Professional Documents
Culture Documents
net/publication/247478566
CITATIONS READS
94 2,645
4 authors, including:
22 PUBLICATIONS 348 CITATIONS
Karl Isler Consulting GmbH
23 PUBLICATIONS 467 CITATIONS
SEE PROFILE
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Karl Isler on 06 October 2015.
Thomas Fiig
is Chief Scientist in the Revenue Management Development department at Scandinavian Airlines System
(SAS). He is responsible for developing methods and strategy for revenue management systems at SAS,
including the overall design and methodologies of the O&D forecasting and optimization systems. His recent
work has focused on methodologies for O&D optimization in semi-restricted fare structures and estimating
price elasticities. Dr Fiig holds a PhD in Mathematics and theoretical Physics and a BA in Finance from the
University of Copenhagen.
Karl Isler
is Head of Operations Research and Strategy in the Revenue Management, Pricing and Distribution
department of Swiss International Airlines. He holds a PhD in Theoretical Physics from ETH Zurich. He
developed the concepts for the integrated O&D pricing and inventory control strategy used by Swiss.
Craig Hopperstad
is currently president of Hopperstad Consulting. Previously, as Project Director in the Boeing Commercial
Airplane Group, he was a principal in the development of passenger preference, fleet planning, scheduling
and revenue management models. He is the author of numerous papers and presentations, many of which
deal with the application of the Passenger Origin/Destination Simulator (PODS), which he developed at
Boeing and for which Hopperstad Consulting now holds a license.
Peter Belobaba
is Principal Research Scientist at the Massachusetts Institute of Technology (MIT), where he teaches
graduate courses on The Airline Industry and Airline Management. He is Program Manager of MIT’s Global
Airline Industry Program and Director of the MIT PODS Revenue Management Research Consortium. Dr
Belobaba holds a Master of Science and a PhD in Flight Transportation Systems from MIT. He has worked as
a consultant on revenue management systems at over 40 airlines and other companies worldwide.
Correspondence: Peter Belobaba, MIT International Center for Air Transportation, 77 Mass. Ave., Room
33–215, Cambridge, MA 02139 USA
E-mail: belobaba@mit.edu
ABSTRACT This paper develops a theory for optimizing revenue through seat inventory control that can
be applied in a variety of airline fare structures, including those with less restricted and fully undifferentiated
fare products that have become more common in the recent past. We describe an approach to transform the
fares and the demand of a general discrete choice model to an equivalent independent demand model. The
transformation and resulting fare adjustment approach is valid for both static and dynamic optimization and
extends to network revenue management applications. This transformation allows the continued use of the
optimization algorithms and seat inventory control mechanisms of traditional revenue management systems,
developed more than two decades ago under the assumption of independent demands for fare classes.
Journal of Revenue and Pricing Management advance online publication, 19 June 2009; doi:10.1057/rpm.2009.18
Keywords: seat inventory control; airline fare structures; network revenue maximization; O-D control;
marginal revenue transformation; DAVN-MR, PODS
& 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
www.palgrave-journals.com/rpm/
Fiig et al
2 r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
Optimization of Mixed Fare Structures
nistic demand (no stochasticity), single airline version became known as ‘EMSRb’, and is also
leg (no network) and a single market (one described in detail in Belobaba and Weath-
specific price-demand relation). The theory is erford (1996). Dynamic programming models
by no means limited by any of these simplifica- have also been applied to the single-leg seat
tions, which are only introduced to present the inventory control problem, by Lee and Hiersch
theory more clearly. We then discuss specific (1993) and Lautenbacher and Stidham (1999),
applications of the optimization theory to among others.
different fare structures. In the section ‘Exten- Network RM systems (or origin-destination
sions to stochastic models’ we show that the controls) use network optimization mathe-
theory extends beyond the simplifications made matics to determine fare class seat availability
above to stochastic models, including dynamic based on paths (that is passenger itineraries),
programming. Simulated revenue results from not by individual leg, making them particularly
the Passenger Origin Destination Simulator valuable to airlines that operate networks of
(PODS) are presented in the section ‘Simula- connecting flights. Smith and Penn (1988)
tion results using PODS’. implemented one of the first O-D control
methods, virtual bucketing, at American
Airlines. Many network RM systems evolved
LITERATURE REVIEW from this initial concept, leading to the
Most revenue management techniques were development of Displacement Adjusted Virtual
developed under a critical assumption that Nesting (DAVN). DAVN adjusts the value of
demand for a given fare class is independent each path’s O-D fare for the estimated revenue
of the demand for other fare classes. The displacement of a passenger on the connecting
effective use of these revenue management leg (Saranathan et al, 1999). Williamson (1992)
(RM) systems by airlines worldwide has been and Vinod (1995) offer more detailed descrip-
estimated to generate on the order of 4–6 per tions of O-D control and specifically the
cent in incremental revenues (Smith et al, DAVN approach.
1992). Barnhart et al (2003) describe the An alternative mechanism for O-D control
evolution of revenue management systems since that makes use of similar network optimization
the early 1980s. McGill and van Ryzin (1999) logic is ‘Bid-Price Control’. Developed by
provide a comprehensive survey of revenue Simpson (1989) and elaborated by Williamson
management literature, focusing on the evolu- (1992), this method requires that the airline
tion of forecasting and optimization models. store a bid-price value (an estimate of the
The optimization of airline seat inventory to marginal network revenue value of each
maximize revenue given multiple fare products incremental seat) for each leg in the network.
on a single flight leg can be traced back to For each itinerary request, seat availability is
Littlewood (1972), who solved the fare class determined by comparing the relevant O-D
mix problem for two nested fare classes. fare to the sum of the bid prices on the legs that
Belobaba (1987, 1989) extended the nested the path traverses. Algorithms for calculating
seat allocation problem to multiple fare bid prices in an airline network include
classes with the development of the Expected mathematical programming models by Talluri
Marginal Seat Revenue (EMSR) heuristic. and Van Ryzin (1998), heuristic approaches by
Curry (1990), Wollmer (1992) and Brumelle Belobaba (1998), and a probabilistic conver-
and McGill (1993) described optimal solutions gence algorithm by Bratu (1998). Dynamic
for multiple nested fare classes. Belobaba (1992) programming can also be used to determine
then modified the EMSR heuristic to better optimal network bid prices, in theory at
approximate the optimal solution to the multi- least, as formulated by Gallego and Van
ple nested fare class problem. The modified Ryzin (1997). In practice, the ‘curse of
r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19 3
Fiig et al
dimensionality’ makes the size of the network price-oriented (or ‘priceable’) demand that
DP problem infeasible to solve for realistic will always purchase the lowest available fare
airline networks. Instead, some network RM and product-oriented (or ‘yieldable’) demand
systems apply DP methods to single flight legs that is willing to buy a higher fare due to
to determine a vector of bid prices, in specific attributes or lack of restrictions. They
conjunction with a mathematical programming suggested that identification of price- vs
approach for network optimization. product-oriented demands in an RM database
The implementation of network RM sys- could enable the airline to generate separate
tems by airlines has been estimated to generate forecasts of each type of demand, and to
an additional 1–2 per cent in total revenues, combine these forecasts into a single ‘hybrid’
above the 4–6 per cent realized from the forecast. While traditional time series forecast-
traditional leg-based RM systems (Belobaba, ing models can be applied to the product-
1998). Yet, virtually all of the optimization oriented demand data, a forecasting approach
models employed by these systems were based that includes estimates of willingness to pay (or
on the assumption of independent demands, sell-up potential) is required to forecast price-
first by flight leg and fare class, and then by oriented demand for each fare class. Belobaba
passenger itinerary (path) and fare class. None and Hopperstad (2004) describe such a model,
of the traditional optimization models des- called ‘Q-forecasting’, which can provide the
cribed in the literature explicitly addresses the estimates of price-oriented demand required
complications caused by the move to less for hybrid forecasts.
restricted fare structures since 2000. As noted Talluri and van Ryzin (2004) formulated a
by Tretheway (2004), the introduction of these DP model for the case when consumer beha-
simplified fare structures ‘has undermined the viour is described by a general discrete choice
price discrimination ability of the full service model, and showed that only strategies on the
network carrier, and is the most important efficient frontier are relevant in the optimiza-
pricing development in the industry in the past tion. In this paper, we show that any
25 years.’ optimization algorithm for an airline network
Without adequate demand segmentation can be extended to a completely general fare
restrictions, simplified fare structures allow structure, as developed and first presented by
passengers that previously would buy higher Fiig et al (2005) and Isler et al (2005). The
fares to purchase lower fare classes. The RM approach, also referred to as ‘fare adjustment’,
system’s historical booking database then makes use of a marginal revenue transformation
records fewer bookings in the higher fare that adjusts the fare and demand inputs to
classes which, in turn, leads to lower forecasts traditional RM optimizers, effectively trans-
of future demand for higher fare types. The forming the choice model representing a
seat allocation optimizer then protects fewer general fare structure into independent demand
seats for higher fare classes and makes more of a fully differentiated fare structure.
seats available to lower booking classes, which
encourages even more high-fare demand to
buy down. A mathematical model that des- THE OPTIMIZATION PROBLEM
cribes this ‘spiral down’ effect was developed by In this section we describe the general
Cooper et al (2006). optimization problem and define the notation
Several works describing methods for rever- to be used throughout the paper. We consider
sing this spiral down effect have focused on the simplest possible example to develop the
modifications to the forecasting models used in optimization theory, that is, a single leg with
traditional RM systems. Boyd and Kallesen capacity cap. We begin with the solution of the
(2004) proposed a segmentation between optimization problem when fares are fully
4 r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
Optimization of Mixed Fare Structures
1400 100000
1200
80000
CAP
1000
800 60000
Fare
TR
600 40000
400
20000
200
0 0
0 50 100 150 0 50 100 150
Q Q
Figure 1: (a) Incremental revenue as function of demand (quantity). Total revenue shown as shaded area below curve.
(b) Total revenue versus total demand (quantity).
differentiated (that is independent or ‘product- Figure 1(b) illustrates the quantities TRk and
oriented’ demand). The airline has published a Qk, denoting the total revenue and the total
set of fares fi, in decreasing fare order: fi>fi þ 1, quantity sold given k is the lowest fare level
i ¼ 1,y, n1. We assume that demand is available, as
deterministic, with demand di for each fare
product. X
k X
k
Qk ¼ dj and TRk ¼ fj dj :
In a fully differentiated fare structure, j¼1 j¼1
traditional RM models assume that the demand
for each fare product is independent of whether The total revenue increases monotonically
other products are available. As shown in with quantity, calculated as the area below the
Figure 1(a), the revenue maximizing solution curve in Figure 1(a).
is to allocate seats for all demands in decreasing Now assume that the restrictions on the fare
fare order until capacity is reached. The seats products are removed, making them fully
available to the lowest fare product must be undifferentiated (that is, ‘priceable’ demand).
limited by capacity. The step-wise decreasing The only distinction between the fare products
curve is the incremental revenue as function of is their price levels and therefore customers will
quantity (Q). The capacity constraint is shown always buy the lowest fare available. Even
as the vertical line at cap ¼ 100. Note that the customers willing to buy the higher fares will
data behind Figures 1 and 2 are are summarized now buy the lowest available fare. The total
in Table 1. revenue and the total quantity sold given k is the
r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19 5
Fiig et al
a 1400 b 100000
1200
1000 80000
CAP
800 diff.
CAP
60000
600
TR
Fare
P(Q)
400
40000
200 MR(Q) un-diff.
0
20000
0 50 100 150
-200 Max
-400 0
-600 0 50 100 150
Q Q
Figure 2: (a) The demand curve P(Q) and marginal revenue MR(Q) for the undifferentiated case. (b) TR(Q) for both
differentiated and undifferentiated case.
lowest fare available can now be expressed as The marginal revenue is illustrated by the
lower curve, MR(Q), on Figure 2(a). Note that
X
k
Qk ¼ dj and TRk ¼ fk Qk : the area below the curve corresponds to the total
j¼1 revenue. As more seats are made available to the
lower fare, the marginal revenue becomes
The total revenue is shown for both negative, meaning the total revenue decreases.
the differentiated and undifferentiated case in Therefore the fare classes below $600 in this
Figure 2. With all passengers buying the lowest example should always remain closed.
available fare, the total revenue for the We can therefore summarize the solution
undifferentiated case falls below that of the principle of the revenue optimization problem
differentiated fare structure. as follows: Order the fares in decreasing marginal
The incremental demand accommodated by revenue and open fares until capacity is reached or the
opening up fare class k is dk. However the marginal revenue becomes negative. This general
incremental revenue is not dk fk but a reduced principle is valid for both differentiated and
amount, which we can calculate by accounting undifferentiated fare structures.
for the revenue lost due to buy down. The loss
from buy-down is given by Qk1( fk1fk),
since the aggregated demand that previously GENERAL FORMULATION
bought class k-1 now will buy class k. The In this section we develop the theory for a
incremental revenue contribution by opening completely general fare structure, which may
up class k can therefore be written as dk have any set of restrictions attributed to each
fkQk1( fk1fk). Another way to state this is fare product. The fare and the restrictions of a
that, instead of receiving the fare fk by opening fare product determine the probability that a
up class k, we obtain an adjusted fare f 0 k ¼ booking occurs for that product, which will
[dk fkQk1( fk1fk)]/dk. The adjusted fare generally depend on what other products are
is also known as the marginal revenue MRk and available as well.
may be calculated as the increment in total An optimization strategy consists of opening
revenue per capacity unit: up a given set of fare products, Z. Examples
include opening all products N ¼ {1, 2, y, n};
or keeping all products closed {}. In the
TRk TRk1 0
previous example, we considered the subsets
MRk ¼ ¼ fk
Qk Qk1 of opening fare products in decreasing fare
6 r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
Optimization of Mixed Fare Structures
order {};{1};{1, 2},y, {1, 2, y, n}, which are Total Marginal Efficient
Sm
Revenue revenue: f’ Si Frontier
i
‘nested’ strategies. If we now allow a strategy to
TRi
be any set of fare classes open, there will be 2n Si-1
TRi-1
different strategies, ZDN.
The demand for a fare product depends in S1
general on the particular optimization strategy All choice sets: Z
d’i
Z. The demand for fare product j given strategy
S0 CAP
Z will be denoted dj(Z). The total demand
Qi-1 Qi Demand Q
P a given strategy is calculated as Q(Z) ¼
for
jAZ dj(Z). The corresponding total revenue Figure 3: Construction of convex hull.
for a given strategy (assumingPno capacity
constraint) is given by TR(Z) ¼ jAZ dj(Z)fj .
The general optimization problem can then Q(Z), which is a weighted average over the
be stated as: max TR(Z) subject to Q(Z)pcap. fares in Z.
The solution to the problem can most easily be The marginal revenue transformation maps
solved by analyzing TR(Z) versus Q(Z) in a the original fare structure into the transformed
scatter plot where each of the 2n different fare structure. The demands for the transformed
strategies are plotted. These correspond to products are given by the marginal demand d 0 k
what we can call ‘pure’ strategies. In addition to while the fares of the transformed fare products
the pure strategies, we may construct ‘mixed’ are given by the marginal revenue.
strategies by combining the pure strategies such
that each of the pure strategies are only used for d 0k ¼ Qk Qk1 ; k ¼ 1; . . . ; m
a fraction of the time.
The mixed strategies trace out the convex f 0k ¼ ðTRk TRk1 Þ=ðQk Qk1 Þ; k ¼ 1; . . . ; m
hull of the pure strategies. This is marked in
Figure 3 as the shaded area, meaning that any The marginal demand associated with each
point within the shaded area is feasible. strategy k is also called the partitioned or
The solution to the optimization problem incremental demand. The marginal revenues
max TR(Z) subject to Q(Z)pcap is identified are called adjusted fares. The transformation is
as the upper boundary on the convex hull. This identical to the previous MR expression. The
portion marked with the solid lines is termed usefulness of the marginal revenue transforma-
the ‘efficient frontier’. Let the subset of tion follows from the following theorem,
strategies that lies on the efficient frontier which we have proven by direct construction
(efficient sets) be denoted S0, S1, y, Sm. Here of the convex hull (since the independent
S0 ¼ {} is the empty set with TR(S0) ¼ 0 and demand model with demand d 0 k and fares f 0 k
Q(S0) ¼ 0. The total demand and total revenue would trace out exactly the same convex hull
of strategy Sk is denoted by Qk and TRk and are therefore equivalent).
respectively. Note that Figure 3 illustrates the Theorem: The marginal revenue transforma-
general case where the optimal solution is tion transforms the original fare structure into
obtained by a mixed strategy. an equivalent model with independent fare
These sets are ordered according to quantity products.
Qk1pQk and TRk1pTRk for k ¼ 1, y, m.
The end-point for the efficient frontier is strategy Implementation of control
Sm, representing the maximum total revenue. mechanism using fare adjustment
The first non-empty set is in general We have shown in the last section that the
S1 ¼ {1}, the highest fare in the fare structure, marginal revenue transformation transforms
since it has the highest value slope, TR(Z)/ the optimization problem for an arbitrary fare
r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19 7
Fiig et al
structure to an equivalent independent demand mechanism suggests that we will in practice apply
model. But how is the booking control mecha- a nested approximation, even in cases where the
nism transformed? A booking control mecha- nesting property is not exactly satisfied.
nism in traditional RM systems calculates This simplification also applies to bid price
numerical seat availability for each booking control. Here, the marginal revenue going from
class based on the solution of the optimi- one strategy to the next can be assigned to the
zation problem and the current seat inventory. newly added classes. The bid price control is
However, in the transformed model the therefore modified by using the marginal
strategies on the efficient frontier can involve revenue associated with the booking class
correspond to a subset of the original booking rather than the fare itself.
classes, but not necessarily all classes in
decreasing fare order. APPLICATIONS TO DIFFERENT
Consider for example an efficient frontier FARE STRUCTURES
with the efficient strategies {1, 2} at capacity 1, In this section, we describe the application
{1, 3} at capacity 2, while {1, 2, 3} is inefficient. of the marginal revenue transformation and
If the capacity is 2, a nested booking limit fare adjustment theory to different airline fare
control would protect 1 seat for {1, 2} and allow structures – from traditional differentiated fares
1 booking for {1, 3}, thus the seat availability for to completely undifferentiated fares, along with
{1, 2} would be 2 and for {1, 3} 1 correspond- ‘hybrid’ fare structures that include both types
ingly. It is not possible to map this to seat of fare products.
availability for the original classes without
opening the inefficient strategy {1, 2, 3}. New Fully differentiated fare products
distribution mechanisms would be required, Under the traditional RM assumption of fare
where the exact number of requested seats has class independence, the fare products are
to be known in advance and the corresponding assumed to be adequately differentiated, such
strategy can be shown accordingly. that demand for a particular fare product will
Many choice models have the desirable only purchase that fare product. Therefore, for
property that the strategies on the efficient any set of fare products with fares fj, and
frontier are nested, SkCSl, kol, for example demands dj, j ¼ 1, y, n, the marginal revenue
{1}, {1, 3}, {1, 2, 3}. Note that the nesting transformation yields
does not necessarily have to follow the fare
order. On such nested efficient frontiers, a class d 0k ¼ Qk Qk1
once contained in an efficient strategy will also X X
be contained in any of the next efficient ¼ dj dj ¼ dk
j¼1;...;k j¼1;...;k1
strategies as capacity increases. It is therefore
possible to map the original booking classes to
the strategies in which they occur first with and
increasing capacity (or going from left to right TRk TRk1
on the efficient frontier). f 0k ¼
Qk Qk1
When the efficient frontier is nested, the P P
marginal revenue transformation can therefore dj fj dj fj
j¼1;:::; k j¼1;:::; k1
be interpreted as the assignment of demands ¼ ¼ fk
dk
and fares to the original booking classes which,
when fed into the optimization algorithm k ¼ 1, y, n. Thus for the differentiated fare
and the control mechanism for independent product case, the marginal revenue transforma-
demand, produces the correct availability con- tion acts as an identity and both demands and
trol. This ease of implementing the control fares are unaffected. We can refer to this
8 r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
Optimization of Mixed Fare Structures
0
demand as being ‘product-oriented’. Let us fk ¼ fk fM ; where
denote the unadjusted fare by f 0 prod
k . expðb DÞ
fM ¼ D
1 expðb DÞ
Completely undifferentiated fare
products The adjusted fare can therefore be expressed a
We now apply the marginal revenue transfor- the original fare minus a positive constant
mation to a single flight leg with a fully called the fare modifier, fM. For a general fare
undifferentiated fare structure, again assuming structure, the fare modifier will differ by class.
deterministic demand. The definitions of fares Because the fare modifier is the cost associated
fj, and demands dj, j ¼ 1, y, n are unchanged. with buy-down, it is also termed the price
Due to the fully undifferentiated fare structure, elasticity cost.
passengers will only buy the lowest available In Table 1, the constant fare modifier can
fare. Let k be the lowest open fare product, be seen fM ¼ $572. Since seats are only
then the demand for all other fare pro- allocated to fare classes with positive adjusted
ducts becomes zero, dj ¼ 0 for jak Hence the fares, we can interpret fM as the lowest open
cumulative demand equals the demand for the fare. In the limit D-0, using l’Hopital’s rule
lowest open fare product dk ¼ Qk, k ¼ 1, y, n. we obtain fM-1/b for D-0, which depends
The demand Qk ¼ Qn psupk (assuming fk is only on the beta parameter in the exponential
the lowest open) can be expressed as a sell-up demand model.
probability psupk from n-k, times a base
demand Qn for the lowest class n. Hybrid demand in mixed fare
The marginal revenue transformation deter- structures
mines the adjusted fares. The demand In this section we show how to optimize a
is calculated as d0 k ¼ QkQk1 ¼ dkdk1 single leg with a fare structure consisting
¼ Qn(psupkpsupk1). of a mix of differentiated and undifferentiated
fare products assuming deterministic demand,
The adjusted fares are given by using the marginal revenue transformation.
TRk TRk1 fk Qk fk1 Qk1 This is the so-called hybrid demand case. The
fk 0 ¼ ¼ fare class demands are decomposed into
Qk Qk1 Qk Qk1
fk psupk fk1 psupk1 contributions from both differentiated (pro-
¼ : duct-oriented) and un-differentiated (price-
psupk psupk1
oriented) demand:
Let us denote this by f 0 price
k . dj ¼ dj
prod
þ dj
price
; j ¼ 1; . . . ; n
Note that the highest fare value is un-
adjusted: The marginal revenue transformation deter-
mines the adjusted fares. The demand for each
0 TR1 TR0 f1 d1 0 fare class is obtained as
f1 ¼ ¼ ¼ f1 !
Q1 Q0 d1 0 0
X prod price
dk ¼ Qk Qk1 ¼ dj þ dk
Consider the special case of exponential j¼1;...;k
sell-up Qk ¼ Qn psupk ¼ Qn exp(b( fkfn)) !
X prod price
and assume for simplicity an equally spaced dj þ dk1
fare-grid with fk1fk ¼ D. The marginal j¼1;...;k1
revenue transformation yields: ¼ dk
prod
þ Qn ðpsupk psupk1 Þ
0
dk ¼Qn expðbfn Þ
Define rk ¼ dkprod/(dkprod þ Qn(psupkpsupk1)) as
ðexpðbfk Þ expðbfk1 ÞÞ: the ratio of product demand to total demand.
r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19 9
Fiig et al
10 r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
Optimization of Mixed Fare Structures
one can write the optimization problem as: Marginal revenue transformation
The optimal strategy as function of the bid-
Jt1 ðxÞ ¼ Jt ðxÞ þ l maxðTRðZÞ price can only consist of one of the vertices on
ZN
ð1Þ the efficient frontier. As long as the bid-price is
BPt ðxÞ QðZÞÞ higher than the highest fare, we should close all
Observe now that for all x and t, this problem is classes.1 We denote this strategy by S0. If the
of the same form, namely bid-price falls below f1 the optimal strategy is
initially S1, until the bid-price falls below the
maxðTRðZÞ BP QðZÞÞ ð2Þ slope of the convex hull between S1 to S2. The
ZN
slopes of the efficient frontier correspond to
where only the value of the bid-price may the critical values of the bid-price, where the
change. optimal strategy changes. The optimal strategy
for a given bid price BP is found by walking
Efficient frontier along the efficient frontier from one strategy to
The problem of optimizing maxðTRðZÞ BP the next, increasing sales volume, as long as
ZN
QðZÞÞ is most easily analyzed by drawing
TR(Z) versus Q(Z) in a scatter plot for all TRðSk Þ TRðSk1 Þ
MRk ¼ XBP:
possible strategies of open classes, together with QðSk Þ QðSk1 Þ
the straight line BP Q for specific values of the
bid-price. For comparison, the plot in Figure 4 The last strategy obtained this way is the optimal
is identical to Figure 3 describing the efficient strategy of open booking classes for bid price BP.
frontier for the deterministic case. Once the efficient frontier is known, the
The optimum for the specific BP in the bid-prices can be calculated from the DP
chart is attained where the parallel shifted formulation recursively. Because the efficient
line BP Q is tangent to the convex hull of the frontier contains only a few of the possible
r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19 11
Fiig et al
strategies, there are potentially many fare independent optimizations of the form (2) for
structures with corresponding choice models each itinerary.
(identified by a set of pj(Z)) which will lead to The DP network model has no practical
the same efficient frontier. All these models will importance, since the number of possible states
have the same bid-price. In particular consider explodes with capL. However the marginal
the independent demand model defined by: revenue transformation holds as well. This leads
us to conclude that any heuristic network
0
pk ¼ QðSk Þ QðSk1 Þ; k ¼ 1; . . . ; m: algorithm which assumes independent demand
by O-D itinerary can be easily extended to deal
with arbitrary fare structures by applying the
0
fk ¼ MRk marginal revenue transformation. This will be
TRðSk Þ TRðSk1 Þ illustrated in the example with DAVN-MR
¼ ; k ¼ 1; ; m below.
QðSk Þ QðSk1 Þ
12 r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
Optimization of Mixed Fare Structures
easily solved since only classes where f 0 kXBP For a general fare structure we need to apply
should be included in the strategy S. If we the formalism presented in the previous
further insert the exponential sell-up with sections, which consists of the following steps:
equally spaced fares we obtain
X 0 0 (1) Determine the strategies k ¼ 1, y, m on
max pi ð fi BPÞ the efficient frontier.
S
i2S
(2) Apply the marginal revenue transformation
¼ max½expðb ðfk fn ÞÞ ðfk BPÞ to both demands and fares.
k
(3) Apply EMSRb in the normal fashion using
We have shown in the section ‘Applications to the transformed demands and fares.
different fare structures’ that the fare modifier
This method, called EMSRb-MR, thus
for exponential sell-up and equally spaced
considers independent normally distributed
fare grid is a constant. Thus the open classes
demands d 0 kBN(mk, s2k), with fares f 0 k, k ¼ 1,
are exactly those which have fifMXBP, a
y, m. Note that the demands are related to the
result which would not be immediately
strategies not the individual fare products. The
clear from optimization problem on the right
aggregated demand, variance, and average fares
hand side.
are calculated analogously using the trans-
formed demands and fares and hence
Leg RM: EMSRb-MR the formulas for the joint protection level
Here we show how EMSRb can be extended and the booking limit become exactly the
0 0
same: pk ¼ m1;k þ s1;k F1 ð1 fkþ1 =f1;k Þ, and
0
to optimize a general fare structure. First we
will briefly consider the results for standard BL0 k ¼ capp0 k.
unadjusted EMSRb, in order to be able to Consider again the case of fully undiffer-
compare with the transformed EMSRb-MR. entiated fare products with equal spaced fares,
In standard EMSRb we consider independent as presented in Table 1 and in section
fare products with fares fi, with normally ‘Applications to different fare structures’. In
distributed demands diBN(mi, s2i ), i ¼ 1, y, n. this case the strategies become nested and we
The aggregated demand for fare product may associate an adjusted fare with each
i ¼ 1, y, k is also strategy. For simplicity (and ease of compar-
Pknormally distributed 2with
¼ ison) we will assume that the demands and
average
Pk 2 m 1,k i ¼ 1 mi, P k s1,k ¼
variance
1
i ¼ 1si and average fare f1;k ¼ m1;k i¼1 mi fi .
standard deviations are identical to those of the
The joint protection level pk is then given by fully differentiated case.
pk ¼ m1;k þ s1;k F1 ð1 fkþ1 =f1;k Þ, and the Table 2 compares the standard EMSRb
booking limit by BLk ¼ cappk. results in the case of differentiated fares with
r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19 13
Fiig et al
the modified booking limits coming from the revenue contribution of accepting an O&D
EMSRb-MR approach in the case of undiffer- passenger by subtracting the displacement costs
entiated fares. Note that the adjusted fares used of carrying a passenger on connecting legs.
as input to EMSRb after the marginal revenue Further, we have used the previous result that
transformation lead to closure of the lowest two the adjusted fare can be expressed as the fare
classes with negative marginal revenues. Classes minus a fare modifier. Thus this formula
2 through 4 also get reduced seat availability, extends the standard DAVN formula by just
given that the estimated potential for sell-up an additional term ‘fM’. Note that fM ¼ 0 for
has been incorporated into the marginal differentiated fare products. The undifferen-
revenue values of each class. tiated fare products are mapped to lower
The adjusted EMSRb-MR overcomes spiral buckets since fM>0 (except the highest un-
down by modifying the resulting booking differentiated fare class). Also note that all
limits to account for the estimated sell-up undifferentiated fare products with fares fofM
potential, and thereby the dependent demand will be closed (regardless of remaining capacity)
forecasts, which results in closing availability since they map to negative adjusted fares.
to the inefficient fare products (that is with DAVN-MR is then applied by running DAVN
negative adjusted fares). in the standard fashion on the transformed fare
structure to produce booking limits by bucket
Network RM: DAVN-MR with (in marginal revenue scale) for each leg. Finally,
differentiated and undifferentiated note that the formula for the adjusted fare is
fare structures valid for an arbitrary fare structure with nested
For a large network airline, it is common efficient frontier, but there the fare modifier
to offer a variety of fare structures in different will vary by class.
O-D markets, each with different restrictions,
as required to respond to competitors in each SIMULATION RESULTS USING
market. However, individual flight legs to
PODS
or from a connecting hub provide seats to
In this section, we present simulation results
multiple O-D markets, each of which can have
from PODS (Passenger Origin Destination
different fare structures. In this section we show
Simulator) to illustrate the revenue impacts of
how to optimize a network with a mix of
applying the marginal revenue transformation
differentiated fare products and undifferenti-
in DAVN-MR, as described above. PODS and
ated fare products. As an example we will
its various components have been widely
use DAVN (Displacement Adjusted Virtual
Nesting).
The marginal revenue transformation ap- PATH/CLASS
AVAILABILITY
plied to DAVN (termed DAVN-MR) results in PASSENGER REVENUE
DECISION PATH/CLASS MANAGEMENT
transforming both demands and fares as MODEL BOOKINGS/
CANCELLATIONS
OPTIMIZER
described in the previous sections. A particu- CURRENT FUTURE
larly simple expression for the adjusted fare is BOOKINGS BOOKINGS
14 r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
Optimization of Mixed Fare Structures
1
2
3 H1(41) 33
5 4
6 37
32 40 26
23 31
7 29
28 24 21
8 27
9 11 22 34
35
10 12 38
14 15 36
16 13
17 18
H2(42)
30
19 20
25
39
Traffic Flows
UNDIFFERENTIATED SEMI-DIFFERENTIATED
described in the literature – see for example, In PODS, each simulated passenger chooses
Hopperstad (1997), Belobaba and Wilson among available path/class alternatives (that is,
(1997), Lee (1998), and Carrier (2003). The itinerary and fare product combinations) with
central concept in PODS is that the models of fares lower than that passenger’s randomly
the revenue management systems and the drawn maximum willingness-to-pay (WTP).
model of passenger choice are contiguous but Business passengers have a higher input mean
independent. That is, the revenue management WTP than leisure travellers. If there exist more
system(s) define the availability of path/classes than one available path/class meeting this WTP
in a market and the generated passengers pick criterion, the passenger then chooses the
from the available path/classes and book. The alternative with the lowest generalized cost,
only information available to the revenue defined as the sum of the actual fare and the
management system(s) of the simulated airlines disutility costs with the restrictions (if any)
for setting path/class availability is the history of associated with each fare class. A passenger will
(previously simulated) passenger bookings. ‘sell up’ from a preferred lower fare class (with a
Figure 5 illustrates this basic architecture. lower generalized cost) to a more expensive
r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19 15
Fiig et al
one, therefore, if the lower class is closed due to of cancellation and a fee for changing reserva-
RM controls and if the higher fare is still lower tions) are applied to lower fare classes (Figure 7).
than the passenger’s WTP. The simulated sell- The focus of the simulation study is on
up behaviour is thus imbedded in the PODS Airline 1 and its use of hybrid forecasting and
passenger decision model, rather than specified DAVN-MR. In the simulations, the effects of
as an input sell-up parameter. each of the two components of the transforma-
Figure 6 shows the geometry of network D, tion were studied separately. Three scenarios
one of the standard PODS networks. In net- were considered:
work D there are two airlines, each with their
own hubs. The traffic flow is from the 20 K Standard: Airline 1 uses standard path/fare
western cities to the hubs and to the 20 eastern class forecasting with DAVN (baseline).
cities. Both airlines employ three connecting K Hybrid: Airline 1 changes to Q/hybrid
banks per day, resulting in each airline operat- forecasting, without fare adjustment.
ing 126 legs, which produce 1446 paths in 482 K DAVN-MR: Airline 1 adds fare adjustment
markets. to its Q/hybrid forecasting.
Six fare classes are assumed, and simulation Note that hybrid forecasting for the fully
results are provided here for both an undi- undifferentiated fare structure corresponds to
fferentiated and a semi-differentiated fare only using the price-oriented forecasting
structure, as shown in Figure 7. In the undi- model (so-called Q-forecasting) [Belobaba
fferentiated case there are no fare restrictions or and Hopperstad (2004)].
advance purchase rules. In the semi-differen- The three scenarios are examined, for the
tiated case, two restrictions (no refunds in case undifferentiated and the semi-differentiated
100.0
Revenue Index
100 100
75 75
50 50
25 25
0 0
Standard Hybrid DAVN-MR Standard Hybrid DAVN-MR
60 60
40 40
20 20
0 0
Standard Hybrid DAVN-MR Standard Hybrid DAVN-MR
Figure 8: Simulation results for the monopoly case with undifferentiated and semi-differentiated fare-structures in
network D.
16 r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
Optimization of Mixed Fare Structures
fare structures, first in a monopoly environment potential of the price oriented demand and
where only Airline 1 is present and then reduces spiral down. Note that load factors
in a competitive environment where the for standard and hybrid forecasting are almost
second airline uses standard path/fare class identical. Hybrid forecasting can be thought
forecasting and traditional DAVN network of as an attempt to ‘re-map’ passengers
RM optimization. that bought down back to their original fare
class.
Applying fare adjustment by moving to full
Monopoly case
DAVN-MR leads to a further revenue increase
The simulation results are presented in
of about 16 per cent in both cases. The revenue
Figure 8. In the panels the revenue index and
benefit of applying fare adjustment is significant
load factor (LF) are shown both for the
since its effect is to close lower fare classes,
undifferentiated and the semi-differentiated
which further restricts buy-down. This effect
fare structures. The revenue index is reported
can clearly be seen from the load factors.
relative to the baseline monopoly case in which
Applying fare-adjustment leads to a significant
the airline employs standard forecasting and
drop in load factor from 85 per cent to 72 per
optimization.
cent–74 per cent.
The results show that applying hybrid
forecasting only, for both fare structures, leads
to significant revenue gains on the order Competitive case
of 10 per cent–16 per cent. The reason is that The simulation results for the two-airline
hybrid forecasting considers the buy-down competitive case are presented in Figure 9,
Load Factor
68.5
64.6
50 50
25 25
0 0
Standard Hybrid DAVN-MR Standard Hybrid DAVN-MR
AL1 AL2 AL1 AL2
Figure 9: Simulation results in the competitive case with undifferentiated and semi-differentiated fare-structures in
network D.
r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19 17
Fiig et al
18 r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19
Optimization of Mixed Fare Structures
York, NY: The Aviation Weekly Group of the McGraw-Hill Yield Management Study Group Meeting, Cape Town, South
Companies, pp. 285–302. Africa, May.
Belobaba, P. and Hopperstad, C. (2004) Algorithms for revenue Lautenbacher, C. J. and Stidham, S. J. (1999) The underlying
management in unrestricted fare markets. Presented at the markov decision process in the single-leg airline yield
Meeting of the INFORMS Section on Revenue Manage- management problem. Transportation Science 33: 136–146.
ment, Massachusetts Institute of Technology, Cambridge, Lee, A. Y. (1998) Investigation of competitive impacts of origin-
MA. destination control using PODS. Unpublished Master’s
Belobaba, P. P. and Weatherford, L. R. (1996) Comparing Thesis, Massachusetts Institute of Technology, Cambridge,
decision rules that incorporate customer diversion in perish- MA.
able asset revenue management situations. Decision Sciences Lee, T. C. and Hirsch, M. (1993) A model for dynamic airline
27(2): 343–363. seat inventory control with multiple seat bookings. Transporta-
Belobaba, P. P. and Wilson, J. L. (1997) Impacts of yield tion Science 27: 252–265.
management in competitive airline markets. Journal of Air Littlewood, K. (1972) Forecasting and Control of Passenger Bookings,
Transport Management 3(1): 3–10. 12thAGIFORS Annual Symposium Proceedings, Nathanya,
Boyd, E. A. and Kallesen, R. (2004) The science of revenue Israel, pp. 95–117.
management when passengers purchase the lowest available McGill, J. I. and van Ryzin, G. J. (1999) Revenue management:
fare. Journal of Revenue and Pricing Management 3(2): 171–177. Research overviews and prospects. Transportation Science 33(2):
Bratu, S. J-C. (1998) Network value concept in airline revenue 233–256.
management. Master’s thesis, Massachusetts Institute of Saranathan, K., Peters, K. and Towns, M. (1999) Revenue
Technology, Cambridge, MA. Management at United Airlines. AGIFORS Reservations and
Brumelle, S. L. and McGill, J. I. (1993) Airline seat allocation Yield Management Study Group, April 28, London.
with multiple nested fare classes. Operations Research 41: Simpson, R. W. (1989) Using network flow techniques to
127–137. find shadow prices for market and seat inventory control,
Carrier, E. J. (2003) Modeling airline passenger choice: Passenger memorandum M89-1, MIT Flight Transportation Laboratory.
preference for schedule in the passenger origin-destination Cambridge, MA: Massachusetts Institute of Technology.
simulator. Unpublished Master’s Thesis, Massachusetts Smith, B. C., Leimkuhler, J. F. and Darrow, R. M. (1992)
Institute of Technology, Cambridge, MA. Yield management at american airlines. Interfaces 22(1):
Cooper, W. L., Homem-de-Mello, T. and Kleywegt, A. J. (2006) 8–31.
Models of the spiral-down effect in revenue management. Smith, B. C. and Penn, C. W. (1988) Analysis of alternative
Operations Research 54(5): 968–987. origin-destination control strategies. AGIFORS Symposium
Curry, R. E. (1990) Optimal airline seat allocation with fare Proceedings 28: 123–144.
classes nested by origin and destinations. Transportation Science Talluri, K. T. and van Ryzin, G. (1998) An analysis of bid-price
24(3): 193–204. controls for network revenue management. Management
Fiig, T., Isler, K., Hopperstad, C. and Cleaz-Savoyen, R. (2005) Science 44: 1577–1593.
DAVN-MR: A Unified Theory of O&D Optimization in a Talluri, K. T. and van Ryzin, G. (2004) Revenue management
Mixed Network with Restricted and Unrestricted Fare under a general discrete choice model of consumer behavior.
Products. AGIFORS Reservations and Yield Management Study Management Science 50(1): 15–33.
Group Meeting, Cape Town, South Africa, May. Tretheway, M. W. (2004) Distortions of airline revenues: Why
Gallego, G. and van Ryzin, G. (1997) A multi-product, multi- the network airline business model is broken. Journal of Air
resource pricing problem and its application to network yield Transport Management 10(1): 3–14.
management. Operations Research 45: 24–41. Vinod, B. (1995) Origin and destination yield management. In:
Guo, J. C. (2008) Estimation of sell-up potential in airline D. Jenkins (ed.) The Handbook of Airline Economics. New York,
revenue management systems. Unpublished Master’s Thesis, NY: The Aviation Weekly Group of the McGraw-Hill
Massachusetts Institute of Technology, Cambridge, MA. Companies, pp. 459–468.
Hopperstad, C. H. (1997) PODS Modeling Update. AGIFORS Williamson, E. L. (1992) Airline network seat inventory
Yield Management Study Group Meeting, Montreal, Canada. control: methodologies and revenue impacts. Ph.D.
Hopperstad, C. H. (2007) Methods for Estimating Sell-Up. Thesis, Massachusetts Institute of Technology, Cambridge,
AGIFORS Yield Management Study Group Meeting, Jeju Island, MA.
Korea. Wollmer, R. D. (1992) An airline seat management model for a
Isler, K., Imhof, H. and Reifenberg, M. (2005) From Seamless single leg route when lower fare classes book first. Operations
Availability to Seamless Quote. AGIFORS Reservations and Research 40(1): 26–37.
r 2009 Palgrave Macmillan 1476-6930 Journal of Revenue and Pricing Management 1–19 19