venture capital investment under uncertainty and asymmetric
Post on 18-Dec-2014
402 Views
Preview:
DESCRIPTION
TRANSCRIPT
Venture Capital Investment under Uncertainty and
Asymmetric Beliefs: A Continuous-Time, Stochastic
Principal-Agent Model∗
Yahel Giat†, Steven T. Hackman‡ and Ajay Subramanian§
October 21, 2008
Abstract
We develop a continuous-time, stochastic principal-agent model to investigate the effects ofasymmetric beliefs and agency conflicts on the characteristics and valuation of venture capitalprojects. In our model, a venture capitalist (VC) and an entrepreneur (EN) have imperfectinformation and differing beliefs about the intrinsic quality of a project in addition to havingasymmetric attitudes towards its risk. We characterize the equilibrium of the stochastic dynamicgame in which the VC’s dynamic investments, the EN’s effort choices, the dynamic compensationcontract between the VC and EN, and the project’s termination time are derived endogenously.Consistent with observed contractual structures, the equilibrium dynamic contracts feature bothequity-like and debt-like components, the staging of investment by the VC, the progressivevesting of the EN’s stake, and the presence of inter-temporal performance targets or milestonesthat must be realized for the project to continue. We numerically implement the model andcalibrate it to aggregate data on VC projects. Our analysis of the calibrated model shows thatEN optimism significantly enhances the value that venture capitalists derive. Entrepreneurialoptimism explains the discrepancy between the discount rates used by VCs (∼ 40%), whichadjust for optimistic payoff projections by ENs, and the average expected return of VC projects(∼ 15%). Our results show how the “real option” value of venture capital investment is affectedby the presence of agency conflicts and asymmetric beliefs.
Key Words: Dynamic Principal-Agent Models, Stochastic Dynamic Games, Incentive Con-tracts, Imperfect Information, Heterogeneous Beliefs.
∗We gratefully acknowledge financial support from the Kauffman Foundation under the “Roadmap for an En-trepreneurial Economy” initiative. We thank two anonymous referees and seminar audiences at the the 2007 Stan-ford Institute for Theoretical Economics (SITE) workshop on “Dynamic Financing and Investment”, the 2007 NorthAmerican Summer Meeting of the Econometric Society (Duke University, Durham, NC), the 2007 Real Options Con-ference (Berkeley, CA), the 2008 Chicago-Minnesota Theory Conference (University of Chicago), the Fields Institutefor Mathematical Sciences (Toronto, Canada), the University of Paris-Dauphine (Paris, France), and ESSEC (Paris,France) for valuable comments. The usual disclaimers apply.
†Department of Industrial Engineering and Management, Jerusalem College of Technology, Jerusalem, Israel‡School of Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, GA 30332§Department of Risk Management and Insurance, J. Mack College of Business, Georgia State University, Atlanta,
GA 30303
1 Introduction
Real-world productive activities are typically characterized by decentralized decision-making. The
agents who control various aspects of production often have different objectives, access to informa-
tion, and beliefs about project qualities (see Chen and Zenios, 2005, Gibbons, 2005). For example,
entrepreneurs (ENs) are often more optimistic about the success of their start-up firms than the
experienced venture capitalists (VCs) who provide capital (see Baker et al, 2005 for a recent sur-
vey). Further, VCs are usually well-diversified and less exposed to firm-specific risk than the less
diversified ENs who have significant human capital invested in their firms. As a result, VCs and
ENs have differing attitudes towards the risks of projects, which leads to agency conflicts that affect
the financing and operation of start-up firms. The interests of ENs (the “agents”) are aligned with
those of VCs (the “principals”) through incentive contracts that are affected by the VC’s and EN’s
heterogeneous beliefs about the outcomes of projects as well as their differing risk attitudes.
We develop a dynamic, stochastic principal-agent model of venture capital investment to ex-
amine the impact of asymmetric beliefs on the characteristics of venture capital projects—their
values, the structures of dynamic contracts between VCs and ENs, the durations of VC projects,
and the manner in which VC investment is staged over time. In our model, VCs and ENs have
asymmetric beliefs about the intrinsic qualities of projects as well as asymmetric attitudes towards
their risk. We characterize the equilibrium of the stochastic dynamic game in which the VC’s
dynamic investments, the EN’s effort choices, the dynamic compensation contract for the EN, and
the project’s termination time are derived endogenously. We calibrate the model parameters by
matching the return distributions of VC projects predicted by the model to their observed values in
the data. We show that the degree of EN optimism is significant enough to explain the discrepancy
between the discount rates used by VCs to value projects (∼ 40%), which adjust for optimistic
payoff projections by ENs, and the average expected return of VC projects (∼ 15%). EN optimism
is a key determinant of the durations and economic values of VC relationships, and could explain
features of observed contracts between VCs and ENs.
Model Overview: In our continuous-time stochastic model, a cash-constrained EN with a
project approaches a VC for funding. The project generates potential value through physical capital
1
investments by the VC and human capital (effort) investments by the EN. We model the evolution
of the project’s termination payoff at each date, which is the total payoff (present value of future
earnings) if the project is terminated at that date. The termination payoff evolves as a Gaussian
process and is contractible. The variance of the termination payoff process is the project’s intrinsic
risk, which remains invariant through time. The drift of the termination payoff process has two
components: a fixed, non-discretionary component that represents the project’s intrinsic quality,
and a discretionary component that is determined by the VC’s investment and the EN’s effort. The
discretionary component is observable, but non-verifiable and, therefore, non-contractible.
The VC and the EN have imperfect information about the project’s intrinsic quality and could
have differing, normally distributed priors. Their respective beliefs are, however, common knowl-
edge, that is, they “agree to disagree” about their respective mean assessments of project quality,
the difference of which represents the degree of asymmetry in beliefs. We consider the general sce-
nario in which the VC’s and EN’s mean assessments of project quality could differ from its true
mean. Further, the EN could be either optimistic or pessimistic relative to the VC. The common
variance of the VC’s and EN’s respective assessments of the project’s quality is the project’s tran-
sient risk. The transient risk is resolved over time as the VC and EN update their assessments of
the project’s quality based on observations of the project’s termination payoff.
The VC has linear preferences whereas the EN is risk-averse with CARA preferences. The VC
offers the EN a long-term contract that specifies her dynamic investment policy, the termination
time (a stopping time) of the project, and the EN’s payoff. The EN dynamically chooses his
effort to maximize his expected utility. The contractually specified payoffs of the VC and EN, the
investment policy, the EN’s effort policy, and the termination time are derived endogenously in
equilibrium of the dynamic game between the VC and EN.
The Equilibrium Contracts: We derive the incentive efficient dynamic contracts between the
VC and EN. Under an optimal contract, the change in the EN’s stake in the project or her promised
payoff (his “certainty equivalent” expected future utility) evolves as an Ito process. The change in
the EN’s stake has a performance-sensitive component that depends on the change in the project’s
termination payoff and a performance-invariant component that does not. The key contractual
parameters—the VC’s investments, the EN’s effort, and his compensation—are determined by the
2
EN’s pay-performance sensitivities, that is, the sensitivities of the change in the EN’s stake to
the change in the project’s termination payoff. Conditional on the project’s continuation, the VC’s
optimal investments and the EN’s pay-performance sensitivities are deterministic functions of time.
The performance-invariant component of the change in the EN’s stake over the period is, however,
stochastic and depends on the project’s termination payoff history through its effect on the VC’s
and EN’s updated assessments of the project’s intrinsic quality.
Consistent with observed contractual structures, (i) the VC’s payoff structure has “debt” and
“equity” components; (ii) the VC optimally stages her investment; (iii) the EN’s stake in the
project progressively vests over time; and (iv) the project is continued if and only if inter-temporal
milestones or performance targets are realized (see Gompers, 1995, Kaplan and Stromberg, 2003).
The Dynamics of Equilibrium Contracts: The time-paths of the VC’s investments and the
EN’s pay-performance sensitivities depend on the relative magnitudes of the degree of asymmetry
in beliefs and the costs of risk-sharing between the VC and EN. If the EN is pessimistic, then
the pay-performance sensitivities and investments increase over time. If the EN is “reasonably
optimistic,” i.e., the EN is optimistic, but the degree of his optimism is below a threshold relative
to the costs of risk-sharing, then the pay-performance sensitivities and investments decrease over
time. If, however, the EN is “exuberant,” i.e., the degree of EN optimism is above this threshold,
then the pay-performance sensitivities decrease over time and investments increase in early periods
and decrease in later periods. Hence, depending on the relative magnitudes of risk-sharing costs
and asymmetry in beliefs, the VC’s investment policy could become more aggressive over time,
less aggressive over time, or vary non-monotonically. The EN’s compensation could become either
more or less sensitive to performance over time.
The intuition for the above results hinges on the interplay among (i) the EN’s effort that is
positively (negatively) affected by his optimism (pessimism); (ii) the costs of risk-sharing due to
the EN’s risk aversion; and (iii) the complementary effects of the VC’s investment and the EN’s
effort on output. The passage of time lowers the degrees of optimism (or pessimism) as successive
project realizations cause the VC and the EN to revise their initial assessments of project quality.
Hence, the beneficial (detrimental) effects of optimism (pessimism) in mitigating the agency costs
of risk-sharing between the VC and the EN decline over time.
3
If the EN is optimistic, the VC can increase the performance-sensitive component of the EN’s
compensation because the EN overvalues this component. Hence, the EN’s pay-performance sen-
sitivity and effort are initially high. As the EN’s optimism declines over time, the positive effect
of optimism on the power of incentives that can be provided to the EN declines so that the EN’s
pay-performance sensitivity and effort decrease. (The opposite implications hold true when the EN
is pessimistic.) If the EN is optimistic, but his degree of optimism is below a threshold, the VC’s
investment also declines over time. Because investment and effort are complementary, the decline
in the power of incentives to the EN and, therefore, his effort over time causes the VC to also lower
her investment over time. If the EN’s optimism is above a threshold, the VC exploits the EN’s
exuberance by initially increasing her investment to compensate for the decrease in effort of the
EN. After a certain point in time when the VC’s investment attains its maximum, the decreasing
effort of the EN makes it optimal for the VC to also lower her investments.
The effects of risk on the investment path also depend on the degree of asymmetry in beliefs. If
the EN is either pessimistic or reasonably optimistic, the time path of optimal investments decreases
pointwise with the EN’s risk aversion, the project’s intrinsic risk, and its initial transient risk. If
the EN is exuberant, however, the investment path is, in general, non-monotonic and increases with
the EN’s risk aversion as well as the project’s intrinsic and transient risks in early periods, and
decreases in later periods. In contrast with traditional real options models in which all decisions are
made by monolithic agents (Dixit and Pindyck, 1994), the interaction between asymmetric beliefs
and agency conflicts could lead to a positive or negative relation between risk and investment.
Calibration and Numerical Analysis: We numerically implement the model and calibrate
the baseline values of its parameters, which include the average intrinsic quality of VC projects, the
degree of asymmetry in beliefs between VCs and ENs, the EN’s risk aversion, and his disutility of
effort. We estimate these parameters by matching the predicted distributions of “round by round”
returns of VC projects to their observed values reported in Cochrane (2005).
Consistent with anecdotal evidence, our “indirect inference” approach shows that ENs are,
indeed, significantly optimistic relative to VCs. (Recall that we do not assume that the EN is
optimistic a priori.) Because the VC exploits EN optimism through the provision of more powerful
incentives, EN optimism significantly enhances the value to the VC. Interestingly, however, EN
4
optimism lowers overall project value because the VC prolongs the project and over-invests to take
advantage of the EN’s optimism and enhance her own value at the expense of project value.
We examine how changes to the degree of EN optimism, the project’s intrinsic risk, and its
transient risk affect the project value and the VC’s stake. The project’s intrinsic and transient
risks have opposing effects on the “speed of learning” about project quality and, therefore, the rate
at which the degree of EN optimism declines over time. As a result, they have differing effects
on the project value and the VC’s stake: the project value and the VC’s stake decrease with the
project’s intrinsic risk, but vary non-monotonically with its transient risk.
Prior empirical and anecdotal literature documents that VCs use discount rates around 40%
to value projects even though the average expected return of VC projects is approximately 15%
(Cochrane, 2005). It has been suggested that higher discount rates could be a mechanism that VCs
use to “adjust” optimistic projections by ENs. Previous research, however, has not ascertained
whether EN optimism is, in fact, significant enough to generate such a large discrepancy between
VC discount rates and the average expected returns of VC projects. We define the implied discount
rate (IDR) as the rate at which the VC would discount the EN’s projections of the project’s payoffs
to conform to her own valuation of the project’s payoffs. The IDRs for a wide range of parameter
values predicted by the model lie between 30% and 50%, which is consistent with the range of
VC discount rates reported in prior empirical research (e.g. Sahlman, 1990, Cochrane, 2005). Our
study, therefore, confirms that entrepreneurial “optimism premia” are indeed high enough to justify
the discount rates used by VCs in reality.
Related Literature: Our study belongs to the growing body of literature that analyzes
dynamic principal-agent models. In a seminal study, Holmstrom and Milgrom (1987) present a
continuous-time principal-agent framework in which the principal and agent have CARA prefer-
ences and payoffs are normally distributed. They show that the optimal contract for the agent is
affine in the project’s performance. Schattler and Sung (1993) and Sung (1995) provide a rigorous
development of the first-order approach to the analysis of continuous-time principal-agent prob-
lems with exponential utility using martingale methods. Following Spear and Srivastava (1987),
a significant stream of the literature applies dynamic principal-agent models to study executive
compensation (Spear and Wang, 2005, Cvitanic et al, 2005, Cadenillas et al, 2006, Sannikov, 2007)
5
and financial contracting (DeMarzo and Fishman, 2006, Biais et al, 2007).
We contribute to this literature by developing and analyzing a dynamic principal-agent model
with imperfect public information and heterogeneous beliefs. The optimal dynamic contract reflects
the effects of Bayesian learning and the resultant dynamic variation of the degree of asymmetry
in beliefs in addition to the usual tradeoff between risk-sharing and incentives. Further, both the
principal and the agent take productive actions in our model.
In the specific context of venture capital, a strand of the literature investigates the importance
of staging in the mitigation of VC-EN agency conflicts. Using a deterministic model, Neher (1999)
shows that staging is essential to overcome the hold-up problem. As in Neher (1999), the manner
in which VC investment is staged over time as well as the number of stages are determined en-
dogenously in our framework. Our framework is, however, stochastic and incorporates asymmetric
beliefs between the VC and EN.1
Another strand of the literature on venture capital analyzes the features of the optimal contracts
that emerge in “double-sided” two-period moral hazard models in which the VC and EN exert effort
(Casamatta, 2003, Cornelli and Yosha, 2003, Schmidt, 2003, Repullo and Suarez, 2004). We too
develop a model in which the VC and EN take value-enhancing actions. Similar to these studies,
the optimal contracts predicted by our analysis have “debt” and “equity” features consistent with
observed contractual structures. Our study focuses on the effects of asymmetric beliefs on the
characteristics of VC-EN relationships in a dynamic principal-agent model.
The principal-agent paradigm is also applied to various operations management contexts. Fol-
lowing the early work of Atkinson (1979), recent studies examine the inefficiencies arising from
either hidden information (for example, Cachon and Lariviere, 2001, Ha, 2001) or hidden action
(for example, Lal and Srinivasan, 1993, Plambeck and Zenios, 2000, 2003, Chen, 2005) in supply
chain contracting. We contribute to this line of research by developing and analyzing a dynamic
principal-agent model with heterogeneous beliefs, and in which both the principal and agent make
value-enhancing decisions over time. Our framework could potentially be applied in supply chain
contexts as well as scenarios such as venture capital investment and R&D in which heterogeneous1Kockesen and Ozerturk (2004) argue that some sort of EN “lock-in” is essential for staged financing to occur.
Egli et al (2006) argue that staging can be used to build an EN’s credit rating. Berk et al (2004) develop an R&Dmodel with a single, monolithic agent in which staging is exogenous.
6
beliefs and agency conflicts play important roles.
In summary, we contribute to the literature by developing and analyzing a dynamic, stochastic
principal-agent model of venture capital investment. The model is parsimonious, yet realistic
enough to be taken to the data to yield quantitative assessments of the effects of the salient aspects
of VC projects, namely, risky payoffs, agency conflicts, uncertainty about project quality and
asymmetric beliefs. The tractability of the model, coupled with the fact that it is able to match
disparate empirical data on the payoff distributions of projects as well as the discount rates used
to value them, suggests that it could be useful as a tool to value risky ventures.
2 The Model
We develop a continuous time framework with time horizon [0, T ]. At date zero, a cash-constrained
entrepreneur (hereafter, EN) with a project approaches a venture capitalist (hereafter, VC) for
funding. The project generates value through physical capital investments by the VC and human
capital (effort) investments by the EN. Both the VC and the EN have imperfect information about
the project and differ, in general, in their initial assessments of the project’s quality.
If the VC agrees to invest in the project, she offers the EN a long-term contract that describes
her subsequent investments in the project, the EN’s compensation, and the termination time of the
relationship. The VC’s investments are made continuously over time. The termination time could
be a random stopping time.
The key state variable in the model is the project’s termination payoff Vt, which is the total
payoff if the VC-EN relationship is terminated at date t. The termination payoff is the only
economic variable that is contractible. For simplicity, we assume the project does not generate any
intermediate cash flows so that all payoffs occur upon termination.
2.1 The Termination Payoff Process
All stochastic processes are defined on an underlying probability space (Ω,F , P ) on which is defined
a standard Brownian motion B. The initial termination payoff of the project is V0. The incremental
termination payoff, that is, the change in the termination payoff over the infinitesimal period
7
[t, t + dt], dVt, is the sum of a base output, a Gaussian process that is unaffected by the actions
of the VC and EN, and a discretionary output, a deterministic component that depends on the
physical capital investments by the VC and the human capital (effort) by the EN. It is given by
dVt =
base output︷ ︸︸ ︷(Θ− lt)dt + sdBt +
discretionary output︷ ︸︸ ︷Acα
t ηβt dt. (1)
The first component, Θ, of the base output represents the project’s core output growth rate,
which we hereafter refer to as the project’s intrinsic quality. The VC and EN have imperfect
information about Θ and could also differ in their beliefs about its value. The second component of
the base change, lt, represents “operating costs,” which could include wages to salaried employees,
depreciation expenses, decline in revenues due to increased competition, fixed costs arising from
increases in the scale of the project, etcetera. These costs are deterministic and increasing over
time, which ensure that termination occurs in finite time almost surely. The third component of the
base change, sdBt, where s > 0 is a constant, represents the “intrinsic” component of the project’s
risk in period [t, t + dt]. It is the component of the project’s risk that remains invariant over time,
and is independent of Θ.
The discretionary output in period [t, t+dt] is a direct result of the VC’s capital investment rate
ct and the EN’s effort ηt, and is described by a Cobb-Douglas production function. The discretionary
output is observable to the VC and the EN. However, as in the literature on incomplete contracting
(see Chapter 6 of Laffont and Martimort, 2002), the discretionary output is non-verifiable and,
therefore, non-contractible. Because the discretionary output is non-contractible, the EN must be
indirectly provided with appropriate incentives to exert effort through her explicit contract with
the VC that can only be contingent on the termination payoff process.
The uncertainty in the value of Θ is the project’s transient risk. The VC’s and EN’s initial priors
on Θ are normally distributed with Θ ∼ N(µV C0 , σ2
0) and Θ ∼ N(µEN0 , σ2
0), respectively. Their
respective beliefs are, however, common knowledge, that is, they agree to disagree (see Morris,
1995, Allen and Gale, 1999). Because the equilibrium does not depend on how the EN’s and VC’s
mean assessments of project quality relate to its true mean, we make no assumptions about the
true mean of the project quality distribution. We consider the most general scenario in which the
8
EN could be optimistic or pessimistic relative to the VC, that is, µEN0 could be greater or less than
µV C0 . While the VC and EN disagree on the mean of the project’s intrinsic quality, they agree on
its variance, σ20.
2
The transient risk is resolved over time as the VC and the EN update their priors on Θ in a
Bayesian manner based on observations of the project’s performance. Define
dξt := dVt − (Φ(ct, ηt)− lt)dt = Θdt + sdBt. (2)
It follows from well-known formulae (Oksendal 2003) that the posterior distribution on Θ for each
date t ≥ 0 is N(µ`t, σ
2t ), ` = V C, EN , where
σ2t =
s2σ20
s2 + tσ20
, µ`t =
s2µ`0 + σ2
0ξt
s2 + tσ20
, ` = V C, EN. (3)
Note that the σt tend to zero. Let
∆t := µENt − µV C
t =s2∆0
s2 + tσ20
=∆0
σ20
σ2t (4)
denote the degree of asymmetry in beliefs at date t. It is resolved deterministically and monoton-
ically over time, and its absolute value, | ∆t |, also declines over time. Consequently, if the EN is
more optimistic (pessimistic) than the VC, the degree of optimism (pessimism) declines.
2.2 VC-EN Interaction
The contract between the VC and the EN describes the VC’s capital investments over time, the
EN’s effort, the termination date, and the EN’s payoff upon termination. The termination time
is, in general, a random stopping time that is contingent on the project’s performance history. We
follow the traditional principal-agent literature by having the contract also specify the EN’s effort
and requiring that the contract be incentive compatible with respect to the specified effort of the2The literature on behavioral economics (see Baker et al, 2005) distinguishes between optimism and overconfidence.
The EN is “optimistic” if his assessment of the mean (the first moment) of the project quality distribution is higherthan that of the VC, while he is “overconfident” if his assessment of the variance (the second moment) of the projectquality distribution is lower than that of the VC. In the terminology of the behavioral economics literature, therefore,the EN could be optimistic, but not overconfident in our framework.
9
EN (see Holmstrom and Milgrom, 1987).
Let Ft denote the information filtration generated by the history of termination payoffs, the
VC’s investments and the project’s discretionary outputs. A contract is described by the quadruple
(Pτ , c, η, τ), where c and η are Ft-adapted stochastic processes, τ is an Ft-stopping time, and
Pτ is a nonnegative Fτ-measurable random variable. Pτ is the EN’s contractually promised payoff
and Vτ −Pτ is the VC’s payoff at the termination time τ of the contractual relationship. In period
[t, t + dt], the VC’s investment rate is ct and the EN’s effort is ηt.
The VC offers the EN a long-term contract at date zero. The VC is risk-neutral whereas the
EN is risk-averse with inter-temporal CARA preferences described by a negative exponential utility
function. Their discount rates are equal and set to zero to simplify the notation. We extend the
model to incorporate nonzero discount rates when we calibrate it to the data in Section 6.
The EN’s expected utility at date zero from a contract (Pτ , c, η, τ) is
−EEN0
[exp
− λ
(Pτ −
∫ τ
tkηγ
t dt)]
. (5)
In (5), EEN0 denotes the expectation with respect to the EN’s beliefs at date zero and the parameter
λ ≥ 0 characterizes the EN’s risk aversion. The EN’s disutility from effort in period [t, t + dt] is
given by kηγt dt with k > 0, γ > 0. For future reference in the derivation of the equilibrium, we
follow Holmstrom and Milgrom (1987) by defining the EN’s certainty equivalent expected future
utility, Pt, from the contract at any date t as
exp(−λPt) := EENt exp
(− λ
(Pτ −
∫ τ
tkηγ
udu))
, (6)
where the notation EENt denotes the EN’s expectation conditioned on the information available at
date t, that is, the σ-field Ft. Note that the EN’s certainty equivalent future expected utility at the
contractual termination date τ is his contractually promised terminal payoff Pτ . For expositional
convenience, we hereafter refer to the EN’s certainty equivalent expected future utility process
Pt, t ≥ 0 as his promised payoff process.
The allocation of bargaining power between the VC and the EN is determined by the certainty
equivalent reservation utility or promised payoff that the EN must be guaranteed at date zero. We
10
allow for all possible allocations of bargaining power that are indexed by different values of P0.
A contract (Pτ , c, η, τ) is feasible if and only if it is incentive compatible for the EN with respect
to his effort choices, that is, given the terminal payoff, Pτ , the VC’s investment policy, c, and
the termination time τ , it is optimal for the EN to exert effort described by the process η. The
risk-neutral VC’s optimal contract choice is a feasible contract that maximizes her expected payoff
net of her investments, i.e., a feasible contract (Pτ , c, η, τ) is optimal if and only if it solves
(Pτ , c, η, τ) = arg max(Pτ ′ ,c′,η′,τ ′)
EV C0
[Vτ ′ − Pτ ′ −
∫ τ ′
tc′tdt
], (7)
where EV C0 denotes the expectation with respect to the VC’s beliefs at date zero and the maxi-
mization is over feasible contracts.
3 The Equilibrium
We assume the following condition on the parameters for the remainder of the paper:
Assumption 1 (1− α)γ/β > 2.
This condition implies that the EN faces decreasing returns to scale from the provision of effort.
Further, the EN’s disutility from his effort is sufficiently pronounced relative to his positive contri-
bution to output that an equilibrium contract between the VC and the EN exists.
3.1 Structure of Optimal Contract
The following two theorems characterize the optimal contract. Proofs are provided in Appendix A.
Theorem 1 (The EN’s Promised Payoff Process)
The EN’s promised payoff evolves as dPt = atdt + btdVt, where the contractual parameters at ∈IR, bt ∈ IR++ are Ft-progressively measurable.
The parameter bt is the EN’s pay-performance sensitivity. It represents the sensitivity of the change
in the promised payoff to performance during the infinitesimal period [t, t + dt]. The parameter at
is the EN’s performance-invariant compensation. It determines the component of the change in the
promised payoff that does not depend on performance during the infinitesimal period [t, t + dt].
11
In light of Theorem 1, a contract is completely specified by the performance-invariant compen-
sation and pay-performance sensitivity parameters, at, bt, the VC’s investment rate, ct, the EN’s
effort, ηt, at each time t, and the termination time τ .
3.2 Existence and Characterization of Equilibrium
We briefly outline the arguments involved in the derivation of the optimal contract, which is formally
characterized in Theorem 2 below and proved in Appendix A.
Fix date t ≥ 0. The derivation proceeds in four steps:
Step 1. The EN’s incentive compatible effort. For a given EN’s pay-performance sensitivity bt and
the VC’s investment rate ct at date t, we show that the EN’s incentive compatible effort is
η(bt, ct) :=(Aβcα
t bt
γk
) 1γ−β
. (8)
Step 2. The EN’s performance-invariant compensation. The VC optimally chooses her investment
rate ct and the EN’s pay-performance sensitivity bt incorporating the EN’s incentive compat-
ible effort given by (8). The performance-invariant compensation parameter at is chosen to
satisfy the “promise-keeping” constraint, that is, the EN’s promised payoff is actually deliv-
ered by the contract. The promise-keeping constraint pins down the contactual parameter at
as a function of the other two contract parameters bt and ct. In particular,
at = at(bt, ct) := 0.5λs2b2t + kη(bt, ct)γ − bt
(Acα
t η(bt, ct)β − lt + µEN
t
). (9)
Step 3. The optimal investment and pay performance sensitivity. Incorporating the EN’s incentive
compatible effort (8) and the functional form for at in (9), the change in the VC’s continuation
value (her expected future payoffs) at date t is
CVt = Λt(bt, ct)dt, where (10)
Λt(bt, ct) := ∆tb− 0.5λs2b2 + φ(b)cα γγ−β − c + µV C − lt. (11)
12
In (11), ∆t is the degree of asymmetry of beliefs at date t, defined in (4), and
φ(b) :=
Aγ
γ−β
(1k
) βγ−β
((βbγ
) βγ−β
(1− βb
γ
)), if 0 ≤ b ≤ γ/β,
0, otherwise.
(12)
The function Λt(bt, ct) is the rate-of-change of the VC’s continuation value; hereafter, we
shall refer to it simply as the continuation rate. The VC chooses the investment rate ct and
the EN’s pay-performance sensitivity bt to maximize the continuation rate. Assumption 1
guarantees a unique solution b∗t , c∗t to this maximization problem. In particular, it implies
that Λt(bt, ·) is strictly concave in ct, since the exponent on ct is less than one. Consequently,
given the pay-performance sensitivity, there is a unique investment rate. We show that
c∗t = c(b∗t ) :=( αγ
γ − β
) γ−β(1−α)γ−β φ(bt)
γ−β(1−α)γ−β , (13)
b∗t = arg max0<bt
Λt(bt, c(bt)), where (14)
Λt(bt, c(bt)) :=[∆tbt − 0.5λs2b2
t +γ − β − αγ
αγc(bt)
]+ (µV C
t − lt). (15)
We refer to the function c(·) as the optimal investment function. We discuss the properties
of the optimal investment function, which plays a central role in our analysis, in Section 3.4.
Step 4. Determination of the optimal termination time. The optimal termination time of the contract
is the solution to the optimal stopping problem
τ∗ = arg maxτ≤T
EV C0
∫ τ
0Λt(b∗t , c
∗t )dt, (16)
where the maximization is over all Ft-stopping times τ ≤ T .
Theorem 2 (Characterization of Equilibrium)
(a) Conditional on the project not being terminated prior to date t ∈ [0, T ]:
– The EN’s pay-performance sensitivity parameter, b∗t , solves (14).
– The VC’s equilibrium investment rate is c∗t = c(b∗t ), where c(·) is defined in (13).
13
– The EN’s performance-invariant compensation parameter is a∗t := at(b∗t , c∗t ), where at(·, ·)is defined in (9).
– The EN’s effort level is η∗t := η(b∗t , c∗t ), where η(·, ·) is defined in (8).
(b) The termination time of the relationship solves the optimal stopping problem (16).
3.3 The VC’s Controllable Rate Function
Let
Ft(b) := ∆tb− 0.5λs2b2 +γ − β − αγ
αγc(b) (17)
denote the “controllable” portion of the continuation rate; we hereafter refer to it as the VC’s
controllable rate function. As summarized in Theorem 2, the equilibrium contract at date t is
determined by b∗t , the solution to (14). An examination of (15) shows that b∗t is also the solution to
b∗t = arg max0<b
Ft(b). (18)
By Theorem 2 and (18), the EN’s pay-performance sensitivity b∗t , the VC’s investment rate c∗t ,
and the EN’s effort η∗t are deterministic functions of time (conditional on the project’s continuation).
The performance-invariant compensation parameter a∗t is, however, stochastic and depends, in
particular, on the EN’s current mean assessment µENt of the project’s intrinsic quality. The proof
of the theorem shows that this parameter adjusts stochastically to ensure that the EN’s promise
keeping constraints are satisfied at each date and state.
The equilibrium contract critically depends on the VC’s controllable rate function, Ft(b). This
function consists of three components:
• Economic rent (cost) from the EN’s optimism (pessimism). When ∆t > 0, the term, ∆tb,
reflects the rents that the VC extracts from the EN by exploiting his optimism about the
project’s intrinsic quality. When ∆t < 0, ∆tb is the cost that the VC must bear to compensate
the EN for his pessimism about the project’s intrinsic quality.
• Cost of risk. The term, 12λs2b2, reflects the VC’s costs of risk-sharing with the risk-averse
EN. We refer to λs2 as the price of risk ; hereafter, we denote it by p.
14
• Return on investment. The “return on investment” term, γ−β−αγαγ c(b), reflects the VC’s
expected return as a result of her investment and the EN’s effort.
The interplay among these three “forces” determines the equilibrium dynamics.
3.4 Uniqueness and Stability of Equilibrium
The characteristics of the contract depend on the optimal investment function c(·) given in (13).
The following properties of c(·) play a central role in our subsequent analysis (see Figure 1).
Proposition 1
(a) The function c(·) is strictly positive and strongly unimodal3 on [0, γβ ], satisfies c(0) = c( γ
β ) = 0,
and achieves its maximum at b = 1.
(b) The function c(·) is strictly concave on [0, bM ] and strictly convex on [bM , γβ ], where bM ∈ (1, γ
β )
is the unique minimum of the function c′(·).
Proof. The proof of this, and all subsequent results in the paper, are provided in Appendix B.
The intuition for the non-monotonicity of the function is that an increase in the agent’s pay-
performance sensitivity affects the principal’s investment in two distinct but opposite ways. On
the positive side, the agent increases his effort. Because investment and effort are complementary,
the increase in the agent’s effort provides an incentive for the principal to increase her investment.
On the negative side, since the agent’s disutility of effort increases, the principal’s cost to maintain
the agent’s participation also increases. For lower values of the pay-performance sensitivity, the
complementarity of investment and effort causes the benefits of increased output to dominate.
Hence, the principal finds it beneficial to increase her investment. However, beyond a threshold
level of pay-performance sensitivity, the costs of inducing high effort from the agent are so high
that the principal lowers her investment. In other words, it is optimal for the principal to allow
output to be dominated by the agent’s effort.
The ratio of the absolute value of the initial degree of asymmetry of beliefs to the price of risk,
namely, | ∆0 | /p, provides an a priori bound on the equilibrium pay-performance sensitivity b∗t .3A function f(·) is strongly unimodal on the interval [a, b], a < b, if there exists an x∗ ∈ (a, b) such that f(·) is
increasing on [a, x∗] and f(·) is decreasing on [x∗, b]. Obviously, the value x∗ maximizes f(·) on [a, b].
15
Figure 1: Optimal investment function
6
- b
c(b)
1
qc′(1) = 0
γβ
c(0) = c( γβ
) = 0 bM
.
...........................
.........................
.......................
......................
....................
...................
.................
................
.............................................................................. ........ ....... ....... ....... ....... ...... ....... ....... ....... ....... ........ ........ ........ ........ ......... ........ ......... ........ .......... ........ ........... ........ ....................... ......... ......... ....................... ........ ........... ......... .......... ......... ......... ........ ......... ........
................
............... ............... .............. ............. ...... ....... .........
c′′(bM ) = 0q¾ convex -
¾ c o n c a v e-
Proposition 2
An optimal solution to (18) is always less than or equal to max∆0p , 1 if ∆0 ≥ 0 and is less than
1 if ∆0 < 0.
In our subsequent analysis we assume that the initial degree of asymmetry in beliefs, ∆0, is below
a threshold relative to the price of risk, p. The assumption ensures that the equilibrium is stable
and the contractual parameters are continuous functions of the primitives of the model.
Assumption 2 ∆0/p ≤ bM .
(The parameter bM above is defined in Proposition 1.) It follows immediately from Proposition
2 and Assumption 2 that a solution to (18) must lie in the interval [0, bM ). By Proposition 1,
the optimal investment function c(·) is strictly concave on the interval [0, bM ]. It follows from
(17) that the VC’s controllable rate function, Ft(b), is also strictly concave on [0, bM ] and hence
strongly unimodal. Consequently, there exists a unique solution b∗t to (18). Moreover, it must also
be positive, since the proof of Proposition 1 shows that the marginal optimal investment c′(0) is
infinite. We summarize these observations with the following proposition.
Proposition 3
Under Assumptions 1 and 2, the function Ft(·) is strictly concave on [0, bM ]. Further, the solution
to (18) is strictly positive and less than bM .
4 Equilibrium Dynamics
We investigate the dynamics of the EN’s compensation, his effort, and the VC’s investment condi-
tional on continuation of the project. Since the degree of asymmetry in beliefs, ∆t, and variance,
σ2t , are deterministic functions of time (see (4)), it follows from Theorem 2 and Proposition 1 that
16
the equilibrium values for the pay-performance sensitivity, investment and effort at each point in
time (conditional upon continuation) are also deterministic. The only component of the contract
that is stochastic and is adjusted based on realizations of the termination payoff Vt of the project
is the performance-invariant compensation parameter a∗t .
Let
Ft(b) = F (b) := −0.5λs2b2 +γ − β − αγ
αγc(b) (19)
be the principal’s controllable rate function in the benchmark scenario in which beliefs are sym-
metric. Since F (.) is time-independent, the agent’s equilibrium pay-performance sensitivities, the
principal’s investments and the agent’s effort are all constant. Let b∗p, c∗p and η∗p denote their values.
It follows from (4) and (17) that the VC’s controllable rate function can be expressed as
Ft(b) =∆0
σ20
σ2t b + F (b), (20)
where F (.) is defined in (19). Since σt → 0, it follows from Berge’s Theorem of the Maximum
that b∗t → b∗p, and thus (c∗t , η∗t ) → (c∗p, η∗p) by continuity where (b∗p, c∗p, η∗p) are the equilibrium pay-
performance sensitivity, investment, and effort in the benchmark scenario with symmetric beliefs.
We now describe the manner in which these economic variables converge to their asymptotic values.
Theorem 3 (The Dynamics of the Equilibrium—Optimistic EN)
Suppose that the EN is more optimistic than the VC so that ∆0 ≥ 0.
(a) The EN’s pay-performance sensitivity b∗t decreases monotonically with t and approaches b∗p
as t →∞.
(b) The value
t∗ := (∆0
p− 1)
s2
σ20
=∆0
λσ20
− s2
σ20
(21)
is the point in time at which the EN’s pay-performance sensitivity and effort, and the VC’s
investment rate equal their values in the “no agency” benchmark scenario. (This interpreta-
tion of t∗ only applies if t∗ ≥ 0, which holds if and only if the initial degree of asymmetry of
beliefs is at least as large as the price of risk.)
17
(c) The EN’s pay-performance sensitivity b∗t exceeds 1 if t < t∗, equals 1 at t = t∗, and less than
1 if t > t∗.
(d) The VC’s investment rate c∗t increases until time t∗ and then decreases monotonically towards
c∗p as t →∞.
(e) For t ≥ t∗, η∗t decreases monotonically towards η∗p as t →∞ .
Note that if ∆0 ≤ p so that t∗ ≤ 0, then the EN’s pay-performance sensitivity, his effort, and the
VC’s investment rate all decrease monotonically over time. Figures 2 and 3 illustrate the results of
Theorem 3. They describe the trajectories of the EN’s pay-performance sensitivity and the VC’s
investment for two different values of the initial degree of asymmetry in beliefs, ∆0.
The intuition for the results of Theorem 3 hinges on the interplay among the positive effect
of optimism on the EN’s effort, the costs of risk-sharing due to the EN’s risk aversion that are
negatively affected by the project’s intrinsic risk, and the complementary effects of investment
and effort on output. The passage of time lowers the degree of optimism as successive project
realizations cause the VC and the EN to revise their initial assessments of project quality.
If the EN is optimistic, the VC can increase the performance-sensitive component of the EN’s
compensation because the EN overvalues this component. Hence, the EN’s pay-performance sen-
sitivity and effort are initially high. The negative effect of the evolution of time on the EN’s
optimism, however, causes the EN’s pay-performance sensitivity and effort to decline over time.
Due to the previously discussed non-monotonic relation between the VC’s investment and the EN’s
pay-performance sensitivity, the VC’s investment initially increases to “compensate” for the de-
crease in effort of the EN. After a certain point in time when investment attains its maximum, the
decreasing effort of the EN makes it optimal for the VC to also lower her capital investments.
Theorem 4 (The Dynamics of the Equilibrium—Pessimistic EN)
Suppose that the EN is more pessimistic than the VC so that ∆0 < 0.
(a) The EN’s pay-performance sensitivity b∗t increases monotonically with t and approaches b∗p as
t →∞.
(b) The VC’s investment rate c∗t increases monotonically towards c∗p as t →∞.
18
6
- t
b∗t
0 2 4 6 8 10 12 14
....................
............................................ ........... .............. ................. ............ ................ .................... ............. ................ ................... ............. ................ .................. ............. ............... .................. .............. ............... ................. ..................... ........................ ............. ............... .................. ............................................. ............. ............... ................. ..................... ........................ ............................................. ............. ............... ................. .............................................
∆0 < 0
b∗p∆0 = 0
.............
.............
......................
....................
......... ............ ............... .................. ............. ................. .................... ............. ................ ................... ............. ................ .................. ............. ................ .................. .............. ................ ................. .............. ............... ................. ...................... ........................ ..................... ........................ ............. ............... ................. ............................................. ...................... ........................ ............. ............... .................. .............................................
∆0 > 0
Figure 2: Possible equilibrium pay-performance sensitivity paths
(c) The EN’s effort η∗t increases monotonically towards η∗p as t →∞ .
If the EN is pessimistic, he under-values the performance-sensitive portion of his compensation
relative to the VC. Hence, the power of incentives that can be provided to the EN is initially
low so that his pay-performance sensitivity and effort as well as the VC’s investment are initially
low. With the evolution of time, the degree of pessimism declines, which has a positive effect on
the power of incentives to the EN so that his pay-performance sensitivity, effort, and the VC’s
investment all increase.
Theorems 3 and 4 describe the paths of the EN’s pay-performance sensitivity and effort, and
the VC’s investment rate conditional on the project’s continuation. Depending on the relationship
between the degree of asymmetry in beliefs and the price of risk, it follows from the theorems
that the VC’s investments until termination (these are the investments that are actually observed
because there is no investment after termination) could either increase, decrease, or vary non-
monotonically (initially increase and then decrease).
4.1 Sensitivity of Equilibrium Dynamics
In light of Theorems 3 and 4, the manner in which the equilibrium dynamics are affected by changes
in the underlying parameters critically depends on the initial value of the degree of asymmetry in
beliefs ∆0. In what follows the EN is said to be pessimistic if ∆0 < 0, reasonably optimistic if
∆0 ∈ [0, p) and exuberant if ∆0 ∈ (p, pbM ], where p = λs2 is the price of risk. (Assumption 2
guarantees that ∆0 ≤ pbM .)
19
6
- t
c∗t
2 4 6 8 10 12 14 16
c(1) “no agency” investment level
.....................
............................................. ........... .............. ................. ............ ................ .................... ............. ................ ................... ............. ................ .................. ............. ............... .................. .............. ............... ................. ............. ............... ................. ...................... ....................... ............ ............... ................... ........................ ..................... ...................... ........................ ............ ............... .................. ............................................. .............................................
∆0 < 0
c∗p∆0 = 0
...............
.....
...................
.................. ......... ............ .............. ................ ............. ................ .................... ............. ................ ................... ............. ................ .................. .............. ................ .................. .............. ................ ................. .............. ............... ................. ..................... ........................ .............. ............... ................. ............................................. ..................... ......................... ...................... ........................ ..................... ........................ .............................................
∆0 < p
.
.................................
..............................
............................
.........................
......................
...................
...................................................... .................. ..... ....... .......... ............ ............... ......... .......... .............. .................. ................ ............. .......... ....... ...... ......................... ........................ ........................ ......................... ....................... ......................... ....................... ......................... ....................... ......................... ....................... ........................ ...................... ......................... .............................................. ..................... ......................... ..............................................
∆0 > p
Figure 3: Possible equilibrium investment paths
The following theorem characterizes the effects of the EN’s risk aversion, λ, the initial transient
risk, σ20, and the intrinsic risk, s2, on the equilibrium dynamics.
Theorem 5 (Effects of Intrinsic Risk, Transient Risk and EN’s Risk Aversion)
(a) If the EN is pessimistic, the paths of the EN’s pay-performance sensitivity and the VC’s
investment are pointwise decreasing in the EN’s risk aversion, pointwise increasing in the
initial transient risk, and pointwise decreasing in the intrinsic risk.
(b) If the EN is optimistic, the path of the EN’s pay-performance sensitivity is pointwise decreas-
ing in the EN’s risk aversion and the initial transient risk.
(c) If the EN is reasonably optimistic, then the path of the VC’s equilibrium investment is
pointwise decreasing in the EN’s risk aversion and the initial transient risk.
(d) If the EN is exuberant, the VC’s investment path changes as depicted in Figure 4 as a result
of a change in the EN’s risk aversion and the initial transient risk. More precisely, let λ1 < λ2
and σ1 < σ2 be two possible values of the EN’s risk aversion and the initial transient risk,
respectively. There exist t∗(λ1, λ2) and t∗∗(σ1, σ2) such that the VC’s investments when the
EN’s risk aversion is λ1 (the initial transient risk is σ1) are higher than her investments when
the EN’s risk aversion is λ2 (the initial transient risk is σ2) for t < t∗(λ1, λ2) (t < t∗∗(σ1, σ2))
and lower for t > t∗(λ1, λ2) (t > t∗∗(σ1, σ2)).
(e) If ∆0 ≤ 4p, parts (b)-(d) hold for the intrinsic risk. If ∆0 > 4p, the effects of intrinsic risk on
20
the pay-performance sensitivity and investment paths is ambiguous.4
Figure 4 demonstrates that the path of equilibrium investment converges to different limiting values
depending on the EN’s risk aversion.
The EN’s pay-performance sensitivity, b∗t , declines with his risk aversion because an increase in
the EN’s risk aversion increases the costs of risk-sharing. An increase in the transient risk lowers
the degree of optimism or pessimism at each date because the “signal to noise ratio” is increased
so that the VC and EN “learn faster.” An increase in the intrinsic risk increases the costs of risk-
sharing and also increases the degree of optimism or pessimism at each date because the “signal to
noise ratio” decreases so that the VC and EN “learn more slowly.”
When the EN is pessimistic, the pointwise decline in the degree of pessimism with the transient
risk increases the power of incentives to the EN so that the EN’s pay-performance sensitivity and
the VC’s investment increase at each date. On the other hand, the pointwise increase in the degree
of pessimism and the costs of risk-sharing with the intrinsic risk decreases the power of incentives
to the EN so that his pay-performance sensitivity and the VC’s investment decrease at each date.
When the EN is optimistic, the decline of the degree of optimism with the initial transient risk
causes the economic rents to the VC in each period from the EN’s optimism to be lowered relative
to the costs of risk-sharing. Hence, the EN’ pay-performance sensitivity declines. Intrinsic risk,
however, has conflicting effects on the power of incentives to the EN. An increase in the intrinsic
risk increases the degree of optimism at each date, which has a positive effect on the power of
incentives. However, an increase in the intrinsic risk also increases the costs of risk-sharing, which
has a negative effect on the power of incentives. When the EN is optimistic and ∆0 ≤ 4p, the
costs of risk-sharing outweigh the benefits of the EN’s optimism so that the EN’s pay-performance
sensitivity also decreases with intrinsic risk. When ∆0 > 4p, the conflicting effects of optimism and
risk-sharing costs cause the effects of intrinsic risk to be ambiguous.
The change in the VC’s investment path when the EN is optimistic critically depends on whether
the EN is reasonably optimistic or exuberant. If the EN is reasonably optimistic, then the costs of
risk-sharing outweigh the benefits of the EN’s optimism so that the VC’s investment path declines4The condition is trivially satisfied when ∆0 < 0. Under Assumption 2, the condition ∆0 ≤ 4p is automatically
satisfied when bM ≤ 4. Since b∗t < bM , this condition implies that the EN’s optimal pay-performance sensitivityshould be less than four, which is easily satisfied in reasonable parametrizations of the model.
21
6
- t
c∗t
t∗2 t∗1
c(1) “no agency” investment level..................................... ........ .......... ...... ........ .......... ...... ......... ........... ...... ......... ........... ........ ........ .......... ............. ......... .......... .............. ........... ............. ............ ............ ............. ......................... ......................... ........................ ............ ............. ........................ ........... ............. ........................ ........................ ........... ............. ........................ ............. ............ ........... ............. ............. ............ ........................ ........... ............. ............. ........... ........... ............. ....................... ........... ............. ............. ........... .......................
c2(t)
.
...............................
............................
........................
.........................
..........................
................
............................................ ........... .............. ........... ............. .......... ............. .......... .............. ........... ............. ........... ............. ........ .......... .............. ................. ................. .............. .......... .............. ........ .......... .......... ........ ....... ............ ............. ........................ ............. ............ ........... ............. ............. ........... ........... ............. ............. ........... ........... ............. ............. ........... ...... ........ .......... .......... ........ ...... ...... ........ .......... .......... ........ ...... ........... ............. .......................
c1(t)
Figure 4: Sensitivity of the equilibrium investment path to a change in the EN’s risk aversion, theinitial transient risk, or the intrinsic risk. Path c2(·) corresponds to an increase in λ, σ2
0 or s2.
with the EN’s risk aversion as well as the project’s intrinsic and transient risk. If the EN is
exuberant, then an increase in intrinsic or transient risk increases the costs of risk-sharing, thereby
partially offsetting the VC’s rents from the EN’s optimism. Early in the project, it is beneficial for
the VC to compensate for the resulting decline in the EN’s effort by increasing investment. As time
passes, however, the EN’s degree of optimism declines thereby reducing the rents to the VC. The
costs of risk-sharing, therefore, dominate in later in the project so that an increase in risk results
in a decline in the VC’s investment.
In stark contrast with traditional real options models with monolithic agents (e.g. Dixit and
Pindyck, 1992), the results of Theorem 5 show that the interactive effects of optimism and agency
conflicts could lead to a positive or negative relation between risk and investment.
Theorem 6 (Effects of Degree of Asymmetry in Beliefs)
(a) The path of the EN’s equilibrium pay performance sensitivity is pointwise increasing in the
initial degree of asymmetry in beliefs.
(b) If the EN is pessimistic or reasonably optimistic, then the path of the VC’s equilibrium
investment is pointwise increasing in the initial degree of asymmetry in beliefs.
(c) If the EN is exuberant, then the path of equilibrium investment by the VC changes as in
Figure 5 as a result of a change in the initial degree of asymmetry in beliefs—the time-path
of investment shifts “to the right” if the initial degree of asymmetry increases.
22
6
- t
c∗t
t∗1 t∗2
c(1) “no agency” investment level.............................. ....... ........ .......... ...... ........ .......... ...... ........ ........... ...... ......... ........... ....... ........ ........... ........... ........ ....... .............. ........... ............. ............ ........................ ........................ ............ ............. ........................ ............ ............. ........................ ........... ............. ........................ ............ ............ ........... ............. ............. ........... .......... .............. ............. ........... .......... .............. ............. .......... ........... ............ ....................... .......... .............. .......... ........ ...... .......... .............. ............. ........... .......................
c1(t)
.
...............................
............................
........................
.........................
..........................
................
............................................ ........... .............. ........... ............. .......... ............. .......... .............. ........... ............. ........................ ...................................... ........ ........... .............. .................. ..................... ........................ ............................ ...... ........ ........... .......... ........ ....... ............ ............. ........................ ............. ............ ........... ............. ............. ........... ........... ............. ............. ........... ........... ............. ............. ........... ...... ........ .......... .......... ........ ...... ...... ........ .......... .......... ........ ...... ........... ............. ....................... c2(t)
Figure 5: Sensitivity of equilibrium investment path to the initial degree of asymmetry in beliefs.Path c2(·) corresponds to an increase in ∆0.
An increase in the initial degree of asymmetry in beliefs increases the power of incentives that
can be provided to the EN so that his pay-performance sensitivity increases at each date. When
the EN is pessimistic or reasonably optimistic, the VC increases her investment at each date. When
the EN is exuberant, however, the investment path is non-monotonic. The intuition for the effects
of risk on investment discussed earlier is reversed so that the degree of asymmetry in beliefs affects
the investment path as described in part (c) of the theorem.
5 Project Duration
The following proposition shows that there exists a trigger level of the project’s mean quality at
each date such that it is optimal for the VC to continue the project if and only if her current
assessment of the project’s quality exceeds the trigger.
Proposition 4 (The Optimal Termination Policy)
The optimal stopping policy for the VC is a trigger policy: there exist µ∗t such that the VC
terminates the project only if µV Ct < µ∗t .
23
Let Y ∗t dt := (c∗t
αη∗tβ − lt)dt denote the equilibrium net discretionary output in period [t, t+ dt].
Since dVt = Y ∗t dt + ξtdt, it follows that
Vt − V0 =∫ t
0dVu =
(∫ t
0Y ∗
u du)
+(∫ t
0ξudu
).
Given the formula for µt given in (3), we may conclude that µt ≥ µ∗t if and only if Vt ≥ V ∗t , where
V ∗t := V0 +
(∫ t
0Y ∗
u du)
+(s2 + tσ2
0)µ∗t − s2µ0
σ20
.
The sequence of the V ∗t may be thought of as the performance targets the project must reach at
each date to prevent termination.
An increase in the EN’s initial degree of optimism about project quality increases the rents to
the VC from the EN’s optimism thereby increasing her expected continuation value at each point
in time. Hence, an increase in the EN’s optimism prolongs the project’s duration. An increase
in the EN’s risk aversion or cost of effort, however, increases the costs of risk-sharing for the VC,
thereby lowering her continuation value at each point in time.
Proposition 5 (Comparative Statics of Project Duration)
The project duration τ increases with the initial degree of asymmetry in beliefs, decreases with the
EN’s risk aversion, and decreases with the EN’s cost of effort.
6 Numerical Analysis
We numerically explore further implications of the model using a discrete-time approximation of
the continuous-time model. We describe the details of the numerical implementation in the on-line
Appendix C. We directly model the evolution of the VC’s current assessment of project quality µV Ct
because, as explained in Section 5, it determines her continuation decision at any date t. In the
first stage of the numerical implementation, we approximate the evolution of µV Ct using a discrete
lattice and derive the termination triggers µ∗t . In the second stage, given the triggers obtained from
the first stage, we use Monte Carlo simulation to model the evolution of µV Ct and to obtain the
key output variables of interest. Gompers (1995) reports that the average length of a round of VC
24
financing is approximately one year. Accordingly, we set the time period between successive dates
in the discrete lattice to one year and assume that it corresponds to a single round of financing.
6.1 Calibration
To obtain a reasonable set of “baseline” parameter values for our numerical analysis, we calibrate
the model to actual aggregate data on the distribution of round by round returns of venture
capital projects reported in Cochrane (2005). We classify the parameters of the model into two
groups: “direct” parameters whose baseline values can be set using guidance from previous empirical
research, and “indirect” parameters whose values are estimated by matching statistics predicted
by the model to their observed values in the data.
In our numerical implementation, we incorporate a nonzero discount rate for the VC (and EN).
Cochrane (2005) finds that the average expected return on venture capital investment in his sample
is 15%. Accordingly, we set the discount rate, Rb, to 15%. Further, we assume that the VC has
experience so that her prior assessment of the project quality distribution is correct. We assume a
production technology with constant returns to scale so that β = 1 − α. We assume a quadratic
form l(t) = l1tl2 for the loss function. We normalize the initial seed capital V0 to one and set P0
to this value. We estimate the remaining parameters of the model (see Table 1) by matching its
predictions to data.
In our estimation, we use statistics on the round by round returns and standard deviations of
VC projects in each of the first four rounds of financing reported in Table 4 of Cochrane (2005).
As the length of a single round of financing is set to one year, the round-by-round returns of a VC
project in the model are Vt−Vt−1−ct−1
ct−1, 1 ≤ t ≤ 4. Cochrane (2005) also reports the overall mean
and standard deviation of the VC project returns and the average number of rounds of financing.
The statistics used for our estimation are displayed in the first rows of the two panels of Table
2. We estimate the values of the indirect parameters of the model by matching the predicted
values of the statistics in Table 2 to their observed values. The standard errors of the estimates are
determined by parametric bootstrapping (see the on-line Appendix C).
As shown in Table 2, the model is able to closely match the observed statistics. The estimated
value of the degree of asymmetry in beliefs ∆0 = 0.504, while the VC’s assessment of the mean
25
project quality µV C0 = 0.113. The data, therefore, suggest that the level of entrepreneurial optimism
is very significant. The baseline values of the average intrinsic risk s and transient risk σ0 are high,
which confirms anecdotal and empirical evidence that venture capital is risky and is characterized
by significant uncertainty about project quality.
6.2 Numerical Results
We first analyze the model when the parameters take their baseline values in Table 1. We then
explore various comparative static relationships by varying parameters about their baseline value.
In our numerical analyses, we compare the actual scenario in which there are asymmetric beliefs
and agency conflicts with two benchmark scenarios: the no agency scenario in which beliefs are
symmetric, and both the VC and EN have linear preferences; and the symmetric beliefs scenario
in which the VC and EN have symmetric beliefs, but the EN has CARA preferences.
We compute two output variables in each of the three scenarios. The
Project V alue := EV C0
[e−RbτVτ −
τ−1∑
t=0
e−Rbtct
]. (22)
is the expected total payoffs to the firm less the capital investments discounted at the rate Rb. The
V C V alue := Project Value− EV C0
[e−RbτPτ
]. (23)
is the Project Value less the termination payoff to the EN discounted at the rate Rb. The expecta-
tions in (22) and (23) are with respect to the VC’s beliefs about project quality, which are assumed
to be correct. The termination payoff process evolves as in (1) with the contractual parameters,
(a∗, b∗, c∗), the EN’s effort, η∗, and the performance targets, V ∗, set to their equilibrium values
for the specific economic scenario (no agency, symmetric beliefs or actual) being analyzed.
6.2.1 Baseline Analysis
Table 3 reports the Project Value and VC Value in the actual scenario and the two benchmark
scenarios. The difference between the project values (VC values) in the no agency and symmetric
benchmark scenarios represent the deadweight agency costs of risk sharing between the VC and
26
EN from the perspective of the firm (the VC fund). The difference between the project values (VC
values) in the actual and symmetric scenarios reflect the effects of EN optimism on the project
value (VC value). We see that the VC significantly benefits from EN optimism. The project value,
however, is lower in the actual scenario than in the symmetric beliefs scenario. This is the due to
the fact that the VC exploits EN optimism by over-investing and prolonging the project’s duration
to increase her value at the expense of the overall value of the project.
Table 4 reports the EN’s pay-performance sensitivities and the investments for the first four
rounds. Consistent with Theorem 3, the EN’s pay-performance sensitivity and the VC’s investments
decline over time. The EN’s pay-performance sensitivity decreases sharply across the four periods.
Successive capital infusions by the VC, therefore, rapidly reduce the EN’s stake in the firm.
6.2.2 Comparative Statics
Figure 6 (a) shows that the project value and VC value both vary non-monotonically with the
initial transient risk—they initially decrease and then increase. To understand the intuition for the
effects of transient risk, note that, by (4), the degree of EN optimism at any date t declines with
the initial transient risk. This has a negative effect on the power of incentives to the EN, his effort,
and the economic rents the VC can extract from EN optimism. From (2) and (3), the standard
deviation σµt of the evolution of the mean assessment of project quality is
σµt =
sσ20
s2 + tσ20
. (24)
From (24), an increase in the initial transient risk increases the standard deviation of the evolution
of the mean assessment of project quality and, therefore, the likelihood of both high and low
realizations. Since the VC can limit her downside by terminating the relationship if intermediate
signals of project quality are sufficiently poor, the “real option value” of continuing the project
increases with the initial transient risk. The interaction between the negative effects of transient
risk on the degree of EN optimism and its positive effects on the real option value of continuation
causes the project value and VC Value to vary non-monotonically with the initial transient risk.
27
Figure 6 (b) shows that the project value and VC value decline with the intrinsic risk, s.5 From
(24), the standard deviation of the evolution of the mean assessment of project quality decreases
with intrinsic risk above a threshold. Hence, the option value of continuing the relationship in any
period also declines. An increase in the intrinsic risk also increases the costs of risk-sharing, which
has a negative effect on the EN’s effort and the VC’s investment. The project value and VC value,
therefore, decline with intrinsic risk.
Figure 6 (c) shows that the VC value significantly increases with the degree of asymmetry in
beliefs, which illustrates the benefits to the VC from exploiting the EN’s optimism by providing
more powerful incentives. The project value, however, initially increases and then decreases with
the degree of asymmetry in beliefs. When the degree of EN optimism is low, the positive effects
of increased optimism on the EN’s effort, the VC’s investment and the project’s output cause the
project value to increase with optimism. When the degree of EN optimism is above a threshold,
however, the VC exploits the EN’s optimism by sub-optimally (from the standpoint of project
value) prolonging the project’s duration and over-investing in the project, thereby increasing her
value at the expense of project value.
6.2.3 Implied Discount Rates
There is considerable empirical and anecdotal evidence that VCs typically use high discount rates
in the range between 35% and 50% to value projects (see Sahlman, 1990, Gladstone and Gladstone,
2002). Sahlman (1990) suggests that high discount rates could be a mechanism that VCs use to
adjust optimistic projections by ENs. To the best of our knowledge, however, it has not been
ascertained whether optimism could indeed explain the high discount rates used by VCs. Do the
levels of EN optimism predicted by our model lead to the discount rates observed in reality?
We calculate the implied discount rate (IDR) of a project as the rate the VC would use to
discount the EN’s projections of the project’s payoffs to equal her own valuation of the project
defined in (23). In other words, the IDR is the discount rate the VC would use to obtain her
valuation of the project if the project’s intrinsic quality were (hypothetically) distributed according5We allow for the discount rate to vary with the intrinsic risk as follows: Rb(s) := r + ((0.15 − r)/s)s, where
r = 0.068 is the risk-free rate reported in Cochrane (2005) and s is the baseline value of s.
28
to the EN’s beliefs. The IDR βV C solves:
E0
[e−βV Cτ (Vτ − Pτ )−
τ−1∑
t=0
e−βV CtctVt
∣∣∣ Θ ∼ N(µEN0 , σ2
0)]
= VC value. (25)
Figure 7 reports the IDR’s for varying values of α, σ0, s, ∆0. The range of IDR’s is consistent with
the discount rates that VC’s use to assess the value of a new venture (see Sahlman, 1990). Our
results, therefore, suggest that entrepreneurial optimism, indeed, explains the discount rates used
by VCs in reality.
7 Conclusions
We develop a continuous-time, stochastic principal-agent model to investigate the effects of asym-
metric beliefs and agency conflicts on the characteristics and valuation of venture capital projects.
We characterize the equilibrium of the stochastic dynamic game in which the VC’s dynamic in-
vestments, the EN’s effort choices, the dynamic compensation contract between the VC and EN,
and the project’s termination time are derived endogenously. Consistent with observed contractual
structures, the equilibrium dynamic contracts feature both equity-like and debt-like components,
the staging of investment by the VC, the progressive vesting of the EN’s stake, and the presence of
inter-temporal milestones or performance targets that must be realized for the project to continue.
We numerically implement the model and calibrate it to aggregate data on VC projects. Our
numerical analysis shows that EN optimism significantly enhances the value that venture capitalists
derive. Entrepreneurial optimism explains the huge discrepancy between the discount rates used
by VCs (∼ 40%), which adjust for optimistic payoff projections by ENs, and the average expected
return of VC projects (∼ 15%). Our results show how the “real option” value of venture capital
investment is affected by the presence of agency conflicts and asymmetric beliefs. Permanent and
transitory components of projects’ risks have differing effects on their values and durations.
29
References
1. Allen, F. and Gale, D. (1999), “Diversity of Opinion and the Financing of New Technologies,”Journal of Financial Intermediation, 8, 68-89.
2. Atkinson, A. 1979. Incentives, Uncertainty and Risk in the Newsboy Problem. DecisionSciences 10 341-353.
3. Baker, M., Ruback, R. and Wurgler, J. (2005), “Behavioral Corporate Finance: A Survey,”Handbook of Corporate Finance: Empirical Corporate Finance (ed: E. Eckbo), Part III,Chapter 5, Elsevier/North-Holland.
4. Berk, J., Green, R. and Naik, V. (2004), “Valuation and Return Dynamics of New Ventures,”The Review of Financial Studies, 17, 1-35.
5. Biais, B., Mariotti, T., Plantin, G., Rochet, J. (2007), “Dynamic Security Design: Conver-gence to Continuous Time and Asset Pricing Implications,” Review of Economic Studies, 74,345-390.
6. Cachon, G.P., M.A. Lariviere. 2001. “Contracting to Assure Supply: How to Share DemandForecasts in a Supply Chain,” Management Sci. 47(5) 629-646.
7. Cadenillas, Abel, Jaksa Cvitanic, and Fernando Zapatero, 2006. “Optimal risk-sharing witheffort and project choice,” Journal of Economic Theory, forthcoming.
8. Casamatta, C. (2003), “Financing and Advising: Optimal Financial Contracts with VentureCapitalists,” Journal of Finance, 58, 2059-2086.
9. Chen, F. 2005. Salesforce Incentives, Market Information, and Production/Inventory Plan-ning. Management Sci. 51(1) 60-75.
10. Chen F., S.A. Zenios. 2005. Introduction to the Special Issue on Incentives and Coordinationin Operations Management. Management Sci. 51(1) 1.
11. Cochrane, J. (2005), “The Risk and Return of Venture Capital,” Journal of Financial Eco-nomics, 75, 3-52.
12. Cornelli, F. and Yosha, O. (2003), “Stage Financing and the Role of Convertible Securities,”Review of Economic Studies, 70, 1-32.
13. Cvitanic, J., Wan. X and Zhang, J. 2005, “Continuous-Time Principal-Agent Problems withHidden Action and Lump-Sum Payment”, Working paper.
14. Davison, A.C. and Hinkley, D.V. (1997). Bootstrap Methods and their Application, CambridgeSeries in Statistical and Probabilistic Mathematics.
15. DeMarzo, P. and Fishman, M. (2007), “Optimal Long-Term Financial Contracting with Pri-vately Observed Cash Flows,” Review of Financial Studies, forthcoming.
16. Dixit, A., R. Pindyck. (1994), Investment Under Uncertainty. Princeton University Press,Princeton, NJ.
17. Egli, D. and Ongena, S. and Smith, D.C. (2006), “On the Sequencing of Projects, ReputationBuilding, and Relationship Finance,” Finance Research Letters, 3, 23-39.
30
18. Gelderen, M. and Thurik, R. and Bosma, N. (2005), “Success and Risk Factors in the Pre-Startup Phase,” Small Business Economics, 24, 365-380.
19. Gibbons, R. 2005. Incentives between Firms (and Within). Management Sci. 51(1) 2-17.
20. Gibbons, R., Murphy, K.J. (1992), “Optimal Incentive Contracts in the Presence of CareerConcerns: Theory and Evidence,” Journal of Political Economy 100(3) 468-505.
21. Gladstone, D. and Gladstone, L. (2002), Venture Capital Handbook: An Entrepreneur’s Guideto Raising Venture Capital, Prentice Hall, Upper Saddle River, New Jersey.
22. Gompers, P. (1995), “Optimal Investment, Monitoring, and the Staging of Venture Capital,”The Journal of Finance, 50, 1461-1489.
23. Gompers, P. and Lerner, J. (1996), “The Use of Covenants: An Empirical Analysis of VenturePartnership Agreements,” Journal of Law and Economics, 39, 463-498.
24. Ha, A.Y. 2001. Supplier-Buyer Contracting: Asymmetric Cost Information and Cutoff LevelPolicy for Buyer Participation. Naval Res. Logist. 48(1) 41-64.
25. Holmstrom, B. (1999), “Managerial Incentive Problems: A Dynamic Perspective,” Review ofEconomic Studies 66, 169-182.
26. Holmstrom, B., Milgrom, P. (1987), “Aggregation and Linearity in the Provision of Intertem-poral Incentive,” Econometrica 55(2) 303-328.
27. Kaplan, S. and Stromberg, P. (2003), “Financial Contracting Theory Meets the Real World:An Empirical Analysis of Venture Capital Contracts,” Review of Economic Studies, 70, 281-315.
28. Kockesen, L. and Ozerturk, S. (2004), “Exclusivity and Over-investment: A Model of Rela-tionship Financing,” Working Paper, Columbia University.
29. Kulatilaka, N., E.C. Perotti. 1998. Strategic Growth Options. Management Sci. 44(8)1021-1031.
30. Lal, R., V. Srinivasan. 1993. Compensation Plans for Single and Multi-Product Salesforces:An Application of the Holmstrom-Milgrom Model. Management Sci. 39(7) 777-793.
31. Morris, S. (1995), “The Common Prior Assumption in Economic Theory,” Economics andPhilosophy, 11, 227-253.
32. Neher, D. (1999), “Staged Financing: An Agency Perspective,” Review of Economic Studies,66, 255-274.
33. Oksendal, B. (2003), Stochastic Differential Equations: An Introduction with Applications,Sixth Edition, Springer-Verlag, Heidelberg, Germany.
34. Plambeck, E.L., S.A. Zenios. 2000. Performance-Based Incentives in a Dynamic Principal-Agent Model. Manufacturing Service Oper. Management 2(3) 240-263.
35. Plambeck, E.L., S.A. Zenios. 2003. Incentive Efficient Control of a Make-to-Stock ProductionSystem. Oper. Res. 51(3) 371-386.
36. Repullo, R. and Suarez, J. (2004), “Venture Capital Finance: A Security Design Approach,”Review of Finance, 8, 75-108.
31
37. Sahlman, W. (1990), “The Structure and Governance of Venture-Capital Organizations,”Journal of Financial Economics, 27, 473-521.
38. Sannikov, Y. (2007), “A Continuous-Time Version of the VC-EN Problem,” Review of Eco-nomic Studies, forthcoming.
39. Schattler, H. and Sung, J. (1993), “The First Order Approach to the Continuous-Time VC-ENProblem with Exponential Utility,” Journal of Economic Theory 61, 331-371.
40. Schmidt, K. (2003), “Convertible Securities and Venture Capital Finance,” Journal of Fi-nance, 58, 1139-1166.
41. Spear, S. and Srivastava, S. (1987), “On Repeated Moral Hazard with Discounting,” Reviewof Economic Studies 54(4), 599-617.
42. Spear, S. and Wang, C. (2005), “When to Fire a CEO: Optimal Termination in DynamicContracts,” Journal of Economic Theory 120(2), 239-256.
43. Sung, J. (1995), “Linearity with Project Selection and Controllable Diffusion Rate in Continuous-Time Principal-Agent Problems,” RAND Journal of Economics 26, 720-743.
44. Trigeorgis, L., 1999, Real Options: Managerial Flexibility and Strategy in Resource Allocation,The MIT Press, Cambridge, MA.
32
Appendix A
Proof of Theorem 1
A rigorous proof of the theorem requires a precise interpretation of equation (1), which describesthe evolution of the termination payoff process. As in the traditional principal-agent literature(see Holmstrom and Milgrom, 1987), we consider the termination payoff process V (·) to be agiven random process on a probability space with investment and effort altering the probabilitydistribution of this process.
The Termination Payoff Process
We consider an underlying probability space (Ω,F) with probability measures Q`, ` ∈ V C, EN,representing the VC’s and EN’s beliefs. Θ is a normal random variable with variance σ2
0 andmean µ`
0 under measure Q` and B is a standard Brownian motion. The complete and augmentedfiltration of the probability space generated by the Brownian motion B(·) is denoted by Ft.Consider the process V (·) = sB(·) where s2 is the intrinsic risk of the project. We will use theGirsanov transformation (see Oksendal, 2003) to obtain new probability measures on (Ω,F) suchthat the process V (·) evolves as in (1).
Suppose that η(·) and c(·) are strictly positive, square-integrable Ft-measurable stochasticprocesses (under the measures QV C and QEN ) defined on the time horizon [0, T ] describing the EN’schoices of effort and the VC’s choices of investments over time. Recall that l(·) is a deterministicprocess describing the operating costs of the firm. Define the processes
ζc,η(t) := exp[∫ t
0(Θ + Ac(u)αη(u)β − l(u))s−1dB(u)− 1
2
∫ t
0(Θ + Ac(u)αη(u)β − l(u))2s−2du] (26)
Bc,η(t) := B(t)−∫ t
0(Θ + Ac(u)αη(u)β − l(u))s−1du . (27)
The process ζc,η(·) is a positive, square-integrable martingale.6 Define the new measure Π`c,η by
dΠ`c,η
dQ`= ζc,η(T ). (28)
By Girsanov’s theorem (see Oksendal, 2003), the process Bc,η(·) is a Brownian motion under the
6The processes are assumed to satisfy the Novikov condition (see Oksendal, 2003):
E` exp[1
2
∫ T
0
(Θ + Ac(u)αη(u)β − l(u))2s−2du] < ∞, ` ∈ V C, EN .
Because the equilibrium investment and effort processes described in Theorem 1 are deterministic and Θ is a normalrandom variable, the Novikov condition is satisfied by these processes.
In fact, we do not need to assume that feasible (not necessarily optimal) investment and effort processes satisfy theNovikov condition for our analysis to be valid; we only require that they be square-integrable. In this case, the processζc,η(·) is only guaranteed to be a local martingale and the measure Π`
c,η is a finite measure, but not necessarily aprobability measure. Our analysis, however, only requires that Π`
c,η be a finite measure. Since, as mentioned earlier,the Novikov condition is satisfied by the equilibrium investment and effort processes, the measure corresponding tothe equilibrium processes is a probability measure.
33
measure Π`c,η. Further, under this measure, the process V (·) evolves as
dV (t) = [Θ + Ac(t)αη(t)β − l(t)]dt + sdBc,η(t) . (29)
Equation (29) describes the evolution of the termination payoff process and is identical to equation(1), but with the Brownian motion and the probability measures representing the VC’s and EN’sbeliefs depending on the investment and effort processes. It is important to keep in mind that V (·)is a fixed process whose sample paths are not affected by investment and effort. Investment andeffort, however, alter the probability distribution of the sample paths of V (·).
For future reference, we make an important observation. The process
dWc,η(t) := s−1[dV (t)− (Ac(t)αη(t)β − l(t))dt− µ`tdt] (30)
is an Ft-Brownian motion with respect to the probability measure Π`c,η. Moreover, the complete
and augmented filtration generated by this Brownian motion is Ft. The EN’s and VC’s meanassessments of project quality Θ at date t, µEN
t , µV Ct are given by (2) and (??).
Utility Related Processes
Let τ ≤ T be an Ft-stopping time denoting the termination time of the VC-EN relationship.Let c(·), η(·) and η(·) be strictly positive Ft-adapted square-integrable processes on [0, τ ].7 Acontract is represented by the quadruple (Pτ , c(·), η(·), τ). We now define some processes that areused frequently in the sequel.
• Cumulative value process of the EN. This is the conditional expected future utility to the EN atany date including the sunk disutilities of prior effort, from a given contract (Pτ , c(·), η(·), τ).Formally,
UP,c,η,τ (t) := EENc,η [− exp(−λ[P (τ)−
∫ τ
0kη(u)γdu]) | Ft] . (31)
Here, E`c,η[· | Ft]; ` ∈ Pr,Ag denotes conditional expectation at date t under the probability
measure Π`c,η defined in (28). For future reference, we note that the cumulative value process
of the EN is a square-integrable Ft-martingale under the measure ΠENc,η .
• Promised payoff process for the EN. The EN’s promised payoff process corresponding to agiven contract (Pτ , c(·), η(·), τ) is defined by (6).
• Adjusted cumulative value process of the EN. This process represents the cumulative valueprocess of the EN from a contract where his effort is η(s); s ≤ t and effort η(s); s ≥ t. Formally,
YP,c,τ (η(·); t; η(·)) := EENc,η [− exp(−λ[P (τ)−
∫ t
0kη(u)γdu−
∫ τ
tkη(u)γdu]) | Ft] . (32)
• EN’s maximum conditional expected utility process. This process represents the EN’s maxi-mum conditional expected utility at date t given that he has exerted effort η(s); s ≤ t and his
7These processes are assumed to satisfy the Novikov condition—see footnote 6.
34
terminal payoff, the VC’s investment process, and the termination time are (Pτ , c(·), and τ),respectively. Formally, we define
ZPτ ,c,η,τ (t) := supη(·)YP,c,τ (η(·); t; η(·)) . (33)
To simplify the subsequent notation, we drop the subscripts denoting the dependence of the pro-cesses defined in (31)-(33) on the contract wherever there is no danger of confusion.
Structure of Incentive Compatible Contracts
A contract (Pτ , c(·), η(·), τ) is incentive compatible with respect to the EN’s effort if and only if,
given the terminal payoff, Pτ , the VC’s investment process, c(·), and the termination time, τ ,
the EN’s optimal effort choices are η(·). The following lemma, from which the theorem follows,
characterizes incentive compatible contracts.
Lemma 1 (Incentive Compatible Contracts)
A contract (Pτ , c(·), η(·), τ) is incentive compatible only if the EN’s promised payoff process P (·)satisfies the following stochastic differential equation:
dP (t) = a(t)dt + b(t)dV (t) (34)
where
b(t) =γk
Aβc(t)αη(t)γ−β (35)
and
a(t) :=λ
2b(t)2s2 + kη∗(t)γ − b(t)
(Ac(t)αη(t)β − l(t) + µEN
t
). (36)
Proof. Define the process η(.) as follows:
η(s) = η(s) for s 6= t, η(t) = η′(t), (37)
where η′(t) is any candidate (possibly sub-optimal) effort choice of the EN at date t. By theprinciple of optimality of dynamic programming (Oksendal, 2003), the effort η(t) is optimal for theEN at date t when his prior effort choices are η(·) only if
η(t) = argmaxη′(t)EENc,η [Z(η(·); t + dt)− Z(η(·); t) | Ft] = argmaxη′(t)Ec,η[dZ(η(·); t) | Ft] . (38)
In what follows, we derive the infinitesimal change dZ(η(·); t) and then use (38) to establish thestatements of the Lemma. It follows from (32), (33), and (37) that
dZ(η(·); t) = dZ(η(·); t) + Z(η(·); t)kλ(η′(t)γ − η(t)γ)dt . (39)
35
Since the process η(·) represents the EN’s optimal effort choices by hypothesis, it follows that
Z(η(·), t) = U(η(·), t), (40)
and hence the process Z(η(·), ·) is a square-integrable Ft-martingale under the measure ΠENc,η .
By the martingale representation theorem (see Oksendal, 2003), there exists a square-integrable,Ft-adapted process ω(·) such that8
dZ(η(·); t) = ω(t)dWc,η(t) = ω(t)s−1[dV (t)− (Ac(t)αη(t)β − l(t))dt− µENt dt] . (41)
Since the expectation in the dynamic programming equation (38) is taken under the measure ΠENc,η ,
it follows from (30) and (41) that Z(η(·); t) evolves under this measure as
dZ(η(·); t) = ω(t)s−1Ac(t)α(η′(t)β − η(t)β)dt + ω(t)dWc,η(t) . (42)
Substituting (42) in (39) yields
dZ(η(·); t) =[ω(t)s−1Ac(t)α(η′(t)β − η(t)β) + kλZ(η(·); t)(η′(t)γ − η(t)γ)
]dt
+ω(t)dWc,η(t) . (43)
Having derived the requisite expression for dZ(η(·); t), we substitute it in (38) to obtain
η(t) = argmaxη′(t)[ω(t)s−1Ac(t)αη′(t)β + kλZ(η(·); t)η′(t)γ ] . (44)
It follows that the effort η(t) is optimal over the interval [t, t + dt] if and only if
ω(t)Z(η(·); t) = − kλs
Ac(t)α
γ
βη(t)γ−β . (45)
From the definition of the promised payoff process in (6), and using (40), we have
P (t) = − logZ(η(·); t)λ
+∫ t
0kη(u)γdu . (46)
Using Ito’s Lemma and (41), we obtain
dP (t) = −ω(t)s−1[dV (t)− (Ac(t)αη(t)β − l(t))dt− µENt dt]
λZ(η(·); t) +ω(t)2
2λZ(η(·); t)2 dt + kη(t)γdt. (47)
Now substituting the expression (45) for the quantity ω(t)Z(η(·);t) in (47) yields
dP (t) = a(t)dt + b(t)dV (t), (48)8Identity (41) is an almost sure relation that holds under all equivalent probability measures on the probability
space. It is only under the measure ΠENc,η defined in (28) that the process [dV (t) − (c(t)αη(t)β − l(t))dt − µEN
t dt] isthe increment of a Brownian motion.
36
where b(t) and a(t) are given by (35) and (36), respectively, as claimed. This completes the proofof the lemma.
Because an incentive compatible contract must have the affine form by Lemma 1, it immediatelyfollows that the optimal contract must also have the same affine form. This completes the proof ofTheorem 1.
Proof of Theorem 2
By (35) and (36), a candidate optimal contract is completely described by the investment processc(·), the processes a(·), b(·) describing the performance-invariant and performance-dependent com-ponents of the EN’s promised payoff process, and the termination time τ . By (35), the contract isincentive compatible if and only if the EN’s effort at date t is
ηt = η(bt, ct) :=(Aβcα
t bt
γk
) 1γ−β
. (49)
DefineMa,b,c,τ (0) = EV C
c,η
[(V (τ)− P (τ)−
∫ τ
0c(s)ds)
](50)
as the VC’s expected future payoff at date 0 if she chooses a contract (Pτ , c(·), η(·), τ) ≡ (a(·), b(·), c(·), τ),where η(.) is given by (49). The VC’s contract choice problem is then the following:
(a∗(·), b∗(·), c∗(·), τ∗) = argmax(a,b,c,τ)Ma,b,c,τ (0). (51)
Let the “state” of the system at any date t be described by the ordered pair (t, µV Ct ). We
first restrict consideration to Markov controls where a(t), b(t), c(t) and the decision to terminatethe relationship only depend on the current state (t, µV C
t ). We derive the optimal Markov controlpolicy. We then appeal to the verification theorem of dynamic programming (see Theorem 11.2.3 ofOksendal, 2003) to conclude that the optimal Markov control policy is, in fact, the optimal controlpolicy over the entire space of admissible Ft-adapted controls.
We note from (35) and (36) that the control a(·) is determined by the controls b(·), c(·) andthe state of the system. Hence, a Markov control policy is completely described by (b(·), c(·), τ).For simplicity, we abuse notation by denoting the VC’s continuation value in state (t, µV C
t ) fromadopting the Markov control policy (b(·), c(·), τ) by
Mb,c,τ (t, µV Ct ) = EV C
t;c,η
[(V (τ)− V (t))− (P (τ)− P (t))−
∫ τ
tc(s)ds
], (52)
where η(·) is determined by (49). Let M∗(t, µV Ct ) be the optimal continuation value within the
space of Markov controls and (b∗(·), c∗(·), τ∗) be the optimal Markov control policy (we derive thispolicy in the following).
Suppose that the VC deviates from the optimal policy over the infinitesimal time interval[t, t+ dt] by choosing the controls (b(t), c(t)), Let M(t, µV C
t ) denote the VC’s continuation value at
37
date t under this deviated policy. It follows from Lemma 1 and (52) that
M(t, µV Ct ) = EV C
t;c,η
[− a(t)dt + (1− b(t))dV (t)− c(t)dt + M∗(t + dt, µV C
t+dt)]
. (53)
By (29), we have
M(t, µV Ct ) = EV C
t;c,η
[−a(t)dt+(1−b(t))[µV C
t +Ac(t)αη(t)β−l(t)]dt−c(t)dt+M∗(t+dt, µV Ct+dt)
], (54)
where η(t) is given by (49) with b(t) replacing b(t). Since the VC’s investment and EN’s effort areobservable, the VC’s assessment µV C
t+dt of project quality at date t+dt is independent of the choicesof controls (b(t), c(t)). Hence, the function M∗(t + dt, µV C
t+dt) is also independent of these choices.By the principle of optimality of dynamic programming (see Oksendal, 2003), the optimal controls(b∗(t), c∗(t)) at date t must maximize the “flow” term in (54), that is,
(b∗(t), c∗(t)) = argmaxb(t),c(t)
− a(t)dt + (1− b(t))[µV C
t + Ac(t)αη(t)β − l(t)]dt− c(t)dt
(55)
By (36) and (49), we can show (after some algebra) that
(b∗(t), c∗(t)) = arg maxb(t),c(t)
Λt(b(t), c(t))dt, (56)
whereΛt(b, c) := (∆tb− 0.5λs2b2 + φ(b)cα γ
γ−β − c + µV Ct − lt) (57)
and
φ(b) := Aγ
γ−β
(1k
) βγ−β
((βb
γ
) βγ−β
(1− βb
γ
)). (58)
We first determine the VC’s optimal investment rate c(b) as a function of the EN’s pay-performance sensitivity b and then simultaneously derive the optimal investment rate and pay-performance sensitivity. By (57) and (58), the optimal investment rate is zero if b ≥ γ/β. Forb ∈ (0, γ/β), Assumption 1 guarantees that the function Λt(b, ·) is strictly concave in the invest-ment rate c (the exponent on c is guaranteed to be less than 1). As a consequence, setting thepartial derivative of Λt(b, ·) with respect to c equal to zero implies that the optimal investment asa function of the pay-performance sensitivity b is given by (13). Substitution in (56), we see thatthe VC chooses the pay-performance sensitivity at date t to solve (14).
By the above arguments, the VC receives a “flow” payoff Λt(b∗t , c∗t )dt in each infinitesimaltime period [t, t + dt]. It immediately follows that the risk-neutral VC chooses to terminate therelationship at the stopping time that solves (16).
The Markov control policy derived above trivially satisfies the conditions of the dynamic pro-gramming verification theorem (see Section 11 of Oksendal, 2003). Hence, it is, in fact, the optimalcontrol policy among the space of all square-integrable Ft-adapted controls. This completes theproof of Theorem 2.
38
Appendix B: Proofs of Remaining Results
For each model parameter “Π” (e.g. σ20, s2, λ, ∆0, k) we let bt(π) denote the solution to (18)
at time t, define ct(π) := c(bt(π)), and let b(π) and c(π) denote the corresponding time pathswhen the parameter Π’s value equals π. We write F ′
t(b, π) when we wish to explicitly indicate thefunctional dependence of the derivative of Ft on the parameter value π. For subsequent reference,the derivative of the VC’s controllable rate function (17) is given by
F ′t(b) = ∆t − pb +
γ − β − αγ
αγc′(b) =
s2
s2 + tσ20
∆0 − pb +γ − β − αγ
αγc′(b). (59)
Proof of Proposition 1.
The marginal optimal investment is given by
c′(b) ∝(1
k
) 1(1−α)
γβ−1 br1(γ − b)r2(1− b) (60)
where
r1 :=2− (1− α) γ
β
(1− α) γβ − 1
and r2 :=α γ
β
(1− α) γβ − 1
,
and where the symbol ∝ means “equal up to a positive multiplicative constant”. Under Assumption2, the parameter r2 is positive and the parameter r1 is negative. Since γ
β > 1 (Assumption 1), thestrong unimodality of c(·) easily follows from (60). Since c(0) = c(γ) = 0 and c′(0) = +∞, it alsofollows from (60) that c(·) achieves its maximum at b = 1. Part (a) has been established.
To establish part (b), we note that the second derivative of the optimal investment function is
c′′(b) ∝ br1−1(γ
β− b)r2−1[r1(
γ
β− b)(1− b)− r2b(1− b)− b(
γ
β− b)].
The expression inside the brackets is a strictly convex quadratic function whose value at 1 is nega-tive, whose value at γ
β > 1 is positive, and whose value at 0 is negative since r1 < 0. Consequently,there is exactly one root bM of the quadratic in the interval (1, γ
β ) such that c′′(bM ) = 0. At bM themarginal investment is at its minimum. Moreover, since c′′(·) is negative on [0, bM ) and is positiveon (bM , γ
β ), the function is strictly concave on [0, bM ] and strictly convex on [bM , γβ ].
Proof of Proposition 2.
Suppose ∆0 ≥ 0. It directly follows from (59) that
F ′t(b) ≤ ∆0 − pb +
γ − β − αγ
αγc′(b),
since ∆t ≤ ∆0 for all t. Figure 1 (p. 16) shows that c′(1) = 0 and c′(b) < 0 for all b ∈ (1, γβ ). It is
straightforward then to check that F ′t(b) < 0 for all b > max∆0
p , 1, which proves the claim when
39
∆0 ≥ 0. Suppose ∆0 < 0 so that ∆t < 0 for all t. Then
F ′t(b) < −pb +
γ − β − αγ
αγc′(b),
By the previous arguments, the right hand side above is less than zero for b ≥ 1 so that b < 1.
Proof of Theorem 3.
a) By (59), for any fixed b > 0, F ′t(b) decreases with t because ∆t = s2
s2+tσ20
∆0 decreases with t for∆0 > 0. By definition, F ′
t(b∗t ) = 0. Hence, F ′
s(b∗t ) < 0 for s > t. Since F ′
s(b∗s) = 0 by definition,
we must have b∗s < b∗t by the strong unimodality of Fs(·). The fact that b∗t → b∗p as t →∞ followsfrom the fact that ∆t → 0 as t →∞ and the Theorem of the Maximum. This establishes part (a).b) The derivative F ′
t(1) = s2
s2+tσ20
∆0 − p is zero when t = t∗ and t∗ ≥ 0. Thus, bt∗ = 1 and ct∗ is atits maximum, which coincides with the “no agency” benchmark case, as required for part (b).c) Since F ′
t(1) > 0 if t < t∗ and F ′t(1) < 0 if t > t∗, part (c) follows from the strong unimodality of
each Ft(·).d) Part (a) establishes that the b∗t decrease with time, and so part (d) follows from part (c) andthe fact that the optimal investment function c(·) increases on [0, 1] and decreases on [1, bM ].e) Since b∗t and c∗t both decrease with time on [t∗,∞), part (e) follows immediately from thefunctional form (8).
Proof of Theorem 4.
a) By (59), for any fixed b > 0, F ′t(b) increases with t because ∆t = s2
s2+tσ20
∆0 increases with t for∆0 < 0. By definition, F ′
t(b∗t ) = 0. Hence, F ′
s(b∗t ) > 0 for s > t. Since F ′
s(b∗s) = 0 by definition,
we must have b∗s > b∗t by the strong unimodality of Fs(·). The fact that b∗t → b∗p as t →∞ followsfrom the fact that ∆t → 0 as t →∞ and the Theorem of the Maximum. This establishes part (a).b) By Proposition 2, 0 < b∗t < 1 when ∆0 < 0. By Proposition 1, the optimal investment functionc(·) is strictly increasing on (0, 1). It follows that, because b∗t increases with t, c∗t = c(b∗t ) alsoincreases. This establishes part (b).c) Since η∗t = η(b∗t , c∗t ), part (c) follows directly from (8) and the fact that both b∗t and c∗t increasewith t.
The following Lemma will be used repeatedly in the proofs to follow.
Lemma 2
If F ′t(b, π) is an increasing (decreasing) function of π, then bt(π) is an increasing (decreasing)
function of π.
Proof. Let π1 < π2. Suppose first that F ′t(b, π) is an increasing function of π. By definition,
0 = F ′t(bt(π2), π2) = F ′
t(bt(π1), π1) < F ′t(bt(π1), π2),
40
which immediately implies bt(π1) < bt(π2) by the strong unimodality of Ft. The proof in thedecreasing case is analogous.
Proof of Theorem 5.
a) If ∆0 < 0, F ′t(b) decreases with λ, decreases with s, and increases with σ0 by (59). By Lemma 2,
b∗t decreases with λ, decreases with s and increases with σ0. Since b∗t ∈ (0, 1) by Proposition 2, andc(·) is increasing in (0, 1) by Proposition 1 that c∗t = c(b∗t ) also decreases with λ, decreases with s
and increases with σ0. This establishes part (a).b) Part (b) follows using similar arguments by observing that F ′
t(b) decreases with λ and decreaseswith σ0.c) Since F ′
t(1, π) = ∆t − p < ∆0 − p < 0, it follows from the strong unimodality of Ft(·, π) thatbt(π) ∈ (0, 1) for all t. Since the function c(·) increases on [0, 1], part (c) follows from part (b).d) We turn to part (d). (Please refer to Figure 4.) Suppose π1 < π2 where π = λ or σ2
0. Firstwe note that by part (a) the path b(π1) lies strictly above the path b(π2). Let t∗j , j = 1, 2, denotethe value of t∗ in (21) corresponding to πj . Clearly, t∗1 > t∗2. By Theorem 3(c), in the interval[0, t∗2) both paths b(π1) and b(π2) lie above one. Since b(π1) > b(π2) and since the function c(·)decreases on [1, bM ], it follows that c(π1) < c(π2) in this interval. Analogously, by Theorem 3(c),in the interval (t∗1,∞) both paths b(π1) and b(π2) lie below one. Since b(π1) > b(π2) and sincethe function c(·) increases on [1, bM ], it follows that c(π1) > c(π2) in this interval. By Theorem3(d), we know that in the interval [t∗2, t
∗1] the path c(π1) increases whereas the path c(π2) decreases.
Moreover, by definition of t∗j we have that ct∗j (πj) = c(1), j = 1, 2, and so
ct∗1(π1) = c(1) > ct∗1(π
2) and ct∗2(π1) < ct∗2(π
2) = c(1).
Thus, the trajectories c(π1) and c(π2) cross exactly once in this interval, as claimed. Part (d) hasbeen established.e) We turn our attention to s2. Since the proofs of parts (c) and (d) when π = λ or σ2
0 are basedon part (b) (and Proposition 1 and Theorem 3), it is sufficient to establish part (b) when π = s2.To this end, we shall prove that for each t and π, ∂F ′t (bt(π),π)
∂π < 0, from which the result will followby direct application of Lemma 1. First, we examine the case when bt(π) ≤ 1. Since
F ′t(bt(π), π) =
∆0π
π + tσ20
− λbt(π)π +γ − β − αγ
αγc′(bt(π)) = 0, (61)
it follows that
π∂F ′
t(bt(π), π)∂π
=∆0tσ
20π
(π + tσ20)2
− λbt(π)π
=∆0tσ
20π
(π + tσ20)2
− [ ∆0π
π + tσ20
+γ − β − αγ
αγc′(bt(π))
]by (61)
= − [ ∆0π2
(π + tσ20)2
+γ − β − αγ
αγc′(bt(π))
]
41
which is less than zero since bt(π) ≤ 1 implies that c′(bt(π)) ≥ 0. (The condition ∆0 ≤ 4p is notrequired in this case.) Now consider the case when bt(π) ≥ 1. Here,
∂F ′t(bt(π), π)
∂π=
∆0tσ20
(π + tσ20)2
− λbt(π) <4λtσ2
0π
(π + tσ20)2
− λ since bt(π) ≥ 1 and ∆0 ≤ 4p = 4λπ
= − λ(π − tσ2
0)2
(π + tσ20)2
≤ 0.
The proof is complete.
Proof of Theorem 6.
The proofs of parts (a) and (b) are identical to the proofs of Theorem 5(a, b). The proof of part (c)follows the same arguments given in the proof of Theorem 5(d), except that here part (a) impliesthat the path b(π1) lies strictly below the path b(π2).
Proof of Proposition 4.
Let
φ(t, µt, τ) := EV Ct
[ ∫ τ
t(F ∗
v − lv + Θ)dv]
(62)
denote the VC’s continuation value function at date t given her current project assessment µt anda given (possibly sub-optimal) stopping time τ . In the above, F ∗
v satisfies (18). The VC’s optimalcontinuation value function is
φ∗(t, µt) := supτ≥t
φ(t, µt, τ), (63)
By standard dynamic programming arguments, the optimal termination time τ∗ (if it exists) mustsolve (63) for any t ∈ [0, T ] and µt ∈ (−∞,∞). Further, the VC continues the project at any date t
and project assessment µt if and only if φ∗(t, µt) > 0. The proof proceeds by showing that φ∗(t, ·) ismonotonic (non-decreasing) and lower semi-continuous. It then follows that at each date t ∈ [0, T ]there exists a trigger µ∗t such that the VC continues the project if and only if µt > µ∗t .
We prove the monotonicity and lower semi-continuity of φ∗(t, ·) by considering the sequence ofdiscrete stopping time problems in which for each fixed positive integer N the VC is constrained toterminate the project only at the discrete set of times 0, T
2N , . . . , (2N−1)T2N , T. We show that the
VC’s optimal value functions φ∗N (t, ·) in the discrete problems are continuous and monotonic. Wethen use a convergence argument to show that φ∗(t, ·) is lower semi-continuous and monotonic.
Pick a positive integer N . We use backward induction to show continuity and monotonicity ofφ∗N (t, ·). To establish continuity we further show there exist positive constants κ1
t , κ2t such that
φ∗N (t, µt) ≤ κ1t + κ2
t maxµt, 0. (64)
42
The assertions of continuity, monotonicity and (64) are trivial at date T since φ∗N (T, ·) ≡ 0. Supposethat the assertion is true for t ∈ [t′ + 1
2N , . . . , T ]. We will establish that the assertion is true fort ∈ [t′, t′ + 1
2N ). Consider first any t ∈ (t′, t′ + 12N ]. By the dynamic programming principle,
φ∗N (t, µt) = Et
[ ∫ t′+ 1
2N
t(F ∗
v − lv + Θ)dv + φ∗N (t′ +1
2N, µt′+ 1
2N)]
. (65)
Since F ∗v is bounded and deterministic by (18), (4), (58) and (13) and Θ is normally distributed,
EV C0
[ ∫ T
0((F ∗
v )2 + (lv)2 + (Θ)2)dv]
< ∞. (66)
We can therefore apply Fubini’s theorem to conclude that
φ∗N (t, µt) =[ ∫ t′+ 1
2N
t(F ∗
v − lv + µt)dv + Etφ∗N (t′ +
12N
, µt′+ 1
2N)]
. (67)
We first establish monotonicity of φ∗N (t, ·). The integral on the right-hand side of (67) obvi-ously increases with µt; it remains to show the expectation on the right-hand side of (67) is alsomonotonic in µt. A bit of algebra applied to (2) and (3) shows that µt′+ 1
2N∼ N(µt, σ
2) is normallydistributed. Further, µt′+ 1
2Nmay be expressed in the form ft(µt, Z) where Z ∼ N(0, 1) and ft(·, ·)
is an increasing, linear function of its arguments. The monotonicity of Etφ∗N (t′+ 1
2N , ·) now followsfrom
Etφ∗N (t′ +
12N
, µt′+ 1
2N) = Eφ∗N (t′ +
12N
, ft(µt, Z)), (68)
since the expectation on the right-hand side of (68) is taken with respect to the standard normaldensity, which is independent of the problem parameters, and since both ft(·, z) and φ∗N (t′ + 1
2N , ·)are monotonic in µt (the latter by the inductive assumption).
Next we show continuity of φ∗N (t, ·). Once again, this property obviously holds for the integralon the right-hand side of (67); it remains to show the expectation on the right-hand side of (67)is also continuous in µt. This result will follow from identity (68) if the limit and expectationoperators may be interchanged, since both ft(·, z) and φ∗N (t′ + 1
2N , ·) are continuous in µt (thelatter by the inductive assumption). By the inductive assumption (64) the function φ∗N (t′ + 1
2N , ·)is bounded above by a positive function whose expectation
E[κ1
t′+ 1
2N+ κ2
t′+ 1
2Nmaxµt′+ 1
2N, 0
]= κ1
t′+ 1
2N+ κ2
t′+ 1
2N σt√
2πe−1/2(
µtσt
)2 + µtP (Z > −µt
σt) (69)
is finite, and thus the interchange is justified by the dominated convergence theorem.To complete the inductive argument we must show that (64) holds for t. The integral on the
right-hand side of (67) is bounded above by (t′ + 12N − t)(F ∗
0 + max(µt, 0))—recall the F ∗t decrease
with t. Since (64) holds for t = t′+ 12N , the inequality (69) shows that the expectation on the right-
hand side of (67) is bounded above by (κ1t′+ 1
2N
+ κ2t′+ 1
2N
σt√2π
) + κ2t′+ 1
2N
max(µt, 0). It is therefore
possible to define positive constants κ1t , κ2
t such that (64) holds for t, as required.
43
Finally, we must establish the inductive step for t = t′. We have
φ∗N (t, µt) = max[0,
[ ∫ t′+ 1
2N
t′(F ∗
v − lv + µt′)dv + Et′φ∗N (t′ +
12N
, µt′+ 1
2N)]]
, (70)
where (70) differs from (67) because the VC can terminate at date t′. It should be clear that theprevious arguments still apply, and hence the inductive step is established.
Because 0, 12N , . . . , (2N−1)T
2N , T ⊂ 0, 12N′ , . . . ,
(2N′−1)T
2N′ , T for all N < N ′, it follows thatφ∗N (t, µt) ≤ φ∗N ′(t, µt). For each (t, µt) ∈ [0, T ]× (−∞,∞) we may therefore define
φ(t, µt) := limN→∞
φ∗N (t, µt) . (71)
We claim that φ = φ∗. Fix (t, µt) ∈ [0, T ] × (−∞,∞). Since φ∗(t, µt) ≥ φ∗N (t, µt) for all N ,φ∗(t, µt) ≥ φ(t, µt). Suppose that φ∗(t, µt) > φ(t, µt). Choose any ε < (φ∗(t, µt) − φ(t, µt))/2.There exists a stopping time τ ε such that φ(t, µt) < φ∗(t, µt)− ε < φ(t, µt, τ
ε) where φ(t, µt, τε) is
defined in (62). Define the stopping time τ εN = i
2N 1 i
2N <τε< i+1
2N . There exists N sufficiently large
such that φ(t, µt, τεN ) > φ(t, µt, τ
ε) − ε. It follows that φ(t, µt) < φ(t, µt, τεN ). By definition of the
function φ∗N (t, µt), however, φ(t, µt, τεN ) ≤ φ∗N (t, µt) ≤ φ(t, µt), which is a contradiction. Hence,
φ(t, µt) = φ∗(t, µt). The monotonicity of φ∗(t, ·) easily follows from the monotonicity of φ∗N (t, ·)and the fact that φ∗(t, µt) = limN→∞ φ∗N (t, µt). The lower semi-continuity of φ∗(t, ·) follows fromthe fact that the supremum of continuous functions is lower semi-continuous.
Proof of Proposition 5.
The controllable rate function Ft(·) (17) is an increasing function of ∆0, which implies that F ∗t
is also a increasing function of ∆0. One may proceed exactly as in the proof of Proposition 4 toestablish that each CVt(·) is a pointwise increasing function of ∆0. Consequently, the trigger valueswill decrease. Since a change in ∆0 has no effect on the sample paths, part (a) follows. The proofof (b) is the same, except that each Ft(·) is now a decreasing function of either λ or k, and thusthe trigger values will increase.
44
On-Line Appendix C
Numerical Implementation of the Model
By Proposition 4, the relationship is terminated at date i if and only if µV Ci is less than a trigger µ∗i .
In our numerical implementation, therefore, we focus on directly modeling the evolution of µV Ci ,
which is hereafter denoted by µi to simplify the notation. In the first stage of our implementation,we approximate the evolution of µi using a discrete lattice and derive the triggers µ∗i that determinewhether the relationship is continued or terminated in each period. In the second stage, given thetriggers obtained from the first step, we use Monte Carlo simulation to model the evolution of µi
and derive the key output variables of interest, e.g., the duration of the relationship, the “rationalexpectations” value of the firm, and the continuation value of the VC.
Lattice design
We use a lattice to approximate the possible evolution of µi, i = 0, 1, 2, . . . , T − 1, which is givenby (3). Gompers (1995) finds that the average time between investments for different investmentstages is 1.09 years. We, therefore, set the length of each period δt to one year. Sahlman (1990)provides empirical evidence there are at most 8 investment stages. We, therefore, choose a finitetime horizon T = 10 in our numerical implementation.
At date 0 the project quality is given by µ0. Let n(i) denote the number of states at date i > 0and let µi,j denote the firm’s quality at the jth state at date i, j = 1, . . . , n(i). The lattice is designedso that the minimal and maximal states at date i, µi,1 and µi,n(i), are κ standard deviations belowand above the minimal and maximal states at date i− 1, respectively. More precisely,
µi,n(i) = µi−1,n(i−1) + κσµi−1 (72)
andµi,1 = µi−1,1 − κσµ
i−1 (73)
In (72) and (73), σµi−1 is the standard deviation of the evolution of the assessment of project quality
over the period (i− 1, i). This can be derived from equation (3) as
σµi−1 =
σ2i−1√
s2 + σ2i−1
. (74)
The values for the remaining n(i)−2 states are equally spaced between the minimum and maximumstates. The number of states in the lattice increases linearly from period to period. That is,n(i) = Mi for i > 0. The value of M is set to 25 and the value of κ is set to 2.5.9
9We found that an increase in M or κ or both did not change the optimal trigger values (to within a 3% tolerance),and that the probability of the firm surviving to the 10th period was zero.
45
The VC’s Continuation Value and the Termination Triggers
Let CVi,j denote the VC’s continuation value at state µi,j . At date T − 1 the continuation value isindependent of the future and is given by CVT−1,j = µT−1,j + F ∗
T−1 − lT−1 for each state µT−1,j ,j = 1, 2, . . . , n(T − 1). At earlier dates the continuation values are given by:
CVi,j = µi,j + F ∗i − li +
n(i+1)∑
k=1
pi+1,ki,j max(CVi+1,k, 0) . (75)
In the above, pi+1,ki,j denotes the probability that the assessment of project quality will transition
from state µi,j at date i to state µi+1,k at date i + 1.10
Starting from the last investment period T and working backwards through time we use dy-namic programming to compute the continuation values for all states and dates. Since the truecontinuation value function is continuous and increasing, we complete the approximation to CVi(·)by linear interpolation. We then determine the optimal trigger µ∗i , which solves CVi(µ∗i ) = 0.
In the second stage, given the termination triggers, µ∗i , we directly model the evolution of(3) using Monte Carlo simulation to compute the various economic statistics of interest. We run50, 000 simulations; the key economic statistics that we derive do not change by more than 1% ifthe number of simulation runs is increased beyond 50, 000.
Estimation of Indirect Parameters
We use the simulated method of moments to estimate the indirect (or the “deep” structural)parameters of the model. Let O denote the vector of the 11 aggregate statistics (see Table 2) weare trying to match.
Phase I: Estimation of Indirect Parameter Values
We simulate a large number N of firms and fix this simulated sample. The number N is chosenlarge enough to minimize simulation errors. The simulation error is negligible when N = 50, 000.For a given candidate indirect parameter vector π, we compute the vector V of simulated values ofthe 11 statistics. Let di(Vi, Oi) = Vi−Oi denote the difference between the simulated and observedvalues of ith statistic and d denote the vector of differences. Define
f(π) := dT Σd. (76)10If µi+1,k is within ±κσµ
i from µi,j , we set
pi+1,ki,j := Φ
[(1
2(µi+1,k + µi+1,k+1)− µi)
1
σµi
]− Φ
[(1
2(µi+1,k + µi+1,k−1)− µi)
1
σµi
],
where Φ(·) denotes the cdf of the standard normal distribution. Otherwise, the transition probability is zero.
46
In (76) the matrix Σ is diagonal. The vector π∗ of parameter estimates solves
π∗ = arg minπ
f(π). (77)
Phase II: Bootstrapping to Determine Confidence Intervals for π∗
We use parametric bootstrapping to determine the confidence intervals for the estimated parameters(see Davison and Hinkley, 1997). This method consists of two steps:
Step 1: Generation of bootstrapped statistics.We fix the parameter vector π = π∗. We generate X samples of NC firms denoted as (S1, . . . ,SX).
We set NC = 7765 to incorporate the fact that the 11 Cochrane statistics are calculated from asample of 7765 firms. We use samples Si to compute the ith vector of the 11 statistics, respec-tively. In this manner, we obtain a set of X vectors of “bootstrapped” Cochrane statistics Vj ,j = 1, 2, . . . , X.
Step 2: Repeat Phase I for each bootstrapped vector of statisticsWe replace the vector O of actual values of the statistics of Phase I with the vector Vj of
bootstrapped values. We solve (76) and (77) to obtain a set of X “bootstrapped” estimates ofthe indirect parameter vector (π∗1, . . . , π
∗X). We use these vectors to obtain standard errors for the
estimated parameters π∗.
Phase III: Computation of Output Statistics and Confidence Intervals
We use π∗ as the baseline vector in our numerical analysis. To compute the confidence intervals forall output statistics of interest when parameters are set to their baseline values, we compute thevalues of the statistics for each of the X bootstrapped samples described above with the parametervector set to the corresponding vector π∗i , i = 1, 2, . . . , X, determined above. In our “comparativestatics” analyses, where one or more parameters are varied while keeping the other parametersat their baseline values, we apply the same percentage changes to the parameter values in thebootstrapped collection π∗i , i = 1, 2, . . . , X, to compute confidence intervals. For example, if weanalyze the effect of varying the degree of asymmetry in beliefs ∆0 by 10% from its baseline valueon some output statistic X, we determine the confidence interval for X by computing its values forthe bootstrapped samples where ∆0 is varied by 10% from its corresponding baseline value for thebootstrapped sample.
47
Table 1: Baseline Parameter ValuesThe table displays the baseline parameter values obtained by the model calibration. The point estimates and their
standard errors are shown in the second and third rows, respectively. The standard errors are obtained by generating
bootstrapped samples using the model as described in Appendix C.
Technology Parameters Belief Parameters Preference Parameters
A α l1 l2 µV C0 ∆0 s σ0 λ k γ
0.6834 0.3212 0.0434 2.1726 0.1128 0.5039 0.7308 0.4401 1.5518 0.0480 4.9614(0.001) (0.005) (0.001) (0.021) (0.002) (0.015) (0.113) (0.083) (0.034) (0.001) (0.077)
Table 2: Predicted and Observed StatisticsThe table displays the observed values of the statistics used to calibrate the model, their predicted values from the
model, and the standard deviations of the statistics obtained by generating bootstrapped samples using the model
as described in Appendix C. The first row shows the round by round returns (RTRR) and standard deviations of
VC projects, the mean overall return (Mean All), standard deviation (Stdev All), and mean duration (Duration All)
of VC projects reported in Cochrane (2005, Table 4). The second row shows the model’s predictions, and the third
shows the corresponding standard errors.
Standard Deviation of RTRR Standard Deviation of RTRR Mean StdDev DurationRd 1 Rd 2 Rd 3 Rd 4 Rd 1 Rd 2 Rd 3 Rd 4 All All
0.26 0.20 0.15 0.09 0.90 0.83 0.77 0.84 0.20 0.71 2.100.26 0.19 0.15 0.09 0.86 0.84 0.82 0.81 0.21 0.71 2.08
(0.008) (0.008) (0.010) (0.020) (0.012) (0.012) (0.013) (0.022) (0.014) (0.006) (0.012)
Table 3: Project Value, VC Value, Duration, and Survival ProbabilitiesThe table shows the Project Value, VC Value, Expected Total Investment (Inv), Expected Duration, and the proba-
bilities of termination in successive round (p∗1, p∗2, p∗3, p∗4) in the Actual Scenario with asymmetric beliefs and agency
conflicts, the Symmetric benchmark scenario in which beliefs are symmetric, and the No Agency benchmark scenario
in which beliefs are risk attitudes are symmetric.
Agency Project VC Duration p∗1 p∗2 p∗3 p∗4Scenario Value Value
Actual 0.954 0.385 2.080 0.250 0.455 0.259 0.035Symmetric 0.960 0.177 1.834 0.401 0.388 0.188 0.023No Agency 1.014 0.288 1.961 0.330 0.411 0.226 0.032
Table 4: Contract Parameter Values (first four periods)The table shows the contractual parameters—the EN’s pay-performance sensitivities and the VC’s investments—inthe first four rounds in the Actual Scenario with asymmetric beliefs and agency conflicts, the Symmetric benchmarkscenario in which beliefs are symmetric, and the No Agency benchmark scenario in which beliefs are risk attitudesare symmetric.
Agency b∗1 b∗2 b∗3 b∗4 c∗1 c∗2 c∗3 c∗4Scenario
Actual 0.477 0.397 0.348 0.315 0.097 0.095 0.093 0.092Symmetric 0.141 0.147 0.150 0.152 0.083 0.083 0.083 0.084No Agency 1.000 1.000 1.000 1.000 0.101 0.101 0.101 0.101
48
Figure 6: Variations of Project Value and VC Value with Transient Risk, Intrinsic Risk, and Degree of Asymmetry in Beliefs
a) Effects of Transient Risk σ0
0
0.2
0.4
0.6
0.8
1
1.2
-100 -80 -60 -40 -20 0 20 40 60 80 100Percent Deviation
V-Project V-VC
b) Effects of Intrinsic Risk
s
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
-100 -80 -60 -40 -20 0 20 40 60 80 100Percent Deviation
V-Project
V-VC
c) Effects of Degree of Asymmetry in Beliefs
Δ0
0
0.2
0.4
0.6
0.8
1
-100 -80 -60 -40 -20 0 20 40 60 80 100Percent Deviation
V-Project V-VC
Figure 7: Variations of Implied Discount Rates (IDRs) with Transient Risk, Intrinsic Risk, and Degree of Asymmetry in Beliefs
Implied Discount Rate
0
0.2
0.4
0.6
0.8
-100 -80 -60 -40 -20 0 20 40 60 80 100Percent Deviation from Baseline Values
sigma s_risk
Delta_0
top related