Dynamic Capital Structure Modelling
under Alternative Stochastic Processes
by
Simon Seir Bjerrisgaard
and
Denis Fedoryaev
Master’s Thesis
Presented to the Faculty of
the Department of Finance of Copenhagen Business School
in Partial Fulfilment of the Requirements for the Degree of
Elite MSc in Advanced Economics and Finance
(Cand.Oecon.)
Supervisor: Kristian R. Miltersen
Copenhagen Business School
July 2011
No. of pages (characters): 125 (242,958)
Executive Summary
A coherent line of development of research focusing on modelling optimal
capital structure is presented. We begin by revisiting the pricing model in-
troduced in the seminal work of Merton (1974), which represents the funda-
mental building block for most capital structure models. The general partial
differential equation is rederived, and it is demonstrated how it is applied to
price debt and equity as contingent claims. The first stepping stone to our
analysis is the static model along the lines of Leland (1994). We present com-
parative statics for optimally chosen leverage, coupon, and default threshold
as well as for tax advantage to debt, and discuss the drawbacks of the model
in detail. It is evident that under our own base case parameters—based on
the most recent Danish data and empirical estimates—which differ from un-
reasonably high values applied in Leland, the optimal leverage is too high
compared to that observed in practice. We proceed by introducing dynam-
ics in the model of capital structure; rather than focusing on any particular
model, we outline a generalised framework to describe the overall family of
existing dynamic capital structure models. The state variable is the operat-
ing income, and it is the value of a claim to the entire payout of the firm that
is modelled as a stochastic process, which restores the no-arbitrage condition
violated in the static model. As the leverage can be readjusted, the restruc-
turing boundary is incorporated, stipulating when refinancing should take
place, and is obtained from the smooth-pasting condition to ensure incen-
tive compatibility. The key insight from the comparative statics analysis is
that the optimal leverage drops substantially compared to that in the static
i
ii S. Bjerrisgaard and D. Fedoryaev
model, and the default boundary is also lowered, reflecting the fact that the
refinancing option, ceteris paribus, enhances the firm valuation.
Further, we propose our own dynamic model which alters the cash flow
process by assuming mean reversion. The vast majority of existing capital
structure studies rests on the assumption that the state variable follows a ge-
ometric Brownian motion since this ensures mathematical tractability while
modelling the capital structure readjustment. We argue that this underly-
ing process due to its properties is not consistent with dynamics of the firm
fundamental and that the assumption of mean reversion appears to be more
suitable. Not only the latter describes the development of the real sector
better, but is also reinforced by empirical evidence. Besides, by assuming
that earnings follow the mean-reverting process, we obtain a better control
over the process and can thus draw a much clearer distinction between the
industries in terms of both profitability and stability—through varying the
long-term mean and the speed of mean reversion, respectively. As a matter of
fact, optimal capital structure has never been modelled dynamically outside
of the scope of a geometric Brownian motion, presumably due to the loss of
homogeneity property. We assume the modified mean-reverting process with
volatility being proportional to the current earnings level, and derive the or-
dinary differential equation used for pricing debt and equity. Further, using
contingent claims analysis and state pricing, we prove that the new process
still possesses the homogeneity property which implies that the mechanics
of the model is unaltered in time. After that we describe what we call the
optimal capital structure decision framework and formulate how the tuple
that closes it is obtained, which fully determines the capital structure choice
of the owner-manager.
Extensive numerical simulations are conducted, with explicit focus on
numerical algorithms and their shortcomings. We bring together the key op-
timal capital structure frameworks and carry out a cross-model comparison
to uncover the implications that different assumptions have for the results.
When benchmarking our model against the static mean-reverting model of
iii
Sarkar and Zapatero (2003), we find lower optimal leverage and optimal
coupon. If compared to the conventional dynamic GBM-based model, our
model suggests lower optimal leverage and higher restructuring frequency,
stipulated by the difference in expectations of equity holders regarding the
future development of cash flow under the two processes. This relationship is
even more pronounced if we correct for the finiteness of the stationary vari-
ance of the mean-reverting process. Overall, given reasonable assumptions
for base case parameters, the optimal leverage in our model is found to be
closer to the empirical regularities than that in the existing models. All re-
lationships are studied through the prism of mean reversion, and it is shown
that the speed of earnings convergence and the long-term mean value of earn-
ings are indeed important parameters as their impact on the key variables
could be rather substantial. Finally, we extend the analysis by modifying
the GBM-based model to take into account fixed operating costs, so that the
resulting cash flow is not bounded at zero. We demonstrate that higher op-
erating leverage implies lower financial leverage, and find that when the firm
is allowed to have negative cash inflow, it will be less levered compared to an
otherwise identical firm whose earnings are instead mean-reverting but al-
ways positive. Further, we incorporate a call premium in the dynamic model
with mean reversion and show that assuming debt being callable at par may
lead to somewhat understated leverage and too high restructuring frequency.
Keywords: Dynamic capital structure; Optimal bankruptcy; Optimal restruc-
turing; Contingent claims analysis; Mean-reverting earnings; Homogeneity
property
JEL classification: G12; G13; G32; G33; C60; C61; C63
Acknowledgements
We wish to acknowledge with profound gratitude the counsel and constant
encouragement of Kristian Miltersen, whom we are much obliged for numer-
ous conversations, constructive suggestions and help on the coding. We are
also thankful to Michael Genser for providing the code for results in Am-
mann and Genser (2004), which gave us some insight on specific aspects of
the sphere considered. Finally, we would like to thank Sudipto Sarkar and
Fernando Zapatero for the comments on their model, from which we could
draw additional inspiration for our own analysis. We are solely responsible
for all remaining errors.
iv
Contents
1 Introduction 1
1.1 Basic Framework . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Contingent Claims Analysis 15
2.1 Stochastic Processes . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Ito’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Note on Risk Neutrality in Continuous Time . . . . . . . . . . 23
2.4 The General Partial Differential Equation . . . . . . . . . . . 25
2.5 Isomorphism to Option Pricing . . . . . . . . . . . . . . . . . 29
3 Static Model of Capital Structure 34
3.1 Time Independence of Cash Flows . . . . . . . . . . . . . . . . 35
3.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.1 Solution to the ODE . . . . . . . . . . . . . . . . . . . 37
3.2.2 Valuation of Claims . . . . . . . . . . . . . . . . . . . . 39
3.2.3 Optimal Leverage . . . . . . . . . . . . . . . . . . . . . 44
3.3 Comparative Statics . . . . . . . . . . . . . . . . . . . . . . . 45
3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4 Family of Existing GBM-based Dynamic Models 50
4.1 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 Comparative Statics . . . . . . . . . . . . . . . . . . . . . . . 57
v
vi S. Bjerrisgaard and D. Fedoryaev
4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5 Proposed Dynamic Model with Mean Reversion 65
5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.2 Homogeneity Property . . . . . . . . . . . . . . . . . . 74
5.2.3 Optimal Capital Structure Decision Tuple . . . . . . . 80
5.3 Numerical Results and Model Implications . . . . . . . . . . . 87
5.3.1 Note on Numerical Simulations . . . . . . . . . . . . . 87
5.3.2 Numerical Analysis . . . . . . . . . . . . . . . . . . . . 90
5.4 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.4.1 Fixed Operational Costs . . . . . . . . . . . . . . . . . 103
5.4.2 Call Premium . . . . . . . . . . . . . . . . . . . . . . . 108
6 Conclusion 111
References 116
Chapter 1
Introduction
And make sure that the capital structure we have in place is the right capital
structure. I think that’s the reason that we’ve been successful.
(Henry R. Kravis, co-founder of KKR)
How to finance operations and new investment projects is a problem that
CFOs and managers of corporations frequently encounter. The primary rea-
son why it is beneficial to develop a specific leverage policy is that the optimal
combination of debt and equity creates value for the owners of business by
lessening the tax base. Exploiting the tax benefits of debt financing in the
optimal way could help managers potentially increase the firm value by as
much as 10% (Graham, 2000).
The importance of the capital structure choice for the company valuation
has been one of the central questions in academia since the very seminal
work of Modigliani and Miller (1958). Since then, numerous authors have
contributed to the formation of the optimal capital structure study sphere
by introducing different methods and utilising various assumptions. Some
of the earlier attempts to price corporate securities were based on the firm
value dynamics, while subsequent models instead assumed that the manage-
rial expectations regarding the future development of earnings should be the
primary driver. Initially, the papers, aimed to determine the optimal com-
1
2 S. Bjerrisgaard and D. Fedoryaev
position of debt and equity, ignored the firm’s option to change the financing
policy in time, and a comprehensive dynamic framework was not presented
until Fischer, Heinkel, and Zechner (1989a). Moreover, a model allowing
the company to change its leverage while also assuming earnings to be the
governing factor was not introduced until Goldstein, Ju, and Leland (2001).
What brings the capital structure models existing hitherto together is that
irrespective of the underlying factor choice the process describing its dynam-
ics implicitly assumes that the value of the factor always stays positive and
has only a positive trend. Since such assumption seems rather unfeasible,
especially when considering earnings, this leaves some scope for developing a
dynamic capital structure model based on an alternative stochastic process
to rectify these shortcomings.
The paper proceeds as follows. In the rest of this chapter we dwell on the
terminology used throughout the thesis and state the assumptions implic-
itly incorporated in the analysis that follows, pose our research questions,
and provide a literature review on capital structure modelling. Chapter 2
presents necessary theoretical insights from stochastic calculus and highlights
the importance of contingent claims analysis for the theory of capital struc-
ture choice. In Chapter 3 we lay out a static model of capital structure
and analyse it in detail. Chapter 4 follows with a study of a dynamic cap-
ital structure model which serves as a classic example of the conventional
dynamic framework. In Chapter 5 we introduce an alternative assumption
regarding the dynamics of earnings, and thoroughly develop a dynamic cap-
ital structure model, providing corresponding analysis and numerical tests.
Chapter 6 concludes.
1.1 Basic Framework
As there might exist ambiguity in interpretation of the notions and concepts
of capital structure theory, we will first briefly outline some terminology that
will apply throughout. One should note, however, that since the terminology
Chapter 1. Introduction 3
within this field of academic research is not uniform, the below should merely
be viewed as a guide to this thesis rather than as a universal introduction.
Capital Structure
Capital structure is the consensus term for a company’s composition of debt
and equity, and thus specifies the sources of the capital from which a firm
finances its operations and investments. Debt must be serviced irrespectively
of the success of the firm, while equity only has a residual claim to the cash
generated by the company’s operations. This difference in obligation means
that the cash is spent servicing debt payments before it can be returned to
equity holders, and results in a higher return on equity than on debt.
Equity in accounting terms is the sum of common stock, preferred stock,
any capital surplus, retained earnings, treasury stock, stock options, and
reserves; throughout the thesis we will not distinguish between these com-
ponents and will instead consider equity an aggregate parameter represented
only by common stock. We implicitly assume the following: there is no
subordination among shares, no part of earnings is retained as they are paid
out in full to shareholders as dividends, no additional paid-in capital is raised
from equity investors, the company is not holding shares in itself or any other
company, and no reserves are held. We also ignore the fact that shares with
different cash flow or control rights might exist.
Corporate debt is normally comprised of bonds and loans, with the latter
being either syndicated or bilateral facilities and having either revolving or
set instalment repayment schedule. Bonds have a fixed maturity, and thus
usually cannot be paid down gradually, but both loan and bond contracts can
include call provisions allowing the debtor to pay the principal back early. In
this thesis we will exclude loans from the analysis, as the pricing technique
of debt securities that will be presented later implies that the security must
be traded. Despite the existence of the secondary loan market, due to its
relatively small size we will treat the value of debt as being synonymous with
the value of outstanding bonds.
Based on the above, capital structure can be summarised to exhibit in
4 S. Bjerrisgaard and D. Fedoryaev
what proportion the company financing is split between common shares and
bonds. The ratio of debt to the sum of market values of debt and equity
is termed “leverage”. Optimality of capital structure stems from the max-
imisation of the equity value as the financing policy is at sole discretion of
shareholders. Note that since in the models presented in further chapters all
existing debt is retired before the optimal capital structure is derived, the
optimal leverage in fact maximises the firm value.
State Variable
In the models considered in this paper, the optimal leverage is calculated by
utilising contingent claims pricing. The underlying variable which the value
of these claims is dependent on is termed the state variable, and typically
represents a firm fundamental, e.g. value of unlevered assets, earnings, or
another proxy. In Chapters 2 and 3 the state variable is assumed to be the
asset value of the firm, while models in Chapters 4 and 5 consider EBIT
instead.
Default
Bankruptcy is a legal proceeding involving a company that is unable to re-
pay its outstanding debt. A default occurs when the firm has not met its
legal obligation according to the bond contract, i.e. when it has not met a
scheduled payment, or alternatively, has violated a covenant—the latter be-
ing termed technical default. Furthermore, a default can occur not only when
a firm is unable to pay, but also when it is unwilling to pay—the situation
known as strategic default. The unwillingness to pay is a choice made by
equity holders and might be regarded as their last resort to save some value
to themselves after repaying the principal and possibly incurring bankruptcy
costs.1
1 Interestingly, even though strategic debt service has been conventionally considered toresult in lower debt value and higher yield spreads (Anderson and Sundaresan, 1996; Mella-Barral and Perraudin, 1997), Acharya, Huang, Subrahmanyam and Sundaram (2006)demonstrate that when managers are not forced to pay out excess earnings as dividends,the ability to service debt strategically could reduce the likelihood of liquidity-driven de-faults (the extent of such reduction is increasing in firm’s cost of capital), leading tonegligible effect on debt value and yield spread.
Chapter 1. Introduction 5
We will make no distinction between default and bankruptcy, despite the
obvious discrepancy in the interpretation of the two terms. It should be
noted, however, that the way bankruptcy is modelled will slightly change
throughout the paper. One of the possible assumptions is that bankruptcy
occurs when EBIT reaches an exogenously given critical value; another ap-
proach is that bankruptcy occurs when the state variable reaches an endoge-
nously determined threshold at which it is no longer in the best interest of
equity holders to keep servicing the debt, and the company is liquidated.
The latter is the assumed bankruptcy procedure under the static model in
Chapter 3. Alternatively, the firm can be taken over post-default as a going
concern, and we consider this option in Chapters 4 and 5.
Let us now turn to formulating the general setup of our analysis, delin-
eated by specific confining assumptions that will be implicit throughout the
thesis.
In practice bond contracts involve a great variety of contractual provi-
sions. Despite that, we will refrain from including the following common
bond indenture provisions and liability characteristics in our framework:
Collateral. Collateral is the borrower’s pledge of specific property that
must be forfeited in case of default, and is often included to serve as
a protection for creditors. By excluding this provision we do not have
to consider additionally the transfer and division of collateral upon
default.
Explicit Maturity. Maturity refers to the final repayment date at
which point the principal must be redeemed in full. As will be shown
later, we assume a sufficiently long maturity, so that coupon payments
are essentially the only debt-related cash outflows.
Subordination. Subordination refers to the order of priorities in
claims for ownership of assets. In our model at no point in time a
firm could have two different debt issues outstanding, and therefore
there is no need to model subordination.
6 S. Bjerrisgaard and D. Fedoryaev
Debt Renegotiation. It has been shown by Gilson, John, and Lang
(1990) in a study of 169 financially distressed US firms that around 53%
of the firms went into bankruptcy, while the rest successfully renegoti-
ated the terms of the debt. In this paper, however, we will abstain from
examining debt renegotiation possibilities since this particular aspect
has already been studied extensively in the literature.2
In addition to the restrictions stated above, we will further limit the scope
of our analysis by making the following assumptions:
No Asset Substitution.3 This implies that the investment deci-
sions are unaffected by capital structure, and particularly that the risk
profile is independent of leverage. Note that we also exclude the un-
derinvestment possibilities of Myers (1977) type from our analysis, i.e.
any value-increasing investment opportunity is being accepted by the
equity holders.4
Full Advantage of Tax Shield. It is assumed that all interest ex-
penses can be deducted from the taxable income, even if the earnings
fall below the coupon level, and thus it is implied that the value of the
tax shield could be obtained as a perpetual stream of savings, condi-
tional on the firm remaining solvent.
Constant Risk-free Interest Rate. The risk-free rate has a bearing
on the discounted value of cash flows. Despite the stochastic nature
observed in practice, we assume that this parameter is constant to
simplify the matters.
2 For a theoretical review of how the possibility of renegotiating either coupon orboth coupon and principal affects debt value, see Anderson and Sundaresan (1996) orChristensen, Flor, Lando, and Miltersen (2002), respectively.
3 Jensen and Meckling (1976) present a theoretical analysis of the asset substitutionproblem in their seminal paper.
4 Various studies have also found weak support for asset substitution and underinvest-ment. See, e.g. Graham and Harvey (2001) or Ju and Ou-Yang (2006).
Chapter 1. Introduction 7
Furthermore, the reader should pay attention to the fact that the abso-
lute priority rule is taken for granted at all times.5 In case assumptions in
addition to those stated above are required, they will be pointed out where
appropriate.
1.2 Problem Formulation
In our analysis we will generally be focusing on two aspects: how to model
optimal capital structure in a dynamic setting, and what effect alternative
assumptions—for instance, regarding the process describing the state variable
dynamics—will have on the composition of debt and equity.
Since empirical evidence shows that firms do indeed alter their financing
policy (e.g. Leary and Roberts, 2005), we find it necessary to incorporate the
option to readjust the capital structure dynamically. Furthermore, allowing
firms to continuously adjust their capital structure according to the changes
in the state variable is more intuitive compared to the static setup, where
leverage is permanent and thus reflects an assumed knowledge of the firm’s
development far into the future. In addition to making the optimal capital
structure dynamic, we aim at utilising an alternative stochastic process to
describe the evolution of the state variable. This is carried out to rectify some
of the drawbacks of the underlying process applied in the majority of existing
models. Apart from that, to make a progression to our model clearer, we
intend to analyse the mechanisms behind existing frameworks, focusing on
their economic interpretation and pointing out the main shortcomings.
Due to multifaceted aspects of corporate financing and restrictive as-
sumptions that need to be made to enable comprehensible derivations, we do
not intend to propose an entirely universal framework, but rather to build a
model that works under certain restrictions. We recognise that a great vari-
5 Even though there exist rational explanations for violation of the absolute priorityassumption (see, for example, Christensen et al., 2002), we choose to adhere to the as-sumption of absolute priority as the effect of deviation on leverage and yield spread hasbeen shown to be marginal, even when different bankruptcy determinants are considered.See, for instance, Leland (1994, Section VI).
8 S. Bjerrisgaard and D. Fedoryaev
ety of factors can contribute to the explanation of observed leverage ratios,
but we will nevertheless limit the scope of our analysis to a specific number
of explanatory parameters. We also respect that the modelling approach
may depend on the type of the state variable considered, but still find it nec-
essary to investigate what implications alternative assumptions have for the
capital structure decision, thereby either reinforcing or undermining previous
findings.
Thus, the main objective of this thesis is to provide answers to the fol-
lowing pivotal questions:
How can a valuation framework to price any corporate security contin-
gent on a firm fundamental be obtained?
What are the main implications of allowing dynamic refinancing in
the traditional model of optimal capital structure which assumes that
earnings follow a geometric Brownian motion?
How can optimal capital structure be modelled dynamically under the
assumption that earnings do not follow a conventional geometric Brow-
nian motion?
Assuming mean reversion in earnings, what inferences could be drawn
from benchmarking the results of the proposed dynamic model against
those in the existing static model?
How does the optimal capital structure decision change when earnings
are assumed to follow a mean-reverting process in a dynamic setup,
compared to that in existing GBM-based models?
If a dynamic GBM-based model with earnings that are allowed to ob-
tain negative values due to the introduction of fixed operating costs is
considered, how does optimal leverage compare to that under a dynamic
model with earnings that exhibit mean reversion?
Chapter 1. Introduction 9
There are naturally numerous ways to address the questions stated above,
both empirically and theoretically. However, it should be emphasised that
we will restrict our focus to the theoretical analysis combined with numerical
simulations, thus leaving out empirical tests of the proposed model.
1.3 Literature Review
With the basic setup and problem formulation in place, a brief review of
literature concerning the topic of choice is necessary to demonstrate how the
research sphere has evolved through time and what theoretical advancements
we could base our analysis upon.
The question of whether capital structure matters is most famously ad-
dressed by Modigliani and Miller (1958), who argue that the market value
of a firm is independent of its capital structure in a frictionless world. The
idea behind this argument is that investors can freely trade in the finan-
cial markets to make their personal portfolio in accordance with a preferred
capital structure, thus eliminating the need for companies to alter their cap-
ital structures. However, after relaxing the assumption of the absence of
frictions, Modigliani and Miller (1963) found that when companies are al-
lowed to deduct interest expenses from taxable income, the optimal capital
structure under certain conditions is to be fully levered, i.e. firms should
be financed solely by debt. As such examples are virtually non-existent in
practice, many of the subsequent research studies were dedicated to prove
that this conclusion does not hold under a variety of circumstances, includ-
ing the introduction of agency costs, bankruptcy costs, transaction costs or
other assumptions that could be considered when trying to explain observed
leverage ratios.
The academic research following Modigliani and Miller provided different
approaches to explain the determinants of capital structure, and in general
they could be divided into two groups: asymmetric information models and
the trade-off theory models. Myers and Majluf (1984) based their pecking
10 S. Bjerrisgaard and D. Fedoryaev
order theory on the concept of asymmetric information between managers
and investors. The idea behind their work was that managers might be forced
to give up positive-NPV projects in case outside financing is required, and
that internal financing (retained earnings) is preferred to external funding
sources. If outside financing is the only option, then debt is favoured due to
the incentives of the manager to issue new shares only when she expects to
receive a high price, which from investors’ viewpoint would mean that the
stock is overpriced. The new shareholders, knowing this, will adjust the price
they are willing to pay for the new shares, reducing the proceeds from the
issue and diluting the value of existing shares. The manager, acting on behalf
of existing shareholders, will therefore prefer to finance the new project with
debt, which is reinforced by the result that the value of debt is less sensitive
to the private information of the manager than the value of equity.6
A classic alternative to the asymmetric information approach is Kraus
and Litzenberger (1973), who state that optimal capital structure is deter-
mined by balancing the costs and benefits of debt. The benefit of debt is
that companies can deduct the interest expenses from their taxable income,
thus shielding their earnings, while the costs of debt financing are generally
recognised as being the expected direct and indirect bankruptcy costs. When
balancing the costs and benefits of moving towards a target capital structure,
another important factor to consider is adjustment costs, which represent a
potential reason for a wide variation in observed leverage ratios among simi-
lar firms, and might help explain why a firm would deviate from its optimal
capital structure (Myers, 1984).
A drawback of the theories focusing on asymmetric information is that
despite giving indications of which factors need to be considered, they do
not provide an explicit advice on what the leverage level should be. Models
based on the trade-off theory, on the other hand, have been more successful
in determining optimal values of debt and equity, and it has been confirmed
6 This result is derived from the option pricing theory, and rests on more strict as-sumptions than the model presented in Myers and Majluf (1984). See Galai and Masulis(1976).
Chapter 1. Introduction 11
that this theoretical foundation works fairly well at predicting typical lever-
age levels (e.g. Ju, Parrino, Poteshman, and Weisbach, 2005).7 Subsequent
continuous-time modelling approaches originating from this strand of liter-
ature can be generally categorised as either structural frameworks or more
recently adopted reduced-form frameworks, with Jarrow, Lando, and Turn-
bull (1997) and Duffie and Singleton (1999) as examples of papers studying
reduced-form models, and Merton (1974), Black and Cox (1976), Leland
(1994) and Goldstein et al. (2001) as those dealing with structural models.
Duffie and Lando (2001) present an example of a structural model that is
consistent with reduced-form representation, and thus provide a link between
the two frameworks.
One of the distinctions between structural and reduced-form models con-
cerns the information available to the modeller. Reduced-form models re-
quire a limited information set similar to what is observed by the market,
compared to structural models which assume availability of more compre-
hensive information, akin to what is observed by the manager. Further-
more, reduced-form frameworks are to a greater extent concerned with the
pricing of corporate debt, rather than with explaining optimal capital struc-
ture or debt structure (maturity, subordination, etc). Finally, but perhaps
most importantly, it should be mentioned that structural models link de-
fault explicitly to the first time that the state variable value falls below a
certain level, while reduced-form frameworks model bankruptcy through the
default probability distribution based on publicly available information, e.g.
accounting information, credit rating or business cycle data. Such feature
of the reduced-form approach makes it particularly appealing for modelling
bonds and credit derivatives. However, since our focus is on modelling the
capital structure choice rather than credit risk, we will restrict our attention
7 It is important to acknowledge that the trade-off theory still cannot explain all theobserved differences in capital structure across firms. For instance, in an internationalcomparison Wald (1999) found that the most profitable firms borrowed the least, theexact opposite of what is predicted by the trade-off theory. For a further comparison ofhow the pecking order and the trade-off theories perform empirically, see Frank and Goyal(2003) or Cotei and Farhat (2009).
12 S. Bjerrisgaard and D. Fedoryaev
to structural models.
Structural models commence with an assumption regarding the dynamics
of the state variable. Given this, it is possible to find the price of debt
and equity as the solution to a partial differential equation; to arrive at
this pricing framework, the pioneering results of Black and Scholes (1973)
and Merton (1974) are applied. From the price of debt- and equity-related
securities it is then possible to derive the coupon level which maximises
the firm value, ultimately resulting in an indication of the optimal capital
structure.
Structural models studied in academic literature could be further split
into two categories—static and dynamic models. Leland (1994) serves as a
classic example of the former one, while various authors have studied the cap-
ital structure in a dynamic setting, e.g. Fischer et al. (1989a) or Goldstein et
al. (2001). The main difference between the two groups is, as the name sug-
gests, that static models are derived assuming that the capital structure is a
one-off decision, and can thus only be applied to a narrow range of companies
with unreasonably stable fundamentals. Therefore, one of the key limitations
of this class of models is that they ignore the firm’s optimal restructuring
policy in response to the state variable fluctuations in time. Conversely,
dynamic models view the capital structure decision as an unceasing process.
Since dynamic capital structure modelling is at the core of this paper,
let us briefly outline the development of academic research studying this
class of models. Kane, Marcus, and McDonald (1984) build one of the first
frameworks which serves as a basis for further analysis. They develop a model
of optimal debt policy which incorporates personal taxes and bankruptcy
costs. The underlying asset in their model is allowed to follow a mixed jump-
diffusion process rather than an ordinary diffusion process, and this extension
precludes the use of Black and Scholes (1973) replication derivation, thus
leading them to employ an equilibrium approach. Besides, they extend the
notion of the rate of return shortfall introduced by McDonald and Siegel
(1984) to model the net tax advantage of debt in terms of rate of return (i.e.
Chapter 1. Introduction 13
measured as a flow variable rather than the standard stock variable). The
key insight from their model is that when tax advantage of debt is strictly
positive, it is never optimal for a firm to be all-equity financed, however large
its bankruptcy costs.
Fischer et al. (1989a) extend the setting of Kane et al. (1984) to allow
a firm to recapitalise at any point in time, but incorporate refinancing costs
to make continuous adjustment costly. They obtain closed-form solutions for
debt and equity as a function of the firm’s dynamic recapitalisation decisions;
further, the optimal recapitalisation policy is derived, which depends on firm-
specific characteristics. They argue that while the leverage declines, the firm
foregoes an increasing amount of debt-related tax shields, and at some point
it is optimal to recapitalise; in the opposite case—while the leverage ratio
increases—it might be optimal to recapitalise to avoid bankruptcy costs.
An important insight from their work is thus that it is useful to study the
leverage ratio range, which helps to explain why similar firms could have
different capital structures—due to the fact that any ratio within the specified
recapitalisation boundaries is in fact optimal.
The next landmark in development of dynamic capital structure mod-
elling is the paper by Goldstein et al. (2001), which introduces a novel model
accommodating the option to increase leverage. As the authors point out,
the fact that Kane et al. (1984) and Fischer et al. (1989a) obtain relatively
simple solutions for capital structure is largely due to the fact that their as-
sumptions effectively reduce the analysis to the one-period framework, and
lead them to argue that the tax advantage to debt is marginal. Goldstein
et al., on the contrary, find that tax benefit of debt increases significantly
as the option to lever up in the future is taken into account. The crucial
novelty of this paper is also in a new interpretation of the state variable as
earnings before interest and taxes, in contrast to the previous studies which
considered the value of unlevered assets. We revisit this assumption in more
detail in Chapter 4.
Since Goldstein et al. (2001) numerous authors have been trying to ex-
14 S. Bjerrisgaard and D. Fedoryaev
tend and refine earlier dynamic capital structure models, and we will only
briefly touch upon some of the issues brought up by them later in our paper.
Among others, Flor and Lester (2002) examine optimal debt maturity and
call premium; Christensen et al. (2002) focus on different aspects of debt
renegotiation; Dangl and Zechner (2004) study the effect of capital structure
dynamics on credit risk; Ammann and Genser (2004) propose a model which
includes finite-maturity debt and multiple bond issues; Francois and Morellec
(2004) investigate how different assumptions regarding bankruptcy affect the
valuation of debt and equity; Titman and Tsyplakov (2007) provide a model
where the firm value is determined by endogenous investment and financing
choices.
Chapter 2
Contingent Claims Analysis
Before proceeding to the discussion of how the optimal capital structure
of the firm could be determined, we will touch upon some fundamentals of
stochastic calculus and introduce the techniques of contingent claims analysis
which represent an important stepping stone to further analysis.
First, we will describe the Wiener process, state the basic properties of
the two stochastic processes—a geometric Brownian motion and the mean-
reverting process with proportional volatility—which will be utilised in the
models we consider, and sketch the basic result of Ito’s lemma as it is critical
for most of the derivations. After that we will elaborate on the question
of attainability of risk neutrality in continuous-time models. Finally, the
method to determine the value of a claim written on a firm fundamental which
evolves according to a diffusion process is presented, and it is also shown
how the obtained valuation framework could be applied to price corporate
securities.
2.1 Stochastic Processes
Let us first show how the Wiener process can be defined as the limit of a
discrete-time process. Consider the value of a stochastic process observed at
time t, z(t). If ∆t is a discrete time change, then the change in z(t) over that
15
16 S. Bjerrisgaard and D. Fedoryaev
time interval is given by
z(t+ ∆t)− z(t) , ∆z =√
∆tε, (2.1)
where ε is a random variable with zero mean and unit variance, and it holds
that Cov[z(t+ ∆t)− z(t), z(s+ ∆t)− z(s)] = 0 for non-overlapping intervals
(t, t+∆t) and (s, s+∆t). It follows that E[∆z] = 0 and Var[∆z] = ∆t, and
that z(t) has serially uncorrelated increments. z(t) represents an example of
a random walk process.
If a change in z(t) over a fixed time interval from 0 to T , consisting of n
intervals of length ∆t, is considered, then
z(T )− z(0) =n∑i=1
∆zi. (2.2)
Applying (2.1), discretised over n time intervals, to (2.2) yields
z(T )− z(0) =n∑i=1
√∆tεi =
√∆t
n∑i=1
εi. (2.3)
Consequently, the first two moments of z(T )− z(0) are
E[z(T )− z(0)] =√
∆tn∑i=1
E[εi] = 0, (2.4)
Var[z(T )− z(0)] = (√
∆t)2
n∑i=1
Var[εi] = ∆t · n · 1 = T, (2.5)
where E[·] and Var[·] denote the mean and variance operators, respectively,
conditional on the information at date t. It should be noticed that, holding
the length of the time interval fixed, the moments of z(T )−z(0) are indepen-
dent of the number of intervals n. This means that when the continuous-time
limit is obtained by letting n go to infinity—or equally, by letting ∆t go to
zero—the central limit theorem, under the assumption that the εi are inde-
Chapter 2. Contingent Claims Analysis 17
pendent and identically distributed, can be applied to state that
plimn→∞
[z(T )− z(0)] ∼ N(0, T ), (2.6)
i.e. z(T ) − z(0) is normally distributed with mean zero and variance T . It
follows that the distribution of a stochastic process z(t) over any finite time
interval [0, T ] could be interpreted as the distribution of a sum of infinitely
many independent increments ∆zi =√
∆tεi, which are drawn from an arbi-
trary distribution. However, since the CLT stipulates that the sum of these
increments converges in distribution to the Gaussian one, we can, without
loss of generality, assume that ∀i εi ∼ N(0, 1).
The limit of an infinitesimal increment could be seen as
dz(t) = limt→0
∆z = limt→0
√∆tε. (2.7)
Due to the previously stated characteristics of ε it holds that E[dz(t)] = 0
and Var[dz(t)] = dt. Given the above, dz is termed a Wiener process and is
often referred to as a (pure) Brownian motion.
The change in z(t) over [0, T ] can now be expressed as
z(T )− z(0) =
∫ T
0
dz(t) ∼ N(0, T ), (2.8)
with the right-hand side being represented by the Ito integral.8 It should be
emphasised that z(t) has an unbounded variation over any finite time interval
and is thus nowhere differentiable.
8 The Ito (stochastic) integral has a more general form I(t) =∫ t
0ϕ(s)dW (s), where
ϕ(s) is an adapted process and W (s) is the Wiener process. This integral is in fact simplya generalisation of a well-known Riemann–Stieltjes integral in the sense that it considersstochastic processes instead of real-valued functions and the integration is carried out withrespect to a (non-differentiable) stochastic process.
18 S. Bjerrisgaard and D. Fedoryaev
0.2 0.4 0.6 0.8 1.0t
-2
-1
1
2
zHtL
Figure 2.1. Sample paths for z(t) as a discrete-time random walk process andas a Wiener process.
From the evolution of the discrete-time process in Figure 2.1 it is evident
that as n → ∞, so that ∆t → 0, the random walk process converges to
the continuous-time Wiener process. The latter, as we will show now, serves
as a fundamental building block for more general continuous-time stochastic
processes.
Consider a new process z′ defined by dz′(t) = σ(t)dz(t), where σ(t) is
a constant. This process has the following distribution over a discrete time
interval [0, T ]:∫ T
0
dz′(t) = σ
∫ T
0
dz(t) ∼ N(0, σ2T ).
If we now additionally introduce a deterministic constant change of µ per unit
of time to the z′(t) process, so that dz′(t) = µdt + σdz(t), the distribution
would change to∫ T
0
dz′(t) = µT + σ
∫ T
0
dz(t) ∼ N(µT, σ2T ). (2.9)
Further, let us consider the time-inhomogeneous analogue of this process, i.e.
Chapter 2. Contingent Claims Analysis 19
with µ and σ both being time-varying, which yields
z′(T )− z′(0) =
∫ T
0
dz′(t)
=
∫ T
0
µ(t)dt+
∫ T
0
σ(t)dz(t). (2.10)
Thus, we have generalised the standard Wiener process z(t) to a new process
z′(t) whose evolution is given by (2.10). This process is often referred to as
the Ito process.
It is important to emphasise here that the described Ito process is a repre-
sentation of a very broad class of stochastic processes with drift and volatility
being dependent not only on time and the current value of the process, but
also possibly on the past process realisations. Therefore, Ito processes gener-
ally do not possess the Markov property which, loosely speaking, states that
given the present, the future is independent of the past. More specifically,
the process described by the stochastic differential equation
dξ(t) = µ(ξ(t), t)dt+ σ(ξ(t), t)dW (2.11)
is a continuous-time Markov process9 if the instantaneous change in the pro-
cess at date t has a distribution that depends only on t and the contem-
poraneous value ξ(t), and not prior values ξ(s), s < t. Moreover, if path
continuity is further assumed, ξ = ξtt≥0 is called a diffusion process.
Geometric Brownian Motion
Perhaps the most widely applied diffusion process is a geometric Brownian
motion, which could be obtain by letting µ(ξ(t), t) = µξ and σ(ξ(t), t) = σξ
in (2.11), so that
dξ = µξdt+ σξdW, (2.12)
9 Markov chain, used primarily in discrete-time analysis, is also an oft-encounteredterm.
20 S. Bjerrisgaard and D. Fedoryaev
where µ ∈ R and σ ∈ R+ are constants. Both drift rate and volatility rate
are proportional to the current value of ξ, which means that (2.12) can be
easily adjusted to express the dynamics in relative terms. The process is
suitable for state variables which grow exponentially at the average rate of
µ, and it exhibits the following properties:
If ξ hits zero (a zero-probability event), it remains at zero
Given a positive starting value, ξ remains positive
At any time s > t, the value ξ(s) is lognormally distributed with
Et[ξ(s)] = ξ(t)eµ(s−t),
Vart[ξ(s)] = ξ2(t)e2µ(s−t)(
eσ2(s−t) − 1
) Given positive drift, both conditional mean and variance are increasing
in time
Due to the above properties, a geometric Brownian motion is commonly
utilised to model stock prices, which can never become negative because of
the limited liability of investors; other examples include asset prices, wage
rates, etc. In Section 2.4 a geometric Brownian motion is used to describe
the dynamics of the state variable, and the importance of this process as a
fundamental building block for finding the optimal capital structure will be
further highlighted in Chapters 3 and 4.
Mean-reverting Process
Another common diffusion process is the mean-reverting process. It should
be noted that—just as with the Brownian motion—there exist different vari-
ations of this process, depending on whether drift and/or volatility are pro-
portional to the current value of the process, the most popular example being
the Ornstein–Uhlenbeck process. However, we will focus on another modifi-
cation which will be underlying our analysis:
dξ = κ(θ − ξ)dt+ σξdW, (2.13)
Chapter 2. Contingent Claims Analysis 21
where θ ∈ R+ is the long-term mean value, σ ∈ R+ is the constant volatility
rate, and κ ∈ R+ is the speed at which the process converges towards the
long-term equilibrium. It can be seen that if the current level of ξ is below
θ, then the drift term is positive, and the process value tends towards the
steady-state value. The opposite relationship takes place when ξt > θ. The
size of increment changes depends on κ; even though the speed of mean
reversion is positive, the process, due to its stochastic nature, can still move
further away from its long-run value, and for higher volatility the potential
deviation is larger. The process has the following characteristics:
If the initial value of the process is positive, then the future value
remains positive
Conditional mean and variance are as follows:
Et[ξ(s)] = θ + (ξ(t)− θ)e−κ(s−t),
Vart[ξ(s)] =θ2σ2
2κ− σ2− e−2κ(s−t)(ξ(t)− θ)2 +
2θσ2(ξ(t)− θ)κ− σ2
e−κ(s−t)
+2κ2(ξ(t)− θ)2 − κσ2ξ(t)(3ξ(t)− 2θ) + σ4ξ2(t)
(κ− σ2)(2κ− σ2)
· e−(2κ−σ2)(s−t)
In the limit the expected value of the process is equal to the long-term
mean θ, and thus the conditional volatility becomes bounded
Given the above, mean-reverting processes are often used to model com-
modity prices, interest rates, currency exchange rates, etc. Later we will
underline the economic intuition behind the utilisation of the described mean-
reverting process with proportional volatility, as opposed to a geometric
Brownian motion. In Chapter 5 the process is applied to describe the evolu-
tion of the state variable which stipulates the optimal capital structure.
22 S. Bjerrisgaard and D. Fedoryaev
2.2 Ito’s Lemma
Since continuous-time Ito processes are not differentiable, we need to apply
Ito’s lemma—known as the fundamental theorem of stochastic calculus—to
differentiate or integrate functions of such processes. Here we will sketch
the intuition behind this result using a Taylor series expansion, omitting the
formal proof.
Let ξ(t) follow the process given by (2.11) and consider a function F (ξ, t)
which is at least twice differentiable in ξ and once in t. We are interested
in finding the total differential of this function, dF . Usual rules of calculus
define this differential in terms of changes in t and ξ:
dF =∂F
∂tdt+
∂F
∂ξdξ +
1
2
∂2F
∂ξ2(dξ)2 +
1
6
∂3F
∂ξ3(dξ)3 + . . . (2.14)
Using (2.11) and the basic rules of Ito calculus (dt)2 = 0, (dW )2 = dt, and
(dtdW ) = 0, (dξ)2 is determined as
(dξ)2 = µ2(ξ, t)(dt)2 + σ2(ξ, t)(dW )2 + 2µ(ξ, t)dtσ(ξ, t)dW
= σ2(ξ, t)dt. (2.15)
Expanding (dξ)3 in the same fashion will result in the expression every term
of which contains dt raised to a power greater than one, so it will go to zero
faster than dt in the limit does. Since the same is valid for higher-order
terms, the differential dF is given by
dF =∂F
∂tdt+
∂F
∂ξdξ +
1
2
∂2F
∂ξ2(dξ)2,
which could finally be expressed as10
dF =
[∂F
∂t+ µ(ξ, t)
∂F
∂ξ+
1
2σ2(ξ, t)
∂2F
∂ξ2
]dt+ σ(ξ, t)
∂F
∂ξdW. (2.16)
10 Equation (2.16) corresponds to the case when F is a function of one Ito process. Multi-
dimensional Ito’s lemma states that dF = ∂F∂t dt+
∑mi=1
∂F∂ξi
dξi+12
∑mi=1
∑mj=1
∂2F∂ξi∂ξj
dξiξj .
Chapter 2. Contingent Claims Analysis 23
The expression in (2.16) is sometimes also referred to as the law of motion
since it exhibits how F (ξ, t) evolves over time, given that ξ(t) follows an Ito
process. It should further be noted that while dF and dξ have different mean
and variance, they both depend on the same source of uncertainty.
2.3 Note on Risk Neutrality in Continuous
Time
Interestingly, despite the remarkable advancement of the theory of stochastic
calculus, the fact remains that the question of risk neutrality still causes
confusion and some of its nuances are frequently overlooked in the existing
literature. In this subsection we will elaborate on how the transition to
the risk-neutral measure in continuous-time modelling could be made, and
pinpoint the specific shortcoming inherent in many pieces of analysis focusing
on optimal capital structure. We will restrict our attention to the question of
the existence of the risk-neutral probability measure, thus leaving aside the
aspect of its uniqueness. Note that here we will only sketch the necessary
key results and their implications, omitting formal proofs and derivations.11
Let X = Xtt∈[0,∞) be a stochastic discounted price process defined on a
filtered probability space (Ω,F ,P). For now assume that filtration F satisfies
what is commonly referred to as the usual conditions in the literature, later
we will elaborate on the specifics of those. The first fundamental theorem of
asset pricing dictates that the absence of arbitrage opportunities is equivalent
to the existence of an alternative probability measure Q such that any X is
a Q-martingale, i.e. X is Ft-adapted and it holds that
∀s Xs = EQ[Xt | Fs], 0 ≤ s ≤ t <∞. (2.17)
Such Q is termed an equivalent martingale measure. Note that the critical
11 See, e.g. Karatzas and Shreve (1991, Chapter 3) or Revuz and Yor (1999,Chapter VIII) for a rigorous analysis.
24 S. Bjerrisgaard and D. Fedoryaev
condition here is the equivalence of the two probability measures, which is
defined as follows. The two measures P and Q are equivalent if and only if
∀A ∈ F P(A) = 0⇔ Q(A) = 0, (2.18)
which is also frequently interpreted as the two measures being mutually abso-
lutely continuous on F . The intuition behind the definition of equivalence is
straightforward: equivalent probability measures always agree on all certain
(i.e. probability-one) and all impossible events.
Now let us turn to the Girsanov theorem which is the most important
result related to the transformation of the probability measure. It should be
mentioned that there are numerous interpretations of the original theorem,
and here we will only outline the general idea behind the basic result of
Girsanov (1960).
Let Y = Ytt∈[0,∞) be an adapted process and
Zt , exp
(∫ t
0
YsdWs −1
2
∫ t
0
Y 2s ds
)under the P-measure. Suppose Z is a martingale.12 Then there exists an
equivalent probability measure Q13 such that the process W = Wtt∈[0,∞)
defined as
WQt , WP
t −∫ t
0
Ysds (2.19)
is a Wiener process under Q. Thus, loosely speaking, a shifted Wiener process
is again a Wiener process provided that the original probability P is replaced
with an equivalent probability Q.
Note that the shift of the Wiener process could also be expressed as
12 It could be shown by applying Ito’s lemma that Z is a local martingale with Z0 = 1.The oft-employed result to verify that Z is a martingale, is the Novikov condition which
states that if E[exp
(12
∫ T0Y 2s ds
)]<∞, then Z(Y ) is a martingale.
13 After Z is shown to be a martingale, the new probability measure could be strictlydefined in terms of the Radon–Nikodym derivative ZT = dQ
dP .
Chapter 2. Contingent Claims Analysis 25
dWt = Ytdt + dWt, which enables the following simple transformation. If
under the probability measure P the process X has the dynamics dXt =
µtdt+σtdWt, then its Q-dynamics would look like dXt = (µt+σtYt)dt+σtdWt,
which implies that in going from the original process to the new one the
diffusion is unchanged.
It should be mentioned that the usual conditions often imposed on filtra-
tion F in the literature could void the Girsanov theorem and consequently
the transition to the risk-neutral probability measure.14 The discussion of
such conditions and the proof of inapplicability thereof are out of the scope of
this thesis.15 However, there is one much less subtle criterion which must be
fulfilled for the Girsanov transformation to hold—the finiteness of the time
horizon. For any finite time T ∈ [0,∞) there exists a unique probability
measure Q on F∞ that is equivalent to P on FT , and the key nuance related
to the Girsanov theorem is that without very limiting restrictions on the
transform process Y the theorem cannot ensure that there exists an appro-
priate probability measure equivalent to P on F∞. Thus, the infinite-time
setup frequently employed in the literature (e.g. Leland, 1994; Goldstein et
al., 2001) may in fact not contain an equivalent martingale measure.16 To
our knowledge, the only paper that acknowledges the problem is Christensen
et al. (2002).
2.4 The General Partial Differential Equation
The optimality of capital structure entails the maximisation of the total
firm value, and is thus dependent on the market values of debt and equity,
which both could be viewed as claims on the underlying state variable. The
14 For example, it could be shown that under the requirement that F0 contains allnegligible sets of F∞, the Girsanov theorem fails. Bichteler (2002, Section 1.3) introducesthe natural conditions which rectify the shortcomings of the conventional usual conditions.
15 Firoozi (2006) provides a note on the martingale property of processes. For a com-prehensive analysis, see Bichteler (2002, Section 3.9).
16 For the formal proof, see Bichteler (2002, Section 3.9). The discrete-time version ofthis result was demonstrated in Back and Pliska (1991).
26 S. Bjerrisgaard and D. Fedoryaev
specific values of these claims can be obtained as the solution to a certain
partial differential equation, which will be derived in this section, with the
applied methods being based on Merton (1974).
The set of assumptions required for the derivation of the general partial
differential equation is as follows.
Financial markets are frictionless (absent transaction costs, taxes or
asset indivisibility issues), arbitrage-free, and complete17
There exists a money market account with a constant risk-free rate r
at which investors can lend and borrow
Assets can be traded continuously in time
The dynamics of the firm value V is described by a diffusion process—a
geometric Brownian motion, i.e.
dV = (µV − ν)dt+ σV dW, (2.20)
where µ is the (instantaneous) expected rate of return on the firm, ν
is the total payout to claim holders if positive, and the net amount re-
ceived by the firm from new financing if negative. The (instantaneous)
variance of the return on the firm is represented by σ2, and W is the
Wiener process.
Suppose there exists a traded security whose market value Y is a function
of V , Y = F (V, t), and whose instantaneous changes in value are described
by
dY = (µY Y − νY )dt+ σY Y dWY , (2.21)
17 Here we imply the conventional definition of absence of arbitrage and completenessthrough existence and uniqueness of an equivalent martingale measure Q, commonly ac-cepted in the asset-pricing field (see, e.g. the seminal work of Harrison and Pliska, 1981),although we do recognise that alternative views exist. For instance, Battig and Jarrow(1999) introduce a definition of market completeness which is independent of any particu-lar probability measure and demonstrate that under this definition arbitrage opportunitiesmay in fact exist in a complete market.
Chapter 2. Contingent Claims Analysis 27
where µY is the expected rate of return, νY is the monetary payout to the
holder of the security, σY is the return volatility, and WY is the Wiener
process. Although in practice companies are often not allowed to liquidate
assets to meet interest payments due to a corresponding negative covenant
frequently observed in debt contracts, we will assume that assets are sold to
finance payouts, following Merton (1974).
Ito’s lemma can now be applied to obtain the differential of the security
value as
dY =
(∂F
∂t+ (µV − ν)
∂F
∂V+
1
2σ2V 2∂
2F
∂V 2
)dt+ σV
∂F
∂VdW. (2.22)
Since both (2.21) and (2.22) describe the evolution in value of the same
security, drift, volatility, and source of uncertainty should be identical, i.e.18
µY Y = Ft + (µV − ν)FV +1
2σ2V 2FV V + νY , µY F (2.23)
σY Y = σV FV , σY F (2.24)
dWY = dW (2.25)
Continuous trading, existence of a risk-free asset, and market complete-
ness stipulate that a replicating portfolio containing the firm and the security
can be formed. The portfolio is financed with a short position in the riskless
asset, implying zero net investment. After rearranging (2.20) to get
dV + νdt
V= µdt+ σdW,
we could express the total return on the portfolio as
dK = M1
(dV + νdt
V
)+M2
(dY + νY dt
Y
)+M3rdt, (2.26)
where M1 is the amount invested in the firm, M2 is the amount invested in
18 For readability purposes we will from now on use the shorthand notation Ft, FV , and
FV V , denoting derivatives ∂F∂t , ∂F
∂V and ∂2F∂V 2 , respectively.
28 S. Bjerrisgaard and D. Fedoryaev
the security and M3 , −(M1 +M2) is the amount borrowed at the risk-free
rate. Substituting from above, we get
dK = M1(µdt+ σdW ) +M2(µY dt+ σY dWY )−M1rdt−M2rdt
= M1(µ− r)dt+M2(µY − r)dt+M1σdW +M2σY dWY ,
which by using (2.25) could be written as
dK = [M1(µ− r) +M2(µY − r)]dt+ [M1σ +M2σY ]dW. (2.27)
The no-arbitrage condition dictates that when there is no uncertainty related
to the portfolio’s expected return, the risk premium must be equal to zero,
and therefore the following system of conditions takes place:
M1σ +M2σY = 0 (2.28)
M1(µ− r) +M2(µY − r) = 0 (2.29)
where (2.28) implies elimination of the source of uncertainty, and (2.29) stip-
ulates zero expected return. Of course, M∗1,2 = 0 is an obvious solution, but
since such a portfolio would be empty, the only non-trivial solution could be
found from
M1σ +M2σY = 0⇔M1 =−M2σY
σ,
−M2σY (µ− r)σ
+M2(µY − r) = 0⇔ µY − rσY
=µ− rσ
. (2.30)
which means that the excess return per unit of risk must be the same for
both assets in the portfolio.
Chapter 2. Contingent Claims Analysis 29
The no-arbitrage condition in (2.30) could be restated using expressions
for µY and σY from (2.23) and (2.24):
µ− rσ
=
Ft+(µV−ν)FV + 12σ2V 2FV V +νY
F− r
σV FVF
⇔
1
2σ2V 2FV V + (rV − ν)FV + Ft − rF + νY = 0. (2.31)
which is a partial differential equation (PDE) validating the absence of arbi-
trage with respect to the replicating portfolio. The sheer beauty of this result
is in the fact that any security whose value is dependent on the firm value
must satisfy this PDE, and we will heavily rely on this condition in further
analysis. Of course, to differentiate between the securities, one should addi-
tionally specify the payout along with the boundary conditions. This will be
demonstrated in the next section when prices of securities are derived.19
From (2.31) it is easy to see that the pricing function F which solves
the general PDE is dependent not only on the current value of the firm and
time, but also on the risk-free rate, the firm’s payout policy, the payout to
holders of the security, and the volatility of the firm value. Interestingly,
F does not depend on the expected return on the firm given by µ, or risk
preferences of investors. This means that two investors who agree on the firm
valuation and its volatility, irrespective of differences in their utility functions
or expectations regarding the firm’s future performance, will accept the same
pricing function to value contingent claims.
2.5 Isomorphism to Option Pricing
Let us now briefly touch upon how corporate securities can be interpreted as
options and therefore valued using the option pricing techniques pioneered
19 An alternative method to obtain the value of a contingent claim is to discount thesecurity’s cash flows at the risk-free rate, and take expectations. This approach assumesthat a risk-neutral measure Q exists. Despite the difference in methodology to the no-arbitrage approach, the Feynman–Kac theorem states that the two methods yield the sameresult.
30 S. Bjerrisgaard and D. Fedoryaev
by Black and Scholes (1973).
In addition to what is stated in the previous section the following as-
sumptions are applied:
There exist two types of claims written on the firm, viz. interest-free
debt and equity, and no other claims can be issued
The bond issue has principal P and must be paid back at time T
The firm cannot pay cash dividends or do share repurchases prior to
debt maturity
If the debt value is denoted as D, then the general PDE from (2.31) can
be written as
1
2σ2V 2DV V + rV DV −Dτ − rD = 0, (2.32)
where νY = 0 as there are no coupon payments, ν = 0 because the firm is
not allowed to attract additional financing or make payouts, τ , T − t is the
time to maturity and thus Ft = −Fτ .To find the function D which solves the PDE, two boundary conditions
and an initial condition need to be laid out. Since the unlevered firm value is
defined as V , D(V, τ) +E(V, τ), where E is the value of equity, then these
conditions are given by
D(0, τ) = E(0, τ) = 0, (2.33)
D(V, τ) ≤ V, (2.34)
D(V, 0) = minV, P, (2.35)
where (2.33) is explained by the fact that debt and equity values always re-
main non-negative, (2.34) is dictated by the limited liability of equity holders,
(2.35) is the initial condition for debt at τ = 0, i.e. at maturity, and is stip-
ulated by the absolute priority rule.
Chapter 2. Contingent Claims Analysis 31
Even though (2.32) could now be solved directly along with (2.33)–(2.35)
by using e.g. standard Fourier transforms, it is possible to ease the procedure
by reducing the problem to the one already solved in literature. Note that
the PDE for equity value is given by
1
2σ2V 2EV V + rV EV − Eτ − rE = 0 (2.36)
s.t.
E(V, 0) = max0, V − P, (2.37)
in addition to conditions corresponding to (2.33) and (2.34).
It is critical to realise that equations (2.36) and (2.37) are identical to
those for a European call option written on a zero-dividend stock, where
V corresponds to the stock price and P—to the strike price. This insight
allows us to write out the solution to (2.36)–(2.37) immediately, using the
Black–Scholes framework:
E(V, τ) = Call(V, P, σ, r, τ) = V Φ(d1)− P e−rτΦ(d2), (2.38)
where Φ(·) is the cumulative distribution function for the Gaussian distribu-
tion, and
d1 =ln(VP
)+(r + 1
2σ2)τ
σ√τ
,
d2 = d1 − σ√τ .
Therefore, from (2.38) and D = V − E, we get
D(V, τ) = P e−rτ −[P e−rτΦ(−d2)− V Φ(−d1)
], (2.39)
and one could notice that in fact the debt value could be represented as the
value of the portfolio comprising the money market account and a short put
32 S. Bjerrisgaard and D. Fedoryaev
position since Put(V, P, σ, r, τ) = P e−rτΦ(−d2)− V Φ(−d1).
The intuition behind the similarities in payoffs to debt and equity owners
and holders of the corresponding contingent claims is as follows. At maturity
equity holders have the option to either to pay the debt principal back or
default. If the value of the firm is sufficiently higher than the principal, then
shareholders will prefer to retire the debt and become the sole claimants of the
then debtless company. This makes the value of their claim at maturity equal
to the difference between the firm value and the principal. If the company
is worth less than the principal, owners of equity will declare bankruptcy
and let debt holders take over the company. The payoff to equity holders is
therefore equivalent to that of an owner of a European call option written
on the firm, with the strike price equal to the principal. Thus, the value of
equity can be obtained by pricing the call option, utilising the Black–Scholes
framework. The value of debt can be derived in a similar fashion, using that
the payoff to the bondholders is equivalent to that of a short position in a
put option written on the firm, where equity holders own the put and can
exercise it at the strike equal to the principal. Therefore, the total value of
debt is equal to the discounted value of the principal less the value of the
put option.
An important implication of the Merton model is that relationships in the
option pricing framework can be directly applied. A typical example is the
impact of an increase in the business risk—measured by the variance of firm
value—which benefits equity holders at the expense of bond holders because
the former have no downside risk. Equity owners will therefore have an in-
centive to transfer value to themselves by taking on more risky projects after
the debt contract has been settled. As a result, if the financing decision is
not separated from the investment decision, the problem of asset substitution
will occur, as extensively discussed in Jensen and Meckling (1976). However,
using numerical analysis, Leland (1998) finds that agency costs related to
the asset substitution problem are small, which is also confirmed by Graham
and Harvey (2001) in an empirical study.
Chapter 2. Contingent Claims Analysis 33
Despite its fundamental significance, Merton’s valuation framework can
only be seen as a basic structural model, and that is amplified by the fact
that several assumptions the model rests upon are rarely observed in practice.
Firstly, the model ignores taxes and thus cannot be used as an appropriate
indication of the optimal capital structure. The lack of applicability is further
emphasised by the possibility of the firm value to either rise to an arbitrar-
ily high level or drop to almost zero without any sort of reorganisation of
the corporate financing. Dynamic models allowing the firm to increase its
leverage depending on the value of the state variable will be presented in
Chapters 4 and 5. Secondly, the model only considers zero-coupon bonds. In
addition to the fact that interest charges on debt are a commonly observed
phenomenon, this assumption implies that the firm cannot default prior to
maturity. The models presented in the chapters that follow, however, do
incorporate a continuous possibility of bankruptcy as well as explicit coupon
payments. Naturally, we will also have to adhere to certain assumptions that
are difficult to justify empirically, e.g. non-stochastic interest rates and con-
stant volatility, but since the primary object of our analysis is the capital
structure, we do not expect the results to change qualitatively because of
that.
Chapter 3
Static Model of Capital
Structure
One of the drawbacks of the Merton model is that it does not consider taxes.
In reality bond investors pay personal taxes on the received interest income
at a certain tax rate τi. Similarly, equity owners incur tax on dividends at
a rate τd. Besides, as owners of equity are subject to double taxation in
the form of both corporate and personal taxes, they face the effective tax
rate τe, which is calculated using (1 − τe) = (1 − τc)(1 − τd). Furthermore,
as companies are generally allowed to deduct interest expenses from the tax
base, they have an incentive to utilise this option by taking on more debt to
finance operations and investments. If the interest expenses are assumed to
be always tax-deductible and it holds that τe > τi, then there is always a tax
advantage to debt. Note, however, that personal taxes will not be explicitly
included in the model described below, following Leland (1994).
In this chapter a model balancing the tax advantage to debt with the
costs of financial distress is presented. The former is seen as a stream of
uncertain tax savings, thereby considering the possibility that the company
might default and thus cannot perpetually utilise the tax shield. The costs
of financial distress are represented by the fraction of the asset base that
is lost due to bankruptcy costs in case the firm defaults. The model yields
34
Chapter 3. Static Model of Capital Structure 35
closed-form solutions, and thus presents explicit guidance on the choice of
optimal capital structure. The underlying assumptions and derivations are
based on the model of Leland (1994).
The chapter is structured as follows. Firstly, we dwell on the assumption
of time independence which is critical for the model derivations. Secondly,
valuation of different contingent claims is presented, following from the spec-
ification of corresponding boundary conditions. Thirdly, the framework is
extended to consider bankruptcy as an endogenous event, as opposed to
modelling default as triggered by the asset value reaching an exogenously
given level. Besides, it is shown how the optimal leverage can be derived
by letting the coupon level be determined from the firm value maximisation
problem. The chapter concludes by presenting the model performance—using
comparative statics analysis—and outlining important drawbacks.
3.1 Time Independence of Cash Flows
Brennan and Schwartz (1978) were the first to take into account the tax
advantage to debt when explicitly modelling optimal capital structure, but
conducted their analysis only numerically. The model developed in Leland
(1994) serves as a general version of that in Brennan and Schwartz as it
provides closed-form solutions for optimal leverage. The former differs by
examining permanent capital structure changes under the infinite maturity
setup, however, as opposed to Brennan and Schwartz who consider finiteness
of debt maturity.
The infinite maturity is in fact one of the key assumptions in Leland
(1994), implying that securities have time-independent cash flows and thus
their valuation is independent of time as well. Generally, however, the value
of a debt security is time-dependent. Corporate bonds, for instance, have fi-
nite maturities, which means that the final cash flow is not the same as those
on interest payment dates and that the impact of the principal repayment
on the bond valuation varies with maturity. The assumption of time inde-
36 S. Bjerrisgaard and D. Fedoryaev
pendence can nevertheless still be justified. Firstly, for sufficiently long bond
maturities the principal redemption is negligible compared to the total payout
value, and can thus be ignored. Therefore, cash flows can be approximated
by interest payments only, essentially making such bond a time-independent
security. Very long bond maturities are not a purely theoretical concept
though: for example, in 2010 DONG Energy, the leading energy company
of Denmark, issued 30-year bonds, in 2008 an American conglomerate Gen-
eral Electric Company placed a 40-year bond issue, and in 2007 a British
retailer Tesco announced an issue of 50-year bonds, among others. Examples
of corporate bond issues with even more extreme maturities do exist as well,
represented by Ford Motor Company’s issue of a 100-year bond in 1997 and
Coca-Cola’s 100-year bond issue in 1993.
Another argument justifying the assumption in question is that revolv-
ing credit agreements—when debt is allowed to be rolled over at maturity
maintaining the interest rate—are often similar to a time-independent setup.
The assumption of time independence is exactly what allows Leland to derive
closed-form solutions, as opposed to preceding authors.20
3.2 The Model
As in Merton (1974), the model considered in Leland (1994) is based on
the dynamics of a firm fundamental. In both models the fundamental is
represented by the unlevered firm value21, and Leland considers the following
20 Note that debt with explicitly finite maturity could also be incorporated in the staticframework. Leland and Toft (1996) develop a model of optimal leverage for a firm thatcontinuously issues finite-maturity bonds, generalising the findings of Leland (1994). Theyshow that not only debt value depends on maturity, but also does bankruptcy whendetermined endogenously. Further, Flor and Lester (2002) extend the results of Lelandand Toft to a dynamic setting. In our analysis we will, however, following the discussion inSection 2.3, assume that all the issued debt has the maturity which is finite, but sufficientlylong to obtain time-independent cash flows.
21 Naturally, the value of the unlevered firm corresponds to the value of assets. Therewill be no distinction between the two terms in the remainder of the thesis.
Chapter 3. Static Model of Capital Structure 37
process to describe its development:22
dV
V= µ(V, t)dt+ σdW, (3.1)
where V is asset value, µ < r is the drift23, and σ is the constant volatility.
V is naturally independent of the financial structure. As the model does
not include any explicit earnings-generating process, and no payout to claim
holders is financed through the sale of assets—as opposed to Merton—Leland
assumes that all claims are serviced by an additional costless issue of equity.
Consider now a claim valued as Y = F (V, t) written on the unlevered
firm, which continuously pays out νY conditioned on the firm being solvent.
Derivations in Chapter 2 suggest that the value F (V, t) of such a security
must satisfy the partial differential equation
1
2σ2V 2FV V + rV FV + Ft − rF + νY = 0. (3.2)
However, due to the assumed time independence and the payout to holders of
the claim, this equation can be reduced to the following ordinary differential
equation (ODE):
1
2σ2V 2FV V + rV FV − rF + νY = 0. (3.3)
3.2.1 Solution to the ODE
We will first solve the homogeneous differential equation
1
2σ2V 2FV V + rV FV − rF = 0. (3.4)
The procedure becomes apparent immediately after one recognises that (3.4)
is a standard Cauchy–Euler equation, i.e. a linear homogeneous ODE with
22 Henceforth, we will be working directly under the risk-neutral probability measureQ, unless otherwise stated.
23 This condition is dictated by the requirement that the state variable has a finite value,and follows from the derivation of the Gordon growth model (Gordon, 1959).
38 S. Bjerrisgaard and D. Fedoryaev
variable coefficients. We will now transform it to the analogous ODE with
constant coefficients to apply the conventional technique of solving a differ-
ential equation using the corresponding characteristic equation.
To deal with the problem of variable coefficients it is common to apply
the following substitution: V = ex. To obtain the first-order derivative in
(3.4), FV , consider
∂F
∂x=∂F
∂V
∂V
∂x=∂F
∂VV, (3.5)
using ∂V∂x
= V . This means that
V FV =∂F
∂x.
The second-order derivative is found in a similar fashion:
∂2F
∂x2=
∂
∂x
(∂F
∂x
)=
∂
∂x
(∂F
∂VV
)=∂F
∂V
∂V
∂x+ V
∂
∂x
(∂F
∂V
)= V
∂F
∂V+ V 2∂
2F
∂V 2, (3.6)
using the chain rule to get ∂∂x
(∂F∂V
)= ∂V
∂x∂2F∂V 2 . Therefore,
V 2FV V =∂2F
∂x2− ∂F
∂x.
Plugging both terms back into (3.4) yields
1
2σ2∂
2F
∂x2+
(r − 1
2σ2
)∂F
∂x− rF = 0, (3.7)
which is a linear homogeneous ODE with constant coefficients, and therefore,
can be solved explicitly using the corresponding characteristic equation
1
2σ2β2 +
(r − 1
2σ2
)β − r = 0. (3.8)
Chapter 3. Static Model of Capital Structure 39
The roots to the characteristic equation are
β1,2 =
(12σ2 − r
)±√
(r + 12σ2)2
σ2,
with
β1 = 1,
β2 = −2r
σ2.
Since both roots are not complex, and due to the fact that r > 0 they
are distinct, the general (and in fact, complete) solution to (3.4) can be
formulated as
F (V ) = B1eβ1x +B2eβ2x,
or, more specifically:
F (V ) = B1V +B2V− 2rσ2 . (3.9)
Adding one particular solution results in the general solution to the inhomo-
geneous ODE in (3.3):
F (V ) = B0 +B1V +B2V− 2rσ2 . (3.10)
Note that the constants B0, B1, and B2 should be determined by the bound-
ary conditions for a specific claim, with B0 taking into account the case when
the security offers a payout νY .
3.2.2 Valuation of Claims
Now we turn to pricing the securities contingent on the unlevered firm value,
using the solution to the ODE derived above.
40 S. Bjerrisgaard and D. Fedoryaev
The two primary claims considered are debt and equity securities, whose
values are given by D(V ) and E(V ), respectively. Two additional claims
could be introduced: tax benefit of debt, whose value is denoted TB(V ),
and bankruptcy costs, whose value is represented by BC(V ). Therefore, the
total firm value can be calculated as A(V ) = V +TB(V )−BC(V ), and since
the equity claim is residual, it could valued as E(V ) = A(V ) − D(V ). As
will be shown below, by applying (3.10) along with the necessary boundary
conditions, closed-form solutions to the values of all these claims could be
derived.
In Leland (1994), owners of corporate bonds are recipients of a perpetual
coupon C, conditioned on the firm being solvent. If the value of assets hits an
exogenously specified threshold Vd, the company defaults and bondholders
pay a fraction 0 < γ < 1 of the unlevered firm value in bankruptcy costs
before taking over the firm, and become the sole owners of the company
valued at (1− γ)Vd.24 Therefore, the boundary conditions for the debt claim
are given by
At V = Vd, D(V ) = (1− γ)Vd, (3.11)
As V →∞, D(V )→ C
r, (3.12)
where the first one states the firm value at default, and the second one implies
that as the firm value increases exponentially, debt essentially becomes risk-
free.
Using condition (3.12) together with (3.10) means that B1 = 0 and
B0 = Cr
. Inserting these expressions in condition (3.11) applied to (3.10),
yields (1 − γ)Vd = Cr
+ B2V− 2rσ2
d , and therefore, B2 =[(1− γ)Vd − C
r
]V
2rσ2
d .
24 The firm value upon bankruptcy is calculated as if the assumption of absolute priorityapplies, which in this model leaves nothing for the equity holders after default. However,in a study of 30 US firms defaulting during the 1980s, Eberhart, Moore, and Roenfeldt(1990) document that equity owners on average received 7.6% of the total value distributedto all claimants. A possible explanation of the seeming contradiction between their findingand the models we present is that we model default differently, assuming the takeover ofthe company, as opposed to restricting the bondholders’ claim only to the principal.
Chapter 3. Static Model of Capital Structure 41
All constants are now known, and the value of the debt security can be
obtained as
D(V ) =C
r+
[(1− γ)Vd −
C
r
](V
Vd
)− 2rσ2
. (3.13)
Consider a new security pd(V ) that has a payoff of 1 DKK if the firm
defaults, and zero in all other states of nature. This claim can thus be
interpreted as an Arrow–Debreu security. Its present value is of the form
pd(V ) = B1V +B2V− 2rσ2 , (3.14)
because B0 = 0 since the claim offers no intermediate cash flows. When
taking into account the boundary conditions
limV→Vd
pd(V ) = 1,
limV→∞
pd(V ) = 0,
it can be seen that
pd(V ) = V2rσ2
d V −2rσ2 =
(V
Vd
)− 2rσ2
. (3.15)
By using this result, debt value can alternatively be written as
D(V ) =C
r+ pd(V )
((1− γ)Vd −
C
r
). (3.16)
The value of debt can therefore be seen as a perpetual stream of coupon pay-
ments, given that the firm is solvent, plus the net payoff in case bankruptcy
takes place, which is equal to the firm value at default less the lost interest
income.
To obtain the levered firm value A(V ) we need to quantify the two key
effects of debt—through bankruptcy costs and tax deductibility of interest
expenses. Since bankruptcy costs are contingent on the unlevered firm value
42 S. Bjerrisgaard and D. Fedoryaev
and time independence of related cash flows also holds, we could apply (3.10)
in the same manner as above to price this claim. This security could be
viewed as the one offering a lump sum payment γVd upon default and zero
otherwise, thus implying the following boundary conditions:
At V = Vd, BC(V ) = γVd,
As V →∞, BC(V )→ 0,
and hence,
BC(V ) = γVdpd(V ). (3.17)
As before, by using (3.15), the value of this claim can be interpreted as the
present value of receiving γVd given that the state of nature when the com-
pany files for bankruptcy is realised. Naturally, the claim value is increasing
in Vd as both the probability of default and the amount received by the se-
curity holder will be larger the higher is Vd, with V counteracting the former
relationship.
The stream of tax savings due to debt financing resembles a claim with a
perpetual payoff τcC contingent on the company remaining solvent, and the
fact that it can fully benefit from the tax deductibility in every period until
default. That brings us to the corresponding boundary conditions:
At V = Vd, TB(V ) = 0,
As V →∞, TB(V )→ τcC
r,
and thus the claim value is
TB(V ) =τcC
r− τcC
rpd(V ). (3.18)
The value of this claim can be interpreted as the value of receiving the tax
benefit continually minus the present value of the cash flows that are lost
Chapter 3. Static Model of Capital Structure 43
after the unlevered firm value falls to the default level.
Now we are able to obtain the total firm value A(V ) as the unlevered firm
value plus tax benefits of debt minus bankruptcy costs, i.e.
A(V ) = V +τcC
r−(τcC
r+ γVd
)pd(V ). (3.19)
It could be noticed that as soon as V = Vd and the bondholders take over
the company, the firm value will be equal to the value of assets less the
bankruptcy costs, as the tax benefits will be lost due to the inability of
the new owners to relever the company after default.25 A less obvious, and
possibly even counterintuitive, implication of (3.19) is that the firm value
is increasing in the corporate tax rate and coupon payment. This can be
explained by the nature of the underlying variable that the model is based
on: since the asset value, unlike e.g. EBIT, is not taxed, the value of the
firm is only affected by taxes and coupons through the value of the tax
benefits, and therefore, the two parameters have no negative impact on the
firm valuation. This problem will be addressed in the models presented in
Chapters 4 and 5.
The equity claim is valued as a residual:
E(V ) = V +τcC
r−(τcC
r+ γVd
)pd(V )
−[C
r+
[(1− γ)Vd −
C
r
]pd(V )
]= V − (1− τc)
C
r+
[(1− τc)
C
r− Vd
]pd(V ). (3.20)
Thus, from the derived claim values we see that the bankruptcy level and the
coupon payment play an important role in determining the capital structure
choice. However, up until now we considered both as being exogenously
given.
25 The value of the firm at bankruptcy can be modelled alternatively, see e.g. Christensenet al. (2002), who allow the new owners to optimally relever the company post-default.
44 S. Bjerrisgaard and D. Fedoryaev
3.2.3 Optimal Leverage
Although the default threshold for the asset value might be given exoge-
nously, as if imposed by a covenant in the debt contract, it can be reason-
ably argued that shareholders will try to choose the boundary optimally to
maximise the value of their claim, stipulating the following smooth-pasting
condition:
∂E(V )
∂V
∣∣∣∣V=Vd
= 0. (3.21)
After simple algebra, the optimal bankruptcy level V ∗d can be found to
be
V ∗d =(1− τc)Cr + 1
2σ2
. (3.22)
Substituting (3.22) into the values of debt and equity claims as well as in the
total firm value found above, yields
D(V ) =C
r+
[(1− γ)(1− τc)C
r + 12σ2
− C
r
][(r + 1
2σ2)V
(1− τc)C
]− 2rσ2
, (3.23)
E(V ) = V − (1− τc)Cr
+
[(1− τc)C
r− (1− τc)C
r + 12σ2
] [(r + 1
2σ2)V
(1− τc)C
]− 2rσ2
,
(3.24)
A(V ) = V +τcC
r−[τcC
r+ γ
(1− τc)Cr + 1
2σ2
][(r + 1
2σ2)V
(1− τc)C
]− 2rσ2
. (3.25)
Note that (3.23)–(3.25) only suggest a certain capital structure given an
arbitrary level of coupon, which in fact does not mean that financing is
optimal. To determine optimal leverage it is necessary to assume that the
firm management can choose the coupon level which maximises the firm
Chapter 3. Static Model of Capital Structure 45
value. Using (3.25) to solve for the optimal coupon yields
C∗(V ) =(r + 1
2σ2)V s
1− τc, (3.26)
where
s =
(τcσ
2
d
)σ2
2r
,
d = 2τcr + τcσ2 + 2rγ − τc2rγ.
Plugging C∗(V ) into (3.23) and (3.25) gives
D∗(V ) =
(r + 1
2σ2)V s
(1− τc)r+
[(1− γ)V s−
(r + 1
2σ2)V s
(1− τc)r
]s
2rσ2 , (3.27)
A∗(V ) = V +τc(r + 1
2σ2)V s
(1− τc)r−
[τc(r + 1
2σ2)V s
(1− τc)r+ γV s
]s
2rσ2 . (3.28)
Note that since C∗(V ) is only a function of V and constant parameters, it
is possible to determine the optimal capital structure simply by knowing the
current asset value. Furthermore, not only the model provides the manage-
ment with the information on the current capital structure given the dynam-
ically changing firm fundamental, but also on how far the current leverage
is from the optimal one, D∗(V )A∗(V )
, thus providing some degree of flexibility and
possibility of application in different scenarios.
3.3 Comparative Statics
The static model described above will serve as a benchmark for the models
in subsequent chapters, and in this section we will look into its performance
using comparative statics analysis. The applied parameter values are based
46 S. Bjerrisgaard and D. Fedoryaev
Table 3.1. Comparative statics for the static model. Base case parameters are:r = 3.5%, γ = 5%, and σ = 25%. The tax advantage of debt is calculated as T = A∗
V − 1.All monetary values are per unit of asset value.
L∗ C∗ χ∗d T
Base case 75.6% 0.042 0.48 0.16
τc = 20% 73.7% 0.039 0.47 0.12
τc = 30% 77.2% 0.046 0.48 0.21
σ = 20% 77.6% 0.039 0.53 0.18
σ = 30% 74.4% 0.047 0.44 0.15
γ = 0% 78.8% 0.045 0.51 0.17
γ = 10% 72.7% 0.040 0.45 0.15
on the Danish data and empirical estimates.
Table 3.1 presents comparative statics for optimal leverage, coupon, and
default threshold, as well as the tax advantage to debt. As it can be noticed,
a higher corporate tax rate increases optimal leverage. This is due to firms
benefitting more from the tax shield, which is illustrated by a higher optimal
coupon payment and a larger tax advantage to debt. From the comparative
statics given in Table 3.1 it is further evident that equity holders of risky firms
will optimally default at lower asset values. Since the value of equity—akin
to the value of a call option—is increasing in volatility, equity holders of risky
companies will continue operations for longer during downturns, compared
to shareholders of safer companies, as the former face a larger potential gain.
Due to limited liability, the expected value of equity in a risky firm is thus
higher than in a safer but otherwise identical company.
The relative amount of debt is higher for less risky firms because the
state variable is less likely to hit the lower boundary, allowing for a more
aggressive leveraging. Interestingly, it can be noted that the optimal coupon
appears to be increasing in the business risk. However, as also mentioned by
Leland (1994), this is due to the specific parameter value: as confirmed by
Figure 3.1(a) below, both high- and low-risk companies optimally choose to
commit to a higher coupon level, while firms with an intermediate volatility
levels pay smaller coupons. However, even though the optimal coupon is
Chapter 3. Static Model of Capital Structure 47
decreasing for some intervals of volatility, borrowing will always be more
expensive for companies with a greater level of uncertainty attached to the
future asset value, as the value of debt is decreasing in volatility. These two
relationships are illustrated in Figures 3.1(e,f).
For an increase in bankruptcy costs the asset value at which the firm
declares bankruptcy drops, cf. Figure 3.1(b). On the other hand, if the
risk-free interest rate rises, the value of equity declines due to the diminished
discounted value of the tax shield, and therefore an increase in r will have
an opposing effect on the default boundary because equity value equals zero
at a higher level of the state variable. These relations are evident from
Figures 3.1(c,d).
3.4 Discussion
The presented model could be criticised on a number of points. Firstly, and
most importantly, if V represents the price of a traded asset after the initial
debt issue, then the model is not arbitrage-free. The risk-free profit can be
earned by investors who buy the unlevered assets of the firm for V , lever up
the company and sell it for a higher value A(V ) > V . To preclude arbitrage,
it hence must hold that A(V ) = V −TC, where TC is the amount required to
be paid in transactions costs to optimally lever the company. If V does not
represent the price of a traded asset, the general PDE from Chapter 2 cannot
be applied. This problem could potentially be circumvented by introducing a
traded derivative written on the assets of the firm, or, alternatively, by using
risk-neutral valuation to arrive at the same ODE.26 One way to rectify this
significant drawback is to model a firm fundamental which is not traded, e.g.
earnings, and we will implement this approach in the next chapter.
A second negative point is the assumption that the firm is liquidated
upon default. In a study of US firms, Alderson and Betker (1995) find that
26 This alternative approach is used in e.g. Cox, Ingersoll, and Ross (1985) to find theprice of contingent claims, although in a different setup than that in Merton (1974).
48 S. Bjerrisgaard and D. Fedoryaev
0.1 0.2 0.3 0.4Σ
0.035
0.040
0.045
0.050
0.055
0.060
C*
(a) Coupon and asset volatility
0.05 0.10 0.15 0.20Γ
0.42
0.44
0.46
0.48
0.50
d
(b) Default boundary and bankruptcy costs
0.02 0.04 0.06 0.08 0.10r
0.24
0.26
0.28
0.30
0.32
E
(c) Equity value and risk-free interest rate
0.02 0.04 0.06 0.08 0.10r
0.35
0.40
0.45
0.50
0.55
0.60
d
(d) Default boundary and risk-freeinterest rate
0.0 0.1 0.2 0.3 0.4Σ
100
200
300
400S
(e) Credit spread and asset volatility
0.1 0.2 0.3 0.4Σ
0.9
1.0
1.1
1.2
1.3
D
(f) Debt value and asset volatility
Figure 3.1. Performance of the static model. Base case parameters are: r = 3.5%,τc = 25%, and σ = 25%. Credit spread is defined as S = C
D − r.
Chapter 3. Static Model of Capital Structure 49
80% of the firms would have created more value to their owners if opera-
tions were continued, compared to if assets were liquidated. Furthermore,
knowledge-based firms with more intangible assets such as pharmaceuticals
and IT companies may have low asset value, and a liquidation of this type
of company thus yields a considerable loss of shareholder wealth, and will
therefore unlikely be seen in practice. Furthermore, if asset specificity is
very high, liquidation costs will surge due to the illiquidity in the secondary
market, which might result in a fire sale (Shleifer and Vishny, 1992). In this
case the decision to continue operations will generally prevail as well.
Lastly, the model could be criticised for not allowing the firm to change
its capital structure after the initial bond issue. This static nature of the
debt level explains why the model suggests a leverage ratio of 76% in the
base case, while empirical research documents much lower estimates.27 This
may be due to the inability to exploit the tax advantage to debt in case
the asset value increases. It would therefore be desirable to allow the debt
level to change dynamically—so that the capital structure is tailored to the
realised value of the firm fundamental continually—and we incorporate that
option in the model presented in the chapter that follows.
27 Gaud, Hoesli, and Bender (2007) in a European study including 172 Danish firmsobserved in the period from 1988–2000 provide evidence that mean leverage in Danishcorporations is around 33%, which falls in the same estimate range as that in most ofother considered countries. For the presented model to yield similar results, bankruptcycosts and asset volatility have to be simultaneously increased to unreasonably high 70%and 45%, respectively, with other base case parameters unchanged.
Chapter 4
Family of Existing GBM-based
Dynamic Models
Changes in capital structure are commonly observed. Leary and Roberts
(2005) conclude that firms on average issue or repurchase their own securities
once a year. The discontinuity in restructuring activity could be explained,
for instance, by the presence of adjustment costs. Consider a firm which
exhibits high operating performance and has positive growth forecasts. To
better exploit the tax advantage to debt, this company has to continually
increase leverage, i.e. to dynamically adjust capital structure according to the
realised earnings. This option is ignored in the class of static models, a classic
example of which was introduced in Chapter 3. In this chapter, however, a
model which does allow for changes in capital structure will be presented.
Optimal leverage in a dynamic setting has previously been investigated by
Fischer et al. (1989a), Goldstein et al. (2001), Christensen et al. (2002),
and Flor and Lester (2002), among others. A distinguishing feature common
to most of existing research papers is the underlying assumption that the
state variable follows a geometric Brownian motion, which allows a simplified
progression from the static setup to the dynamic one due to the homogeneity
property inherent in this particular stochastic process. Therefore, we will try
to lay out a dynamic capital structure model which will describe the whole
50
Chapter 4. Family of Existing GBM-based Dynamic Models 51
family of dynamic GBM-based models in general, rather than represent any
particular setup.
The structure of the chapter is as follows. Firstly, a general introduction
to the model is provided. Afterwards, the set of equations together with
the maximisation problem needed to obtain the optimal capital structure is
presented. The chapter concludes with the comparative statics analysis and
the discussion of the model.
4.1 The Model
As mentioned in Chapter 3, the question of whether to treat the unlevered
firm value as corresponding to the price of a traded asset or not, is a critical
point. The problem is avoided by considering earnings before interest and
taxes as the state variable. Moreover, this makes the model more realistic in
the sense that coupons paid to bondholders are now financed with earnings
rather than by issuing new equity. Thus, the state variable χ denotes the
value of a claim to the entire payout produced by the firm, and its develop-
ment is given by a geometric Brownian motion:
dχ = µχdt+ σχdW, (4.1)
where µ is an exogenously given constant risk-neutral drift rate, σ is a con-
stant volatility of earnings, and r is a constant interest rate on money market
account.
The management will choose to default when the state variable hits a
lower boundary χd , dχ0. However, in addition to models described in previ-
ous chapters, there will now also be introduced an upper boundary χu , uχ0,
at which the management will decide to call all outstanding debt at par and
relever the firm optimally afterwards.28 It is precisely the latter option that
28 Restructuring could in principle be modelled in different ways. For example, a possibleoption would be to model debt as non-callable, and let the firm keep the existing debtand carry out an extra issue in addition. However, in this case additional assumptions
52 S. Bjerrisgaard and D. Fedoryaev
stipulates the dynamics of capital structure. Note that the boundaries for the
state variable could be determined either exogenously, by the management’s
precommitment in the debt contract, or endogenously, by incentive compati-
bility constraints imposed on the equity holders. The retirement of debt will
briefly make current equity holders the sole claimants of the company. Sub-
sequently, the firm will issue new corporate bonds with principal and coupon
exceeding the previous level. It can be shown that due to the properties of
a geometric Brownian motion optimal values for decision variables in case χ
hits the upper boundary χu could be easily obtained by scaling with a factor
u = χuχ0
. This implies that, for instance, optimal default level after the first
refinancing will be equal to uχd, after the second restructuring—u2χd, and
so on, where χd is the lower boundary in the initial period. The same scaling
factor can be applied to find the optimal coupon and restructuring threshold.
Even though both upward restructurings and downward debt repurchases are
observed in practice, we abstain from including the option to reduce the debt
level when earnings reach the lower boundary to simplify the analysis.
Figure 4.1 below captures the management’s option to increase the debt
level in case EBIT increases sufficiently. It can be noticed that when χ is fluc-
tuating within the range (χd, χu), the management will choose to maintain
the capital structure. A possible explanation of this phenomenon is given in
Fischer et al. (1989a), who document that there is a leverage range in which
capital structure will not be changed due to adjustment costs.
Due to the lognormal dynamics of χ, it is not necessary to scale the
parameters µ and σ when reoptimising the capital structure, and the EBIT
process in the second period will have the same properties as in the first
period. This feature implies that the mechanism of the capital structure
decision is in fact unchanged. It can be shown by backward induction that
scaling by a factor u is indeed optimal.29
To obtain the pricing function F of securities contingent on the state
regarding the seniority and credit risk of several debt issues need to be made, which maycomplicate the analysis without changing results qualitatively.
29 See Goldstein et al. (2001, Appendix B).
Chapter 4. Family of Existing GBM-based Dynamic Models 53
t
d× Χ0
u×d× Χ0
u2×d× Χ0
Χ0
u× Χ0
u2× Χ0
u3× Χ0
Χ
Figure 4.1. A simulated sample path of the firm’s earnings with lognormaldynamics. Initial earnings level is χ0, and period zero ends with earnings reaching χu,at which point the company calls the entire existing debt and reoptimises the capitalstructure; in the subsequent periods the procedure is repeated.
variable χ, we apply the framework developed in Chapter 2, which boils
down to the following ODE:
1
2σ2χ2Fχχ + µχFχ − rF + νY = 0, (4.2)
where νY denotes the payout specific to a security Y , and Fχ and Fχχ denote
the first- and second-order partial derivatives of F with respect to χ. The
general solution to the corresponding homogeneous differential equation can
be formulated as
F (χ) = B1χβ1 +B2χ
β2 , (4.3)
where β1 and β2 are the roots to the characteristic equation corresponding
to (4.2).
The system of equations describing the model should embrace the bound-
ary conditions for debt and equity and two smooth-pasting conditions. Two
additional equations for the maximum potential firm value and the value of
the principal will be constructed to supplement the final system.
54 S. Bjerrisgaard and D. Fedoryaev
The two value functions that need to be determined are
D(χ) = d1χβ1 + d2χ
β2 +(1− τi)C
r(4.4)
and
E(χ) = e1χβ1 + e2χ
β2 +(1− τe)χr − µ
− (1− τe)Cr
, (4.5)
where the inclusion of τd ensures that the value of equity is decreasing in the
tax rate, which corrects the inconsistency observed under the static model,
when the relationship was positive.
Let us consider the boundary conditions. When EBIT reaches the lower
boundary χd, the firm is taken over by bondholders, after they pay the
bankruptcy costs, γ. Debt value at bankruptcy can thus be expressed as30
D(χd) = (1− τe)(1− γ)χdr − µ
. (4.6)
As shown in (4.6), after the bankruptcy costs have been paid the old bond-
holders become the new equity holders of an unlevered firm. However, we
assume that a company cannot issue new bonds if it has previously defaulted
on debt payments, and therefore the new shareholders simply own a riskless
perpetuity of earnings, and firm value can thus be described by the Gordon’s
growth model (Gordon, 1959). Naturally, the stream of earnings would not
be riskless if, for example, fixed operating costs were considered in the model.
However, as no EBIT-financed payouts except coupon payments are consid-
ered, the firm cannot default in the future as it has no outstanding debt, and
hence becomes risk-free. It can also be seen from (4.6) that the expected
growth in EBIT is unaltered after the firm is taken over by bondholders, as
opposed to e.g. Goldstein et al. (2001) who assume that a firm is liquidated
at default. The assumption of maintaining the firm operations is supported
30 Note that values of debt and equity claims at the boundaries are dependent on theinitial EBIT level χ0, but we suppress it as an argument for convenience in the currentand subsequent chapters.
Chapter 4. Family of Existing GBM-based Dynamic Models 55
by empirical evidence that documents that only 5% of financially distressed
firms are eventually liquidated (Gilson et al., 1990). If the new firm owners
are allowed to optimally relever the company post-default, optimal leverage
is expected to increase (Christensen et al., 2002), because debt holders will
be able to exploit the tax shield in the future, thereby increasing the pre-
default value of their claim. This option is, however, not considered in our
model.
At the lower boundary, equity holders will have given up their claim to
residual earnings, and hence
E(χd) = 0 (4.7)
to comply with the absolute priority rule.
At the upper boundary, management will call all outstanding debt, and
under the assumption that no call premium needs to be paid on bonds retired
prematurely,31 the value of the claim of debt holders is
D(χu) = D(χ0). (4.8)
After equity holders have paid back the old debt D(χ0), they will issue
new bonds D(χu) with higher principal and coupon. The value of equity at
restructuring point is thus
E(χu) = E(χu) + (1− q)D(χu)−D(χ0), (4.9)
where q denotes restructuring expenses paid as a fraction of the new debt,
and can be interpreted as issuance costs. As the repayment of the old prin-
cipal is financed by the equity holders, the company will be instantaneously
unlevered, and at this point in time shareholders will choose the new optimal
debt level.
In addition to the two boundary conditions stated, smooth-pasting con-
31 We do not consider any liquidity issues stemming from the presence of this provision.The assumption of the absence of a call premium is relaxed in Subsection 5.4.2.
56 S. Bjerrisgaard and D. Fedoryaev
ditions need to be invoked at both lower and upper EBIT thresholds. These
conditions ensure that it is always in the best interest of equity holders to
either declare bankruptcy or refinance. Even though one or both thresholds
could be determined exogenously—for instance, Goldstein et al. (2001) as-
sume that management precommits to the restructuring level χu, while the
default level χd is determined by the incentive compatibility condition—we
choose to let both upper and lower boundaries be determined endogenously
for consistency. The two smooth-pasting conditions are
∂E(χ)
∂χ
∣∣∣∣χ=χd
= 0, (4.10)
∂E(χ)
∂χ
∣∣∣∣χ=χu
= E(χ0) + (1− q)D(χ0). (4.11)
The high-contact condition is needed to ensure that the decision to refinance
is in fact optimal for the equity owners at that specific EBIT level. If this
condition were not included in the model, equity holders would not maximise
the potential firm value by restructuring at the upper boundary, resulting
in a discrepancy between E(χu) and the shareholders’ incentive-compatible
equity level. The low-contact condition states that it must be optimal for
owners of equity to default exactly when the lower boundary χd is reached.
The system now consists of six equations (4.6)–(4.11), with d1, d2, e1, e2,
d, u as unknowns. This system cannot be solved analytically, and thus we
have to resort to numerical analysis. To simplify expressions utilised in the
numerical procedure, we construct the following extra variables:
Π , E(χ0) + (1− q)D(χ0), (4.12)
Σ , D(χ0). (4.13)
We will refer to the first variable as the “maximum potential firm value”.
This value can be interpreted as the price the owner-manager, who solely
owns the EBIT-generating machine or production technology, can sell the
Chapter 4. Family of Existing GBM-based Dynamic Models 57
company for.32 This price is what equity holders want to maximise. The
second definition restates the debt value at issuance as the principal Σ. This
reflects the assumption that bonds are issued at par, with no discount or
premium. Note that both definitions imply pricing per unit of EBIT.
With the two extra equations introduced, the complete system is now
comprised of eight equations, and must be solved for eight unknowns. It could
be reduced to two equations for a pair d, u containing the known parameters
r, σ, µ, γ, q, τe, τi and the coupon C. Due to the non-linearity of these two
equations, we have to utilise numerical methods to obtain values for d and
u, assuming a certain coupon level, and then insert them in the maximum
potential firm value Π, which is further maximised to find the optimal coupon
C∗, i.e.
C∗ = arg maxC
Π. (4.14)
The optimal coupon is then plugged into the algebraic expressions of debt
and equity, which ultimately determine the optimal capital structure. The
results obtained from this optimisation procedure are presented in the next
section. We dwell on the applied numerical methods in the next chapter.
4.2 Comparative Statics
Before proceeding to the numerical results of the model, let us briefly under-
line some common issues associated with simulations. A widespread problem
is an ambiguous choice of the base case parameters. Of course, one might
argue that the exact set of parameters does not matter for the study of
general relationships or relative comparisons, but it is nevertheless obvious
that model predictions in absolute terms do rely (and possibly heavily) on
32 This interpretation is reminiscent of that of Christensen et al. (2002). However,to emphasise that this value could in fact reflect a different leverage level if the owner-manager has initially taken on a suboptimal amount of debt, we prefer to think of Π as thepotential amount the entrepreneur can sell his company for, which implies the optimalityof capital structure. We dwell on the interpretation of this variable in Subsection 5.2.3.
58 S. Bjerrisgaard and D. Fedoryaev
Table 4.1. Comparative statics for the dynamic GBM-based model. Base caseparameters are: r = 3.5%, γ = 5%, µ = 2%, q = 1%, σ = 25%, τi = 37%, τd = 28%,and τc = 25%. The tax benefit of debt is calculated as T = Π
χ0(1−τe)r−µ
− 1. All monetary
values are per unit of EBIT.
L∗ C∗ χ∗d T
Base case 36.1% 1.04 0.19 0.12
τc = 20% 28.9% 0.77 0.15 0.04
τc = 30% 39.7% 1.23 0.21 0.22
σ = 20% 41.7% 1.11 0.25 0.13
σ = 30% 32.0% 1.02 0.15 0.11
γ = 0% 39.3% 1.17 0.21 0.13
γ = 10% 33.3% 0.93 0.18 0.11
the initial choice, and therefore, extensive sensitivity analysis should be con-
ducted. Another shortcoming oft-encountered in the literature is drawing
inferences from some specific values produced by a model (e.g. in Goldstein
et al., 2001) which does not seem appropriate as certain relationships are
in fact not monotonous, and thus graphical analysis is needed to provide a
complete picture in continuity. Apart from the monotonicity issue, there is
also a problem of ignoring strong influence of a third model parameter on the
nature of the relationship considered. Studying cross-relationships of such
kind could provide important insights and that is why we will introduce a
third dimension when necessary.
To carry out a comparison to the static model we preserve the base case
parameters defined in Section 3.3. Note, however, that personal taxes are
now included in the model. The tabulated results are presented below.
The key insight from the comparative statics is that the initial leverage
is now much lower than that observed in the static model. This is a direct
consequence of the fact that the management, while endowed with an option
to relever the firm, will take on less debt at the beginning and instead adjust
the capital structure dynamically in the future to make it correspond to
the earnings level. In the static setting, the initial debt level will have to
reflect the expected future development of the firm, and thus may in fact
Chapter 4. Family of Existing GBM-based Dynamic Models 59
be implicitly suboptimal from a dynamic perspective. Moreover, we see that
the bankruptcy threshold is lowered considerably, reflecting the fact that,
ceteris paribus, the value of a firm possessing a restructuring option will be
higher than the value of the firm without one. Thus, in some cases when the
earnings level is low enough for the equity holders to declare default in the
static framework, it will still not be sufficient to prompt them to do so in the
dynamic setting. While comparing the tax advantage to debt in the dynamic
framework with that in the static model, one should note that the latter does
not consider personal taxes. Besides, since Leland (1994) models asset value
and not income, an increase in the tax rate has only a positive influence
since the deduction of interest expenses becomes more valuable. This is not
the case in the dynamic model, where the introduction of a dividend tax
ensures that the value of equity is decreasing in taxes. Therefore, due to
this difference in the assumed state variable and thus in the tax regime, it is
not completely correct to compare the tax benefit of debt directly. The tax
advantage to debt in the dynamic framework is expected to be higher due
to the ability to exploit the benefits of debt financing better because of the
restructuring option; indeed, we do obtain a tax advantage to debt of 33%
in the dynamic setting when we exclude personal taxes from the model.
The effects of changes in parameters on the selected dependent variables
are similar to those obtained in the static model. However, a few merit
comments. The initial earnings level is lower than the optimal coupon in the
base case, but still higher than the bankruptcy threshold, which implies that
the firm may not default even though earnings are insufficient to cover the
interest payments on outstanding bonds. As asset sale or additional debt
issues are not allowed in our model, the deficit has to be financed by equity
owners as long as EBIT fluctuates above χd.
In the dynamic model the spread in optimal values for varying parameter
values is wider. Hence, not only optimal values per se differ, but so does the
impact of slight changes in input parameters. For instance, if bankruptcy
costs increase by five percentage points from the base case, optimal coupon
60 S. Bjerrisgaard and D. Fedoryaev
is reduced by more than 10% in the dynamic model, while in the static frame-
work it only decreases by 6%. Furthermore, if volatility deviates from the
base case by five percentage points in the dynamic model, optimal leverage is
in the range 32–42%, compared to the interval 74–78% in the static setting.
Thus, as opposed to the much smaller effort required to determine optimal
leverage in a static setting, precision of estimates and calibration of input
parameters are essential if firms wish to fully benefit from the opportunity
to adjust capital structure dynamically.
As the dynamic model allows for debt increase, management, when decid-
ing on capital structure changes, also needs to take readjustment costs q into
consideration, in addition to the parameters in a static setup. If restructur-
ing is more costly, this will result in new bonds being issued less frequently
and will prolong the commitment to the coupons related to the current debt
level. Therefore, the equity holders will prefer lower coupon payments in the
future, and accordingly lower leverage and decreased probability of default.
There is also a counter-effect present, however: as the cost of rebalancing the
capital structure rises, the equity holders may prefer to have a higher initial
leverage and higher coupon payments to better utilise the tax advantage to
debt. These opposing effects help explain why the incremental change in
leverage in Figure 4.2(a) is so narrow—the level drops by a mere 0.5 percent-
age points when restructuring costs increase from 0.5% to 2%. Overall, the
interplay of the described relations is evident from Figures 4.2(a,e).
The advantage of taking on debt to finance operations and investments
is decreasing in the tax rate on interest income. As the inclusion of τi only
affects the debt value, the maximum potential firm value decreases when the
tax on interest income increases. This results in a lower tax benefit of debt,
making equity financing more favourable. This leads to the optimal leverage
being lower for higher personal taxes, as can be seen from Figure 4.2(b).
Figure 4.2(f) demonstrates that as the effective tax rate increases, the
optimal restructuring level falls. This is because the tax benefit of debt
is positively related to τe, and thus for higher taxes there is a wider net
Chapter 4. Family of Existing GBM-based Dynamic Models 61
0.005 0.010 0.015q
0.357
0.358
0.359
0.360
0.361
0.362
L*
(a) Leverage and restructuring costs
0.38 0.39 0.40 0.41Τi
0.30
0.31
0.32
0.33
0.34
0.35
0.36
L*
(b) Leverage and tax on interest income
0.25 0.30 0.35 0.40 0.45 0.50 0.55Σ
400
500
600
700
800
900
1000S
(c) Credit spread and earnings volatility
0.25 0.30 0.35 0.40 0.45 0.50 0.55Σ
1.05
1.10
1.15
C*
(d) Coupon and earnings volatility
0.005 0.010 0.015q
1.1
1.2
1.3
1.4
1.5u
(e) Restructuring boundary andrestructuring costs
0.44 0.46 0.48 0.50 0.52Τe
1.28
1.30
1.32
1.34
1.36
1.38
1.40u
(f) Restructuring boundary andeffective tax rate
Figure 4.2. Performance of the the dynamic GBM-based model. Base caseparameters are: r = 3.5%, τc = 25%, τd = 28%, τi = 37%, γ = 5%, q = 1%, and σ = 25%.
62 S. Bjerrisgaard and D. Fedoryaev
refinancing gain which equity owners can realise by adjusting the capital
structure. Therefore, the restructuring level u is lowered to accommodate
more frequent readjustment.
Results in Table 4.1 imply that riskier firms pay a higher coupon. How-
ever, conclusions based on tabulated comparative statics have the obvious
drawback of implying monotonicity. From Figure 4.2(d) it can be seen that
the coupon is increasing in volatility only for higher-risk firms. As in the
static model, both safer and riskier companies will optimally commit to
high coupons, when compared to an otherwise similar firm with a mid-range
volatility level. Safe firms will prefer higher coupons to better exploit the
tax advantage of debt, whereas risky companies will choose high coupon pay-
ments to attract bond investors. Furthermore, the absence of monotonicity in
the relationship between earnings volatility and optimal coupon means that
the change in coupon cannot provide an unambiguous guidance on whether
the price of borrowing has increased. Indeed, we see from Figure 4.2(c) that
firms with higher business risk always have to face a wider credit spread—
because the debt value is strictly decreasing in volatility—even though an
increase in volatility may in fact imply a lower coupon payment.
If there is a greater uncertainty attached to a firm’s future earnings level,
such company will, in addition to the higher credit spread, also have a sub-
stantially lower optimal leverage than an otherwise identical firm, but the
relationship is less pronounced if the effective tax rate faced by equity hold-
ers is increased, as illustrated by flattening curvature of the optimal leverage
plane for higher τe in Figure 4.3(a). This could be explained by the fact that
higher tax benefit of debt counteracts the negative impact of volatility, which
makes debt financing more favourable.
Further, an increase in the effective tax rate prompts the management
to commit to higher optimal coupon. If bankruptcy is expensive, however,
interest payments will be lowered due to a decreased earnings level at which
equity holders declare bankruptcy and therefore lowered default probability.
This balancing of the positive and negative effects of debt exactly repre-
Chapter 4. Family of Existing GBM-based Dynamic Models 63
(a) Leverage, earnings volatility andeffective tax rate
(b) Coupon, bankruptcy costs andeffective tax rate
Figure 4.3. Optimal leverage for varying earnings volatility and effective taxrate, and optimal coupon for varying bankruptcy costs and effective tax rate.Base case parameters are: r = 3.5%, τc = 25%, τd = 28%, τi = 37%, γ = 5%, q = 1%,and σ = 25%.
sents the classic trade-off that managers encounter when considering optimal
capital structure, as depicted in Figure 4.3(b).
4.3 Discussion
The dynamic model presented in this chapter allows the firm management
to increase the debt level, enabling a better utilisation of the tax shield for
growing earnings. The company’s option to reduce the outstanding debt is
excluded, which is further emphasised by the inability of firms to accumulate
cash savings. This assumption might be justified by some empirical evidence
which suggests that financially distressed firms will not reduce debt levels due
to transaction costs (Gilson, 1997). Furthermore, strategic debt service mod-
els indicate that equity holders do not have the ability to reduce debt close
to bankruptcy, as earnings are used to bargain with creditors and the level
will be too low for a successful renegotiation of the coupon (Mella-Barral and
Perraudin, 1997). Following the arguments of Jensen and Meckling (1976)
and Jensen (1986), the holding of cash encourages “empire building” and
aggravates agency costs of free cash flow, respectively, and the assumption of
full payout and thus zero retained earnings can be seen as a way to mitigate
64 S. Bjerrisgaard and D. Fedoryaev
these issues. Hence, excluding the downward restructuring option is not a
critical limitation.
One should note that most of the dynamic capital structure models ignore
the dynamics at the lower earnings threshold by modelling default through
liquidation or by restricting the firm to be only all-equity financed post-
bankruptcy, as is the case with the model presented in this chapter. This
might, of course, lead to understated leverage since the debt holders will not
be able to reoptimise the capital structure, which leads to a lower initial
valuation of the debt claim.
As the value of the state variable reaches the upper threshold, the com-
pany increases the debt level by calling all outstanding debt, and issuing new
debt with a larger principal and coupon. In the presented framework, how-
ever, no premium is paid on bonds retired prematurely. This will give equity
holders an incentive to continuously call the debt to exploit the tax benefit
of debt, as long as the tax advantage is sufficient to cover restructuring costs.
Consequently, the model may underestimate the optimal leverage as upward
restructurings will occur too often.
One of the cornerstones of the model is its scalability, which allows opti-
mal future coupon, bankruptcy and restructuring thresholds to be calculated
by simply scaling the current values. This property rests on the assumption
that the state variable follows a geometric Brownian motion. Even though a
closer analysis does reveal some drawbacks of this process, it has neverthe-
less been commonly used for mathematical tractability. To circumvent some
of the known issues, the underlying state variable can instead be modelled
with an alternative process. In the next chapter we propose a novel dynamic
capital structure model which addresses some of the shortcomings inherent
in the family of GBM-based models.
Chapter 5
Proposed Dynamic Model with
Mean Reversion
Having analysed the fundamental models of capital structure in the academic
literature, we now turn to the development of our own model, which extends
existing research in several dimensions described below.
The chapter is organised as follows. Section 5.1 dwells on the motiva-
tion behind the model and emphasises the key points of differentiation from
existing dynamic capital structure models. Section 5.2 focuses on the deriva-
tion of the model and its properties, and lays out the system to be solved to
close the model. Section 5.3 considers tests of the model, elaborates on the
numerical methods employed, and presents the numerical results and their
implications. Section 5.4 extends the analysis by incorporating additional
assumptions into the basic setting and illustrates obtained results in com-
parison with those from the main mean-reverting model and the dynamic
GBM-based model.
5.1 Motivation
It is important to stress from the outset that in this paper we do not aim
to build a completely new framework for analysing optimal leverage; rather,
65
66 S. Bjerrisgaard and D. Fedoryaev
our novelty is in providing a new perspective on the modelling of capital
structure decisions, and thus our model will hopefully serve as an extension
of existing research and give some significant insights on the matter.
The overwhelming majority of existing capital structure studies—as a
matter of fact, almost every one of them—is based on the assumption that
the underlying state variable follows the stochastic process represented by a
geometric Brownian motion. Such a widespread adherence to this assump-
tion is not surprising: the process embraces some very attractive properties33
that ensure mathematical tractability and in some special cases even allow
analytical solutions for the optimal capital structure. Only very few papers
consider the arithmetic Brownian motion as the underlying process, for in-
stance, Ammann and Genser (2004)—Genser (2006) for further extension—
and Bank and Lawrenz (2005), and they only study capital structure in a
static setting. Deviation from the Brownian motion in general is even rarer.
In this thesis we present a dynamic model of capital structure with the op-
erating income as the state variable, which follows a mean-reverting process,
and show the implications for the study of capital structure choice. Admit-
tedly, as all the other authors in the field considering dynamic modelling,
we are also unable to derive analytical solutions, but nevertheless we get re-
markably far. The model is analysed in great detail and extensive numerical
simulations are carried out, with the shortcomings and potential areas of
improvement being highlighted.
One of the key contributions of this paper is the analysis of capital struc-
ture under an alternative process which exhibits mean reversion. The intu-
ition behind our approach is as follows. Firstly, the assumption of either ge-
ometric or arithmetic Brownian motion implies that the state variable could
in time reach infinitely high values, and that due to the commonly assumed
positive drift rate the value constantly increases, with short-run falls being
restricted by the volatility and not having any impact in the medium or long
term. In practice, however, such variable as operating income or, more gen-
33 In particular, the process possesses the homogeneity property, which permits dynamicmodel transformation. We will elaborate on this property later in the chapter.
Chapter 5. Proposed Dynamic Model with Mean Reversion 67
erally, earnings is very often consistent with a certain long-term mean value
which is industry-specific, and ignoring that fact may severely undermine the
application of a model; a potential caveat here is that the long-term mean will
change through time, and we take that into account in our model. Secondly,
by assuming mean reversion we obtain more control over the process as we
are able to directly specify not only the speed of mean reversion which is in
a way similar to the drift rate of a geometric Brownian motion, but also the
long-term mean. That effectively allows the modeller to distinguish between
different industries both in terms of their profitability, through the mean
value of earnings, and their stability, through varying how fast the earnings
will converge. Finally, there exists empirical evidence of mean-reverting pat-
terns in the evolution of earnings, e.g. Bajaj, Denis, and Sarin (2004), Fama
and French (2000), Lipe and Kormendi (1994).
To the best of our knowledge, the only paper that models mean-reverting
earnings while considering capital structure choice in a continuous-time frame-
work is Sarkar and Zapatero (2003).34 However, they only develop a static
model which is akin to that of Leland (1994), and as such it does not allow for
changes in leverage, which is a fairly restrictive setting. In fact, the dynamics
of capital structure has never been considered so far outside of the scope of
the geometric Brownian motion assumption. Here we specifically emphasise
that by the notion “dynamic” we mean the fact that capital structure can
be varied through time at the discretion of management, which acts in the
interests of existing shareholders. Although modelling in a static setting may
still provide interesting insights and serves as a good starting point, it is the
dynamics that really makes the model relevant by bringing more flexibil-
ity to it and making its predictions more consistent with observed financing
practices. Moreover, we further differ from Sarkar and Zapatero (2003) by
employing a more complex setting to our model. For instance, we explicitly
model the complete tax structure by considering corporate and personal tax-
ation imposed on both equity and debt holders, while they only incorporate
34 Raymar (1991) also considers mean reversion, but works in the discrete-time setting,assuming that EBIT follow a first-order autoregressive process.
68 S. Bjerrisgaard and D. Fedoryaev
the net tax advantage of debt financing—an aggregate parameter, supposed
to account for everything at once, which leads to certain ambiguities. For
example, that parameter does not affect the debt claim valuation, which ac-
tually makes the setup identical to simply ignoring tax on interest income.
Thus, the tax advantage to debt is always present in their model, which could
ultimately lead to overstated leverage.
Another important dimension of our analysis is explicit focus on numerical
methods—a part of research that is frequently being ignored in the literature.
In fact, numerical simulations are often presented without even mentioning
which specific algorithm has been applied, and, as will be shown later, since
different methods bear different implications one should take great care to
avoid inherent caveats. The two mathematical problems embedded in the
analysis of dynamic capital structure are the following: solving a system of
nonlinear algebraic equations and running nonlinear local optimisation. The
general algorithmic toolbox for tackling those is not as extensive as one could
initially imagine, and we will try to discuss the main methods used.
The model derived in the next section is extended afterwards to incorpo-
rate a call premium, which relaxes the assumption that equity holders are
able to retire the old debt at par at the restructuring point. Apart from
that, we also extend the dynamic GBM-based model to take into account
fixed operational costs, which allows us to resolve another widespread issue
of capital structure analysis—the cash flow being restricted to take on only
positive values. Since most start-ups do exhibit negative operating income
as do mature firms during economic downturns, this issue is fairly limiting
from a practical angle.
Chapter 5. Proposed Dynamic Model with Mean Reversion 69
5.2 The Model
5.2.1 Preliminaries
We assume that the operating income of a firm follows the mean-reverting
stochastic process under the risk-neutral probability:
dχ = κ(θ − χ)dt+ σχdW, (5.1)
where κ is the speed of mean reversion, θ is the long-term mean value of
earnings35, σ is volatility of earnings, and W is the Wiener process.36 Note
that, as opposed to the conventional arithmetic Ornstein–Uhlenbeck process,
the volatility term is not constant and is proportional to the current earnings
level. In the special case of θ = 0 the process becomes a geometric Brownian
motion with drift µ = −κ. The non-negativity of earnings is implicitly
assumed due to the nature of the process, but we relax this assumption in
Section 5.4.
We will now derive the Gordon growth model for the assumed mean-
reverting stochastic process, which will be required in the valuation of claims
that follows in Subsection 5.2.3.
Lemma 1. Let ξ be a random variable following (5.1), then ∀s > t
Et
[∫ ∞t
ξse−r(s−t) ds
]=θ
r+ξt − θr + κ
. (5.2)
Proof. Since the corresponding integral is finite, by Fubini’s theorem we have
Et
[∫ ∞t
ξse−r(s−t)ds
]=
∫ ∞t
e−r(s−t)Et [ξs] ds. (5.3)
35 Henceforth we will use the notions “operating income”, “EBIT” and “earnings” in-terchangeably, ignoring possible differences stipulated by accounting practices.
36 The specific functional form of mean reversion with proportional volatility was intro-duced by Bhattacharaya (1978).
70 S. Bjerrisgaard and D. Fedoryaev
It could be shown that the expectation of the mean-reverting process is37
Et [ξs] = θ + (ξt − θ)e−κ(s−t), (5.4)
and thus substituting in the right-hand side of (5.3) yields∫ ∞t
e−r(s−t)(θ + (ξt − θ)e−κ(s−t)) ds
=
∫ ∞t
(θer(t−s) + (ξt − θ)e(t−s)(r+κ)
)ds
= −(θ
rer(t−s) +
(ξt − θr + κ
)e(t−s)(r+κ)
)∣∣∣∣∞t
=θ
r+ξt − θr + κ
.
In fact, the result in Lemma 1 is quite intuitive: if earnings exhibit mean
reversion, the expected value of future cash inflows will be equal to the long-
term mean value discounted in perpetuity, which is represented by the first
term, plus what we call the convergence term—the extra cash flow related
to the fact that during a certain period of time, actual earnings will deviate
from their long-term mean.
The general partial differential equation for pricing contingent claims is
obtained by the absence of arbitrage condition as in previous chapters. We
omit most of the derivations.
Let Υ(χ) be the value of a claim written on the total earnings χ of the
firm. Assume Υ(χ, t) is twice differentiable in χ and once in t, and apply
Ito’s lemma to (5.1) to get
dΥ =
(Υt + κ(θ − χ)Υχ +
1
2σ2χ2Υχχ
)dt+ σχΥχdW. (5.5)
If the claim additionally offers a continuous payoff νY , then the expected
return it yields must equal the risk-free rate to preclude arbitrage opportu-
37 See, e.g. Tsekrekos (2010).
Chapter 5. Proposed Dynamic Model with Mean Reversion 71
nities, i.e.
rΥdt = EQ [dΥ + νY ]
= EQ
[(Υt + κ(θ − χ)Υχ +
1
2σ2χ2Υχχ
)dt+ σχΥχdW + νY
].
Due to the common properties of the Wiener process described in Chapter 2
and the fact that Υ(χ, t) and its derivatives are known at time t, the above
could be rewritten as the following partial differential equation:
1
2σ2χ2Υχχ + κ(θ − χ)Υχ − rΥ + Υt + νY = 0. (5.6)
Note that in order for an equivalent martingale measure to exist, we have
to assume that the time horizon is finite,38 but sufficiently long to neglect
the relative value of the principal and thus obtain time independence of the
payouts. So, the value of a claim on the operating income will then satisfy
the ordinary differential equation
σ2
2χ2Υχχ + κ(θ − χ)Υχ − rΥ + νY = 0. (5.7)
To derive the general solution to this ODE let us first solve the corre-
sponding homogeneous differential equation
σ2
2χ2Υχχ + κ(θ − χ)Υχ − rΥ = 0 (5.8)
by demonstrating that it could be represented as a special case of a specific
differential equation, whose solution form is commonly known.
Consider the general confluent equation (Abramowitz and Stegun, 1972,
Eq. 13.1.35), which is the generalised form of the well-known Kummer’s
38 See Section 2.3 for more details on the question of existence of an equivalent martingalemeasure.
72 S. Bjerrisgaard and D. Fedoryaev
equation:
ω′′(Z) +
[2A
Z+ 2f ′(Z) +
(bh′(Z)
h(Z)− h′(Z)− h′′(Z)
h′(Z)
)]ω′(Z)
+
[(bh′(Z)
h(Z)− h′(Z)− h′′(Z)
h′(Z)
)(A
Z+ f ′(Z)
)+A(A− 1)
Z2
+2Af ′(Z)
Z+ f ′′(Z) + (f ′(Z))2 − a(h′(Z))2
h(Z)
]ω(Z) = 0. (5.9)
The general solution to this differential equation is given by
ω(Z) = B1Z−Ae−f(Z)Φ(a, b, h(Z)) +B2Z
−Ae−f(Z)Ψ(a, b, h(Z)), (5.10)
where B1 and B2 are arbitrary constants, Φ(a, b, h(Z)) is the Kummer’s con-
fluent hypergeometric function, and Ψ(a, b, h(Z)) is the Tricomi’s confluent
hypergeometric function. The Kummer’s confluent hypergeometric function
could be expanded in a generalised hypergeometric series
Φ(a, b, h(Z)) =∞∑0
(a)n(h(Z))n
(b)nn!,
where (a)n and (b)n are the rising factorials with (a)0 = 1 and (b)0 = 1.
The Tricomi’s confluent hypergeometric function is most frequently defined
in terms of the Kummer’s function:
Ψ(a, b, h(Z)) =π
sin πb
(Φ(a, b, h(Z))
Γ(1 + a− b)Γ(b)
− (h(Z))1−b Φ(1 + a− b, 2− b, h(Z))
Γ(a)Γ(2− b)
),
where Γ(·) is the gamma function.39
Let A = −β, f(Z) = 0, and h(Z) = 2κθσ2Z
. Then h′(Z) = − 2κθσ2Z2 and
39 To be completely strict, one should additionally specify the domains of Φ(·, ·, ·) andΨ(·, ·, ·), which is a somewhat delicate matter, given that the functions are undefined inseveral special cases, e.g. when a and b are non-positive integers.
Chapter 5. Proposed Dynamic Model with Mean Reversion 73
h′′(Z) = 4κθσ2Z3 . Substituting into (5.9) gives after simple algebra
ω′′(Z) +
[−2β
Z+
(− bZ
+2κθ
σ2Z2+
2
Z
)]ω′(Z)
+
[(− bZ
+2κθ
σ2Z2+
2
Z
)(− βZ
)+β(β + 1)
Z2− 2aκθ
σ2Z3
]ω(Z) = 0.
Multiplying both sides of the above by Z yields
Zω′′(Z) +
[−2β − b+
2κθ
σ2Z+ 2
]ω′(Z)
+
[−2κθ(β + a)
σ2Z2+β2 + β(b− 1)
Z
]ω(Z) = 0. (5.11)
The homogeneous ODE of interest (5.8) could be rewritten as
χΥχχ +
(2κθ
σ2χ− 2κ
σ2
)Υχ −
2r
σ2χΥ = 0, (5.12)
and one could immediately notice that if we let Z = χ and ω(Z) = Υ(χ),
then this differential equation would be identical to (5.11) if and only if the
following system of equations holds:
− 2β − b+2κθ
σ2χ+ 2 =
2κθ
σ2χ− 2κ
σ2
− 2κθ(β + a)
σ2χ2+β2 + β(b− 1)
χ= − 2r
σ2χ
This system is equivalent to the following one:
b = −2β + 2 +2κ
σ2
a = −β
1
2σ2β(β − 1)− κβ − r = 0.
74 S. Bjerrisgaard and D. Fedoryaev
The roots of the quadratic polynomial are
β1,2 =2κ+ σ2 ±
√(2κ+ σ2)2 + 8rσ2
2σ2,
completing the two sets of solutions to the system of equations.
Thus, the general solution to our homogeneous ODE becomes
Υ(χ) = B1χβ1Φ
(a(β1), b(β1),
2κθ
σ2χ
)+B2χ
β2Ψ
(a(β2), b(β2),
2κθ
σ2χ
),
(5.13)
where a(·) and b(·) symbolically denote values of parameters depending on
the root of the polynomial. B1 and B2 will be determined by the corre-
sponding boundary conditions. Note that we will not present the solution
to the inhomogeneous ODE in its general form here, but will formulate the
particular solutions in Subsection 5.2.3, depending on the type of a claim
considered.
5.2.2 Homogeneity Property
Now that the solution to the ODE is derived, we are ready to formulate the
result which is central to the dynamics of capital structure in our model—the
homogeneity property of the assumed mean-reverting process.
Lemma 2. The process given by equation (5.1) is homogeneous of degree one
in the pair (χ, θ).
Proof. We will demonstrate the validity of this statement using contingent
claims analysis. Let χ ∈ [χd, χu] follow (5.1), χd and χu being exogenously
determined levels. Consider a simple claim contingent on χ paying 1 DKK
when χ falls to χd before the upper threshold χu has been reached. Following
the approach from the previous subsection, the price Pd of such claim could
Chapter 5. Proposed Dynamic Model with Mean Reversion 75
be obtained as a solution to the following differential equation:40
σ2
2χ2∂
2Pd∂χ2
+ κ (θ − χ)∂Pd∂χ− rPd = 0 (5.14)
subject to the two boundary conditions
Pd(χd) = 1 and Pd(χu) = 0.
The solution to this system is
Pd(χ) =
(1
χ
)− 12− κσ2−√
4κ2+4κσ2+8rσ2+σ4
2σ2(
1
χd
) 12
+ κσ2
+
√4κ2+4κσ2+8rσ2+σ4
2σ2
·
( 1
χ
)√4κ2+4κσ2+8rσ2+σ4
σ2
Φ(χu) Ψ(χ)−(
1
χu
)√4κ2+4κσ2+8rσ2+σ4
σ2
· Φ(χ) Ψ(χu)
)/
( 1
χd
)√4κ2+4κσ2+8rσ2+σ4
σ2
Φ(χu) Ψ(χd)−
(1
χu
)√4κ2+4κσ2+8rσ2+σ4
σ2
Φ(χd) Ψ(χu)
where Φ(χ) and Ψ(χ) are the Kummer’s and Tricomi’s confluent hypergeo-
metric functions, respectively, and due to space constraints they are symbol-
ically denoted as
Φ(χ) = Φ
(−1
2− κ
σ2−√
4κ2 + 4κσ2 + 8rσ2 + σ4
2σ2,
1−√
4κ2 + 4κσ2 + 8rσ2 + σ4
σ2,
2κθ
σ2χ
)
40 For simplicity we ignore taxes as the inclusion of them does not impact the line ofderivation.
76 S. Bjerrisgaard and D. Fedoryaev
and
Ψ(χ) = Ψ
(−1
2− κ
σ2+
√4κ2 + 4κσ2 + 8rσ2 + σ4
2σ2,
1 +
√4κ2 + 4κσ2 + 8rσ2 + σ4
σ2,
2κθ
σ2χ
)
Utilising the notation from the previous subsection, the solution could be
rewritten in a more convenient form
Pd(χ) =
(1χ
)−β1 (1χd
)β1 [(1χ
)β1−β2Φ(χu)Ψ(χ)−
(1χu
)β1−β2Φ(χ)Ψ(χu)
](
1χd
)β1−β2Φ(χu)Ψ(χd)−
(1χu
)β1−β2Φ(χd)Ψ(χu)
,
(5.15)
where β1 > 0 and β2 < 0.
Let ρ ,(
1χd
)β1−β2Φ(χu)Ψ(χd) −
(1χu
)β1−β2Φ(χd)Ψ(χu), then the above
becomes
Pd(χ) =1
ρ
[(χ
χd
)β1 (χβ2−β1 Φ(χu) Ψ(χ)− χβ2−β1u Φ(χ)Ψ(χu)
)]. (5.16)
By analogy, we can derive the price Pu of a symmetric claim which pays
1 DKK when χ reaches χu before the lower boundary has been reached:
Pu(χ) =1
ρ
[(χ
χu
)β1 (χβ2−β1 Φ(χd) Ψ(χ)− χβ2−β1u Φ(χ) Ψ(χd)
)]. (5.17)
The two claims priced above could be seen as Arrow–Debreu securities in
the dual state of nature system i ∈ d, u, with either low or high earnings
being realised. Thus, the existence of such claims implies the completeness
of our system, and we are able to price any security in this market.
Define a new claim that continuously pays aχt+b ∀t. Using Lemma 1 it is
easy to see that the price of this claim is given by aχt+δθ+η, with α = ar+κ
,
Chapter 5. Proposed Dynamic Model with Mean Reversion 77
δ = aκr(r+κ)
, and η = br. Consider now a general claim which is contingent on
χ and pays out a continuous dividend aχt + b ∀t as long as χ remains within
(χd, χu). When χ either falls to χd or reaches χu, the claim yields the closing
one-off payment Υi. The price of this claim could be derived as41
Υ(χ, θ) = aχ+δθ+η+[Υd−(aχd+δθ+η)]Pd(χ)+[Υu−(aχu+δθ+η)]Pu(χ).
(5.18)
The intuition behind this equation is trivial. As long as the earnings oscillate
within the specified interval, at any time t the value of the payoff offered by
the claim is aχt + δθ + η; as soon as the state variable reaches any of the
boundaries, the claim pays out Υi, but the net inflow is lower due to the lost
future value of the dividend stream and is thus reduced to Υi−(aχi+δθ+η).
The discounted value of the net payoff is obtained through the price of the
Arrow–Debreu security corresponding to that state of nature, Pi(χt).42
Inserting expressions for Arrow–Debreu prices from (5.16) and (5.17) into
the price function gives
Υ(χ, θ) = aχ+ δθ + η + [Υd − (aχd + δθ + η)]
· 1
ρ
[(χ
χd
)β1 (χβ2−β1Φ(χu)Ψ(χ)− χβ2−β1u Φ(χ)Ψ(χu)
)]+ [Υu − (aχu + δθ + η)]
· 1
ρ
[(χ
χu
)β1 (χβ2−β1Φ(χd)Ψ(χ)− χβ2−β1u Φ(χ)Ψ(χd)
)].
(5.19)
To formally demonstrate the required homogeneity property it needs to be
41 Note that even though the pricing function depends only on the current level ofearnings χ, wwe also make the long-term mean value of earnings θ an explicit argumentsince this parameter is fundamental in our model.
42 We will omit the time subscript in further derivations to make notation less confusingand reserve the subscript solely for denoting the state of nature.
78 S. Bjerrisgaard and D. Fedoryaev
proven that Υ(λχ, λθ) = λΥ(χ, θ).43 Since the second and third terms in
(5.19) are symmetric, it suffices to show that this relation holds for any one
of them. We will consider the downside case, ignoring the other term.
Υ(λχ, λθ) , Υ(λχ, λθ;α, δ, λη, λχd, λχu, λΥd, λΥu)
= aλχ+ δλθ + λη + [λΥd − (aλχd + δλθ + λη)]
· 1
ρ′
[(λχ
λχd
)β1 ((λχ)β2−β1Φ(λχu)Ψ(λχ)
−(λχu)β2−β1Φ(λχ)Ψ(λχu)
) ],
where ρ′ is equal to ρ with the respective parameters scaled up, i.e.
ρ′ =
(1
λχd
)β1−β2
Φ(λχu)Ψ(λχd)−(
1
λχu
)β1−β2
Φ(λχd)Ψ(λχu)
= λβ2−β1[χβ2−β1d Φ(λχu)Ψ(λχd)− χβ2−β1
u Φ(λχd)Ψ(λχu)].
Plugging that into Υ(λχ, λθ) gives after some simplification
Υ(λχ, λθ) = λ(aχ+ δθ + η) + λ[Υd − (aχd + δθ + η)]
(χ
χd
)β1· χ
β2−β1Φ(λχu)Ψ(λχ)− χβ2−β1u Φ(λχ)Ψ(λχu)
χβ2−β1d Φ(λχu)Ψ(λχd)− χβ2−β1
u Φ(λχd)Ψ(λχu).
After noticing that Φ(λχ, λθ) = Φ(χ, θ) and Ψ(λχ, λθ) = Ψ(χ, θ), the above
43 It should be noted that the parameters expressed in monetary units, e.g. η thatenters the claim price, are also scaled up accordingly by the same factor of λ.
Chapter 5. Proposed Dynamic Model with Mean Reversion 79
becomes
Υ(λχ, λθ) = λ
(aχ+ δθ + η + [Υd − (aχd + δθ + η)]
·1ρ
[(χ
χd
)β1 (χβ2−β1Φ(χu)Ψ(χ)− χβ2−β1u Φ(χ)Ψ(χu)
)])
= λΥ(χ, θ),
which implies
Υ(λχ, λθ) = λΥ(χ, θ).
The formal statement of Lemma 2 could be interpreted as follows. When-
ever the upper earnings boundary is reached and the capital structure is read-
justed, the management of the firm will face the same decision in the sense
that the mechanics of the optimal capital structure model will be unchanged;
moreover, not only earnings will be scaled as in the case of a geometric Brow-
nian motion, but also the long-term mean earnings will be shifted accordingly.
A straightforward implication behind this result is that the model vari-
ables and parameters are independent of the underlying currency. For in-
stance, if we redefine the units of the model in terms of EUR instead of
DKK, this would not have any bearing on the pricing function as the latter
will just be corrected accordingly. More generally, the homogeneity property,
loosely named a “scaling feature” in Goldstein et al. (2001), allows a simple
transformation in time of the value of any claim through a time-invariant
factor which we will define later. The importance of this result is mainly
expressed in the salient simplicity of the derivations in the dynamic setting,
which is the main reason behind the ubiquitous assumption of a geometric
Brownian motion in the academic literature.
80 S. Bjerrisgaard and D. Fedoryaev
5.2.3 Optimal Capital Structure Decision Tuple
The setup is effectively the same as the one we introduced in the previous
chapter, with all the notation preserved. At any point in time the firm can
have only a single class of callable debt outstanding, with sufficiently long
finite maturity. Bankruptcy and restructuring44 decisions are endogenised
through the lower and upper earnings thresholds, χd and χu, respectively. In
case of bankruptcy, the debt holders take over the firm as a going concern
after incurring the cost of bankruptcy procedure, γ. While restructuring,
equity holders are assumed to call the entire debt at par, and issue new
debt of a larger amount at a cost q; the case of downward restructuring is
omitted. The tax structure is as follows. The corporate earnings are taxed at
the corporate tax rate τc, and debt holders face a tax rate of τi on the received
interest payment. After the coupon is paid out, the rest of the earnings is
distributed to equity holders as a dividend, and thus is taxed at τd. The
effective tax rate borne by the firm owners is denoted τe.
As a note aside, let us dwell on the factors that determine the tax advan-
tage to debt. Choose a moment in time when the firm belongs in its entirety
to the owner-manager, which means that she owns claims on both debt and
equity payouts, and consider the aggregate after-tax inflow that is entitled
to her. The debt claim would bring a coupon payment C, taxed at τi, and
thus the after-tax payoff would be (1− τi)C. The equity claim is subject to
the coupon payout and dual taxation on corporate and personal levels, which
overall yields (1− τd)(1− τc)(χ−C). Thus, the total after-tax payoff of the
owner-manager becomes
(1− τe)χ+ (τe − τi)C. (5.20)
We see that the first term in (5.20) is independent of the capital structure
decision, but the second term is strictly increasing in the factor (τe− τi) and
44 The term “restructuring” sometimes refers to the bankruptcy procedure in the liter-ature; throughout the paper we will, however, utilise it implying debt refinancing, or thereadjustment of capital structure.
Chapter 5. Proposed Dynamic Model with Mean Reversion 81
represents the tax advantage to debt (the increasing relation to the coupon
is obvious as the manager owns the debt claim). Therefore, the tax benefit
of debt exists if and only if the debt is coupon-bearing and the effective tax
rate on equity holders’ payout is higher than the tax on interest income, with
the benefit increasing in the tax difference. We will assume throughout our
theoretical analysis that τe > τi, and will demonstrate the impact of the tax
rates on optimal capital structure while presenting our numerical results in
Section 5.3.
It is important to note that the expression for the tax advantage to debt
from above has an important implication: it could potentially result into
an overstated leverage. The caveat stems from the fact that the assumed
tax schedule is perfectly symmetric, meaning that coupon payments are de-
ductible even when the firm’s current earnings are insufficient to service the
debt, i.e. χt < C. Such taxation scheme favours debt and thus may lead
to the situation when the firm is so levered that the coupon exceeds even
the initial level of earnings. The key problem is that the tax advantage to
debt is independent of the earnings level, and the conventional method to
resolve this issue is to introduce another parameter ε ∈ [0, 1], which repre-
sents the proportional reduction of the effective tax rate when the earnings
net of interest (EBT) become negative. The tax rate in that case will be ετe,
thereby diminishing the benefit of leverage. However, although Christensen
et al. (2002) confirm that in a model with debt renegotiation the assumption
of full tax deduction (ε = 1) is unreasonable and very frequently results in
negative net operating income, they also show that under the general setting
this effect is somewhat minor, and changes in the main results are negligible.
Therefore, in our analysis we will assume the symmetric tax schedule.
To determine the optimal capital structure we will first derive values of
debt and equity, utilising Lemma 2. In actual fact, both of these claims
are special cases of a general claim considered in the previous subsection,
and thus could be easily priced according to Υ(χ, θ), with the corresponding
model parameters that need to be determined.
82 S. Bjerrisgaard and D. Fedoryaev
Let χ0 be the initial level of earnings defined when the first debt issue
occurred, and d ∈ (0, 1) and u ∈ (1,∞) denote the factors for obtaining the
downside and upside earnings levels, respectively. Note that χ0 will natu-
rally change through time when subsequent restructurings will take place, so
it should generally be denoted as χtj , where tj is the time of the jth read-
justment of leverage; we will nevertheless consider only the case of the very
first change in capital structure, since all the other ones will be identical.
It follows from the complete solution to the ODE derived in Subsection 5.2.1
that the value of the debt claim is given by
D(χ) = d1χβ1Φ
(−β1,−2β1 + 2 +
2κ
σ2,
2κθ
σ2χ
)
+ d2χβ2Ψ
(−β2,−2β2 + 2 +
2κ
σ2,
2κθ
σ2χ
)+
(1− τi)Cr
, (5.21)
where
β1,2 =2κ+ σ2 ±
√(2κ+ σ2)2 + 8rσ2
2σ2, β1 > 0, β2 < 0.
The equity claim is priced in a similar fashion by applying Lemma 1:
E(χ) = e1χβ1Φ
(−β1,−2β1 + 2 +
2κ
σ2,
2κθ
σ2χ
)
+ e2χβ2Ψ
(−β2,−2β2 + 2 +
2κ
σ2,
2κθ
σ2χ
)
+ (1− τe)(θ
r+χ− θr + κ
)− (1− τe)C
r. (5.22)
Now we need to specify the appropriate boundary conditions to find the
constants d1, d2, e1, and e2. It should be emphasised that the derivations
that follow rely heavily on Lemma 2, whereby we are able to scale optimal
coupon as well as restructuring and default boundaries by u = χuχ0
.45 For now
45 Since the downward restructuring option is omitted—as is the case when debt holders
Chapter 5. Proposed Dynamic Model with Mean Reversion 83
the coupon C is taken as given; we will state the optimisation programme
it solves after specifying the system of equations needed to price the claims.
Without loss of generality, we let χ0 = 1 and consider all the model variables
in monetary terms to be per unit of earnings.
When earnings fall to χd, the equity holders will choose to default and the
firm will be taken over by debt holders, as we rule out any debt renegotiation
possibilities. An important point is that the company is not liquidated at
default, but is rather maintained to continue operations after the transfer of
control. Thus, the debt value at the lower boundary after correcting for the
bankruptcy costs is
D(χd) = (1− γ)(1− τe)(θ
r+dχ0 − θr + κ
). (5.23)
The equity holders will give up all control rights, so the value of their
claim at default is zero:
E(χd) = 0. (5.24)
When earnings rise to χu, the current debt is called in its entirety. Note
that here we assume that the equity holders can always call the debt at par,
but relax this assumption later. The value of the debt claim at the upper
earnings threshold is
D(χu) = D(χ0). (5.25)
Immediately after the debt is paid back the equity holders will face the same
capital structure choice as they did initially, apart from the fact that the
earnings would now amount to χu. They will optimally determine the new
leverage and debt structure (interest payment in our case) so that the equity
are allowed to relever the firm afterwards and effectively become the new equity holders—scaling down by a corresponding factor d = χd
χ0is not considered.
84 S. Bjerrisgaard and D. Fedoryaev
claim will be valued as
E(χu) = [E(χ0) + (1− q)D(χ0)]u−D(χ0), (5.26)
taking into account the restructuring costs q, which are calculated as a per-
centage of the principal of the new debt and could be interpreted as issuance
expenses paid to a broker, e.g. an investment bank arranging the placement.
The interpretation of the equity claim may seem a little ambiguous at this
stage as it appears that it embraces both equity and debt instruments, but
we will elaborate on this below.
To make further analysis more intuitive, we also introduce two additional
variables defined as
Π , E(χ0) + (1− q)D(χ0), (5.27)
Σ , D(χ0). (5.28)
As has been briefly mentioned in Chapter 4, Π denotes the maximum
potential firm value, which is the price the owner-manager who solely owns
the EBIT-generating machine can sell the firm for. We prefer to call this
value “potential” as the entrepreneur due to e.g. high risk aversion or per-
sonal preferences might choose a suboptimally low leverage or even not have
any debt at all, and thus forego the tax benefit of debt entirely. In case an
acquirer with different preferences, e.g. a PE fund, takes over the company,
the offer price will reflect the potential to increase the total firm value by
taking on debt and exploiting the tax shield, and this is the price the current
equity owners would like to maximise. This principle is central to the LBO
valuation technique commonly applied by investment banks nowadays, but
might also be encountered in other deals where debt financing can potentially
play an important role after the target is acquired. Thus, Π is an abstract
variable which explicitly takes into account the full potential of value cre-
ation from the optimal financing viewpoint. Variable Σ in (5.28) denotes the
Chapter 5. Proposed Dynamic Model with Mean Reversion 85
principal of the current debt, so the definition implies that debt is issued
at par. Finally, it is convenient to express leverage in terms of these two
variables: L = ΣΠ
.
A careful reader would notice that in the boundary conditions d and u
were assumed to be deterministic, but in fact they must be obtained en-
dogenously through the smooth-pasting conditions, implicit in the capital
structure decision of the equity holders. The two smooth-pasting conditions
stipulate the exact points in time when the shareholders should choose to
default or restructure. From a game theory perspective, they can be seen
as the binding incentive compatibility constraints, ensuring that it is always
optimal to declare bankruptcy when the earnings fall to χd and to lever up
the firm when the earnings reach χu. The conditions are as follows:
∂E(χ)
∂χ
∣∣∣∣χ=χd
= 0, (5.29)
∂E(χ)
∂χ
∣∣∣∣χ=χu
= Π. (5.30)
The interpretation of these conditions is fairly straightforward if one recog-
nises that the decision faced by the equity holders represents a classic optimal
stopping problem. The solution to such problem is the set of the smooth-
pasting conditions complemented by the corresponding value-matching con-
ditions (represented by the respective boundary conditions from above). Con-
sider, for example, the determination of the lower earnings threshold, χd. If
the low-contact condition did not hold and the value function were minimised
at some χd < χd, then it would have been optimal for the equity holders to
default after the earnings fall to χd; the reasoning in the opposite case is
analogous, and the two cases together reveal the logic behind (5.29): if the
condition holds, it is optimal for the equity holders to default exactly when
the earnings equal the lower boundary level χd.
Overall, equations (5.23)–(5.30), when combined, form the system of eight
nonlinear algebraic equations in eight unknowns, d, u, d1, d2, e1, e2, Π,
86 S. Bjerrisgaard and D. Fedoryaev
Σ, which could be reduced to two equations in the pair of unknowns d,
u. Due to the nonlinearity of the system as well as numerous confluent
hypergeometric functions being involved in each equation, it is not possible
to obtain analytical solutions, and we will solve the system using numerical
methods in the next section.
Having described the procedure for obtaining prices of the debt and equity
claims, we can state the simple optimisation programme that the owner-
manager will have to solve to find the optimal initial coupon:
C∗ = arg maxC
Π, (5.31)
which means that the owner-manager will choose the debt structure so as to
maximise the potential firm value. (5.31) cannot be solved numerically for
the same reasons as the system of equations described earlier, and we will
apply nonlinear local optimisation methods to solve it in Section 5.3.
Thus, we are now ready to formulate what we call the optimal capital
structure decision framework and the tuple that closes the model and pro-
vides the final solution. The operations in the decision framework are as
follows. Before the initial debt issuance the owner-manager computes the
optimal coupon C∗ in accordance with (5.31); she also finds the optimal
earnings thresholds for restructuring and default as the solution to the sys-
tem of equations (5.23)–(5.30), thus implicitly determining the horizon of
the capital structure decision; finally, she solves for the respective debt and
equity values and determines the optimal leverage L∗. Simultaneous exe-
cution of the decision framework yields the tuple 〈C∗, d∗, u∗,Σ∗,Π∗〉, which
represents the generalised solution to the problem of the owner-manager.46
46 Note that the explicit decision variables are only C∗, d∗, and u∗, while the optimalleverage is directly obtained once they have been determined.
Chapter 5. Proposed Dynamic Model with Mean Reversion 87
5.3 Numerical Results and Model Implica-
tions
The aim of this section is to compare the performance of the dynamic cap-
ital structure framework presented above to that of existing models, and to
investigate whether the assumption of an alternative underlying stochastic
process brings the optimal leverage closer to observed ratios. We carry out
extensive numerical simulations and show both cross-model comparisons as
well as general relationships under our framework.
5.3.1 Note on Numerical Simulations
Before presenting the results of our model we will consider some of the aspects
underlying the numerical procedure utilised to obtain them. We should stress
that since numerical methods form a vast independent research field, we do
not aim to cover all the nuances of numerous advanced techniques involved,
but rather try to sketch the algorithms applied in our numerical simulations
and point out the main inherent drawbacks. We will focus primarily on the
methods that are realised through known built-in functions in traditional
technical computing programmes, e.g. Mathematica or MATLAB, as these
are the ones most commonly used by researchers.
The two mathematical problems in the numerical procedure we carry out
are solving a system of nonlinear algebraic equations and running nonlinear
local optimisation. We will omit the discussion of the optimisation methods
as there is plenty of systematised literature on the topic47; moreover, our local
optimisation programme nests the system of nonlinear algebraic equations,
and the evaluation procedure revealed that most of the computational issues
resulted from the latter.
Numerical algorithms utilised to solve the above problems could be broadly
split into gradient-based methods and direct search methods. The former
ones utilise gradients (or Jacobians, in vector calculus) or Hessians, while the
47 See, for example, Nocedal and Wright (2006).
88 S. Bjerrisgaard and D. Fedoryaev
latter ones do not. Such general distinction is logical as the use of derivatives
could give rise to significant complications as will be discussed below. Direct
search methods, on the other hand, tend to be computationally more expen-
sive, but are more robust. The most widespread fundamental root-finding
algorithms for dealing with nonlinear systems are: Newton’s method, the
secant method, and Brent’s method, with only the first one belonging to the
family of gradient-based techniques. Interestingly, Newton’s method, despite
having a very long history, is still underlying in its different variations the
majority of built-in functions, and is considered a very powerful technique
with quadratic convergence at a simple root and linear convergence at a
multiple root; yet, it still bears certain limitations.
The method can be loosely described as follows. Assume ∃p ∈ [a, b] | f(p) =
0. If f ′(p) 6= 0, then ∃δ > 0 such that the sequence pn∞n=0 defined by the
recurrence relation
pn+1 = g(pn) = pn −f(pn)
f ′(pn), n = 0, 1, 2, . . .
will converge to p for any initial approximation p0 ∈ [p− δ, p + δ].48 One of
the central problems related to the non-convergence of the algorithm is its
dependence on the initial approximation to the root, p0. It is important to
stress here that since equations in our system involve complicated expressions
with nonlinearity of varying degree and numerous confluent hypergeometric
functions, there is no systematic procedure for finding all solutions, even
numerically. Therefore, if the starting approximation is not sufficiently close
to the desired root, the sequence is likely to converge to some other root.
Moreover, an improper starting point may cause an infinite cycle with the
sequence being repeated or almost repeated, thus preventing convergence
entirely. Extensive numerical tests reveal that the behaviour of the functions
in our system is rather unstable, and as a result even a slight change in initial
approximation may indeed have a tangible impact on the solution. Another
48 Since f(p) = 0 it is easy to see that Newton’s method is executed by computing afixed point of the iteration function g(x), i.e. g(p) = p.
Chapter 5. Proposed Dynamic Model with Mean Reversion 89
aggravating issue stems from the use of a derivative of a function in the
iteration. The derivative may turn out to be zero at a particular iteration,
yielding the division-by-zero error, and the value obtained may not be an
acceptable approximation to the root; the derivative at the root may not
exist, and the sequence might in fact be diverging; if the derivative at the
root is discontinuous, the convergence may never take place in an acceptable
neighbourhood of the root; further, there are other less obvious examples that
could lead to potentially distorted results, e.g. in the case when |g′(x)| ≥ 1
on an interval containing the root p, there is a chance of divergent oscillation
(Mathews and Fink, 1999, Section 2.4).49
The secant method is effectively a finite difference approximation of New-
ton’s method and thus has the advantage of avoiding the use of derivatives:
pn+1 = g(pn, pn−1) = pn − f(pn)pn − pn−1
f(pn)− f(pn−1), n = 1, 2, . . .
However, being a two-point iteration mechanism, it exhibits a serious draw-
back as well, as both of the initial points should be close to the root. Besides,
the convergence of this method is normally slower, and the tests on our sys-
tem do not discover its superiority to Newton’s method.
Brent’s method combines the bisection method with the secant method,
imposing some extra conditions at each iteration. The validity of this method
in our case is severely undermined, however, due to the absence of function
continuity, and the results are clearly inferior to those produced under New-
ton’s method. We also try to implement other algorithms, viz. Muller’s
method and Steffensen’s method, but irregularities in the model output still
persist. It appears that the key problem lies in the somewhat high sensitivity
of discussed methods to the initial approximation(s) and thus the algorithm
49 Usually most of the built-in functions offer additional options to Newton’s method,supposed to mitigate some of the mentioned pitfalls. Examples include varying the numberof iterations, controlling for tolerance, introducing a damping factor—which should ideallybe chosen to be the multiplicity of the desired root, to speed up the convergence—orincreasing the precision beyond machine precision. However, as the tests show, if the levelof function complexity is too high, applying these techniques would still not guaranteethat all the issues are rectified.
90 S. Bjerrisgaard and D. Fedoryaev
that is less sensitive or insensitive to the starting point would be desired.
A possible example of such algorithm—the modified Newton’s method—is
described in Atluri, Liu, and Kuo (2009), but its implementation is outside
of the scope of this thesis. Thus, in our final programme module we ap-
ply the damped Newton’s method with additional specifications, and even
though we do conduct an extensive search procedure to choose a correct ini-
tial approximation, the obtained results should nevertheless be interpreted
carefully.
Another important issue related to the numerical procedure is the si-
multaneity of the optimal capital structure decision framework, which may
create additional noise in the results. The problem is that the system (5.23)–
(5.30) should be solved for a given coupon value, while the optimal coupon
is obtained from the optimisation programme (5.31) which in turn implies
exogenously given default and restructuring boundaries. Since the entire de-
cision framework has to be solved numerically, the described simultaneity
stipulates a high dependence of results on the initial value of the coupon.
In an attempt to rectify this issue, we iteratively search for an appropriate
starting point which produces a relatively stable output.
5.3.2 Numerical Analysis
We start out by benchmarking our results against those produced under
the mean-reverting process in the static setting, considered in Sarkar and
Zapatero (2003). Note that to carry out a correct comparison we apply
the same base case parameters that are used in their model, even though
the choice is not easily justifiable.50 Utilising the notation from previous
50 Consider, for instance, bankruptcy costs of 50%, which are the same as in Leland(1994). The value is unreasonably high, contradicting most empirical estimates, and isespecially surprising given that, for example, a well-developed bankruptcy procedure inthe US clearly does not reflect the fact that on average half of the firm value is lost upondefault. Further, the risk-free rate of 7% does not seem to match the historical data eitheras the 30-year US Treasury bond yield, which is a conventional proxy for the risklesssecurity, averaged roughly 5% in 1998–2003.
Chapter 5. Proposed Dynamic Model with Mean Reversion 91
chapters, we set r = 7%, γ = 50%, τc = 15%, τi = 0%51, σ = 40%, κ = 0.1,
and θ = 1.0.
Hypothesis 1. The dynamic mean-reverting model predicts, ceteris paribus,
lower optimal leverage and optimal coupon than the static mean-reverting
model does.
Static MR
Dynamic MR
0.3 0.4 0.5 0.6 0.7 0.8Σ
0.65
0.70
0.75
0.80
0.85
L*
(a) Leverage and earnings volatility
Static MR
Dynamic MR
0.06 0.08 0.10 0.12 0.14 0.16Κ0.2
0.3
0.4
0.5
0.6
0.7
0.8C*
(b) Coupon and speed of mean reversion
Dynamic MR
Static MR
0.04 0.06 0.08 0.10 0.12 0.14 0.16Κ0.2
0.3
0.4
0.5
0.6
0.7
0.8
L*
(c) Leverage and speed of mean reversion
Figure 5.1. Comparison between dynamic and static mean-reverting models.Base case parameters are: r = 7%, γ = 50%, τc = 15%, τi = 0%, σ = 40%, κ = 0.1, andθ = 1.0. Numerical results of the static model in Sarkar and Zapatero (2003) are taken asgiven. Note that (a) assumes a different value for the speed of mean reversion: κ = 0.2(as in Sarkar and Zapatero), which explains a difference in leverage levels.
The juxtaposition in Figure 5.1 demonstrates that the optimal leverage is
indeed lower in the dynamic model—the result reminiscent of the comparison
51 Even though parameter τ in Sarkar and Zapatero (2003) is meant to take into accountpersonal and corporate taxation, it effectively plays the role of the corporate tax in theclaim valuation, and thus in the capital structure decision. Moreover, tax on interestincome does not enter the value of the debt claim, and debt holders receive the couponpayment in full.
92 S. Bjerrisgaard and D. Fedoryaev
between dynamic and static GBM-based models.52 The interpretation is also
similar: in the static model the leverage level is too high as the management
needs to optimise given the infinite expected earnings path and thus chooses
to commit to the capital structure that would yield the largest net gain in
the long term; however, as soon as the option to relever the firm in the future
is granted, a more conservative capital structure is chosen upfront, optimally
balancing bankruptcy costs of debt and the tax shield benefits already in
the short run. Thus, the dynamics of the model enables a more accurate
calibration of the debt level, and the latter is lower as a consequence. Note
that the difference between the leverage levels in dynamic and static models
is not as sharp as in the case when earnings follow a geometric Brownian
motion, which could be explained by the inherent stability of the mean-
reverting process dictated by reversion to mean and bounded conditional
long-term variance. Applying the same reasoning as above it is easy to see
why the optimal coupon is lower under the dynamic setting.
The direction of relationships is analogous between the two models, as
expected. The optimal leverage is decreasing in earnings volatility, which
implies that firms try to neutralise the effect of increased business risk, to
prevent the default probability from soaring. Interestingly, the gap between
the optimal leverage levels widens as volatility grows substantially. This re-
flects the logic described above: the implicit time horizon of the optimisation
problem faced by the management is so long under the static setup that the
volatility elasticity of leverage is very low. On the other hand, in the dynamic
model managers are much more cautious—knowing that the opportunity to
lever up will still be present in the future—which is further amplified by the
fact that in this case there is also an upper earnings threshold present and
52 Here we should stress that since the mathematical representation of the programmewe are solving is extremely computationally-intensive due to the presence of numerousconfluent hypergeometric functions and nonlinearity of varying degree, it is not possible tocreate continuous graphical forms, and therefore we resort to the procedure of discretisationand interpolation. The unstable behaviour of these functions also explains occasionalnarrowness of considered intervals.
Chapter 5. Proposed Dynamic Model with Mean Reversion 93
restructuring is costly.53 Furthermore, as the speed of mean reversion aug-
ments, the earnings converge to their long-term mean faster, so the cash flow
stream becomes more predictable. This effect shows that the mean reversion
speed counteracts the influence of volatility which explains why the optimal
leverage rises. Hence, direct relationships in Results 2 and 3 from Sarkar and
Zapatero (2003) are confirmed under the dynamic setting.
Model dynamics also has an important implication for the lower earnings
threshold:
Hypothesis 2. The static GBM-based model offers a low default bound-
ary, while the static mean-reverting model suggests a much higher bankruptcy
threshold. The intermediate value is reached when the dynamics is introduced
under the assumption of mean reversion, i.e.
χStaticGBMd < χDynamicMRd < χStaticMR
d
and the threshold is further decreased when the earnings converge slower.
It should firstly be noted that the default boundary in the GBM-based
model is predicted to be lower than that in the static mean-reverting model.
This is due to the fact that a short-term downswing in earnings of the same
magnitude would be perceived differently in the two cases, the former being
generally less sensitive. Equity holders are aware that in the static GBM-
based model the earnings will keep following a positive trend, dictated by
the drift, while in the static mean-reverting model they are bounded by the
occasional fluctuation above the long-run mean, and thus it will be unlikely
to recoup the incurred losses as fast. Besides, the default earnings threshold
in the dynamic mean-reverting model is expected to be lower than that in
the static mean-reverting model. The intuition behind this is simple: the
possibility of altering the capital structure dynamically will, ceteris paribus,
53 It should be pointed out that the comparison of the discrepancy in e.g. optimalleverage across comparative statics for different parameters would not be valid as there isno bijective mapping between them.
94 S. Bjerrisgaard and D. Fedoryaev
Static GBM
Dynamic MR
Static MR
0.2 0.3 0.4 0.5 0.6 0.7Σ0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7d
(a) Default boundary and earningsvolatility
Static GBM
Dynamic MR
Static MR
0.10 0.15 0.20 0.25Κ
0.1
0.2
0.3
0.4
0.5
0.6
0.7d
(b) Default boundary and speed of meanreversion
Figure 5.2. Comparing default boundary in the dynamic mean-revertingmodel, the static mean-reverting model, and the static GBM-based model.Base case parameters are: r = 7%, γ = 50%, τc = 15%, τi = 0%, σ = 40%, κ = 0.1,and θ = 1.0. Results under the static mean-reverting model are simulated by solvingnumerically the system in Sarkar and Zapatero (2003).
enhance the valuation of the firm and the shareholders would try to salvage
such firm even if its earnings fall to the bankruptcy level of the static case.
Consequently, what is left to determine is the relationship between the
threshold values in the case of the dynamic mean-reverting and the static
GBM-based models. This is easy to see if one recognises that even though
the earnings are no longer bounded when the capital structure dynamics is
incorporated in the mean-reverting setup—since the long-term mean will be
shifted in time—the overall growth will still be slower due to the fact that
the drift becomes negative as soon as the current value exceeds the long-term
mean. In fact, the diverging nature of the geometric Brownian motion process
stipulates a much lower boundary value: even with a somewhat moderate
volatility estimate of 25%, equity holders will declare bankruptcy only after
the earnings drop to a quarter of their initial value.
Figure 5.2 confirms the statement of Hypothesis 2 for varying earnings
volatility and speed of mean reversion. From the direct relationship between
the latter and the default threshold it is clear that as the speed of earn-
ings convergence to the long-term mean slows down, our model suggests a
lower boundary value. Moreover, we also find that if the industry outlook is
Chapter 5. Proposed Dynamic Model with Mean Reversion 95
improved, i.e. if analysts predict higher expected long-term earnings, then
the default boundary is lowered. This effect could be explained by the fact
that given a more positive forecast, equity holders will become more patient
and thus even when earnings fall to the level when default would have previ-
ously been declared, they will still keep the company alive, anticipating the
optimistic prospects.
The subsequent numerical results are based on our set of parameters
which is represented by the latest available Danish data as well as empiri-
cal findings. We set r = 3.5%, which corresponds to the average yield for
2008–11 on the Danish government bonds with longest maturity available,
10 years; Danish corporate tax rate is τc = 25%54, and dividends are taxed at
τd = 28%, yielding an effective tax rate for equity holders of τe = 46%; inter-
est income in Denmark is taxed at τi = 37%. The rest of the parameters are
determined as follows. Recent empirical studies estimate bankruptcy costs to
vary within the interval 2–15% of the firm value at bankruptcy (Strebulaev,
2007; Bris, Welch, and Zhu, 2006; Bris, Schwartz, and Welch, 2005). We only
consider direct bankruptcy costs as the indirect costs of financial distress have
a very high variation; even direct costs of bankruptcy procedure could change
substantially depending on firm-specific characteristics, e.g. size (Grinblatt
and Titman, 2002, p. 560), and since the capital structure optimisation is
most relevant for larger corporations, we choose a more conservative esti-
mate of γ = 5%. Note that the difference in assumed bankruptcy costs is
striking—in the dynamic model we do not have to assume such unreasonably
high values as in Sarkar and Zapatero (2003) or Leland (1994) to obtain re-
sults consistent with practice. We assume readjustment costs q to comprise
1% of the new debt issue, consistent with empirical estimates (Davydenko
and Strebulaev, 2007; Kim, Palia, and Saunders, 2003) and fee structures
proposed by investment banks for larger financings across debt capital mar-
kets. There are almost no papers that document the volatility of operating
income and since it is known to be generally lower than that of net income
54 To preserve the homogeneity property of the model we ignore the variability of taxrates across firms or time.
96 S. Bjerrisgaard and D. Fedoryaev
(Petrovic, Manson, and Coakley, 2009), we apply a conservative estimate of
25%. Finally, the two key parameters describing the mean-reverting nature
of the process—the long-term mean and the convergence speed—are set to
θ = 1.0 and κ = 0.1, respectively. It is important to emphasise that values of
mean reversion parameters play more an indicative role, since they are very
industry-dependent per se and as such are merely meant to facilitate mod-
elling of particular industries. Widely-varying discrepancy of mean reversion
patterns in earnings and profitability between industries has been confirmed
in the literature (Nordal and Næs, 2010; Altunbas, Karagiannis, Liu, and
Tourani-Rad, 2008).
Let us now turn to comparing the output of numerical simulations under
the dynamic model with mean reversion and the dynamic GBM-based model
presented in Chapter 4 to draw some more subtle inferences.
Hypothesis 3. The dynamic mean-reverting model suggests lower optimal
leverage and higher restructuring frequency than the dynamic GBM-based
model does, and the former relationship is more pronounced if the long-term
industry outlook is worsened or the speed of earnings convergence is slowed
down.
Conclusions in Hypothesis 3 are reinforced by Figure 5.3. First, we ob-
serve that assuming the underlying mean-reverting process brings down the
optimal leverage compared to that under a geometric Brownian motion. This
phenomenon has an intuitive explanation: GBM-based models always imply
a positive earnings trend due to the nature of the drift, thus offering a more
favourable setup as opposed to the mean-reverting model, where drift could
take both positive and negative direction depending on the current level of
earnings. Moreover, it is important to point out that assuming the same
volatility of earnings for both models may downplay the difference in opti-
mal leverage. This follows from the fact that our process has an inherently
bounded variance as a result of reversion to mean, by contrast with a geo-
metric Brownian motion whose conditional variance is infinite. Therefore, to
carry out a more accurate comparison the earnings volatility assumed in the
Chapter 5. Proposed Dynamic Model with Mean Reversion 97
Dynamic GBM
Dynamic MR
0.26 0.27 0.28 0.29 0.30 0.31 0.32Σ
0.24
0.26
0.28
0.30
0.32
0.34
0.36
L*
(a) Leverage and earnings volatility
Dynamic MR, Σ = 40 %
Dynamic MR, Σ = 25 %
Dynamic GBM
0.03 0.04 0.05 0.06 0.07Γ
0.25
0.30
0.35
0.40L*
(b) Leverage and bankruptcy costs
Dynamic MR
0.10 0.11 0.12 0.13 0.14 0.15Κ
0.35
0.40
0.45
0.50
0.55
0.60
0.65L*
(c) Leverage and speed of mean reversion
Dynamic MR
0.98 0.99 1.00 1.01 1.02 1.03Θ
0.315
0.320
0.325
0.330
0.335L*
(d) Leverage and long-term mean
Dynamic GBM
Dynamic MR
0.004 0.006 0.008 0.010 0.012 0.014q1.05
1.10
1.15
1.20
1.25
1.30
1.35
1.40
u
(e) Restructuring boundary andrestructuring costs
Figure 5.3. Comparison between dynamic mean-reverting and dynamic GBM-based models. Base case parameters are: r = 3.5%, τc = 25%, τd = 28%, τi = 37%, γ =5%, q = 1%, σ = 25%, θ = 1.0, κ = 0.1, and µ = 2%. (b) additionally exhibits thedynamic mean-reverting model with an increased volatility level of 40%, while all otherparameters kept unchanged.
98 S. Bjerrisgaard and D. Fedoryaev
base case should be higher for the mean-reverting model. It could be easily
shown that the stationary variance of the process given by (5.1) is
V[χs] =θ2σ2
2κ− σ2, (5.32)
and so we increase the initial volatility level accordingly, to bring the vari-
ance of the process closer to that of a geometric Brownian motion. The
outcome is demonstrated in Figure 5.3(b): we see that the leverage in our
model is lowered, making the gap between the two considered models more
pronounced.
Further, one should note that the above discrepancy was observed for
the base case with implied neutral outlook on earnings and average mean
reversion speed. If the long-term earnings forecast turns negative, i.e. the
predicted mean value of operating income is lower than the current value,
then the optimal leverage would be decreased, thus also widening the gap
between the GBM-based and mean-reverting models. The same impact has
slowing down the earnings convergence speed which reduces the stability of
the earnings stream. Both effects are evident from Figures 5.3(c,d), and
could be observed in practice as well, e.g. during the period of economic
stagnation: at its beginning, sector analysts would publish more negative
forecasts, and after that would maintain them for longer-than-usual as the
economy keeps stagnating and recovery of the real sector is slow. Besides, we
also see that the optimal leverage is much more sensitive to the changes in the
mean reversion speed. This could be explained by the fact that altering the
convergence speed has an immediate bearing on the firm’s activities through
increasing the certainty of cash inflows, as opposed to the long-term forecast
that only has a partial effect at the current point in time and comes into play
more gradually.
Higher restructuring frequency for the mean-reverting model could be ob-
served in Figure 5.3(e).55 If the earnings deviate too much from the long-term
55 Our interpretation of frequency is akin to that in physics—viz. frequency of waves—toemphasise the continuity of the process. Therefore, higher frequency does not necessarily
Chapter 5. Proposed Dynamic Model with Mean Reversion 99
0.05 0.10 0.15 0.20 0.25 0.30 0.35Σ0.6
0.7
0.8
0.9
1.0
1.1
1.2
d,u
(a) Earnings boundaries andearnings volatility
0.04 0.06 0.08 0.10 0.12 0.14 0.16Κ0.4
0.6
0.8
1.0
1.2d,u
(b) Earnings boundaries and speed ofmean reversion
Figure 5.4. Upper and lower boundaries in the dynamic mean-reverting model.Base case parameters are: r = 3.5%, τc = 25%, τd = 28%, τi = 37%, γ = 5%, q =1%, σ = 25%, θ = 1.0, and κ = 0.1.
mean, they will start reverting back straight away, and thus the restructuring
boundary should be lower compared to that in the GBM-based model, oth-
erwise the threshold may never be reached. It is also worth mentioning that
the difference is quite substantial: in the base case, for example, the expected
earnings growth up to the restructuring point is almost 15 percentage points
higher in the GBM-based model.
Finally, we will point out some general relationships under the dynamic
mean-reverting framework, so henceforth the focus is restricted solely to the
results under our model.
Figure 5.4(a) presents the dependence of the lower and upper earnings
thresholds on the earnings volatility. We see that the gap is diverging, which
reflects the fact that when the business risk is increased, both boundaries
will be altered accordingly as equity holders prefer to keep their strategic
decisions unaffected by short-run fluctuations in operating income. Higher
speed of earnings convergence has the opposite effect, as expected: more
stable businesses will decrease the upper boundary to keep exploiting the tax
advantage to debt; the other relationship is more subtle, however, because it
may not be instantly clear why it is optimal to hike the default boundary. As
imply higher number of times the boundary has been hit.
100 S. Bjerrisgaard and D. Fedoryaev
0.370 0.375 0.380 0.385 0.390Τi
0.319
0.320
0.321
0.322
0.323
0.324
0.325
0.326
L*
(a) Leverage and tax on interest income
0.46 0.47 0.48 0.49 0.50 0.51 0.52Τe
0.33
0.34
0.35
0.36
0.37L*
(b) Leverage and effective tax rate
Figure 5.5. Optimal leverage and tax structure in the dynamic mean-revertingmodel. Base case parameters are: r = 3.5%, τc = 25%, τd = 28%, τi = 37%, γ =5%, q = 1%, σ = 25%, θ = 1.0, and κ = 0.1.
the speed of mean reversion augments, equity holders will expect to finance
the debt service deficit56 from their own pockets more and more often as the
optimal coupon will be rising—cf. Figure 5.1(b)—and the situation when
the earnings are insufficient will be occurring more frequently. Therefore,
the lower boundary will be pushed up, diminishing the spread between the
boundaries, just as would be the case if the earnings volatility went down.57
From Figure 5.5 it is evident that the optimal leverage is decreasing in
the tax on interest payments and increasing in the effective tax rate imposed
on equity holders. These relationships are a direct consequence of the form of
the after-tax payoff of the owner-manager, described in (5.20). When the tax
rate on interest income rises or, alternatively, the effective tax rate declines,
the tax advantage to debt shrinks, thus reducing the incentive to utilise debt
financing which ultimately results into lower optimal leverage. The direct
impact of changes in the tax structure could also be seen from the valuation
of debt and equity claims.
56 Since asset sale or extra debt issues are not allowed in the model, consistent withprotective covenants utilised in practice, every time the situation χt < C occurs, therequired cash is provided by equity holders.
57 Perhaps one could understand this relationship a bit easier if a decrease in the speed ofmean reversion is considered. Slower earnings convergence means that they will stay belowthe long-term mean for longer, and thus, given the same volatility, the default probabilityis increased, prompting equity holders to decrease the bankruptcy threshold.
Chapter 5. Proposed Dynamic Model with Mean Reversion 101
0.015 0.020 0.025 0.030 0.035r
1.145
1.150
1.155
1.160
1.165
1.170
1.175
u
(a) Restructuring boundary andrisk-free rate
0.015 0.020 0.025 0.030 0.035 0.040 0.045 0.050r0.0
0.2
0.4
0.6
0.8
C*
(b) Coupon and risk-free rate
0.015 0.020 0.025 0.030 0.035 0.040 0.045 0.050r0.5
0.6
0.7
0.8
0.9d
(c) Default boundary and risk-free rate
0.015 0.020 0.025 0.030 0.035 0.040 0.045 0.050r
400
450
500
550
600
650S
(d) Credit spread and risk-free rate
Figure 5.6. Effect of varying risk-free rate on restructuring boundary, optimalcoupon, default boundary, and credit spread (in bps) in the dynamic mean-reverting model. Base case parameters are: r = 3.5%, τc = 25%, τd = 28%, τi =37%, γ = 5%, q = 1%, σ = 25%, θ = 1.0, and κ = 0.1.
A consequence of the risk-free rate hike is that the discounted value of the
tax benefit of debt received in future periods is lowered, and hence the owner-
manager chooses a higher coupon to counterbalance this effect, which in turn
enhances the debt valuation. The overall impact on the debt value—through
both increased risk-free rate and optimal coupon—is more pronounced com-
pared to that on the coupon payment and therefore debt becomes relatively
cheaper when the risk-free interest rate rises, as illustrated by the credit
spread in Figure 5.6(d).
To interpret the relationship between the restructuring boundary and the
risk-free rate first notice that as the latter rises, the optimal coupon is in-
creased, and the effect is somewhat strong. Therefore, a low riskless rate
102 S. Bjerrisgaard and D. Fedoryaev
results in a low optimal coupon, meaning that the utilisation of the tax
shield is quite weak. This induces the equity holders to set the restructuring
threshold low to begin exploiting the tax benefit of debt earlier. On the other
hand, if the risk-free interest rate is high, this implies that the tax shield is
already utilised heavily, and that the firm may already be running a high
risk of bankruptcy because the debt service constitutes a large proportion of
the earnings. This stipulates a high value of the upper earnings boundary.
The monotonicity of the relationship between the two extremes we consid-
ered is also easily shown: whenever the risk-free rate rises, this eventually
leads to increased default probability58, so the equity holders will prefer not
to restructure as often and the upper threshold will be raised. Another ar-
gument in favour of the positive relationship stems from the intertemporal
discrepancy in cash flows: the costs of restructuring are incurred now, while
the present value of the benefits is diminished as they are discounted with a
higher interest rate. This result is interesting as it contrasts with that ob-
tained in the dynamic GBM-based model of Goldstein et al. (2001), whose
tabulated simulations show that the restructuring boundary declines as the
interest rate rises. The authors do not comment on the relationship, and
even though it could be partially justified by the nature of the process—for
low values of the risk-free rate, a slight increase may lead to the decline in
the restructuring threshold as the equity holders anticipate a positive drift
in earnings and a slightly higher default probability might not matter as
much—we still find such relationship quite puzzling generally, and expect
our result to be more consistent with practice.
5.4 Extensions
In this section we will extend the analysis of dynamic models under alterna-
tive cash flow processes. Of course, possible directions of extension are nu-
58 This follows from the fact that when a coupon is higher, the low-contact conditioncombined with the boundary condition for equity at default will stipulate a higher defaultthreshold. This effect is demonstrated in Figure 5.6(c).
Chapter 5. Proposed Dynamic Model with Mean Reversion 103
merous, and examples include considering endogenous finite maturity, debt
subordination, post-bankruptcy restructuring option, and so forth; technical
extensions such as e.g. stochastic interest rates or non-constant volatility
could also be considered. Nevertheless, we choose to restrict our attention to
the following two: incorporating fixed operating costs in the dynamic GBM-
based model and including a call premium in the dynamic mean-reverting
model. The former one plays a dual role as it enables us to resolve one of the
well-known issues of capital structure modelling—cash flow being bounded
at zero—and to directly compare the effects of the two changes in the nature
of the cash flow process, viz. allowing cash flow to become either negative or
mean-reverting. The second extension shows that the basic dynamic model
with mean reversion introduced in this chapter might understate the optimal
leverage if equity holders possess the option of redeeming the entire debt at
par.
5.4.1 Fixed Operational Costs
Let us first consider the model with the earnings being unrestricted. An
obvious shortcoming of GBM-based models as well as of the mean-reverting
model proposed above is that they always imply a positively-valued state
variable. Negative earnings are fairly often observed in practice, e.g. in
start-ups or in mature firms during the periods of recession, and could also
be attributed to other industry-specific factors, for instance, to the prevalence
of fixed costs in the cost structure combined with demand cyclicality or
intensified competition.
One of the ways to introduce a possibility of non-positive earnings is to
assume that the state variable follows an arithmetic Brownian motion, but
that would lead to the loss of homogeneity property, significantly complicat-
ing the modelling of capital structure dynamics. Therefore, we apply another
solution—redefining the state variable as the firm revenue and incorporating
fixed operating costs in the model. These costs, denoted F , are continuously
paid by equity holders, and thus the resulting net cash flow is χ−F −C, so
104 S. Bjerrisgaard and D. Fedoryaev
that the earnings, due to the stochastic nature of the state variable, could
take on negative values even before the debt is serviced.
In this subsection we explore the implications for optimal capital structure
that incorporating fixed operating costs in the dynamic GBM-based model
described in Chapter 4 has. The obtained results are benchmarked against
those under the dynamic mean-reverting model. Note that the inclusion of
fixed costs would lead to the following changes in the system of equations.
First, the value of the equity claim becomes lower as the costs are entirely
borne by equity holders:
E(χ) = e1χβ1 + e2χ
β2 +(1− τe)χr − µ
− (1− τe)(C + F )
r. (5.33)
Further, the boundary condition for debt at default is also altered: to pre-
clude any strategic considerations, debt holders are assumed to manage the
company just as well as equity holders did and thus have to pay the operating
costs of the same amount after taking over the firm upon bankruptcy, i.e.
D(χd) = (1− τe)(1− γ)dχ0
r − µ− (1− τe)F
r. (5.34)
After implementing the above changes, we solve for the optimal capital
structure decision tuple numerically in the same fashion as before, and apply
the same set of base case parameters, viz. r = 3.5%, τc = 25%, τd =
28%, τi = 37%, γ = 5%, q = 1%, σ = 25%, and µ = 2%. For the fixed
costs we use a moderate estimate of F = 0.2.
First we study the dependence of the key variables on operating fixed
costs in the dynamic GBM-based model, using the framework introduced in
Chapter 4 as a benchmark.
Hypothesis 4. In the dynamic GBM-based model with fixed costs, higher
operating leverage59 implies higher optimal coupon, increased default and re-
59 Strictly speaking, for operating leverage to be defined, the model should embracevariable costs as well. However, for the purposes of comparative statics analysis variablecosts are assumed to be zero, so that varying fixed operating costs would unambiguously
Chapter 5. Proposed Dynamic Model with Mean Reversion 105
structuring boundaries, higher credit spread, and lower financial leverage.
Figure 5.7(a) demonstrates the reverse impact on the financial leverage.
We see that when fixed costs are assumed to be zero, the two models yield
the same optimal leverage, and when we increase the costs, the financial
leverage gradually drops. Comparison between the two base cases reveals
that even when the fixed costs amount to a moderate value of 0.2, the optimal
leverage drops by almost 10 percentage points. The intuition behind the
relationship is as follows. An increase in fixed costs implies a higher default
probability because the equity value falls to zero at a higher earnings level,
i.e. the bankruptcy threshold is augmented. Bondholders, ceteris paribus,
will demand a higher coupon as well as yield since debt becomes riskier: with
higher fixed costs even a short-term downswing dictated by volatility could
already trigger default.60 As a consequence of the interplay between these
effects, the optimal leverage goes down as well. Further, equity holders will
naturally raise the restructuring boundary to balance out the readjustment
frequency with riskiness of debt.
The cross-model comparison leads to the following hypothesis:
Hypothesis 5. The dynamic GBM-based model that accommodates fixed
operating costs, exhibits lower optimal leverage than the dynamic model with
mean reversion does, i.e.
L∗DynamicGBMFC< L∗DynamicMR < L∗DynamicGBM .
The hypothesis is reinforced by Figure 5.8 which shows the dynamics of
the optimal leverage for varying bankruptcy costs and cash flow volatility.
This result can be explained by the different nature of cash flow process mod-
ifications that we consider. Incorporating fixed operating costs in the model
stipulate the change in operating leverage.60 It may appear odd to the reader that the optimal leverage and the optimal coupon
payment are in fact moving in the opposite directions. However, we emphasise that this isprimarily due to the fact that they are both determined simultaneously from the optimalcapital structure decision framework; moreover, this relationship does not always takeplace—cf. Figures 5.1(a,b).
106 S. Bjerrisgaard and D. Fedoryaev
Dynamic GBM
Dynamic GBM with FC
0.0 0.1 0.2 0.3 0.4 0.5 0.6F
0.20
0.25
0.30
0.35
L*
(a) Leverage and operating costs
Dynamic GBM with FC
Dynamic GBM
0.0 0.1 0.2 0.3 0.4 0.5 0.6F
1.0
1.2
1.4
1.6
1.8
2.0C*
(b) Coupon and operating costs
Dynamic GBM
Dynamic GBM with FC
0.1 0.2 0.3 0.4 0.5 0.6F
0.2
0.3
0.4
0.5
d
(c) Default boundary and operating costs
Dynamic GBM with FC
Dynamic GBM
0.0 0.1 0.2 0.3 0.4 0.5 0.6F
1.2
1.4
1.6
1.8
2.0
u
(d) Restructuring boundary andoperating costs
Dynamic GBM with FC
Dynamic GBM
0.0 0.1 0.2 0.3 0.4 0.5 0.6F300
400
500
600
700S
(e) Credit spread and operating costs
Figure 5.7. Effect of varying operating leverage in the dynamic GBM-basedmodel with fixed costs (FC). The standard dynamic GBM-based model serves as abenchmark. Base case parameters are: r = 3.5%, τc = 25%, τd = 28%, τi = 37%, γ =5%, q = 1%, σ = 25%, µ = 2%, and F = 0.2.
Chapter 5. Proposed Dynamic Model with Mean Reversion 107
Dynamic GBM with FC
Dynamic MR
Dynamic GBM
0.035 0.040 0.045 0.050 0.055 0.060 0.065 0.070Γ
0.25
0.30
0.35
L*
(a) Leverage and bankruptcy costs
Dynamic GBM
Dynamic MR
Dynamic GBM with FC
0.25 0.26 0.27 0.28 0.29 0.30 0.31 0.32Σ
0.20
0.25
0.30
0.35
L*
(b) Leverage and earnings volatility
Figure 5.8. Comparison between the dynamic GBM-based model with andwithout fixed costs and the dynamic mean-reverting model. Base case parametersare: r = 3.5%, τc = 25%, τd = 28%, τi = 37%, γ = 5%, q = 1%, σ = 25%, θ = 1.0, κ =0.1, µ = 2%, and F = 0.2.
has an immediate negative bearing on the cash flow process and there is no
uncertainty in the effect as the costs are constant and continuous; the effect
of introducing mean reversion in earnings is much more subtle: firstly, it is
not unambiguous—depending on the current earnings value, mean reversion
can have both positive and negative influence on the cash flow development,
and secondly, it is less pronounced if a reasonable assumption regarding the
magnitude of earnings convergence speed is made. Consequently, the opti-
mal capital structure is more sensitive to the introduction of fixed operating
costs in the model. It is worth noting, however, that the optimal leverage
predicted by this model seems to provide estimates that are lower than em-
pirical regularities, even for a moderate value of fixed costs. In the base case,
for example, it yields the leverage of 26%, as opposed to 33% observed un-
der the dynamic mean-reverting model, whereas studies investigating Danish
corporates document estimates in the range 30–35% (among others, Gaud
et al., 2007; Chen and Hammes, 2003). Plus, as shown above, the optimal
leverage drops somewhat fast when fixed costs are increased, bringing the
model forecast even further from the magnitude consistent with practice.
108 S. Bjerrisgaard and D. Fedoryaev
5.4.2 Call Premium
The dynamic mean-reverting model presented in Section 5.2 considered debt
being callable, so that equity holders could change the leverage level before
the outstanding bonds matured. More importantly, it was assumed that this
option could be exercised at no cost, potentially resulting in too frequent
restructuring. Furthermore, empirical evidence suggests that corporations
indeed do pay a premium when bonds are retired early (Mitchell, 1991; Fis-
cher et al., 1989b; Vu, 1986). Therefore, the model will now be adjusted so
that a call premium needs to be paid by equity holders when the old debt
is redeemed prior to maturity. Since the inclusion of a call premium only
affects the value of debt and equity claims at the restructuring point, only
these boundary conditions will be stated—other equations remain unchanged
from the previous section.
When the upper boundary is reached and equity holders will optimally
decide to retire old bonds, the value of debt will be given by
D(χu) = (1 + Λ)D(χ0),
where Λ denotes a call premium, modelled as a fraction of the bond value.61
Following the same logic, the equity value at restructuring is now
E(χu) = [E(χ0) + (1− q)D(χ0)]u− (1 + Λ)D(χ0).
Note the distinction between the new debt level, uD(χ0), and old debt level,
D(χ0), as restructuring costs are a fraction q of the new bond value, while
the call premium is paid only on top of the old principal. Further, as the call
premium is paid before equity holders decide on the optimal capital structure,
61 We acknowledge that empirical research documents call premia being not constant,but rather increasing in time to maturity and dependent on firm characteristics, e.g.volatility (Fischer et al., 1989b). However, for the purposes of tractability we assume a callpremium to be an exogenously given constant. Additionally, we abstain from includingany deferral period and thus allow equity holders to call outstanding bonds at any timeafter issuance.
Chapter 5. Proposed Dynamic Model with Mean Reversion 109
Π and Σ from (5.27) and (5.28) are unaltered, as is the maximisation problem
in (5.31). The optimal bankruptcy and readjustment thresholds d and u are
still found using numerical procedures, and leverage is calculated through
the coupon which optimises the maximum potential firm value.
It should also be mentioned that even though some studies consider a
possibility of choosing the call premium optimally by equity holders (Hen-
nessy and Tserlukevich, 2008; Flor and Lester, 2002; Fischer et al., 1989b),
we exclude this option and instead assume that this parameter is exogenously
given—following Christensen et al. (2002) and Dangl and Zechner (2004)—
to maintain the focus on those decision variables which are more directly
associated with the capital structure decision.
To see what implications the call premium has for the optimal composi-
tion of debt and equity, let us consider some numerical results.
Hypothesis 6. If retiring the outstanding debt prior to maturity entails pay-
ing a call premium, then equity holders will choose a higher optimal leverage
and will readjust the capital structure less often.
0.00 0.01 0.02 0.03 0.04 0.05L
0.30
0.35
0.40
0.45
0.50L*
(a) Leverage and call premium
0.00 0.01 0.02 0.03 0.04 0.05L1.10
1.15
1.20
1.25
1.30
1.35
1.40u
(b) Restructuring boundary and callpremium
Figure 5.9. Effect of varying call premium on optimal leverage and restructur-ing boundary in the dynamic mean-reverting model. Base case parameters are:r = 3.5%, τc = 25%, τd = 28%, τi = 37%, γ = 5%, q = 1%, σ = 25%, θ = 1.0, andκ = 0.1.
Introducing the call premium has a positive effect on the optimal leverage.
Since the capital structure decision is made at the point where both debt
110 S. Bjerrisgaard and D. Fedoryaev
and equity claims belong to the owner-manager, the latter would favour debt
financing as the value of a premium-bearing debt is higher. Moreover, the
benefit from the enhanced debt valuation comes today, while the downside of
having to pay the premium lies only in the future, when restructuring takes
place, and thus it is less pronounced due to discounting. The mathematics
of this relationship could be easily seen from the expression for leverage, viz.
L = ΣE(χ0)+(1−q)Σ , as the nominator is always increasing faster in the call
premium because of the presence of restructuring costs in the model and the
reverse impact of the call premium on the equity value.
Furthermore, as the call premium rises, it will become more and more
expensive for equity owners to retire the old debt prematurely, and thus
they will have less incentive to restructure. Therefore, when a premium to
calling outstanding bonds before maturity is introduced, the firm will decide
to refinance less frequently, and this effect is illustrated by Figure 5.9(b).
Chapter 6
Conclusion
In this thesis we analyse in detail the development of research focusing on
modelling optimal capital structure, and attempt to extend it by proposing
the dynamic mean-reverting model that rectifies some of the shortcomings
of existing studies. The scope of analysis is further broadened to consider
an alternative modification of the cash flow process—allowing earnings to
obtain negative values. We bring together the key dynamic models and try
to uncover the implications that different frameworks have for the results
through comprehensive numerical simulations.
We start out by revisiting the pricing model introduced in the seminal
work of Merton (1974), which represents the fundamental building block for
most capital structure models. The general partial differential equation is
rederived along the Black–Scholes lines, and it is shown how this valuation
framework could be applied to price debt and equity as contingent claims;
an intuitive analogy with the option pricing model is further provided. We
also demonstrate that the oft-employed assumption of time independence—
used to make a transition from the partial differential equation to the ordi-
nary differential equation—may violate the conditions for the existence of an
equivalent martingale measure.
The first stage of our analysis is considering the basic static model akin
to that in Leland (1994). The valuation of claims is derived using the ODE
111
112 S. Bjerrisgaard and D. Fedoryaev
and corresponding boundary conditions for the case when the bankruptcy
decision of equity holders is endogenised and the optimal coupon is found
from the firm value maximisation problem. We present comparative statics
for optimally chosen leverage, coupon, and default threshold as well as for tax
advantage to debt. It is evident that under our own base case parameters—
based on the most recent Danish data and empirical estimates—which differ
from unreasonably high values applied in Leland, the optimal leverage is
too high compared to that observed in practice. Moreover, the prominent
inconsistency of this model lies in the fact that the state variable is the value
of unlevered assets which represents the price of a traded asset, thus opening
up arbitrage opportunities. Apart from that, the model embraces a very
simple setup, ignoring e.g. personal taxation, and the assumed static nature
of capital structure is very limiting as well.
Therefore, the natural progression of our analysis is introducing dynam-
ics in the model of capital structure. Rather than focusing on any particular
model, we outline a generalised framework to describe the overall family of
existing dynamic capital structure models. The state variable is the operat-
ing income, which restores the absence of arbitrage in the model. The firm
value is reinterpreted as the maximum potential firm value, which reflects the
debt financing potential at the point in time when both claims on debt and
equity belong to the owner-manager. Further, some other previous inconsis-
tencies are rectified due to a more elaborate setup, e.g. equity value is no
longer increasing in the effective tax rate. As the leverage can be readjusted,
the upper earnings boundary is incorporated, stipulating when restructur-
ing should take place; it is obtained from the high-contact condition which
makes the refinancing decision incentive-compatible from the viewpoint of
equity holders. The key insight from the comparative statics analysis of the
model is that the optimal leverage drops substantially compared to that in
the static model, and the default boundary is also lowered, reflecting the fact
that the refinancing option, ceteris paribus, enhances the firm valuation.
In the rest of the thesis we develop our own model which alters the cash
Chapter 6. Conclusion 113
flow process by assuming mean reversion. The vast majority of existing cap-
ital structure studies rests on the assumption that the state variable follows
a geometric Brownian motion since this ensures mathematical tractability
while modelling the capital structure readjustment. However, this underly-
ing process implies that the state variable—represented by earnings—could
in time reach infinitely high values, and moreover, that due to the positive
drift, its value is always expected to increase, with short-term falls being
governed only by volatility. We argue that such behaviour is perhaps more
consistent with the development of the stock price, rather than with dy-
namics of the firm fundamental, and that the assumption of mean reversion
appears to be more suitable. Not only the latter is generally more consistent
with the realities of the business world, but is also confirmed by a number of
empirical studies. Furthermore, by assuming that operating income follows
the mean-reverting process, we are able to draw a much clearer distinction
between the industries through a better control over the process—by varying
the long-term mean and the speed of mean reversion, which correspond to
the industry profitability and stability, respectively.
To our knowledge, optimal capital structure has never been modelled dy-
namically outside of the scope of a geometric Brownian motion, presumably
due to the loss of homogeneity property. We assume the modified mean-
reverting process with volatility being proportional to the current earnings
level, and derive the ordinary differential equation used for pricing debt and
equity. Further, using contingent claims analysis and state pricing, we prove
that the process still possesses the homogeneity property, which implies that
the mechanics of the model is unaltered in time. The importance of this
result is expressed in the salient simplicity of the derivations in the dynamic
setting, which is the main reason behind the ubiquitous assumption of the ge-
ometric Brownian motion in the academic literature. After that we describe
what we call the optimal capital structure decision framework—comprised
of the system of boundary conditions and smooth-pasting conditions—and
formulate how the tuple that closes it is obtained, which fully determines the
114 S. Bjerrisgaard and D. Fedoryaev
capital structure choice of the owner-manager.
We conduct extensive numerical tests, focusing explicitly on the numeri-
cal methods applied in the procedure and pointing out inevitable shortcom-
ings, the main ones being the high dependence on the initial approximation
and the difficulty of dealing with the simultaneity of the system. Numerical
results could be split into three sets: comparing our model to the static mean-
reverting model developed in Sarkar and Zapatero (2003), benchmarking our
model against the conventional dynamic GBM-based model, and examining
general relationships under our own setup. Firstly, we find that the dynamic
mean-reverting model predicts lower optimal leverage and optimal coupon
than the static mean-reverting model does. The interpretation of this result
is as follows. In the static model the leverage is too high as equity holders
are forced to choose the capital structure given the infinite expected earnings
path and thus have to optimise over the long-run horizon, while if the relev-
ering option is granted, a more conservative leverage is chosen a priori, to
optimally balance the costs of debt financing with the tax benefits already in
the short term. Secondly, the dynamic model with mean reversion suggests
lower optimal leverage and higher restructuring frequency than the dynamic
GBM-based model does, which could be explained by the difference in ex-
pectations of equity holders regarding the future development of the cash
flow under the two processes. When earnings follow a geometric Brownian
motion, the trend always remains positive, while in case of mean reversion it
varies depending on the current earnings value. Therefore, equity holders in
the GBM-governed setup are excessively optimistic and thus set the optimal
leverage too high. Moreover, we emphasise that since the stationary variance
of the mean-reverting process is finite, as opposed to that of a geometric
Brownian motion, we need to hike the base case volatility in the former case
to make a comparison more correct. This reveals an even more pronounced
gap in the optimal leverage levels. Overall, the optimal leverage in our model
is found to be in the range 30–35%, given reasonable assumptions for base
case parameters, which brings it closer to the empirical estimates than the
Chapter 6. Conclusion 115
existing models. All relationships are studied through the prism of mean
reversion, and it is shown that the speed of earnings convergence and the
long-term mean value of earnings are indeed important parameters as their
impact on the key variables could be rather substantial.
Besides the above, we consider two additional extensions: first, the GBM-
based model is modified to take into account fixed operating costs, so that
the resulting cash flow can obtain negative values, and second, the mean-
reverting model is changed to incorporate a call premium paid on the debt
that is retired prematurely. We demonstrate that higher operating leverage
implies lower financial leverage, and find that when the firm is allowed to
have its accounts in the red, it will be less levered compared to an otherwise
identical firm whose earnings are instead mean-reverting but always positive.
Finally, it is shown that assuming debt being callable at par may lead to
understated leverage and too high restructuring frequency.
References
Abramowitz, M. and I.A. Stegun (1972). Handbook of Mathematical Func-
tions, Ninth Edition, Dover Publications Inc.
Acharya, V., J. Huang, M. Subrahmanyam, and R.K. Sundaram (2006).
When Does Strategic Debt Service Matter? Economic Theory, Vol. 29 (2),
363–378.
Alderson, M.J. and B.L. Betker (1995). Liquidation Versus Continuation:
Did Reorganized Firms Do the Right Thing? Ohio State University working
paper No. 9512.
Altunbas, Y., A. Karagiannis, M-H. Liu, and A. Tourani-Rad (2008). Mean
Reversion of Profitability: Evidence From the European-Listed Firms, Man-
agerial Finance, Vol. 34 (11), 799–815.
Ammann, M. and M. Genser (2004). A Testable Credit Risk Framework with
Optimal Bankruptcy, Taxes, and a Complex Capital Structure, University of
St. Gallen working paper series in finance No. 1.
Anderson, R.W. and S. Sundaresan (1996). Design and Valuation of Debt
Contracts, The Review of Financial Studies, Vol. 9 (1), 37–68.
Atluri, S.N., C-S. Liu, and C-L. Kuo (2009). A Modified Newton Method
116
Chapter 6. Conclusion 117
for Solving Non-Linear Algebraic Equations, Journal of Marine Science and
Technology, Vol. 17 (3), 238–247.
Back, K. and S.R. Pliska (1991). On the Fundamental Theorem of Asset
Pricing With an Infinite State Space, Journal of Mathematical Economics,
Vol. 20 (1), 1–18.
Bajaj, M., D.J. Denis, and A. Sarin (2004). Mean Reversion in Earnings
and the Use of E/P Multiples in Corporate Valuation, Journal of Applied
Finance, Vol. 14 (1), 4–10.
Bank, M. and J. Lawrenz (2005). Informational Asymmetry Between Man-
agers and Investors in the Optimal Capital Structure Decision, SSRN working
paper series.
Battig, R.J. and R.A. Jarrow (1999). The Second Fundamental Theorem of
Asset Pricing: A New Approach, The Review of Financial Studies, Vol. 12
(5), 1219–1235.
Bhattacharya, S. (1978). Project Valuation with Mean-Reverting Cash Flow
Streams, The Journal of Finance, Vol. 33 (5), 1317–1331.
Bichteler, K. (2002). Stochastic Integration With Jumps, Cambridge Univer-
sity Press.
Black, F. and J.C. Cox (1976). Valuing Corporate Securities: Some Effects
of Bond Indenture Provisions, The Journal of Finance, Vol. 31 (2), 351–367.
Black, F. and M. Scholes (1973). The Pricing of Options and Corporate Li-
abilities, The Journal of Political Economy, Vol. 81 (3), 637–654.
118 S. Bjerrisgaard and D. Fedoryaev
Bradley, M., G.A. Jarrell, and E.H. Kim (1984). On the Existence of an
Optimal Capital Structure: Theory and Evidence, The Journal of Finance,
Vol. 39 (3), 857–878.
Brennan, M.J. and E.S. Schwartz (1978). Corporate Income Taxes, Valua-
tion, and the Problem of Optimal Capital Structure, The Journal of Business,
Vol. 51 (1), 103–114.
Bris, A., A. Schwartz, and I. Welch (2005). Who Should Pay for Bankruptcy
Costs? The Journal of Legal Studies, Vol. 34 (2), 295–341.
Bris, A., I. Welch, and N. Zhu (2006). The Costs of Bankruptcy: Chapter 7
Liquidation versus Chapter 11 Reorganization, The Journal of Finance, Vol.
61 (3), 1253–1303.
Chen, Y. and K. Hammes (2003). Capital Structure Theories and Empirical
Results—A Panel Data Analysis, SSRN working paper series.
Christensen, P.O., C.R. Flor, D. Lando, and K.R. Miltersen (2002). Dynamic
Capital Structure with Callable Debt and Debt Renegotiations, SSRN work-
ing paper series.
Cotei, C. and J. Farhat (2009). The Trade-Off Theory and the Pecking Order
Theory: Are They Mutually Exclusive? North American Journal of Finance
and Banking Research, Vol. 3 (3), 1–16.
Cox, J.C., J.E. Ingersoll, and S.A. Ross (1985). An Intertemporal General
Equilibrium Model of Asset Prices, Econometrica, Vol. 53 (2), 363–384.
Dangl, T. and J. Zechner (2004). Credit Risk and Dynamic Capital Struc-
ture Choice, Journal of Financial Intermediation, Vol. 13 (2), 183–204.
Chapter 6. Conclusion 119
Davydenko, S.A. and I.A. Strebulaev (2007). Strategic Actions and Credit
Spreads: An Empirical Investigation, The Journal of Finance, Vol. 61 (2),
565–608.
Dixit, A.K. and R.S. Pindyck (1994). Investment under uncertainty, Prince-
ton University Press.
Duffie, D. and D. Lando (2001). Term Structures of Credit Spreads with
Incomplete Accounting Information, Econometrica, Vol. 69 (3), 633–664.
Duffie, D. and K.J. Singleton (1999). Modeling Term Structures of Default-
able Bonds, The Review of Financial Studies, Vol. 12 (4), 687–720.
Eberhart, A.C., W.T. Moore, and R.L. Roenfeldt (1990). Security Pricing
and Deviations from the Absolute Priority Rule in Bankruptcy Proceedings,
The Journal of Finance, Vol. 45 (5), 1457–1469.
Fama, E. and K.R. French (2000). Forecasting Profitability and Earnings,
Journal of Business, Vol. 73 (2), 161–175.
Fink, K.D. and J.H. Mathews (1999). Numerical Methods Using MATLAB,
Third Edition, Prentice Hall.
Firoozi, F. (2006). On the Martingale Property of Economic and Financial
Instruments, International Review of Economics and Finance, Vol. 15 (1),
87–96.
Fischer, E.O., R. Heinkel, and J. Zechner (1989a, March). Dynamic Capital
Structure Choice: Theory and Tests, The Journal of Finance, Vol. 44 (1),
19–40.
120 S. Bjerrisgaard and D. Fedoryaev
Fischer, E.O., R. Heinkel, and J. Zechner (1989b, December). Dynamic Re-
capitalization Policies and the Role of Call Premia and Issue Discounts, The
Journal of Financial and Quantitative Analysis, Vol. 24 (4), 427–446.
Flor, C.R. and J. Lester (2002). Debt Maturity, Callability, and Dynamic
Capital Structure, University of Southern Denmark working paper.
Francois, P. and E. Morellec (2004). Capital Structure and Asset Prices:
Some Effects of Bankruptcy Procedures, The Journal of Business, Vol. 77
(2), 387–411.
Frank, M.Z. and V.K. Goyal (2007). Trade-off and Pecking Order Theories
of Debt, SSRN working paper series.
Galai, D. and R.W. Masulis (1976). The Option Pricing Model and the Risk
Factor of Stock, Journal of Financial Economics, Vol. 3 (1-2), 53–81.
Gaud, P., M. Hoesli, and A. Bender (2007). Debt-Equity Choice in Europe,
International Review of Financial Analysis, Vol. 16 (3), 201–222.
Gilson, S.C., K. John, and L.H.P. Lang (1990). Troubled Debt Restructur-
ings: An Empirical Study of Private Reorganization of Firms in Default,
Journal of Financial Economics, Vol. 27 (2), 315–353.
Girsanov, I.V. (1960). On Transforming a Certain Class of Stochastic Pro-
cesses by Absolute Continuous Substitution of Measures, Theory of Proba-
bility and its Applications, Vol. 5 (3), 285–301.
Goldstein, R., N. Ju, and H.E. Leland (2001). An EBIT-Based Model of
Dynamic Capital Structure, The Journal of Business, Vol. 74 (4), 483–512.
Chapter 6. Conclusion 121
Gorbenko, A.S. and I.A. Strebulaev (2010). Temporary versus Permanent
Shocks: Explaining Corporate Financial Policies, The Review of Financial
Studies, Vol. 23 (7), 2591–2647.
Gordon, M.J. (1959). Dividends, Earnings, and Stock Prices, The Review of
Economics and Statistics, Vol. 41 (2), 99–105.
Graham, J.R. and C.R. Harvey (2001). The theory and practice of corporate
finance: evidence from the field, Journal of Financial Economics, Vol. 60
(2-3), 187–243.
Grinblatt, M. and S. Titman (2002). Financial Markets & Corporate Strat-
egy, Second Edition, McGraw–Hill.
Harrison, J.M. and S.R. Pliska (1981). Martingales and Stochastic Integrals
in the Theory of Continuous Trading, Stochastic Processes and their Appli-
cations, Vol. 11 (3).
Hennessy, C.A. and Y. Tserlukevich (2008). Taxation, Agency Conflicts,
and the Choice Between Callable and Convertible Debt, Journal of Eco-
nomic Theory, Vol. 143 (1), 374–404.
Jarrow, R.A., D. Lando, and S.M. Turnbull (1997). A Markov Model for the
Term Structure of Credit Risk Spreads, The review of Financial Studies, Vol.
10 (2), 481–523.
Jensen, M.C. (1986). Agency Costs of Free Cash Flow, Corporate Finance,
and Takeovers, The American Economic Review, Vol. 76 (2), 323–329.
Jensen, M.C. and W.H. Meckling (1976). Theory of the Firm: Managerial
122 S. Bjerrisgaard and D. Fedoryaev
Behavior, Agency Costs and Ownership Structure, Journal of Financial Eco-
nomics, Vol. 3 (4), 305–360.
Ju, N. and H. Ou-Yang (2006). Capital Structure, Debt Maturity, and
Stochastic Interest Rates, The Journal of Business, Vol. 79 (5), 2469–2502.
Ju, N., R. Parrino, A.M. Poteshman, and M.S. Weisbach (2005). Horses and
Rabbits? Trade-off Theory and Optimal Capital Structure, The Journal of
Finance and Quantitative Analysis, Vol. 40 (2), 259–281.
Kane, A., A.J. Marcus, and R.L. McDonald (1984). How Big is the Tax
Advantage to Debt? The Journal of Finance, Vol. 39 (3), 841–853.
Karatzas, I. and S.E. Shreve (1991). Brownian Motion and Stochastic Cal-
culus, Second Edition, Springer–Verlag.
Kim, D., D. Palia and A. Saunders (2003). The Long-Run Behaviour of Debt
and Equity Underwriting Spreads, NYU Stern working paper No. S-DRP-
03-03.
Kraus, A. and R.H. Litzenberger (1973). A State-Preference Model of Opti-
mal Financial Leverage, The Journal of Finance, Vol. 28 (4), 911–922.
Leary, M.T. and M.R. Roberts (2005). Do Firms Rebalance Their Capital
Structures? The Journal of Finance, Vol. 60 (6), 2575–2619.
Leland, H.E. (1994). Corporate Debt Value, Bond Covenants, and Optimal
Capital Structure, The Journal of Finance, Vol. 49 (4), 1213–1252.
Leland, H.E. (1998). Agency Costs, Risk Management, and Capital Struc-
ture, The Journal of Finance, Vol. 53 (4), 1213–1243.
Chapter 6. Conclusion 123
Leland, H.E. and K.B. Toft (1996). Optimal Capital Structure, Endogenous
Bankruptcy, and the Term Structure of Credit Spreads, The Journal of Fi-
nance, Vol. 51 (3), 987–1019.
Lipe, R. and R. Kormendi (1994). Mean Reversion in Annual Earnings and
Its Implications for Security Valuation, Review of Quantitative Finance and
Accounting, Vol. 4 (1), 27–46.
McDonald, R. and D. Siegel (1984). Option Pricing When the Underlying
Asset Earns a Below-Equilibrium Rate of Return: A Note, The Journal of
Finance, Vol. 39 (1), 261–265.
Mella-Barral, P. and W. Perraudin (1997). Strategic Debt Service, The Jour-
nal of Finance, Vol. 52 (2), 531–556.
Merton, R.C. (1974). On the Pricing of Corporate Debt: The Risk Structure
of Interest Rates, The Journal of Finance, Vol. 29 (2), 449–470.
Mitchell, K. (1991). The Call, Sinking Fund, and Term-to-Maturity Features
of Corporate Bonds: An Empirical Investigation, The Journal of Financial
and Quantitative Analysis, Vol. 26 (2), 201–222.
Modigliani, F. and M.H. Miller (1958). The Cost of Capital, Corporation Fi-
nance and the Theory of Investment, The American Economic Review, Vol.
48 (3), 261–297.
Modigliani, F. and M.H. Miller (1963). Corporate Income Taxes and the
Cost of Capital: A Correction, The American Economic Review, Vol. 53 (3),
433–443.
124 S. Bjerrisgaard and D. Fedoryaev
Myers, S.C. (1977). Determinants of Corporate Borrowing, Journal of Fi-
nancial Economics, Vol. 5 (2), 113–269.
Myers, S.C. (1984). The Capital Structure Puzzle, The Journal of Finance,
Vol. 39 (3), 575–592.
Myers, S.C. and N.S. Majluf (1984). Corporate Financing and Investment
Decisions When Firms Have Information That Investors Do Not Have, Jour-
nal of Financial Economics, Vol. 13 (2), 187–221.
Nocedal, J. and S.J. Wright (2006). Numerical Optimization, Second Edi-
tion, Springer.
Nordal, K.B. and R. Næs (2010). Mean Reversion in Profitability of Non-
Listed Firms, European Financial Management, Wiley Online Library, ac-
cessed June 2011.
Petrovic, N., S. Manson and J. Coakley (2009). Does Volatility Improve UK
Earnings Forecasts? Journal of Business Finance & Accounting, Vol. 36 (9),
1148–1179.
Raymar, S. (1991). A Model of Capital Structure when Earnings are Mean-
Reverting, The Journal of Financial and Quantitative Analysis, Vol. 26 (3),
327–344.
Revuz, D. and M. Yor (1999). Continuous Martingales and Brownian Mo-
tion, Third Edition, Springer–Verlag.
Sarkar, S. and F. Zapatero (2003). The Trade-off Model with Mean Revert-
ing Earnings: Theory and Empirical Tests, The Economic Journal, Vol. 113
(490), 834–860.
Chapter 6. Conclusion 125
Shleifer, A. and R.W. Vishny (1992). Liquidation Values and Debt Capac-
ity: A Market Equilibrium Approach, The Journal of Finance, Vol. 47 (4),
1343–1366.
Strebulaev, I.A. (2007). Do Tests of Capital Structure Theory Mean What
They Say? The Journal of Finance, Vol. 62 (4), 1747–1787.
Titman, S. and S. Tsyplakov (2006). A Dynamic Model of Optimal Capital
Structure, McCombs research paper series No. FIN-03-06.
Tsekrekos, A.E. (2010). The Effect of Mean Reversion on Entry and Exit
Decisions under Uncertainty, Journal of Economic Dynamics & Control, Vol.
34 (4), 725–742.
Vu, J.D. (1986). An Empirical Investigation of Calls of Non-Convertible
Bonds, Journal of Financial Economics, Vol. 16 (2), 235–265.
Wald, J.K. (1999). How Firm Characteristics Affect Capital Structure: An
International Comparison, The Journal of Financial Research, Vol. 22 (2),
161–187.