Securitizing Congestion The Congestion C

Available online at www.sciencedirect.
com
Transportation Research Part B 42 (2008) 407–437

www.elsevier.com/locate/trb
Securitizing congestion: The congestion call option q

Terry L. Friesz a,*, Reetabrata Mookherjee b, Tao Yao a
a
Department of Industrial and Manufacturing Engineering, The Pennsylvania State University,
University Park, PA 16802, United States
b
Zilliant Inc., 3815 S Capital of Texas Highway, Suite 300, Austin, TX 78704, United States
Received 13 February 2007; received in revised form 13 October 2007; accepted 13 October 2007
Abstract
We introduce the idea of a European-type congestion call option to value commuting to work along a given path for a
given departure time selected by automobile drivers who are modeled as Cournot–Nash non-cooperative agents competing
for limited roadway capacity when the alternative of telecommuting exists. We show how to compute flow patterns based
on congestion call options and contrast these to traditional user optimized flow patterns. In particular, a numerical exam-
ple is presented that shows congestion options have the potential to lower the network-wide social costs of congestion.
Ó 2007 Elsevier Ltd. All rights reserved.
Keywords: Congestion securities; Congestion pricing; Congestion call; Congestion option; Financial engineering; Option pricing; Dyn-
amic games; Dynamic user equilibrium; Stochastic dynamic user equilibrium
1. Introduction
Dynamic traffic assignment (DTA) is the positive (descriptive) modeling of time-varying flows of automo-
biles on road networks consistent with established traffic flow theory and travel demand theory. Dynamic user
equilibrium (DUE) is one type of DTA wherein the unit cost, including early and late arrival penalties, of travel
for the same purpose is identical for all utilized route and departure time pairs. In the context of planning, DUE
is usually modeled for the within-day time scale based on demands established on a day-to-day time scale. In
this paper we introduce the idea of a European-type congestion call option to value strategies for commuting to
work when telecommuting is allowed. We further show how a DUE model may be modified to describe traffic
flow patterns that arise when a market for congestion calls is established. We also provide a numerical example
illustrating that the European congestion call option may lower the social costs of congestion.
q
This is an expanded version of a paper presented at the XIII Panamerican Conference on Traffic and Transportation Engineering,
Albany, NY, September 26–29, 2004. The proceedings of that conference contain a non-mathematical summary of the notion of a
congestion option. This manuscript is, however, the first and only complete mathematical explication of the European congestion call
option by the authors.
*
Corresponding author. Tel.: +1 814 863 2445; fax: +1 814 863 4745.
E-mail addresses: tfriesz@psu.edu (T.L. Friesz), reeto.mookherjee@gmail.com (R. Mookherjee), taoyao@psu.edu (T. Yao).
0191-2615/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trb.2007.10.002
408 T.L. Friesz et al. / Transportation Research Part B 42 (2008) 407–437
2. Some general remarks about dynamic user equilibrium models
In the last several years, much effort has been expended to develop a theoretically sound formulation of
dynamic network user equilibrium (DUE) which is also a canonical form acceptable to scholars and practi-
tioners alike. DUE models tend to be comprised of four essential submodels:
1. a model of path delay;

2. flow dynamics;
3. flow propagation constraints; and
4. a route/departure-time choice model.
It is not our intention nor does space permit us to review the entire DUE literature. Rather we simply com-
ment that Peeta and Ziliaskopoulos (2001), in a comprehensive review of DTA and DUE research, note that
there are several published models comprised of the four submodels named above. Friesz and Mookherjee
(2006) have shown how one such DUE model may be numerically solved using the theory of infinite dimen-
sional mathematical programming and the notion of a fixed-point in Hilbert space focusing on two infinite
dimensional variational inequality formulations of the DUE problem reported in Friesz et al. (1993) and
Friesz et al. (2001) that have much in common with other published models. In fact the Friesz et al. (1993)
and Friesz et al. (2001) formulations are more computationally demanding than most if not all other DUE
models because of the complicated path delay operators, equations of motion and time lags they embody.
As such the algorithmic perspective reported herein should work as well or better when adapted to other
DUE models, including those for which path delay is determined by a nonlinear response surface or by sim-
ulation for a so-called rolling horizon.
The advent of advanced information technologies has made it possible to consider market mechanisms such
as dynamic tolls and auctions as instruments for influencing traffic congestion. Here we consider a market-
based approach for controlling congestion, namely congestion securities. We focus on the idea of creating
a market for congestion rights in which congestion call options are traded. A congestion call option, as defined
herein, assures the holder the right to depart at a specific time and travel over a specific route. Using notions
from the theory of option pricing, we show that the value of any departure-time/route-choice combination
may be derived and employed within a DUE model to ascertain how flow patterns evolve when congestion
options are traded. Securitizing congestion is a novel idea and we know of no published literature that directly
addresses congestion options of any kind. However there are other areas of transportation for which a small
literature on options exists. For example, Marcus and Anderson (2006) put forward a real option analysis for
pricing of rental cars. In a similar vain Ching et al. (2007) present a real option approach for improving rev-
enue management by considering fluctuating commodity prices and time-varying strike prices with application
to airline revenue management. Sahin and Gallego (2006) show from a dynamic pricing standpoint that selling
call options on capacity may result in significantly higher expected revenues than those arising from low-to-
high pricing. Kober and Baumel (1990) reported the applications of transportation forward contracting and
options to railroads. Wilson et al. (1998), on the other hand, developed a dynamic stochastic simulation frame-
work to help shippers quantify alternative options for shipping grains by rail. Yan et al. (1995) put forward an
alternative risk based approach to intermodal pricing by considering and estimating various opportunity costs
that constitute, in effect, option prices.
3. A brief review of dynamic user equilibrium
We will consider a finite time horizon T 2 R1þþ and treat time t 2 ½0; T R1þ as a continuous variable. We
will employ the following notation, most of which is familiar to the dynamic traffic assignment community:
p an abstract path of the network of interest

(i, j) an abstract origin–destination (OD) pair of the network of interest
P set of all network paths
W set of all OD pairs
T.L. Friesz et al. / Transportation Research Part B 42 (2008) 407–437 409
Pij set of paths connecting OD pair ði; jÞ 2 W

A the set of arcs in the network
Ep exercise price for call p 2 P
Da unit delay for arc a 2 A
Dp unit delay for path p 2 P
v known, constant value of time for travelers
hp flow on path p (departure rate from the origin node of p)
h ðhp : ði; jÞ 2 W; p 2 Pij Þ0
Qij fixed number of trips undertaken for t 2 [0, T] between OD pair (i, j)
Sp(t) asset price attached to each departure time t and route choice p 2 P
lp(t, h) deterministic drift for path p 2 P depending on time of day and path flows
rp(t, h) volatility associated with path p 2 P depending on time of day and path flows
L the planning horizon
s one typical day in the planning horizon such that s 2 [0, 1, . . . , L]
D length of the day
t clock time such that t 2 [(s 1)D, sD] for all s 2 [1, . . . , L]
Other specialized notation will be introduced as needed in the course of our exposition.
We employ a variational inequality representation of dynamic user equilibrium due to Friesz et al.
(1993); analyzed by Friesz et al. (1996) and Friesz et al. (2001); and similar to the models of Perakis
(2000), Kachani and Perakis (2002); and Dial (2003a,b). In particular an arbitrary path p 2 P of the network
of interest is
:
p ¼ fa1 ; a2 ; . . . ; ai1 ; ai ; aiþ1 ; . . . ; amðpÞ g
where m(p) is the number of arcs comprising path p. We also let te denote the time at which flow exits an
arc, while td is the time of departure from the origin for the same flow. The exit time function spai therefore
obeys
te ¼ spai ðtd Þ
while
1 if ai 2 p

dai p ¼
0 if ai 62 p
is an element of the arc-path incidence matrix. The relevant arc dynamics are
dxpai ðtÞ
¼ gpai1 ðtÞ gpai ðtÞ 8p 2 P; i 2 ½1; mðpÞ ð1Þ
dt
xpai ð0Þ ¼ xpai ;0 8p 2 P; i 2 ½1; mðpÞ ð2Þ
where xpai is the traffic volume of arc ai contributed by path p, while gpai is the flow exiting arc ai and gpai1 is the
flow entering arc ai of path p 2 P. Also, gpa0 is the flow exiting the origin of path p; by convention we call this
the flow of path p and use the symbolic name
hp ¼ gpa0
Furthermore
X
xa ðtÞ ¼ dap xpa ðtÞ 8a 2 A
p2P
is the total arc volume.

Each arc is characterized by an arc delay function
Dai ðxai ðtÞÞ
which stipulates delay as a function of arc volume at the instant of entering an arc. The following recursive
identities then follow immediately:
spa1 ¼ t þ Da1 ½xa1 ðtÞ 8p 2 P
spai ¼ spai1 ðtÞ þ Dai ½xai ðspai1 ðtÞÞ 8p 2 P; i 2 ½2; mðpÞ
These in turn give rise to the nested path delay operators proposed by Friesz et al. (1993):
mðpÞ
X
Dp ðx; tÞ dai p Uai ðx; tÞ 8p 2 P
i¼1
where
x ¼ ðxpai : p 2 P; i 2 ½1; mðpÞÞ
and
Ua1 ðx; tÞ ¼ Da1 ½xa1 ðtÞ
Ua2 ðx; tÞ ¼ Da2 ½xa2 ðt þ Ua1 Þ
Ua3 ðx; tÞ ¼ Da3 ½xa3 ðt þ Ua1 þ Ua2 Þ
..
. " #
i1
P
Uai ðx; tÞ ¼ Dai ½xai ðt þ Uai 1 þ þ Ua1 Þ ¼ Dai xai ðt þ Uaj Þ
j¼1
To ensure realistic behavior, we employ asymmetric early/late arrival penalties

F ½t þ Dp ðx; tÞ T A
where TA is the desired arrival time (TA > T) and
t þ Dp ðx; tÞ > T A ) F ½t þ Dp ðx; tÞ T A ¼ vL ðx; tÞ > 0
t þ Dp ðx; tÞ < T A ) F ½t þ Dp ðx; tÞ T A ¼ vE ðx; tÞ > 0
t þ Dp ðx; tÞ ¼ T A ) F ½t þ Dp ðx; tÞ T A ¼ 0
vL ðx; tÞ > vE ðx; tÞ
We combine the actual path delays and arrival penalties to obtain the effective delay operators
Wp ðx; tÞ ¼ Dp ðx; tÞ þ F ½t þ Dp ðx; tÞ T A 8p 2 P ð3Þ
Elementary manipulations based on the chain rule and the assumption that arc delay operators are differen-
tiable lead to the flow propagation constraints
ga1 ðt þ Da1 ½xa1 ðtÞÞð1 þ D0a1 ½xa1 ðtÞ_xa1 Þ ¼ hp ðtÞ 8p 2 P ð4Þ
gpai ðt þ Dai ½xai ðtÞÞð1 þ D0ai ½xai ðtÞ_xai ðtÞÞ ¼ gpai1 ðtÞ 8p 2 P; i 2 ½2; mðpÞ ð5Þ
These flow propagation constraints are consistent with the chosen dynamics and model of arc delay.
The flow conservation constraints are
X Z T
hp ðtÞ dt ¼ Qij 8ði; jÞ 2 W ð6Þ
p2Pij 0
where Qij is the fixed travel demand for OD pair ði; jÞ 2 W; ½0; T R1þ is the departure-time window; Pij is
the set of paths connecting that OD pair; and W is the set of all OD pairs. Consider the following two-point
boundary value problem:
dy ij X
¼ hp 8ði; jÞ 2 W ð7Þ
dt p2Pij
y ij ð0Þ ¼ 0 8ði; jÞ 2 W ð8Þ
y ij ðT Þ ¼ Qij 8ði; jÞ 2 W ð9Þ

The two-point dynamics (7)–(9) are equivalent to the flow conservation constraints (6).
Of course, we also have
P ¼ [ði;jÞ2W Pij
We also enforce the following non-negativity restrictions:
x P 0; g P 0; hP0 ð10Þ
for all times within the analysis period, where x has been defined previously and
g ¼ ðgpai : p 2 P; i 2 ½1; mðpÞÞ
h ¼ ðhp : p 2 PÞ
The non-negativity restrictions together with the flow propagation constraints ensure, for strictly positive arc
exit flows, the condition
1 þ D0ai ½xai ðtÞ_xai ðtÞ > 0
guaranteeing the first-in-first-out (FIFO) queue discipline is enforced for every p 2 P and i 2 [1, m(p)]. As a
consequence the set of physically feasible arc exit flows and path flows
X ¼ fðh; gÞ : ð1Þ; ð2Þ; ð4Þ; ð5Þ; ð7Þ; ð8Þ; ð9Þ and ð10Þ holdg
is the relevant feasible region.
The infinite dimensional variational inequality formulation for dynamic network user equilibrium itself is:
find (h*, g*) 2 X such that
XZ T

hWðt; xðh ; g ÞÞ; ðh h Þi ¼ Wp ½t; xðh ; g Þ½hp ðtÞ hp ðtÞ dt P 0 ð11Þ
p2P 0
for all (h, g) 2 X, all of whose solutions Friesz et al. (2001) show are dynamic user equilibria.1 In particular the
solutions of (11) obey
Wp ðt; xðh ; g ÞÞ > lij ) hp ðtÞ ¼ 0 ð12Þ
hp ðtÞ
> 0 ) Wp ðt; xðh ; g ÞÞ ¼ lij ð13Þ
for p 2 Pij . where lij is the lower bound on achievable costs for any ij-traveler, given by
lij ¼ minflp : p 2 Pij g P 0
and
lp ¼ ess inffWp ðt; xÞ : t 2 ½0; T g P 0
We call a flow pattern satisfying (12) and (13) a dynamic user equilibrium. The behavior described by (12) and
(13) is readily recognized to be a type of Cournot–Nash non-cooperative equilibrium. It is important to note
that these conditions do not describe a stationary state, but rather a time-varying flow pattern that is a Cour-
not–Nash equilibrium (or user equilibrium) at each instant of time. See Friesz et al. (1993) and Friesz et al.
(2001) for the detailed derivation of (11).
1
Although we have purposely suppressed the functional analysis subtleties of the formulation, it should be noted that (11) involves an
inner product in a Hilbert space, namely ðL2 ½0; T ÞjPj .
4. The philosophy behind the congestion call option
In this section we describe several issues that must be considered in constructing the envisioned congestion
call option that is the focus of this paper.
4.1. Multiple time scales
Let s 2 [0, 1, . . . , L] be one typical day within the planning horizon, and take the length of each day to
be D, while the clock time within each day s is represented by t 2 [(s 1) D, sD] for all s 2 [1, . . . , L]. The plan-
ning horizon consists of L consecutive days.
4.2. A submodel of demand reduction through telecommuting
We assume all travelers are headed to work and each traveler using the transport network has a quali-
tatively identical real option; he/she could have stayed home and worked via the Internet. Moreover, pro-
ductivity of the worker varies depending on whether he/she is telecommuting or traveling to work. Travelers
decide whether or not to telecommute based on their moving average of congestion experienced over pre-
vious days and also on restrictions imposed on the maximum number of people who may telecommute on a
given day. We postulate that the volume of network travelers Rsij who decide to telecommute between a
given OD pair ði; jÞ 2 W on day s + 1 is determined by the following system of constrained difference
equations:
Ps1 R ðjþ1ÞD
Wp ½t; xðh ; g Þ dt
"P #
p2Pij j¼0 jD
Rijsþ1 ¼ Rsij þ gsij vij ð14Þ
jPij j s D
!
X X
R0ij ¼ 0; Rsij 6 cj Qij and Rsij P 0 8s 2 ½0; L ð15Þ
i2Wo i2Wo
where Q ¼ ðQij : ði; jÞ 2 WÞ is a known trip table, Pij is the set of paths connecting OD pair (i, j) and Wo is the
set of origin nodes considered. The term in parenthesis on the right-hand side of (14) is the moving average of
effective travel delay as of day s (today) when travelers decide whether or not to telecommute on day s + 1
(tomorrow); vij is the threshold of congestion characterizing the (i, j)-travelers, and cj is a policy determined
parameter that expresses an upper bound on the fraction of workers who may telecommute. Note that more
people will switch to telecommuting as the moving average of congestion exceeds their threshold and vice ver-
sa. The constraints (15) articulate the requirement laid down by the central authority (or by the employer in
some cases, especially that of the Federal workforce in Washington, DC) limiting the total number of people
who can telecommute. Note that the total number of people who travel to work at node j from node i on day s
is
Qsij ¼ Qij Rsij ð16Þ
4.2.1. Cases when the regulatory constraints are binding

On a given dayPif more travellersP to a destination node j want to telecommute than allowed by the regula-
tory constraint ð i2Wo Rsij 6 cj ð i2Wo Qij ÞÞ; then the travelers who are denied the right to telecommute must
make physical trips to work, even though the trip might have a value less than the value of time expended
during the commuting experience. This class of commuters may be treated as a separate class. For these com-
muters telecommuting is not an option, although they may still purchase the call option if the expected net
cost (which is the travel delay related cost minus expected financial gain) from purchasing the option, evalu-
ated at the time of departure, is less than the cost of taking alternative routes where options are not traded. In
a nutshell, the option evaluation framework for this class of commuters is the same as that of the other class
who are willingly making the physical trip.
4.2.2. A possible model for rejecting demand for telecommuting

For the notion of a congestion option to be useful, appropriate technology facilitating the decision-making
of telecommuting must be in place. Our purpose is not to design or describe the implementation of such tech-
nology but rather to develop a mathematical model that can explore the impacts of telecommuting combined
with the congestion call option on the social cost of congestion. Designing such a mechanism is to use a net-
work capacity control mechanism, such as partitioned booking limits, virtual nesting controls, and bid-price
controls – all of which are widely used in airline revenue management applications (Talluri and Ryzin,
2004). The basic idea is to have a fixed number of telecommuting slots available for every destination. The
commuters having offices located at a given destination node and willing to telecommute, bid for such a right.
Every individual will have a fixed number of points allocated by his/her employer every month, possibly in the
form of a telecommuting voucher. On a given day the employee, if willing to telecommute, will bid with all or
some of the points he/she has remaining. If we assume that the willingness to pay is uniformly distributed
around the mean of the prevailing moving average of effective travel delay, then some efficient capacity control
mechanisms may be designed that will deny some or all of the bidders the right to telecommute based on a
threshold set and updated periodically.
4.3. Option trading
We postulate there exists an information system that identifies routes and provides path delay information,
as well as key statistical information about volatility, a notion made more precise below. A mechanism based
on that information determines market prices faced by all commuters for their journeys to work. Each such
journey is described by a departure time, a route choice, and – once the journey is completed – the price of that
journey. There are no commuters who can avoid market pricing and there are no travelers who are not com-
muters. It is not necessary to our analysis that the journey prices determined by the pricing mechanism be effi-
cient; indeed, implementable dynamic congestion pricing mechanisms are necessarily ad hoc and will produce
inefficient prices in practice owing to the computationally intractable nature of real-time dynamic traffic
assignment on large networks.
So network users (commuters) who decide and are allowed to make a physical trip to work have two
choices:
he/she may purchase a congestion option to travel on a given route for a given departure time, or
he/she may travel on the same route for the same departure time by paying the market price upon reaching
his/her work destination.
The congestion option, a contract to buy or sell congestion rights, is sold to an individual user for a given
route and a given departure time. A central authority, such as a municipal planning organization (MPO), will
serve as a market maker for the commuter route and departure decisions we have described. Within that mar-
ket, we further postulate that trip prices are characterized by a time-of-day drift and a delay-dependent vol-
atility, giving rise to random price walks that are a geometric Brownian motion. As a consequence and as we
shall see explicitly below, an option pricing decision by the central authority may be predicated on a valuation
of the option of interest obtained by solving a Black–Scholes type partial differential equation. We also assume
that option trading starts at the beginning of the morning commute to work and continues until a pre-specified
daily time of expiry for all option contracts that have been written in a given day and by which time every
option purchaser has completed his/her travel. At the time of expiry, a traveler who has purchased an option
makes the decision whether or not to exercise the option based on his/her actual payoff.
We focus in this paper on call options. A congestion call option is a contract between the seller (writer of
the call, the MPO) and traveler (buyer of the call) that allows the buyer to have access to a travel route and a
departure time as well as information on the travel time to his/her destination. Moreover, the call options con-
sidered for this example are of the so-called European type, meaning that they may be exercised only at a pre-
specified moment of expiry. For example, if the traveler decides to exercise a call option, he/she gets a payoff
which is equal to the difference between the current asset (trip) price and strike price, unless of course that
difference is negative, in which case no payoff occurs. Uncertainty concerning option prices impart, according
to Itô’s lemma, randomness to the departure rates and to the arc exit flows even if the route and departure-
time choice mechanism is deterministic. Fluctuations of option prices arise from a myriad of causes. Perhaps
the most obvious factor influencing option price is the level of congestion as expressed by effective path delay.
Indeed, in our analysis below, congestion option price volatility is assumed to vary with the level of congestion
encountered. We also assume in the analysis to follow that the time of day drift as well as delay-dependent
volatility are available in identical form to all travelers for use in decision-making via the aforementioned
information system.
Clearly, for the notion of a congestion option to be useful, appropriate technology facilitating the writing
and exercising of congestion options must be in place. Our purpose is not to design or describe the implemen-
tation of such technology but rather to develop a mathematical model that can explore the impacts of conges-
tion call options and – perhaps at some future date – to form the basis of a decision support system for
managing a market for congestion puts2 as well as congestion calls. Owing to the well known put-call parity
relationship, the results of our mathematical analysis of the call option that is the focus of this paper may be
immediately adapted to the study and pricing of congestion puts. In order to evaluate the potential of conges-
tion calls to influence the net social costs of congestion, we need to value such options. We follow the now
standard analysis to value an option whose dynamics constitute a Wiener process,3 but make important mod-
ifications to that analysis in order to capture the essential features of a congested transportation network. In
fact the partial differential equation we obtain for congestion call value is not the much studied Black–Scholes
equation but rather a different partial differential equation.
Since transport network users could have telecommuted, given he/she is not denied the option because of
policy constraints, when an individual decides to make a physical trip to work, that trip must have a value
equal to or higher than the value of time expended during the commuting experience. To value time we use
the commuters’ wage rate for each origin–destination pair, although any other value of time could be used.
In reality only a fraction of the total population of travelers would belong to the class of potential telecom-
muters we have described. However, we have purposely striven to keep our exposition simple so that those
unfamiliar with option pricing may fully grasp the notion of a congestion option. In particular, modeling sev-
eral distinct transportation network user classes for each origin–destination pair is quite possible, but it would
gain us little in understanding how to model congestion call options and would present notational complexity
hampering understanding. For these reasons we consider only a single user class for each origin–destination
pair, namely telecommuters with an identical value of time.
4.4. Stochastic differential equations describing asset prices
The aforementioned access to path p along with a departure time is thought of as an ‘‘asset” with price Sp(t)
at time t evolving according to the geometric Brownian motion
dS p ðt; h Þ
¼ lsp ðtÞ dt þ rsp ½Dsp ðt; xðh ; tÞÞ; t dBp 8p 2 P; t 2 ½0; LD ð17Þ
S p ðtÞ
S p ð0Þ ¼ K p 2 R1þþ 8p 2 P ð18Þ
where lp(t) is the deterministic drift that depends on the time of day, and rsp ðÞ is the associated volatility dur-
ing day s, and dBp is a Brownian (white noise) disturbance. The volatility for path p will in general depend on
the time of day as well as delays for that path. Note also that the initial asset prices K p 2 R1þ for all p 2 P are
known constants. Furthermore, S is the vector of all asset prices; that is
S ðS p : p 2 PÞ0 ð19Þ
It is widely known that the stochastic differential equation (17) with the initial condition (18) has the following
solution:
2
A put option is a contract to sell an asset at a previously agreed price.
3
There are many published expositions of option valuation analysis based on Itô’s lemma. One such reference is Wilmott et al. (1993).
2
! !
t ðrsp ½Wsp ðt; xðh ; tÞÞ; tÞ
Z

S p ðt; h Þ ¼ K p exp lsp ðtÞ dt þ rsp ½Wsp ðt; xðh ; tÞÞ; tBp
0 2
See for example Wilmott et al. (1993).

We have imagined a situation where network users have perfect and complete information about everything
except call option prices. These users are described by a dynamic user equilibrium that devolves from deter-
ministic arc dynamics and deterministic arc and path delays as well as the from the values of relevant conges-
tion call options. Uncertainty exists concerning the call option prices, which are described by Brownian
motion. Naturally these call option prices impart, according to Itô’s lemma, randomness to the departure rates
and the arc exit flows. Fluctuations of call option prices arise from a myriad of causes. For a given origin–
destination pair ði; jÞ 2 W, perhaps the most obvious factor influencing volatility for a specific path p 2 Pij
is own traffic expressed as hp; also important is non-own traffic hq for paths q 2 Pij that have arcs in common
with path p 2 Pij . In Section 7 we present a simple model for congestion option price volatility that is based on
the above observations.
4.4.1. Estimation of drift and volatility

The functions lsp ðÞ and rsp ðÞ may be based on theoretical arguments or inferred statistically. For example,
since the time scales likely to be considered will be short, the drift may be thought of as changing very grad-
ually. In contrast, the volatility will climb sharply with rising congestion (effective path delay) until a jam den-
sity is reached at which time volatility will drop sharply toward zero.
4.5. The value function for the congestion call option
As mentioned earlier, a congestion call option is sold for a given route when departure from the origin
occurs. A congestion call option is exercised at the end of the day (at a pre-specified time) when all the network
users return from work. To evaluate the potential of congestion calls to influence the net social costs of con-
gestion, one may follow the now standard analysis, based on the Itô calculus, of option valuation but with
important modifications explained below.
Moreover, we are now in a position to analyze the path specific value function of the congestion call option
using Ito’s Lemma and the standard arbitrage-does-not-occur condition, provided we properly define the
notion of arbitrage associated with telecommuting. In turn, we will be able to derive a partial differential equa-
tion with proper initial and boundary conditions, solution of which provides the value function of the option
for a particular path.
4.5.1. Itô’s lemma

It is useful to remember that since the underlying stochastic process for asset prices is a Wiener process, we
know
pffiffiffiffiffi
ðdBp Þ2 ! dt as dt ! 0 so that dBp ¼ O dt ð20Þ
for all p 2 P. As a consequence, our dynamics (17) lead to

2 2 2 2 2 2
ðdS p Þ ¼ ðS p lp dt þ S p rp dBp Þ ¼ ðrp S p Þ ðdBp Þ þ 2lp rp S 2p dt dBp þ ðlp S p Þ ðdtÞ
’ r2p S 2p ðdBp Þ2 ! r2p S 2p dt ð21Þ
when we retain only first order terms. On the basis of this result, Itô’s lemma tells us that for
all p 2 P
2
oF ðS p ; tÞ oF ðS p ; tÞ 1 2 o F ðS p ; tÞ oF ðS p ; tÞ
dF ðS p ; tÞ ¼ rp S p dBp þ lp S p dt þ ðrp S p Þ 2
dt þ dt ð22Þ
oS p oS p 2 oðS p Þ ot
for any function F(Sp, t) that has the first and second derivatives appearing on the right-hand side
of (22).
4.5.2. The value function for a path

With the preceding background we are ready to begin construction of the congestion call value function. In
particular we express the net value of the congestion portfolio to travelers using path p instead of other routes
as
pp ¼ V p Dp S p for ði; jÞ 2 W; p 2 Pij ð23Þ
where Vp is the valuation of the path p call option and Dp is a scaling factor chosen subsequently to eliminate
stochasticity (the negative of which has the interpretation of the number of underlying assets held in the port-
folio). Note that the value function depends on the asset price and the time of departure; that is
V p ¼ F p ½S p ðtÞ; t for ði; jÞ 2 W; p 2 Pij ð24Þ
The value measure Vp captures far more that the traditional unit cost of flow notions emphasized in the trans-
portation network modeling literature; in particular, it implicitly reflects the risk associated with volatile con-
gestion levels. Clearly, taking Dp to be fixed during the time step dt, one may write
dpp ¼ dV p Dp dS p
We apply Itô’s Lemma to obtain
oV p oV p 1 o2 V p oV p
dV p ¼ rp ½t; Wp S p dBp þ lp ðtÞS p dt þ ðrp ½t; hS p Þ2 2
dt þ dt
As a consequence
2
oV p oV p 1 2 o Vp oV p
dpp ¼ rp ½t; Wp dBp þ lp ðtÞS p dt þ ðrp ½t; Wp S p Þ 2
dt þ dt Dp dS p
2
1

oV p oV p 2 o Vp oV p
¼ rp ½t; Wp Dp dBp þ lp ðtÞS p Dp dt þ ðrp ½t; Wp S p Þ 2
dt þ dt
We now pick the scale to obey
oV p
Dp ¼ 8p 2 P ð25Þ
oS p
so that
2
1 2 o Vp oV p
dpp ¼ ðrp ½t; hS p Þ 2
dt þ dt 8p 2 P ð26Þ
2 oðS p Þ ot
4.5.3. The arbitrage-does-not-occur condition

We postulate that people have distinct in-office and at-home marginal productivities with respect to time.
We assume no productive work occurs during the commute, so the value of the option to travel must equal the
net productivity loss associated with commuting; that is
dpp ¼ ðMP office
ij MP home
ij Þ dt 8p 2 Pij ð27Þ
for each p 2 Pij and MP office
ij and MP home
ij are the in-office and at-home marginal productivities, respectively.
Combining (26) and (27) leads to the following system of partial differential equations that describe evolution
of the congestion call for every origin–destination pair ði; jÞ 2 W:
1 o2 V p oV p
ðrp ½t; hS p Þ2 2
þ ðMP office
ij MP home
ij Þ¼0 8p 2 Pij ð28Þ
2 oðS p Þ ot
Initial and boundary conditions. To keep the exposition simple, we assume that the expiry time of the option is
fixed at midnight (sD) for all potential commuters. Therefore the relevant boundary conditions for the Euro-
pean congestion call option are
V sp ½S p ðsDÞ; sD ¼ m max½S p ðsDÞ ap ; 0 8ði; jÞ 2 W; p 2 Pij ð29Þ
V p ½0; t ¼ 0 8p 2 P ð30Þ
and V sp ½S p ðsDÞ; sD ! S sp ðsDÞ as S sp ðsDÞ ! 1 8p 2 P ð31Þ

where ap is the strike price of the option associated with path p and m is the known value of time. Note that (29)
ensures that at the time of expiry the option would not be exercised if the net value of the option Sp(T) ap is
negative. Similarly, if asset price becomes 0 at any time t, it will stay at zero thereafter per the geometric
Brownian motion stipulated by (17) and (18). Under such circumstances, the option becomes meaningless,
and hence value of option becomes 0 as per (30). Furthermore, (31) tells us that if the underlying asset price
becomes very high, the exercise price becomes very small compared to the asset price and, hence, the value of
the option is, in the sense of a limit, the asset price.
5. Dynamic equilibrium and the stochastic differential variational inequality (SDVI) problem
We assume that the network travelers are risk neutral in selecting a route and departure time. A risk neu-
tral user purchases an option only if the expected financial liability arising from the purchase of the option is
negative (which implies a positive net cash flow.) Otherwise, the user simply takes the same route without pur-
chasing the option. We use a dynamic Cournot–Nash equilibrium to describe the decisions of all agents, both
those purchasing options and those not (still take the same route and pay the prevailing asset price at the time
of arrival). However, because of the Brownian motion we consider in describing the price fluctuations of tra-
vel alternatives arising from the unobservable aspects of individual travelers’ utility functions over time, the
formulation becomes a stochastic differential (that is, dynamic) variational inequality (SDVI). Let j be the
premium charged by the central authority to decide the purchase price of an option, that is
j V sp ðS p ðt; h Þ; tÞ: This is the price one has to pay at time t to purchase the option. When j > 1, a premium
is charged whereas j < 1 corresponds to a scenario where the central authority provides a subsidy to encour-
age more travelers to purchase an option. Therefore, j is a decision variable in the hands of the central
authority.
5.1. Net trip value for the option purchasers
Expected net trip value for the option purchasers corresponding to a departure time t can be expressed in
terms of their expected return at the time of expiry and price they paid to purchase the option:
þ
EV s1 ðt; h Þ ¼ Et ½ðS p ðsDÞ ap Þ jS p ðtÞ j V sp ðS p ðtÞ; tÞ Et ½S p ðsDÞjS p ðtÞ ð32Þ
where the first term on the right-hand side is the expected payoff at the time of expiry conditioned over the
current asset price and the second term is the price to be paid at time t to purchase the option (and depart
from the origin). We use the usual notation (y)+ max(y,0). Note that since asset price follows geometric
Brownian motion, the asset price distribution follows a log-normal distribution, and thus the conditional
expectation may be obtained as a closed form expression. However, one can also obtain the conditional expec-
tation numerically using a finite element approximation.
5.2. Net trip value for non-option purchasers
If a traveler does not purchase an option, he/she would pay a price equal to the prevailing market price at
the time he/she reaches the intended destination. Therefore, the net expected trip cost for such travelers is the
expected asset price at the time of their arrival conditioned over the current asset price. Thus, the expected net
trip value in this case is
EV s2 ðt; h Þ ¼ Et ½S p ðt þ Dp ðt; h ÞÞjS p ðtÞ ð33Þ
where Dp(t, h*) is the path delay along path p corresponding to a departure time t.
5.3. Properties of modified dynamic user equilibrium
Travelers departing at time t will purchase the option when EV s1 ðt; h Þ > EV s2 ðt; h Þ and vice versa. There-
fore, the definition of a dynamic user equilibrium is modified to consider the net expected trip value. In par-
ticular, the modified DUE problem must obey:
1. Non-trivial path flow requires maximal net value:

hp > 0 ) t F ½t þ Dp ðx ; tÞ T A þ maxðEV s1 ðt; h Þ; EV s2 ðt; h ÞÞ ¼ nij
8ði; jÞ 2 W; p 2 Pij ; s 2 ½1; L and t 2 ½ðs 1ÞD; sD
2. Net trip value has a well-defined maximum:
F ½t þ Dp ðx ; tÞ T A þ maxðEV s1 ðt; h Þ; EV s2 ðt; h ÞÞ 6 nij
8ði; jÞ 2 W; p 2 Pij ; s 2 ½1; L and t 2 ½ðs 1ÞD; sD
3. Flows, volumes, and maximal trip value are non-negative:
hP0
xP0
nP0
where
0
n ¼ ðnij : ði; jÞ 2 W; p 2 Pij Þ
Articulation of an infinite dimensional variational inequality consistent with the above formulation of stochas-
tic DUE requires that we define
K = {(h,g): (1), (2), (4), (5), (7), (8), (9), (14), (15), (10), (19), (17), (18), (28), (29), (30), (31) hold}
Consequently we may characterize dynamic user equilibrium in the presence of congestion options as follows:
for a sample path x, which is a realization of the geometric Brownian disturbances on the right-hand sides of
(17), we solve the following stochastic differential variational inequality (SDVI): find {h*(x), g*(x)} 2 K such
that
X X X L1 Z ðjþ1ÞD
ftF ½t þ Dp ðx ; tÞ T A max½EV s1 ðt; h Þ; EV s2 ðt; h Þ ½hp hp ðxÞg dt P 0 ð34Þ
ði;jÞ2W p2Pij j¼0 jD
for all (h, g) 2 K where t is the value of time; S(t, x), h(x) and g(x) denote respectively the realizations of asset
prices, path flows and arc departure rates for sample trajectory x 2 X; and X is the set of all possible sample
trajectories. Furthermore the conditional expectations EV s1 ðt; h Þ and EV s2 ðt; h Þ appearing in (34) were defined
in (32) and (33). Clearly, (34) is a SDVI wherein uncertainties are associated with the asset price S p ðt; xÞ and
the value function of the congestion option V p ½S p ðt; xÞ; t.
5.4. Fokker–Planck equation and expected net congestion costs
The probability density function of asset price Sp(t), conditioned over Sp(0) = Kp, is
P fz < S p ðtÞ < z þ DzjS p ð0Þ ¼ K p g
wðz; t; K p Þ ¼ lim ð35Þ
Dz!0 Dz
for every path p 2 P. The Fokker–Planck equation is a forward partial differential equation whose solutions
are the conditional transition probabilities (35). The Fokker–Planck equation4 and boundary conditions asso-
ciated with the geometric Brownian motion (17) are
4
See Montroll and Badger (1974) for an explanation and derivation of the Fokker–Planck equation.
1 2 o2 o o
rp ðt; h ÞS 2p ðh ; tÞ 2 wða; t; K p Þ ¼ lp ðt; h ÞS p ðh ; tÞ wða; t; K p Þ þ wða; t; K p Þ for all p 2 P ð36Þ
2 oS p oS p ot
wða; 0; K p Þ ¼ dða K p Þ for all p 2 P ð37Þ
wð1; t; K p Þ ¼ 0 for all t 2 ½0; L; p 2 P ð38Þ
wð0; t; K p Þ ¼ 0 for all t 2 ½0; L; p 2 P ð39Þ
where d() is the dirac delta function.
5.4.1. Closed form solution of the Fokker–Planck equation

For a given sample path x, the path flows h*(x) can be expressed as a pure function of time as explained in
Friesz and Mookherjee (2006). Hence the volatility and drift coefficients r(h*(x), t) and l(h*(x), t) may also be
expressed as pure functions of time. With the above observations in mind, it is possible to show that the con-
ditional probability density function of asset price Sp(t), conditioned by the initial value Sp(0), follows a log-
normal distribution; therefore ln(Sp(t)) follows the normal distribution with mean
Z t
1
mp ðh ðxÞ; tÞ lnðK p Þ þ lp ðh ðxÞ; jÞ r2p ðh ðxÞ; jÞ dj
0 2
and variance
Z t
g2p ðh ðxÞ; tÞ r2p ðh ðxÞ; jÞ dj
0
The closed form solution of (36)–(39) is5

0 " #2 1

1 1 logðaÞ m p ðh ðxÞ; tÞ
wða; t; K p Þ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp @
A for a > 0
2
a 2p gp ðh ðxÞ; tÞ 2 g p ðh ðxÞ; tÞ
¼ 0 otherwise
Making substitutions for w(a, t, Kp) in the previously derived expressions for mp(h*(x),t) and g2p ðh ðxÞ; tÞ, we
obtain for a > 0 the following expression:
1
wða; t; K p Þ ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
hR i
t
a 2p 0 g2p ðh ðxÞ; jÞ dj
0 h Rt i2 1
S p ðt; xÞ lnðK p Þ 0 ðlp ðh ðxÞ; jÞ 12 r2p ðh ðxÞ; jÞÞ dj
exp @ ð40Þ
B C
Rt
2 0 r2p ðh ðxÞ; jÞ dj
A
5.4.2. Expressions for conditional expectations

To proceed we need to find computable expressions for the expectations (32) and (33). Consequently, we
þ
need to obtain computable expressions for Et ½ðS p ðsDÞ ap Þ j S p ðt; xÞ and Et ½S p ðt þ Dp ðt; h ÞÞ j S p ðt; xÞ, each
of which we next address.
Expression for Et ½ðS p ðsDÞ ap Þþ j S p ðt; xÞ: The distribution of S p ðsDÞ conditioned by S p ðt; xÞ follows a log-
normal distribution; therefore lnðS p ðsDÞÞ follows a normal distribution with mean
Z sD
1 2

mp lnðS p ðt; xÞÞ þ lp ðh ðxÞ; jÞ rp ðh ðxÞ; jÞ dj ð41Þ
t 2
and variance
5
See Wilmott et al. (1993) for the derivation of the closed form solution using the Fourier transform.
Z sD
g2p r2p ðh ðxÞ; jÞ dj ð42Þ
t
We use an indicator function I() such that for any event A, I(A) = 1 if the event occurs and I(A) = 0 other-
wise. For any such event A we have
EðIðAÞÞ ¼ P ðAÞ
We can hence write
þ
Et ½ðS p ðsDÞ ap Þ jS p ðt; xÞ ¼ Et ½ðS p ðsDÞ ap Þ IðS p ðsDÞ P ap ÞjS p ðt; xÞ
¼ Et ½S p ðsDÞ IðS p ðsDÞ P ap ÞjS p ðt; xÞ ap P ððS p ðsDÞ P ap ÞjS p ðt; xÞÞ ð43Þ
Now, to derive an expression for P ððS p ðsDÞ P ap Þ j S p ðt; xÞÞ; let us define z as the standard normal random
variable with CDF U(). Therefore S p ðsDÞ can be expressed using mp and gp as defined respectively in (41)
and (42) as
S p ðsDÞ ¼ expðmp þ gp zÞ ð44Þ
It follows immediately that
( )
lnðap Þ mp
P ððS p ðsDÞ P ap ÞjS p ðt; xÞÞ
¼ P fmp þ gp z P lnðap Þg ¼ P z P
gp
! !
lnðap Þ mp mp lnðap Þ
¼1U ¼U ð45Þ
gp gp
Expression for Et ½S p ðsDÞ IðS p ðsDÞ P ap Þ j S p ðt; xÞ: Using (44), we have
1
Z
j2
Et ½S p ðsDÞ IðS p ðsDÞ P ap ÞjS p ðt; xÞ ¼ pffiffiffiffiffiffi emp þgp j e 2 dj
2p mp þgp jPlnðap Þ
1
Z
gp
¼ pffiffiffiffiffiffi e12ðjgp Þ2 emp þ 2 dj
2p jP mp lnða gp
pÞ
( !) !
g
mp þ 2p mp lnðap Þ g
mp þ 2p mp lnðap Þ
¼e P z þ gp P ¼e U gp þ ð46Þ
gp gp
Therefore, substituting (45) and (46) into in (43), we obtain
! !
þ
g
mp þ 2p mp lnðap Þ mp lnðap Þ
Et ½ðS p ðsDÞ ap Þ jS p ðt; xÞ ¼e U gp þ ap U Et ½S p ðsDÞjS p ðtÞ ð47Þ
gp gp
where mp and gp are defined in (41) and (42), respectively.

Expression for Et ½S p ðt þ Dp ðt; h ÞÞ j S p ðt; xÞ: The expression for Et ½S p ðt þ Dp ðt; h ÞÞ j S p ðt; xÞ can also be
obtained in closed form after observing that S p ðt þ Dp ðt; h ÞÞ conditioned by S p ðt; xÞ also follows a lognormal
distribution; so lnðS p ðt þ Dp ðt; h ÞÞÞ follows a normal distribution with mean
Z tþDp ðt;h Þ
1 2

m~ p lnðS p ðt; xÞÞ þ lp ðh ðxÞ; jÞ rp ðh ðxÞ; jÞ dj ð48Þ
t 2
and variance
Z tþDp ðt;h Þ
g2p
~ r2p ðh ðxÞ; jÞ dj ð49Þ
t
Therefore,
Et ½S p ðt þ Dp ðt; h ÞÞjS p ðt; xÞ ¼ m
~p
6. A fixed-point algorithm
To solve (34) for a specific sample path realization x we employ a fixed-point algorithm, which is stated
below.
Fixed-point algorithm for solving (34)
Step 0. Realization of random noise.

Generate random noise components dBxp ðtÞ N ð0; dtÞ corresponding to the asset-price-sample-path x
for all t 2 [0, LD].
Step 1. Initialization.
Identify an initial feasible solution (h0, g0) 2 K and set k = 0.
Step 2. Sample path realization.
Generate an asset-price-sample-path realization using (17) and (18) with hk(x) as the path flow vector,
and thereby update the asset prices, S kp ðt; xÞ for 0 6 t 6 LD and all p 2 P.
Step 3. Solve the PDE.
Solve the PDE (28)–(31) with the volatility rp(t, hk(x)) to obtain the current path specific congestion
call value operators V p ½S kp ðt; xÞ; t for all (s 1)D 6 t 6 sD, s 2 [1, L] and all p 2 P.
Step 3a. Compute demand reduction.
Using hk(x) as the path flow vector, solve the constrained difference equations (14) and (15) and
thereby compute Rsij and Qsij .
Step 4. Form and solve fixed-point problem.
Re-state the DUE problem as a fixed-point problem based on the call value operators and effective
travel demand as found in Steps 3 and 3a, respectively; solve the resulting DUE problem iteratively.6
Call the solution (hk+1(x), gk+1(x)).
Step 5. Stopping test.
If

! !

hkþ1 ðxÞ hk ðxÞ

6e

kþ1

g ðxÞ gk ðxÞ
where e 2 R1þþ is a preset tolerance, stop and declare h*(x)

hk+1(x) and g*(x)
gk+1(x). Otherwise
set k = k + 1 and go to Step 2.
Note that by generating different random noise components, dBxp ðtÞ and thereby different sample paths x in
Step 0, we will get different solutions [h*(x), g*(x)]. It is also important to realize that the fixed-point algorithm
described above may be carried out in continuous time provided we employ a continuous time representation
of the solution of each subproblem from Step 4. This may be done using a continuous time gradient projection
method, as in Friesz and Mookherjee (2006), to deal with time-shifted controls. We further comment that it is
unlikely that the path value operators will always be strictly monotonic. In that event, as explained in Friesz
and Mookherjee (2006), the fixed-point algorithm used for Step 4 does not enjoy a proof of convergence, and,
consequently, the algorithm we have proposed for solving (34) is a heuristic.
6.1. Approximating the expected value of net congestion costs
To compute the average net social congestion costs, we use the transition probabilities obtained through the
solution of Fokker–Planck equation; that is, we solve (36)–(39). The transition probabilities together with the
6
See Friesz and Mookherjee (2006) for the details of converting a differential variational inequality into a fixed-point problem in Hilbert
space.
net values of the congestion calls corresponding to each sample path of interest allow the needed expectations
to be computed. Note that
n o
s
V net
p ðx; tÞ ¼ tF ½t þ Dp ðx ; tÞ T A CðtÞ max½S p ðsD; xÞ ap ; 0 jV p ðS p ðt; xÞ; tÞ
þ f1 CðtÞg S p ðt þ Dp ðt; h Þ; xÞ
is the congestion cost at time t 2 [0, LD] associated with the sample path x 2 X, where C(t) is the indicator
function expressing whether individuals departing at time t purchased a call option; in particular the indicator
function obeys
CðtÞ ¼ 1 if EV s1 ðt; h Þ P EV s2 ðt; h Þ

¼ 0 otherwise
Consequently, the expected net congestion cost is

L1 Z
X X X ðjþ1ÞD Z z¼þ1
EV net
¼ V net
p ðx; tÞwp ðz; t; K p Þ dz dt ð50Þ
ði;jÞ2W p2Pij j¼0 jD z¼1
Using an ensemble comprised of a large number of samples, the approximation

L1 Z ðjþ1ÞD
1 X X XX
EV net
V net
p ðx; tÞwp ðS p ðt; xÞ; t; K p Þ dt ð51Þ
R ði;jÞ2W p2Pij x2X j¼0 jD
may be reliably employed, where R is the number of sample path realizations considered.
7. Numerical example
Numerical solution of the differential variational inequality (34) using the fixed-point algorithm introduced
in Section 6 requires that we know how the value function evolves over time for a given sample path of
congestion security prices. Our postulate that volatility depends on departure rates means that, for a given
sample path realization, each partial differential equation (28) is coupled to the differential variational inequal-
ity (34).
We developed a software code that employs current path specific volatility together with the current value
of departure rates to solve (28) with appropriate boundary and initial conditions. The value function Vp(Sp, t)
thereby obtained is returned to a main program that iteratively solves the DUE fixed-point problem. That is,
at each major fixed major iteration, the stochastic differential equations (17) and (18) along with the partial
differential equation (28) are solved using the current path flow vector (hk) and arc exit flow vector (gk) where
k is the index of major (fixed-point) iterations. In particular, the partial differential equation (28) is solved
using Matlab’s PDEPE parabolic partial differential equation solver.
We provide numerical results for four distinct decision environments:
1. the dynamic user equilibrium of a 5 arc, 4 node network having a single OD pair and 3 feasible paths when
there is no telecommuting, no market for congestion rights, and no option trading;
2. the same network with no market for congestion rights and no option trading but limited
telecommuting;
3. the same network with both telecommuting and a market for congestion rights but no option trading; and
4. the same network with telecommuting, a market for congestion rights and European-type call option
trading.
In particular, we provide a comparison of the total congestion costs with and without call option trading. Our
example shows that congestion call option trading can lower the social costs of congestion. By intent, we make
no effort in this paper to ascertain what conditions assure that options will lower congestion costs – that is the
topic of a future paper.
7.1. Multi-period DUE model: no telecommuting and no options
When telecommuting is not a viable choice to the travelers, OD travel demand stays unchanged; hence flow
patterns remain identical for all days. In this case each of the within-day subproblems are decoupled and may
be solved independently as outlined by Friesz and Mookherjee (2006). In what follows, we consider the 5 arc, 4
node network shown in Fig. 1. The forward star array and arc delay functions Da(xa(t)) for all 5 arcs of the
network are contained in the following table:
Arc name From node To node Arc delay, Da(xa(t))

1 x
a1 1 2 2 þ 70a1
x a2
a2 1 3 1 þ 150
1 x a3
a3 2 3 2 þ 100
x a4
a4 2 4 1 þ 150
1 xa 5
a5 3 4 2 þ 100
There is a travel demand of Q = 75 units from node 1 (origin) to node 4 (destination). There are 3 paths con-
necting nodes 1 through 4, namely
p1 ¼ fa1 ; a4 g
p2 ¼ fa2 ; a5 g
p3 ¼ fa1 ; a3 ; a5 g
so that the set of all paths is
P14 ¼ fp1 ; p2 ; p3 g
a1 a4
1 a3 4
a2 a5
Fig. 1. The 5-arc 4-node traffic network with (1, 4) being the OD pair.
We take the planning horizon to be 4 days (i.e., L = 4) and the length of each day to be D = 24 h. The desired
arrival time for commuters is TA = 13 (1:00 PM of every day). The controls (path flows and arc exit flows) and
states (arc traffic volumes) are enumerated in the following table:
Paths Path flows Arc exit flows Traffic volume of arcs

p1 hp 1 gpa11 ; gpa14 xpa11 ; xpa14
p2 hp 2 gpa22 ; gpa25 xpa22 ; xpa25
p p p p p p
p3 hp 2 ga31 ; ga33 ; ga35 xa31 ; xa33 ; xa35
We consider the symmetric early/late arrival penalty

2
F ½t þ Dp ðx; tÞ T A ¼ ½t þ Dp ðx; tÞ T A
Furthermore, without any loss of generality, we take the initial traffic volumes on every arc to be zero:
xpai ð0Þ ¼ 0 8p 2 P; i 2 ½1; mðpÞ
We forgo the detailed symbolic statement of this example and instead provide numerical results in graphical
form for an essentially exact solution achieved after 29 major (fixed-point) iterations. Figs. 2–4 depict the com-
puted departure rates and arc exit flows for paths p1, p2 and p3, respectively.
Cumulative traffic volumes on the 5 different arcs are plotted against time in Fig. 5 where
xa1 ðtÞ ¼ xpa11 ðtÞ þ xpa31 ðtÞ
xa2 ðtÞ ¼ xpa22 ðtÞ
xa5 ðtÞ ¼ xpa25 ðtÞ þ xpa35 ðtÞ
for all time t 2 [0,LD].
Note that the effective path delay operator in (3) gives the unit travel cost along a path p at time t. Fig. 6
analyzes the effective delay and flow for path p1 by plotting both for the same time scale which shows that path
flow is maximal when the associated unit travel cost (effective path delay) is at its well-defined minimum. Sim-
ilar comparisons are also made for paths p2 and p3 in Figs. 7 and 8, respectively; these plots make clear that a
dynamic user equilibrium has been attained.
7.2. Multi-period DUE model: telecommuting and no options
When telecommuting is an alternative for some of the travelers, the number of travelers who decide to tele-
commute at the beginning of day s is given by the constrained difference equations (14) and (15). We set the
proportionality constant (gsij ) associated with (14) to 0.06. The threshold of congestion set by the (1, 4)-trav-
18
16
14
12
Flows
10
8
6
4
2
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
Flow p1 g1p1 g4p1
Fig. 2. Path and arc exit flows for path 1.

18
16
14
12
Flows
10
8
6
4
2
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
Flow p2 g2p2 g5p2
16
14
12
10
Flows
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
Flow p3 g1p3 g3p3 g5p3
18
16
Cumulative Arc Volume
14
12
10
0
0 6 12 17 23 29 35 40 46 52 58 64 69 75 81 87 93
Time (Hour)
Arc 1 Arc 2 Arc 3 Arc 4 Arc 5
Fig. 5. Cumulative arc volume vs. time.
elers is v14 = 20. The policy determined parameter c4 is set at 0.2, which indicates that employers located at
node 4 allow up to 20% of their total work force to telecommute on any given day. We once again forgo
the detailed symbolic statement of this example and instead provide numerical results in graphical form for
an essentially exact solution achieved after 32 major (fixed-point) iterations. Figs. 9–11, respectively depict
departure rates and arc exit flows for paths p1, p2 and p3. We also plot cumulative traffic volumes on the 5
different arcs against time in Fig. 12.
As before, we analyze the effective delay and flow for paths p1, p2and p3 by plotting both for the same time
scale in Figs. 13–15, respectively; these plots show that path flow is maximal when the associated unit travel
200 18
180 16
160
14
140
Unit Travel Cost

12
120
10
100
8
80
6
60
4
40
20 2
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 1 Flow p1
Fig. 6. Path flows and unit travel costs for path p1.
60 18
16
50
14
Unit Travel Cost
40 12
10
30
8
20 6
4
10
2
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 2 Flow p2
Fig. 7. Similar comparison for path p2.
60 16
14
50
12
Unit Travel Cost
40
10
Path Flow
30 8
6
20
4
10
2
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 3 Flow p3
Fig. 8. Comparison of path flows and associated unit travel costs for path p3.
cost (effective path delay) is at its well-defined minimum, again establishing that a dynamic user equilibrium
has been attained.
18
16
14
12
Flows
10
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
Flow p1 g1p1 g4p1
18
16
14
12
Flows
10
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
Flow p2 g2p2 g5p2
16
14
12
10
Flows
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (hour)
Net travel demand and demand reduction are plotted below against the same time scale (day) which clearly
demonstrates that more commuters switch to telecommuting as their rolling average of congestion costs
increases with the passage of time.
18
16

14
12
10
0
0 6 12 17 23 29 35 40 46 52 58 64 69 75 81 87 93
Time (Hour)
200 18
180 16
160 14
Unit Travel Cost
140
12
120
10
100
8
80
6
60
4
40
20 2
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 1 Flow p1
60 18
16
50
14
Unit Travel Cost
40 12
10
30
8
20 6
4
10
2
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 2 Flow p2
7.3. Market for congestion with telecommuting but no options
When we assume that there exists a market for congestion, travelers who take a particular route need to
purchase that right at the prevailing asset price. We consider a market for trading rights to travel along path
60 16
14
50
12
Unit Travel Cost

40
10
Path Flow
30 8
6
20
4
10
2
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 3 Flow p3
p1; no such market exists for paths p2 and p3. We assume the deterministic drift lp1 ðtÞ and path specific vol-
atility rp1 ðt; hÞ have the following form:
lp1 ðtÞ ¼ ap1 þ bp1 t ð52Þ
2
rp1 ðt; Dsp ðt; xðh; tÞÞÞ ¼ cp1 þ gp1 Dsp ðt; xðh; tÞÞ þ jp1 ½Dsp ðt; xðh ; tÞÞ ð53Þ
where ap1 ; cp1 ; gp1 2 R1þþ 1
and bp1 ; jp1 2 R : We choose ap1 ¼ 0:03, bp1 ¼ 0:001, cp1 ¼ 0:2, gp1 ¼ 0:02 and
jp1 ¼ 3 105 . From our choice of parameters, it is evident that deterministic drift is overshadowed by
the path specific volatility. Also, the volatility associated with a path has a nonlinear relationship with
own-path delay. Our model of volatility also expresses the tendency of volatility to drop after a jam density
is reached; see Fig. 16. We generate sample paths using a traditional Euler scheme with time discretization:
S p1 ;t ¼ S p1 ;t1 þ Dt lp1 ;t þ Bp1 ;t rp1 ;t 8t 2 ½1; N
S p1 ;0 ¼ K p1
where the entire planning horizon is discretized using the fixed time steps Dt ¼ 14 : The noise, Bp1 ;t follows a nor-
mal distribution with mean 0 and standard deviation Dt. We choose K p1 ¼ 4 and generate a number of sample
paths. We plot variations of asset price against time for six such sample paths in Fig. 17.
76 5
75 4.5
4
Demand Reduction
Net Travel Demand
74
3.5
73
3
72 2.5
2
71
1.5
70
1
69
0.5
68 0
0 1 2 3
Time (Day)
Net Q Reduction
Fig. 16. Net travel demand and demand reduction.

25
20
Asset Price ($)

15
10
0
0 10 20 30 40 50 60 70 80 90
Time (Hour)
Fig. 17. Plot of asset price (Sp1) vs. time associated with path p1 for 6 sample path realizations.
The numerical results we provide here in graphical form corresponds to an essentially exact solution
achieved after 18 major (fixed-point) iterations. Figs. 18–20 depict departure rates and arc exit flows for paths
p1, p2 and p3, respectively. Cumulative traffic volumes on the five different arcs are plotted against time in
Fig. 21.
Furthermore, we analyze the effective delay and flow for paths p1, p2and p3 by plotting both for the same
time scale in Figs. 22–24, respectively; these figures show that path flow is maximal when the associated unit
travel cost (effective path delay) is at its well-defined minimum, so clearly a dynamic user equilibrium has been
attained. Net travel demand and demand reduction are plotted below against the same time scale (day) for
four sample paths in Fig. 25.
18
16
14
12
Flows
10
8
6
4
2
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
Flow p1 g1p1 g4p1
9
8
7
6
Flows
5
4
3
2
1
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
Flow p2 g2p2 g5p2

12
10
Flows
6
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
18
16
14
12
10
0
0 6 12 17 23 29 35 40 46 52 58 64 69 75 81 87 93
Time (Hour)
1600 18
16
1400
14
1200
12
Unit Travel Cost
1000
10
800 8
6
600
4
400
2
200
0
0 -2
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 1 Flow p1
250 9
200 7
Unit Travel Cost

6
150
5
4
100
3
2
50
1
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 2 Flow p2
250 12
10
200
Unit Travel Cost
Path Flow
150
100
4
50
2
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 3 Flow p3
76 12
74
10
72
Demand Reduction
Net Travel Demand
70 8
68
6
66
64 4
62
2
60
58 0
0 1 2 3
Time (Day)
Net Q S1 Net Q S2 Net Q S3 Net Q S4 Reduction S1

Reduction S2 Reduction S3 Reduction S4
Fig. 25. Net travel demand and demand reduction for 4 sample paths.
7.4. Congestion options with telecommuting
Now we consider the same network, but with a European call option traded for path p1; no options are
traded for paths p2 and p3. Drift and volatility associated with the asset price dynamics are assumed to be
the same as described in Section 7.3. We again forgo the detailed symbolic statement of this example and
instead provide numerical results in graphical form for an essentially exact solution after 34 major (fixed-
point) iterations. In each major iteration we solve the partial differential equation
2
1 2 o V p1 oV p1
ðrp1 ½t; hk Þ 2
þ ðMP office
ij MP home
ij Þ¼0
2 oðS p1 Þ ot
with appropriate initial and boundary conditions where k is the iteration index. Fig. 26 below depicts the solu-
tion of the above partial differential equation as a response surface showing how option value varies as a func-
tion of time and traffic activity. In the above, we use the constant value of time m = $8, while the difference in
the marginal productivity of each traveler commuting from origin 1 to destination 4 is 5 and the call exercise
price is Ep1 ¼ 3:5. When call option trading occurs as described above, Figs. 27–29, respectively depict depar-
ture rates and arc exit flows for the paths p1, p2 and p3. Cumulative traffic volumes on the five different arcs of
60
Value of the Option
40
20
-20
-40
10
50
40
5 30
Time 20
10
0
Asset Price
0
Fig. 26. Plot of value of the option (V p ðt; hp Þ), asset price (S p ðt; hp Þ) and time (t) for path p1.
18
16
14
12
Flows
10
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
Flow p1 g1p1 g4p1

Flows
5
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
Flow p2 g2p2 g5p2
5
Flows
0
0 8 16 24 32 40 49 57 65 73 81 89 96
Time (Hour)
the network are plotted against time in Fig. 30. We compare net unit congestion cost and path flow in
Figs. 31–33 below, which make explicit that a dynamic user equilibrium has been attained. Furthermore,
20
18
16
14
12
10
0
0 6 12 17 23 29 35 40 46 52 58 64 69 75 81 87 93
Time (Hour)

1200 18
16
1000
14
12
Unit Travel Cost

800
10
600 8
6
400
4
2
200
0
0 -2
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 1 Flow p1
250 9
200
7
Unit Travel Cost
6
150
5
4
100
3
2
50
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 2 Flow p2
250 8
7
200
6
Unit Travel Cost
5
Path Flow
150
100
3
2
50
1
0 0
0 7 14 21 28 35 42 49 56 62 69 76 83 90 96
Time (Hour)
Cost 3 Flow p3
76 9
8
74
7
Net Travel Demand

72
6
70 5
4
68
3
66
2
64
1
62 0
0 1 2 3
Time (Day)
Net Q S1 Net Q S2 Net Q S3 Net Q S4
Reduction S1 Reduction S2 Reduction S3 Reduction S4
Fig. 34. Net travel demand and demand reduction for 4 sample paths.
net travel demand and demand reduction are plotted in Fig. 34 against the same time scale (day) for four
sample paths.
7.5. Comparison of congestion costs
For a specific sample path, we compute the net congestion costs for each sample path. This process is
repeated for 10 sample paths keeping all the parameters unchanged. For computing the transition probabil-
ities w(a, t, Kp) corresponding to a sample path x, we use the closed form expression (40) for the solution of the
partial differential equation (36) through (39) with initial condition Kp = 4, while drift and volatility are based
on hp1 ðt; xÞ as previously explained. The table immediately below expresses the efficacy of congestion call
options in reducing the net social costs of congestion by considering the four scenarios we have described
in detail above; those scenarios, it will be recalled, are: (1) no telecommuting, no market for congestion
and no options, (2) telecommuting with no congestion market and no options, (3) a market for congestion
with telecommuting but no options; and (4) telecommuting and a market for congestion with options. Our
findings are summarized below:
Cases Cong. Q14 Rs14 ~s

Q Qualitative analysis
14
cost 1 2 3 4 1 2 3 4
Scenario (1) 1196 75 0 0 0 0 75 75 75 75 Identical flow pattern
Scenario (2) 1185 75 0 1.45 3.45 4.56 75 73.5 71.5 70.4 Increased local congestion; slightly
decreased global congestion
Scenario (3) 1104 75 0 2.76 5.81 8.94 75 72.2 69.2 66.1 Decreased global congestion
(avg) (avg) (avg)
Scenario (4) 1077 75 0 2.78 5.36 7.92 75 72.2 69.6 67.1 Further decreased global congestion,
fewer telecommuters
(avg) (avg) (avg)
We see that introduction of telecommuting only decreases the total congestion cost by about 1%, primarily
because local congestion on some paths increases in some days. However, introduction of the call option low-
ers the expected net social cost of congestion (calculated over 10 sample paths) by 9% and also induces more
telecommuting. On the other hand, if we have only a market for congestion (scenario c), that too has the
potential to reduce the social costs of congestion.
8. Conclusions and future work
We have shown that there is an infinite dimensional variational inequality representation of dynamic user
equilibrium in the presence of European congestion call options. This formulation opens the door to adapta-
tion of computational schemes from the dynamic traffic assignment literature to the numerical analysis of con-
gestion options and their impacts. We have demonstrated by example that
1. the coupled computation of option prices and DUE flows may be accomplished with relatively simple meth-
ods (Euler’s scheme for stochastic differential equations and a fixed-point algorithm); and
2. congestion call options have the potential to lower the net social costs of congestion.
In our future work we intend to develop sufficiency conditions, which, when met, assure congestion securities
will in fact lower the net social costs of congestion. The numerical examples presented herein set the stage for
at least three new directions in transportation network flow modeling and financial engineering: markets for
congestion rights, the design of exotic congestion securities and improved algorithms for pricing those
securities.
References
Ching, W.-K., Li, X., Siu, T.-K., Wu, Z.-Y., 2007. Improving revenue management – a real option approach, pre-print.
Dial, R., 2003a. Towards faster algorithms for dynamic traffic assignment. part i: On arc-loading continuous origin–destination demands,
pre-print.
Dial, R.B., 2003b. Toward faster algorithms for dynamic traffic assignment. part ii. pre-print.
Friesz, T.L., Mookherjee, R., 2006. Solving the dynamic network user equilibrium problem with state-dependent time shifts.
Transportation Research Part B 40 (3), 207–229.
Friesz, T.L., Bernstein, D., Smith, T.E., Tobin, R.L., Wie, B.W., 1993. A variational inequality formulation of the dynamic network user
equilibrium problem. Operations Research 41, 179–191.
Friesz, T.L., Bernstein, D., Suo, Z., Tobin, R.L., 2001. Dynamic network user equilibrium with state-dependent time lags. Networks and
Spatial Economics 1, 319–347.
Friesz, T.L., Bernstein, D., Stough, R., 1996. Dynamic systems, variational inequalities and control theoretic models for predicting time-
varying urban network flows”. Transportation Science 30, 14–31.
Kachani, S., Perakis, G., 2002. Fluid dynamics models and their application in transportation and pricing. Tech. Rep., Sloan School of
Management, MIT.
Kober, R., Baumel, C.P., 1990. Transportation forward contracting or options: A description and analysis of an application to railroads.
Transportation Practitioners Journal 58 (1), 51–63.
Marcus, B., Anderson, C., 2006. Low prices guarantees – a real options analysis. Operations Research 54 (6), 1041–1050.
Montroll, E.W., Badger, W.W., 1974. Introduction to Quantitative Aspects of Social Phenomena. Gordon and Breach Science Publishers.
Peeta, S., Ziliaskopoulos, A., 2001. Foundations of dynamic traffic assignment. Networks and Spatial Economics 1 (3), 233–265.
Perakis, G., 2000. The dynamic user equilibrium problem through hydrodynamic theory. Tech. Rep., Sloan School of Management, MIT .
Sahin, O., Gallego, G., 2006. Inter-temporal valuations, product design and revenue management, pre-print.
Talluri, K.T., Ryzin, G.J.V., 2004. The Theory and Practice of Revenue Management. Springer.
Wilmott, P., Dewynne, J., Howison, S., 1993. Option Pricing. Oxford Financial Press.
Wilson, W., Priewe, S.R., Dahl, B., 1998. Forward shipping options for grain by rail. Journal of Agricultural and Resource Economics 23,
526–544.
Yan, S., Bernstein, D., Sheffi, Y., 1995. Intermodal pricing using network flow techniques. Transportation Research Part B 29 (3), 171–
180.

Securitizing Congestion The Congestion C

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Securitizing Congestion The Congestion C

Uploaded by

Copyright:

Available Formats

Available online at www.sciencedirect.

Transportation Research Part B 42 (2008) 407–437

Securitizing congestion: The congestion call option q

2. Some general remarks about dynamic user equilibrium models

1. a model of path delay;

3. A brief review of dynamic user equilibrium

p an abstract path of the network of interest

Pij set of paths connecting OD pair ði; jÞ 2 W

is the total arc volume.

To ensure realistic behavior, we employ asymmetric early/late arrival penalties

y ij ð0Þ ¼ 0 8ði; jÞ 2 W ð8Þ

y ij ðT Þ ¼ Qij 8ði; jÞ 2 W ð9Þ

4. The philosophy behind the congestion call option

4.1. Multiple time scales

4.2. A submodel of demand reduction through telecommuting

4.2.1. Cases when the regulatory constraints are binding

4.2.2. A possible model for rejecting demand for telecommuting

4.3. Option trading

4.4. Stochastic diﬀerential equations describing asset prices

See for example Wilmott et al. (1993).

4.4.1. Estimation of drift and volatility

4.5. The value function for the congestion call option

4.5.1. Itô’s lemma

for all p 2 P. As a consequence, our dynamics (17) lead to

4.5.2. The value function for a path

4.5.3. The arbitrage-does-not-occur condition

V sp ½S p ðsDÞ; sD ¼ m max½S p ðsDÞ  ap ; 0 8ði; jÞ 2 W; p 2 Pij ð29Þ

and V sp ½S p ðsDÞ; sD ! S sp ðsDÞ as S sp ðsDÞ ! 1 8p 2 P ð31Þ

5.1. Net trip value for the option purchasers

5.2. Net trip value for non-option purchasers

5.3. Properties of modiﬁed dynamic user equilibrium

1. Non-trivial path ﬂow requires maximal net value:

5.4. Fokker–Planck equation and expected net congestion costs

5.4.1. Closed form solution of the Fokker–Planck equation

The closed form solution of (36)–(39) is5

5.4.2. Expressions for conditional expectations

where mp and gp are deﬁned in (41) and (42), respectively.

Fixed-point algorithm for solving (34)

Step 0. Realization of random noise.

where e 2 R1þþ is a preset tolerance, stop and declare h*(x)

6.1. Approximating the expected value of net congestion costs

CðtÞ ¼ 1 if EV s1 ðt; h Þ P EV s2 ðt; h Þ

Consequently, the expected net congestion cost is

Using an ensemble comprised of a large number of samples, the approximation

7.1. Multi-period DUE model: no telecommuting and no options

Arc name From node To node Arc delay, Da(xa(t))

Paths Path flows Arc exit flows Traffic volume of arcs

We consider the symmetric early/late arrival penalty

7.2. Multi-period DUE model: telecommuting and no options

Fig. 2. Path and arc exit ﬂows for path 1.

Fig. 3. Path and arc exit ﬂows for path 2.

Fig. 4. Path and arc exit ﬂows for path 3.

Fig. 5. Cumulative arc volume vs. time.

Unit Travel Cost

Fig. 7. Similar comparison for path p2.

Fig. 9. Path and arc exit ﬂows for path 1.

Fig. 10. Path and arc exit ﬂows for path 2.

Fig. 11. Path and arc exit ﬂows for path 3.

Cumulative Arc Volume

Fig. 12. Cumulative arc volume vs. time.

Fig. 14. Similar comparison for path p2.

7.3. Market for congestion with telecommuting but no options

V sp ½S p ðsDÞ; sD ¼ m max½S p ðsDÞ ap ; 0 8ði; jÞ 2 W; p 2 Pij ð29Þ

CðtÞ ¼ 1 if EV s1 ðt; h Þ P EV s2 ðt; h Þ