European Journal of Operational Research 251 (2016) 586599

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Decision Support

Dynamic pricing of primary products and ancillary services

Fredrik degaard, John G. Wilson1
Ivey Business School, Western University, 1255 Western Road, London, ON N6G 0N1, Canada

a r t i c l e

i n f o

Article history:
Received 26 October 2013
Accepted 16 November 2015
Available online 8 December 2015
Keywords:
Revenue management
Ancillary services
Single-leg ight
Dynamic pricing
Customer choice

a b s t r a c t
Motivated by the growing prevalence for airlines to charge for checked baggage, this paper studies pricing
of primary products and ancillary services. We consider a single seller with a xed capacity or inventory of
primary products that simultaneously makes an ancillary service available, e.g. a single-leg ight and checked
baggage service. The seller seeks to maximize total expected revenue by dynamically setting prices on both
the primary product and the ancillary service. In each period, a random number of customers arrive each of
whom may belong to one of three groups: those that only want the primary products, those that would buy
the ancillary service if the price is right, and those that only purchase a primary product together with the
ancillary service. A multi-period dynamic pricing model is presented with computational complexity only
of order equal to the number of periods. For certain distributions, close to analytical results can be obtained
from which structural insights may be gleaned.
2015 Elsevier B.V. All rights reserved.

1. Introduction
Consumers expectations of what should be included in a purchased service vary greatly. At sporting or entertainment events,
nobody expects to be served complementary food or beverages. Similarly, for car-rentals, hotels, and cruises (including all-inclusive), consumers generally have little expectation of receiving ancillary services, such as car insurance, GPS, meals or wi, for free. However,
at the opposite spectrum, for air travel, consumers are, or rather
were, generally accustomed to receive complementary ancillary services (such as food and beverages, inight entertainment, and the
most prominent bone of contention: checked baggage service). For
example, KLM, Air France, Delta Air Lines, American Airlines, and Air
Canada now all charge for checked baggage on their respective domestic European and North American economy class ights; KLM
and Air France charges Euro15 for each piece, while Delta Air Lines,
American Airlines, and Air Canada charge $25 and $35 for the rst
and second checked bag, respectively.2 Even the US low-cost carrier Southwest Airlines, which markets ight service with two free
checked bags charges for additional bags beyond two (Delta charges
$125, American $150, Air Canada $100, and Southwest $50.) In sharp
contrast to Southwest Airlines, the US ultra-low-cost carrier Spirit

Corresponding author. Tel.: +1 519 661 4278.

E-mail addresses: fodegaard@ivey.uwo.ca (F. degaard), jwilson@ivey.uwo.ca
(J.G. Wilson).
1
Tel.: +1 519 661 3867.
2
On domestic Canadian routes, Air Canada charges $25 for the second.
http://dx.doi.org/10.1016/j.ejor.2015.11.026
0377-2217/ 2015 Elsevier B.V. All rights reserved.

Airlines charges not only for checked baggage ($30) but also for carryon baggage ($35).
Revenue from ancillary services such as baggage fees is of growing importance to the airline industry. For example, the total baggage fee revenues for US airlines has over the past six years grown
to a multi-billion dollar industry; $464M (2007), $1,149M (2008),
$2,729M (2009), $3,395M (2010), $3,361M (2011), $3,487M (2012),
$3,350M (2013), $3,529M (2014).3 Common arguments from airlines
for charging for checked baggage are increasing fuel costs and to allow customers greater price exibility. It is lesser known that many
passenger airlines also transport cargo. Therefore low or zero baggage fees can represent an opportunity cost since fewer checked bags
means more cargo space. The operational revenue from cargo transportation varies considerably among passenger airlines. For instance,
for Delta and American, cargo represents only about 3% of the total
operating revenue (AMR Corp, 2011; Delta Air, 2011), while reportedly for the top eight Asian airlines, cargo represents on average 30%
of total revenue (Wong, Zhang, Hui, & Leung, 2009).
In this paper, we present a model of a single seller with a xed capacity or inventory of homogeneous products who also provides an
ancillary service, e.g. an airline that offers checked baggage service
or a hotel that provides wi. We consider a nite, discretized timehorizon over which the seller for each period seeks to set the prices
of the primary items and ancillary service in order to maximize expected total revenue. We assume in each period a random number of
customers arrive and that customers randomly belong to one of three

3
http://www.rita.dot.gov/bts/sites/rita.dot.gov.bts/les/subject_areas/airline_
information/baggage_fees/index.html; accessed 2015-08-27.

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

groups: those who only want a primary item and who would not consume the ancillary service even if it were free (Group 1), those who
would potentially be willing to pay for both the primary item and
ancillary service but who equally well would settle for just the primary item (Group 2), and those that only want a primary item as long
as they also receive the ancillary service (Group 3). The objectives of
the paper are: (1) to discuss how to derive optimal prices in order to
maximize the expected total revenue, and (2) to analyze optimality
conditions and the potential gain in charging separately for ancillary
services versus a single charge for the bundle of the primary item and
the ancillary service.
Given the importance and growing trend to charge for ancillary
services, the paper makes several contributions. First, we provide a
dynamic pricing model for evaluating the expected total revenue and
propose an easy to implement algorithm for deriving the optimal dynamic prices. For the special case with uniform distributions and no
capacity constraint, we derive close to closed form solutions. Second, we show that the optimal prices for the unconstrained capacity problem can often serve as a good approximation for the capacity constrained problem. Third, through numerical analysis based on
the uniform distribution we show that the incremental gain in charging separately for the ancillary service is higher when the proportion
of Group 3 type customers is lower. In other words, the incremental
gain for airlines/hotels to impose ancillary fees is smaller when the
proportion of customers with demand for the ancillary checked baggage/wi service is high. Instead the largest incremental gain for a
dual pricing strategy is when the majority of customers do not demand the ancillary service.
1.1. Literature review
Revenue Management (RM), formerly referred as Yield Management, has over the last forty years experienced a tremendous growthboth in industry and academic research. For a general overview and
background, see Bitran and Caldentey (2003); Boyd and Bilegan
(2003); McGill and van Ryzin (1999); Talluri and van Ryzin (2004b).
Although there is a rich literature on both capacity based and pricing
based revenue management, as rms and industries develop innovative business solutions new research opportunities arise. A prime
example is the sale of ancillary or secondary products and services a relatively undeveloped research area.
The modeling framework presented in this paper contributes to
three major components of the RM literature. First, we are considering dynamic pricing over time but extend the setting to include ancillary services. To the best of our knowledge this is the rst paper to
do so. Thus far the main focus has been on the dynamic pricing of the
primary product, e.g. Zhang and Cooper (2009) who present a Markov
Decision Process model for dynamic airline pricing. Second, we consider a stochastic customer arrival process. Traditionally, much of
the literature regarding airline pricing or capacity allocation is based
on Poisson arrivals of customers and formulated as a multi-period
Network Revenue Management problem; see Gallego and van Ryzin
(1997); Ch. 3 of Talluri and van Ryzin (2004b); Maglaras and Meissner (2006); Kunnumkal and Topaloglu (2010); Meissner and Strauss
(2012). The general idea is to consider a number of small time periods
where the probability of a single customer is small such that no more
than one arrival is possible. Optimal decisions and total expected revenue are then found by recursively solving the Bellman equations. In
this paper, we model the problem in a different manner. The benet
with our probabilistic approach is that it allows us to formulate the
problem as a straightforward one dimensional optimization problem
from which structural properties can be derived, e.g. uniqueness of
the optimal prices, etc. Indeed, in special cases, we can obtain results
that are close to closed form. Additional benets is that our model allows for a general arrival process (not just Poisson), while not adding
any computational complexity. Third, our model framework is based

587

on an underlying customer choice model. Recent development within

the RM literature include to directly model underlying consumer behavior or choices, see Talluri and van Ryzin (2004a) and chapters 2.6
and 7 of Talluri and van Ryzin (2004b); Kunnumkal and Topaloglu
(2010); Chen and Homem-de-Mello (2010); and Cooper and Li (2012).
Two papers directly related to our research objective include
Shulman and Geng (2012) and Allon, Bassamboo, and Lariviere
(2011). Shulman and Geng (2012) analyze a duopoly game between
two asymmetric rms that offer a base good and an add-on to a set of
heterogeneous consumers. They segment the consumers into three
groups: (i) a base group that does not buy the add-on, (ii) a group of
knowledgeable consumers who buys the add-on if the add-on price
is less than their willingness-to-pay, and (iii) a group of bounded rational consumers who were expecting to receive the add-on for free
and who will only buy the add-on as long as their willingness-topay is higher than the total price. It is assumed all three consumer
groups purchase a base good. Based on the assumption that consumers willingness-to-pay is uniformly distributed, Shulman and
Geng (2012) derive equilibrium prices for the base good and the
add-on.
The paper by Allon et al. (2011) looks at the problem from a different perspective, namely from a coordinated social perspective of
pricing the main service and ancillary service such that both consumer surplus and rm prot is maximized. They characterize, under both homogeneous and heterogeneous customer settings, when
a two-part tariff versus bundling is socially optimal. They extend their
results to a setting with multiple sellers and show that, at equilibrium, pricing of the ancillary service is set at the marginal cost of
providing the service. One of the main differences between the modeling framework of Allon et al. (2011) versus our paper and Shulman
and Geng (2012) is that customers do not have an explicit value for
the ancillary service. Instead, Allon et al. (2011) assume all customers
have a cost-driven likelihood for wanting/needing the ancillary service regardless of price (where the probability is decreasing in an
effort level which indirectly depends on the price of the ancillary
service).
Although our framework shares some common features with
Shulman and Geng (2012) and Allon et al. (2011), there are ve signicant differences. First, while both Shulman and Geng (2012) and
Allon et al. (2011) restrict attention to a single-period setting, we consider a multi-period dynamic pricing setting. Second, we consider the
case of a possible capacity constraint on the number of primary items
(i.e. base good or main service). In other words, we do not assume all
consumers are guaranteed a primary item. Third, we extend the heterogeneity of consumers so that not all consumers are assumed to
purchase a primary item if one is available. Fourth, although we include numerical illustrations based on the uniform distribution, our
main discussion centers on a general modeling framework without
any specic assumption regarding the distribution of the willingnessto-pay. Finally, our overall research objectives are different. For instance, Shulman and Geng (2012) seek to address how add-on pricing
affects rm prots and consumer surplus in an asymmetric duopoly
game with the three consumer groups. Implicit in their modeling
framework and explicit in their results is that rms charge for the
add-on. In contrast, one of our objectives is to analyze under what
conditions a rm should charge for the ancillary item (or add-on).
The motivation for this is to analyze the Delta/American Airlines
versus Southwest Airlines business strategies of checked baggage
service.
Framing the airline baggage fee problem as a bundling and twopart tariff problem, there is a vast literature spanning economics (e.g.
the classic Oi (1971), and McAfee, McMillan, and Whinston (1989)),
management science (e.g. Banciu, Gal-Or, & Mirchandani (2010)), and
marketing (e.g. Essegaier, Gupta, & Zhang (2002)). From the four
listed, the one most related to our paper is Essegaier et al. (2002),
who considers a single-period pricing strategy of access services (e.g.

588

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

gym membership, phone and internet service). Similar to our paper,

Essegaier et al. (2002) are interested in analyzing when at fee versus two-part tariff pricing is optimal. Although there are differences
to our modeling framework (e.g. we consider multi-period dynamic
pricing problem while they consider a single-period problem), we
share some similarities. For instance, Essegaier et al. (2002) also consider two types of heterogeneous customers (heavy vs. light users),
and a seller (service provider) constrained by limited capacity. On the
other hand, Essegaier et al. (2002) exclusively consider uniformly distributed willingness-to-pay (preferences), while we consider a general framework. Interestingly, even for the case with uniform distributions, we have some contrasting results, such as, when the seller
has no capacity constraint, for Essegaier et al. (2002) a at fee pricing
is then optimal, while we show that imposing a two-part tariff (e.g.
charging for the baggage service separately ) may generate substantial revenue gain.
It may at rst seem the bundling problem is different from the airline baggage fee problem, since it might seem odd to consider an airline selling baggage service unbundled, i.e. an airline selling checked
baggage service without the need to also purchase an air fare ticket.
However, as mentioned earlier, passenger airlines do in fact sell excess carrying capacity for cargo transportation on spot markets. Since
cargo transportation is a separate business issue and currently a relatively minor contribution to the operating revenue for many airlines, we restrict our analysis to the special partial mixed bundling setting where the ancillary service can only be sold as part of a bundle
with the primary product. For general air cargo RM, see Amaruchkul,
Cooper, and Gupta (2007); Han, Tang, and Huang (2010); and Ch.10.12
of Talluri and van Ryzin (2004b). Regarding the joint operation of passenger and air cargo revenue management, we are only aware of the
paper by Wong et al. (2009).
Two related empirical papers include Brueckner, Lee, Picard, and
Singer (2015) and Nicolae, Arkan, Deshpande, and Ferguson (2013).
Brueckner et al. (2015) empirically analyze whether the air fares were
affected by the introduction of checked baggage fees. In particular,
whether air fare decreased due to the unbundling of the ancillary service. Nicolae et al. (2013), motivated by the effect on operational performance by (presumably) having to handle fewer bags, investigates
how the introduction of baggage fees affects the departure delays and
scheduled block times.
Finally, in addition to modeling based research there is also a
vast literature within behavioral economics, marketing and consumer
psychology which through empirical and experimental studies investigate the effect of prices on consumer choices. For an overview
of some behavioral issues related to the problem we study, see
Hamilton, Srivastava, and Abraham (2010).
2. Problem description
We consider a seller who over a xed time-horizon of T periods
wants to sell a xed inventory (capacity) of C homogeneous primary
items; T < . In addition to the primary items, the seller also makes
available an optional secondary or ancillary service to those who have
already purchased a primary item. For instance, an airline may offer checked baggage service at an extra fee or a hotel might provide
wi for an additional charge. We index the periods by t and order
them in reverse such that the rst period is t = T, and the nal period is t = 1. At the end of period t = 1 the resource perishes, e.g. the
ight departs. The sellers objective is to maximize total expected revenue by setting optimal prices of the primary item p1, t and ancillary
service p2, t , for t = 1, . . . , T . We let pt represent the vector of prices
in time-period t, pt = ( p1,t , p2,t ), t = 1, . . . , T . No additional inventory decision or constraints on the ancillary service are imposed, i.e.
the available capacity for the ancillary service is no less than the

capacity C of the primary items. Note that a special case of our problem formulation is when p2, t is xed throughout the planning horizon T (and only p1, t is dynamically updated each period). For instance,
currently most airlines and hotels tend to charge a xed baggage and
wi fee, respectively, and only dynamically change the air fare and
room rate, respectively. In Section 3.3 we show that this pricing strategy is in fact optimal when the underlying willingness-to-pay of customers is uniformly distributed.
It is assumed the customer base consists of three groups: those
that are only interested in purchasing a primary item (Group 1), those
that in addition to the primary item are interested in purchasing the
ancillary service if the total price p1,t + p2,t is attractive but who
would be ne with just the primary item (Group 2), and those that
are interested in purchasing a primary item only if they can purchase an ancillary item as well (Group 3). For example, one could
think of airline passengers as those who only want a seat and do
not need to check a bag, those who would check a bag if the total
price is within their willingness-to-pay but who are able to do without checked baggage service, and those who must check a bag. Denote the probability a customer belongs to Group i by i , i = 1, 2, 3,
where 1 + 2 + 3 = 1. It should be noted that we are not considering price or fare class segmentation of customers. For example, for the
airline illustration, we are assuming a single fare class (p1, t ) but with
three distinct groups of customers: all types of travelers pay p1, t for
the ticket but in addition a Group 2 customer might pay and Group
3 customer will pay p2, t for checked baggage service if a ticket is
purchased.
We assume customers in all groups are heterogeneous with respect to their willingness-to-pay (wtp). The distribution function of
the willingness-to-pay for Group i is denoted by Fi (), i = 1, 2, 3. We
assume Fi () are (strictly) increasing and continuous and that all customers are independent. For notational convenience we restrict the
analysis to the case when Fi (), i = 1, 2, 3, (as well as i ) are stationary
and do not depend on time t. The extension to non-stationary distribution functions is straight-forward and all results remain tractable;
more details are provided in the Conclusion. If a primary item is available, then a customer in Group 1 who has a wtp at least as high as
p1, t will always purchase a primary item, a customer in Group 2 who
has a wtp at least as high as p1, t will also always purchase a primary
item (if in addition the Group 2 customers wtp is at least as high as
p1,t + p2,t then they will purchase both a primary item and the ancillary service), and customers in Group 3 with a reservation price of at
least p1,t + p2,t will purchase both a primary item and ancillary service. Note that for Group 2 and Group 3 customers the distribution
functions F2 () and F3 (), are for the combined bundle of the primary
product and ancillary service, but with the distinction that a Group 2
customer with a wtp between p1, t and p1,t + p2,t will only purchase
the primary product (without the ancillary service) while a Group 3
customer with the same wtp would not purchase anything.
We assume payments for both the primary item and ancillary service are made at the time of booking. Although many airline passengers pay for checked baggage service at the time of check-in,
most airlines provide opportunities and incentives for passengers to
pre-pay at the time of booking. We also assume that customers arrive at random and that the seller cannot distinguish or discriminate between customers from the three groups upon arrival. So, for
instance, it might be the case that a higher paying customer from
Group 3 is turned away because a lower paying Group 1 customer
has taken the last available item or vice-versa. Furthermore, we assume the seller cannot ration or reserve capacity among the three
groups.
If a primary item is available in period t then, since customers arrive at random, the probability the next sale is a Group i customer is
given by i (1 Fi ( p1,t )), for i = 1, 2, and 3 (1 F3 ( p1,t + p2,t )), for

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

i = 3, t = 1, . . . , T . We dene the overall probability of a sale in period

t by,

zt (pt ) =

1 (1 F1 ( p1,t )) + 2 (1 F2 ( p1,t ))
+ 3 (1 F3 ( p1,t + p2,t )),

Transition Probabilities In each period t, for a given state c and set

of prices p, the transition probabilities are given by (2),

Pr{ct1 = c x|ct = c, pt = p}

x = 0, 1, 2, . . . , c 1,
gt (x|p )
= (1 Gt (ct 1|p )) x = c,

(1)

and the probability of no sale by 1 zt (pt ), t = 1, . . . , T .

We assume that in each period t a random number Nt of potential
customers arrive, t = 1, . . . , T . We impose no specic arrival process
and allow for any general stochastic arrival process (including nonhomogeneous with respect to time). We denote the probability that
exactly n potential customers arrive in period t by Pr[Nt = n], n =
0, 1, 2, . . .; t = 1, . . . , T . To avoid trivial cases we assume E[Nt ] < ,
for all t. Let Mt (pt ) be the total (random) demand in period t given the
prices pt , t = 1, . . . , T . In other words, Mt (pt ) is the (random) subset
out of the Nt customers who are willing to buy a primary item (possibly with the ancillary service) if a primary item is available, i.e. the
total number of Group 1, 2, and 3 customers who are willing to purchase a ticket. Consequently, the probability mass function for Mt (pt )
is given by,

gt (x|pt ) Pr[Mt (pt ) = x]

 
 n
zt (pt )x (1 zt (pt ))nx Pr[Nt = n],
=
x
nx

589

(2)

for t = 1, . . . , T, with an associated cumulative distribution function

given by Gt (x|pt ) Pr[Mt (pt ) x].
Note that we do not know exactly how many (potential) customers will arrive each period, to which group they belong or their
willingness-to-pay. Furthermore, we do not assume the seller has unT
limited capacity but instead include the possibility that C < t=1
Nt
and hence the seller cannot necessarily serve all demand. In other
words, the seller may incur lost sales and, in particular, may lose
sales to customers willing to pay more. In Section 4 we analyze the
simpler case when capacity is not a constraining factor and the
seller could sell a primary item to everyone. Since customers from
the three groups arrive at random, there are many arrival combinations to consider. Consequently, formulating the sellers expected total prot is not immediate. However, a work-around is to condition on
the realized demand and evaluate the expected revenue per demand
realization. Finally, although the dynamic pricing over several timeperiods provides additional combinatorial complexity which usually
suffers from the curse of dimensionality, we provide a solution that
enables the computational complexity to only grow linearly in the
length of the time-horizon T.
3. The multi-period model
In this section, we formulate a multiperiod dynamic pricing
model. By concentrating only on the probability of a sale during a
period, we show that only a one dimensional decision variable need
be considered. This is achieved by looking at the maximum expected
revenue per sale at a given probability of sale. Section 3.3 contains an
illustration of the approach. Multiperiod optimization and an illustrative example are provided in Section 3.2. The computational complexity of obtaining solutions only increases linearly in the number
of periods.
We formulate the problem as a Dynamic Program: Decision
Epochs We assume the planning-horizon consists of a xed and
nite number or periods T, indexed by t = 1, . . . , T . We assume decisions are made at the beginning of each period before a random
number Nt of potential customers arrive. State space At the beginning of each period t the seller observes the amount of remaining
capacity ct , ct = 0, 1, 2, . . . , C. Actions At the beginning of each period t the seller has to set non-negative prices for both the primary
item, p1, t , and ancillary service, p2, t . We let pt denote the action in
period t = 1, . . . , T . Rewards In each period t, for a given state ct
and set of prices pt , the sellers expected revenue is given by t (pt , ct ).

otherwise.

Optimality equations Let Vt (c) be the optimal expected revenue from the last t remaining periods given a capacity c at the beginning of period t, for t = 1, . . . , T,

Vt (c ) = sup

t (pt , c ) +

p1,t ,p2,t

c1


Vt1 (c x )gt (x|pt )) ,

(3)

x=1

with the boundary conditions that Vt (0 ) = 0 for all t, and V0 (c ) = 0

for all c.
Using standard backward induction we could solve for optimality and derive the optimal prices for each period t and every state c.
Unfortunately, this entails the standard curse of dimensionality associated with many pricing problems solved using dynamic programming. Therefore, we now explicitly transform the expressions to reect the approach that the optimization equations can be reduced
from a two dimensional optimization over pt to a one dimensional
optimization over [0, 1].
Dene Dt (z) to be the (random) number of people willing to buy
a primary item during period t when one is available given that the
probability of any primary item sale equals z, t = 1, . . . , T . Since the
Fi () are (strictly) increasing and continuous, i = 1, 2, 3, it is always
possible to nd at least one set of prices pt = ( p1,t , p2,t ) such that,
for a given z [0, 1],

z = 1 (1 F1 ( p1,t )) + 2 (1 F2 ( p1,t )) + 3 (1 F3 ( p1,t + p2,t )).

(4)
For a given z, we denote P(z) as the set of price pairs (p1 , p2 ) such that
(4) holds,

P(z ) = {( p1 , p2 ) | 1 (1 F1 ( p1 )) + 2 (1 F2 ( p1 ))
+ 3 (1 F3 ( p1 + p2 )) = z}.
The probability mass function of Dt (z) is given by, for t = 1, . . . , T,

 

ht (x|z ) Pr[Dt (z ) = x] =

 n
zx (1 z )nx Pr[Nt = n].
x
nx

(5)

Let Ht (x|z ) Pr[Dt (z ) x] denote the corresponding cumulative distribution function. Note that Expressions (2) and (5) have identical
functional form but involve different arguments, which is why we use
the two different notations gt (x|p) and ht (x|z) to distinguish them. In
(2) the argument is the prices p1, t and p2, t , whereas in (5) the argument is the probability of sale z.
Dene (z) as the maximum expected revenue per customer sale for
a given likelihood of a sale z, i.e. for z [0, 1],

(z )
=


sup

( p1 ,p2 )P(z )

p1 + p2

2 (1 F2 ( p1 + p2 )) + 3 (1 F3 ( p1 + p2 ))
z


. (6)

Note that the supremum is taken over all (p1 , p2 ) that result in the
given probability of sale z. Furthermore, the optimality Eq. (3) can
now be reduced to one-dimensional optimizations over the probability of a sale. We formally summarize this result in the following
proposition.4
4
This result is similar to Prop.1 by Maglaras and Meissner (2006) regarding optimal
aggregate consumption rates for multi-product dynamic pricing problems. We would
like to acknowledge an anonymous reviewer for bringing this to our attention.

590

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

Proposition 1. The optimal expected total revenue for the last t =

1, . . . , T periods is given by,

(z ) E[min(Dt (z ), c )] +

Vt (c ) = sup

0z1

c1


Vt1 (c x )ht (x|z ) ,

x=1

(7)
with the boundary conditions Vt (0 ) = 0, for all t, and V0 (c ) = 0, for all c,

and E[min(Dt (z ), c )] = c1
x=0 xht (x|z ) + c (1 Ht (c 1|z )), and (z)
given by (6).
Observe that (z) does not depend on the time-index t and that
E[min(Dt (z ), c )] only depends on t to the extent that the arrival process is non-stationary. If the arrival process is stationary so Nt does
not depend on the time-index t, then (z ) E[min(Dt (z ), c )] does
not depend on t which helps to simplify the multiperiod problem.
Another consequence of this formulation is that no assumption with
regard to the arrival order of the three groups needs to be made. Next
we show how the prices that maximize (z) can be derived.

Proposition 2. For a given z [0, 1], if lz uz , then

(z ) = pz1 + pz2
where,

2 (1 F2 ( pz1 + pz2 )) + 3 (1 F3 ( pz1 + pz2 ))

z

pz ( pz1 , pz2 ) =

(qz1 , qz2 ) if l z < uz ,

(uz , 0 ) otherwise.

(10)

(11)

Note that neither the capacity C nor the number of potential customers Nt affects the calculation of pz for the expected revenue per
customer sale. Furthermore, for certain distributions, (z ), pz1 , and
pz2 can be provided in closed form. An illustration of this is provided
in Section 3.3 for uniformly distributed valuations. Next we discuss
how we solve the optimality Eq. (7) using backward induction to derive the optimal prices p1,t and p2,t for each period t = 1, . . . , T .
3.2. Solving for optimal dynamic prices
Given a zero terminal reward V0 (c ) = 0, for all c, the optimality
equation for the nal period is given by,

3.1. Expected revenue per customer sale

The objective of this section is to, for a given period t and probability of sale z, derive the set of prices pt = ( p1,t , p2,t ), that maximizes
the expected revenue per customer sales (z ) = p1,t + p2,t (2 (1
F2 ( p1,t + p2,t )) + 3 (1 F3 ( p1,t + p2,t )))/z. Since the distribution
functions Fi (), i = 1, 2, 3, are stationary with respect to time, and we
are only interested in maximizing the conditional expected revenue
given a sale within a period, we may omit the subscript t.
For a given value of z [0, 1], dene the single price uz implicitly
via the following equality,

V1 (c ) = sup

0z1

{ (z ) E[min(D1 (z ), c )]}.

(12)

This is an one-dimensional optimization problem that can be solved

using a standard numerical search. Let z1 be the argument that
maximizes (12), i.e. z1 arg maxz { (z ) E[min(D1 (z ), c )]}. In other
words, for a given remaining capacity c of primary items in the nal
period, z1 is the optimal sales probability level, and the optimal valueto-go is given by,

z = 1 (1 F1 (uz )) + 2 (1 F2 (uz )) + 3 (1 F3 (uz )).

V1 (c ) = (z1 ) E[min(D1 (z1 ), c )],

In other words, uz is the price of the primary item that together

with p2 = 0 results in the probability of a sale to be z. Note that uz
is well-dened since the Fi () are strictly increasing and continuous,
i = 1, 2, 3. There may be other combinations of p1 and p2 that result
in the probability of a sale being z but any such pair of prices must satisfy p1 uz , since 1 (1 F1 ( p1 )) + 2 (1 F2 ( p1 )) + 3 (1 F3 ( p1 +
p2 )) is strictly decreasing in p1 and p2 . In order to simplify notation,
but without loss of generality, dene F() as the following mixture
distribution,

where (z1 ) is given by (10), and the optimal prices for the nal period p1 = ( p1,1 , p2,1 ), are given by (8) and (9). By backward induction, we solve the optimality Eq. (7) by nding the optimal probability
of a sale for each period t,

F (x ) =

1 F1 (x ) + 2 F2 (x )
.
1 + 2

 + z 
1
2
,0 .
1 + 2

Then any p1 [l z , uz ], is a feasible primary price, since p1 and the

ancillary price q2 (z, p1 ) dened by (as shown in the Appendix),

q2 (z, p1 )

F31

1 z (1 + 2 )F ( p1 )

p1 ,

is feasible, unique and satises (1). Now dene qz1 to be the result of
a one-dimensional maximization:

qz1

arg max p1 + q2 (z, p1 )

c1


Vt1 (c x )ht (x|z )}.

x=1

(13)
Solving these one-dimensional optimization problems generate the
optimal probability of a sale zt for each period t, which will correspond to a set of optimal prices pt = ( p1,t , p2,t ). Consequently, for
t = 1, . . . , T,

Given z [0, 1], dene a lower-bound lz by,

l z = F 1 max

zt arg maxz { (z ) E[min(Dt (z ), c )] +

Vt (c ) = (zt ) E[min(Dt (zt ), c )] +

and dene qz2 to be the value of q2 (z, p1 ) evaluated at qz1 ,

qz2 q2 (z, qz1 ).

(9)
qz1 ,

qz2

Given that the probability of a sale is set to z, then

are the primary and ancillary prices, respectively, that provide the highest expected revenue per customer, i.e. the prices that maximize the right
hand side of (6). We summarize this in the following result.

Vt1 (c x )ht (x|zt ).

x=1

The computational benet of solving the optimality equations

with respect to z rather than pt is clear: we only require to derive
the prices associated with each level of z once, since in each period
t the prices that maximize (z) do not change. Note too that capacity level c and (random) number of potential customers Nt only

2 (1 F2 ( p1 + q2 (z, p1 ))) + 3 (1 F3 ( p1 + q2 (z, p1 )))

p1

c1


(8)

c1
affect the calculation of E[min(Dt (zt ), c )] and
x=1 Vt1 (c
x )ht (x|z ), and that these computations scale very well. We would
therefore expect that for reasonable C and Nt (e.g. in the hundreds
or thousands), computing power and time will not be a limiting
factor. To provide further intuition and illustrate the proposed
approach we next discuss the specic case when the underlying
willingness-to-pay is uniformly distributed.

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

3.3. Uniformly distributed willingness-to-pay

In order to simplify the notation and provide better intuition
we ignore Group 2 by assuming 2 = 0; hence, 3 = 1 1 . Let the
willingness-to-pay for both Group 1 and Group 3 customers be uniformly distributed. Without loss of generality, we standardize the
problem with Group 1 and Group 3 customers willingness-to-pay being U[0, 1] and U[0, a], a 1, respectively. There are two cases for the
upper-bound uz depending on the value of z:

u =

a
11

a
a1 +11

a
a1 +11

if 0 < z (1 1/a )(1 1 ),

z

if

(1 1/a )(1 1 ) < z 1.

 z 

The lower-bound is given by l z = max 1 ,0 . Thus, given z, there

1
are three feasible regions of p1 to consider: (I) 0 < z min{(1
1/a )(1 1 ), 1 }, (II) min{(1 1/a )(1 1 ), 1 } < z max{(1
1/a )(1 1 ), 1 }, (III) max{(1 1/a )(1 1 ), 1 } < z 1.
For a given z, the ancillary price corresponding to a feasible p1 is
given by,

q2 (z, p1 ) =

a
11

a
11

z p1

a
11

 a1 +11 
11

if 1 p1 a
p1

if 0 p1 < 1.

Incorporating the above results in the following two expressions for

the expected revenue per customer sale,

 a 
if 1 p1 a
a 11 z
11
p1 ) p1 q2 (z, p1 )
p1 + az (aq2 (z,
(q2 (z, p1 ))2
if 0 p1 < 1.

Note that the rst case is independent of p1 , while the second case is
quadratic in p1 . In other words, any pair of prices ( p1 , p2 ) = ( p1 , a
az/(1 1 ) p1 ), where p1 1, generates the same expected revenue (z ) = p1 + p2 . This is intuitive since for p1 1 all sales are to
Group 3 customers who only purchase the primary product with the
ancillary service. The quadratic expression is maximized at,

a
1


p1 ( z ) = +
2
2(a1 + 1 1 )

a
z.
a1 + 1 1

As expected, we see that 

p1 (z ) is decreasing in z; a lower price results from a higher probability of a sale. A perhaps less intuitive result
is that the maximizing p1 is linearly decreasing in z. It can be verip1 (z ) uz , and hence that 
p1 (z ) is within the permissied that l z 
ble region. However, there are two additional thresholds with respect
to z,

1
1
z = (1 1 ) 1
,
2
a


1
1 1
z=
1 + 1 +
,
2
a
that denes the region for which 
p1 (z ) [0, 1], i.e. if z z then

p1 (z ) 0. Note that z (1 1 )(1
p1 (z ) 1, and if z z then 
p1 (z ) 0 we
1/a ), and 1 z (and of course that z z). When 
p1 (z ) 1, we set pz1 = a az/(1 1 ) (a
set pz1 = 0, and when 
a1 1 + 1 )/(2(1 1 )). In other words, there are three cases for
the maximizing primary price,

pz1 =

a
1

aa1 1+1
a
(11 ) z 2(11 )
a
a
a1 +1
1
2(a1 +11 )

if 0 z z
z

if z < z < z
if z z 1,

and two cases for the maximizing ancillary price,

pz2

aa1 1+1
2(11 )
a
1a1
11

if 0 z < z
z

if z z 1.

591

To summarize, for positive maximizing primary prices the maximizing ancillary price is a xed constant and independent of z, while
when the maximizing primary price is zero the maximizing ancillary price is linearly decreasing in z. Note that the pricing solution
for z z, is not unique, and we have presented the pricing strategy with a xed constant ancillary price and variable primary price.
Other possible pricing strategies for z z, include to not charge for
the ancillary service and instead charge a higher price for the primary product (pz1 = a az/(1 1 ), pz2 = 0), and charging a xed
primary price and instead varying the ancillary price (e.g. pz1 = 1,
pz2 = a az/(1 1 ) 1).
Given pz1 and pz2 we can simplify the expression for (z),

(z ) =

 a 
a 11 z
if 0 z z

a
1 + a+(a1)(11 ) 
z
2
2(a1 +11 )
2(a1 +11 )
1 )(a1 +21 )

+ (a1)(4(1

a1 +11 )z

a(1+1 ) a z a1
11

11

if z < z < z
if z z 1.

(11 )z

We can see that for the two cases when z < z, that (z) is strictly
decreasing in z, while for the case when z z further consideration
regarding the particular value of 1 must be given. We illustrate the
above discussion with a numerical example.
Numerical Example - Let a = 5/4, 1 = 3/4, and 3 = 1/4. Then,
from above,

uz =

5/4 5z

if 0 < z 1/20,

20/19 20z/19 if 1/20 < z 1.

For a given z, the possible ranges for p1 are given by,

[1 4z/3, 5/4 5z]

if 0 < z 1/20,

[1 4z/3, 20/19 20z/19]

if 1/20 < z 3/4,

[0, 20/19 20z/19]

if 3/4 < z 1,

and the ancillary price corresponding to a feasible p1 given by,

q2 (z, p1 ) =

5/4 5z p1

if 1 p1 5/4

5 5z 19p1 /4 if 0 p1 < 1.

From this we nd,


p1 (z ) = 39/38 20z/19.
By inspection, we see that if z [1/40, 39/40], then 
p1 (z ) [0, 1], resulting in the following three cases of maximizing prices,

if 0 z 1/40
(9/8 5z, 1/8 )
( pz1 , pz2 ) = (39/38 20z/19, 1/8 ) if 1/40 < z < 39/40

(0, 5 5z )
if 39/40 z 1,

and corresponding expression for the expected revenue per customer

sale,

(z ) =

5/4 5z

if 0 z 1/40

40/38 20z/38 + 35/4864z

if 1/40 < z < 39/40

35/4 5z 15/4z

if 39/40 z 1.

Recall that for z 1/40 the maximizing pair of prices ( pz1 , pz2 ) =
(9/8 5z, 1/8 ) is not unique, and that any pair of prices ( pz1 , pz2 ),
such that pz1 1 and pz2 = 5/4 5z pz1 , will also maximize the expected revenue per customer sale; e.g. ( pz1 , pz2 ) = (1, 1/4 5z ).
Fig. 1 illustrates the maximizing prices pz1 and pz2 for two numerical examples: (I) a = 5/4, 1 = 3/4, 3 = 1/4, and (II) a = 2, 1 =
1/4, 3 = 3/4. The rst example, with its relatively low ancillary
price, is in line with current practice of most North American legacy
airlines (with a $25 baggage fee). The second example, where the ancillary price is rather high and in particular for large z even above the

592

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

Fig. 1. Numerical illustration of conditional prices pz1 and pz2 given probability of sale z when Group 1 wtp is U[0,1], and Group 3 wtp is U[0,a].

Fig. 2. Numerical illustration of the value function in period 6 for different capacity level c; Group 1 wtp is U[0,1], and Group 3 wtp is U[0,a].

primary price, is consistent with the pricing strategy of many lowcost airlines such as Spirit Airlines and Ryanair (where the baggage
fee may exceed the air fare).
Based on the two numerical examples we next illustrate the value
function and dynamic pricing strategies. The illustration is based
on a time-horizon of ten periods (T = 10), a capacity of 100 (C =
100), and exactly 20 customers arriving each period (Nt = 20, t =
1, . . . , 10).5 In Fig. 2 the value function for period t = 6 is shown for
four possible states, c = 100, 50, 25, 10. In both graphs the x-axis represents the probability of sale z, and the y-axis represents (z )

E[min(D6 (z ), c )] + c1
x=1 V5 (c x )h6 (x|z ). We observe that the value
function is unimodal and depending on current capacity c maximized
5
Calculations were done in R (R Development Core Team, 2013) using a standard
laptop PC. Given the closed form solutions, computing time for the maximizing prices
was innitesimal. Computing time for solving the MDP, for all periods and all states,
was negligible.

at different values of z. For instance, in Example I with a current capacity of c = 50 (i.e. half-way through the time-horizon, the plane is
half sold-out), the optimal probability of sale is about .45 (z6 = .45)
with the optimal value V6 (50 ) = 30. Based on the right panel in Fig. 1,
we see this corresponds to optimal prices of ( p1,6 , p2,6 ) = (.55, .13 ).
Fig. 3 shows, for a given time-period (x-axis), the optimal probability of sale zt (top graphs) and the optimal dynamic prices
( p1,t , p2,t ) (bottom graphs) for a given capacity level c. We note, for
a given capacity c, that the optimal likelihood of sale zt is increasing over time, and consequently the optimal primary price p1,t is decreasing over time. Similarly, for a given period t, the optimal likelihood zt is increasing in the available capacity c, and hence the optimal
primary price p1,t is decreasing in c. A corollary of the last observation is that prices will increase if sales are high in earlier time periods.
We also observe that, for the four capacity scenarios illustrated, the
ancillary price never changes. This is anecdotally consistent with current practice of only dynamically changing the air fare and keeping

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

593

Fig. 3. Numerical illustration of optimal dynamic probability of sale zt (top) and primary prices p1,t (bottom) for different capacity levels c; Group 1 wtp is U[0,1], and Group 3 wtp
is U[0,a]. The bottom panels include the optimal dynamic ancillary price p2,t .

the baggage fee constant. Recall though that when the optimal probability of sale zt is less than the lower threshold z = 1/40 (Example I)
and 3/16 (Example II), the pricing strategy is not unique.

( ) represent the normal density and CDF approx ( ) and H

where h
imation, respectively, to Dt (z), with mean z and variance z2 . Integrating by parts, Eq. (14) can be written as,

3.4. Fixed number of arriving customers

E[min(Dt (z ), c )]
c 
c 
c 
z
z
z
z 
z 
+c 1
,

If the problem is simplied such that the number of arriving customers is xed Nt = N (and not stochastic), then the normal approximation to the binomial distribution can further improve the computational complexity. Specically, let Nt = N such that gt (x|pt ) and
ht (x|z) in (2) and (5), respectively, reduce to the binomial distribution. Let z and z2 represent the expected number and variance of
customers who would buy an item if one is available (at a given likelihood of a sale z); z = E[Dt (z )] = N z, z2 = Var (Dt (z )) = N z
(1 z ). Based on the normal approximation to the binomial distribution, the expected number of sales can be approximated by, for a
given z,

E[min(Dt (z ), c )]

(c|z )),
(x|z )dx + c (1 H
xh

where () is the standard normal CDF. Note for a given likelihood of
sale z, z and z2 are xed parameters and hence the evaluation of
E[min(Dt (z ), c )] is a simple calculation.
A further specialized case is when N = 1 (a common assumption
in the extant literature). In this case, the distributions given by (2)
and (5) reduce to the Bernoulli distribution, and E[min(Dt (z ), c )] =
z. Consequently the value functions drastically simplify to, for t =
1, . . . , T, c > 0,

Vt (c ) = sup

0z1

(14)

{ (z )z + Vt1 (c 1 )z + Vt1 (c )(1 z )},

the boundary conditions Vt (0 ) = 0, for all t, and V0 (c ) = 0, for all c.

594

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

4. The single-period model

To gain further insights regarding ancillary pricing and in particular the benet of dual price versus single price strategies, we next focus on the single period problem. Besides providing valuable insight
to the multi-period problem, the single-period problem is important
in its own right as there are many applications for which it applies,
e.g. sporting venues, theaters, etc. Since we are only considering a
single period problem we omit the subscript notation t. Furthermore,
in order to isolate the source of the stochasticity we limit the analysis
to a xed number of customer arrivals N. Consequently, for a given
likelihood of sale z, demand D(z) is binomially distributed with parameters N and z.
We start the discussion with the case when the seller has
no capacity constraint, C N, and sales are not censored, i.e.
E[min(D(z ), C )] = E[D(z )] = N z. The sellers expected total revenue is then given by,

V (C ) = sup { (z ) N z},

(15)

where (z) is given by (6).

The seller has to determine the optimal probability of sale z (and
corresponding set of prices p1 and p2 ) that maximize (15). However,
note that the optimal probability z (and prices p1 and p2 ) are independent of C and N. Below we show how these optimal prices can be
used to provide near-optimal solutions for the constrained capacity
case C < N.
For specic distributions Fi (), i = 1, 2, 3, the non-linear optimization problem given in (15) is numerically straightforward and can
be solved using a variety of optimization routines. On the other
hand, for arbitrary distributions, structural properties regarding the
optimal set of prices are not immediate. For example, to address
one of the motivations of this paper when p2 should be positive
and when it should be zero, requires additional assumptions regarding Fi (), i = 1, 2, 3. An immediate method to determine if p2 >
0, is to verify if pH arg max p { p(2 (1 F2 ( p)) + 3 (1 F3 ( p)))} >
arg max p { p(1 (1 F1 ( p)) + 2 (F2 ( pH ) F2 ( p)))} pL , and if so
simply set p1 = pL , and p2 = pH pL . Similarly, if pL pH , then it
follows that p2 = 0. However, note that in this case the p1 that maximizes over both customer groups is not necessarily pL .
A special case is when 2 = 0 and there are only Group 1 and
Group 3 type customers. In this case, one approach to derive structural properties regarding the optimal price for the ancillary service
is to impose assumptions on the ordering of the failure (or hazard)
rates of the willingness-to-pay. Let i (x ) = fi (x )/(1 Fi (x )) denote
the failure rate of wtp for Group i, for i = 1, 3. If both Group 1 and
Group 3s wtp have Increasing Failure Rates (IFR), and the failure rate
for Group 3 is larger than the failure rate for Group 1 then the optimal price for the ancillary service is zero (e.g. the airline should skip
the baggage fee). We formally summarize this result in the following
proposition.
Proposition 3. Let C N, 2 = 0, and 1 () and 3 () have increasing
failure rates, with 3 (x) 1 (x), for all x, then p2 = 0.
A trivial corollary to Proposition 3 is when the wtp for the two
customer groups are identical (i.e. F1 (x ) = F3 (x ), x), and the seller
rather intuitively should not charge separately for the ancillary service and set p2 = 0.
Although Proposition 3 provides insight into when the ancillary
fee should be zero, the seller must know or be able to estimate the
failure rates and determine if the one for Group 3 is larger than
the one for Group 1. This may be as challenging as simply determining the optimal set of prices. However, many of the common
distribution functions have the IFR property. Therefore, based on
historical demand data the seller could estimate F1 () and F3 () using
a parametric approach and validate the assumptions in Proposition 3.

Examples of distributions that have IFR include uniform, normal, exponential and gamma with shape parameter greater than 1 (Bagnoli
& Bergstrom, 2004). Note too that Proposition 3 would also hold
if rather than IFR, the assumption was made regarding Increasing
Generalized Failure Rate (IFGR). The added benet is that this enables
a more general setting and larger set of distributions (Banciu &
Mirchandani, 2013). Assumptions regarding IFR/IFGR are common
within the RM literature since they ensure basic revenue functions
of the form x (1 Fi (x )) are unimodal, and hence admit unique
revenue maximizing price.
We now return to the case when the seller has a capacity constraint and C < N, and show how the optimal prices for the unconstrained capacity case (C N) can provide a near-optimal solution.
For the constrained capacity case, the optimal expected revenue is
given by (12): V (C ) = supz { (z ) E[min(D(z ), C )]}. Recall that for
the unconstrained capacity problem the expected demand and variance is given by E[D(z )] = z N, and Var[D(z )] = z (1 z ) N, respectively. Therefore,
for a given C < N, as long as, for instance,

C E[D(z )] + 3 Var[D(z )], then the optimal prices for the unconstrained capacity setting are near-optimal for the capacity constrained setting. That is, for a given N, the seller can set a threshold
and determine the upper-limit d such that Pr{D(z ) >d} . For
instance, a rough 99% condence interval is E[D(z )] 3 Var[D(z )].

If C d = E[D(z )] + 3 Var[D(z )] then the optimal prices for the unconstrained capacity setting are approximately optimal even considering the capacity constraint.
4.1. Numerical analysis - uniform distribution
To further evaluate the potential benet of charging for ancillary
services, we next analyze the case when there are only Group 1 and
Group 3 customers (i.e. 2 = 0). We assume customers willingnessto-pay is uniformly distributed and scale the problem such that wtp
U[0, 1] for Group 1, and wtp U[0, a] for Group 3, a > 0. Note
that unlike Section 3.3 we include cases a < 1. The main goal is to
compare the gain in expected total revenue between the optimal
dual pricing policy p = ( p1 , p2 ) and the optimal single price policy
p . The optimal prices are found using a search maximizing V (C ) =
supz { (z ) E[min(D(z ), C )]}, without and with the constraint that
p2 = 0. For ease of exposition we will slightly modify the notation and
denote VD (C) and VS (C) as the corresponding maximum expected total revenue for the Dual pricing and Single price policy, respectively.
To measure the improvement of a dual-pricing strategy, we denote

V(C) as the incremental gain,

V (C ) = (VD (C ) VS (C ))/VS (C ). We
set N = 20 and consider four cases regarding the capacity constraint:
(I) C = 5, (II) C = 10, (III) C = 15, and (IV) C = 20.
Fig. 4 illustrates the gain in expected revenue from the optimal dual-price policy versus the optimal single-price policy. In each
graph, the y-axis represents
V(C), and the x-axis represents the
upper-bound a of Group 3 customers willingness-to-pay. The three
curves correspond to different values of 1 : .2 (solid black), .5 (dashed
red), and .8 (dotted blue). Recall we are assuming 2 = 0 and hence
3 = 1 1 . There are a number of interesting observations to note.
The rst and foremost interesting feature to note is that there can
be a substantial gain in charging for the ancillary service. We observe
potential revenue gain of up to 40% for certain cases of a and 1 . On
the other hand, we also observe that there are situations when there
is no gain in imposing ancillary fees. More specically, when a 1, i.e.
when Group 3 customers wtp is stochastically dominated by Group
1 customers wtp, then it is always optimal to set p2 = 0. Note that although we are considering a general capacity constrained setting, the
feature for when a 1 is consistent with the result in Proposition 3
regarding the unconstrained capacity setting.
The second observation to note is that depending on the value
of 1 , the highest incremental gain of using a dual-price strategy

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

595

Fig. 4. Numerical illustration of the gain in expected total revenue versus upper-bound a of Group 3 wtp; Group 1 wtp U[0, 1], Group 3 wtp U[0, a].

changes and is not monotone in a. In particular we see that when

1 is low (high chance of a Group 3 customer) then the highest incremental gain is for relatively low values of a (i.e. between 1 and 2),
and when 1 is high (low chance of a Group 3 customer) then the
highest incremental gain is for the high values of a (i.e. greater than
2). This is a bit counter-intuitive as one would have expected that
as the wtp for Group 3 increases the dual-price strategy should be
able to capitalize even more. However, this is explained by observing that when the wtp of Group 3 is large then the optimal singleprice strategy is to exclude sales to Group 1 and only focus on Group
3 customers.
A third observation to note is that the graphs for cases (III) and (IV)
are very similar. Although they may appear identical there are small
differences between them. The closeness between the bottom two
graphs in Fig. 4 illustrates the discussion regarding the near-optimal
pricing solutions based on the optimal prices for the unconstrained
capacity case: even if the seller is technically constrained, the constraint does not limit the expected revenue potential. Next we analyze the gain in the expected revenue as a function of 1 .

Fig. 5 illustrates the gain in expected revenue as a function of the

proportion of Group 1 customers. In each graph, the y-axis again represents
V(C), but now the x-axis represents 1 . The four curves correspond to different values of a: 1 (solid black), 1.5 (dashed red), 2
(dotted blue), and 4 (dash-dotted purple). There are two striking observations to highlight: (1) the gain is not always symmetric in 1 ,
and (2) the gain is higher for large values of 1 .
Regarding the rst observation, although one may not have expected the gain to be symmetric in 1 , it might seem counterintuitive that for low values of a the gain is symmetric, and for high
values of a the gain is not symmetric. One might have expected that
for large values of a there is greater exibility for the dual-price strategy to adapt to the particular value of 1 , and thus for a symmetric change around 1 = .5, to be able to derive the same expected
revenue. However, even if this is the case, the problem is that the
single-price strategy cannot adapt as easily. For a large value of a,
the single-price strategy can compete quite well with the dual-price
strategy when 1 is low (high chance of a Group 3 customer), and
hence the gain is small, but if 1 is large (small chance of Group 3

596

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

Fig. 5. Numerical illustration of the gain in expected total revenue versus proportion 1 of Group 1 customers; Group 1 wtp U[0, 1], Group 3 wtp U[0, a].

customers) then the single-price strategy cannot capture the higher

revenues from Group 3 and thus the gain is large. This also explains
the second feature why the gain is higher for large values of 1 . Intuitively, one might at rst consider that if there is a high chance of
having Group 3 customers (low 1 ) then there should be a high gain
of introducing the ancillary fee p2 . As noted above, the issue is that
if there is a high chance of Group 3 customers then the single-price
strategy performs almost as well since it prices for Group 3 by sacricing sales from Group 1. This has a very relevant managerial insight.
Namely, the main incremental gain from ancillary fees is not on, for
instance, routes where customers generally check baggage, but instead on routes where customers rarely check baggage. Furthermore,
the result is anecdotally consistent with what we observe in many
RM settings. For instance, premium down-town business hotels usually charge extra for wi service, while lower rated business hotels
located outside the city center usually provide complementary wi.

there is no capacity constraint. The reason we focus on this is twofold: (1) we can easily derive closed form solutions, and (2) it can
provide a good approximation to the constrained capacity setting.
Following the previous discussion, we assume Group 1 and Group 3
customers willingness-to-pay follow U[0, 1] and U[0, a], respectively
(and that there are no Group 2 customers, 2 = 0). Consequently, the
sellers dual pricing problem becomes,

VD (C ) = max N (1 (1 p1 )+ p1
p1 ,p2

+ 3 1

In Figs. 4 and 5 we observed that for a > 1 the optimal dual-price

strategy always performed better than the optimal single-price strategy. We conclude this section with a more detailed analysis when

(16)

where (1 p1 )+ = max{1 p1 , 0}. From Proposition 3, and our observation in Fig. 4 regarding C = 20 and N = 20, we know that for a
1, it is never optimal to set p2 > 0. Therefore, for the ensuing analysis
we only consider cases where a > 1. The two rst-order conditions of
(16) with respect to p1 are,6


4.2. Unconstrained capacity - uniform distribution

p1 + p2
( p1 + p2 ))| p1 + p2 a ,
s

1 21 p1 2(1 1 )( p1 + p2 )/a

for p1 1,

(1 1 ) 2(1 1 )( p1 + p2 )/a

for p1 > 1,

6
Note that the value functions for both the dual-price and single-price case are concave and it is sucient to only consider rst-order conditions.

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

where 1 1 = 3 . The corresponding rst-order condition with respect to p2 is given by (1 1 ) 2(1 1 )( p1 + p2 )/a = 0. Solving for p1 results in p1
= 1/2, and p2
= a/2, for the two cases re1
1
= (a 1 )/2, and
spectively. Which we use to solve for p2 , and get p1
2
=
0,
for
the
two
cases
respectively.
By
comparing
all the possibilip2
2
ties with respect to p1 and a (1 a < 2, and a > 2) it can be shown that
the optimal set of prices is p = (1/2, (a 1 )/2 ), and consequently
the maximum expected total revenue is given by, for a 1,

VD (C ) =

N (1 + 3 a )
.
4

(17)

The extension to the general case with Group 1 customers wtp

U[0, a1 ], Group 3 customers wtp U[0, a2 ], and a1 < a2 is immediate and results in p1 = a1 /2, p2 = (a2 a1 )/2, and VD (C ) = N (1 a1 +
2 a2 )/4. The resulting maximum expected total revenue (17) can be
evaluated against the expected total revenue from the single-price
policy that only optimizes over p1 , for p1 a,

VS (C ) = max N
p

 

p
1 (1 p)+ p + 3 1 ( p) | p a .
a

(18)

There are again two rst-order conditions to consider,

1 (1 2p) + (1 1 )(1 2p/a ) for p 1,

(1 1 )(1 2p/a )
for p > 1,

which yields the optimal single price,

p =

a/(2 + 21 a 21 ) for a (2 1 )/(1 1 ),

for a > (2 1 )/(1 1 ).

a/2

Substituting

VS (C ) =

p

into (18) results in,

Ns/(4 + 41 a 41 ) for a (2 1 )/(1 1 ),

N (1 1 )a/4

for a > (2 1 )/(1 1 ).

It should be rather intuitive that VD (C) > VS (C).

Next we consider a special dual-pricing policy where p2 is a xed
proportion of p1 , i.e. p2 p1 . In other words, we assume the seller
has decided a priori to set p2 as a given factor of p1 . We denote the
maximum expected total revenue by V (C) and the optimal price for
the primary product by p1 ( ). We do not restrict to be less than or
greater than 1. A motivation for 1 may be the baggage fee that
is roughly 5%10% of a US domestic air fare (one way), while a motivation for 1 may be insurance for rental cars. Given these structural assumptions, the expected total revenue becomes, for p1 1,
and p1 (1 + ) a,

V (C ) = max N1 (1 p1 ) p1 + N (1 1 )
p1

p1 + p1
1
a


( p1 + p1 ) .

(19)

From Eq. (19) the price of the primary product that maximizes (19) is
given by,

p1 ( ) =

1 + (1 + )(1 1 )/1
.
2(1 + (1 + )2 (1 1 )/1 a )

We see that the optimal price p1 ( ) depends on , a, and the odds
against a Group 1 customer (1 1 )/1 . Although it is clear that
p1 ( ) is increasing in a, note that p1 ( ) is not monotone in either
or (1 1 )/1 . That is, for a given (1 1 )/1 and a, p1 ( ) may be
increasing or decreasing in . Likewise with respect to (1 1 )/1 ,
for a given a and .
We conclude this section by evaluating the impact from using the
unconstrained capacity optimal pricing strategy for the constrained capacity problem. In Section 4.1, we derived the optimal prices p1 = 1/2
and p2 = (a 1 )/2 when valuations are uniformly distributed and
a > 1. Given these prices, the probability of a sale z = 1/2, and hence
total demand D(z) Bin(N, 1/2). Consequently, an approximate 99%

597

condence interval on the total demand is N/2 (3 N )/2. For N >

9, this implies the lower limit is non-negative and the upper-limit
is less than N. Therefore, a near-optimal pricing heuristic for the
constrained capacity problem when (N + 3 N )/2 C N, is to simply set prices according to the unconstrained capacity problem. This
highlights why the bottom two graphs in Figs. 4 and 5 are so similar. With N = 20 and 2.5 standard deviations, the upper bound of
d 16. Since case (III) in the two gures is for C = 15 it should not
be surprising that the results are close to the unconstrained capacity
case (IV).
5. Conclusions
Ancillary services have become an important revenue source in
many industries using RM. A prime example is the growing trend in
the airline industry to offer unbundled or add-on services such as
preferred seating, early boarding, drinks, checked baggage service;
see news articles McCartney (2013); White (2013). Consequently, the
next generation of RM systems need to take into consideration pricing and demand of offered bundles and add-on services, and not be
based solely on revenue optimization from the sale of primary products or services. Motivated by the growing prevalence in the airline
industry to charge for checked baggage service, this paper has considered a seller with a xed capacity or inventory of primary items
that also makes an ancillary service available. The issue facing the
seller, and main research question of the paper, is how to dynamically
set prices of the primary items and the ancillary services in order to
maximize expected prot.
The model framework was based on three groups of heterogeneous customers: those that only want the primary product, those
that could do with only the primary product but given a right price
will also purchase the ancillary service, and those that only want
the primary product provided they also receive the ancillary service.
The sellers problem is to dynamically set prices for both the primary product and ancillary service. Our formulation does not require
specic distributional assumptions on the arrival process or the underlying customers willingness-to-pay. In this general capacity constrained setting with numerous customer combinations and allowing
for lost sales, a rst challenge is to formulate a useful expression for
the expected total revenue. The problem was then transformed into
a one dimensional optimization over the probability of a sale. Given
specic distributional assumptions, this formulation allows for the
exploration of structural properties. For example, for uniform distributions, we demonstrated the existence of unique optimal prices for
the primary product and ancillary service. Indeed, for this example,
the one dimensional formulation led to a close to closed form expression for the total expected revenue as a function of the probability of
a sale.
In our development we have assumed that both the distribution
of the willingness-to-pay Fi () and the proportion of each group i are
time independent, i = 1, 2, 3. This has purely been to keep the notational complexity to a minimum. Both stochastic components can be
made functions of time with minor modication. Indeed the only real
change would be to allow (z) to be t (z) and different for each period t = 1, . . . , T . In the MDP recursion the value of t (z) (and corresponding maximizing prices ( pz1,t , pz2,t )) would be called on in period
t, rather than the current value (z) (and corresponding maximizing
prices ( pz1 , pz2 )).
Assuming uniform distributions is common in the extant literature and mainly done for reasons of mathematical tractability.
Although this entails that the insights obtained are limited by the
particular form of the uniform distributions, there are nevertheless
interesting and valuable conclusions to draw. Consequently, we have
included a number of analyses where the underlying distributions
are assumed to be uniform. We demonstrated how our formulation
can lead to close to closed form results in such cases. Based on a

598

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

numerical analyses, we also investigated when pricing separately

for the ancillary service is benecial. The numerical analysis revealed some interesting and seemingly counterintuitive results. For
instance, we found that the incremental gain from an optimal dualprice strategy versus an optimal single-price strategy is higher when
the proportion of customers demanding the ancillary service is lower.
This has important RM consequences. For instance, consider an airline with two ights: one primarily used by travelers who do not
check baggage, and one primarily used by travelers who do check
baggage. The incremental revenue gain in charging for the checked
baggage is higher on the rst ight. So, when most travelers check
baggage, offering a single-price is almost equally good to charging
separately for bags and the airline does not need to appear to nickeland-dime their customers.
There are many areas to further enrich the proposed modeling
framework and we hope the paper will motivate more research related to revenue management and ancillary services. Two immediate extensions include considering empirical distributions and the effect of competition. Although the presented model framework is very
general and in particular imposes no distributional assumptions, the
results were based on strictly increasing distribution functions for the
willingness-to-pay. Extending the framework to a discrete probability space, including empirical distributions, is straight-forward and
mainly entails keeping track of the point-mass at the critical thresholds. In fact, the problem becomes easier to numerically solve. The
other extension regarding competition effects is a bit more challenging and would require further assumptions on the customer behavior and choice model. We leave this and other extensions for future
consideration.
Acknowledgment
We wish to thank the three anonymous reviewers for their valuable comments and feedback which greatly helped improve the paper. We would also like to acknowledge comments from seminar
participants at Smith School of Business, Queens University. The research project was partially funded by Professor degaards NSERC
Discovery Grant.
Appendix. Proofs of results

Proof of Proposition 1. - For the moment, x the primary and ancillary prices, p1, t and p2, t , respectively. Suppose that there are x < c
customers (or sales) during period t. Then the expected number of
ancillary sales during period t is the expected value of a binomial
random variable with parameters x and (2 (1 F2 ( p1,t + p2,t )) +
3 (1 F3 ( p1,t + p2,t )))/zt (pt ), since this latter is the proportion of
time a ancillary sale is made.
Suppose that there are x c customers during period t who
would buy a ticket were one available. Then the actual number
of ancillary sales during period t is a random choice of c from x
where the probability of success is (2 (1 F2 ( p1,t + p2,t )) + 3 (1
F3 ( p1,t + p2,t )))/zt (pt ). The expected value of this hypergeometric random variable is equal to c (2 (1 F2 ( p1,t + p2,t )) + 3 (1
F3 ( p1,t + p2,t )))/zt (pt ). Thus, noting that gt (x|pt ) may be replaced by
ht (x|zt (pt )),

(zt (pt )) p1,t + p2,t

2 (1 F2 ( p1,t + p2,t )) + 3 (1 F3 ( p1,t + p2,t ))

,
zt (pt )
for all values of pt = ( p1,t , p2,t ):

(z ) E[min(Dt (z ), c )] +

Vt (c ) sup
z

c1


x p1,t + p2,t

c1


Vt1 (c x )ht (x|z ) .

x=1

On the other hand, for a given value of z, we can nd an exhaustive set of values for p1, t and p2, t when zt (pt ) = z. Elements of this
set are simply particular feasible values and thus, since Vt (c) is the
supremum over all possible values of p1, t and p2, t :

Vt (c ) sup
z

(z )E[min(Dt (z ), c )]+

c1


Vt1 (c x )ht (x|z ) .

x=1

Proof of Proposition 2. - The rst step is to show that q2 (z,

p1 ) is well-dened by showing that the argument (1 z (1 +
2 )F ( p1 ))/3 , of F31 ( ), is between 0 and 1, and then that q2 (z, p1 )
0. Note that,

1 z (1 + 2 )F ( p1 )
1,
1 (1 + 2 )
if and only if,

F ( p1 )

1 + 2 z
,
1 + 2

which has been assumed by p1 lz . Now note that, by denition of

uz , only values of p1 uz need be considered and that 1 z (1 +
2 )F ( p1 ) 3 F3 ( p1 ) is decreasing in p1 and equals zero at p1 = uz .
Thus,

1 z (1 + 2 )F ( p1 )
0.
1 (1 + 2 )
Also 1 z (1 + 2 )F ( p1 ) 3 F ( p1 ), from which it follows that
q2 (z, p1 ) 0.
+ z
For p1 F 1 (max{ 1 +2 , 0} ), the unique value of p2 that satis1
2
es (4) is given by q2 (z, p1 ). Therefore replacing p2 with q2 (z, p1 ) in
(6) and maximizing over p1 provides the result. 
Proof of Proposition 3. - Since there is no capacity constraint we
can determine the price that maximizes the expected revenue per
customer. Let pI be the price that maximizes p(1 F1 ( p)), i.e. pI =
(1 F1 ( pI ))/ f1 ( pI ) = 1/1 ( pI ). Let pII be the price that maximizes
p(1 F3 ( p)), i.e. pII = (1 F3 ( pII ))/ f3 ( pII ) = 1/3 ( pII ). That is pII is
the total price (pII = p1 + p2 ) that maximizes the expected revenue
for a Group 3 customer. Since 3 (x) 1 (x), for all x, it follows that
1/1 (p) 1/3 (p) p, for all p pII . Since 1/1 (x) is decreasing it
follows there will be a pI pII where 1/1 ( pI ) = pI . Therefore, since
p1 is the same for both Group 1 and Group 3 customers, it follows that
p2 = max{ pII pI , 0} = 0. 

2 (1 F2 ( p1,t + p2,t )) + 3 (1 F3 ( p1,t + p2,t )) 

ht (x|zt (pt ))
zt (pt )
x=0

2 (1 F2 ( p1,t + p2,t )) + 3 (1 F3 ( p1,t + p2,t )) 
+
c p1,t + p2,t
ht (x|zt (pt ))
zt (pt )
x>c

2 (1 F2 ( p1,t + p2,t )) + 3 (1 F3 ( p1,t + p2,t )) 
= p1,t + p2,t
E[min(Dt (zt (pt )), c )].
zt (pt )

t (pt , c ) =

From (4), noting that E[min(Dt (zt (pt )), c )] and ht (x|zt (pt )) depends
only on the z = zt (pt ) derived from pt and that,

F. degaard, J.G. Wilson / European Journal of Operational Research 251 (2016) 586599

References
Allon, G., Bassamboo, A. Lariviere, M. (2011). Would the social planner let bags
y free?Working paper, available here: http://www.kellogg.northwestern.edu/
faculty/bassamboo/htm/Papers/Allon_Bassamboo_Lariviere_Would_The_Social_
Planner_Let_Bags_Fly_Free.pdf, accessed 2015-08-27.
Amaruchkul, K., Cooper, W., & Gupta, D. (2007). Single-leg air-cargo revenue management. Transportation Science, 41, 457469.
AMR Corp (2011). AMR Corporation form 10-k for the scal year ended december
31, 2011. Available here: http://phx.corporate-ir.net/phoenix.zhtml?c=117098&p=
irol-sec&control_selectgroup=Annual%20Fillings, accessed 2015-08-27.
Bagnoli, M., & Bergstrom, T. (2004). Log-concave probability and its applications. Economic Theory, 26, 445469.
Banciu, M., Gal-Or, E., & Mirchandani, P. (2010). Bundling strategies when products are
vertically differentiated and capacities are limited. Management Science, 56, 2207
2223.
Banciu, M., & Mirchandani, P. (2013). New results concerning probability distributions
with increasing generalized failure rates. Operations Research, 61, 925931.
Bitran, G., & Caldentey, R. (2003). An overview of pricing models for revenue management. Manufacturing & Service Operations Management, 5, 203229.
Boyd, E., & Bilegan, I. (2003). Revenue management and E-Commerce. Management
Science, 49, 13631386.
Brueckner, J., Lee, D., Picard, P., & Singer, E. (2015). Product unbundling in the travel
industry: The economics of airline bag fees. Journal of Economics & Management
Strategy, 24(3), 457484.
Chen, L., & Homem-de-Mello, T. (2010). Mathematical programming models for revenue management under customer choice. European Journal of Operational Research, 203, 294305.
Cooper, W., & Li, L. (2012). On the use of buy up as a model of customer choice in
revenue management. Production and Operations Management, 21, 833850.
Delta Air (2011). Delta Air Lines, Inc. form 10-k for the scal year ended december
31, 2011. Available here: http://www.delta.com/content/dam/delta-www/pdfs/
about-nancial/DeltaAirLines_10K_2011.pdf, accessed 2015-08-27.
Essegaier, S., Gupta, S., & Zhang, Z. (2002). Pricing access services. Marketing Science,
21, 139159.
Gallego, G., & van Ryzin, G. (1997). A multiproduct dynamic pricing problem
and its applications to network yield management. Operations Research, 45(1), 24
41.
Hamilton, R., Srivastava, J., & Abraham, A. (2010). When should you nickel-and-dime
your customers? MIT Sloan Management Review, 52(1), 5967.

599

Han, D., Tang, L., & Huang, H. (2010). A markov model for single-leg air cargo revenue
management under a bid-price policy. European Journal of Operational Research,
200, 800811.
Kunnumkal, S., & Topaloglu, H. (2010). A new dynamic programming decomposition
method for the network revenue management problem with customer choice behavior. Production and Operations Management, 19, 575590.
Maglaras, C., & Meissner, J. (2006). Dynamic pricing strategies for multiproduct revenue management problems. Manufacturing & Service Operations Management,
8(2), 136148.
McAfee, R., McMillan, J., & Whinston, M. (1989). Multiproduct monopoly, commodity
bundling, and correlation of values. The Quarterly Journal of Economics, 104(2), 371
383.
McCartney, S. (2013). A new bundle jungle for travelers. The Wall Street Journal. Jan 30,
2013
McGill, J., & van Ryzin, G. (1999). Revenue management: Research overview and
prospects. Transportation Science, 33, 233254.
Meissner, J., & Strauss, A. (2012). Network revenue management with inventorysensitive bid prices and customer choice. European Journal of Operational Research,
216, 459468.
Nicolae, M., Arkan, M., Deshpande, V. Ferguson, M. (2013). Do bags y free? an empirical analysis of the operational implications of airline baggage fees. Working paper,
Moore School of Business - University of South Carolina, available from fourth author on request, received 2013-10-17.
Oi, W. (1971). A disneyland dilemma: two-part tariffs for a mickey mouse monopoly.
The Quarterly Journal of Economics, 85(1), 7796.
R Development Core Team (2013). R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0,
http://www.R-project.org
Shulman, J. D., & Geng, X. (2013). Add-on pricing by asymmetric rms. Management
Science, 59(4), 899917.
Talluri, K., & van Ryzin, G. (2004a). Revenue management under a general discrete
choice model of consumer behavior. Management Science, 50(1), 1533.
Talluri, K., & van Ryzin, G. (2004b). The Theory and Practice of Revenue Management.
Norwell, Massachusetts: Kluwer Academic Publishers.
White, M. (2013). Airlines cash in on every inch, even the jammed bins overhead. The
New York Times,.Oct 11, 2013
Wong, W., Zhang, A., Hui, Y., & Leung, L. (2009). Optimal baggage-limit policy: Airline
passenger and cargo allocation. Transportation Science, 43, 355369.
Zhang, D., & Cooper, W. (2009). Pricing substitutable ights in airline revenue management. European Journal of Operational Research, 197, 848861.