Discounting, value and dollar beta matching in flexible cashflow

Discounting, value and dollar beta matching
in flexible cashflow systems.
Steinar Ekern, Mark B. Shackleton and Sigbjørn Sødal
Autumn 2014
Abstract
Discounting, value and dollar beta matching in flexible cashflow
systems.
The flexibility to switch from one cashflow to another has option value.
We develop the betas of these option values and use them in first order
conditions for optimal exercise. Because option functions smooth paste to
their targets, along with their values at exercise the dollar weighted betas of
options must match too.
For operational flexibility which allows switching between distinct cashflow modes, we use a discount function representation to relate an option’s
value at creation to its end value and to derive option betas.
Using three sets of equations, a system is solved for all option and investment sums (strikes) as a function of given investment thresholds and the
betas there. Essential option values are placed in vectors and the investment
network (graph) is captured by a matrix of discount factors that allow large
structures with many operating modes to be tracked.
C61, G31. Key Words; Value matching, smooth pasting, dollar beta and
rate of return equalisation, cashflow networks and bipartite directed graphs.
1
Introduction
When a financial option is exercised, tradeable quantities are exchanged; for
instance upon exercise a call option’s value coupled with the strike price is
swapped for a stock. This is a simple example of the benefit of flexibility
which allows the present value of one cash flow (the certain interest on a
strike price) to be converted into another (the uncertain dividend flow on a
stock).1
This article follows those in the literature on management of real assets,
where there is flexibility to switch between operational cashflow types. Many
articles have made the analogy between operational and financial options,
where capital investment including running costs, is viewed as a strike price
and operational value a stock price.2 These techniques are employed to
evaluate capacity and other investment decisions.3
Although less has been written on the acquisition or creation4 of such
option flexibility, their worth is also based on an exercise condition at investment. Managers maximise flexibility value by ensuring optimal timing for
these investment decisions. After an investment, if further flexibility remains
(i.e. the use of one option creates another) optimal timing should be driven
by maximisation of the joint value of both initial and subsequent options.
When an option is exercised and an investment transition made, we therefore recognise that flexibility is unlikely to end and this article accounts for
the simultaneous application and creation of operational options. Furthermore these follow-on options influence the timing of prior exercise; a decision
to proceed can not be taken without anticipating follow on flexibility and
its optimal decision too. Our valuation procedure tracks sequential flexibility through operating modes, requiring the values at optimal investment
1
See Black and Scholes [4], Merton [17]; also Cox, Ross and Rubinstein [7] for the risk
neutral valuation of options.
2
See Myers [18], Brennan and Schwartz [5], McDonald and Siegel [16] and the texts of
Dixit and Pindyck [10] and Trigeorgis [24]. Baldwin and Clark [3] with Gamba and Fusari
[14] motivate and value the real options of project design using six principles of modularity
(splitting, substitution, augmenting, excluding, inversion and porting). Latterly these are
valued using multidimensional stochastic processes and methods.
3
Including capacity (Pindyck [19]), land (Capozza and Li [6]), costly reversibility (Abel
and Eberly [2] and Eberly and van Mieghem [12]), marginal cost of capital (Abel, Dixit,
Eberly and Pindyck [1]), information (Lambrecht and Perraudin [15]), capital structure
and debt valuation (Sarkar and Zapatero [20]).
4
Simple calls (and puts where the exchange is reversed) in financial markets are sold
by option traders to clients for a premium or fee which is designed to cover the hedging
strategy and some profit. In return the client enjoys the financial flexibility described.
Operational or real options are not explicitly purchased in this manner but are embedded
in ownership of any physical asset.
2
or divestment times, i.e. when the options are applied and created, to be
tracked.
For an option’s value to be maximised by choice of its decision threshold
and time, so called smooth pasting (or tangency) must also hold. At an
optimal threshold, the value functions of the option and that of the net
payoff must converge (value matching) and at that transition point, do so
tangentially (smooth pasting). That is the two value functions must meet
and join with equal first derivatives there.
When flexibility is exercised at a decision threshold, we make use of the
intuition embedded in the first order condition. Take for example an American style call option on a stock (i.e. one that can be exercised early). At the
critical threshold that represents the optimal exercise point, the derivatives
(slope of functions with respect to the stock price) for both the option and
its payoff must match. The first derivative of an option price with respect to
its stock is known as delta, so this condition says that upon exercise a call
option should have a delta of one (unity being the slope of a call).
Options also have an elasticity, defined as this delta (or slope) multiplied
by the ratio of the stock price to the option price. Although similar to a
profit mark-up (as shown by Dixit, Pindyck and Sødal [11]), we interpret
the option elasticity as a (CAPM) beta relative to its underlying investment.
For instance a call option that is exercised in the money at a critical price
to strike ratio of two, will have a relative beta of two.5 Because it contains
a 2:1 levered position, this means that the local rate of return (or drift) is
double that of the underlying. Thus the option beta (or local rate of return
in Shackleton and Sødal [21]) can be used as a metric in optimal exercise
timing.
When the use of one option generates the creation of another, we apply
value and weighted beta matching to account for their interactions at exercise. In addition to equations that track the value before and after each
transition, this is done by representing first order conditions including the
beta of options weighted by the dollar value of the options.
The other element required to complete our analysis, is a discount factor
representation of option values.6 This provides both the means to determine
betas and also to link option values that are separated by threshold and time.
Section two gives examples of how dollar beta matching is useful in the
early exercise of American style put and call options, both individually and
when they interact in an operational setting. Section three incorporates
5
In the next section a $4 stock price threshold and an option value of $2 commands $4
of stock with $2 of borrowing implicit in the option hedge; this has a relative beta of 2.
6
See Dixit, Pindyck and Sødal [11] and Sødal [23] for the discount factor approach.
3
interactions with a third threshold and mode, using geometric Brownian to
illustrate. Section four uses matrix algebra to solve these systems. Section
five introduces a fourth threshold and solves a more advanced problem with
two way switching. Sections six and seven discuss conditions for maxima and
conclusions.
2
Linking slopes and betas at thresholds
2.1
Stand alone options at individual thresholds
For an option  ( ) its sensitivity to underlying  is known as its delta,
∆ ( )  All assets, not just stocks with respect to the market, have a beta;
for options this quantity may be dynamic like delta. The beta (or elasticity)
( ), tracks the sensitivity of an option’s percentage change  ( ) ( )
to the return  of its stochastic underlying driver  (assumed to have
unit beta). The beta is closely linked to the delta and a third quantity shown
in (1).
∆ ( ) =
 ( )

( ) =
  ( )
 ( ) 
( ) ( ) = 
 ( )
(1)

The last quantity in (1) is the dollar beta, i.e. the regular beta scaled
by the dollar value  ( ) of the option itself; it measures the dollar impact
of the beta, i.e. weighted by its value and is equivalent to the delta scaled
by the underlying  . Since the delta at early exercise of an option has a
particular value (often 1 or −1), we use these values with knowledge of the
beta and dollar beta at early exercise to infer other unknowns.
To quantify the example in the introduction, take an American style
call with value7  ( ) on a stock  . The delta of this option is ∆ ( ) =
 ( ) ; multiplying this by the underlying price gives the dollar beta,
then dividing this by the option value yields its beta   ( ) =  ∆ ( )  ( ).
Note that an option’s scale plays a role in its dollar beta but not in its beta.
Suppose upon the stochastic price  reaching the fixed point 4 = $4 we
see this option being exercised; since the delta there is unity ∆ (4 ) = 1 the
dollar beta at this point is $4. We will assume that the location of thresholds
such as 4 are known (or observed).8
7
Option values and betas are labelled  ( ) and ( ) but have subscripts that distinguish their type,  for calls,  for puts and later  for idle,  for full. Apart from this
labelling, in the GBM case constants   are also used for their particular beta values.
8
This threshold 4 = 4 is the highest of all those considered; 1 = 1 the lowest, is
motivated next and interim 2 = 2 and 3 = 3 in following sections.
4
Suppose we also know that the beta   (4 ) at this threshold is 2; it might
have been derived from an uncertainty model or measured by regression.
From these facts we can infer  (4 ) = $2
¯
4
  ( ) ¯¯
4
= 
(2)
=
  (4 ) = 2 =
¯
 ( )   =4  (4 )
2
Typically it is the strike price that is a given and the exercise threshold
4 (4 ) that is determined as a function of 4 , but here treating the threshold
4 = $4 and option beta   (4 ) = 2 as given we intend to solve strike price
4 (4 ) (the reverse problem is a numerical task). At this critical underlying
price 4 , the call option value converts into a payoff. With a strike labelled
4 (because it is applied at 4 ) and from 4 = $4 we then can solve for
4 (4 ) = $2 along with  (4 ) = $2
 (4 ) = 4 − 4 = 2
  (4 ) (4 ) = 4 = 4
(3)
At this threshold, smooth pasting implies that the slope or delta of the
option and its payoff match so that the differential of the first line in (3) at 4
confirms this. However the second line of (3) is a dollar beta representation
(from (2)), where the beta   (4 ) has been applied on the left to  ( )
and the unit beta (of  ) to  on the right. At the point of exercise, these
equations say that the beta of  ( ) is double that of the underlying and
that the option ($4 of stock net of $2 borrowing) has twice the local rate of
return (or drift) of  in the CAPM world.
As well as a call option to capture revenue once potential value reaches
4  there is an analogous put. This operating argument for puts rests on
the flexibility to suspend operations from running to the idle condition. For
a stand alone American style put  ( ) we consider exercise at 1 = $1
Smooth pasting requires a delta at 1 of −1 and (from an uncertainty model
or empirically by regression) if we also know the put’s beta is −1, we can
infer the put value itself  (1 ) = $1 at exercise
¯
1
  ( ) ¯¯
1
=− 
(4)
=−
  (1 ) = −1 =
¯
 ( )   =1
 (1 )
1
This is consistent with an exercise price 1 = $2 for the put (labelled 1
because it occurs at 1 ; it could be different to 4 e.g. if   (1 ) were less
than −1) because the payoff and dollar beta (−1) conditions are given in two
lines of (5)
 (1 ) = 1 − 1 = 1
  (1 )  (1 ) = −1 = −1
5
(5)
The first line of (5) smooth pastes with a delta of ∆ (1 ) = −1 i.e. the
put payoff has unit negative slope. The negative beta indicates a put return
rate or local drift of the risk free rate less the market risk premium ($2 on
deposit with a $1 liability on the short hedge). The second line of (5) is
the dollar beta matching condition, and this is written beneath the value
matching condition to indicate that (given   (1 )  1 ) it is as important in
the determination of 1   (1 ) as the value condition (first line of 5).
Thus if the betas and thresholds of options are known or observable, they
can be used to infer option prices and strikes at exercise.9
2.2
Follow on options and two thresholds
As suggested, the put may not be financial but operational in nature and
therefore not stand alone. Now we represent the present value of revenues in
 and costs in  whilst option value    represent the flexibility value of
launch or closure decisions.
Upon its exercise into an idle state (when the operational value  is
sacrificed in return for retrieving fixed value 1 from a prior investment), it
is possible that the call option to relaunch may be acquired. The condition
at the put exercise point must then embrace the acquisition of  ( ).
On the left hand side of the next equation (6), the put value  ( ) is
added to the present value of perpetual revenue  less a present value cost,
1 =  − 1  where  is the perpetual cost rate of operations saved on
suspension and 1 is a one time and irrecoverable cost that diminishes these
savings. These sums are foregone upon suspension but a call option to relaunch  ( ) is acquired and this goes on the right of an equation which
matches value when  = 1 i.e.
 (1 ) + 1 − 1 =  (1 )
(6)
This differs from (5) because now more flexibility is present. Until the optimal
point  = 1 values on left and right are dynamic in  and equality only
holds at 1 . However up to and including that time, the slopes and betas
with respect to  can be calculated to portray proximity to the first order
condition through betas.
Smooth pasting is akin to examining slopes on either side of such value
equations. Differentiating equation (6) by  (in the limit as it reaches 1 )
and multiplying by  represents   ( ) = ( ) ( ) the dollar beta
9
These beta values, e.g.   = 2 and   = −1 later become fixed under GBM and
carried with constants   that match call and put options.
6
of  ( ) on either side of (6).10 Therefore as well as (6), if smooth pasting
holds, dollar (or value weighted) betas must match at  = 1 in (7)
  (1 )  (1 ) + 11 =   (1 ) (1 )
(7)
Now compared to (6), the put value  ( ) has been weighted by its beta
  ( ), the value of operations  (foregone) by a beta of 1 and the call value
 ( ) weighted by its beta   ( ) Since its beta is zero, 1 has disappeared.
For decisions with follow on flexibility, smooth pasting demands rate of
return equalisation across value including the option acquired. Optimal exercise of the put  requires that the total dollar betas must match at 1 including options acquired  , so this call has an influence through   ( ) ( )
at 1 .
Now suppose that this system were to be closed within the two thresholds
mentioned, 1  4 ; this standard hysteresis has been studied before (Dixit [9]).
We apply the same logic at 4 where the call is exercised (not as in (3)) but
including recovery of the put we have motivated for use at 1 . Value matching
is easy to monitor by inclusion of the put on the right hand side of (3), and
smooth pasting via the means of dollar beta matching just described (for
both options and the payoffs at the boundary, differentiation and  scaling)
 (4 ) =  (4 ) + 4 − 4
  (4 ) (4 ) =   (4 )  (4 ) + 14 
(8)
Now hysteresis equations (6-8) give four conditions for the six unknowns
 (4 )  (1 )  (4 )  (1 ) 4  1 as a function of known thresholds 4 
1 and assumed betas   (4 )   (4 )   (1 )   (1 ) Two more conditions
are required, i.e. six in total. The four conditions presented so far apply at
transitions; the two extra required equations link between these transitions.
They involve a relationship between  (1 ) &  (4 ) and another between
 (4 ) &  (1 )
For a given scale, the means to link the relative size of option values at
two different transition times (i.e. thresholds) is using discount functions.
Discount functions are option solutions with a standardised, unit payoff.
They also provide the means to determine   ( ) etc. and are described in
the next subsection.
10
Taking the beta of the each side of (6), would make sense for the right ( ( )) but not
for the left hand side ( ( )+ −). The dollar beta operator however produces the sum
of dollar components directly (i.e.   (1 )  (1 ) + 11 on the left of 7). This is also important if an option subject to knock out can attain a zero value (due to complementarity,
in section 5). The dollar beta operator   ( ) is robust to this situation.
7
In solving option problems, a stochastic process for  is chosen first and
then solutions are customised at boundaries that complete the problem. By
referring to dollar betas at thresholds 1 etc., for a given process the aim
is include beta specifications in boundary conditions (such as 7) alongside
those of value (i.e. jointly in 8).
By treating value and dollar beta matching as equally important, we place
all conditions in a simultaneous system which allows option values  (1 )
 (1 ) and corresponding 1  to be solved from 1    (1 )   (1 ) etc. This
is done by using standardised (unit payoff) functions to separate the form of
the option values  ( ) and  ( ) from their scales  (1 )  (1 ) which are
measurements at the boundary of  ( )  ( ).
The process chosen for  determines the form of functions  ( )  ( )
and these in turn fix   ( ) and   ( ). From these functional forms, the
boundary conditions require certain fixed values   (1 )    (1 ) alone and
for given 41 these can be treated as knowns in (8) etc. This means that the
remaining unknowns  (1 )  (1 ) and 1 etc. can be solved using linear
algebra.
2.3
Discount functions
The exercise time and life of each option depends on an unknown time taken
for its state variable  to travel from its current value to a given threshold.
For time homogeneous problems, valuation can occur using stochastic stopping time methods. Although the time taken for  to evolve from 1 to 4
is unknown, expectations can be formed with respect to  and its continuous
discount − where11  ( ) = 4 
In common with Dixit, Pindyck and Sødal [11], we consider the present
value of $1 paid at the boundary 4  From the perspective of  = 1 ¿ 4 =
 ( )  this threshold is far away in both terms of  and time  . Thus the
present value of the future sum is much less than $1 but increases12 toward
$1 as  approaches 4 . For several stochastic processes, the probability
distribution of the stopping time  can be evaluated and an expected discount
factor can be found analytically (see Table 1 in next section) but numerical
methods are also possible in other cases.
Excluding changes from cash flows (with present values 4 − 4 ), the
payoff to timing flexibility at a threshold is taken as the option value. Values
11
To separate them for subscripts indicating fixed thresholds, indices for stopping times
 () indicate a value of the uncertain variable  at time  These are usually omitted for
the initial (known) level  (0) which is abbreviated to  (similarly for ).
12
Note that if the boundary is above  then the discount factor increases with  (call)
but it decreases with  if the boundary to be hit is below (put).
8
at the end of an option’s life (e.g.  (4 ) or  (1 )) assume the role of
scaling constants and the functional form of the flexibility value is carried in
a multiplier which is labelled as a discount  ( 4 ) or  ( 1 ).
Having previously labelled the call  ( )  we relabel it  ( ) to indicate
the flexibility present when operations are in idle mode . The put  ( )
is relabelled13  ( ) since it is active when operations are full, hence  .
Discount functions    are labelled with   notation to associate put or
call with full or idle states.
Also since more than one call or put may present, it allows labels  
rather than   to identify the state in which the option is active. Furthermore a third state  ( ) is introduced later that contains an option that is
neither uniquely a put nor a call but a combination of  ( ) and  ( )
and which therefore cannot be labelled by  or .
2.4
Call option discount
For a unit payoff at the random time   conditional on current value 
the valuation of discount 0   ( 4 ) 6 1 is given by the risk neutral
expected14 , risk free discounted time  when  ( ) = 4
¯
£
¤
 ( 4 ) =  − ¯  ( ) = 4 
(9)
This fraction becomes 1 at the time when  reaches 4 and is stochastic up
to that point. Applied to a dollar payoff at the stopping time, from a prior
point (and time)  it gives the present value of that one off future cashflow
at an uncertain time  .
From the perspective of call  ( ) creation at 1  the discount function
has fixed value  (1  4 ) ; the first argument represents the initial time point
and the second the final one when the option is used. The first argument is
set to dynamic  when a beta formula is required but a prior threshold when
the call’s initial fraction of its final value is required (both shown in 10)
 ( ) =  ( 4 )  (4 )
 (1 ) =  (1  4 )  (4 ) 
(10)
In (10), the first interpretation gives the means to separate an option’s scale
 (4 ) from its dynamics  ( 4 ) whilst the second relates two option scaling values. Rather than one scaling constant, two are needed; one each in
13
Their betas   too are labelled with  
For a discrete outcome case, if there were risk neutral chances of 1/3 for the process
stopping at  = 1 1/3 of  = 10 and 1/3 of it never stopping ( → ∞) with risk free
 = 5% the expected discount factor is (−05 + −05 + 0)3 = 0519 Discount factors can
thus adapt to structures where stopping is never certain or where   [ ] is unbounded.
14
9
different value matching and dollar beta equations. However one is determined from the other using discounting (e.g. on the right in 10).
Although discounting links  (1 ) to  (4 )  it only provides one third of
the necessary conditions. Value equations (first line of 8) join option values
of different types at a threshold e.g.  (1 ) and  (1 ) and another set of
conditions apply there too. The dollar betas (in the second line of 8) must
match and these equations result from a derivative scaled by  (or a beta
and value), both of which come from the dynamic discount factor.
With option values portrayed via discounting, the end value  (4 ) can
be treated as a scaling constant. Therefore differentiating the option  ( )
or discount function  ( 4 ) (in  ) is equivalent and gives the same beta
function. Since boundary conditions include both value matching and smooth
pasting, from (1) the beta at 1 that is important to rate of return matching can be expressed in  ( 4 ) (first line of 11) and evaluated at certain
thresholds (in the second line where a similar result holds for   (4 ) (4 )).
2.5

 ( 4 )
  ( )
=
=   ( )
 ( ) 
 ( 4 )

¯
 ( 4 ) ¯¯

  (1 ) =
¯
 ( 4 )

 =1
(11)
Put option discount
The put depends on a discount factor that increases if  falls toward a lower
threshold 1   Its discount factor  ( 1 ) applies to $1 paid at a random
time occurring when  falls to 1
¯
£
¤
 ( 1 ) =  − ¯  ( ) = 1 
(12)
From this representation the put value function  ( ) is given as a dynamic
fraction of its final value or for  (4 ) a static fraction of  (1 )
 ( ) =  ( 1 )  (1 )
 (4 ) =  (4  1 )  (1 ) 
(13)
The put value and discount functions have negative slope and beta in  .
Since  ( 1 )   0 differentiating  ( ) by  and scaling by 
gives a negative beta; the beta at 1 is necessary for rate of return matching
at the point of put usage 1
  ( )

 ( 1 )
=
=   ( )
 ( ) 
 ( 1 )

¯
 ( 1 ) ¯¯


  (1 ) =
¯
 ( 1 )

 =1
10
(14)
To reiterate, discount functions fulfil two useful roles; first they capture the
dynamics of values and their betas and second they fix the ratios and scales
of options at the time of their creation compared to their use. Even though
discount factors do not appear in the value matching condition, their first
property is used at that instant in smooth pasting or dollar beta matching
(8).
Although defined by 4  1 the flexibility system described so far could
not have been solved without the discount factors; these provided both the
means to determine the betas and also to link beginning and end options.15
Before solving (in section four) we include another possible operating
state and intervention point 2 . In order to do this, a third cashflow mode is
included. Whilst this is easiest to incorporate using a Geometric Brownian
motion, the results of this paper also pertain for other stochastic processes.
The advantage of GBM is that the claims    ( ) are iso-elastic, i.e. the
betas are constant for all  . Even if betas are not constant but depend on 
the diffusion dynamics and the solution to its asset pricing equation provide
the means to calculate particular   (1 ) and   (1 ) etc.
3
Three threshold GBM example
So far we have not yet utilised a specific diffusion. We will do this now for a
stochastic flow  and value  that follows a Brownian motion  geometrically


=
= ( − )  + 


(15)
Under this risk neutral diffusion, if the revenue flow  has drift  −  (where
 is the risk free rate and  the yield)  =  follows the same diffusion.16
This is convenient for a third cashflow mode.
15
Equations (6, 7, 8, 8, 10, 13) represent classical hysteresis across two operational
modes, full  and idle . Two known thresholds 1  4  the fixed discount factors between
them  (4  1 )   (1  4 ) (from 11, 14) and the option betas at these points   (1 ) 
  (4 )    (1 )    (4 ) (which depend on the  process rather than option constants)
determine the values of  (4 )  (1 )   (4 )  (1 ) and the two strike prices 14 
The system has six unknowns and due to the discounting equations, six equations in total;
two value matching, two weighted beta matching and two discounting. However, rather
than stop here, the flexibility system is first expanded to include a third and later a fourth
state operating between overall hysteresis levels 1 and 4 .
16
 =
Z
0
∞
  [ ()] −  =
Z
∞
0
11
h 
i∞ 
−  = − −
= 


0
Claims such as discount factors ( ) that are contingent (on reaching
certain levels) must satisfy a Bellman equation. This condition is derived
from the risk neutral expectation of a local change generating a risk free
rate of return over an interval . Applying expectations to the Ito’s lemma
expansion17 , we require   [( )] = ( )
For a discount claim ( ) that depends on the proximity of  to a
boundary, an asset pricing equation generates option solutions  ( ) ∝  
and beta constants  =  (for the call) or  =  (for the put) given by
 ( )
1 2 2  2  ( )
 
−  ( )
+ ( − ) 
2
2
 s

µ
¶2
− 1
1 −
2
− 2 ±
  =
−
+ 2
2
2


2

0 =
(16)
This says that the discount factors  ( ) ∝   are iso—elastic claims,
either depending on a positive or negative power (beta)  =   1 or  =  
0 Their dollar beta is a constant multiple of their value i.e.   ( )  =
 ( ).
These claims must also satisfy no—bubble conditions ( ( ) → 0 as
 → 0 and  ( ) → 0 as  → ∞); these match  to the idle state  (the
call) and  (the put) to . Both discount factors reach a unit value, but at
their own boundary;  (4  4 ) = 1 and  (1  1 ) = 1
These complete the specification of standard unit discount factor claims
with their constant betas
 idle, call  6 4  ( 4 ) = (4 )
 full, put  > 1  ( 1 ) = (1 ) 
(17)
These forms appear in real options paper with GBM processes; they can be
combined to form other discount factors that satisfy boundary conditions
at both of 41 (in section 5, terminology    labels these discounts that
depend on    ).
3.1
Other processes and discount factors
The same procedure can be applied to other stochastic processes who’s discount factors are available; Table 1 shows standardised discount factors for
17
Unlike risk neutral, using physical or real world (CAPM) expectations   []
to determine local expected returns of  generates a beta dependent risk premium,
i.e.   [ ( )]  ( ) = ( +  ( ) ∗ risk premium)  However, for the isoelastic
GBM, discounting can occur using CAPM expecations (see Shackleton and Wojakowski
[22]).
12
RN diffusion 
( − ) 
+
Call/put discount functions
 (  4 ) = ( 4 ) :  6  4
 (  1 ) = ( 1 ) :  >  1
 (  4 ) = (−4 ) :  6  4
 (  1) = (−1 ) :  >  1


 (−2+2+22  2
2 )
1
 (  1 ) =  −2+2+22  2
(
1 2 )
 + 
 ( − ) 
+
 quadratic for  
2
 ( − 1) +
2
( − )  −  = 0
2 2
 +
2
 −  = 0
2
 ( − 1)
2
− −  = 0
Table 1: Discount factors for Geometric, Arithmetic and Mean Reverting
processes.
Arithmetic and Mean Reverting flows  along with the GBM claims just discussed. Whilst these processes have similar solutions   to a fundamental
quadratic, unlike GBM they are not iso-elastic (constant beta). However the
techniques developed here are robust for processes other than GBM (e.g. the
mean reversion process used in Sarkar and Zapatero [20]).
3.2
Power flow
For a different function of the cashflow driver  (which still follows (15)), we
now motivate the present value of a third cashflow mode. Before the full state
with cashflow  and value  is launched, suppose that there is operational
flexibility to engage with    the  power of underlying cashflow . This
might arise for instance when a test market is possible during which time the
economies to scale  are different to those in full operations 
For a GBM flow, an expectation of  () (a uncertain cashflow  years in
the future raised to power  (0    1)) is also proportional to the power
of the current value ; i.e.   [ () ] ∝   
Like that of the full flow ( = ), the perpetuity value for a power of
cashflows (proportional to   ) has a fixed dividend yield18 with respect to its
18
The yield  0 on a perpetuity of the power cashflow arises through
Z
∞



−
[ () ] 
 = 
0

Z
0
  0
∞
"
#∞
− 0 
1 2
−

(−1)+(−)−


(
)
 2
 = 
=
0
0
1
=      0 where  0 =  −  ( − ) −  2 ( − 1)
2
Note that 0    1 implies that  0 lies between  and  all are positive. The PV of
power flows      0 would revert to  for  = 1 or to a constant 1 for  = 0
Finally, note that the beta of the present values of the power of these flows with respect

to the driver  is  
 = . Although this does not affect the beta of the options
themselves, due to the presence of a cashflow with value    it is needed in smooth
13
flow    Before going to full value  , we thus compare a mode of operations
engaging in a cashflow with value   . Since this occurs with scales to return
 we label this power state gamma with subscript  Options whilst in this
mode are labelled  ( )
This mode is beneficial for  ¿ 1 but detrimental for  À 1 so the new
cashflow pattern and its present value are scaled to align at a value of  = $1
(i.e. when  = $1 then the present value of the power cashflow and full
cashflow are equal   =  at  = $1) Thus  = $1 is also a very rough
(without option value) indicator of the desired switching level, the actual
switching level is designated 2 so that the call from idle19 operates from 1
to 2 and the call within the new gamma state from 2 to 4 
3.3
Three policy points; 1  4 and 2
If 1 has already been encountered and switching out of the full state has
occurred, we assess the conditions under which is optimal to switch from
idle (zero cashflow but with non—zero call value) to cashflow  with value
  by incurring a PV cost of 2 (the strike, which will be solved for, must
be aligned to a perpetual cost rate   and a one time cost 2 ). We label
option values in the power state  ( ) operations continue in this mode
after  reaches 2 until 4 (when switching to full occurs  ( )). Thus the
key separation for  ( ) in section three is 2  4 .
To summarise, rather than waiting to go full at 4  we assume the first
call  ( ) only takes operations into the power mode described, (cashflow
  and value   ) at 2 . Having started this power mode at 2 we presume
 continues to 4 ( 2 ) whereupon it is possible to exercise a second call
into full production . At 4 this would cause   to be relinquished but
 to be gained when  ( ) is used to claim the full flow along with its put
(back to idle again). Thus compared to section 2, the call has been split into
two separate calls,  ( ) and  ( ) (hence the need for notation other than
 ( )).
Although we label the value of the second call option in the gamma
cashflow mode  ( ) this does not mean that  ( ) has dependency on 
( ( ) 6=   nor do options in this state depend on ). If the uncertainty
driver is the same  and   process (15) throughout then  ( ) will be
another straight call  ( ) with beta also given by  from (16); i.e. under
GBM a call on   still has the same beta as a call on  .20
pasting and dollar beta matching.
19
In this case, (10) needs revising to  (1 ) =  (1  2 ) (2 )
20
That is to say that claims  () can either be formulated as functions of  or (say 
where)  =    Under GBM,  ( ) and  ( ( )) generate asset pricing equations with
14
Using the discount framework, this extra call also has beta  and is used
to relate its dynamic or initial value to final  (4 ) through
 = 
 ( ) =  ( 4 )  (4 )
 (2 ) =  (2  4 )  (4 ) (18)
When the first call from  idle  ( ) is used to get into power mode at 2 ,
the present value of the power cashflow is acquired along with the second call
 ( ) (to go from power to full mode). If this operational transition requires
a present value cost 2 =   + 2 (to be solved) then the value matching
condition at 2 reflects use of the call from idle into an operational value net
of costs plus the option in the next (power) state 
 (2 ) =  (2 ) + 2 − 2 
(19)
In addition to the six conditions in standard hysteresis, note that another
two option values and another strike 2 have been introduced. However (18
& 19) gives two of the extra three conditions required maintain solvability.
Another condition is given by weighted beta matching at 2  this says that
not only must the values either side of (19) match but to do so smoothly i.e.
with dollar betas matching too. The betas are known because the option
function takes the same form in this region  as idle . Taking value matching equation (19) and applying the dollar beta operator (differentiate with
respect to  and scale by  ; alternatively evaluate the beta and apply that
to the option value) gives
 (2 ) =  (2 ) + 2 
(20)
Now we close the cycle in this section; allowing transfer from power 
mode to full  i.e. from present value of revenue   to  by incurring or
saving some cost/benefit 4 . Thus the present value cost 4 must allow for
the addition of  , the saving of   net of a another one time cost or
friction 4 ; the net cost being 4 = 4 +   −  
When the power call is used to exit the power mode  and enter the full
mode  this generates a value condition (21). This happens at  = 4 when
the  call exercises into full mode, upon which 4 (the present value of the
power flow) is lost (left hand side) along with  (4 ) but 4  the full present
value, is gained (right hand side)
 (4 ) + 4 =  (4 ) + 4 − 4 
(21)
0
betas   0 that are different, but scaled by  Their solutions  ( ) ∝   and  () ∝  
0
0
are equivalent because  =  0  ensures that   =    =    This is true for calls or
puts.
15
In equation (21), items on the left represent the power state, including
the option  and net present values of the cashflows there; on the right
the full state contains the next option  ( ) a put, net present values of
cashflows there less the one time friction 4 (this is embedded in 4 and
causes the transition to be irreversible). In (21), all fixed present values have
been consolidated on the right into one sum 4 . This equation too must
match by dollar beta (again where 4 disappears) and since   =  and
  =  we get (22)
 (4 ) + 4 =  (4 ) + 4 
(22)
The option  (4 ) is then valued as a function of its discounted payoff, the
discount factor requiring betas at 41
 (4 ) =  (4  1 ) (1 )
  = 
(23)
Now we have three value matching equations, three discountings, and
three weighted beta matchings.21 Given thresholds 1  2  4  present value
cashflows 4 etc. and discount factors  () (used twice),  () and betas,
the three value equations, three discounting equations and three weighted
beta equations can all be solved for the following variables  (1 )  (2 )
 (2 )  (4 )  (1 )  (4 ) and 1  2  4  These are most easily summarised after describing them in a network or graph and its associated matrix.
3.4
The three level problem
These option value can be placed in a graph or network as in Table 2. The
underlying value of full operations  is mapped onto the horizontal axis,
whilst the vertical axis increases with values  . In particular it is important to track the passage of  toward each of the three thresholds 421 which
are marked and the investment jumps 421 that occur there (upward when
calling, incurring a cost and downward when putting, realising a saving).
From the full state  (in the top row of the network) when  reaches
1 from above i.e. moving from the right to the left, full operations are
suspended with loss of present value revenue 1 =  1  but also with a
saving of present value of operating and transition costs 1 =   − 1 
In the idle state  (on the bottom row), no operational cashflow is present;
however there is the call option  ( ) there. This is gained at 1 with value
 (1 ) and is used at 2 where its value is  (2 )
21
From within equations (10 (as amended in footnote 19), 18, 23 for discounting), (7,
20, 22 for dollar betas) and (6, 19, 21 for value matching).
16
 (1 ) + 1
←
 (4 ) + 4
↑ −4
 (4 ) + 4
Full flow 
 (2 ) + 2 →
Power flow 
↑ −2
 (1 )
→
 (2 )
Idle no flow 
1 threshold
2 threshold
4 threshold
↓ +1
Table 2: Investment graph at three thresholds (horizontal) for three state
system (vertical, idle power full). Value matching at investment occurs vertically, diffusion and discounting occurs horizontally.
Due to required investment 2  exercise of the call option from idle, lifts
value upward. This represents the present value of running costs  in  the
gamma state as well as transition costs 2  i.e. 2 =   + 2 . In return
for sacrificing  (2 ) and committing this present value cost, the call in the
power state  (2 ) is gained along with the present value of operations in
the power mode 2 
This cycle completes with the power state (middle row); at 4 this second call option is used with value  (4 ). This is combined with a third
transition cost 4  which now comprises three items; first the cost saving of
stopping power operations   secondly the present value cost of engaging
in full operations −  and finally another one off transaction cost −4 so
that 4 =   + 4 −   (to be solved). Along with the acquisition of the
put option in the full  state,  (4 ) this completes the augmented hysteresis cycle in this section (Ekern [13] presents another extension to switching
hysteresis, one with a finite number of transitions that would admit matrix
solutions).
To reiterate, normally 421 and    would have been specified by the
characteristics of the assets and their flexibility, but here for given thresholds
421 and betas   we will solve them through 421 .
The total value at each stage is given by taking the value of these options
and adding the present value of operations in all states, i.e.  ( )  ( )+  
or  ( ) +  . The vertical dimension is augmented with quantities 421
which determine their separation (this is apparent in the Figure in Section 5
which displays solved and interim values for a four level system).
4
Solving with vectors and matrices
With many option constants to consider (this method can encompass arbitrarily large systems and is generalised to a fourth threshold in Section 5), it
17
is useful to collect similar items into vectors and then use matrix and vector
equations to represent their linkages.
Given the three discount equations, rather than using one common vector
of values (say V) a more sensible grouping is to separate option values into
one of two vectors U or W Option values at the beginning of their life (on the
|
right hand side of a value equation) are put into U = [ (4 )  (2 )  (1 )]
and22 those at the end of their life (left hand side of a value transition equa|
tion) W = [ (4 )  (2 )  (1 )] 
This allows the value of options at their birth to be expressed succinctly
as a matrix multiplication D of values at their use, i.e. the relationship
between beginning (left U in 24) of their life compared to the end (right W
in 24)
U ⎤ = ⎡
D
⎤ ⎡ W ⎤
 (4 )
 (4 )
0
0
 (4  1 )
⎦ ⎣  (2 ) ⎦ 
⎣  (2 ) ⎦ = ⎣  (2  4 )
0
0
 (1 )
0
 (1  2 )
 (1 )
0
(24)
⎡
Note that the call discount factor  () has been used twice, between
intervals 1  2 (idle ) and 2  4 (gamma ) whilst with only one put present
(in the full state ),  () appears once (between 4 and 1 ).
Two other vectors capture the present values of operational cash flows
before (i.e. preceding on the left of the value transition (25)) option exercise
|
Z = [4  0 1 ] and after (proceeding, i.e. on the right) the transition at
|
each level Y = [4  2  0] . These flows are treated as potentially perpetual
in nature but their life will depend upon the potential exercise and timing
of the next option (when terminated the loss of the flow is accounted by
deduction at the next exercise threshold).
Thus the present values of cost rates and option strikes are also grouped
into vectors, one for the present value of costs in the preceding state K =
|
[  0  ]  one for the present value of cost in the proceeding state
|
K = [  −  0] and finally one for the frictions borne at the exer|
cise time K =[4  2  1 ]  0 It is these final frictions that ensure that
each transition is irreversible (as elements of K go to zero then reversibility
could appear).
Collectively in (25), these vectors represent all three value matching equa22
The transpose of a vector is indicated by | .
18
tions, one line for each threshold
W ⎤ + ⎡ Z − K ⎤ = ⎡ U ⎤ + ⎡ Y − K − K ⎤
 (4 )
4 −  
 (4 )
4 −   − 4
⎣  (2 ) ⎦ + ⎣
⎦ = ⎣  (2 ) ⎦ + ⎣ 2 −   − 2 ⎦ 
0
1 −  
 (1 )
 (1 )
−1
(25)
This can be made more compact by grouping all fixed (zero beta) costs
into one vector containing 421 ; i.e. X = K + K −K to be subtracted
|
from value after exercise. This X = [4  2  1 ] has positive strike elements
for calls and negative for puts.
Furthermore subtracting Z from both sides means that the value of options in the preceding state W can be expressed as the sum of the new,
proceeding options U, plus a cash flow difference between modes (Y − Z)
less a net present value of costs −X Most succinctly we can use the expression for W with U = DW and inputs D Y Z X to solve for W as an
optimisation across 421
⎡
W = U+Y−Z−X
max W = [I − D]−1 [Y − Z − X] 
(26)
421
That is to say, if all costs and rates (K etc.) in X were given then
W = U + Y − Z − X can be used with U = DW to eliminate U (note that
I is the identity matrix, three by three in this case). Then the end of life
option values in W would be found by a maximisation problem in 421 . For
fixed X (and K’s) this is a search procedure in three dimensions 421 .
However if we assume 421 with the first order conditions, we can calculate W U X. Since vectors Z Y and matrix D are all given as functions of
421  taking these inputs as optimal, we can calculate U W X directly.
To do this the dollar betas of operational flows are also required; these are
|
|
given by β Z = [4  0 1 ] and β Y = [4  2  0] which are combined
|
into β Y − β  Z = [4 − 4  2  −1 ] and used in (27) for dollar beta
matching.23
Having Z on the right hand side and differentiating the items of (27) line
by line in  and then multiplying by  gives the dollar betas both sides
23
Although only the vector β Y − β Z is required, matrices β  β are given by
⎡
⎤
⎡
⎤
1 0 0
 0 0
β = ⎣ 0  0 ⎦  β = ⎣ 0 1 0 ⎦ 
0 0 1
0 0 1
19
of a value transition. This is succinctly expressed using two diagonal beta
matrices β  β  defined in (27)
β ⎤ ⎡ U
β ⎤ ⎡ W ⎤ = ⎡
⎤ + ⎡β Y − β Z⎤
 0 0
 (4 )
 0 0
 (4 )
4 − 4
⎣ 0  0 ⎦ ⎣  (2 ) ⎦ = ⎣ 0  0 ⎦ ⎣  (2 ) ⎦ + ⎣
⎦ 
2
0 0 
 (1 )
 (1 )
−1
0 0 
(27)
In order to separate the option constants in W U X from their (known)
betas, matrices β  β were used. These diagonal matrices operate on the
vectors of option values to weight them with their betas. They are evaluated
using the option beta at each threshold; the flexibility options are iso-elastic
so they contain either  or  ( and  are both calls with the same function
 and power   1 and  is a put with power   0 with factor on  ).
⎡
4.1
Solution
In particular, note that the beta matching and discounting equations do not
involve the transition costs X This is another reason why it is easier to solve
for those at the end. Direct elimination from (24 & 27) gives W U that are
consistent with smooth pasting i.e. dollar beta matching
W = [βW − βU D]−1 (β Y − β Z)
U = D [βW − βU D]−1 (β Y − β Z)
X = U − W + Y − Z
(28)
In the first two lines of (28), an inverse of a matrix combination is applied to a
beta weighted difference (in brackets). The expression Y − Z is a difference
of present values of revenues before and after each transition (one in each
row), however when weighted by their betas β Y − β Z it is akin to a
rate of return on this difference and thus acts like a flow in the definitions of
W U The matrix β −β D reflects the net beta of the change in flexibility
value, β for the immediate W and β D in anticipation of the demise of
U at the end of the next stopping time. Thus option solutions W U can be
thought of as aperiodic perpetuities on beta weighted cashflows β Y − β Z
with factor [βW − βU D]−1  That is to say that their benefit is linked to the
cashflow gain carried by β Y − β Z which comes along with random future
timing reflected in a perpetuity factor (matrix inverse).
As the system cycles through    episodes, the expected timing is carried though this inverse [β − β D]−1 . Although idle, power and full states
last for different lengths of time, the matrices carry the expected timings and
present value of each through the discount factors embedded in D.
20



 (1 ) + 1
↓ +1
 (1 )
1 thresh.
←  (2 ) + 2
↑ −2
→
 (2 )
2 thresh.
 (3 ) + 3
↓ +3
↔  (3 ) + 3
3 thresh.
←
 (4 ) + 4
↑ −4
→  (4 ) + 4
4 thresh.
Table 3: Investment graph with four thresholds (horizontal) and three states
(vertical). Value matching at investment occurs vertically, diffusion and discounting occurs horizontally.
This method can now be extended to encompass two way flexibility with
four thresholds.
5
Four thresholds & two way flexibility
In this section we illustrate a more complex investment scenario involving
the same interim (power) state  Now as well as the chance to invest further
if things go well there is also the chance to recoup and partially divest from
full by rejoining  rather than  if the situation worsens. This splits the
put into two, creating a high level put join  and a second lower level put
now embedded in  (which becomes more complex as a result). The network
of investment transitions, is given in Table 3 which is similar to Table 2
other than retreat (like advance) has two stages. The middle stage  has
reversion on the left (low  = 1 rather than the full) and the full state 
has reversion to the  state at an additional level 3 (upward moves occur
at even thresholds 24 and downward odd 13 ).24
Since the power state no longer has a unique direction of propagation
(up, i.e. a call  as in section 3), first we need to construct new appropriate
discounts, ones that can accommodate two different diffusion destinations.
This is done with    discount factors that are subject to knockout features (i.e. where a claim value of zero is attainable). These are combinations
of fundamental    the simple one way factors already used. These are
best explained using growth factors which we describe first.
24
Note that horizontal arrows ↔ indicate diffusions these take time to transit and involve
possible movement to left or right in  . Vertical arrows ↑ ↓ indicate either investment or
divestment and occur instantaneously at a threshold; these are not reversible other than
by further forward passage through the network and its graph.
21
5.1
Growth factors and their betas
It is more convenient from now on, to use compact notation for elements that
are static and non—dynamic, i.e. to use 12 rather than  (1  2 ) and 43
rather than  (4  3 ) also  4 rather than  (4 ) etc. However  ( 4 )
and  ( ) will be maintained for quantities that are dynamic in  and not
fixed at a threshold.
The two simplest options in Table 3 are the straight call (bottom row)
and put (top row) that can be discounted easily and whose dollar betas are
already known (these form part of the D matrix required in this section)
∙
¸ ∙
¸∙
¸
∙
¸ ∙
¸∙
¸
 4
43
 3
  4  4
43
 3
0
0
=
=

0
12
0
12
1
2
 1 1
2
The first (discount) matrix can be inverted to form a growth factor which
expresses end values  3  2 as a function of  4  1 at the beginning (this
forms part of a growth matrix G used in 42)
¸ ∙
¸∙
¸ ∙
¸∙
¸
∙
0
0
(43 )−1
 4
34
 4
 3
=
=
 (29)
0
21
2
1
1
0
(12 )−1
This naturally suggests that the inverse discount can be used as a growth
factor in relationships such as  3 = (43 )−1  4 = 34  4 where it is now
understood that the put growth factor 34  1 from 4 to 3 is the inverse of
the discount factor from 4 to 3  Thus the subscripts fixing the thresholds,
not only determine which thresholds are under scrutiny but also whether
 represents a discount or a growth.25 Due to the definition of the beta,
growth factors have the same absolute elasticity as their discounts, but with
a different sign.
5.2
Top & bottom    complementary discounts
For the values in Table 3 not yet identified, we require discount factors which
have a unit payment at one threshold but are subject to a knockout possibility, becoming strictly worthless at a second attainable threshold.
25
For GBM the call growth factor is 21 = (2 1 )  1 Not only do we assume
that (12 )−1 = 21  i.e. that inverse discount factors are growth factors with the same
functional form applied to a growth ratio 2 1 , but also implied are relationships which
chain discounts (or growths) together across intermediate thresholds i.e. 12 23 = 13
etc. This is possible due to concatenation of two stopping times; one from 1 to 2 and the
other from 2 to 3.
Note that    have different elasticities so these discount/growth and pass through
relationships only hold within common option type and 12 23 cannot be represented
as a single discount.
22
In this section, starting from  = 3 there are two conversion boundaries
of interest 4  3 and 1  3  For conversion at the top i.e. 4  exercise is
completed but at 1 flexibility at the top becomes worthless because another
conversion occurs there instead. The complementary option  used at the
bottom causes  to be knocked out for zero value. Since  is used at the
top of the flexible power state, we use notation , also three arguments are
required  ( 4  1 )  At the bottom  ( 1  4 ) is used; this represents a
discount factor that knocks in at the bottom but out at the top. Normalised
for unit and zero payoffs, in static notation these claims are subject to payoff
or boundary conditions.
¸ ∙
¸
∙
1 0
441 414
=
(30)
0 1
141 114
For dynamic  the first argument in  ( 4  1 ) is the current level of
the dynamic variable  the second a level (above or below  ) at which the
discount factor becomes one and the third an attainable level at which it becomes zero. Formally these claims are risk neutral expectations (conditioning
on current  ) but subject to an extra condition on acceptable paths
[ ( 4  1 )  ( 1  4 )]
(31)
¯
£ − ¯
¤

− ¯
¯
 ( ) = 4  min  1 
 ( ) = 1  max  4
=  
¸
∙
[ ( 4 )  ( 1 )]
1
−41

=
−14
1
1 − 14 41
We already have two claims  ( 4 ) and  ( 1 ) (straight call and put
discount factors without knockouts) that satisfy single conditions. For 4 >
 > 1 the claims  ( )  ( ) are fixed linear combinations of these
that reflect two pathways. For  the direct (or successful) path is via a call
from  to 4 but the indirect (and unsuccessful) path is from  to 1 under
the put multiplied by the path from 1 to 4 with the call; this must be
subtracted (see Darling and Siegert [8] for complementary stopping times,
similarly for  the direct put path is from  to 1 but the indirect call
path from  via 4 then a put from 1 to 4 must be subtracted). In their
valuation, both routes must be scaled (by a matrix determinant) to account
for the discounted round trip, i.e. call from 1 to 4 times put from 4 to
1 .
Equations in (31), say that a discount factor that pays off $1 at 4 but
knocks out and becomes worthless at 1 is a long position (with magnitude (1 − 14 41 )−1 ) in the call  ( ) and a short position (14 (1 −
14 41 )−1 ) in the put  ( )
23
As  increases, the value of the straight call increases and that of the
straight put decreases so  ( ) increases in value until it reaches 1 when
 = 4 . Conversely as  decreases, the call decreases and the put increases
its short contribution, thus  is monotonic in  . As  reaches the knockout
level 1  even though the straight call  still has some residual value (14 )
this is eliminated by the negative element from the straight put and the whole
claim  becomes worthless (for similar reasons,  ( ) is monotonic and
decreasing in  achieving 0,1 at 41 )
5.3
Dollar betas of  and 
Since the new discount factors are a linear combinations of the straight
put and call, their slopes at any point  are too. Rescaling the slopes
by  gives (33) the dollar weight of their betas   ( )  ( 4  1 ) and
  ( )  ( 1  4 )
[  ( )  ( 4  1 )   ( )  ( 1  4 )]
∙
¸
1
1
−41
 (32)
= [ ( 4 )  ( 1 )]
−14
1
1 − 14 41
Even under GBM, these are dynamic in  ; although the component betas
  are fixed, the weights  ( 4 ) and  ( 1 ) change with  and in
(32) the dollar beta of claims   ( ) are not constant weights of  and .
Furthermore, the top call with knock out  is more levered than the straight
call  having a higher slope and dollar beta ( has a more negative beta
than  ).
In order to value the two way flexible state  ( ) and its dollar beta,
we develop equations (37) and (39) next; these will need evaluation at all
thresholds 4321 .
5.4
Two way flexibility  ( )
Within the power state, knock out options    allow flexibility to have
one of two possible uses; at 1 a value of  (1 ) may be realised but at 4 a
different value  (4 ) will result (from a perspective within 1  4 the option
value  ( ) at  may be less than  (1 ) or  (4 )  i.e. it is a combination of
both put and call    through    ). From the perspective of a dynamic
point  from within this range, the current expected value of flexibility is a
probabilistic and discount weighted sum of two options, only one of which
will be attained
 ( ) =  ( 4  1 ) 4 +  ( 1  4 ) 1 
24
(33)
This dynamic version of the flex value in the power state carries dependency on  through  ( 4  1 ) and  ( 1  4 )  which are the dynamic
discount factors subject to knockouts. The payoffs 4  1 at 41 are static.
Having determined that the option in the power state has two possible
outcomes at two end thresholds, we can focus on the values at the beginning
of the option’s life i.e. 3  2 . Since these are static we also represent
them with non dynamic notation for items occurring at fixed thresholds, e.g.
 (2  4  1 ) = 241 etc.
We know that the  ( ) option is acquired at either of the thresholds 2 or
3  so can specialise the dynamic equation (33) to those two particular values
of  Using compressed notation the option value in the interior (3  2 ) of
the power state can be expressed as a function of the extreme values (4  1 )
∙
¸ ∙
¸∙
¸
3
341 314
4
=

(34)
2
241 214
1
For 241 in the second row, the first threshold in the subscripts is the level
there e.g.  = 2 , the second 4 the target or call level and the third e.g. 1
the level at which the flexibility dies or is knocked out. The complementary
option in the second row 214 does the opposite (put at 1 but knock out at
4).
In order to work out how 41 depend on 32 this last equation must be
inverted to produce growth factors comparable to (29). This is easiest to do
by decomposing the matrix (34) using the following
∙
¸ ∙
¸∙
¸
1
341 314
34 31
1
−41
=
 (35)
241 214
24 21
−14
1
1 − 14 41
The inverses of both matrices are required and applied in reverse order
∙
¸−1
¸
∙
1
1
41
1
−41
=
(36)
−14
1
1
1 − 14 41 14
¸−1
∙
¸
∙
1
21 −31
34 31

=
24 21
34 21 − 24 31 −24 34
Now 4  1 can be expressed in 2  3
∙
¸ ∙
¸∙
¸
¸∙
1
4
1
41
3
21 −31
=

1
14
1
−24 34
2 34 21 − 24 31
(37)
This can be simplified back to a similar form to the standard discounts subject to knock out. Multiplying the denominator in the determinant’s fraction
25
by a growth factor 43 12 (each row of the last equation must also be multiplied by 43 12 to compensate) and using discount and growth concatenation yields a round trip determinant between 32 (resting on 1 − 23 32 =
(34 21 − 24 31 ) 43 12 and the production of a new determinant).
Applying these and evaluating the matrix product yields the simplification
(38)
¸
∙
¸∙
¸
∙
1
43 − 42 23 42 − 43 32
3
4
=
(38)
1
2
1 − 23 32 13 − 12 23 12 − 13 32
∙
¸∙
¸
432 423
3
=

132 123
2
Thus not only are two way discount factors  easily expressed as a linear
combination of standard put and call discounts, but the inverse or growth
functions are too26 . The growth elements reflect the value at extreme points
41 as a function of starting from one or other of the interior points 32 , but
for these end values the weights are fixed as a function of the internal points
(and their round trip determinant (1 − 23 32 )−1 ). The resultant growth
matrix contains both positive and negative quantities; these result from the
long and short positions present in the knock out discount factors.
This formulation also holds true for dynamic  ( ) i.e. any  within
the region, from which the dynamic dollar beta   ( ) ( ) is easy to obtain
from (38) and apply to 43
∙
¸
∙
¸∙
¸
1
 4 4
43 − 42 23 42 − 43 32
3
=

 1 1
2
1 − 23 32 13 − 12 23 12 − 13 32
(39)
The weighted betas of 4  1 and 3  2 are necessary for smooth pasting and dollar beta matching, the inversion presented here is useful because
it allows both the discount and its corresponding growth matrix to be represent dollar betas for matching. Before doing this, the ladder network is
formalised.
5.5
A four level ladder network across three stages
We can now formalise the system with flexibility to ratchet up or down a
“ladder” with at least three investment states; idle  full  and intermediate
26
In the bottom left corner of the last matrix, 132 would normally imply (for a discount) achieving 3 from a start point of 1 without hitting 2  Because 1 cannot grow
to 3 without hitting 2  a similar and literal interpretation is hampered for the growth
factor by the relative levels which is why the growth matrix contains negative quantities.
26
power state  with elasticity  of an underlying flow.27 We use the same
linear algebra but vary the matrix and vector contents in order to examine
the extended flexibility setup from Table 3.
To recap, operational value is again nil in the idle state, varies with   in
the power state and full (power 1) in the highest state. i.e. non-flex values
varying with level 0   and  . Starting with the transition at the highest
threshold, from power to full, one control occurs at threshold 4 with a payoff
net of exercise cost of 4 − 4 − 4 . That is to say at the top threshold on
going to full, 4 is gained but 4 is lost along with incremental investment
cost 4 
Reversion can occur at 3 yielding 3 −3 +3 a reverse payoff including
a partial return of fixed investment cost 3  4 (this imposes a restriction
on 3  4 ). After this two outcomes are possible. Similar transitions occur
at 2 into the power state from the idle and 1 back to the idle from the
power. These last two have investment strikes of 2 and 1 (1  2 ) 
The value matching conditions within this system are shown in (40), with
contents specific to this section in Table 3
⎡ W ⎤ = ⎡ U
 4
4
⎢ 3
⎢  3 ⎥
⎢
⎢
⎥
⎣ 2 ⎦ = ⎣ 2
1
1
⎤ + ⎡ Y − Z
4 − 4
⎢ 3 − 3
⎥
⎥ + ⎢
⎣
⎦
2
−1
⎤ − ⎡
X
  −   + 4
⎢
⎥
⎥ − ⎢   −   + 3
⎣
⎦
  + 2
−  + 1
⎤
⎥
⎥
⎦
(40)
The network depicted in Table 3 is a mathematical graph, furthermore it is
bipartite and directed.28 Note that with two new possible outcomes from the
power state, branching of pathways is now possible.
From the power state, since reversion to the idle state is possible (at 1 )
as well as elevation to the full state (at 4 ), the discount matrix is populated
with six elements, and in particular two rows now contain complementary
discount factors with mutual exclusivity, i.e. conditional upon each other not
paying off. For example in row three (within the power state) the element
341 represents from the point of view of threshold 3  the discounted value
of the chance of paying a dollar at 4 knowing that if 1 is reached the
27
More ladder regions can be considered using different  1   2 and more regions etc. Note
that unless two thresholds converge, the up (even) and down (odd) rungs of this ladder
are different. However reversibility can be achieved by letting two thresholds become
arbitrarily close.
28
Bipartite (see Wilson [25]) in that investment nodes are either beginning U or end
W of state and directed in the sense that W maps forward (one to one) to U upon value
matching and smooth pasting/dollar beta. Beginning of state values U are linked to W
(one to two) by discounting but this also has a backward growth interpretation.
27
opportunity dies (zero value).
The second 314 in row two is complementary and pays off at 1 assuming 4 is not hit. Other discount factors are interpreted similarly but simple
43  12 are one way factors in full and idle states respectively. The detailed discounting equations U = DW are thus given in (42) which contains
equation (34) in rows two and three
D
⎡ U ⎤ = ⎡
0
43
 4
0
0
⎢ 341
⎢ 3 ⎥
0
0
314
⎢
⎢
⎥
⎣ 2 ⎦ = ⎣ 241
0
0
214
1
0
0
12
0
⎤ ⎡ W ⎤
4
⎥ ⎢  3 ⎥ 
⎥
⎥ ⎢
⎦ ⎣ 2 ⎦
1
G
⎡ W ⎤ = ⎡
0
432 423
4
0
⎢ 34
⎢  3 ⎥
0
0
0
⎢
⎢
⎥
⎣ 2 ⎦ = ⎣ 0
0
0
21
1
0
132 123
0
⎤ ⎡ U ⎤
 4
⎥ ⎢ 3 ⎥ 
⎥
⎥ ⎢
⎦ ⎣ 2 ⎦
1
(41)
We also summarise the inverse of D as G = D−1 (who’s elements, which
can be greater than one and positive or negative, are drawn from (29), (38)).
This is a matrix that values W as a function of U
5.6
(42)
Example
Now we have assembled two thirds of the equations required to solve the four
level system; in particular from the value matching condition (40), on the
right we can represent U as a discounted version of W and on the left we
can represent W as a discounted version of U This means that line by line
implicit differentiation and  scaling  () | =41 is carried in the D G
matrices alone. These dollar beta components are denoted with primes e.g.
D0 , indicating that they have been implicitly differentiated by  and scaled
by  at the boundary in each row.29 These comprise the matrices required
to form β  and β from first principles (rather than adhoc in (27))
GU = DW + Y − Z − X
⇒
G0 U = D0 W + (Y − Z)0
(43)
Finally to retrieve the beta matrices that we desire, the forward and back
projections can be unwound for U W to produce a dollar beta matching set
29
To form dollar betas, this says perform a line by line  scaled differentiation ()0 of
U = DW (and W = GU). Since D and D0 (G G0 ) have no diagonal elements the
operation is carried in the first matrix, i.e. U0 = (DW)0 = D0 W and dependence on 
in line  is carried in D etc. alone. These need rescaling by value to form betas.
28
of equations
G0 DW = D0 GU + (Y − Z)0
β W = β U + (β  Y − β  Z)
(44)
where30 β  = G0 D and β = D0 G and the dollar weighted, threshold
scaled payoff (Y − Z)0 = [4 − 4  3 − 3  2  −1 ]| is captured as
β Y − β Z31
Results for this system under GBM were evaluated32 for 4321 = 4 3 2 1
with elasticity/beta parameters    = 2 −1 05 Figure 1 plots the relationship between the flexibility values (on the vertical) and underlying 
value
Using one way factors, the system accommodated uncertain stopping (e.g.
in section 3, there was a chance that if  remained high the value of flexibility would become small because switching would cease). With two way
discount factors, this system also embraces the potential for different scenarios and branching. Not only is switching uncertain, but it is also not clear in
which direction switching from the power state will go; different sequences
i.e. 2,4,3,1 or 2,1,2,1 etc. are possible.
The advantages of this solution system are apparent. It has building
blocks in the sense that individual cashflow components and their intertwined
flexibilities can be placed individually within a system governed by a common framework and solution method (these could then be analysed using
modularity33 ). This is an important aspect of sequential investment. Secondly, although the example has used GBMs, other processes could be used
30
For the numerical example that
⎡
−1
0
0
⎢ 0 3263 −2842
β = ⎢
⎣ 0 1895 −2263
0
0
0
follows, the beta matrices
⎤
⎡
0
2048
⎢ 0
0 ⎥
⎥
β = ⎢
⎣ 0
0 ⎦
2
0190
31
are given by
0
−1
0
0
⎤
0 −0762
⎥
0
0
⎥
⎦
2
0
0 −1048
See footnote 27 to form the comparable four by fours β  β .
In this example 4321 = 4 3 2 1 were fixed, resultant discount growth and beta
|
matrices were evaluated and X = [4 : 1369 3 : −1359 2 : 1338 1 : −1304] calculated. If these do not match known or target policy costs 41 , 41 should be varied until
41 matches those required. Alternatively as mentioned earlier for input X arbitrary
−1
(non beta matched) values are given by W = [I − D] [Y − Z − X] when 41 = 41
can be varied numerically to find the turning point for W values. Either way the second
order conditions should be checked to ensure these values are maxima.
33
Baldwin and Clark (00) [3] and Gamba Fusari (09) [14] motivate and value modularity.
The design rests on six principles; splitting, substituting, augmenting, excluding, inverting,
porting and the implementation is via multivariate least squares Monte Carlo. These are
open to network or graphical investment approaches.
32
29
Investment ladder
Idle, power and full switching
6
5
P+Vf (full inc put)
P (full only)
4
P^0.5 + Vp (power + flex)
Total
value 3
incl.
flex
P^0.5 (power only)
Vi (idle inc call)
2
X1=‐1.304
X2=1.338
1
X3=‐1.359
X4=1.369
0
0
1
2
3
4
5
Underlying revenue value P
Figure 1: Investment ladder example with 41 = 4 3 2 1 for GBM and
power values    = 2 −1 05. Table 3 shows the investment graph whilst
this plot shows the idle, power and full flex values with their investment
quantities 41 (value matching occurs vertically at  = 4 3 2 1).
30
by replacing the discount/growth components within D G The merit of the
GBM is that its option elasticities are constant which simplifies the formation of β  β etc. For 0    1 it also permits the same option solution
for all  powers      
6
Restrictions for maxima
There are several different types of restriction on thresholds e.g. 41 in
Section 5. Firstly, for GBM (and some other diffusions but not arithmetic
Brownian motion) thresholds must be positive. Secondly, in order for the
discount functions to make sense, pairs must satisfy inequalities such as 2 
1 (4  3 but 3  2 ). This immediately constrains the flex values U
W that depend on 41 and the outputs 41  We would typically require
flexibility to be an asset (not a liability with a negative value) and therefore
solutions should be restricted to W  0 U  0 (vectors of zeros).
As a first order condition, smooth pasting is ensured by dollar beta matching, but this only defines the turning point of a unique value set and does
not say if that turning point is a maximum, minimum or inflection. Any proposed solution W U should be confirmed as maxima and neither a minima
nor saddle point of value. Without constructing more matrices of differentials, the second order conditions required to establish maximisation can be
conducted via numerical investigation.
For an assumed policy P = 41 , the candidate cost vector X(P) can be
calculated and fixed. Then with this fixed X the second order conditions can
be checked by examining (26) W = [I − D]−1 [Y − Z − X] when perturbing
away from P = 4321 . The effect on values within W (or U) will indicate
if the proposed policy and solutions are maxima, minima or saddle points.
Bearing in mind that there are multiple thresholds and flex values, as a function of each threshold change we must take care to identify which individual
flex values within W U we would expect to have concave behaviour and
zero/non zero slopes.
Smooth pasting and dollar beta matching row by row, means that thresholds 41 were chosen to equalize slopes across value matching. Although
smooth pasting says that pairs of values within each value matching line are
optimally co—determined by smoothness, it does not say that the values (in
rows of W U) have turning points with respect to their own threshold.
For example at 1 in the last row of (40), 1 transforms into 1 less 1
(plus 1 ). For fixed 1 as 1 moves from its local optimum, smooth pasted
components will change in a manner that preserves total sensitivity across
the transition 1 = 1 − 1 + 1 . When 1 is varied, flex values at other
31
thresholds however (e.g. 4 to 2 inclusive) will display turning points in
1 . Due to the smooth pasting status of each line, value changes must be
the same on both sides of (for fixed 42 ). Consequently, it is a flexibility’s
sensitivity with respect to thresholds other than its own that should be zero,
with concavity for a maximum. A local perturbations analysis of the system
in this Section 5 in this manner, confirmed values in Figure 1 as maxima.
Given policy P the existence of invertible matrices D G mapping between U W ensures uniqueness of the solution. There are however also
restrictions on 4321 that arise from the economics of desired target (i.e. input) values X = [4  1 ]| i.e. investment costs. With an analytic method
for mapping from a (restricted) set of 4−1 to solved 4−1  the system can be
used numerically via search to find the 4−1 that map to specific investment
costs. If a match of X(P) against a target values (for K’s) can be found and it
is one that satisfies the numerical maximisation test in this subsection, then
the solution can be said to be optimal given the costs X = K + K −K .
7
Conclusion
Flexibility to time the launch, suspension or other transformation within
a policy of thresholds P are valued as discount claims on an underlying
uncertainty process. Their values at the beginning and end of option lives
can be separately identified on a mathematical graph or network which then
lends itself to discounting.
Discounting has been used before with the factors and values appropriate
in each flexibility state depending on diffusion dynamics. We extended this
approach by capturing these features in a matrix that carries the appropriate
diffusion dynamics at each point of the investment network.
The rate of return of each flexibility value depends on the elasticity of its
discount factor. Not only must discounting hold between flex values but for
optimal smooth pasting, their rates of return must match too. This can be
captured using beta matrix conditions derived from, but complementary to,
those carrying fixed discounts. For fixed thresholds, these conditions allow
solution and interpretation of flex values as a series of discounted flexibility
payoffs.
For a given policy of thresholds P flexibility values at the beginning and
end of each state are driven by two considerations; discounting and smooth
pasting or dollar beta matching. For given thresholds, the third and most
natural condition of value matching only contributes to the solution of fixed
investment costs X(P) (which have zero beta). To find policy thresholds as
a function of costs P(X) a numerical search must be undertaken but the
32
approach and equations presented here facilitate this task greatly.
References
[1] Andrew B. Abel, Avinash K. Dixit, Janice C. Eberly, and Robert S.
Pindyck. Options, the value of capital, and investment. The Quarterly
Journal of Economics, 111(3):753—777, 1996.
[2] Andrew B. Abel and Janice C. Eberly. Optimal investment with costly
reversibility. Review of Economic Studies, 63:581—593, 1996.
[3] Carliss Y. Baldwin and Kim B. Clark. The Power of Modularity. Vol.
1, Design Rules. MIT Press, Cambridge, MA., 2000.
[4] Fischer Black and Myron Scholes. The pricing of options and corporate
liabilities. Journal of Political Economy, 81(May—June):637—659, 1973.
[5] Michael John Brennan and Eduardo S. Schwartz. Evaluating natural
resource investments. Journal of Business, 58(2):135—157, 1985.
[6] Dennis Capozza and Yuming Li. The intensity and timing of investment:
the case of land. American Economic Review, 84(4):889—904, 1994.
[7] John C. Cox, Stephen A. Ross, and Mark Rubinstein. Option pricing: a
simplified approach. Journal of Financial Economics, 7:229—263, 1979.
[8] D. A. Darling and A. J. F. Siegert. The first passage problem for a
continuous markov process. Annals of Mathematical Statistics, 24:624—
639, 1953.
[9] Avinash K. Dixit. Entry and exit decisions under uncertainty. Journal
of Political Economy, 97:620—638, 1989.
[10] Avinash K. Dixit and Robert S. Pindyck. Investment under Uncertainty.
Princeton University Press, 1994.
[11] Avinash K. Dixit, Robert S. Pindyck, and Sigbjørn Sødal. A markup
interpretation of optimal investment rules. The Economic Journal,
109(455):179—189, 1999.
[12] Janice C. Eberly and Jan A. Van Mieghem. Multi-factor dynamic investment under uncertainty. Journal of Economic Theory, 75:345—387,
1996.
33
[13] Steinar Ekern. Entry and exit decisions with restricted reversibility.
NHH Working Paper 93-08, 1993.
[14] Andrea Gamba and Nicola Fusari. Valuing modularity as a Real Option.
Management Science, 55(11):1877—1896, 2009.
[15] Bart M. Lambrecht and William R.M. Perraudin. Real options and preemption under incomplete information. Journal of Economic Dynamics
and Control, 27(4):619—643, 2003.
[16] Robert L. McDonald and Daniel R. Siegel. Investment and the valuation
of firms when there is an option to shut down. International Economic
Review, 26(2):331—349, 1985.
[17] Robert C. Merton. The theory of rational option pricing. Bell Journal
of Economics, 4(1):141—183, 1973.
[18] Stewart C. Myers. Determinants of corporate borrowing. Journal of
Financial Economics, 5:147—176, 1977.
[19] Robert S. Pindyck. Irreversible investment, capacity choice and the
value of the firm. American Economic Review, 78:969—985, 1988.
[20] Sudipto Sarkar and Fernando Zapatero. The trade-off model with mean
reverting earnings: Theory and empirical tests. The Economic Journal,
113(490):834—860, 2003.
[21] Mark B. Shackleton and Sigbjørn Sødal. Smooth pasting as rate of
return equalization. Economics Letters, 89(2 November):200—206, 2005.
[22] Mark B. Shackleton and Rafal Wojakowski. The expected return and
time to exercise of Merton style Real Options. Journal of Business
Finance and Accounting, 29(3—4):541—555, 2002.
[23] Sigbjørn Sødal. Entry and exit decisions based on a discount factor
approach. Journal of Economic Dynamics and Control, 30(11):1963—
1986, 2006.
[24] Lenos Trigeorgis. Real options: Managerial flexibility and strategy in
resource allocation. MIT Press, 1996.
[25] Robin J. Wilson. Introduction to Graph Theory. Longman, third edition,
1985.
34