are obtained with the measure of target leverage obtained from the estimation of the model of Flannery and Rangan (2006) in Equation 19, with firm fixed effects and at least 5 observations for each firm. The speed of adjustment λ in (19) has a fixed component and a component related to size (total assets AT), i.e., λ = λ0 + λ1 log(AT ). Model-implied figures are based on a regression of leverage adjustments on RL, Ki,t , and their interaction, as described in Equation 31. Both in the model and in the data, adjustment speeds for all firms are computed in correspondence of average size. Adjustment speeds for large (small) firms are computed in correspondence of one standard deviation above (below) average size. All model-implied quantities are based on simulations of 10,000 firms and 2000 time periods. Panel A: Leverage Adjustments Group of RL Low 2 3 RL Adjustment -0.029 0.023 -0.002 0.002 Panel B: Adjustment Speeds Model All Firms Small Firms Large Firms 0.17 0.22 0.12 41 0.000 0.001 4 High 0.015 -0.008 0.065 -0.019 Data 0.21 0.27 0.15 Table 8 Average Returns and Capital Structure Imbalances The table reports average annualized value-weighted returns of portfolios sorted by terciles of relative leverage (defined as the difference between market and target leverage) and then by terciles of leverage. All figures are are based on simulations of 10,000 firms and 2000 time periods under the baseline calibration of Table 5. Panel A refers to the entire simulated economy. Panels B and C refer, respectively, to firms above and below the median value of size Ki,t . The cells labeled “OL minus UL” report the difference between the average return of the highest- and the lowest-tercile portfolio of stocks sorted by relative leverage, for each leverage group. The cells labeled “High minus Low” report the difference between the average return of the highest- and the lowest-tercile portfolio of stocks sorted by leverage, for each relative leverage group. Panel A: All Firms Leverage High Relative Leverage Low 2 Underlevered (UL) 2 Overlevered (OL) 7.21 7.45 7.48 6.86 6.72 7.53 7.06 8.18 14.89 OL minus UL 0.27 0.67 7.83 High minus Low -0.15 0.73 7.41 Panel B: Large Firms Leverage High Relative Leverage Low 2 Underlevered (UL) 2 Overlevered (OL) 8.18 8.26 9.38 7.66 8.39 10.31 5.67 9.35 20.80 OL minus UL 1.20 2.65 15.13 High minus Low -2.51 1.09 11.42 Panel C: Small Firms Leverage High Relative Leverage Low 2 Underlevered (UL) 2 Overlevered (OL) 7.09 6.90 7.57 6.36 6.89 7.52 6.28 6.94 7.86 OL minus UL 0.48 1.16 1.58 42 High minus Low -0.81 0.04 0.29 Table 9 Average Returns and Capital Structure Imbalances: Regressions The table reports estimated coefficients from cross-sectional regressions of stock returns on relative leverage, market leverage, target leverage, market capitalization, and book-to-market equity. All figures are are based on simulations of 10,000 firms and 2000 time periods under the baseline calibration of Table 5. Columns (1) to (3) refer to the entire simulated economy, columns (4) to (6) and (7) to (9) refer, respectively, to firms above and below the median value of size Ki,t . (1) Relative Leverage Leverage 0.58 All Firms (2) (3) (4) Large Firms (5) (6) (7) Small Firms (8) (9) 0.70 0.71 1.12 1.06 1.45 0.37 0.36 0.46 0.13 0.11 -0.45 -0.66 0.18 0.13 Size -0.03 0.00 -0.05 Book-to-Market 0.02 -0.10 0.01 43 Appendix Proofs of Propositions Proposition 1. The FOC for the firm is: ∂ E[M ((1 + r2D (d1 )(1 − τ2 ))d1 + β2 (d1 )d1 )] ∂Θ (d0 , d1 ) = 1− ∂d1 ∂d1 . (32) From the no arbitrage condition in the debt market, we have that E[M (1 + r2D (d1 ))] = 1. Observe that upon bankruptcy occurring β2 (d1 ) = βd1 . Then, the following holds Z Z∞ A A A A M 1 + ρD 2 (d1 ) f (r2 |r1 )dr2 dM + Z Zr2 M −∞ r2A (1 + r2A )k − β2 (d1 )d1 f (r2A |r1A )dr2A dM = 1, d1 that is 1 + ρD 2 (d1 ) 1 − qD E[M (1 + r2A )|r2A < r2A ]k qD + − βd1 = 1. 1 + rF d1 1 + rF Solving for the coupon rate yields ρD 2 (d1 ) = βd1 E[M (1 + r2A )|r2A < r2A ]k 1 + rF rF + qD qD − + . 1 − qD d1 1 − qD 1 − qD (33) Observe that taxes are only paid when there is no default. Thus, the tax shield is τ ρD 2 (d1 )d1 if no default occurs, τ2 r2D (d1 )d1 = 0 if default occurs. Observing that ∂ ρD rF + qD + 2βd1 qD 2 (d1 )d1 = ∂d1 1 − qD one obtains ( D +2βd1 qD ∂ r2D (d1 )d1 . if no default occurs τ rF +q1−q D τ2 = ∂d1 0 if default occurs. Substituting in equation (4) yields h i rF + qD + 2βd1 qD d1 − d0 = E M τ |r2A ≥ r2A − E M (2d1 β) |r2A < r2A , 1 − qD which simplifies to θ (d1 − d0 ) = τ Collecting all terms with d1 k rF + qD + 2βd1 qD 1 − qD 1 − qD qD − 2d1 β . 1 + rF 1 + rF on the left-hand side and all the rest on the right-hand side one obtains qD qD d1 rF + qD θk − 2βkτ + 2βk =τ + θd0 , 1 + rF 1 + rF k 1 + rF which, after dividing by the coefficient of dk1 on the left hand side and subtracting rearranged as ! qD rF +qD τ 2β(1 − τ ) 1+r d1 d0 d 0 1+r F F − = . qD qD − k k θ + 2β(1 − τ ) 1+r 2β(1 − τ )k 1+r k F F 44 d0 k from both sides can be Using the definitions of target leverage and adjustment speed given in the proposition gives the result. 755 Proposition 2. Observe that ∂Θ (d0 , d1 ) λ1 = θ (d1 − d0 ) = θk − RL1 . ∂d1 1 − λ1 where the first equality follows from the derivative of the adjustment cost, and the second equality follows from equation (7). Then, equation (4) can be written as λ1 RL1 kθ − 1 − λ1 " ∂ r2D (d1 )d1 = E M τ2 ∂d1 # ∂ (β2 (d1 )d1 ) −E M . ∂d1 (34) From the definition of λ1 one obtains qD 2β(1 − τ ) 1+r λ1 F . = 1 − λ1 θ (35) Combining (34) with (35), simplifying the left-hand side of (34), and expanding the derivatives on its right-hand side, we obtain qD ∂rD (d1 ) −2β(1 − τ )k RL1 = E M τ2 r2D (d1 ) − E [M β2 d1 ] + E M τ2 2 d1 − E [M β2 d1 ] , 1 + rF ∂d1 or, equivalently, δD k 1 + rF qD d1 + + E [M β2 d1 ] . β 1 − qD d1 1 − qD E M τ2 r2D (d1 ) − E [M β2 d1 ] = −αRL1 − τ (36) qD d1 we obtain Inserting (36) into the definition of the correction factor (10) and observing that E [M β2 ] d1 = β 1+r F γ(d1 ) = 1 + αRL1 + τ β qD d1 δD k 1 + rF + 1 − qD d1 1 − qD Therefore, the correction factor can be written as γ(d1 ) = 1 + αRL1 − =1+ α d1 2 k − d∗ α k1 The proposition follows immediately. 45 α d1 2 k + τ δD + τ δD d1 k d1 k . −β qD d1 . 1 + rF 760 Variable Definitions The following table summarizes variable definitions with reference to Compustat and CRSP items. Variable Construction Value of Preferred Stocks (P S) Book Equity (BE) Market Leverage (M L) Market Leverage Growth (∆M L) Real Assets (K) Investment/Capital Ratio (I/K) Return on Assets (ROA) Tobin’s Q (Q) Market Capitalization (SIZE) Book-to-Market Equity (B/M ) If available, in this order: P ST KRV , P ST KL, P ST K. CEQ + T XDIT C (if available) − P S DLT T +DLC AT −BE+P RCC F ·CSHO F.M L−M L 0.5(F.M L+M L) AT ·CP I2000 1000 AT −L.AT 0.5(AT +L.AT ) OIBDP AT P RCC F ·CSHO+DLT T +DLC BE+DLT T +DLC T log |P RC|·SHROU (in June) 1000 BE log |P RC|·SHROU T /1000 σt [Mt+1 ] In the model, following Zhang (2005), we define the average Sharpe Ratio as St ≡ E and the risk-free rate as t [Mt+1 ] 1 Et [Mt+1 − 1. We define the equity premium as the average value-weighted return in the simulated economy, where V 765 i,t realized stock returns are computed ex-dividends as Ri,t ≡ Vi,t−1 −D . Profitability is the ratio of operating profits i,t−1 Πi,t to capital Ki,t , investment is the ratio of Ii,t to capital Ki,t , leverage is the ratio of Bi,t to Bi,t + Vi,t − Di,t , book-to-market is the ratio of Ki,t to Bi,t + Vi,t − Di,t , debt adjustments are changes in debt stock Bi,t − Bi,t−1 divided by capital Ki,t−1 , debt issuance is max{0, Bi,t −Bi,t−1 } divided by Ki,t−1 , and equity issuance is max{0, Ei,t } scaled by capital Ki,t . Numerical Solution Method 770 We solve the model using a combination of value function iteration (VFI) and simulation. Each firm i’s problem is characterized by five state variables, i.e. xi,t ≡ (Ki,t , Bi,t , ci,t , zi,t , At ). We approximate the value function V (xi,t ) with piece-wise linear interpolation on a grid 7 × 7 × 5 × 5 × 3, respectively. Note that the projection of V (xi,t ) onto an interpolated structure allows for a precise solution with a relatively parsimonious number of grid points. We check the robustness of our numerical solution by experimenting with finer grids. 775 780 785 Given xi,t , each firm faces three continuous choices, i.e. (Ki,t+1 , Bi,t+1 , ci,t+1 ), and one discrete choice, i.e. whether to default or not I(xt ). Choices Ki,t+1 and Bi,t+1 are evaluated on a grid 30 × 30, and we solve numerically at each step of the VFI for ci,t+1 given each fix evaluation of (Ki,t , Bi,t , ci,t , zi,t , At , Ki,t+1 , Bi,t+1 ). This requires to solve for 7 × 7 × 5 × 5 × 3 × 30 × 30 non-linear equations at each step of the VFI, given the guess of the value function and future default. We solve for the coupon using golden search. The solution with VFI proceed in two steps. First, we find the equilibrium value function Ṽ (i,t ) associated with an identical model but without endogenous default. Second, we use Ṽ (i,t ) as an initial guess for the model with endogenous default and we progressively increase the fixed cost F , re-converge the value function and re-initialize. The convergence criteria on the value function is defined as a max absolute difference between the value functions in two consecutive iterations of the VFI of 10−4 , or lower. We then use the four policy functions {K(xi,t ), B( xi,t ), c(xi,t ), I(xi,t )} to perform a long simulation of our economy (T = 2000) with a panel of i = 1, · · · , 10000 firms, given random draws of idiosyncratic shocks {{zi,t }Tt=0 }10000 and i=1 aggregate shocks {at }Tt=0 . 46 Online Appendix (Not for Publication) OA1. Two-Period Example: Additional Content 790 Notable Special Cases: Sticky Leverage and Modigliani-Miller Equation (9) nests a notable special case. Absent taxes and bankruptcy costs, the first-order condition in (32) simplifies to ∂ E[M (1 + r2D (d1 )]d1 ∂Θ dk0 , dk1 = . (OA1) 1− ∂d1 ∂d1 If debt is fairly priced, the right hand side of (OA1) is equal to one. Therefore, to maximize firm value the firm should not change its level of debt and set d1 = d0 . With this policy the firm does not bear any costs of adjustment and the first derivative of the adjustment cost function is zero. In this case γ(d1 ) = 1 and equation (9) collapses to the celebrated second proposition of MM, that is R2E = R2A + 795 d1 R2A − R2D . k − d1 (OA2) Equation (OA2) should not be read to mean that firms can costlessly change capital structure and returns (as instead is the case in MM). Instead, the meaning of equation (OA2) is that if there is cross-sectional heterogeneity in k and d0 across firms, firms with higher leverage will earn higher returns. Firms are “stuck” at their level of leverage and returns. A simple subcase of this special case is when there are no adjustment costs, in addition to the exclusion of taxes and bankruptcy costs. Equation (OA1) holds trivially and equation (OA2) also holds. We have the original setup of MM, where capital structure is irrelevant and firms are indifferent in their choice of d1 . Geometric Interpretation 800 Figure OA1 provides a representation of the trade-offs in our model and offers a geometric interpretation of the correction factor. To start with, it is useful to reinterpret target leverage as the leverage ratio kd that equalizes the net F +qD ) 34 marginal bankruptcy cost of leverage α kd and the net marginal tax shield of leverage τ (r1+r . In Panel A, which F d refers to the case of an overlevered firm, the net marginal cost of debt α k is represented by the straight line through F +qD ) A and C. The straight line through B and C represents the present value of the marginal tax shield τ (r1+r . The F d∗ 805 810 two lines intersect at point C, in correspondence of target leverage k1 . ∗ d Consider now the representations of the correction factor γ k1 , dk1 in (12), as a function of target leverage and leverage, and γ RL1 , dk1 in (13), as a function of relative leverage and leverage. In (12) and (13), the first term (“1”) is the present value of one dollar of debt in the absence of frictions. The second and third term in both (12) and (13) illustrate the economic tradeoffs at work and their effect on the d∗ correction factor. Notice that their sum α2 dk1 − α k1 in (12) and − α2 dk1 + αRL1 in (13) can both be decomposed as d∗ − α2 k1 + α2 RL1 by simply exploiting the definition of relative leverage in (6). With reference to Figure OA1 Panel A, d∗ d∗ the term α2 k1 is equal to the area of the triangle ABC divided by its base k1 , and accordingly captures the average net surplus from debt that the firm receives when its leverage is equal to target leverage. The surplus arises from the fact that the marginal tax shield is always greater than the marginal bankruptcy cost up to the target. This term enters the definition of γ ( ) with a negative sign, because it decreases the effective cost of debt for the firm. The term α2 RL1 is equal to the area of the triangle CDE divided by the length of its base, RL1 . The term with relative 34 To d∗ 1 k see this, notice that the first-order condition (4) can be simplified using the definition of α in Proposition 2, and target leverage can be obtained as the solution of α d∗ 1 k = τ (rF +qD ) . 1+rF Target leverage is the leverage ratio that firms would choose in the absence of adjustment frictions, in correspondence of which the present value of the marginal tax shield qD the net marginal bankruptcy cost α = 2β(1 − τ )k 1+r . F 47 τ (rF +qD ) 1+rF is equal to the present value of 815 820 leverage captures the average lost surplus at the optimum, which is different from the target. This term carries a positive sign in γ ( ) , because if the optimum is different from the target, the firm is not internalizing all the surplus from its debt and some potential surplus is lost. All else equal, the lost surplus increases the effective cost of debt. The difference between the areas ABC and CDE (or equivalently CD0 E’) is therefore the net benefit to leverage, depicted as the trapezoid ABE 0 D0 in Panel A. If the firm is sufficiently overlevered, the area of CDE becomes larger than that of ABC, resulting in a correction factor that is greater than one. In this case there is a net loss to carrying so much debt, because the average tax shield is smaller than the average net bankruptcy cost. 825 Finally, the term τ δdD1k in (12) and (13) is the product of the tax rate times the recovery rate per unit of debt in case of bankruptcy.35 All else equal, the recovery rate of debt in bankruptcy states lowers the coupon that lenders would otherwise require, and reduces the interest expense and the tax shield in solvency states. This term carries a positive sign, because it captures a loss of tax shields and thus, all else equal, a higher cost of debt. 830 Panel B of Figure OA1 instead refers to the case of an underlevered firm with the same target leverage of the firm in Panel A. Following the same logic, the triangle ABC represents the total surplus that the firm gets from debt when leverage is set equal to the target, and the triangle CDE represents the surplus that the firm loses by choosing an optimum leverage that is below target leverage. For an underlevered firm, the correction factor is smaller than one, because the firm always internalizes a net benefit to leverage, which lowers its effective average cost of debt. Cost of Debt and Leverage 835 The realized return on debt in a given state is ρD i,2 (di,1 ) in case of solvency, and the recovery rate in case of default. By the expression of the coupon in (33) it is immediate to conclude that the return on debt is decreasing with leverage in default states, and increasing in leverage in solvency states. Intuitively, the lender requires a nominal interest rate on the loan such that it is fairly priced, that is E[M (1 + r2D (d1 ))] = 1. The latter no arbitrage condition implies that lenders require higher coupons to offset lower recovery rates. Following standard arguments (see, for example, Cochrane (2001), Chapter 1), the no-arbitrage condition implies that expected return on debt is given by the following expression: E[r2D (d1 )] − rF = −(1 + rF )Cov(M, r2D (d1 )), 840 which emphasizes that, in the presence of risk-averse investors, the expected return on debt increases with leverage if bankruptcy occurs in bad times, in which M takes higher values. In other words, low recovery rates occur in more valuable states, and must be offset by disproportionately higher coupons in solvency states to preserve no arbitrage in the market. Accordingly, higher leverage drives the expected return on debt up, as common intuition suggests. 35 This term and the term (“1”) are not depicted in Figure OA1 because they do not capture marginal effects. 48 OA2. Empirical Evidence: Robustness Checks Alternative Measures of Target Leverage 845 We implement three additional estimations of the target leverage ratio. The first, which we denote as TL2, is still based on the partial adjustment model of Flannery and Rangan (2006) (FR) with firm fixed effects. The second and third alternative measures of target leverage, TL3 and TL4, are non-parametric. We, respectively, construct them using the rolling median leverage at the firm level and at the industry level. More precisely: TL2: Flannery-Rangan with convergence in firm fixed effects. As for TL1, we run a regression specification as in equation (19) that contains fixed effects for each individual firm. Differently from TL1, we guarantee reliable estimates of the fixed effects by excluding firms for which the estimate of the fixed effects does not converge to a stable value. More specifically, stability of the estimate for firm i in year t is achieved if and only if there exists a period t∗ such that the fixed effect estimate Fi,t∗ can be computed and satisfies36 |Fi,t∗ − Fi,t∗ −1 | < 0.05 and |Fi,t∗ − Fi,t∗ −2 | < 0.05. 850 Once the stability criterion is satisfied for firm i, the fixed effect estimate Fi,t∗ at t∗ is used to compute target leverage T Li,t for every year following t∗ and for which equation (19) produces a valid estimate of target leverage for firm i.37 TL3: Firm median leverage with at least five observations. The third measure of target leverage is computed as the rolling median at the firm level for all firms in the sample with at least five observations. TL4: 4-digit SIC code industry median leverage. The fourth measure of target leverage is computed as the four-digit SIC rolling median ML. In this case there is no requirement on the minimum number of observations. 855 860 865 870 A possible concern with the measures of target leverage based on FR is that target and consequently relative leverage are potentially functions of previously identified determinants of equity returns (indeed some of these variables like profitability, size and market-to-book are related to well known asset pricing anomalies). This may make it difficult to interpret our results in the empirical section. To address this concern, TL3 and TL4 are based on a non-parametric approach. Table OA1 shows that descriptive statistics are consistent across our baseline measure of target TL, TL2, and TL3 (0.16-0.17). TL4 yields a lower predicted target (0.13). The TL, TL2, and TL3 have a mean that is close to zero, while RL4 has a slightly larger mean (0.03), which reflects the difference in average between TL4 and ML. Also, consistency in results across measures is reassuring, because it indicates that our estimates of TL are not very sensitive to the exact specification of the leverage measure.38 TL4 is the least stringent measure in terms of data requirements. Accordingly, the sample for which TL4 can be measured is the one that is most similar to the Compustat universe. Panels A, B and C of Table OA2 broadly confirm the patterns observed in Table 2, thus providing a robustness check. Similarly, Table OA3 and Table OA4 are replicas of Table 3 and Table 4, respectively. Overall, the results in Table OA3 and Table OA4 indicate that the relationship between returns and capital structure imbalances is robust to different empirical proxies for target leverage. require that there are no gaps in the firm i’s time series of Fi,s , for t∗ < s < t. ∗ for all years following due to missing data for one or more regressors in Xi,t−1 , it is not possible to estimate M DRi,t t∗ . When this occurs, we temporarily remove the firm from the analysis, and start checking again the stability criterion. 38 In unreported results we find that the average of individual firm-level correlations is high (about 0.87) between TL-TL2 and TL3, and much lower for TL4 (about 0.54). TL and TL2 have a correlation of 1, which stems from the fact that both are estimated with FR. More precisely, TL is estimated only if there are 5 observations, and TL2 is estimated only if stability is achieved. If both measures can be estimated, they yield the same value. This does not mean that having two measures is redundant, because the number of observations in the sample is different across measures. 36 We 37 Occasionally, 49 Book and Net Leverage Different strands of the literature highlight the importance of distinguishing between leverage measured at market or at book values, and the importance of accounting for cash holdings in the balance sheet.39 Bates, Kahle, and Stulz (2009) show that firms in the lower quintiles of size consistently have larger cash-to-assets ratios. Therefore, when leverage is measured net of cash (both for book and market leverage) smaller firms will appear to have less leverage. Additionally, book leverage tends to differ significantly from market leverage for firms with high market-to-book ratios. Therefore, it is possible that by adjusting for cash and by using book values, the distribution of relative leverage changes unevenly in a way that correlates with size and market-to-book ratios. We define book leverage ratio as Di,t , (OA3) BLi,t = Di,t + BEi,t where BEi,t is the book value of equity of firm i in period t. 880 In this spirit, Table OA5 replicates the analysis of Tables 4 and OA4 by looking at market leverage measured net of cash and at book leverage, both gross and net of cash. More precisely, for both debt at market and at book values, we subtract cash and marketable securities (CHE) from total debt. TL and RL are also re-estimated accordingly. Table OA5 reports the coefficients of leverage and relative leverage obtained from a set of FMB regressions that include Size, BM (unreported due to space considerations). Across the three panels, the coefficient of leverage is volatile, taking both positive and negative signs, and is generally insignificant. Instead, relative leverage is positive and significant in all (but one) specification. Similarly, when target leverage is included along with leverage, the results are in line with those in Table 4. 885 From this analysis, we conclude that the measure of leverage that one chooses to employ does not have a first-order impact on the results. This is an important robustness check, because the previous literature has found contrasting results on the relation between leverage and returns, depending on the measure of leverage adopted. Our results suggest that the instability in the results is not driven by the measure of leverage, but by the fact that the relation between leverage and returns is a priori unclear, as discussed in the model section. 875 Additional Controls 890 895 As a robustness check, we also include post-ranking CAPM betas (β), momentum, profitability, and investment in our tests. Post-ranking betas are computed following the procedure in Fama and French (1992). We measure momentum as the continuously compounded return from month m − 12 to month m − 2. Consistent with Novy-Marx (2013) and Hou, Xue, and Zhang (2014) we measure investment (INV) as the growth rate of total assets (AT) and profitability (PROF) as the differences between operating revenues (REVT) and the cost of goods sold (COGS), divided by total assets (AT). Including investment and profitability as controls helps alleviate concerns that the effect of capital structure imbalances on returns is driven by omitted variables that affect cross-sectional differences in E[R2A ] and that are not already captured by the control variables in the baseline specification (20). Table OA6 shows that relative leverage remains largely unaffected by the introduction of the new controls. Both investment and profitability are significant and with the expected sign. Instead, the little explanatory power of CAPM betas (β) should on interpreted with caution. As standard40 , our dynamic model has one source of aggregate risk and predicts a conditional one-factor model, in which leverage variables affect conditional betas. Instead, the rolling estimates of post-ranking betas are unconditional. 39 See Fama and French (1992), Fama and French (2002) and Flannery and Rangan (2006) for the distinction between market and book leverage. See Opler, Pinkowitz, Stulz, and Williamson (1999) and Bates, Kahle, and Stulz (2009) for the evidence on cash holdings. 40 See, for example, Zhang (2005) and Gomes and Schmid (2010). 50 Figure OA1 Correction Factor Decomposition Figure OA1 illustrates the decomposition of the correction factor γ(di,1 ) as discussed in Section 2. The model is solved numerically with three possible cash-flow states (H, M, and L) at time t2 that can occur with probability 0.3, 0.5 and 0.2 respectively. Panel A and Panel B depict two firms with the same target leverage but different initial leverage ratios, namely 0.75 in Panel A and 0.1 in Panel B. In the low state L firms are insolvent with βi,2 (L) = 0.5. In the non-default states, the A tax rate is τ = 0.3. The remaining parameters are as follows: θi = 1, M (H) = 0.8, M (M ) = 1.05, M (L) = 1.1, ri,2 (H) = 0.9, A A ri,2 (M ) = −0.1, Ri,2 (L) = −0.45205, ki = 1. Panel A: Overlevered Firm 0.14 Marginal bankruptcy cost Marginal tax shield D Benefits and Cost ($ per unit of leverage) 0.11 B 0.07 E’ C ”Lost Surplus” (CDE) E ”Surplus at Target” (ABC) Net Benefit to Leverage (ABE’D’) D’ A 0.46 Target Leverage 51 0.71 Optimal Leverage 0.9 Panel B: Underlevered Firm 900 0.14 Benefits and Cost ($ per unit of leverage) Marginal bankruptcy cost Marginal tax shield B 0.07 E C ”Surplus at Target” (ABC) ”Lost Surplus” (CDE) Net Benefit to Leverage (ABED) 0.02 A D 0.15 Optimal Leverage 0.46 Target Leverage 52 0.9 Table OA1 Leverage Decomposition: Descriptive Statistics The table provides summary statistics for our four measures of target leverage and for key firm characteristics. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013. Financial firms and utilities are excluded. ML is the market debt ratio, TL is the estimated target debt ratio, RL is relative leverage, defined as the difference between ML and RL. We consider three alternative measures of target leverage. TL2 is from the estimation of the model of Flannery and Rangan (2006) with firm fixed effects and at least 5 observations for each firm. The estimation of TL2 begins in 1970 (using data since 1965) and is done on a rolling basis as described in Section 3, each year using all the accounting information available until that year. TL2 requires convergence on the estimate of firm fixed effects. TL3 is computed as the rolling median firm leverage for firms with at least 5 observations. TL4 is computed as the rolling median industry leverage at the 4-digit SIC code level. Reported descriptive statistics refer to firm-year observations for which, TL2, TL3, and TL4 are not missing. Panel A reports means, standard deviations, and autocorrelation coefficients for market leverage and for the four measures of target and relative leverage. Panel B reports average market leverage, target leverage, relative leverage, capital stock (K), investment (I/K), profitability (ROA), market capitalization (SIZE), Tobin Q (Q) and the number of observations for the four measures of target leverage. All nominal magnitudes are deflated by the consumer price index, to express all nominal values in 2000 dollars. All variables are winsorized at the 1 percent level and are measured as described in the appendix. Panel A: Leverage Decomposition Mean St. Dev. AC(1) ML: Observed Leverage 0.16 0.17 0.87 TL2: FR with Firm FE (convergence) TL3: Median Firm Leverage (min 5 obs) TL4: Median Industry Leverage 0.16 0.16 0.13 0.15 0.15 0.12 0.90 0.99 0.91 RL2: ML - TL2 RL3: ML - TL3 RL4: ML - TL4 0.00 0.01 0.03 0.10 0.12 0.14 0.66 0.78 0.83 53 Panel B: Firm Characteristics and Sample Selection Measure of TL LEV Leverage TL RL 0.16 K Investment I/K ROA Valuation SIZE Q N 2.49 0.10 0.10 2.64 2.26 81978 TL2 0.16 0.16 0.00 3.00 0.09 0.11 3.17 2.18 62868 TL3 0.17 0.16 0.01 3.04 0.09 0.11 3.22 2.10 61940 TL4 0.16 0.13 0.03 2.42 0.11 0.10 2.60 2.29 75441 54 Table OA2 Firm Characteristics across Quintiles of Relative Leverage The table provides summary statistics for key firm characteristics across quintiles of firms sorted by estimated relative leverage. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013. Financial firms and utilities are excluded. ML is the market debt ratio, TL is the estimated target debt ratio, RL is relative leverage, defined as the difference between ML and RL and ∆ML is the rate of change of market leverage in the following year. We consider three alternative measures of relative leverage. Panel A refers to TL2, which is from the estimation of the model of Flannery and Rangan (2006) with firm fixed effects and at least 5 observations for each firm. The estimation of TL2 begins in 1970 (using data since 1965) and is done on a rolling basis as described in Section 3, each year using all the accounting information available until that year. TL2 requires convergence on the estimate of firm fixed effects. TL3 is computed as the rolling median firm leverage for firms with at least 5 observations. Panel B refers to TL3, which is computed as the rolling median firm leverage for firms with at least 5 observations. Panel C refers to TL4, which is computed as the rolling median industry leverage at the 4-digit SIC code level. The table reports average market leverage, target leverage, relative leverage, capital stock (K), investment (I/K), profitability (ROA), market capitalization (SIZE), Tobin Q (Q) and the number of observations for the three measures of target leverage. All nominal magnitudes are deflated by the consumer price index, to express all nominal values in 2000 dollars. All variables are winsorized at the 1 percent level and are measured as described in the appendix. Panel A: RL2 - FR with Firm FE (convergence) Group of RL ML Leverage TL RL ∆ML K Low 0.11 0.23 -0.12 0.04 1.92 0.09 0.14 2.33 2.18 2 0.11 0.15 -0.04 0.00 1.94 0.09 0.12 2.33 2.43 3 0.11 0.11 0.00 -0.01 1.85 0.08 0.10 1.96 2.47 4 0.15 0.12 0.03 -0.07 3.19 0.07 0.10 2.32 2.19 High 0.30 0.17 0.13 -0.13 6.09 0.08 0.09 2.40 1.57 55 Investment I/K ROA Valuation SIZE Q Panel B: RL3 - Median Firm Leverage (min 5 obs) Group of RL ML Leverage TL RL ∆ML K Investment I/K ROA Valuation SIZE Q Low 0.12 0.26 -0.14 0.03 2.59 0.08 0.15 2.24 2.02 2 0.11 0.15 -0.04 -0.01 2.64 0.08 0.13 2.85 2.37 3 0.09 0.09 0.00 0.07 1.44 0.09 0.10 1.73 2.74 4 0.19 0.15 0.04 -0.10 3.76 0.08 0.11 2.78 1.94 High 0.33 0.16 0.17 -0.15 4.91 0.08 0.09 1.88 1.36 Panel C: RL4 - Median Industry Leverage Group of RL ML Leverage TL RL ∆ML K Low 0.06 0.21 -0.15 0.13 2.37 0.09 0.14 2.11 2.45 2 0.07 0.10 -0.03 0.06 1.89 0.10 0.08 1.91 2.90 3 0.08 0.08 0.00 -0.01 1.75 0.11 0.08 2.03 2.98 4 0.20 0.14 0.06 -0.11 2.97 0.09 0.10 2.28 1.82 High 0.38 0.14 0.24 -0.14 3.13 0.08 0.09 1.11 1.26 56 Investment I/K ROA Valuation SIZE Q 57 6.09 2.03 0.46 8.36 4.35 0.89 7.95 3.79 0.87 8.59 4.56 0.93 Re [t] SR Re [t] SR Re [t] SR ML RL2 RL3 RL4 H-L Re [t] SR Sorting Variable 12.74 3.57 0.72 12.60 3.30 0.65 11.56 2.91 0.56 15.20 3.41 0.66 L 13.68 3.57 0.71 14.15 3.88 0.77 14.97 3.89 0.75 14.55 3.65 0.73 15.71 3.48 0.66 14.66 3.80 0.75 15.31 4.02 0.79 14.60 3.92 0.77 Equally-Weighted 3 5 15.44 3.82 0.76 14.74 3.81 0.75 15.72 4.09 0.82 15.85 4.09 0.82 7 21.33 4.31 0.93 20.55 4.23 0.90 19.92 4.32 0.93 21.29 4.10 0.91 H Excess Returns 2.69 1.15 0.19 5.31 2.14 0.37 5.24 1.85 0.33 1.73 0.51 0.10 H-L 8.52 2.80 0.51 6.34 1.78 0.33 5.11 1.30 0.24 7.80 1.99 0.35 L 7.48 2.62 0.47 8.83 2.77 0.50 8.91 2.68 0.49 9.07 3.11 0.54 8.51 2.69 0.46 7.77 2.28 0.43 9.71 3.46 0.61 7.56 2.84 0.49 Value-Weighted 3 5 9.26 3.13 0.55 9.02 3.23 0.53 7.88 2.50 0.46 11.33 3.76 0.67 7 11.21 2.65 0.49 11.65 3.01 0.55 10.35 3.13 0.56 9.53 2.30 0.44 H In the table stocks are sorted every June in deciles based on their values of market leverage (ML) and relative leverage (RL). RL is obtained from the estimation of TL from the model of Flannery and Rangan (2006) with firm fixed effects and at least 5 observations for each firm. The estimation of TL begins in 1970 (using data since 1965) and is done on a rolling basis as described in Section 3, each year using all the accounting information available until that year. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013 and covered by the Center of Research in Security Prices (CRSP). Financial firms and utilities are excluded. The breakpoints for portfolio sorts are computed on the subset of firms traded on the NYSE market. We consider three alternative measures of relative leverage, as described in Table OA1. The table reports returns in excess of the risk-free rate, t-statistics and Sharpe ratios for the bottom decile (L), the top decile (H) and for the third, fifth and seventh decile. We also report the difference for the excess returns, the t-statistics and the Sharpe ratio between the top decile and the bottom decile (H-L). The left panel reports equally-weighted returns, while the right panel reports value-weighted returns. The sorting variables are matched to monthly returns as described in Section 4. Table OA3 Univariate Portfolio Sorts Table OA4 Levered Returns: Relative, Observed and Target Leverage For each month between July 1980 and December 2013, we estimate cross-sectional regressions of stock returns on relative leverage (RL), market leverage (ML), target leverage (TL), market capitalization (SIZE), and book-to-market equity (B/M). We consider three alternative measures of target and relative leverage, as described in Table OA1. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013 and covered by the Center of Research in Security Prices (CRSP). Financial firms and utilities are excluded. De-listing returns are included in monthly returns. The table reports Fama-MacBeth coefficient estimates. t-statistics are in parentheses. R2 and N denote the cross-sectional R-squared and the number of observations respectively. The independent variables are matched to monthly returns as described in Section 4. All variables are described in the appendix. The symbols (***), (**) and (*) denote statistical significance at the 1, 5 and 10 percent levels respectively. Panel A: Observed and Relative Leverage (1) ML (2) (3) 0.14 (0.47) 0.78∗∗∗ (2.78) RL2 0.86∗∗∗ (3.54) RL3 -0.24∗∗∗ (-5.28) 0.17∗∗ (2.46) -0.20∗∗∗ (-4.78) 0.16∗∗ (2.11) -0.20∗∗∗ (-4.86) 0.14∗ (1.90) 0.56∗∗∗ (2.72) -0.23∗∗∗ (-5.19) 0.17∗∗ (2.09) 0.02 931438 0.02 738975 0.02 730793 0.02 856307 RL4 SIZE B/M R2 N (4) Panel B: Relative Leverage vs Observed Leverage ML RL2 (1) (2) (3) -0.09 (-0.27) 0.85∗∗∗ (2.86) -0.09 (-0.28) -0.55 (-1.00) 0.89∗∗∗ (2.98) RL3 905 -0.20∗∗∗ (-4.90) 0.16∗∗ (2.44) -0.20∗∗∗ (-4.94) 0.15∗∗ (2.24) 1.03∗∗ (2.03) -0.23∗∗∗ (-5.30) 0.20∗∗∗ (3.05) 0.02 738975 0.02 730793 0.02 856307 RL4 SIZE B/M R2 N 58 Panel C: Observed Leverage vs Target Leverage ML TL2 (2) (3) (4) 0.77∗∗ (2.37) -0.85∗∗∗ (-2.87) 0.80∗∗∗ (2.81) 0.49∗∗ (2.22) -0.89∗∗∗ (-2.98) TL3 -0.20∗∗∗ (-4.90) 0.16∗∗ (2.44) -0.20∗∗∗ (-4.94) 0.15∗∗ (2.24) -1.03∗∗ (-2.03) -0.23∗∗∗ (-5.30) 0.20∗∗∗ (3.05) 0.02 738975 0.02 730793 0.02 856307 TL4 SIZE B/M R2 N 59 60 TL RL LEV TL RL LEV 0.02 (0.09) (LEV) -0.14 (-0.58) (LEV) -0.14 (-0.52) 0.70∗∗∗ (2.66) 0.02 (0.08) 0.52∗∗ (2.03) (RL2) 0.41∗ (1.81) 0.47∗∗ (2.23) (RL1) -0.06 (-0.24) (RL2) -0.06 (-0.26) (RL1) 1.01∗∗ (2.34) -0.84∗ (-1.66) (RL4) 0.48∗∗ (2.21) -0.03 (-0.09) (RL3) 0.91∗∗ (2.04) -0.57 (-1.16) (RL4) Panel B: Book Leverage 0.63∗∗∗ (2.84) -0.20 (-0.74) (RL3) Panel A: Net Market Leverage 0.56∗∗ (2.03) -0.70∗∗∗ (-2.66) -0.52∗∗ (-2.03) (TL2) -0.41∗ (-1.81) 0.35 (1.41) (TL2) 0.54∗ (1.96) (TL1) -0.47∗∗ (-2.23) 0.41∗ (1.70) (TL1) -0.48∗∗ (-2.21) 0.45∗ (1.96) (TL3) -0.63∗∗∗ (-2.84) 0.43∗∗ (2.07) (TL3) -0.91∗∗ (-2.03) 0.34∗ (1.91) (TL4) -1.02∗∗ (-2.34) 0.18 (1.07) (TL4) For each month between July 1980 and December 2013, we estimate cross-sectional regressions of stock returns on relative leverage (RL), market leverage (ML), target leverage (TL), market capitalization (SIZE), and book-to-market equity (B/M). On top of our baseline measure of TL from the model of Flannery and Rangan (2006), which we denote as TL1, we consider three alternative measures of target and relative leverage, as described in Table OA1. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013 and covered by the Center of Research in Security Prices (CRSP). Financial firms and utilities are excluded. De-listing returns are included in monthly returns. The table reports Fama-MacBeth coefficient estimates. t-statistics are in parentheses. The independent variables are matched to monthly returns as described in Section 4. All variables are described in the appendix. The symbols (***), (**) and (*) denote statistical significance at the 1, 5 and 10 percent levels respectively. Table OA5 Book and Net Leverage 61 TL RL LEV -0.13 (-0.63) (LEV) -0.17 (-0.86) 0.53∗∗∗ (2.93) 0.54∗∗∗ (3.04) (RL2) -0.10 (-0.46) (RL1) 0.52∗∗∗ (3.21) -0.12 (-0.53) (RL3) 0.78∗∗∗ (2.74) -0.58∗ (-1.68) (RL4) Panel C: Net Book Leverage 0.37∗ (1.80) -0.53∗∗∗ (-2.93) -0.54∗∗∗ (-3.04) (TL2) 0.45∗∗ (2.26) (TL1) -0.52∗∗∗ (-3.21) 0.41∗∗ (2.50) (TL3) -0.78∗∗∗ (-2.74) 0.19 (1.53) (TL4) 62 R2 N Momentum β B/M SIZE PROF INV RL4 RL3 RL2 RL1 ML 0.04 925112 -0.26∗∗∗ (-5.26) 0.06∗∗∗ (5.22) -0.22∗∗∗ (-5.59) 0.20∗∗∗ (3.45) 0.02 (0.09) 0.12 (0.68) -0.09 (-0.39) (1) 0.04 726056 -0.26∗∗∗ (-4.24) 0.05∗∗∗ (3.14) -0.19∗∗∗ (-4.98) 0.16∗∗∗ (2.76) 0.18 (0.18) 0.05 (0.87) -0.17 (-0.74) 0.95∗∗∗ (3.41) (2) 0.04 737667 -0.22∗∗∗ (-3.54) 0.06∗∗∗ (4.74) -0.19∗∗∗ (-4.99) 0.19∗∗∗ (3.35) 0.06 (0.26) 0.11 (0.59) 0.94∗∗∗ (3.32) -0.33 (-1.29) (3) 0.04 729644 -0.28∗∗∗ (-4.71) 0.05∗∗∗ (3.19) -0.19∗∗∗ (-4.99) 0.17∗∗∗ (2.87) 0.02 (0.08) 0.15 (0.83) 1.01∗∗∗ (3.99) -0.32 (-1.28) (4) 0.04 850563 1.14∗∗∗ (2.91) -0.26∗∗∗ (-5.20) 0.06∗∗∗ (5.33) -0.22∗∗∗ (-5.51) 0.22∗∗∗ (4.05) -0.05 (-0.22) 0.10 (0.53) -0.87∗∗ (-2.03) (5) For each month between July 1980 and December 2013, we estimate cross-sectional regressions of stock returns on relative leverage (RL), market leverage (ML), target leverage (TL), investment (INV), profitability (PROF), market capitalization (SIZE), book-to-market equity (B/M), post-ranking CAPM beta (β) and momentum (Momentum). Investment is measured as the growth of firm assets. Profitability in measured as the ratio of operating profits to book equity. On top of our baseline measure of TL from the model of Flannery and Rangan (2006), which we denote as TL1, we consider three alternative measures of target and relative leverage, as described in Table OA1. The sample includes all Compustat firms traded on NYSE, AMEX and NASDAQ between 1965 and 2013 and covered by the Center of Research in Security Prices (CRSP). Financial firms and utilities are excluded. De-listing returns are included in monthly returns. The table reports Fama-MacBeth coefficient estimates. t-statistics are in parentheses. R2 and N denote the cross-sectional R-squared and the number of observations respectively. The independent variables are matched to monthly returns as described in Section 4. All variables are described in the appendix. The symbols (***), (**) and (*) denote statistical significance at the 1, 5 and 10 percent levels respectively. Table OA6 Additional Controls