Belobaba - Airline Seat Inventory PDF
Belobaba - Airline Seat Inventory PDF
Belobaba - Airline Seat Inventory PDF
REPORT R87-7
ABSTRACT
Many airlines practice differential pricing of fare products that share a common
inventory of available seats on an aircraft. Seat inventory management is the process
of limiting the number of seats made available to each fare class. The objective
of both strategies is to maximize the total revenues generated by the mix of fare
products sold for a flight.
This dissertation first examines the evolution of airline marketing and seat in-
ventory management practices. A demand segmentation model is developed to help
explain current airline fare structures. A conceptual model of the consumer choice
process for air travel is then presented, and extended to describe the airline reserva-
tions process and the probabilistic elements that can affect seat inventory control.
A survey of current airline practice in this area revealed that seat inventory
control is an ad-hoc process which depends heavily on human judgement. Past work
on quantitative approaches has focused on large-scale optimization models that solve
simple representations of the problem. A primary objective of this research was the
development of a quantitative approach based on the practical constraints faced by
airlines.
The Expected Marginal Seat Revenue (EMSR) model developed in this thesis
is a decision framework for maximizing flight leg revenues which can be applied to
multiple nested fare class inventories. It is applied to a dynamic process of booking
limit revision for future flight departures, and overbooking factors as well as fare
class upgrade probabilities are incorporated. Examples of EMSR model results are
presented, and a critical analysis of the demand assumptions and sensitivity of the
model is performed.
The EMSR model was implemented as part of an automated seat inventory con-
trol system at Western Airlines and tested on a sample of actual flights. Compared
to flights managed by existing manual methods, flights for which fare class booking
limits were set and revised automatically on the basis of the EMSR decision model
carried more passengers at a lower yield, and generated higher total revenues.
Acknowledgements
Research funding and technical assistance was also provided by the M.I.T.-
Industry Cooperative Research Program at the Flight Transportation Laboratory.
As a member of the Program, Trans World Airlines furnished reservations data for
empirical analysis early in this research. McDonnell-Douglas representative John
Stroup assisted in the initial development of computer code for seat allocation mod-
els.
Personal thanks go to my closest friends: Mike Meyer, for being the first to
encourage me to pursue a Ph.D.; Brad Martin, for providing last-minute help with
tables and figures; Carole Marks, for her genuine interest in my progress and feigned
interest in my research; and Eric and Liane Schreffler, for their hospitality in Los
Angeles.
Contents
Introduction
6.1 EMSR Protection Levels for Different Estimate of Demand Variation 155
6.2 EMSR Protection Levels Plotted Against Input Variables . . . . . . 156
6.3 The Virtual Nesting Concept . . . . . . . . . . . . . . . . . . . . . . 160
Having seen a newspaper advertisement for air travel between Boston and Los
Angeles for $ 198 round-trip, a potential traveler calls the airline and is told that
the lowest available fare for the specific dates and flights requested will be $ 358.
After agreeing to accept this higher fare, the traveler ends up sitting next to another
traveler who has paid over $ 500 for one-way passage. This scenario has become
increasingly familiar to air travelers since the deregulation of the U.S. airline industry
in 1978. Many airlines offer a wide range of fares for travel in a single city-pair
market as they attempt to fill empty seats and realize as much revenue as possible
from each seat sold.
Seat inventory management is the process of balancing the number of seats sold
at each of the fare levels offered so as to maximize total passenger revenues. This
practice enables airlines to influence their total revenues on a flight-by-flight basis,
within a given price structure. Seat inventory control and pricing are two distinct
strategies that comprise what can be referred to as airline revenue management.
This dissertation provides a comprehensive discussion of the airline revenue man-
agement problem, with an emphasis on seat inventory control.
The evolution of the theory and practice of seat inventory management is re-
viewed first in Part One. The economic principles behind airline pricing and market
demand segmentation practices are used to explain the airline fare structures cur-
rently in place. A model of how consumers make air travel decisions is developed,
and the implications for the airline reservations process are discussed. The seat
inventory control problem faced by airlines is then defined and a survey of current
practice allows the practical constraints for solution approaches to be identified.
Part Two is devoted to quantitative methods for seat inventory control, begin-
ning with a review of previous work in the area. The shortcomings of applying past
approaches to the practical problems faced by airlines lead to the development of a
quantitative decision framework for revenue maximization that can be applied di-
rectly to existing fare structures and reservations systems. The Expected Marginal
Seat Revenue (EMSR) model is applied to the problem of revising booking lim-
its on different fare types dynamically, and extended to account for overbooking
and passenger upgrades when reservations requests are denied. The EMSR model
is examined critically with respect to its demand assumptions and sensitivity to
input variables, and its application to origin-destination seat inventory control is
considered.
The final part of this dissertation describes the application of the EMSR frame-
work at an airline, as the development of an automated seat inventory control system
at Western Airlines is discussed. A test of the automated system relative to existing
manual seat inventory control methods was performed, and the results demonstrate
the potential benefits of employing quantitative decision models in the process. The
dissertation concludes with a discussion of future seat inventory control system de-
velopment, and directions for further theoretical and empirical research.
PART ONE
The Evolution of Seat
Inventory Management:
Theory and Practice
In coping with these unique properties of their product "inventory", airlines have
implemented pricing and marketing schemes designed to minimize the "spoilage"
of unsold seats. For many years, carriers pursued incremental demand with "dis-
counted" fares and services while striving to retain their original full-fare traffic
bases. With the increased freedoms of deregulation in the United States, this effort
to stimulate demand and fill empty seats has intensified. A wide variety of differen-
tiated service/price combinations are offered in most markets. With the emergence
of differential pricing of seats that can share a common aircraft cabin, and therefore
the same seat "inventory", the problem of managing these seat inventories so as to
maximize revenue and operating profits has also emerged.
Many airlines have therefore become extremely interested in what is commonly
referred to as "yield management". Yield is the revenue per passenger-mile of traffic
carried by an airline. An airline's yield will be a function of both the prices it charges
for its differentiated service options and the number of seats actually sold at each
price. Yield management thus involves two major components - pricing and seat
inventory control. The emphasis of this dissertation is on the latter. Seat inventory
controlis the process of limiting the number of seats to be made available to service
options with different price levels on a future flight departure.
Part One of this thesis is devoted to the theory and practice of airline seat
inventory control. It introduces the basics of the seat inventory control problem
and outlines its implications from the perspective of both the airline wishing to
manage its seat inventory and the consumer faced with a series of travel decisions
given a choice of service options. The current state of seat inventory control practices
is also reviewed.
Chapter One reviews the development of airline practices and the inherent char-
acteristics of airline operations that led to the need for more sophisticated seat inven-
tory control. The economic concepts of differential pricing and market segmentation
are presented in the airline context as the underlying rationale for the evolution of
current marketing practices in the airline industry. Included is an overview of the
wide range of fare levels, classes of service, and ticket/travel conditions found in the
U.S. airline marketplace.
Chapter Two opens with a model of the decision process for air travel consumers
that has developed in response to changing airline marketing and pricing strategies.
Given a range of price and service options, and an ever-increasing choice of carriers
and routing alternatives, consumers have changed the way in which they make travel
decisions. This model of consumer behavior is extended to illustrate the reservations
process as viewed from the airline's perspective. Greater choice on the part of the
consumer has meant even less predictable demand for the individual airline flight
than was the case a decade ago.
The complexity of the seat inventory control problem for large airlines is ex-
plored in Chapter Three. The characteristics of airline operations and reservations
procedures that contribute to the size and definition of the problem are discussed.
An overview of current seat inventory control practices is then presented. Based on
a survey of nine large North American airlines, this discubsion of current practice
highlights both the range of sophistication found in the industry and the apparent
need for better mathematical and modeling techniques for seat inventory control.
Chapter 1
Determining the price a potential passenger would pay to fly between two cities
used to be a relatively simple exercise. The biggest choice facing the passenger was
generally between the first class and regular economy or coach fares, which were
identical on all carriers and were derived on the basis of mileage-based fare struc-
tures imposed by the Civil Aeronautics Board, or IATA for international services. In
certain markets, price discounts might have been available to those meeting certain
age requirements (children, senior citizens), traveling as a family or group, belong-
ing to specific organizations (military, travel industry), or flying on selected "night
coach" flights. Apart from "night coach" fares, these special fares represented true
discounts in price for a standard service. Potential passengers had only to choose
between the two or three carriers serving the city-pair market desired on the basis
of reputation, service and schedule convenience.
Needless to say, the airline pricing situation has changed dramatically. The prac-
tice of discounting prices on standard service options has given way to the creation
of differentiated service options offered at prices lower than the standard coach fare.
This chapter describes the evolution of airline marketing strategies that has led to
the wide range of price and service options offered in the current airline market-
place. The theoretical economic concepts on which these strategies have been based
are presented first, and the development of a need to control seat inventories more
effectively is outlined. The second section of this chapter explains how airlines have
applied these theoretical concepts and adapted them to theepractical constraints of
their reservations and ticketing processes. Finally, a description of the price and
service options available in most domestic airline markets at the present time pro-
vides an up-to-date illustration of the extent to which the simple pricing structure
of the 1960's has changed.
1.1 Empty Seats and Capacity-Controlled Fares
Restrictions on the use of these lower fares were imposed by the airlines to limit
their use to leisure or vacation travelers. The original objective of these fares was to
stimulate new demand in the discretionary travel market segment, which would in
turn fill otherwise empty seats and generate additional revenues for the airline. By
applying advance purchase, length of stay and round-trip itinerary requirements,
airlines hoped to keep full-fare business travelers from making use of these lower
fares, and thereby introduced a new type of air travel service option to the market.
Research done by the Boeing Aircraft Company into the nature of empty seats
on the larger wide-body aircraft that entered service in the 1970's encouraged air-
lines to apply such restrictions to reduced fares to prevent the diversion of existing
demand from higher fare levels [1]. Boeing argued that, due to an inherent variabil-
ity in demand for air travel and an inability on the part of the airline to provide
exactly the number of seats required to satisfy all full-fare demand, empty seats are
inevitable. The inevitability of having unsold seats and the potential for recover-
ing additional revenues from them led to Boeing's development of an approach for
"surplus seat management"[2], which involved forecasting the number of seats that
would otherwise remain empty.
In the case of air transportation, determining the relevant marginal cost for
pricing purposes is difficult. The marginal unit of production for an airline in the
short-run is a flight departure with a fixed seating capacity, whereas the marginal
unit sold to the consumer is a single seat on that flight. For a scheduled future
flight departure, fixed aircraft operating costs per passenger decrease as the number
of passengers actually carried increases, up to aircraft capacity. Assuming a single
class of service, variable costs per passenger are constant, independent of the number
carried on a particular flight. Total costs per passenger are thus minimized when
the fixed costs of operating the flight are spread over a full planeload of passengers.
Given these conditions, a single price charged for all seats based strictly on
marginal variable costs might not cover the total operating costs of the flight. For
that matter, depending on the cost structure of the airline and the level of demand,
there might not be any one price that will cover total operating costs of the flight.
It might be possible, however, for firms in an oligopolistic industry to segregate the
purchases of individual consumers and charge each consumer as much as he/she is
willing to pay. In theory, each incremental purchaser would be charged less and
less as more output is sold, up to the point at which price equals marginal cost. In
practice, a firm might be able to identify distinct groups of consumers and charge
each group a different price for a homogeneous product.
Figure 1.1 illustrates how differential pricing can enable the firm to cover to-
tal costs with total revenues, whereas strict marginal cost pricing would not. For
an isolated flight scheduled to operate with a given aircraft and fixed number of
identical seats, the marginal variable costs are constant per passenger carried. The
fixed aircraft operating costs per passenger decrease with the number of seats filled,
meaning total costs per passenger also decrease, as shown. In Figure 1.1, the de-
mand curve was drawn below the total cost per passenger curve to illustrate a case
Figure 1.1: Differential Pricing of Airline Seats: Single Flight Example
($/PAX)
P1
P2'
A AVERAGE
P3 TOTAL
-COSTS
___P MARGINAL
4 VARIABLE
COSTS
DEMAND
Q2 Q3 Q4 (SEATS)
in which no one price can generate total revenues to cover total flight operating
costs.
If price is set equal to marginal variable costs and applied to all passengers,
Q, passengers would be carried at a total revenue equal to the area of OP 4 BQ4 .
Total operating costs would be equal to the area OP3 AQ 4 , resulting in an operating
loss equal to P4 P3 AB. With the differential pricing strategy shown, Qi passengers
would pay P1, Q2 - Q1 passengers would pay P2, and so on. Under this strategy,
total revenues exceed total operating costs of the flight as long as the striped area
is greater than the shaded area.
1. the relevant costs of taking on incremental passengers are less than the average
total costs of the flight;
3. the incremental demand is sufficiently elastic that reduced fares increase total
revenues by more than the increase in total costs.
Differential pricing of identical seats on a flight departure can enable the airline
to generate the additional revenue required to cover the total costs of the flight from
those passengers using the reduced fare options. Airline efficiency can be enhanced,
as better use of existing equipment with its sunk costs can be realized by filling
otherwise empty seats. It is also possible for the airline to reduce cyclical variations
in demand over time through selective time-dependent price reductions and effective
seat inventory control, as will be discussed.
For the purposes of demand segmentation, airlines have recognized that differ-
entiated fare products can in fact be offered at different price levels because of the
price and service elasticities of air travel demand. Travelers very sensitive to price
will base air travel decisions almost exclusively on a lowest fare criterion and will
be willing to contend with the restrictions, reduced service amenities, and perhaps
less convenient flight times or routings associated with the lowest-priced fare prod-
ucts. Conversely, passengers with time constraints and service-sensitive travelers
will value level of service factors such as schedule convenience, travel flexibility and
inflight amenities, to the point that price might not be a decision factor in selecting
air travel. Of course, there exists a continuum between these two extremes, along
which the majority of air travelers are likely to fall.
Given this continuum between extreme price sensitivity and extreme service level
sensitivity, it can be difficult to divide the total demand in an airline market into
well-defined segments. Most airlines have in the past attempted to use various socio-
economic and travel characteristics of air travelers to identify distinct travel groups.
A distinction was made between the business and leisure segments of total market
demand, under the assumption that persons traveling on business are far more likely
to be sensitive to time or level of service factors and relatively price-insensitive. At
the same time, leisure travelers have been assumed to be more concerned about
price, and perhaps less sensitive to certain service-related characteristics.
Airlines attempted to ensure that only the leisure travel segment purchases low-
priced fare products through the imposition of ticketing and travel restrictions de-
signed to "fence out" the full-fare business segment. The most commonly imposed
fences were conditions requiring advance ticket purchase and round-trip travel with
a minimum length of stay at the destination, defined so as to preclude those on
business trips from making use of the discounted fares.
The use of "super-saver" -type fare products spread industry-wide in the United
States in the mid-1970's, and airlines found that properly fenced low fares could in
fact generate incremental leisure traffic. The problem that airlines faced, however,
was that the seats being sold to leisure passengers at lower fares were not necessarily
those that would have gone empty. Low-fare passengers wanted to choose the same
peak period flights that were already popular with business travelers. Even worse,
because advance purchase conditions required reduced fare passengers to book ear-
lier than most full-fare passengers, passengers paying lower fares in fact displaced
full-fare passengers from the most desirable flights. Airlines experienced flight loads
that were not significantly higher, total revenues that were actually lower on peak
period flights, and a loss of goodwill by business travelers frustrated by their inability
to obtain seats on desired flights.
First, market demand segments have become more difficult for airlines to iden-
tify. The business versus leisure classification of passengers adopted initially by air-
lines offering differentiated fare products overlooked all other types of non-business
travel. Leisure (vacation) travel is but one component of the non-business or per-
sonal travel segment, which includes travel to visit friends or relatives ("VFR"),
and personal emergency travel. Increases in the proportion of personal relative
to business air travel have made the identification of price- versus service-sensitive
demand segments more complicated. A survey of passengers conducted by the Cana-
dian Transport Commission (CTC) in 1979 found that up to 45 percent of business
and mixed business-pleasure trips in domestic markets could be considered price-
sensitive, while 39 percent of non-business trips were in fact non-discretionary and
thus service-sensitive [14].
Findings such as these indicate that market demand segmentation based solely
on trip purpose might not be the most appropriate. The characteristics of price-
sensitive versus on-demand travel suggest that, in most cases, a trade-off is made
by the passenger between the price of an airline ticket and the need to have travel
flexibility and service amenities, or even the need to travel at all. The CTC has
undertaken numerous passenger surveys in an attempt to identify more clearly the
main determinants of the price versus service trade-off, examining both trip and
traveler characteristics. While traveler characteristics had "very little impact on
price sensitivity independent of the type of trip being taken", trip characteristics
showed a strong relationship with the price versus service trade-off [151.
Trip purpose, trip distance and length of stay were all examined for their influ-
ence on price and service sensitivity, and the results of the various CTC surveys can
best be summarized in diagram form (see Figure 1.2). Trip purpose exhibited the
strongest relationship with the price-service trade-off, although considerable over-
lap by price-sensitive business travelers and service-sensitive non-business travelers
was evident in the results, as mentioned earlier. Trip purpose also showed a strong
relationship with length of stay, which in turn had only a moderate influence on
the price-service trade-off. Trip distance showed little relationship with the other
factors, apart from a moderate influence on length of stay.
Given that trip purpose seems to be the most important determinant of where
a passenger lies along the price-service trade-off continuum, it is not surprising
Figure 1.2: Trip Characteristics and the Price vs. Service Trade-off
PRICE/
SERVICE
TRADE-OFF
LOW
01 LOW
TRIP PURPOSE TRIP DISTANCE
2. Moderately discretionary consumers, who will travel only with a discount from
the full coach fare and are able to meet less stringent advance purchase re-
quirements (up to 7 days in advance), but are unable to meet minimum stay
requirements of a week or more, and are not willing to accept significant de-
viations from their desired flight times;
This choice is related to the price-sensitivity of the consumer for a given trip,
and is not independent of the time-sensitivity of the trip being considered. The
time-sensitivity of a trip is determined by the length of the "time window" over
which a trip may be taken and still provide the consumer with a certain "value"
of being at the desired location. A totally non-discretionary trip is one that must
be taken at a specified time, meaning the acceptable time window for the trip is
very short. On the other hand, a totally discretionary trip is one for which the
acceptable time window for travel is extremely long. For the purposes of market
demand segmentation, then, the time-sensitivity continuum incorporates the notions
of discretionary and non-discretionary travel, without reference to trip purpose.
Given that any one trip will lie somewhere along the time-sensitivity continuum,
and given that the potential traveler must fall somewhere along the price-sensitivity
continuum for each trip being considered, these two scales provide the basis for a
demand segmentation model that includes all possible trip/consumer characteristics.
We can define four generalized demand segments by dividing each continuum into
two sections, as shown in Figure 1.3.
The market demand segments defined by Figure 1.3 are relevant only to the
extent that differentiated fare products are offered to each segment. If only one
fare product is available on all flights in a market, all four segments shown would
be forced to purchase that fare product, or not travel at all. Given that a range of
differentiated fare products is offered in most markets, we can identify the promi-
nent characteristics of consumers in each quadrant of this segmentation model. The
dividing lines between the segments are by no means exact, and each segment still
encompasses a wide range of consumer characteristics and trip itinerary require-
ments. As such, no single definition of the typical consumer in each segment will
apply to all consumers in that group. Nevertheless, the characteristics of most of
the travelers falling into each segment can be described as follows:
4. Type 4: Insensitive to both price and time constraints. This group includes the
relatively few consumers that have little or no time constraints for travel, yet
are willing to pay for high levels of service, including the flexibility of flying
with little advance notice on preferred flights.
Figure 1.3: Market Demand Segmentation Model
PASSENGER:
PRICE-SENSITIVITY
LOW ( ) HIGH
- -
I- ~-
-
TRIP: HIGH I -
-- 0-- -- --- -mM
i
TYPE 1 II TYPE 2
I I I
TIME- - ---
I - - - - I
SENSITIVITY
It I
TYPE 4 11 TYPE 3
I
LOW ISm
w....U.0.0
For demand segmentation purposes, Type 4 consumers can be combined with the
Type 1 group, since both are willing to purchase a high-priced fare product to secure
a high level of service and/or travel flexibility, regardless of their trip purpose.
This market demand segmentation model is still broad in its definitions, and
many airlines currently have in place fare structures with more than three types of
product, as will be described. Nevertheless, the model includes all consumer and
trip characteristics and as such can be used to classify smaller demand segments
more precisely, if necessary.
Airlines abandoned the original two-segment market demand model due in part
to a realization that the characteristics of consumer demand for air travel had
changed. At the same time that defining demand segments was becoming more
difficult, airlines recognized that differential pricing strategies targeted at multiple
demand segments offered the potential of even greater increases in total flight rev-
enues. The greater the number of segments that can be identified and differentiated
fare products that can be offered for sale, the closer airlines can come to the ideal-
istic model of a different price for each seat sold. In spite of the practical problems
of identifying distinct demand segments, most airlines instituted multiple-tier fare
structures.
The problem of identifying market demand segments and the obstacles to ef-
fective differential pricing practices, as discussed, lead to a third way in which the
original surplus seat concept has changed. With an almost universal availability
and usage of low-priced fare products, the term "surplus seats" may no longer be
appropriate in the context of seat inventory control. The original surplus seat con-
cept considered low-fare seats to be a by-product of the airline's operations, which
had traditionally been devoted to serving the full-fare passenger.
The high proportion of passengers now carried at reduced fares indicates that
the original surplus seat concept has been eroded severely. Airlines appear to be
providing capacity for low-fare as well as full-fare passengers, in which case the
low-fare products are a part of the airline's primary output. If this is in fact the
case, economic efficiency would require the low-fare products to contribute in a
substantial amount to airline capacity (fixed) costs, in addition to covering their
marginal variable costs. The inability of airlines to segment market demand precisely
or to charge radically different fares for different products in many markets suggests
that such a contribution is being made.
While the differential pricing concepts and demand analysis techniques of the
original surplus seat management model are still valid, their application in the cur-
rent airline marketplace has become substantially more complex. Rather than es-
timating the demand for a primary product and allocating all remaining capacity
to a marginal product, airlines must recognize that each of the several products
they offer on a flight makes a revenue contribution to fixed costs. A seat inventory
management system in today's airline environment must estimate the demand for
multiple fare products and then allocate seats to each on the basis of the relative
revenue contribution of each and the likelihood that seats will ultimately go empty.
This approach underlies the mathematical models presented in Part Two of this
dissertation.
Airlines wishing to segment the total market demand for air travel to their
revenue-maximizing advantage must design a range of fare "products" that will
appeal to, and be used exclusively by, each of the demand segments. An ideal fare
structure will minimize any "seepage" between segments, particularly the diversion
of those willing to pay higher fares to the lower-priced fare products. Airlines
use restrictions on reduced fares as disincentives to prevent downward diversion of
higher fare passengers, and increased service amenities as an incentive for consumers
to purchase higher-priced fare products. In fact, the absence of purchase or travel
restrictions has come to be regarded as a service amenity of full fare products.
These restrictions on the purchase and use of low-priced fare products are often
accompanied by capacity controls or limits on the number of seats available to
particular fares or types of fares. The specific limits on low-fare seat availability
are not known to consumers, and as such are not explicit attributes of different
low-priced fare products as perceived a prioriby the consumer. Rather, it is the set
of conditions and service amenities associated with a fare product that determines
the market demand segment to which it will appeal.
Advance purchase and minimum stay requirements are designed to keep Type 1
and 2 consumers from making use of the lowest fares. Since the first super-savers
were introduced with a 7-day advance purchase requirement, advance purchase peri-
ods have been lengthened gradually, culminating with the iutroduction by American
Airlines in 1985 of the "Ultimate Super-Saver", which requires a 30-day advance pur-
chase. Over the same period, the minimum stay conditions have been reduced to
the point that most excursion-type fares now require only a Saturday night away
from the originating point before return travel may commence. This requirement
replaced previous 7-day minimum stay conditions because airlines found that the
reduction led to little or no increase in diversion of high-fare travelers. At the same
time, it increased accessibility of the low-fare product to price-sensitive (Type 3)
consumers able to stay away over a weekend but not for a full week [18].
The three traditional fences - advance purchase, round-trip travel and minimum
stay - are effective against diversion only if they are enforced by the airline. While
it might be possible for consumers to side-step an advance purchase requirement by
dealing with a travel agent who is willing to backdate ticket issuance transactions,
the minimum stay rule is also susceptible to abuse by travelers. Many airlines allow
a passenger to stand by for the return portion of a round-trip reduced fare, and
some passengers will routinely stand by for the inevitable empty seat on return
flights, never having had the intention of meeting the fare's original minimum stay
conditions. Although it is up to the airline ticket and gate agents to prevent such
abuse, few agents will deny a passenger with a ticket as long as empty seats remain on
a departing flight, especially if the passenger was ticketed originally for a competing
carrier's flight.
The abuse of both the minimum stay and advance purchase requirements, pri-
marily by Type 1 and 2 consumers, prompted airlines to introduce monetary penal-
ties for changing or cancelling flight bookings for the lowest fares. The objective of
these penalties is to prevent those passengers who are unable to commit to exact de-
parture and return flights far in advance from purchasing several advance purchase
low-fare tickets, using the one that proves to be most convenient, and then obtain-
ing a full refund for the unused tickets. The imposition of a monetary penalty also
makes it more difficult for passengers to confirm reservations for an alternate return
flight without being asked for an additional fare payment by the airline, although
it is still possible to stand by without penalty in many case&.
A price-sensitive consumer planning a trip with little time sensitivity will likely
accept all three of these constraints on his/her travel itinerary in order to obtain
the lowest possible fare. The degree to which a time-sensitive and price-sensitive
(Type 2) consumer is able to conform will depend on how price-sensitive he/she is,
and the degree to which the restrictions impinge on the desired travel times. Survey
findings indicate that, apart from advance purchase conditions, restrictions limiting
the availability of low-fare products to particular travel days and/or times are the
most effective market demand segmentation technique [201. In addition to perform-
ing a demand segmentation function, time of travel restrictions also help the airline
to fulfill another goal of the original differential pricing strategy - spreading peak
period demand to fill otherwise empty seats on less popular flights. The availability
of the lowest 30-day advance purchase fares for travel on Tuesdays and Wednesdays,
for example, helped to reduce day-of-week variation in total demand tremendously
[21].
Business class options are offered in almost all international markets by most
airlines, where long-haul flights make more legroom and a higher level of inflight
service very attractive to those already paying a full coach fare. Trans World Air-
lines offers its Ambassador Class product domestically as well, on its wide-bodied
aircraft. Air Canada has recently introduced a similar Executive Class on long-haul
domestic and Canada-U.S. routes. Both carriers price their business class products
just slightly above the full coach fare, such that the premium is generally less than 10
percent. In many respects, these business class products have replaced the standard
full-fare coach products designed originally for business travelers.
Several carriers have also experimented with providing full-fare coach passen-
gers with added service amenities, ranging from separate check-in counters and
VIP lounge privileges at airports to complimentary ground transportation services.
United Airlines at one point guaranteed that a full fare passenger would not be de-
nied boarding due to overbooking and, more recently, stopped giving advance seat
assignments to passengers with the lowest discount fares. While U.S. carriers have
met with resistance from passengers and travel agents in their attempts to enhance
the full coach fare product, Canadian carriers have been doing just that for years.
Air Canada's Connaisseur Class and CP Air's Empress Class are services to which
only full coach fare passengers are entitled. Both carriers seat full-fare passengers
in a separate area of the economy cabin and provide them with amenities like free
drinks and movies, first choice of meals, and advance seat assignment.
The use of an increasing variety of low-fare restrictions and the inclusion of ser-
vice amenities with full fare tickets represent attempts by the airlines to differentiate
the range of fare products they offer, particularly those that will share seating in
the coach compartment of the aircraft. Significant product differentiation is a major
determinant of effective demand segmentation, which in turn allows the airline to
practice differential pricing to maximize revenue. If the fare products offered by an
airline are differentiated to the point that consumers understand why a reduced fare
product is priced far lower that the full fare product, the airline's fare structure is
entirely rational.
Unfortunately, this type of ideal fare structure can be found in very few markets.
If one is found, it is unlikely to remain stable for any length of time. With the growth
of hub-oriented route systems, most domestic markets are now served by several
airlines, and the choice of carriers in a market often includes at least one low-fare,
single-product competitor. Even if the price leader does not offer a level of service in
a particular market, experience has shown that most of the competing carriers will
match the low price with a similar product. Furthermore, sporadic "fare wars" and
regional price battles serve to undermine the possibility of offering a standardized
differentiated range of fare products in all markets. It is under such circumstances
that airlines and consumers alike marvel at the complexity and incomprehensibility
of airline industry pricing practices.
First class, business class and full coach fare products are targeted at the Type
1 consumer, and are differentiated only by their service amenities and relative price
levels. The most important attribute of all three of these fare products is their lack
of restrictions. A passenger may purchase a ticket at any time prior to departure and
make virtually unlimited changes to both the outbound and return flight itineraries,
all without paying a penalty. At any time, unsold seats will always be available to
passengers purchasing these unrestricted products.
Super coach fares and excursion fares are moderately priced products targeted
at the Type 2 consumer. Super coach fares appear primarily in markets where a low-
fare competitor offers unrestricted fares below the established coach fare. Depending
on the extent to which a carrier wants to respond to the low-fare competitor, it might
attach a minimal (3-day) advance purchase requirement or a one-stop/connecting
flight limitation to its super coach fares to prevent Type 1 consumers from diverting
to this lower fare. Most of the established carriers place capacity controls on their
super coach fares to keep some seats available for full coach'fare passengers right up
to departure time.
Excursion fares as defined in Figure 1.4 are essentially the traditional round-trip
reduced discount fare products introduced in the mid-1970's. Originally targeted at
Figure 1.4: Typology of Airline Fare Products, Spring 1987
SUPER COACH Type 2 0-3 day advance purchase; Coach cabin service. 70-85
FARES 1-stop or connecting flights;
limited seat availability.
EXCURSION Type 2 7-14 day advance purchase; Sat. Coach cabin service. 50-70
FARES , night minimum stay; round-trip
purchase required.
SUPER SAVER Type 3 21-30 day advance; Sat, night Coach cabin service. 30-50
FARES minimum stay; cancel/change
penalties.
"MAX" SAVER Type 3 2-day advance purchase; Sat. Coach cabin service. 20-30
FARES night minimum stay; non-
refundable; very limited seats.
leisure travelers, these fares are being used more often by business travelers willing
to extend a business trip over a weekend or to combine business and pleasure travel
to obtain a lower fare.
The lowest-priced fare product category includes the "MaxSaver" fares intro-
duced recently and available at least through May 1987 on all major U.S. carriers.
These fare products are intended for Type 3 consumers, as they require ticket pur-
chase within 24 hours after a reservation is confirmed and are totally non-refundable
once purchased. The advance purchase requirement, however, is only 2 days prior
to departure, meaning that price-sensitive Type 2 consumers will try to obtain these
fare products just days before a trip. The Saturday night minimum stay require-
ment will not be a deterrent to diversion in the many markets where the round-trip
"MaxSaver" is priced below the lowest available one-way fare. The advance purchase
requirement for these fare products is to be increased to 7 days for travel during the
peak summer period.
Each of the product categories presented in Figure 1.4 can in fact contain several
specific fare basis codes , or published fares, with their own set of rules, effective
dates, and price levels, all in the same O-D market. For example, a "QE30X23S"
fare basis code might be defined by an airline to describe a specific super-saver ("Q")
excursion ("E") fare product that requires a 30-day advance purchase ("30") and
applies for travel on Thursdays through Mondays ("X23"), but only on one-stop
or connecting flights ("S"). A "QWE21" fare basis might be available in the same
market at the same time, describing a different fare product that requires a 21-day
advance purchase, weekend travel, and is available on all flights (subject to capacity
controls). The coding of fare bases is entirely at the discretion of creative minds in
airline pricing departments, although imitation has led to some degree of uniformity
of fare basis codes among airlines and across markets.
Fare bases are grouped by the airline into fare classesfor the purpose of Accepting
and controlling bookings in its reservations system. Ideally, the reservations system
fare classes would correspond exactly to the fare product categories described above,
such that distinct seat inventories could be made available for each category. Most
airline reservations systems, however, are limited to five primary fare classes for
any one flight. The implications of this limitation for seat inventory control are
discussed in Chapter Three. At this point, it suffices to note that airlines offering
more fare product categories than they have reservations system fare classes must
combine more than one product category into one fare class.
All the fare basis codes that comprise a particular fare class begin with the same
letter, by convention. As a historical rule, separate physical compartments on an
aircraft have been assigned their own fare class. Thus, a carrier offering first and
business classes in addition to coach uses up three of five fare classes on Type 1 fare
products. The letter codes used for each fare class can differ by carrier, although "F"
for first, "C" for business and "Y" for full coach fare are industry standards. The
lowest-priced fare products are commonly in Q-class, while all intermediately priced
fare products might be listed as M, B, K, V, S, L, or H-class fares. Because not all
fare products and fare classes are associated with distinct physical compartments on
board the aircraft, passengers booked in two or more fare classes can in fact receive
the same class of service in flight. On most carriers, all passengers paying full coach
fare and less travel in the coach cabin of the aircraft and receive identical on-board
service.
For the consumer wishing to make a reservation for a future flight, the availability
of a specific fare product on the preferred flight departure depends on the availability
of a seat in the associated reservations fare class inventory. If the fare class has been
closed down to further reservations, the consumer is then faced with the option of
accepting a different fare product in another fare class on the same flight or making
inquiries into the availability of the same or similar fare product on other flights
and/or carriers. This decision process is discussed in detail in the following chapter.
Figure 1.5: Fare Structure Relationships
FARE BASIS
CODES
F C
CN
Y
YN
M
Ml
BE70
BR
QE30
QR14I
QE2NR
QE2X57
YCH MX67 BEl4 QWE30I QE2Z57
YD MAP3 BAP7 QXE30
FARE C
CLASSES
This chapter explores the consumer decision process that has evolved with the
institution of multiple-tier pricing strategies by airlines. With reference to classical
micro-economic utility theory, it presents a conceptual model of how individuals
choose from among the available fare products, alternative flights and competing
carriers in an air transportation market. This choice process for individual con-
sumers is then used to develop a description of the aggregate demand for a fare
product on a particular flight, as viewed by the airline through its reservations sys-
tem. The patterns of reservations demand described by the aggregate model are
those that will influence the airline's ability to forecast demand and allocate seats
to different fare products in a revenue-maximizing manner.
A plausible model of individual choice for air travel decisions is one that is based
on a valid theory of consumer decision-making and one which can be adapted to the
context of transportation decisions in general and air travel options in particular.
This section begins with a brief overview of the decision theories that seem to be
best suited to the air travel choice problem, and then extends these theories into a
qualitative choice model that deals explicitly with the selection of fare products and
flights.
The "price" of a travel option is therefore not simply the monetary cost of the
ticket. Various travel options can involve additional disutilities and monetary costs
associated with the timing of the travel components of the trip, as well as the value
of the time actually spent traveling. Given the derived nature of transportation
demand, consumers will want to minimize travel time, cost, discomfort and incon-
venience for a trip with a given level of perceived benefits. The concept of utility
thus represents a generalized function that takes into account the pleasant and un-
pleasant components of making a trip and can form the basis of consumer choice
[29].
A descriptive model of individual choice for the air travel decision process must
take into account not only the availability of heterogenous alternatives, but the
fact that individual decision-makers will have different choice sets and can assign
different utility values to the same attributes of the same alternative. The choice set
for each individual will include the alternatives that are feasible to him/her, given
the constraints on trip timing and costs. This choice set might only include a few
alternatives if the consumer is unable or unwilling to obtain information about more
options, or if the first option considered is found to be acceptable.
The air travel choice process for an individual is triggered by a perceived benefit
of being at a particular destination to participate in activities there. The individual
evaluates the travel options available for a given trip, and the costs or disutilities of
each. If the benefits of being at the destination exceed the disutility of one or more
travel alternatives, the alternative that minimizes disutility will be chosen. If the
travel disutility of all alternatives exceeds the perceived benefits at the destination,
the individual will choose not to make the trip.
Each fare product is defined by its attributes, which contribute to the total
disutility of that fare product. Negative attributes create added inconvenience for
the traveler, resulting in real or perceived costs that are higher than the dollar price
paid for the fare product. The relevant attributes of each fare product may be
grouped as follows:
2. Service Amenities, (S;), the positive attributes related to the quality of service
received in conjunction with the purchase of F;. On-board service amenities
such a meal quality, complimentary cocktails and headsets, and preferred seat
assignments can differ among fare products, as can airport amenities like sepa-
rate check-in facilities and VIP lounge privileges. Currently, the fare products
offered by many domestic U.S. airlines in a shared economy cabin involve few
distinctions in service amenities.
3. Restrictions, (X;), the conditions associated with F;, including required ad-
vance booking and ticket purchase, cancellation and refundability limitations,
effective and discontinued dates, minimum and maximum stay requirements,
and applicability to particular times, days of the week and/or flight itineraries.
Fare products priced lower than the standard full fare generally have one or
more of these restrictions, the severity of which increases as the price decreases.
The total disutility of each fare product under consideration can be represented
by the notion of its "generalized cost", denoted Z(F) for fare product Fi. The gen-
eralized cost of each fare product to the consumer will be a function of its attributes,
for example:
Z(Fi) = Z(Pi,S,Xi) (2.1)
Each passenger will place different values or utilities on the relative importance of
each of these attributes in deriving the generalized cost of each fare product option,
depending on his or her set of values, as determined by both the price-sensitivity
of the individual and the time-sensitivity of the trip being considered. The relative
values of Z(F), Z(F2 ), and Z(Fs) may be different, requiring each individual to
evaluate and rank the alternatives with respect to generalized cost, in order to
identify the fare product that minimizes perceived disutility.
This simple example is unrealistic in that it assumes that all fare products for
the passenger's first choice of departing and/or return flights are available. The
choice model must therefore be expanded to incorporate the added generalized cost
and reduced utility associated with the consumer's not being able to obtain the
preferred fare product on his/her first choice of flights. There will be separate costs
and disutilities associated with displacement from the preferred outbound and return
flights and times. Both the outbound and return flights are included because most
air travel decisions involve round-trip travel. The benefits of being at a particular
destination must be weighed against the disutility of both outbound and return
travel.
We assume that each consumer has determined the time period over which being
at the desired destination and/or away from the origin will generate a benefit. On
the basis of this time period, a "time window" of departure and return times can
be deduced. Within each time window, there is some "ideal" set of departure times
for the outbound and return flights, independent of any knowledge of actual flight
schedules or seat availability.
The consumer begins the process of obtaining information (from a travel agent
or from the airlines directly) about the actual alternatives that might be available
for a particular trip. The specific way in which this choice set is developed will
depend on the consumer's approach to selecting particular flights. For example,
an infrequent air traveler might contact a travel agent to determine which flights
and fare products are available. A more experienced air traveler might establish
a choice set of preferred flight alternatives from published airline schedules before
determining the availability of fare products on each.
Given a range of outbound and return flight itineraries offered by competing air-
lines in a market, there is a large number of air travel alternatives, T = {F, D, R},
that can be constructed from the components F, D, and R. In terms of individual
choice, however, all but a relatively small subset of these combinations will be elim-
inated by the timing constraints established by the consumer for a particular trip.
These constraints are determined by the "time window" established for activities at
the destination, over which the traveler will be able to realize a benefit, as well as
by any time constraints on being away from the origin.
For each individual, then, the generalized cost of each travel alternative is a
function not only of Z(F;), but of the disutility of the D and R components of the
alternative:
Z(T) = Z{F, D, R} (2.3)
The generalized cost of the F component includes the dollar price, Pi, of the fare
product, and the disutility of meeting all of the restrictions, Xi, of that fare product,
perhaps adjusted by the utility of any differences in service amenities, S;, associated
with the fare product. The generalized costs associated with both the D and R
components include the dollar costs and disutilities of displacement from the "ideal"
flight departure times, ltd - td, and of enroute time, t,. These costs can include
monetary expenditures in addition to the value of travel time, for example, increased
hotel or meal expenditures.
The critical assumption of the choice model developed here is that individual
passengers implicitly (or perhaps explicitly) rank the set of feasible air travel pack-
ages in their choice set with respect to the relative disutility of each. As mentioned
earlier, an individual might not consider a full range of alternatives and is likely
to seriously consider only the current option with the lowest disutility. The choice
process can therefore be iterative, in that the consumer can be induced to expand
his/her choice set as information about the (un)availability of preferred travel alter-
natives is gathered. The availability of a particular D or R component depends on
the airline having seats to sell over the flight itinerary in question, in the fare class
inventory from which fare product F must be sold. The ultimate choice is made
by the consumer so as to minimize the total disutility of the travel required for a
particular trip, aubject to the availability of each successively less attractive travel
alternative.
The first six travel alternatives in this hypothetical consumer's ordered choice
set might look like this:
We can infer from this ordered choice set that, for this hypothetical passen-
ger, the marginal disutility of accepting the second and third-ranked return flight
itineraries is smaller than the marginal disutility of accepting the second choice of
outbound flight itineraries. Only after determining that D2 is not available at F1
will this consumer accept what is most likely a higher-priced F2 , but only on the first
choice outbound and return flight itineraries. T6 illustrates how a shift in the fare
product component of the air travel alternative might rule out one or more flight
itineraries (in this case R 2 ). The restrictions associated with F2 might eliminate R2
as a component of a feasible travel alternative.
Given that one or more alternatives might not be available because all F seats
have been sold on the D or R components of the trip alternatives in question, the
choice process is simply one of systematically checking each ranked travel alternative
in ascending order of disutility until a feasible and available combination is found.
The extent to which an individual is willing to accept a lower-ranked travel alter-
native will be a function of the generalized cost of that alternative relative to the
perceived benefits of being at the destination point. If we define the total of these
perceived benefits to be W, then the individual will make the trip as long as the
generalized cost of the highest-ranked feasible and available travel alternative, T*,
is less than W. That is, the trip will be made only if:
If more than one feasible and available travel alternative has a generalized cost less
than W, then the individual will choose T* so as to maximize W - Z(T*).
The value of W for a particular trip thus establishes a "cut-off point" in the
set of ranked alternatives, below which an alternative will not be accepted by the
consumer even if it is available. At this point the disutility of the travel alternative,
due to displacement from preferred flight times and/or much higher monetary costs,
is so great that the trip itself becomes of no value to the individual. The location
of this cut-off point on the ranked list of alternatives will be determined by the
characteristics of the trip and of the consumer, as defined by the market demand
segmentation criteria described in Chapter One. That is, each Z(T*) is a function
of the time-sensitivity of the trip in question and of the price-sensitivity of the
individual with respect to that trip.
The criteria that define the market demand segments also affect the willingness
of the individual to accept lower-ranked trip alternatives and the disutilities of shifts
in the F relative to the D and R components. For example, a consumer wishing
to take a trip for business purposes to attend a meeting at a specified time and
who cannot depart before a certain time nor return after a certain time will place
a greater disutility on accepting lower-ranked flight itineraries than lower-ranked
fare products. In the demand segmentation model, this trip would be classed as
time-sensitive, and the consumer would be relatively insensitive to price.
This formulation allows for the possibility that similar fare products offered by com-
peting carriers in the same market will have different price levels and/or different
restrictions, in addition to any perceived differences in service amenities. A prefer-
ence for a particular carrier on the basis of "brand loyalty" or frequent flyer program
considerations can also be reflected in different utilities assigned to Sik for the same
fare product offered by different carriers.
If the values of Z(Fik) are determined to be different for the same fare product
offered by different carriers, the consumer would then select the fare product/carrier
combination that minimizes Z(Fk), with all else being equal. It is more likely that
distinguishing between identical fare products offered by competing carriers will
involve different D and R components in the travel alternatives being considered.
The consumer must then take into account the relative disutilities of the different
flight itineraries on different airlines. It is possible that increased disutility of the D
and R components involving the preferred airline will be outweighed by the lower
value of Z(Fik) relative to that of other airlines. That is, the utility of traveling on
the preferred carrier can exceed the disutility of less convenient flight itineraries.
The individual choice process for air travel described in this section relies heavily
on the notions of relative utilities and rational consumer behavior. It is important to
emphasize, however, that very few potential air travelers will undergo an exhaustive
enumeration and evaluation of alternatives before making a completely informed
utility-maximizing choice. As mentioned, the number of alternatives considered
and the way in which information is gathered can lead to the consumer making a
less than optimal choice. Nonetheless, the choice process described remains valid
insofar as the consumer will select the trip package alternative providing the lowest
perceived disutility from among those considered, given feasibility and availability of
that alternative.
If all fare products were always available, aggregate demand for each fare class
on each future flight leg to be operated by an airline would be the sum of all requests
from potential passengers for whom the relevant travel alternative {F, D, R} ranks
first in perceived utility. In reality, there is a limited number of seats available for
each alternative in the choice set. Requests for different alternatives arrive over
time before flight departure. If either the D or R components of an individual's
preferred travel alternative are unavailable in conjunction with the desired F at the
time the request is made, the potential passenger will be forced to consider and
perhaps accept successively less desirable alternatives from his/her ordered choice
set. For the airline, each shift by a potential passenger from one travel alternative
to the next has important implications with respect to total bookings and expected
revenues for future flights.
The airline receives a request from a potential passenger for a particular out-
bound flight (D), return flight (R), and fare product (F). If the passenger's pre-
ferred travel alternative is not available in its entirety, the passenger consider the
next-best alternative. This next-best choice will include one or more components
different from those in the initial alternative requested. Depending on which of these
components change in the shift to a less desirable alternative, the impacts for the
airline will differ.
For the airline's purposes of seat inventory control, it is the upward shift in the
fare product component associated with the next-best alternative in the consumer's
choice set that is of most interest. Given the objective of maximizing total flight
revenues, the airline must consider the possibility that a passenger denied a reser-
vation in a low fare class will be willing to step up to the next highest available
class for the same D and R flight itineraries. Stepping up to the next highest fare
class represents what we shall call a vertical shift or "upgrade" in the reservations
request, from the airline's perspective. Such a shift is desirable for the airline, since
no potential traffic is lost and total revenues actually increase.
Unfortunately, not all potential passengers will be willing to make this verti-
cal shift when denied a reservations request, depending on the composition of the
next-best travel alternative in their ordered choice sets. Those not willing to ac-
cept a vertical shift might be willing to accept different departure and/or return
flight components in their trip packages but with the same fare product component
as the one requested initially. This will be the case for extremely price-sensitive
passengers (Type 3) in particular. For the airline denying a reservations request,
then, the probability that the denied passenger will make a vertical shift in his/her
choice process is directly affected by the passenger's price-sensitivity and changes in
the utility of the fare product components between the first-ranked and next-best
alternatives.
If a passenger's request for a particular travel alternative is denied, the next best
choice might involve the same fare product component, but a different departure
and/or return flight itinerary. Acceptance of the next-best alternative in such a case
can involve a shift of flights and/or carriers. If the passenger selects the same fare
product for another flight on the same airline, the airline retains this passenger's
business and revenue. Furthermore, the airline may sell seats that would otherwise
have gone unsold on lower demand flights. This type of movement in the passenger's
choice process can be termed a horizontal shift in preference, or "displacement" of
the passenger from the originally requested flights.
1. a vertical shift, v, ("upgrade") from one fare class to another, on the same
outbound and return flights as those requested initially;
Mathematically, the denial of a reservations request by the airline will elicit one
of the three consumer responses, v, h, or 1, with various probabilities, such that:
For the airline deciding whether to accept or deny a particular request for a
fare class and flight, the magnitudes of these probabilities are extremely important,
especially P(l). Recapture eliminates the losses associated with refused requests.
American Airlines has developed a model of reservations recapture and has made
attempts to measure it. Their theoretical model assumed that "recapture can be
estimated by redistributing demand from a closed or cancelled flight based on pas-
senger preference models."[31]
The problem with these observations concerning recapture rate, apart from the
small sample size and the drawbacks associated with inferring passenger behavior
from a telephone transaction, is they do not recognize that recapture rates will differ
depending on the market demand segment characteristics of the passenger involved.
The willingness of a potential passenger to shift vertically or horizontally from the
travel alternative requested initially will depend on the passenger's price sensitivity,
the time sensitivity associated with the benefits of the trip, and on the specific point
in the passenger's ordered choice set at which the reservations request is denied.
We can extend the individual choice model presented in Section 2.1 to incorpo-
rate the choice shifts that occur when a preferred travel alternative is not available.
The objective is to "explain" the relative probabilities of each type of choice shift
described earlier, with reference to the travel disutilities of the attributes of different
travel alternatives, as evaluated by the consumer. The choice shift that an individ-
ual makes will be determined by the specific alternative(s) being considered in place
of the unavailable alternative and the disutility associated with the change in one
or more components of the travel alternative. The expected choice shift is therefore
a function of the time constraints on the trip involved and of the price-sensitivity of
the consumer for this trip, which in turn define the demand segment to which the
consumer belongs over the course of the choice process for the trip at hand.
For any one denied reservations request, the choice shift that will be made by the
consumer depends on the composition of the "next-best" travel alternative. This
next-best alternative may be the one that offers the smallest increase in travel disu-
tility over the unavailable alternative, as derived from a ranked set of alternatives
by a completely informed consumer. In practical terms, the "next-best" alterna-
tive is often an alternative selected by either the airline reservations agent or a
travel agent as being potentially acceptable to the consumer. In many cases, the
consumer's ordered choice set will consist of as few as one alternative presented to
him/her as being the "next best" option. If the consumer believes an alternative
with a lower disutility than that of the alternative being presented might be avail-
able, the suggested "next best" option will be refused. The consumer will then call
another airline, or deal with another travel agency.
We examine first the simple case in which the consumer deals directly with
only one airline, and assume that no competing airlines are involved in the choice
process. When this consumer is denied a reservations request for a preferred travel
alternative, the airline will offer one or more "next-best" alternatives, different from
the preferred alternative in the fare product component and/or one or both flight
itinerary components. Given a next-best travel alternative which differs from the
initially requested alternative only in terms of the fare product component, the
consumer will accept this alternative only if the disutility of the change in fare
product components is less than the increased disutility associated with all other
known next-best alternatives. If no other alternatives are being considered by the
consumer, or if no other alternatives are available, the next-best alternative will be
accepted as long as the increased disutility of the new fare product component does
not increase the total disutility of travel to the point that it exceeds the perceived
benefits of the trip.
With reference to the notation introduced previously, a shift in the fare product
component of a travel alternative from F1 to F2 on the same airline for identical
flight itineraries will involve an increase in travel disutility for the consumer equal
to:
Z(F 2 ) - Z(F) (2.9)
This increase in disutility will be a function of the price difference between the two
fare products, any changes in the restrictions involved, as well as any perceived or
actual changes in service amenities:
Given only the option of accepting the less desirable fare product component or not
making the trip, the consumer will make a vertical choice shift ("upgrade") if:
where Ti is the initially requested (first choice) travel alternative and W represents
the total perceived benefits of making the trip.
As was the case with an isolated vertical shift, the consumer presented with only
the option of making the specified horizontal shift or not traveling will accept the
changes in the D and R components of the travel alternatives as long as:
Note that a horizontal shift can involve a change in either D or R, or both. Even if
the same R component is included in both alternatives, for example, the disutility
of the R component can change when it is paired with a different D component.
The disutility of a travel itinerary is thus expressed as a non-separable function of
both flight itinerary components.
In the absence of any information about travel alternatives involving other air-
lines, the consumer will accept a vertical choice shift from fare product F1 to F2
rather than a horizontal shift from (D1 , R 1 ) to (D 2 , R2) if:
A horizontal shift will be preferred if the relative disutilities in the above inequa-
tion are reversed. Equality of the two changes in travel disutility would indicate
indifference between the two options.
The market demand segmentation criteria of price and time sensitivity can also
be used to postulate relationships between the probabilities of vertical and horizon-
tal choice shifts when consumers belonging to different segments are presented with
the same "next-best" travel alternatives. Given the price and time sensitivity char-
acteristics associated with Type 1, 2, and 3 consumers, respectively, the following
relationships would be expected:
Although horizontal and vertical shifts enable the airline to retain, or even in-
crease, the revenue it receives from a denied passenger, airlines in competitive mar-
kets must be more concerned with P(l), the probability that a denied request will
cause the consumer to take a seat on a competing carrier, or not travel at all. The
complexity of the choice shift process increases dramatically as we introduce the
possibility of booking loss, which includes "defections" to competing carriers.
First, we must distinguish between consumers who choose to determine the avail-
ability of their initial or preferred alternative by dealing directly with an airline and
those who deal with a travel agency. When dealing with a travel agency, the con-
sumer still evaluates the next-best alternatives offered on the basis of minimizing
increases in travel disutility, as described earlier. In this case, however, the next-best
alternatives presented to the consumer are more likely to involve different airlines or
even combinations of airlines. Furthermore, these next-best alternatives will gener-
ally involve lesser increases in disutility, both in terms of the fare product and flight
itinerary components, than those presented by a single airline.
The same conditions lead to the conclusion that P(Q) is very difficult to measure.
The probability of booking loss is affected by whether the request is being denied by
the airline or by a travel agent. All else being equal, there must be some reduction
in P(l) in the former case, since the airline reservations agent can encourage the
consumer to accept a vertical or horizontal shift. This is one reason for the preference
of most airlines that passengers book directly with them (the savings in travel agent
commissions being another).
Like the probabilities of horizontal and vertical choice shifts, the probability of
booking loss is determined by the specific attributes of the changed components of
each pair of travel alternatives being considered. The travel alternatives involved are
in turn affected by the source of information on availability (airline or travel agent),
as described above. The probability of booking loss will also be a function of how
many of the consumer's successive requests have been denied already, particularly
when dealing directly with the carrier. Generally, each less desirable travel alterna-
tive offered by the same airline will represent a greater increase in travel disutility
than would be the case when alternatives involving several carriers are involved. All
else being equal, the consumer denied several requests by the same airline is likely
to try another airline.
The mathematical formulations of the utilities relevant to consumer choice when
a preferred travel alternative is unavailable can be extended and generalized to ac-
count for most, if not all, possible "next-best" alternative scenarios. The utility
valuations and decision rules will be similar to those presented earlier for the sim-
plified one-airline case, since the consumer will be making the same decisions, but
with more information about travel options.
For each travel alternative TA of which the consumer is aware, the increase in
travel disutility over that of the initially request but unavailable alternative is given
by:
Z(TA) - Z(T*) = Z(FA, DA, RA) - Z(F*, D, R*) (2.18)
As described earlier, the disutility of different fare products will be related primarily
to the price-sensitivity of the consumer for a particular trip. The disutility associated
with any changes to the fare product restrictions and/or service amenities will also
have an effect, as will consumer preference for a particular carrier. The disutilities
associated with different D and R components of travel alternatives will be primarily
time-related.
The market demand segment to which a consumer belongs for a particular trip
can therefore be related to the consumer's expected choice shift behavior in the case
of unavailability of one or more requested alternatives. Market segment criteria
alone, however, may not be sufficient in predicting choice shift behavior, because
the "choice set" with which the consumer is presented or of which he/she is aware
will differ from one case to the next and have an impact on the ultimate choice shift
made.
Passenger choice shift behavior can differ not only from one individual to another,
but for the same individual planning different trips, and even for the same individual
and trip at different points in the process of determining the availability of various
travel alternatives. The probabilities of recapture and booking loss represent an
average rate or expectation that must be estimated from an analysis of actual choice
shifts. The difficulty lies not only in the survey methods required to sample intended
behavior, but also in the need to identify similar consumers who are planning similar
trips. Further theoretical consideration of the consumer choice process is required,
along with much more extensive empirical studies of choice shifts.
Figure 2.1: Consumer Choice Shift Process
INITIAL CONSUMER
REQUEST
(Preferred Travel Alternative)
AVAILABLE UNAVAILABLE
REQUEST
DENIED
1
P(v) ' P(h) P(l)
I
PASSENGER "RECAPTURE"
RR = P(v) + P(h)
"NEXT-BEST"
ALTERNATIVE
"NEXT-BEST"
ALTERNATIVE "
""n
AVAILABLE UNAVAILABLE
The choice shift process described here can be summarized in a reservations
decision tree framework, as shown in Figure 2.1. A request for a particular travel
alternative can either be accepted or refused by the airline, based on the availability
of the requested fare product class and flight itinerary. If accepted, a booking is
confirmed for the consumer. If denied, there is a probability that the consumer will
be recapturedby accepting either a vertical or horizontal shift in travel alternatives,
RR = P(v) + P(h). There is also a probability, P(I) = 1 - RR, that the denied
passenger will represent a booking lose to the airline.
To this point, our airline reservations framework has focussed on denied requests
stemming from unavailable travel alternatives. The assumption made implicitly is
that each accepted request, or booking, will result in a ticket sale and, ultimately,
a revenue passenger carried on a future flight. In reality, an accepted reservations
request does represent a booking that decreases the seat inventory allocated to a
fare class and flight. Each booking, however, will translate into a filled seat that
generates revenue for the airline only if the passenger actually purchases a ticket
and shows up for the booked flight.
Our model of the airline reservations process should therefore incorporate the
probabilities associated with each booking becoming a revenue passenger on a future
flight. The probabilities of relevance are:
2. the probability that a booked passenger will fail to appear for a scheduled
flight departure (i.e., be a "no-show").
Both of these probabilities will be affected by whether the passenger has pur-
chased a ticket subsequent to having a booking confirmed for a future flight. A
passenger that accepts a "next-best" alternative suggested by one airline may still
call another airline to make another booking, and ultimately buy a ticket and travel
on the competing carrier. As was the case with choice shift probabilities, these
probabilities can also be affected by the demand segment of the passengers and the
restrictions associated with the fare products booked.
The length of time before departure that a booking is made, however, is no longer
the only determinant of that booking's cumulative cancellation probability. The
imposition of advance purchase requirements and cancellation penalties after ticket
purchase on certain low-priced fare products has introduced additional elements
that can affect cancellation probabilities. Furthermore, many airlines now stipulate
time limits before which either tickets must be purchased or the booking must
be reconfirmed, even for "unrestricted" fare products. If the ticketing time limit,
be it arbitrary or part of the fare product's restrictions, is not met, the airline's
reservations computer will cancel the booking automatically.
We can extend the decision tree of Figure 2.1 past the point at which a book-
ing is confirmed to examine cancellation and no-show probabilities under different
scenarios. As shown in Figure 2.2, the first major determinant of the cumulative
cancellation probability for a booking is the presence or absence of a ticketing time
limit. In either case, ticket purchase leads to the same set of subsequent branches
in the decision tree, and the same set of probabilities. A failure to purchase a ticket
for the booking, however, leads to distinct results in the two cases. Given a ticket-
ing time limit, the consumer who does not purchase a ticket will have the booking
cancelled, whereas the consumer who fails to purchase a ticket in the absence of any
time limit will ultimately be counted as a "no-show" should he/she decide not to
make use of the reserved space.
BOOKING
CONFIRMED
The same reasoning applies to no-show rates across fare classes at departure
time. No-show probabilities are bound to be higher for reservations made for fare
products for which tickets are less likely to have been purchased and those with no
cancellation penalties. A reservation for a very low-priced fare product, on the other
hand, is more likely to have been made and tickets purchased well in advance. The
price-sensitivity of the "typical" low-fare passenger, combined with a cancellation
penalty on the low fare, should result in a lower no-show rate for the more restricted
fare products.
As discussed in the preceding two chapters, differentiated fare products are tar-
geted at distinct market segments of the total demand for air travel. Individuals
in each of these segments will emerge from the air travel choice process with pref-
erences for different types of fare products. To make the most effective use of the
differential pricing strategies that they have adopted, most airlines practice seat
inventory control to limit the number of seats that may be sold as each of the fare
products offered. In airline reservations systems, limits are placed on the number
of seats available in each fare class or booking class, which can contain several fare
products. Controlling the mix of fare products to be sold for a particular flight has
come to be regarded by some airline managers as by far the most important aspect
of marketing, more important than the actual prices charged for each fare product
[35].
This chapter provides an overview of the seat inventory control problem from
the airline's perspective. The problem is described in the context of current airline
operations, given existing route systems, fare structures and reservations systems.
The types of data available to most airlines for dealing with this problem are outlined
and alternative approaches to limiting the seats available to various fare classes are
discussed. The second section of this chapter then summarizes the state of current
practice in seat inventory control in the airline industry, and identifies the constraints
and limitations that airlines are working to overcome.
3.1 Problem Definition and Context
While yield is an important airline revenue measure, it should not be used as the
only indicator of an airline's ability to generate revenues and, in turn, its profitabil-
ity. High yields and low load factors could be far less profitable for an airline than
lower yields and higher load factors. Yields and traffic levels tend to be inversely
related, such that the attainment of both high yields and high load factors can be
an elusive goal for airlines, particularly in competitive markets.
The yield or average revenue per passenger-mile collected by an airline that offers
a multiple-tier fare structure will depend on both the price levels associated with,
and the number of seats made available to, each fare class. Yields can be increased
either by increasing price levels or by reducing the proportion of seats sold in the
lowest fare product categories. Neither action, however, will guarantee an increase
in total operating revenues for the airline. An increase in yield caused by reducing
the availability of low-fare seats must be weighed against a potential loss of traffic.
Revenue management involves both pricing and seat inventory control. Under
ideal conditions, airlines would have sufficient market information to maximize total
revenues by finding an optimal combination of prices and seats associated with each
of the fare products to be offered on a future flight. In practice, airlines must take
into account the pricing actions of their direct competitors, particularly those actions
that affect the price-sensitive market segments. In the past, airlines have matched
competitors' fares in a very predictable manner, especially the lowest fares. More
recently, some airlines have recognized that some differences in service levels (e.g.,
non-stop versus connecting flights) offered in a market can allow them to maintain
a price differential, even for the lowest-priced fare products. Nonetheless, matching
of price levels for all fare products is virtually universal among direct competitors.
The remainder of this dissertation examines the seat inventory control compo-
nent of the airline revenue management problem. For the reasons cited above, the
price levels associated with each of the fare products, and thus the fare classes in the
reservations system, are assumed to be given and constant over the period during
which most of the reservations are accepted for a future flight departure. The seat
inventory control problem, then, is to determine the optimal (revenue-maximizing)
limits on the sale of available seats in the various fare classes being offered on a
future flight departure.
The seat inventory control problem for a particular flight is defined to a great
extent by the equipment utilization decisions made in the airline scheduling process.
From the outset, the decision to operate a particular type of aircraft on a flight rout-
ing can have implications for seat inventory control, given that anticipated demand
and aircraft size are unlikely to match exactly. Routing constraints and passenger
flow imbalances can mean that the most appropriate size of aircraft cannot always
be physically assigned to a flight departure. A significant excess or shortage of seats
relative to total demand for a particular flight departure will reduce or increase
the need, respectively, for strict control of low-fare sales. Seat inventory control
in essence provides an opportunity to adjust for imperfections in the longer-range
schedule design process. Ironically, seat inventory control can at the same time
hide true demand data from those responsible for long-range schedule planning,
as increased availability of low-fare seats can increase total traffic and overall load
factors.
Seat inventory control can be approached with respect to individual flight legs,
the airline's entire route network, or some sub-unit of the schedule such as a partic-
ular connecting complex at a given time at an airline's hub airport. The flight leg
approach is by far the simplest and is currently used by most airlines, as discussed
in the following section. Demand on the flight leg for each fare class is estimated
and seats are allocated among the classes for each future flight leg separately. The
demand for a fare class is considered to be the total number of passengers that
will request a seat on a particular flight leg and fare class as part of their origin-
destination itinerary.
The complexity of the problem, even at the single-leg level of analysis, has in-
creased tremendously with the development of hub-and-spoke route networks by
most large airlines. A large air carrier can operate over 1000 flights per day, serve
several thousand O-D markets and offer five fare classes in each market. American
Airlines, for example, now serves over 2700 origin-destination markets, compared
with about 800 before deregulation [36]. For any one particular flight departure to
its Dallas/Ft. Worth hub, passengers typically can be booked into one of at least
five fare classes to one of more than fifty destinations. There can thus be over 250
possible fare class/destination combinations for each seat on such a flight leg, each
of which will generate different revenue levels for the airline. With reservations
for future flights being accepted up to 11 months in advance, the size of the seat
inventory control problem can become unmanageable.
At the level of the individual flight leg, seat inventory decisions must be made
within the constraints imposed by the airline's network, schedule and reservations
system capabilities. The aircraft assigned to a particular flight departure is known
and, in turn, the number of seats available in each compartment (first, business,
coach) can be regarded as fixed. Seat inventory decisions must also take into account
the overbooking levels to be used for each flight departure, as determined from an
analysis of the costs associated with oversales and subsequent denied boardings
of confirmed passengers, as well as the costs in lost revenue associated with "no-
shows" and unused seats. Because the problem involves the management of available
reservations "spaces" as opposed to physical seats on the aircraft, the interaction
between the fare class mix of passengers booked and the number that ultimately
show up for a flight can have significant revenue implications.
Finding the optimal limits on the number of bookings that may be accepted in a
particular fare class on a future flight leg requires estimates to be made of both the
expected demand for each fare class and the average revenue associated with each
class. Whether these estimates are based entirely on historical patterns or derived
from a forecasting model, data from past flights are required. For a forecast of leg
demand, information on booking levels prior to departure and actual boardings by
fare class, by flight leg, and by day of week must be extracted from the reservations
system and stored for seat inventory control decision support purposes. Most air-
lines have implemented database management systems that perform this function.
Estimating the revenue associated with bookings in each fare class can be more
difficult, since revenue data are usually collected and summarized independently of
the reservations system. Thus, the sophistication of an airline's seat inventory con-
trol system depends not only on the particular approach used to determine optimal
fare class limits, but on the quality of the data retrieval, forecasting and estimation
methods that provide the input data required for such calculations.
The actual process of seat inventory control for a future flight departure can be
as simple as a one-time setting of booking limits on discount fare classes at the start
of the reservations process for that flight, and taking no further action as reser-
vations are accepted. A more sophisticated approach would take into account the
information provided by actual reservations as they are accepted, through monitor-
ing of booking data, and then adjust fare class booking limits as flight departure
approaches. A comprehensive effort to manage seat inventories and improve yield
would be a dynamic one in which traffic and booking histories are used to set initial
fare class booking limits, actual bookings in each fare class are monitored relative
to these limits, and frequent adjustments are made on the basis of an analysis of
past data, current bookings, and forecasts of future bookings for the flight.
Setting, monitoring and adjusting fare class booking limits is a process which
requires both technical capabilities and human expertise. Management information
systems are required to retrieve, summarize and analyze historical reservations and
traffic data. The judgement of seat inventory control analysts assigned by the airline
is also necessary to determine the extent to which historical data can be applied
validly to current conditions in each market. Clearly, rapid changes to competitive
conditions in the industry or in particular markets, or unusual operational events in
a market, reduce the value of historical data. Still, there is a wealth of information
available to carriers in their reservations systems which can be used to assist in seat
inventory control by reducing the manual effort required by analysts while improving
their decision making process.
The process employed by an airline to manage its seat inventories will be affected
by the range of tools and resources that it has available for this purpose, which will
in turn be determined by the importance that corporate decision-makers attribute to
improving revenue management. From the outset, the seat inventory control system
employed will also depend on the characteristics of the airline's network and on its
fare structure. An airline offering a single fare level for all seats on flights serving
point-to-point markets on a non-stop basis clearly need not be concerned about
sophisticated seat inventory control techniques. At the other extreme, a carrier
with multiple fare classes on flights into and out of large connecting hub complexes
can benefit immensely from improved seat inventory control.
Most established (i.e., pre-deregulation) airlines are closer to the latter extreme,
and as such are extremely interested in revenue management. Among such airlines,
there is a range of effort devoted to, and a range of sophistication in, seat inventory
control. As described in the following section, current practice in this area is evolving
quickly, yet a strong emphasis on human expertise in making seat inventory control
decisions remains.
Seat inventory control and revenue management are closely related to a range of
other functions found in current airline corporate structures, such as pricing, mar-
keting, reservations, overbooking and payload control. None of the airlines surveyed
have been able to combine all the functions critical to revenue management into a
single unit, although several major carriers have moved in this direction. Pricing
and overbooking control were the most frequently named functions to have been
incorporated into the revenue management unit, which in turn was most commonly
found in the airline's marketing or market planning department. Coordination with
the remaining related functions that for various reasons could not be included in
the same department, poses a problem for most of the airlines surveyed.
There seems to be a consensus among those involved in the process that the
preferred place for a seat inventory/revenue management group is in the airline's
marketing department. The close relationship of seat inventory control to the reser-
vations, pricing and fare product development functions make the marketing depart-
ment a logical place for the revenue management function. At many of the airlines
surveyed, however, those responsible for setting and monitoring flight overbooking
levels remain in separate departments responsible for reservations and sales in some
cases and payload control/operations in others. There can thus be substantial re-
sistance and turf-based conflicts with respect to the transfer of the overbooking
function to an ever-expanding marketing department.
The personnel responsible for monitoring and adjusting booking limits through-
out the'reservations process - seat control analysts - make up the largest group
in most revenue management units. There were substantial differences among the
carriers surveyed in the use of seat control analysts, both with respect to the num-
ber of analysts employed and the degree to which the analysts are responsible for
specific markets.
Although the number of analysts employed does not necessarily reflect the so-
phistication of an airline's seat inventory control process, it does reflect the resources
devoted to the problem. The number of agents "required" to manage a carrier's seat
inventory would seem to be related to the number of flight legs operated each day
and to the proportion of a carrier's flight legs that operate in highly competitive
markets or during peak periods of demand.
Table 3.1 illustrates the range of human resources devoted by the nine airlines
surveyed to the flight monitoring and booking limit adjustment process. When the
carriers were compared in terms of daily flights per seat control analyst, the largest
carriers proved to be in the middle of this range, while the high and low extremes
Table 3.1: Comparison of Seat Control Analyst Staffing Levels
A 25 200 + 1,000 + 60
B 35 200 + 1,000 + 30
C 35 200 + 1,000 + 40
D 15 150-200 1,000 + 90
E 20 150-200 500-1000 30
F 10 100-150 500-1000 110
G 25 100-150 0-500 20
H 5 0-100 0-500 150
I 20 0-100 0-500 15
Approximated to nearest 5.
Ranges are given to protect airline anonymity.
were provided by the smaller carriers. Carriers F and H realize that the resources
they currently devote to seat inventory control are inadequate, and have plans to
expand their efforts in the near future. Carrier G had recently undergone such an
expansion, along with a major reorganization of its revenue management group, in
order to monitor its flight leg bookings more closely.
For those carriers with relatively few analysts working on seat inventory con-
trol, booking limit monitoring and adjustment is almost entirely an ad-hoc process,
perhaps targeted at selected markets and flight legs. The carriers with proportion-
ately more analysts generally have a more systematic process in which teams of
analysts are responsible for groups of markets and/or flight legs. The carrier with
the highest number of analysts relative to its daily departures has taken the notion
of specialization to the extreme, making each seat control analyst responsible for all
flight legs that serve a particular market or set of routes. Each agent must consider
historical data, competitors' actions and current trends to both set initial fare class
booking limits and make adjustments as reservations are accepted. These agents
are then held accountable for the traffic mixes and revenue levels achieved on their
own routes.
Overall, the airlines surveyed face very similar problems in organizing their rev-
enue management groups. While some carriers have progressed more rapidly than
others in unifying seat inventory control and related activities, coordination of func-
tions, resistance to organizational change and even intra-departmental organization
are still issues that all face as their revenue management functions evolve.
Most of the airline reservations systems in place today are deficient in several
areas of relevance to seat inventory control. For example, the number of fare class
"buckets" into which bookings can be logged is a major limitation for most of the
carriers surveyed. At least one bucket is required for each of the physical compart-
ments on an aircraft. Any remaining buckets are used by most carriers as sub-classes
of a shared coach compartment. Currently, most systems limit fare class bookings
by flight leg and some have the capability of limiting sales to local passengers in
favor of through and connecting passengers that generate more total revenue. The
large number of fare product/O-D market combinations possible on a single flight
leg makes it desirable for the airline to be able to take reservations in a large number
of fare classes and to limit sales in specific markets when necessary.
Airline attempts to make use of more fare class buckets to control their seat
inventories by passenger O-D itinerary have been impeded by the existing standard-
ized interline distribution system. The major reservations systems used by travel
agents currently display seat availability and accept bookings in a maximum of five
classes. Given the need to exchange availability and booking information with other
reservations systems, airlines are constrained in improving their own systems by the
need to maintain this standardization.
Several carriers are nonetheless changing their own reservations systems to ac-
cept bookings in a much larger number of reservations buckets. In one example of
what will be the "new generation" of reservations systems, eight primary fare classes
will have up to five subordinate buckets for use in controlling low-fare bookings in
particular origin-destination markets. The expanded reservations systems will allow
an airline to limit, for example, the number of extremely low-priced seats in selected
markets without closing down the entire fare class on a flight leg to additional book-
ings. Existing reservations systems do not permit the airline to differentiate between
passengers requesting similar fare products for very different itineraries on the same
flight leg, as these passengers are booked in the same fare class.
The implication of the five fare class limit imposed by the interline standards is
that all airlines hoping to upgrade their own systems are constrained by the need
to remain compatible. Carriers that operate flights with both First Class and Busi-
ness Class compartments are particularly constrained by this compatibility problem.
With distinct reservations classes required for each of the physical compartments
on the aircraft (including the coach cabin), only two fare classes remain for control
of reduced fare sales. In markets where there exist, for example, 30-day advance
purchase fares, 14-day advance purchase fares, and unrestricted "super-coach" dis-
counted one-way fares, the availability of only two discount fare categories can force
the airline to aggregate rather dissimilar fare products and to give up some of its
control over reduced fare sales.
Over half of the carriers surveyed currently have a totally or partially nested
fare class hierarchy in their reservations systems. Other carriers have their coach
cabin fare classes independently nested in a cumulative Y fare class, with little or no
interdependence among the reduced fare classes within the coach cabin. The airlines
planning to upgrade and expand their reservations systems agree that a nested fare
class hierarchy would be preferable from the seat inventory control perspective. In
the expanded systems, each sub-class could be nested within the respective primary
class, which would in turn be nested within the cumulative coach cabin reservations
bucket.
The decision support function of the airline reservations system is distinct from
its capabilities in terms of the number of fare classes and their relationships. Con-
sequently, the data management and analysis functions required for decision sup-
port in seat inventory control can be developed and even used independently of the
reservations system itself, although linkages for recovering raw data from the system
must be maintained. Of the carriers surveyed, those with computer systems large
enough to handle the extra load are developing their decision support tools for use
in conjunction with their reservations systems. On the other hand, several carriers
expressed concern that the decision support functions would tax their already over-
burdened reservations systems and significantly increase delays for both reservations
agents and customers. These carriers are contemplating the purchase of separate
software and hardware for seat inventory control decision support purposes.
At the simplest level, setting the initial booking limits for reduced fare classes
can be done on an "across-the-board" or default value basis. The use of default
values for all flights operated in certain types of markets or with particular air-
craft types requires little in the way of resources, but does not take into account
important differences in passenger mixes and booking patterns between markets or
even between flights in the same market. Greater precision can be achieved by set-
ting lower fare class limits by market, day-of-week, even time-of-day of future flight
departures, but this requires substantial analytical effort.
If the airline's seat inventory control process consists solely of setting fare class
limits at the start of the booking process, then a more detailed approach is certainly
preferable. All the carriers surveyed have progressed beyond this simplest level
of seat inventory control, although several have done so only in the recent past.
Most airlines have recognized that greater benefits from their seat inventory control
efforts can be realized by developing a reservations monitoring and booking limit
adjustment process than by intensifying their efforts to improve the accuracy of the
initial limits.
That is not to say that the initial limits placed on different fare classes are not
important. Any differentiation among markets or flights can help to reduce the
amount of intervention required later in the process and is better than no differenti-
ation at all. One of the larger carriers surveyed uses some default values for groups
of flights and then identifies specific markets, flights, and/or periods of operation
that might require more careful attention. Another carrier that assigns responsi-
bility for different routes to individual agents leaves the setting of initial limits to
them. The default limits automatically entered into the reservations system for this
same carrier are the actual passengers boarded by fare class for the same flight,
one year earlier. Most of the remaining carriers use aggregate approaches in which
limits are initially set according to aircraft type and/or broad market groupings.
More sophisticated computer routines can be designed to flag flights for which
actual reservations levels meet or approach one of several different types of booking
limit values. For example, a flight might be flagged by the monitoring system if
the number of bookings accepted in a lower fare class approaches the class limit, if
total low-fare bookings exceed some percentage of total seats, and/or if the total
number of unbooked seats drops below some pre-set value. The most advanced of
the automated monitoring systems found among the airlines surveyed makes use of
all three monitoring methods. That is, flights are flagged when any one reduced
fare class reaches it own booking limit, when the nested low fare classes reach a
cumulative limit, or when the total number of unbooked seats remaining in the
coach cabin falls below some targeted level.
The airline reservations systems surveyed all monitor actual bookings relative
to pre-set and static fare class limits, with varying degrees of sophistication. An
improvement to the monitoring process is provided by the CDC yield management
package, among others. On the basis of historical booking trends for a flight or group
of flights, the CDC package generates "booking threshold curves" which show the
expected range for cumulative bookings at any point before departure (see Figure
3.1). The flights for which actual bookings stray outside the range of these curves are
then flagged by the system and listed in periodic reports. The dynamic monitoring
capability provided by the CDC package can be used in conjunction with the static
fare class booking limits described above.
The least advanced aspect of revenue management and seat inventory control at
all of the airlines surveyed is that of booking limit adjustment to maximize flight
revenues. This task involves modifying, when necessary, the initial booking limits
on each fare class to take into account both actual bookings received and additional
bookings expected over the time remaining to flight departure. This is the most
important component of seat inventory control, yet it remains dependent on ad-hoc
human judgement rather than systematic analysis. Resource limitations and a lack
of practical models for making optimal decisions about changes to fare class booking
limits as the flight departure approaches have kept the adjustment task at a low level
of sophistication.
When an airline's reservations monitoring system flags a flight for which actual
bookings approach any one of the limits or the threshold established for that flight,
a decision must be made either to increase the availability of seats in the relevant
fare class or to allow the system to close it down to additional reservations. This
decision is currently being made by seat control analysts on the basis of experi-
ence and judgement at every airline surveyed, although the development of decision
support tools is designed to reduce the amount of guesswork involved. Differences
among airlines in adjusting booking limits stem to a greater extent from the number
and responsibilities of the seat control analysts than from differences in the actual
techniques used to make adjustment decisions.
At -least two carriers are hoping to make the adjustment process more systematic
by developing an automated routine for maximizing expected flight revenues at
various points in time before departure. One such algorithm would cause lower-
yield fare classes to be closed down automatically when the expected revenue from
bookings in higher-yield classes exceeds the revenue from selling additional seats in
the lower-yield class. Some computational problems must be overcome, however,
before such an algorithm can be implemented and be of practical use.
Figure 3.1: Example of "Threshold Curve" Monitoring
200 100
/ /
/ /
"THRESHOLD /
R 150 CURVES -75 P
E E
S R
E C
R FLIGHT E
V "EXCEPTION REPORT" N
A 100 GENERATED /50 T
T
I10 0
0 / / F
N
S S
50 25 E
A
T
ACTUAL BOOKINGS
- S
0 0
45 30 21 14 7 0
DAYS BEFORE DEPARTURE
Even if an optimization algorithm were to be used by an airline, the need for
human judgement in seat inventory control cannot be eliminated entirely. Any op-
timization models would be probabilistic in nature and would make use of forecasts
based on historical data. Changes in the competitive environment of airline markets
and the occurrence of unexpected events that affect recent flight booking patterns
are but two examples of variables which cannot be accounted for in such algorithms.
The objective in developing optimization models for seat inventory control is to al-
low analysts to focus their efforts on these exceptional variables by making routine
tasks more systematic.
The above discussion of seat inventory control practice suggests that a range of
approaches are being used by airlines to address essentially the same problem. The
seat inventory control problem, at the level of the single flight leg, is to determine
the booking limits on each fare class that will maximize total revenues for a future
scheduled flight departure. The static problem is to establish these fare class limits
once at the start of the booking process, taking into account the uncertainty asso-
ciated with expected bookings by fare class, to the extent possible. The dynamic
problem is to revise these initial limits on the basis of the additional information
provided by actual bookings as departure day approaches, once again incorporating
the probabilistic nature of future demand.
In Part Two, mathematical approaches for solving the seat inventory control
problem are considered. Throughout the review of past work and the development
of new approaches, the distinction between the static and dynamic problems will
be emphasized, as will the importance of incorporating probabilistic demand into
any solution approach. The discussion of mathematical approaches will be based on
the concepts developed thus far. The demand segmentation and pricing practices
described in Chapter One define the current environment in which seat inventory
control is practised. The consumer choice and airline reservations frameworks pre-
sented in Chapter Two will be used in the development of solution approaches that
take into account realistic demand behavior. Above all, the practical constraints on
seat inventory control, as described in Chapter Three, will guide the development
of a seat inventory control framework that can be adopted readily by most airlines.
PART TWO
Mathematical Models for Seat
Inventory Management
This "marginal seat" approach is the basis for the development of a more cur-
rent mathematical framework in Chapter Five. A decision model for determining
revenue-maximizing booking limits for multiple fare classes in a nested reservations
system is presented for the single flight leg inventory control problem. Applications
of this model to both static and dynamic seat inventory management decisions are
described. The model is then extended to incorporate airline overbooking of seats
due to passenger no-shows. A method for including the possibility of passenger
choice shifts in the booking process into the model is also described. Finally, exam-
ples of model results are provided for a hypothetical flight leg.
Chapter Six examines the EMSR model in the context of actual applications to
airline reservations systems. The demand inputs and the assumptions required when
the model is applied to the data provided by existing leg-based airline reservations
systems are discussed first. The impacts on the EMSR output results of the most
important demand assumptions are described, and the sensitivity of these outputs
to each of the input variables is considered. Finally, the potential application of the
EMSR approach to seat inventory control systems designed to limit seat availability
by passenger itinerary, currently under development, is introduced.
Chapter 4
An Overview of Previous
Research
Although similar to both of these problems, seat allocation among distinct fare
class buckets that will share a common compartment involves some subtle differ-
ences. Because no distinction is made between passengers booked in different fare
classes in terms of on-board service, there is no possibility of upgrades or downgrades
from one fare class to another in the case of oversales. The problem of oversales
involves the total capacity of the shared cabin. Another difference is that a single
seat may be allocated to a fare class inventory, whereas at least a row of seats must
be allocated to a physical compartment, even with moveable bulkheads. Finally, an
airline may offer seats in as many fare classes as can be handled by its reservations
system, while there is a practical limit on the number of physical compartments and
levels of on-board service that may be provided on a single aircraft.
Seat allocation among distinct fare classes that share a common inventory of
identical seats is nonetheless essentially a problem of distributing available seats
among "invisible" compartments that are defined only in the reservations system.
In fact, the seat allocation problem has been referred to as the "electronic bulkhead
problem". Once a seat is assigned to a fare class inventory, it may be booked only
in that fare class, or remain unsold.
In contrast, a "nested" reservations system is one in which the fare class in-
ventories are structured such that a high-fare request will not be refused as long
as any seats remain available in lower fare classes. A nested reservations system
is thus binding in its limits on lower fare classes, but its limits are "transparent"
from above (for higher fare classes). The nesting applies to each successively lower
fare class, such that the limit on the lowest class, say Class 4, applies to bookings
in Class 4 only, whereas the limit on Class 3 applies to the total of Class 3 and 4
bookings. Sales to Class 3 can in essence take over the inventory set aside for Class
4. The booking limits that apply in a completely nested reservations system are
illustrated for a four-class example in Figure 4.1.
The distinction between separate and nested fare class inventories is important
both to the way in which the seat inventory control problem is represented and to
the mathematical methods that are used to determine optimal seat allocations or
fare class booking limits. Equally important is the distinction between the static
problem, in which fare class booking limits are applied at the start of the booking
process for a future flight, and the dynamic problem, in which booking limits may
be revised as actual bookings are accepted. In the latter case, the length of the
interval between revisions determines the degree to which the differences between
distinct and nested fare class inventories affect optimal booking limits. Request-
by-request revision of limits on distinct inventories eliminates the expected revenue
advantages associated with nested inventories. The longer the interval between
revisions, however, the greater the importance of finding the optimal booking limits,
particularly in the case of nested fare classes.
The following section reviews the solution approaches that have been presented
in past work for seat allocation among distinct fare class inventories. The discussion
centers on the static problem, although applications of the methods described to the
dynamic problem are introduced. The subsequent section then presents an overview
of the "marginal seat" model for limiting seats sold in low fare classes. This approach
has been developed in the context of the dynamic seat inventory control problem,
but is equally applicable to the static problem for nested fare class inventories.
90 = BOOKING LIMIT ON
2 30 PROTECTED
FOR CLASS 11 CLASS 2 (+3+4)
120 0
SEATS
Apart from the characteristics of demand for different fare classes, assumptions
with respect to passenger no-shows or cancellations as well as the expected reactions
of passengers denied a reservations request are required when mathematical models
for finding optimal seat allotments are considered. In general, past work has been
based on the following simplifying assumptions:
2. A booking accepted by the airline will translate into a revenue passenger car-
ried, meaning that there are no cancellations or no-shows;
In the allocation of seats among distinct fare class buckets, no assumption as to the
booking rates in the different fare classes is required.
It is possible to find the optimal seat allotments for distinct fare class inventories
through a relatively straightforward application of differential calculus when only
two fare classes on a single flight leg are involved. The seats allocated to each
fare class (S1 and S2) must share a total capacity of C seats. To find the value
of Si = C - S2 that will maximize total expected revenues, R, for the flight, we
differentiate X with respect to Si and set to zero:
aR/|s 1 = 0 (4.2)
That is, seats are allocated between the two fare classes such that the marginal
expected total revenue with respect to additional seats in one class or the other is
equal to zero. At optimality, total expected flight revenues cannot be increased by
taking a seat from class 1 and allocating it instead to class 2. The expected marginal
revenue of the last seat allocated to each class will be equal across the two classes:
This analysis can be extended to three or more distinct fare class inventories,
where S;* for each fare class can be found to maximize total revenues. With more
than two S; variables, the problem becomes more difficult to solve analytically, but
can be formulated as a constrained revenue maximization problem and solved with
the Lagrangian multiplier method. The optimality conditions for the multiple fare
class problem would then be:
8R aR
-- = -- =A, for all fare classes i :j (4.4)
aS;a 0Si
where A, the Lagrangian multiplier, equals the expected marginal revenue for the
last seat allocated to each fare class. As in the two-class case, seats are allocated
among fare classes such that the total expected marginal revenue is equal across all
relevant fare classes. Under the Lagrangian multiplier formulation, a single, non-
linear equation must be solved for A, and it can be shown that this solution is a
unique optimum [42].
Further expansion of the problem to multiple fare classes and multiple-leg flights
or even connecting flight operations requires the application of mathematical pro-
gramming and related techniques to find the optimal fare class seat allotments nu-
merically. All of these methods, as reviewed below, make use of the principle of
equating marginal revenues of the last seats allocated to each inventory in deter-
mining the revenue-maximizing seat allotments.
The network flow approach was pursued by Glover et. al. [45], who postulated
that there is "some number of passengers at each fare class on each flight segment
that will optimize revenue" for the airline. They developed a network-based seat
allocation model for Frontier Airlines, which was designed to find the mix of pas-
senger itineraries (PIs) flowing over the airline's network in various fare classes that
would maximize total daily revenues. The model required inputs of the demand and
revenues expected for each PI from j to k in fare class i, and produced outputs of
the optimal number of seats to be allocated to that PI/fare class combination.
In this network formulation, a set of forward arcs represented the flight legs to
be operated between points in the airline's system (nodes on the network). Flow
on these arcs was associated with passenger loads on each leg, and was constrained
by the leg capacities. A set of backward arcs represented all possible PI/fare class
combinations. Flow on the backward arcs consisted of the number of seats to be
allocated by PI/fare class, constrained only by the total demand for each PI/fare
class combination. The network was solved as a maximum flow (revenue) problem
with side constraints to maintain feasibility (i.e., eliminate cycles) when connecting
flights were included.
Glover's model for Frontier accommodated up to 600 daily flights and 30,000 PIs
in five fare classes. The formulation required 200,000 variables and 3,000 constraints,
and took several hours to solve, making interactive use impossible. A more serious
shortcoming of the model is its reliance on deterministic demand estimates for each
PI/fare class combination. No consideration of probabilistic demand or spill was
included.
Analysts at both Boeing Aircraft [46] and McDonnell-Douglas [47] have ad-
dressed the problem of incorporating probabilistic demand into mathematical pro-
gramming formulations of the seat allocation problem. Because the expected revenue
function for each class is non-linear given a constant fare and stochastic demand, a
simple linear program is an inaccurate representation of the problem. McDonnell-
Douglas analysts proposed a formulation of the single-leg seat allocation problem
that makes use of binary decision variables in a linear integer programming frame-
work. Each variable, Xak, represents the combination of fare class i and seat k on
Table 4.1: Optimal Seat Allocation - Deterministic vs. Stochastic Demand
With an available total capacity of n seats, the formulation of this linear program
is as follows:
MAX W(n) = ( (jX-i - m;(k) (4.5)
i k
subject to:
(E Xi n
i k
The solution to this linear program will be integer, with the Xik variables corre-
sponding to the n largest values of m;(k) equal to 1, and all other Xik = 0.
When the linear program was enlarged to the two-leg problem, each Xi, repre-
sented a seat k which could be allocated to a specific fare class/passenger itinerary
(PI), i. A two-leg flight offers three possible passenger itineraries (A-B, B-C, A-C).
With four fare classes, there would thus be 12 feasible fare class/PI combinations.
We define A(1), A(2), and A(3) to be the set of fare classes available to passenger
itineraries A-B, B-C, and A-C, respectively. Then, with ni and n2 unbooked seats
available on flight legs 1 and 2, the linear program becomes:
subject to:
SX4
A(1) A(3)
<ni
E E Xk 5 n2
A(2) A(3)
0 5 Xk < 1
The two-leg case presented above would require 1800 arcs (3 PIs, 4 classes, 150
seats) in a complete network formulation for a 150-seat aircraft. As the formulation
is expanded to multiple flight legs and/or passenger itineraries involving connect-
ing flights, the number of Xik variables (arcs) required increases rapidly. By using
the network characteristics and the decreasing m,(k) property of the seat allocation
problem, however, Wollmer was able to reduce dramatically the size of the formu-
lation required for solution purposes. He suggested that only the shortest (lowest
Figure 4.2: Network Representation of Seat Allocation Problem
A solution algorithm for this network problem therefore need only solve a series
of longest path problems for a relatively small network. At each iteration, the
expected marginal revenue of each fare class/PI combination in the current reduced
network must be determined, the longest path identified, and a seat allocated to
that path. The marginal expected revenues (lengths) associated with the arcs on
this longest path would then be updated and the procedure repeated. Wollmer and
Stroup at McDonnell-Douglas have continued work on the seat allocation problem,
extending the mathematical programming formulations to an airline connecting hub
operation [50]. The network characteristics of these formulations and the expected
marginal revenue approach to optimal seat allocation are being used to develop
solution algorithms more efficient than the integer programming methods described
above.
where fi is the revenue associated with the current request. In other words, the
marginal expected network revenue from retaining the nth seat on the flight leg
requested for possible sale in higher fare classes and/or for other itineraries is com-
pared to the revenue associated with the current request.
This comparison of the expected revenue differential for each incremental un-
booked seat with a "certain" revenue from the current reservations request is ex-
tendable in conceptual terms to the most complex of network formulations. The
drawbacks of such an approach, however, make its practical use by airlines unlikely
in the near future. The computing capabilities required to run network optimiza-
tion models in order to make expected revenue assessments for each reservations
request received by an airline would be enormous. Even with the reduced network
formulations proposed by Wollmer, finding MAX W(n) and MAX W(n -1) for each
fare class/PI request would be a major burden on even the most advanced airline
computer systems.
Given that most airlines in practice only make revisions to fare class booking
limits periodically, the solution to the static seat allocation problem must apply over
the entire interval between revisions. For applications in which this interval is sub-
stantial, the mathematical programming and network formulations described in this
section are in fact solving a simplified seat inventory control problem. All of the for-
mulations described find the revenue-maximizing allocation of seats to distinct and
separate reservations "buckets", either by fare class or by fare class/PI combination.
As described in the first section of this chapter, most airlines have the lowest fare
class buckets nested within the higher fare classes, such that a high-fare request will
not be denied as long as lower fare classes are still open. Mathematical program-
ming approaches will therefore not necessarily give the true revenue-maximizing seat
allotments for a nested reservations system.
The optimization models reviewed in Section 4.2 all equate the marginal revenues
of incremental seats allocated to distinct fare class inventories so as to maximize total
expected flight revenues. This principle of equating marginal seat revenues has been
applied by numerous researchers to the problem of dynamic booking limit revision
and incorporated into iterative solution approaches. The assumptions of separate
demands for each fare class, no cancellations or no-shows, and no possibility of
recapturing a refused reservations request have been made in virtually all of the
work reviewed in this section. The assumption of probabilistic demand, however,
has been incorporated effectively in most cases. Dealing with the dynamic problem
requires an assumption concerning the order in which and the rates at which requests
for different fare classes are received. The assumption that all low-fare requests are
received first has been the most common, although alternatives have been examined,
as will be discussed.
The marginal seat model was applied in a dynamic reservations context to the
two-class, single flight leg seat inventory control problem by Littlewood in 1972 [51].
He suggested that revenues could be maximized by "closing down" the low-fare class
to additional bookings when the certain revenue from selling another low-fare seat
is exceeded by the expected revenue from saving that seat for a potential high-fare
passenger. That is, low-fare passengers paying f2 should be accepted as long as:
2 71(SI) - f (4.8)
where 71 (S1 ) is the probability of selling all Si remaining seats to high-fare passen-
gers, and fi is the higher fare level. The smallest integer value of Si that satisfies
the above condition is the number of seats that should be protected for class 1 in a
nested fare class system in order to maximize expected revenues.
This formulation includes the probability density of reservations requests for class
1, p1(ri), which when integrated from C - S 2 = Si to infinity produces 11(31), as
defined in Littlewood's model.
DR = f2 -P 2 (S 2 ) - fi 1- 1 (C - S 2 +1) (4.10)
The expected marginal seat revenues for the two fare classes are equal when DR is
set to zero. The optimal value of S 2 must therefore satisfy:
fi
f- =
_
(2 (SS22) )
1 (4.11)
A2 Pil(C - S2 + 1)
This "differential revenue" model in fact generates the optimal seat allotments
for two distinct fare classes, in which a seat allocated to the lower fare class can
actually deny a seat to a high-fare passenger. Richter demonstrated, however, that
the optimal limit on the lower fare class in a dynamic application is a function of the
relative fare levels, total capacity, and the demand for the high-fare class only. It
is not a function of low-fare demand, although the distribution of low-fare demand
and the booking process assumed will influence the expected total revenue for the
flight. For dynamic applications, the optimal value of S2 must satisfy:
f 2 /f 1 = 7 1 (C - S2 + 1) (4.12)
This optimality condition is thus equivalent to Littlewood's "simple" model.
The conclusion that the density of low-fare demand will not affect the optimal
number of seats to be allocated to the higher fare class is important in the nested
fare class problem. The implication is that, at any point in time, the airline should
protect Sj seats for potential high-fare demand, to the point at which the expected
revenue from an additional protected high-fare seat equals the actual fare level
in the lower fare class. At this optimal point, the airline is indifferent between
the "certain" revenue of accepting another low-fare booking and the "expected"
revenue from protecting that seat for a high-fare passenger who is yet to appear.
This concept will be central to the development of a mathematical framework for
more complex representations of the nested fare class problem in Chapter Five.
A sensitivity analysis of the simple model under the assumption that low-fare
passengers book first was performed by Mayer of El Al in 1976 [54]. He showed that
the greater the difference between fi and f2, the more sensitive the total expected
flight revenue will be to a non-optimal allocation of seats. The decrease in expected
revenue will be smaller when too many seats are allocated to low fare passengers
than when too few seats are offered.
Mayer also changed the "early-bird" assumption of the simple model, assuming
instead that low-fare passengers would book first in each of many periods before
departure. He further assumed that demand in period t is independent of demand
in previous periods t+ 1, t+2, ...etc., and defined V.(t) to be the maximum revenue
from period t until departure, given n seats remaining at the beginning of the
period. Mayer then applied a dynamic programming framework to find V"(t) and
the corresponding optimal allotment of low-fare seats for every relevant value of n
and t. Dynamic programming allowed for the possibility that not all low-fare seats
would be booked first, or in any particular period t. The optimal booking policy
derived from this formulation would allow as many high-fare seats as requested to
be booked in each period, as long as any seats remain available.
The question of how the assumption that low-fare passengers book first affects
the optimal seat allotments and expected revenue levels was also addressed by Titzer
and Griesshaber in 1983 [55]. The "early-bird" low-fare booking process was com-
pared with a simulated parallel process of booking in which the rate of low-fare
requests decreases and the rate of high-fare requests increases as departure time
approaches. The simulation varied the means and variances of the low-fare and
high-fare demand distributions for both booking behavior assumptions so that re-
sults from several scenarios could be compared.
The results of the sensitivity analysis showed the maximum expected revenue
level to be reached at approximately the same optimal low-fare seat limit, regard-
less of which booking behavior assumption is made. The parallel booking process
simulation produced higher total expected flight revenues at the optimal seat allot-
ment, simply because high-fare passengers were able to book early and fill a greater
proportion of seats on heavily- booked flights, an illustration of the benefits asso-
ciated with nested fare classes. The total revenue curves as a function of low-fare
seat limits were found to be relatively flat around the optimal points, such that
a 20 percent deviation from the optimal limit was required to cause a noticeable
revenue decrease. The revenue curves for the parallel booking process were even
flatter than those for the "low-fare first" scenario, confirming Mayer's findings that
total expected revenues are less sensitive to erroneous seat allotments when a mixed
booking process is involved.
where fAc is the average fare for A to C passengers and 5AC (SAC) is the probability
of selling more than SAC seats to passengers from A to C.
Demand for each origin-destination market - A-C, A-B, and B-C on a two-leg
flight - was assumed to be independent, and conditions were imposed such that:
EAC(SAC) = EA
A(SAB) + EBC(SEC) (4.15)
subject to the capacity constraint. An iterative solution method was used to find
the optimal values of SAC and SAB.
Buhr suggested that the allotment of seats among fare classes be performed as
a second step in the process, given the optimal allotments for local and through
passengers on each leg. He acknowledged that varying the allotment of low-fare
seats for different passenger itineraries could change the expected revenue levels on
each leg of the flight, but did not address this problem.
for the (j,k) pairs in the string, where fit is the fare yield for fare class k in O-D
market j, and P[rgt > bi] is the probability that another request for (j, k) will be
received given bpt bookings accepted.
The terms of the right hand side are summed over all feasible combinations of
sequential (j, k) pairs for the flight. For example, on a flight operating A to B to
C, a feasible sequence would be to allocate the marginal seat to a low-fare class
from A to B and to save that same seat for a high-fare passenger from B to C. In
total, there would be 12 feasible combinations for a one-stop flight with three fare
classes. Wang's approach requires the feasible combinations to be ranked in terms
of expected marginal revenue as each seat is allocated incrementally. This iterative
solution process clearly has limitations in terms of its application to larger-scale
formulations of the problem.
The applications of the "marginal seat" principle described above have succeeded
in incorporating probabilistic demand explicitly into the seat inventory revenue max-
imization problem. The simple decision rule presented by Littlewood and subsequent
researchers determined optimal fare class limits for dynamic revision of booking lim-
its, for two fare classes on a single flight leg. As mentioned, the "marginal seat"
principle can be applied to the static problem when nested fare class inventories are
involved. Applications of the same principle to multiple fare classes and flight legs,
however, have not addressed the problem of nested fare classes, as evidenced by the
works of Buhr and Wang.
Problem formulations with multiple fare classes, flight legs and passenger itiner-
aries have seen the application of mathematical programming and network solution
methods for revenue maximization. The biggest shortcomings of these approaches
relate to their inability to deal in a practical way with setting static limits for nested
fare classes and to incorporate probabilistic demand. Both of these shortcomings can
be overcome with sheer computer power. Optimization models that generate optimal
seat allotments for distinct inventories can be applied dynamically by re-running the
optimization at frequent intervals, thereby taking into account the properties of a
nested reservations system. The same models can incorporate probabilistic demand
for each O-D and fare class combination if enough variables and/or arcs are added
to the formulation. It would be impractical, however, for most airlines to commit
this magnitude of computer resources to the problem.
None of the mathematical models reviewed here address the practical consider-
ations introduced in Part One that can play an important role in determining the
optimal booking limits for different fare classes. The relationship between overbook-
ing and seat inventory control has been overlooked. Furthermore, refused requests
are not necessarily lost to the refusing airline. The refused passenger may be ac-
commodated in the requested fare class on another flight of the same airline, or
might in fact agree to accept a reservation in a higher fare class on the originally
requested flight. There is also the possibility of significant correlation of demand
levels among fare classes. All of these considerations can have a considerable im-
pact on the optimal fare class booking limits, and complicate the derivation of these
limits. Nonetheless, a realistic solution model must take these factors into account.
The size of the seat inventory control problem and the volatility of the airline
competitive environment dictate that any optimization model be both efficient and
adaptable to changing conditions. Decision rules that can be used dynamically
to limit bookings on flight legs or in specific markets might be a more practical
approach to improving seat inventory control than large-scale optimization models.
The emphasis in any future model development must be on matching the solution to
the practical constraints of the problem, including reservations system capabilities,
data availability and the nature of airline competitive practices. Above all, efforts
must be made to incorporate more realistic demand assumptions and to find optimal
booking limits for reservations systems with interdependent or nested fare classes.
100
Chapter 5
The probabilistic nature of the seat inventory control problem is examined first,
and the use of probabilistic concepts is illustrated in a very simple model for seat
allocation between two distinct fare classes. The "marginal seat" principle discussed
in the previous chapter is then extended to account for multiple fare classes in a
completely nested reservations system, and an "expected marginal seat revenue"
model for making revenue-maximizing seat inventory decisions on the basis of his-
torical demand is presented. This framework is then expanded to take account of
actual bookings accepted for a future flight leg, making it a dynamic seat inventory
control model.
Section 5.4 begins the process of incorporating more realistic demand behavior
into the EMSR decision framework. The uncertainty associated with passenger no-
shows is introduced and the model is extended to account for flight overbooking.
The probability of a refused passenger accepting a vertical shift to a higher fare
class is incorporated to illustrate the impact of passenger choice behavior on rev-
enue maximization methods. Finally, examples of EMSR model results for set of
hypothetical demand and revenue inputs are presented.
101
5.1 A Probabilistic Approach
The total demand for a particular flight, on average, will fluctuate systematically
in cycles described by day of week and season of the jear. There will also be
stochastic variation in demand around the expected values, among similar flights
sampled consistently over a homogeneous period of time. This stochastic demand for
a future flight departure can be represented by a probability density function. Past
analyses have generally assumed a Gaussian (normal) distribution of total demand
for a flight, with means and variances that depend on the market being studied and
on the nature of its traffic [581.
For the purposes of this discussion, we will continue to assume Gaussian demand
densities. Furthermore, we assume initially that the demand densities for different
fare classes are not correlated significantly and the number of requests received
for the various fare classes during different periods before flight departure are not
correlated. These assumptions will be discussed further and tested empirically in
Chapter Six.
The notion of probabilistic demand is central to the airline seat inventory control
problem, as the expected number of requests for each fare class must be estimated
from historical distributions of demand. We define pi(ri) to be the probability
density function for the total number of requests for reservations, rg, received by the
airline for seats in fare class i by the close of the booking process for a scheduled
flight leg departure. This probability density function is derived from historical data
for the same or similar flights, and is assumed to remain valid for future operations
of the flight, at least in the short run and in the absence of changes to exogenous
factors.
Also important is the capacity constraint on the total number of seats available
- on a flight leg. The number of seats allocated to a particular fare class, Si, might
not exceed the number of requests for that fare class, resulting in rejected demand,
or "spill". We can thus define a cumulative probability that all requests for a fare
class will be accepted, as a continuous function of S;:
102
Conversely:
P[r > Si] = p,(r,)dri = 1 - P(Si) = 5i(Si) (5.2)
The probability of receiving more than Si requests for fare class i, or the probability
of spill occurring, is therefore FP(S;). An illustration of the relationship between
p;(r;) and P;(S;) is provided in Figure 5.1.
= f r - p,(r;)dri - S -P;(S;)
i + i = Fi (5.6)
The probability of any given request being refused in class i (the refusal rate) is
simply:
PRR = li/Fi (5.7)
It should be emphasized that PRR, the expected rate of request refusals, does not
equal 1;(S;), the probability that at least one request will be refused and the fare
class will sell out.
103
Figure 5.1: Typical Probability Distributions of Requests
PROBABILITY
DENSITY
FUNCTION
p (r )
AREA = P.(S.)
*I 'I
max
REQUESTS PER
FLIGHT FOR
CLASS i
OF DENIED
1.0
REQUESTS
0.5
S,
SEATS
AVAILABLE
104
can expand on the static model for seat allocation introduced at the start of section
4.2 to include probabilistic demand explicitly. As before, we want to determine the
optimal allocation of seats between distinct (i.e., non-nested) fare classes, subject to
the total capacity constraint. The total capacity of the cabin to be shared among i
fare classes is C, such that:
C =>2S (5.8)
Let f; be the average fare received by the airline when a reservation request for fare
class i is accepted. We want to find the integer values of Si that will maximize total
expected revenue, R, for a flight:
In a two-class example, we have fare classes 1 and 2 with relative fares fi and
f2. To find the value of Si = C - S 2 that will maximize total expected revenues, R,
for the flight, we differentiate 7 with respect to Si and set to zero:
The expected marginal seat revenue of the Sith seat in fare class i, EMSR(S,),
is simply the average fare level in that class multiplied by the probability of selling
Si or more seats:
EMS Rs(S;) = f; P;(S;) (5.11)
The optimal values of Si and S2 in the case of two distinct fare class inventories
must satisfy:
EMSR1 (S*) = EMSR2 (S2*) (5.12)
These optimal values of Si and S2 will depend on the parameters of the probabil-
ity densities of expected demand for each fare class, the relative fares or revenue
levels, and the total capacity available. The relationships between W, Si and C are
illustrated for the simple two-class model in Figure 5.2. R = 7i + W 2 is maximized
when Si*out of a total of C seats are allocated to the higher-priced fare class 1.
105
Figure 5.2: Optimal Seat Allocation, Two Demand Classes
p . (r )
REVENUE
($)
S2 , 2
SEATS
106
When this analysis is extended to three or more distinct fare classes, S* for each
fare class can be found to maximize expected total revenues given the densities of
demand and relative fare levels, subject to a capacity constraint. With more than
two S; variables, the problem becomes more difficult to solve analytically, but can
be formulated as a constrained revenue maximization problem and solved with the
Lagrangian multiplier method. The optimality conditions for the multiple fare class
problem would then be:
where A, the Lagrangian multiplier, equals the expected marginal seat revenue for
each fare class. As before, seats are allocated among fare classes such that the
expected marginal revenue is equal across all fare classes.
The values of Si derived from this model represent the optimal allotments of
the available seats to different fare classes based on expected demand for each fare
class. The demand for each fare class is assumed to be independent of that for other
classes, and the optimal seat allocation is made only once, at the beginning of the
booking period for a flight. In reality, demand for different types of fares might
not be independent, as high demand for one fare class could be associated with
high demand for another. More importantly, this static seat inventory management
model does not account for the dynamic nature of the reservations process in which
actual bookings accepted for a particular flight might provide valuable additional
information, as the time to departure decreases, about the ultimate number of
requests that can be expected for that flight.
This simple model allocates a given number of shared seats among distinct or
"stand-alone" reservations fare classes so as to maximize total expected fRight rev-
enues. Optimal seat allotments were derived under the assumption that any one
seat could be made available to one fare class or another, but not both. In a nested
reservations system, the possibility that more than the "allotted" number of seats
may be sold to higher fare classes suggests that these formulations do not accurately
represent the problem faced by many airlines. Any model that is used to find a fixed
"allotment" of seats among fare classes in a nested reservations system overlooks the
lost revenue potential when unexpectedly high demand causes high-fare requests to
be refused when seats are available in lower classes.
107
management concept was based on the protection of seats for high-fare passengers
and made only those seats that would ultimately remain empty available to discount
passengers. While many of the assumptions of the surplus seat system have become
obsolete, this basic notion of protecting seats for higher-fare passengers remains
valid.
The seat allocation models of the previous chapter assumed independent fare
class demand densities to correspond with the distinct fare class inventory assump-
tion. For the time being, the assumption of no relationship between demand levels
for different fare classes will be retained. In other words, we will assume that, along
the lines of the market demand segmentation model developed earlier, demand for a
particular fare class can be identified and distinguished from other classes. Further-
more, we assume that there is no possibility of vertical or horizontal choice shifts on
the part of a consumer denied a flight/fare class request, meaning a refused request
is a booking loss to the airline. On the other hand, an accepted booking repre-
sents certain revenue for the airline, as we assume that no booking cancellations or
passenger no-shows occur.
108
With the help of the probabilistic concepts introduced in Section 5.1, the ex-
pected marginal revenue approach to seat allocation can be extended to multiple
fare classes in a nested reservations system. Given the historical density of requests
for a fare class i and, in turn, the expected bookings as a function of S;, the expected
revenue from S; seats available in class i is:
R; = f; -Ei(S;) (5.14)
where fA is the net revenue or average fare collected from passengers booked in class
i.
We defined EMSR; to be the expected marginal seat revenue for class i when
the number of seats available to that class is increased by one. The relationship
between expected marginal seat revenues and expected total revenues for a fare
class is illustrated in Figure 5.3. Note that EMSR(S;) depends directly on Pi(S;),
the probability that the number of requests will exceed the number of seats available.
To begin with a simple example, we consider a single-leg flight for which bookings
will be accepted in two nested fare classes, 1 and 2, having average fare levels fi and
f2, respectively. In order to maximize total expected flight revenues, the reservations
process should give priority to class 1 passengers. Class 1 will have the total available
capacity of the shared cabin, C, as its booking limit, BLi. The seats protected from
class 2 and available exclusively to class 1 will be denoted S2'
The optimal protection level S2 for class 1 is the largest integer value of S2 that
satisfies the following condition:
EMSRi(S2) 2 f2 (5.15)
This optimal protection level is not a function of the lower fare class demand density
in the static case where nested buckets are involved. It is a function of the ratio of f2
to fi and of the parameters of the high-fare demand density assumed from historical
109
Figure 5.3: Total and Marginal Seat Revenue Curves
($/FLIGHT) =f.
EMSR .
($/SEAT)
SEATS
110
. Figure 5.4: Maximizing Expected Revenues, Two Class Example
(a) SOLUTION FOR NESTED RESERVATIONS SYSTEM
EMSR($/SEATk- - BL.
S, SEATS AVAILABLE
EMSR($/SEAT)
f1
f2
EMSR 1 (S ) =
EMSR 2 (S
This solution will maximize expected revenues in cases where the booking limit is
set only once at the start of the reservations process (static seat inventory control).
A class 2 request will be rejected only when BL 2 is reached, at which point the
expected revenue for all remaining seats will be greater than the class 2 average
fare. If class 2 requests never reach BL 2 , the unsold seats will be available for
unexpectedly high class 1 demand. In any event, the expected revenue per seat from
class 1 requests in excess of S2 is below f2, in the absence of additional information
on actual bookings for the future flight being managed. Total expected revenues
therefore cannot be increased by making BL 2 smaller. If BL 2 is increased, the
airline will lose the difference EMSRi(SI) - f2 in expected revenue for each seat
less than S2 protected for class 1.
The difference between this nested fare class solution and the "stand-alone" fare
class solution is apparent in Figure 5.4(b). In the latter, seats are allocated between
classes such that:
EMSR(Si) = EMSR 2 (S 2 ) (5.17)
at optimality, subject to the capacity constraint. The result is that the final seat
allocated to class 1 could well have an EMSR lower than f2, in which case unex-
pectedly high class 2 demand would be refused in favor of keeping that seat open
for possible class 1 demand even though it has a lower expected revenue in the class
1 inventory.
Solution methods that treat class 1 and class 2 seats as distinct inventories can
therefore produce less than optimal results when applied in a nested reservations
system context. The differences in expected revenues between the two solutions,
applied to nested fare classes, can be illustrated with a simple example. Table 5.1
summarizes the optimal limits and expected revenues for three solutions, for the
case of a two-class shared cabin with a capacity of 10 seats. Identical independent
Gaussian demand densities are assumed for the two classes. Fares are $100 and $60
for class 1 and class 2, respectively.
The "stand-alone" optimal solution allocates 6 seats to class 1 and 4 seats to class
2. When applied to distinct fare class inventories, the total expected revenue of this
solution is $684. When the "stand-alone" solution is applied to nested fare classes,
determining the total expected revenue requires an assumption to be made with
respect to the order in which bookings are made. In this example, it is assumed
that all class 2 seats are requested before any class 1 requests are made. This
assumption defines the lower bound on the estimate of total expected revenues in
112
Table 5.1: Optimal Limits and Expected Revenues
Optimal Allocation 6 4 10
Expected Revenue $465 $219 $684
Class 1 Protection 6 -
Class 2 Limit - 4
Expected Revenue $471 $219 $690
Class 1 Protection 5
Class 2 Limit - 5
Expected Revenue $477 $255 $702
113
a nested system, since any assumption involving a parallel booking process would
give class 1 requests greater access to class 2 seats.
Under the assumption that class 2 passengers book first, we must consider the
joint probabilities of not selling class 2 seats to class 2 passengers and selling these
seats instead to excess class 1 demand. Thus, application of the stand-alone solution
to nested fare classes will result in a total expected revenue higher than $684. The
increment is equal to the expected marginal seat revenues of the 7th, 8th, 9th and
10th seats available to class 1, weighted by the probability that a maximum of
3, 2, 1, and 0 seats, respectively, will be sold to class 2 passengers. Total expected
revenues therefore increase to $690, or by 0.9 percent, when the stand-alone solution
is applied to nested fare classes.
The total expected revenue will generally be higher for nested fare classes than
for distinct fare class inventories, given the same booking limits on the lower fare
class, as long as there is some probability that not all class 2 seats will be booked
by class 2 passengers and that class 1 demand will exceed the class 1 allotment. As
mentioned earlier, total expected revenues will be higher still when classes 1 and 2
are assumed to book concurrently.
The third and final case summarized in Table 5.1 involves the EMSR model
solution for the same simple example. The EMSR solution is different from the
stand-alone solution, protecting 5 seats for class 1 and limiting class 2 sales to 5
seats. The total expected revenue for the EMSR solution in a nested system is
$702, an increase of 1.7 percent over the previous case.
This simple example demonstrates how the EMSR solution for nested fare classes
can in fact generate higher total expected revenues than the application of the
"stand-alone" optimal allotments to a nested fare class system. The extent of this
revenue difference will depend on the demand densities and booking process as-
sumed. The logic of accepting all lower fare requests as long as the lower fare ex-
ceeds the expected marginal revenue of the higher fare class seat being "displaced"
suggests that the EMSR solution will generate at least as much in total expected
revenues as the stand-alone solution applied to nested fare classes, for a range of
demand and booking assumptions.
Extension of the EMSR model to more than two fare classes on a single flight
leg simply requires that more comparisons of expected marginal revenues be made
among the relevant classes. Looking first at a 3-class example, we assume that
fi > f2 > f/ and that these fare classes are hierarchically nested, as before. In
this problem, we must find not only the number of seats to be protected for class 1
relative to class 2, we must also establish protection levels for classes 1 and 2 relative
to class 3.
114
No change is required in the determination of S2. We need only protect seats for
the exclusive use of class 1 up to the point at which EMSR1(S') = f2, as before.
Only S2 seats will have an expected revenue greater than f2, meaning all remaining
seats may be made available to class 2. The optimal booking limit on class 2 (and
lower classes) thus remains:
BL 2 = C - S2 (5.18)
With 3 fare classes, however, additional revenue comparisons involving class 3 must
be made to determine BL3 . Using the EMSR approach, we determine two protection
levels:
respectively. The total protection level for classes 1 and 2 relative to class 3 can be
taken as the sum of the two, and:
BL 3 =C SI- S (5.20)
That is, all seats with an expected marginal revenue greater than f3 should be held
back from sale to class 3. Otherwise, any request for a class 3 seat may be accepted.
The values S' and S3 are used in the EMSR model to determine an optimal BLs,
but will not actually appear in the reservations system. In essence, the incremental
number of seats protected for class 2 above that required for class 1 alone is equal
to BL 2 - BL 3 . We shall refer to this number as the nested protection level for class
2, and denote it NP 2 . The nested protection level for class j is thus given by:
The nested protection levels can also be expressed in terms of the EMSR model seat
protection outputs, Sj3:
N P; = 3'+
S1- S3 (5.22)
isi i<j
These expressions for NP apply to all fare classes higher than the lowest fare
class. In the EMSR framework, the lowest fare class does not have a protection level
115
per se, but rather a booking limit equal to the number of seats that remain after all
upper classes have been protected. With a capacity of C seats and k fare classes,
then, the values of NPy must satisfy:
C= ( NP + BLk (5.23)
i<k
The EMSR protection levels S!, nested protection levels NP, and booking limits
BL,, are shown graphically for a three-class example in Figure 5.5(a). Figure 5.5(b)
shows the relationship between the booking limits as derived from the EMSR model
and the "stacked" configuration of total expected revenues by fare class in a nested
reservations system. Total expected revenues are shown as a function of total seats
available on the aircraft. As available seats are increased beyond S2 , seats become
available to both classes 1 and 2, and total expected revenues reflect a sharp increase
over those of Class 1 alone. The slope of the total expected revenue curve increases
suddenly at these points due to the increased probability of selling the marginal
seat made available to the lower fare class. The result is an "envelope" of stacked
expected revenue curves.
This analysis of total expected revenue as a function of available seats can also
be applied to aircraft sizing considerations. As shown in Figure 5.5(b), if operating
costs per flight as a function of seating capacity on an "elastic" airplane can be
estimated, an optimal seating capacity can be found. With three fare classes offered,
operating income will be maximized when marginal flight revenues equal marginal
flight costs, or where the total expected revenue curve has the same slope as the
operating cost curve (i.e., at the point C*). This slope is the marginal seat operating
cost.
Given that many U.S. domestic airlines make use of at least four fare classes to
manage their coach cabin seat inventories, we shall present the EMSR model solu-
tion to the nested 4-class problem. Six comparisons between upper class expected
revenue curves and lower class fare levels are required. The optimal S values satisfy
the following conditions:
116
Figure 5.5: EMSR Solution for Nested 3-Class Example
EMSR
($/SEAT)
SLOPE OF
OC(S)
(b) TOTAL EXPECTED REVENUES
R(S)
EXPECTED POPERATING
FLIGHT COST/
REVENUE FLIGHT
($/FLIGHT) OC(S)
117
The resultant optimal booking limits on each fare class are then:
BL1 = C (5.25)
BL 2 = C -Si
BL 3 = C -S -S2
BL 4 = C -S4 -S4 -S4
In the general case of k fare classes offered on a flight leg, the optimal values of
S' must satsy
EMSR 1 (S) = fi, i < j, j= 1,..., k (5.26)
The total number of comparisons required for k nested fare classes is given by:
k(k 2 1)
(5.27)
These optimal protection levels in turn determine the optimal booking limits on
each fare class j:
BLi = C - ZSj (5.28)
It is possible that one or more values of BLi derived from these equations might be
negative, in which case class j should not be offered at all if expected revenues are
to be maximized. In such a case:
BLi = MAX[0, C- E Sp (5.29)
'<
The EMSR model was presented in the context of a reservations system with
complete hierarchical nesting of fare classes that share a common seat inventory. The
decision rules can be applied to any number of nested fare classes with no change
to the model's basic structure. It is important to recognize that not all reservations
systems are nested in this way, or might not be nested at all. Nesting is preferable
in seat inventory management, because there is no difference in the physical seats
or on-board service being sold to different fare classes. Airlines without nested fare
class systems are denying themselves the flexibility of accommodating unexpectedly
high demand levels in high-fare classes and, in turn, are losing potential revenues.
118
5.3 Dynamic Applications of the EMSR Model
The EMSR model outlined in the preceding section provides a decision frame-
work for finding the revenue-maximizing booking limits on nested fare classes for
a single flight leg. In its simplest application, this decision framework provides a
solution to the static seat inventory control problem, in which booking limits must
be set for a future flight at a particular point in time prior to departure, usually
at the start of the reservations process. The static application requires estimates of
the demand densities (from a sample of historical flights) and of the relative revenue
levels associated with each of the fare classes to be offered on the future flight.
The same EMSR decision framework can be applied to a seat inventory control
context in which booking limits may be revised on a regular basis as the flight
departure day nears. In such a situation, additional information is available in the
form of actual bookings already accepted for the future flight. Because an actual
booking in any fare class will almost certainly (barring cancellations and no-shows)
translate into a revenue passenger occupying a seat, incorporating actual bookings
into the EMSR decision framework can reduce the uncertainty associated with the
estimates of expected demand used as input.
In developing a dynamic seat inventory control decision model that makes use
of the EMSR calculations, we will continue to make the simplistic assumptions with
respect to demand densities, refused requests, cancellations and no-shows, as out-
lined in the previous section. Furthermore, we add the assumption that there is no
relationship between the booking rates of different fare classes or among time peri-
ods before departure. As before, the reservations buckets are nested hierarchically
in descending order of fare level.
119
assessment of expected fare class revenues from that point in time, based on the
most recent available demand information for the flight leg.
In the previous section, the EMSR decision model derived the optimal booking
limits, BLi, for each fare class on the basis of the input values of f, and P;(Sj). For
the dynamic problem, the estimates of bookings to come must be generated from
densities of demand by fare class from a historical sample of requests made between
day t before departure and the day of departure. We define ri to be the number of
requests made for class i between days t and 0 before departure, meaning:
r! < r; (5.30)
by definition. The probability density of requests from day t onward is then p;(r!),
and the probability of receiving S or more requests for class i in the time remaining
to departure is -(S).
On any day t prior to flight departure, the inputs required by the EMSR model
are the average fare or revenue levels for each fare class, f;, which may or may not
remain constant over the booking period, and the estimates of t-(S) for all relevant
values of S, derived from the pi(r) densities. The optimal seat protection level for
class 1 relative to class 2 for the period remaining before departure is Sj-(t), such
that:
EMSR'[S2(t)] = f -(S ) = f2 (5.31)
This optimal protection level for day t can be used to find the revised optimal
booking limit on class 2, as follows:
For more than two fare classes, comparisons between the EMSR(S) values of
all upper classes relative to the average fare levels of lower classes are required, as
before. These comparisons involve the demand densities of future requests from the
current day t. Each comparison of a higher fare class i with a lower fare class j
generates an optimal value of S (t) that satisfies:
120
The revised booking limits for day t take actual bookings into account:
The nested protection levels, NPi(t), for successively lower fare classes are derived
as in the initial case, but with actual bookings included:
Figure 5.6 illustrates how actual bookings in a class receiving protection can
cause the booking limit on a lower class to be revised downward in a dynamic seat
inventory control process. The initial booking limit on Class 2, BL 2 , shown in
Figure 5.6(a), is reduced to BL 2 (t) in Figure 5.6(b), due to the actual bookings on
hand in class 1 at day t and the need to protect additional seats for the expected
future requests for class 1 before departure.
The EMSR decision framework has to this point overlooked the possibility that
a reservation might not translate into a passenger being carried on the fight in
question. With each accepted booking assumed to represent certain revenue, the
model has focused on the problem of managing the physical seat inventory on the
aircraft. As described in Chapter Two, there is some probability that a booked
passenger might not be carried on the flight for which the reservation was made.
Whether the original booking is cancelled prior to departure or the passenger simply
fails to appear at departure time, the outcome from the airline's perspective is the
121
Figure 5.6: Dynamic Application of EMSR Model, Two Nested Classes
fi
($/ SEAT)2
f2EREMSR
L2
0 C
2
t
2
(b 1 S2( t).-y -- y
0 C
SEATS
122
loss of a revenue passenger for that particular flight. This loss is the opportunity
cost of having removed a seat from the available inventory and then not receiving
any revenue for it.
The airline industry practice of controlled overbooking of flights above the phys-
ical capacity of the aircraft has evolved with the objective of minimizing such costs.
Overbooking analysis is performed to determine the extent to which a future flight
should be overbooked so as to minimize the sum of the lost revenues associated
with empty seats and the costs of denying boarding to passengers with confirmed
reservations.
Early work on this problem treated the overbooking level as a fixed number that
applied throughout the booking process for a future flight [59]. More sophisticated
models allow for bookings to be cancelled throughout the reservations process, such
that the optimal overbooking level for a future flight is a function of the time left
to departure. These overbooking models require estimates of the demand densities
and cancellation probabilities for future flights. Another approach was proposed by
Rothstein, who developed a model for maximizing passenger revenues subject to a
limit on the proportion of denied boardings [60].
For the purposes of seat inventory control, it is important to recognize the po-
tential interaction between the fare class mix of passengers booked for a flight and
overbooking, one which could have significant revenue implications for the airline.
In this section, the EMSR decision framework is extended to account for flight over-
booking. It is not the intent of this work to develop an overbooking model. Rather,
the objective is to illustrate how overbooking proportions by flight and even by fare
class can be incorporated into the EMSR framework for deriving revenue-maximizing
fare class booking limits.
With the introduction of the possibility that a booking made in fare class i
will not generate revenue fh with certainty, the EMSR decision framework must be
adjusted. The demand inputs required are still estimates of the densities of requests
123
for each fare class. The difference here is that each accepted request (booking) in a
fare class cannot be treated as if the revenue associated with that class will always
be realized. The problems of estimating request densities from actual booking data
will be discussed in Chapter Six. At this point, it suffices to note that the expected
revenue associated with accepting a request will be lower than the actual fare level
in that class, because of the possibility of cancellation or no-show.
The overbooking factors for a fare class determine the extent to which the ex-
pected revenue from a booking is reduced due to this uncertainty. Given overbooking
factors OV1 and OV2, the optimal protection level for class 1 from class 2, S2, must
satisfy:
P51(S2) - fih- = f2 - --V(5.37)
The revenue levels of the two classes are in essence "deflated" by the assumed
overbooking factors. We are thus reducing the expected marginal revenue of each
incremental request for class 1 for which the possibility of protection is being eval-
uated.
The value of S2 that satisfies (5.37) is the protection level for class 1 from class 2,
expressed in terms of the number of reservationsspaces (as opposed to physical seats)
that should be protected for exclusive use of class 1 passengers. Because the demand
inputs to the EMSR framework for all fare classes will be in terms of expected
requests, the calculated protection levels for all upper fare classes will be in terms
of reservations spaces, with the relevant overbooking factors already incorporated.
The generalized decision rule for the EMSR framework with overbooking factors is
to protect S' seats for class i from class j such that:
1 -1 1
EMSRi(S;) - -- = P,(S;) - fi - -= fi - (5.38)
OV, OV, 0V1
The derivation of optimal overbooking limits, BL!, from this revised EMSR de-
cision rule is complicated by the fact that the optimal protection levels are expressed
in terms of reservations spaces (overbooking factors included), while the capacity of
the aircraft is in terms of physical seats. The simplest case occurs when all the OV
values are equal across fare classes, so that the OV factors drop out of the above
formulation, reducing it to the original formulation in (5.26). The overbooking limit
on the total capacity of the shared aircraft cabin, C, is simply:
Booking limits for each subordinate fare class in the coach cabin may be derived
after an overbooking target is established for the total capacity of the cabin. The
124
protection levels generated by the EMSR model are for reservations apaces rather
than physical seats, and the overbooking limits on each fare class j are given by:
The net result is that each fare class may be overbooked by the same percentage
and C* will be the same regardless of the fare class mix actually booked for any
particular flight.
The fact that fare classes are designed to appeal to different air travel demand
segments suggests that passengers booked in each fare class might exhibit different
no-show behavior. As discussed in Section 2.2, it is plausible that passengers in
lower fare classes will be more likely to show up for booked flights than those in
higher fare classes for several reasons, the most important being the cancellation
penalties associated with the lowest excursion fares. Few airlines currently have the
detailed no-show data required to confirm this hypothesis empirically.
The application of different overbooking levels, OV, to each fare class j would
be relatively straightforward in a non-nested fare class structure, as each BL! would
simply be the number of reservations spaces allocated to class j. In a nested struc-
ture, it is possible for all seats to be booked in the highest fare class, for example,
in which case the total overbooking limit derived from different OV values might
not be correct. That is, a total cabin overbooking limit based on an expected fare
class mix of bookings might not be high enough in the extreme case where all seats
are taken by the highest fare class with the highest no-show rate.
In a nested fare class structure with protection levels Sj derived from the EMSR
model, it is not possible to determine analytically the fare class mix that can be
expected, in the absence of assumptions concerning the relative booking rates of
each of the fare classes. Application of different overbooking factors to each fare
class at the start of the booking process, before any bookings are accepted, requires
some estimate of an ultimate fare class mix of bookings. If we assume that, for
a high demand flight, the lowest fare classes will book up first, then the simplest
estimate of total bookings by fare class, given the EMSR protection levels, would
be the nested protection level derived for each successively lower fare class, NP.
125
cabin. The difference is then inflated by 0V4 to find the overbooking limit on class
4:
1 1 1
BL* = OV 4[C -- NP1 - -- NP 2 -- - NP] (5.41)
o0v 0V2 0V3
Note that, with equal OVi values for all fare classes, this equation reduces to (5.40),
for j = 4.
Having found BL*, we can derive the overbooking limits on each of the upper fare
classes by adding the incremental protection levels, as calculated from the EMSR
decision model:
The total authorized overbooking limit, in the absence of actual bookings for
the future flight, is thus determined by the total protection level calculated for the
upper fare classes. The number of physical seats that remain for the lowest class
must be determined, inflated by the relevant overbooking factor, and added to the
summed upper class protection levels. It is possible that the sum of all the upper
class protection levels, when deflated to physical seats, will still exceed the physical
capacity of the shared cabin. In this case, BLi must be constrained to be greater
than or equal to zero. Furthermore, in the subsequent derivation of BL; for upper
classes, the same constraint applies.
These overbooking limits take into account the possibility of different no-show
rates by fare class, and are based on an estimate of the ultimate fare class mix
of bookings. As bookings are accepted in different fare classes, these overbooking
limits can be refined to reflect the actual fare class mix for a particular flight. The
dynamic adjustment capabilities of the EMSR model, as described in the previous
section, enable us to revise both the fare class protection levels and overbooking
limits as departure day nears.
Recall that the dynamic application of the EMSR model revises fare class book-
ing limits for the physical seats on the aircraft periodically to account for actual
126
bookings accepted and to protect additional seats as necessary. Without no-shows
or overbooking, the booking limit on class j at day t before departure was given by:
With the introduction of overbooking factors by fare class, both the protection
levels and actual bookings are expressed in terms of reservations spaces, not physical
seats. The total incremental seat protection level for each successively lower fare
class at day t can be defined as the sum of the nested protection level calculated from
the EMSR model for the period remaining to departure, NPi(t), and the number
of bookings already on hand in that fare class, b .
Starting with class 1, the EMSR formulation generates a protection level for the
period from day t to departure, S2(t), based on estimates of requests still to come in
class 1. Actual bookings are added to this protection level, and the total protection
for class 1 on day t is then the difference between the overbooking limits on class 1
and class 2:
BL*(t) - BL;(t) = S2(t) + bt (5.45)
For class 2, the nested protection for future requests above that required for class 1
is given by NP 2 (t). This incremental protection level and actual bookings in class 2
are added to give the difference between the overbooking limits on class 2 and class
3:
BL;(t) - BL*(t) = NP2 (t) + b2 (5.46)
Similarly:
BL;(t) - BL* (t) = NP(t) + b' (5.47)
To this point, the differences between adjacent overbooking limits have been
defined in terms of nested protection levels for future requests plus actual bookings
already on hand for the fare class in question. The overbooking limit on the lowest
fare class, class 4 in this example, depends on the total of the protection levels and
actual bookings in the upper fare classes. Thus, we inflate whatever seats remain
unprotected and unsold by OV4. Along with upper class nested protection levels,
actual bookings in the top 3 classes must be deflated by their respective OV values
to determine how many physical seats remain unprotected and unsold. The optimal
overbooking limit on class 4 on day t is then given by:
Note that class 4 does not have any "protection level" for its own use, and that
BL (t) is a function of protection levels and bookings in upper classes, relative to
127
C. Actual bookings in class 4 are also not included in the derivation of BL (t). It is
possible that b' could prove to be higher than BLi(t) at some point in time prior to
departure in cases where BL (t) has been revised downward in the dynamic adjust-
ment process. Most airline reservations systems will allow such an inconsistency to
be retained so that cancelled class 4 bookings do not "open up" spaces for additional
class 4 bookings.
The coach cabin physical capacity, C, does not enter into the overbooking limit
computation except in conjunction with class 4. This is a result of the EMSR
approach of allowing any unprotected seats to be made available to the lowest fare
class. The complete expression for the total cabin overbooking limit as derived from
the above equations is as follows:
This formulation for C*(t) can readily be modified to accept different OVi values
as a function of time before departure. Several overbooking models developed in the
past make use of cancellation probabilities to adjust overbooking factors downwards
as departure day approaches. If such an overbooking approach is to be used, the
relevant OVi(t) factors may be included in the above formulation.
Having extended the EMSR formulation to incude the possibility that an ac-
cepted booking might not translate into a revenue passenger carried, we now turn
to the issue of passenger choice shifts during the booking process. As described in
Section 2.2, a refused reservations request for a particular future flight and fare class
does not always result in a booking loss to the airline. Depending on the individual
consumer's choice process, the unavailability of a desired flight and fare class can
lead to:
128
For the purposes of managing the seat inventory for a single flight leg, the
probability of interest is that of a vertical shift in fare classes on the same flight.
The probability that a passenger refused a request for fare class i will accept a
booking in the next highest fare class (i - 1) was defined in Section 2.2 to be Pi(v).
Although the probability of horizontal shift, P(h), and the overall recapture rate
is also important to the airline wishing to maximize system revenues, our focus on
flight leg inventories requires us to assume that a horizontal choice shift is equivalent
to a booking loss. In this section, the EMSR formulation is extended to include the
possibility of vertical shift on the part of the refused passenger. Only shifts from
the requested fare class to the next highest fare class will be considered. We assume
that, given a competitive market and a fare structure that segments market demand
effectively, a vertical choice shift of two or more fare classes will be unlikely.
The EMSR formulation finds the optimal number of seats to be protected for
the exclusive use of the highest fare class, and in turn for each successively lower
nested fare class. There is some probability that a refused class 2 request will in fact
become a class 1 booking, due to a vertical choice shift on the part of the denied
consumer. Defined as P2 (v), this probability is assumed to apply equally to all class
2 requests, regardless of time before departure.
From the basic EMSR formulation, a protection level for class 1 from class 2 of
S2 will still be required. Additional seats protected for class 1, V21, can be taken
either by a refused class 2 passenger or a class 1 passenger. The expected revenue
from a class 1 passenger in the V21th additional protected seat is:
If the upgrade probability, P2(v), is greater than zero, the incremental expected
revenue associated with potential vertical shifts from class 2 may be realized if the
seat is not purchased by a class 1 passenger.
The combined expected marginal seat revenue for the V2 'th seat protected for
class 1 is thus equal to fi multiplied by the probability that a class 1 request will
129
be received for that seat or a vertical choice shift is accepted, given that BL 2 is
reached. The optimal value of V' must therefore satisfy:
where S2 is the protection level for class 1 in the absence of vertical choice shifts,
as before. Factoring and re-arranging the probability terms in the above equation
produces the following equivalent condition:
The combined expected revenue from each additional seat protected for class 1 will
be greater than or equal to f2, given that BL 2 is reached. If BL 2 is not reached,
this additional protection will have no impact.
The total protection level for class 1 from class 2 is then (S2 + V1) and the
booking limit on class 2 (without overbooking) is:
BL 2 = C 1
- S -V
2 (5.54)
Overbooking factors may be incorporated as before, in which case the total EMSR
protection level (S2 + V1) would be treated the same as S2 in the calculation of
overbooking limits.
BL 3 = C - S3 - S3 -V3 (5.56)
In general terms, the incremental protection required for any class i to account
for the probability of vertical choice shift from class j i+1, is V' such that:
The general formula for the resultant booking limit on class j is:
- 130
As before, the nested protection levels are defined by the differences between
successively lower booking limits. Overbooking factors may be applied to these
protection levels to find the optimal overbooking limits, as described in the previous
section, with (Sj + Vj') replacing the protection levels for the relevant classes.
The impact of including one or more P(v) values in the EMSR formulation
will be an increase in the protection levels for each of the higher fare classes. Each
lower fare class will see its booking limit decrease by the incremental protection level
required to account for the possibility of vertical choice shifts to the next-highest fare
class. The magnitude of this decrease will depend on the relative magnitudes of the
P(v) values estimated or assumed, highlighting the importance of this probabilistic
element to the EMSR framework.
The previous sections of this chapter have presented the single-leg Expected
Marginal Seat Revenue model in terms of the mathematical formulations required
to find optimal fare class seat protection levels and booking limits. This final section
illustrates the application of these mathematical formulations to a future flight leg
departure. Based on historical demand and revenue data for a hypothetical flight
leg, the outputs of the EMSR model are presented for the static case (initial booking
limits), in a dynamic limit adjustment example, with overbooking factors included
in the calculations, and, finally, with upgrade probabilities taken into account.
The example used throughout this description is that of a medium-haul flight leg
to be operated with a B-737 aircraft. Four fare class inventories (Y, M, B, Q) are
to share the 107 seats in the coach cabin, and the reservations system is assumed to
be completely nested from the lowest to the highest fare class. Table 5.2 shows the
demand and revenue estimates used as inputs to the EMSR framework to establish
"initial" fare class booking limits. These initial limits are set before any bookings are
accepted for the future flight, or at least well before the number of actual bookings
in any fare class would threaten the most conservative authorized booking limits.
The prorated flight leg revenue estimates by fare class shown in Table 5.2 reflect
the hierarchy and range of revenue levels characteristic of an actual competitive
market. These revenue inputs will remain constant throughout this example, unlike
the demand inputs. The demand estimates shown are for total expected requests
by fare class, as would be derived from a historical sample of past operations of the
same flight leg under similar conditions. In the dynamic application of the EMSR
model, these demand estimates will change as the time to departure decreases. The
131
Table 5.2: EMSR Example - Initial Booking Limits
INPUTS:
Y M B
Total Expected Requests 20.3 33.4 19.3 29.7
Standard Deviation 8.6 15.1 9.2 13.1
Average Revenue ($) 105 83 57
CALCULATIONS:
Y M B
Nested Protection 14 33 27
Nested Booking Limits 107 93 60 33
132
means and standard deviations in each case define an assumed Gaussian demand
density for each fare class.
Table 5.2 also shows the EMSR protection levels, nested protection levels and
initial booking limits for the flight. These outputs are derived from preliminary
estimates of total expected requests by fare class, and do not take into account
actual bookings received. The limits would apply to this flight until bookings to
date and estimates of future requests begin to differ significantly from the original
estimates of total requests. That is, the dynamic revision process should begin
when, according to historical information, more bookings should have been received
in one or more of the upper fare classes. It should also be initiated when actual
bookings in one or more of the upper fare classes exceed or fall below the number
expected up to that point in time. Because the EMSR model does not protect seats
for the lowest fare class, expected booking rates for that class need not be taken
into account, and actual bookings will be limited to the initial authorized level (i.e.,
33) until the first revision run is made.
In this example, day 35 before departure was judged to be the first time that
booking limits should be revised on the basis of a change in the estimates of future
requests. The estimates of future requests used as inputs reflect the assumption of no
correlation between actual bookings and future requests. The possibility of deriving
conditional estimates of future requests based on actual bookings is explored in
Chapter Six.
Table 5.3 shows the mean number of requests for each fare class expected from
day 35 to departure, which, by definition, will be lower than or equal to the initial
total expected requests. New protection levels are derived from the EMSR frame-
work to accommodate this "remaining" demand at day 35. Actual bookings on hand
at day 35 are then added to these protection levels to find the "total protection"
required for each upper fare class. The actual bookings shown in Table 5.3 for day
35 reflect a higher than expected demand for M class. The revised booking limits
are thus lower than the initial limits for both B and Q classes, due to the M-class
demand.
Table 5.4 shows the progression of expected requests still to come and actual
bookings on hand, along with revised protection levels and booking limits for a dy-
namic application of the EMSR model on days 28, 21, 14 and 7 before departure.
Higher than expected M-class bookings in each period lead to successive increases
in the total M protection derived by the EMSR framework and, in turn, successively
lower booking limits on B class. The total protection required for B class remains
relatively stable. The booking limits on Q class decrease due to the increased pro-
tection for M and the stable protection for B.
133
Table 5.3: EMSR Example - Day 35 Booking Limit Revision
Y M B Q
INPUTS:
Nested Protection 13 27 22 -
Actual Bookings 1 9 5 18
Total Protection 14 36 27 -
Nested Booking Limits 107 93 57 30
(Revised)
Seats Available (Nested) 74 61 34 12
134
Table 5.4: EMSR Example - Dynamic Revisions Day 28 to Day 7
Y M B Q
DAY 28:
Expected Future Requests 16.2 23.8 12.6 4.1
Standard Deviation 7.9 13.3 5.5 6.6
Nested Protection 10 24 20 -
Actual Bookings 3 15 8 25
Total Protection 13 39 28 -
DAY 21:
Expected Future Requests 12.9 22.0 11.0 3.3
Standard Deviation 6.9 11.9 6.2 7.1
Nested Protection 8 22 18 -
Actual Bookings 6 19 10 27
Total Protection 14 41 28 -
DAY 14:
Expected Future Requests 9.7 19.4 7.9 0.8
Standard Deviation 5.8 11.6 5.9 5.4
Nested Protection 6 18 15 -
Actual Bookings 6 24 13 27
Total Protection 12 42 28 -
DAY 7:
Expected Future Requests 6.0 17.3 5.8 0.3
Standard Deviation 3.7 10.9 4.8 4.3
Nested Protection 4 14 13 -
Actual Bookings 8 30 15 27
Total Protection 12 44 28 -
135
The downward revision of the Q booking limit demonstrates how relatively in-
frequent revision runs can allow too many seats to be sold in a low fare class. The
Q booking limit is revised from its previous level of 27 down to 24 on day 21, by
which time 27 Q-class bookings have already been accepted. The result is essentially
"negative availability" in Q class, a condition which worsens closer to departure. In
an actual booking process in which cancellations occur, bookings on hand in Q class
might decrease to the authorized limit. No additional bookings could be accepted
in Q class, unless Q cancellations reduce actual bookings to less than the relevant
Q booking limits, at which point Q class would "open up" again.
The occurrence of "negative availability" in this example stems from the assumed
interval between revisions of 7 days. The booking process is in essence left unat-
tended between revisions. More frequent revisions would enable the EMSR model
to be more responsive to changes in expected booking trends and actual bookings
on hand. The problem with more frequent revisions is one of deriving more frequent
accurate estimates of future requests from historical data. Frequent revisions would
also not eliminate the possibility of too many low-fare seats being sold early in the
process, with the result that unexpectedly high demand for higher fare classes might
not be accommodated. Because the EMSR protection levels are based on maximiz-
ing expected revenues, it is inevitable that the booking limits for an individual flight
might not result in optimal revenues.
Dynamic application of the EMSR framework allows the airline to manage the
booking process, taking into account bookings on hand and expected requests still
to come. This management process becomes more responsive and precise with more
frequent revisions driven by demand estimates from more detailed historical data.
Demand estimates based on forecasts rather than simple historical averages would
provide further improvement, as will be discussed in Chapter Six.
To illustrate how overbooking factors may be incorporated into the booking limit
calculations, we employ the same ffight leg example as above. The application of
different fare class overbooking factors is demonstrated for the calculation of initial
limits (without actual bookings) in Table 5.5(a), and for the day 21 revision in Table
5.5(b), as one example of how actual bookings affect the overbooking limits. Table
5.5(a) shows the assumed overbooking factors by fare class and the calculation of
optimal initial overbooking limits for the example flight leg. In the absence of actual
bookings, the nested protection levels, NPi, are added to the overbooking limit on
the lowest fare class (Q), which is derived as explained in Section 5.4. Note that
the nested protection levels are slightly lower than those in Table 5.2, because each
of the upper fare class demand estimates have been subjected to adjustment by an
overbooking factor that is higher than that assumed for Q-class requests. Fewer
"reservations spaces" are thus protected for these upper classes. Furthermore, the
136
Table 5.5: EMSR Example - Overbooking Limits
M B
Overbooking Factors, OV 1.30 1.25 1.20 1.10
Nested Protection, NP 13 31 25
Overbooking Limits, BL; 126 113 82 57
B
Overbooking Factors, OVr 1.30 1.25 1.20 1.10
Nested Protection, NP(21) 7 21
Actual Bookings, b21 6 19
Overbooking Limits, BL;(21) 127 114 74 50
137
transition from physical seats to reservations spaces, while keeping the same inputs
of expected requests as used initially, frees up both seats and reservations spaces
for Q class, relative to Table 5.2. The "weighted" overbooking factor on the entire
shared coach cabin at this point is 1.178.
Table 5.5(b) shows the calculation of revised overbooking limits at day 21 prior to
departure, taking into account the seat protection levels from the EMSR model and
actual bookings. Actual bookings are added to the nested protection levels, and
overbooking limits are calculated starting with BL . The weighted overbooking
factor on the coach cabin increases to 1.187, primarily because the M-class bookings
are higher than expected.
The final step in this example of EMSR results involves the inclusion of upgrade
probabilities to account for vertical choice shifts in the consumer decision process.
The upgrade probabilities are included in the calculation of EMSR seat protection
levels for the initial case in which no overbooking factors are assumed (Table 5.2).
Table 5.6 shows how these initial limits would change if we could estimate the
probability of vertical choice shift from each of the three lowest fare classes. The
upgrade probability is assumed to be lowest for shifts from Q to B, higher to shifts
from B to M, and highest for shifts from M to Y.
Relatively low upgrade probabilities are assumed for Case 1 in Table 5.6. The
nested booking limit on M class is nonetheless reduced by 2 seats when compared
with the initial limits from Table 5.2, and by a total of 4 seats for B class. The
booking limit on Q drops by only 2 seats, because Q class requests are assumed to
have a lower probability of upgrading. Case 2 shows the results when the upgrade
probabilities are assumed to be higher. The nested booking limits decrease further
for all fare classes lower than Y, with the protection for B class remaining constant
due to the low probability of Q upgrades relative to B and M upgrades.
This chapter has presented the Expected Marginal Seat Revenue framework for
determining optimal protection levels and booking limits in a nested multiple fare
class reservations system for a single future flight leg. The basic model was extended
to dynamic applications, to include overbooking factors, and to take into account the
probability of vertical choice shifts on the part of the consumer denied a request for
a low fare class. The model and its extensions were applied to a hypothetical future
flight leg and a set of historical demand and revenue data. The EMSR framework is
based on certain assumptions with respect to the demand for different fare classes
and over time. These assumptions and the sensitivity of the model to them are
examined in the following chapter.
138
Table 5.6: EMSR Example - Initial Limits With Upgrade Probabilities
Y M B Q
CASE 1:
Upgrade Probabilities, Pi(v) - 0.30 0.20 0.10
EMSR Model Results:
Nested Protection Levels 16 35 25 -
Nested Booking Limits 107 91 56 31
CASE 2:
Upgrade Probabilities, P(v) - 0.40 0.30 0.20
EMSR Model Results:
Nested Protection Levels 17 37 25 -
Nested Booking Limits (107| 90 53 28
139
Chapter 6
The preceding chapter presented the Expected Marginal Seat Revenue decision
framework, its revenue-maximizing rationale, and the mathematical formulations
that determine optimal seat protection levels and fare class booking limits. Its em-
phasis was on how the EMSR framework can be used to manage the booking process
for a future flight leg dynamically, and examples of results based on hypothetical
data demonstrated the model's application. The intent was to describe the EMSR
model and its outputs in a relatively simple (i.e., single flight leg) context.
In this chapter, the EMSR model is examined in greater detail, specifically the
effects of input data and assumptions on the validity and performance of the model.
The demand assumptions required in practical applications of the EMSR model
are considered first. The sensitivity of model outputs to variations in the input
variables (both demand and revenues) is then examined. The chapter concludes
with a discussion of the potential application of the EMSR approach to the "virtual
nesting" seat inventory control systems being developed by several major airlines.
Throughout the presentation of the basic EMSR decision framework and its
extensions to include overbooking and consumer choice shift probabilities, no explicit
mention was made of the assumptions concerning the demand for different fare
classes. The model was presented under the assumption that there exists some
"correct" estimate of the demand density for each fare class for each point in time
140
at which the EMSR model is to be run. This density represents the random variation
around the expected number of requests which cannot be explained by systematic
patterns in demand such as seasonal, day of week, or time of day.
In practice, it is the process of estimating the demand densities for total expected
requests or requests to come for each fare class that triggers the need to make as-
sumptions about the characteristics of these densities. Analysis of empirical data
from historical information will inevitably comprise a major part of this estimation
process, and will contribute to the airline's estimate of the density of expected re-
quests by fare class for input into the EMSR framework. The estimation process,
including the determination of the most relevant historical sample from which es-
timates should be derived, the separation of systematic variation from stochastic
variation, and perhaps the adjustment of preliminary estimates through the use of
forecasting methods, can in fact become a far more complicated task than actually
deriving optimal limits from the EMSR model.
In practice, this ideal market scenario might not be replicable or might change
with changing airline marketing and pricing policies. Making use of the EMSR
framework thus requires the airline to make assumptions about the characteristics
of the estimated demand densities to be used as inputs to the model, and to consider
the possibility that the fundamental assumption of distinct and separable demands
for different fare classes might not hold.
The specific estimation issues that must be addressed on the demand side of the
EMSR equation include:
1. The shape of the probability density assumed for each estimate of expected
requests;
2. The extent to which any correlation between demand levels in different fare
classes exists and should be accounted for in the EMSR model;
3. The extent to which correlation between booking activity in the same class
over time is important to the dynamic application of EMSR.
This section examines each of the above issues briefly, considers their impact on the
EMSR model and presents some empirical evidence in each case.
141
6.1.1 Demand Density Shapes
No explicit mention of the assumed shape of the demand density was made in
Chapter Five simply because virtually any probability density function will work in
the EMSR formulations. As long as EMSR,(S) is decreasing in S, the comparison
of the expected marginal revenue from protecting a seat for possible sale in a higher
fare class with the certain revenue from the relevant lower fare class will result in
an optimal protection level. Recall that the value of EMSR(S) is determined by:
where
P,(S) = P[r; > S] (6.2)
For any probability density function, the probability of receiving ezactly S requests
for fare class i is by definition greater than or equal to zero, meaning that the
cumulative probability of receiving more than S requests will always be a decreasing
function of S.
For the purposes of deriving optimal protection levels with the EMSR frame-
work, an assumption concerning the shapes of the demand densities is nonetheless
required. As mentioned previously, Boeing studies have in the past assumed Gaus-
sian (normal) densities of total demand for a particular flight operated on a given
day of the week over a homogeneous period of time [61]. The EMSR diagrams in
the previous chapter were in fact drawn to approximate Gaussian demand densities.
For the EMSR model, however, we need estimates of the densities of expected
requests by fare class, as opposed to the total requests for a flight. Previous empirical
research by this author addressed the question of whether demand densities by fare
class differ significantly from total flight demand densities, at least to the point
of making the Gaussian assumption invalid [62]. Day of departure reservations
totals by fare class were analyzed from a sample of Trans World Airlines (TWA)
transcontinental non-stop flights in 1983. The objective of the study was to accept
or reject the assumption of Gaussian demand densities for each fare class by using
a sample of total reservations by fare class as a proxy for total requests. All flights
for which fare class booking limits were reached, thereby constraining the estimate
of total requests, were edited from the original sample.
142
simply by aggregation over dissimilar periods of time. As mentioned at the start
of this discussion, the estimated density of demand for a particular future flight
and fare class should represent the stochastic or unpredictable variation in expected
requests, and should not reflect systematic variation over time or across markets.
Disaggregation of the dataset to a more homogeneous time period (e.g., low sea-
son) by flight number and day of week allowed a more detailed statistical evaluation
of the distribution of data points within each subset. Extreme values were removed
from each subset when they could reasonably be associated with travel peaks sur-
rounding major holiday periods (e.g., Thanksgiving ). Many of the smaller sample
subsets conformed reasonably well to the Gaussian model, or at least did not deviate
enough from the model to permit statistical rejection of the normality assumption.
Combining similar days of the week to increase sample sizes allowed goodness of fit
tests to be performed. In general, in the absence of capacity constraints (either fare
class booking limits or physical capacity), the data fit the Gaussian model well.
An important finding of this analysis was that fare class demand density shapes
as determined by the standard deviation relative to the mean (i.e., the coefficient
of variation or "k-factor") differed noticeably among days of the week, even for the
same fare class. Furthermore, removing outliers from day of week subsets resulted
in a significant reduction in k-factors, in many cases to levels lower than the 0.33
approximation that has traditionally been assumed for total flight demand. Table
6.1 shows the differences in means and coefficients of variation by day of week for the
"original" and "edited" subsets of full fare coach class (Y) reservations from three
flights. The general pattern demonstrated is one of smaller k-factors on high demand
days of the week and vice versa . For the purposes of seat inventory control and
the EMSR framework, then, it is important that airlines recognize the systematic
patterns in demand by fare class across days of the week for a particular flight leg,
not only in terms of the level of demand (mean), but also in terms of its variation
(standard deviation) relative to that mean.
The conclusion of this analysis of demand densities by fare class was that the
empirical evidence did not allow statistical rejection of the assumption of Gaussian
densities. The analysis highlighted the importance of deriving an estimate of the
density that represents the random variation in demand for a particular fare class
on a given future flight by disaggregating historical data to remove as much of the
systematic variation as possible.
The fundamental assumption of distinct and separable demand for each fare
class on a future flight leg is reflected in the EMSR framework by an assumed
143
Table 6.1: Editing of Day of Week Data Subsets
Original X 28 28 35 30 45 8 36
k 0.24 0.35 0.33 0.38 0.35 0.36 0.32
Edited X? 28 30 39 31 47 8 37
k 0.24 0.29 0.19 0.22 0.29 0.30 0.24
Original X 43 28 34 37 48 23 39
k 0.33 0.36 0.41 0.44 0.33 0.52 0.42
Edited X 45 29 33 37 53 23 37
k 0.22 0.31 0.27 0.22 0.22 0.38 0.38
Original X 50 50 52 53 72 36 63
k 0.40 0.28 0.28 0.29 0.27 0.52 0.24
Edited X 49 51 56 56 76 34 65
k 0.29 0.27 0.19 0.20 0.20 0.44 0.28
144
independence (i.e., no correlation) between the total expected demand levels in each
of the fare classes. The possibility that individual denied requests for a low fare class
will ultimately be "upgraded" to the next higher fare class can be incorporated into
the EMSR model, as described in Section 5.5. Nonetheless, the EMSR approach is
based on the notion that there exists an initial or "natural" demand for each fare
class.
This subsection considers the impact of correlated fare class demands on the
EMSR model outputs, presents some empirical evidence related to this issue, and
discusses approaches to overcoming the problems presented by such correlation,
should it exist. We begin this discussion by noting that under the assumption of
no correlation between the demand densities for, say, classes 1 and 2, the optimal
protection level for both classes from class 3 could be derived from the EMSR
equivalently by assuming a joint demand density for the sum of class 1 and class
2 demands. As an example, assume that the class 1 and 2 individual demands are
identically distributed, as described in Table 6.2(a), and that the relative revenue
levels of the three classes are $100, $80, and $60, respectively.
If we run the EMSR model for the individual densities for classes 1 and 2, we
find the total protection from class 3 to be:
S +3 = 18+17= 35 (6.3)
To determine the effect of correlation between these two demand densities on the
EMSR protection levels, we can input a joint density of demand to find the total
protection from class 3. If no correlation is present, then the parameters of the
joint density, assuming a normal distribution, are simply the sums of the individual
density means and variances:
where Fi is the estimated mean expected demand for class i and &2 is the estimated
variance of demand for class i.
The problem with using a joint demand density for two fare classes in the EMSR
model involves determining the appropriate revenue level, fA+2, to be used in cal-
culating the expected marginal revenue of protecting an additional seat from class
3. In the example shown in Table 6.2, a simple average of class 1 and 2 revenues
is used as an estimate of the revenue associated with the joint demand density. No
weighting was performed because the individual densities are identical. The prob-
lem of joint revenue levels becomes more complicated when this is not the case.
145
Table 6.2: Impact of Correlation of Fare Class Demands
+1.0
+ 0.75
+ 0.5 13.9
+ 0.25 12.65
11.3 35
- 0.25 9.8 35
- 0.5 8 36
- 0.75 5.66 37
- 1.0
146
The EMSR total protection for the joint demand density in this case of no correla-
tion turns is 35 seats, the same as the summed protection in the case of individual
demand densities.
If the correlation between the demand levels for classes 1 and 2 is in fact signif-
icantly different from zero, the mean of the joint density will still be the sum of the
individual means, but the joint variance will be given by:
where Cov(ri, r2) is the covariance between rl and r2, which determines the value
of the correlation coefficient, p:
Cov(ri, r 2 )
p = . .(6.7)
In Table 6.2, a range of values for the estimated correlation coefficient and the
corresponding EMSR joint protection levels are presented. The lower and upper
limits on the joint protection level are defined by perfect positive and negative
correlation, respectively. The effect on optimal protection levels of significant corre-
lation between the demand densities is determined by the standard deviation of the
joint demand density for classes 1 and 2. With positive correlation, the standard
deviation of the joint density is higher than it would be with no correlation, and the
optimal protection levels are lower due to this increased uncertainty. With strong
negative correlation, the standard deviation of the joint density drops, causing the
optimal protection levels to increase. In the example shown, perfect negative cor-
relation results in a zero standard deviation for the joint density, and the optimal
EMSR joint protection level is simply the sum of the mean demands.
Given that significant correlation between fare class demands will affect the
optimal seat protection levels, two questions should be addressed:
2. How can the EMSR demand inputs be adjusted to account for such correla-
tion?
A sample of total reservations by fare class from flights in three O-D markets over
a six-month period was used in the analysis. A low season period was selected to
minimize the number of flights that would have to be edited from the sample because
147
of fare classes having reached their booking limits. This editing was necessary to
isolate the data points that were not constrained by the imposition of booking limits.
The objective of the analysis was to determine whether any significant correlation
between fare class demand levels could be found for the types of homogeneous sample
subsets that would be used to estimate demand densities for the EMSR model in
practical applications.
Correlation matrices were generated for each O-D market and a wide variety of
subsets of the total data sample in each market. As was the case with the test of the
normality assumption for estimated demand densities, this correlation analysis could
only be applied validly to sample subsets in which little or no systematic variation
remained. As an example of why systematic variation would bias the estimated
correlation coefficients, consider the case of grouping two dissimilar days of week
into one sample subset-a very high demand day and a very low demand day, in
all fare classes. A correlation matrix for this sample subset would inevitably show
very high positive correlation coefficients between fare classes because low class 2
demand would always be paired with low class 3 demand on one day of the week,
while high class 2 demand would be paired with high class 3 demand on the other
day of the week.
Table 6.3 summarizes the results of this correlation analysis for several of the
subsets tested. The general findings, as reflected by the examples shown, were a
preponderance of zero or insignificant correlation coefficients between fare class de-
mand levels. A few significant positive and negative results were obtained, but with
no apparent pattern. Aggregation of data subsets to include different days of the
week and/or different flights in the same market resulted in significantly positive
correlation between fare class demands in virtually all cases tested. Disaggregation
of data to eliminate systematic variation, on the other hand, led to almost no evi-
dence of significant correlation, positive or negative. It should be pointed out that
at least some of this absence of significant correlation is attributable to the small
sample sizes of the disaggregated subsets.
The implication for the EMSR model is that it may be used as described in
Chapter Five on individual fare class demand densities, unless empirical analysis
148
Table 6.3: Empirical Analysis of Fare Class Demand Correlation
October
February
October, Flt. 048
October, Flt. 354
Flt. 048, Mondays 8 0
Flt. 048, Wednesdays 12 0
Flt. 048, Fridays 8 0
October
February
October, Flt. 593
October, Flt. 595
October, Tuesdays 10 0 0 -
October, Thursdays 10 0 0 0
October, Sundays 8 0 0 0
149
reveals significant and consistent correlation patterns between fare class demand
levels. On the other hand, the presence of some significant correlation levels even in
the disaggregated subsets shown in Table 6.3 and the presence of others not statis-
tically significant but substantially different from zero in the disaggregate matrices
suggests that it might be necessary to take whatever small amount of correlation
that exists into account. Forecasting methods that make use of dependent variables
related both to time (systematic variation) and actual demand in other fare classes
to predict demand for a given fare class might be used to incorporate the correla-
tion of fare class demands into the EMSR inputs. We shall return to the issue of
reservations forecasting at the end of this section.
150
independence assumption on the EMSR framework, presents some empirical analy-
sis results, and discusses possible solutions to the problem.
In the initial (static) application of the EMSR framework, the density of total
expected requests by fare class is estimated by the mean, Fi, and standard deviation,
01, derived from a historical sample. In the dynamic application of the model, these
inputs are replaced by an estimate of future requests and actual bookings on hand.
This revised estimate includes b, the number of bookings on hand in class i at day
t, and ri, the mean number of requests still to come from day t to departure, with
a standard deviation of 04.
At any point t, the number of bookings on hand is a given quantity for the par-
ticular future flight under consideration. If there is significant correlation between
b and r!, we can use this information to improve our estimates of Fi and a! over
those derived from historical information alone. For normally distributed bookings
and future requests, the expected value and variance of the latter given a value of
b is determined by:
Significant correlation between actual bookings and future requests can thus
substantially affect the conditional estimates (mean and standard deviation) of re-
quests still to come and, in turn, the revised EMSR protection levels. The impact
on ?! of a given value of b is determined by the correlation coefficient j, as well
as by the value of b relative to its own mean, as shown in equation (6.8). Just as
important, the information about correlation could be used to improve the estimate
of requests to come by reducing the variance of the estimate. As shown in equation
(6.9), any significant correlation will act to reduce the variance of r! in a conditional
probability density.
151
more speculative, given the difficulty of estimating "true" demand from constrained
flights.
When the sample subsets were disaggregated by day of week, the overall finding
was one of zero or positive correlation of booking levels between time periods prior
to departure. Significant negative correlation was evident in some cases, particularly
for the more restricted B fare class. Mean bookings in B class actually decreased
in many instances from day t to 0, for t : 14, due to advance booking and ticket
purchase limitations of 14 or more days on B class fare products during the sample
period. Such negative correlation was evident consistently on days of the week that
experienced relatively high demand for these reduced fares (e.g., Saturdays).
152
It is therefore possible to make use of information on the correlation between
booking activity in different time periods prior to flight departure, as well as the
historical correlation of demand levels in different fare classes, to develop a model to
forecast expected future requests more accurately. The success of such a forecasting
effort will depend on the strength of these relationships, which empirical analysis
has shown can vary substantially from market to market and over time.
The assumptions made with respect to the demand for different fare classes can
have an impact on the estimates of request densities used as input to the EMSR
model, as discussed above. Different estimates of demand can in turn change the
outputs of the EMSR model, namely, the optimal protection levels and booking
limits for each fare class. This section continues the discussion of the performance
of the EMSR framework by examining the sensitivity of the results to the input
variables, those related to both demand and revenue.
This discussion centers on the sensitivity of the optimal seat protection level for
class 1 from class 2, S2. In the EMSR decision rule, the value of Sj that maximizes
expected revenues in class 1 will be determined by the ratio f2/fl (where f2 < fi),
and by the parameters of the demand (request) density for class 1, estimated by F;
and ai. The optimal value of S2 must satisfy:
That is, if the fare ratio in equation (6.10) is 0.75, then we must find the value
of Z which has a probability of 0.75 of being exceeded, based on the estimated
153
parameters of the class 1 request density. From a table of standardized cumulative
normal probabilities, we determine Z = -0.675, and we solve the following for S:
S-
Z (6.11)
or:
S =F+ Z& (6.12)
We can vary each component of the above equation separately to determine how
the seat protection level S will be affected. The impact of variations in F are the
most straightforward to interpret. For a fixed Z and &, each unit change in F will
result in a one unit (i.e., one seat) change in the optimal number of seats protected.
All else being equal, then, an increase in the mean demand for a fare class will result
in an equivalent increase in EMSR protection level.
In most airline applications, however, the absolute value of & will increase with
F, while the coefficient of variation remains more stable. If we assume a constant
coefficient of variation, k = 0/F, then the expression for S becomes:
S = F+ ZkF (6.13)
S = F(1+Zk)
The optimal value of S thus remains a linear function of F for a constant Z and k.
For f2/fl = .50, Z will equal zero and the optimal S will simply equal F. Varying
a will have no effect.
154
Figure 6.1: EMSR Protection Levels for Different Estimates of Demand Variation
EMSR 1 (S)
($)
f
CLASS 1 DEMAND
WITH LOWER 2r
f2'> 0.5f
f2= 0.5f
\ CLASS 1 DEMAND
\WITH HIGHER 0
SEATSED S, (SEATS)
PROTECTED
155
Figure 6.2: EMSR Protection Levels Plotted Against Input Variables
/
f2H = 0.40
.f/f = 0.50
f2 / 1 = 0.60
f2 /f = 0.80
5 10 15 20 a
PROTECTED 40
SEATS (S)
30
20
101
a
0.20 0.40 0.60 0.80 1.0
156
The importance of reducing the estimated standard deviation of expected de-
mand as much as possible by editing out systematic variation and/or making use
of forecasting models is demonstrated in Figure 6.2. With F fixed at 20, optimal S
values are plotted in Figure 6.2(a) against a range of 8 estimates, for specific fare
ratios. Clearly, the range of optimal S values decreases dramatically for coefficients
of variation below 0.5.
The same example can be used to illustrate the impact of different fare ratios
(and in turn Z values) on the optimal S. As introduced above, the value of Z
decreases with fz/fi, meaning that the closer f2 is to fi proportionately, the smaller
the optimal protection level for class 1. Figure 6.2(b) shows how, for a given k-factor
or value of o, S is a non-linear function of the fare ratio. It also confirms that for a
range of fare ratios between 0.90 and 0.10, the range of optimal protection levels is
much smaller for lower estimates of the standard deviation of demand.
To summarize, both the demand and revenue inputs to the EMSR framework
can affect the calculated optimal seat protection levels significantly. The estimates
of mean demand by fare class have the greatest direct impact on the model results.
The estimated variance of expected fare class demand in turn determines the range
of output results for given fare ratios, reflecting the impact of increased uncertainty
about future fare class demand. The sensitivity analysis of the EMSR model results
presented in this section highlights the importance of accurate estimates of fare class
demand levels and their variation.
The flight leg-based approach to seat inventory control was selected for the devel-
opment of the EMSR framework both because of its relative simplicity and because
airline reservations systems maintain seat inventories and determine seat availabil-
ity by fare class for each flight leg. As mentioned in Chapter Three, several major
airlines are in the process of re-programming their reservations systems to manage
seat inventories on the basis of the different passenger itineraries that might use
a particular flight leg. The objective of such system development is to allow the
airlines to take into account the impact of requests for different fare class/passenger
itinerary combinations on total system-wide revenues rather than simply the pro-
rated revenues associated with an individual flight leg.
This section considers the application of the EMSR approach to the more sophis-
ticated inventory management structures currently being programmed into airline
reservations systems. The systems being developed make use of a concept known as
157
"virtual nesting" in determining the availability of seats in different fare classes for
a specific origin-destination itinerary.
The fare class/flight leg approach to seat inventory control does not distinguish
between passengers in the same fare class on different itineraries with different ticket
values. The example of a $29 local Q-class fare product and a $99 Q-class fare prod-
uct for a longer itinerary that must share the same initial flight leg was introduced
earlier. In such situations, it is possible for all Q fare class seats to be booked by
the local demand, denying space to the through demand that might have generated
a much higher total revenue for the airline. This can occur under a leg-based in-
ventory control approach when the demand for Q-class seats for local travel on the
continuing leg of the flight is low. Seats on the second leg can remain unsold at
the same time that all the Q-class seats on the first leg have been booked by local
passengers. Total revenues for both legs could have been increased by protecting
some Q-class seats on the first leg for through passengers.
158
have "local sales inhibitors" that allow limits to be placed on local passengers on a
flight leg. This ability to limit local sales has become especially important with the
development of large connecting hubs by the major airlines.
Under the virtual nesting concept, the availability of a particular fare class on
a leg will be determined by the availability of the corresponding "virtual inventory
class", or VIC, associated with a specific flight leg and O-D itinerary. Thus, Q-
class seats might be available on a flight leg from Boston to New York when an
ultimate destination of Atlanta is involved, whereas they might not be available in
conjunction with connecting travel to, say, Miami. As shown in Figure 6.3, this
determination of seat availability depends on the assignment of actual fare classes
by O-D pair to VICs for each flight leg. In the example shown, the value of a Q-class
booking in the BOS-ATL market has been judged by the airline to be higher than
that of either a Q-class or B-class booking in the BOS-MIA market, and is assigned
to VIC-6. Therefore, when the booking limit on VIC-8 for the leg BOS-NYC is
reached, no more Q-class seats for travel from Boston to Miami would be available.
Q-class seats for travel from Boston to Atlanta would remain available until the
booking limit on VIC-6 is reached.
The basic EMSR framework can be applied with little modification to the man-
agement of flight leg seat inventories defined by a virtual nesting system. For each
flight leg, the VICs must be nested hierarchically such that the top inventory class
cannot be shut down to further bookings without closing down all lower-ranked
classes, and the VICs must be ranked according to some measure of the value of
each to the airline in terms of revenue. The EMSR approach of protecting seats
for the top classes and establishing nested booking limits on the lower classes can
then be applied directly to the virtual nesting system. A booking limit would be
determined for, say, VIC-4 in Figure 6.3, and any request for a fare class/itinerary
combination corresponding to VIC-4 will be accepted as long as the total bookings
in VIC-4 and all lower VICs do not exceed BL 4 as derived from the EMSR model.
The EMSR approach may be applied to almost any aggregation of fare products
into virtual inventory classes, as long as a density of expected demand can be es-
timated for each inventory class that must share a common inventory of seats and
each category of demand can be associated with a particular revenue value to the
airline. In virtual nesting based on O-D destination ticket revenues, for example,
the VICs are defined for each flight leg on the basis of actual fare levels charged for
each fare class/itinerary combination. Thus, a Boston to Miami M-class fare level
of $250 is assigned a value of $250 for both the BOS-NYC and NYC-MIA flight
legs to be traversed by the passenger, in the example introduced in Figure 6.3. No
proration of revenues by flight leg is performed. Each possible O-D itinerary and
159
Figure 6.3: The Virtual Nesting Concept
MIA
ATL
BOS-ATL
BOS-MIA
BOS-NYC
VIC-10
160
fare class combination that uses the initial flight leg from Boston to New York City
will be assigned to virtual inventory class.
The EMSR framework can then be applied directly to this hierarchy of virtual
inventory classes, given estimates of expected demand for, and the average revenue
level associated with, each VIC. Optimal protection levels and booking limits can be
derived for each VIC, for each flight leg. Dynamic booking limit adjustment could
also be performed by using the EMSR framework as before. Actual bookings in each
VIC and incremental seat protection levels required for expected future requests in
the same VIC are summed to determine the nested protection levels and booking
limits. The only difference is that the demand and revenue inputs are estimates for
each VIC offered on a flight leg, not simply each fare class.
The aggregation of different fare classes and therefore fare products into virtual
inventory classes can complicate the inclusion of overbooking factors and upgrade
probabilities into the EMSR framework. In the case of overbooking factors, aggre-
gation of O-D itineraries into a single VIC can combine demand from different fare
classes that might exhibit different no-show and cancellation behavior. The over-
booking factors used as input to the EMSR framework will have to be adjusted to
reflect the mix of bookings by fare class expected and accepted in each VIC. Further
theoretical development of the model is required to address the question of optimal
overbooking limits on virtual inventory classes.
Incorporating fare class upgrade probabilities into the EMSR framework as ap-
plied to a virtually nested inventory structure can be even more problematic. Fare
class upgrades occur between adjacent fare classes on a given O-D itinerary request
but adjacent fare classes for a specific O-D itinerary are not likely to be assigned
to adjacent VICs. The approach for incorporating passenger choice shift probabili-
ties into the EMSR framework described in Section 5.5 therefore cannot be applied
directly to a virtually nested system. It might be possible, however, to adjust the
availability levels determined from the EMSR formulations for a particular itinerary
to account for upgrade potential.
161
the through fare being requested. Further theoretical development of the EMSR
approach should be directed at making this intervention process more systematic.
In summary, the EMSR framework can be applied to the virtually nested in-
ventory systems currently under development as a heuristic solution approach for
system revenue maximization. Modifications and further development to account
for overbooking factors, fare class upgrade probabilities and cases with high lev-
els of local demand could improve this heuristic substantially. Even without major
modifications, the EMSR approach can give an airline increased control of its seat in-
ventories and the ability to approximate the impacts of different fare class/itinerary
requests on system revenues. Until airlines are able to make practical use of the
large network optimization routines described in Chapter Four, the use and further
development of this heuristic approach offers the potential for increasing total rev-
enues through seat inventory control, one which can be realized in a relatively short
time.
The Expected Marginal Seat Revenue decision model developed and analysed in
the preceding chapters offers airlines a systematic approach to determining fare class
booking limits under existing reservations system capabilities. The basic principles
of the model may be applied with varying levels of sophistication and automation
to a comprehensive seat inventory control system. The way in which the EMSR
framework is applied in practice will be determined to a large extent by the avail-
ability of data required as inputs, as will be discussion in the following chapter. The
underlying decision approach of the EMSR framework can nonetheless be applied
in conjunction with a range of estimation methods and demand assumptions. With
the development of more sophisticated systems for managing seat inventories, the
basic principles of the EMSR approach can be extended to take into account system
revenue impacts.
162
PART THREE
System Development and
EMSR Model Applications
The final part of this dissertation returns to the practical side of airline seat
inventory management. Having developed a quantitative decision model for seat
inventory control in Part Two, we now examine the application of the EMSR frame-
work in an actual airline environment. Several of the practical considerations in
seat inventory management were introduced in Chapter Three, namely, the need for
human intervention in the process, the limitations imposed by existing reservations
system capabilities, and the importance of data availability. It is the objective of
Part Three to discuss the implementation of the EMSR model as part of an au-
tomated seat inventory management system, in light of these and other practical
considerations.
163
Chapter 7
The development of the EMSR decision framework, and its dynamic adjustment,
overbooking, and choice shift components in particular, was guided to a large extent
by the constraints and capabilities of the existing fare class/flight leg approach to
seat inventory control used by virtually all airlines. Many of these constraints were
discussed in the survey of current practice presented in Chapter Three. The impact
of these constraints on the development of a more systematic method for setting fare
class limits and its incorporation into an automated booking limit control system,
however, did not become apparent until the EMSR model was actually applied to
an airline operating environment.
This chapter first examines seat inventory control system development and im-
plementation issues, based on experiences at Western and discussions with represen-
tatives of several other major airlines. The impacts of airline marketing strategies
164
and operating procedures on system development are outlined. A separate sub-
section is then devoted to the reservations and revenue data requirements of the
EMSR framework, with an emphasis on the ways in which the operating policies
and procedures discussed previously can affect data availability. The second section
of this chapter describes the Automated Booking Limit System developed at West-
ern during 1986, which incorporates the EMSR approach to seat protection. Finally,
a revenue impact test of the EMSR dynamic framework conducted on a sample of
Western flight legs is described, including the test methodology and the results.
Airline policies and operating procedures can in turn affect the quantity and
the characteristics of the demand and revenue data available for input to decision
models like the EMSR framework. We have seen that the validity and accuracy of
the data used as input is extremely important to the EMSR model outputs. While
it might be possible to improve data availability simply by improving computerized
data collection and retrieval capabilities, changes to the policies and procedures that
generate the input data might also be required.
165
strategy adopted by the airline, the effectiveness of its seat inventory control process
determines the extent to which the objective of revenue maximization is attained.
A well-defined objective for the airline's seat inventory control strategy is thus an
essential basis from which improvements to the methods employed can be made.
Yet, it is still the case within the managements of several major airlines that this
type of single clear objective for seat inventory management does not exist.
In recent years, the industry and several carriers in particular have realized that
revenue maximization, not yield maximization, is the most appropriate overall objec-
tive for the relatively short run strategic focus of seat inventory control. In fact, the
director responsible for pricing and inventory management at one of these carriers
has forbidden any references by his staff to "managing yields". The development by
several airlines of inventory systems designed to manage seats by total O-D revenues
provides further evidence of the acceptance of the revenue maximization objective.
The objective of applying the EMSR model to a leg-based fare class inventory
system is to maximize total expected revenues for each future flight departure.
The model calculates the number of seats that should be protected for each upper
fare class and not made available to fare classes with lower expected per-passenger
revenues. The concept is straightforward, yet each of the critical functions listed
above can act to reduce the effectiveness of the EMSR model in maximizing expected
flight revenues.
Pricing and fare product "development" (i.e., the definition of the restrictions
and amenities associated with each fare product) is the function that is most closely
related to seat inventory control, and the most likely to be found in the same or-
ganizational unit of the airline. Nonetheless, it is possible for the objectives of seat
166
inventory control to be undermined by uncoordinated pricing actions and definitions
of fare product attributes. For example, inconsistency in the categorization of fare
products into the fare class inventories can reduce the effectiveness of seat inventory
control.
For seat inventory control to have an impact on total revenues, the seats made
available to a particular fare class inventory should be associated with a similar
average revenue level, relative to the other fare classes. In the case of a system
based on prorated leg revenues by fare class, categorizing different fare products
into fare classes should be done so that all fare products with similar prorated
average revenues are grouped into the same fare class. Grouping fare products on
the basis of similar prorated revenue levels can result in similar fare products for
different O-D markets being grouped into different fare classes, causing apparent
inconsistencies in the classifications.
Using the criterion of relative revenue averages alone can also result in inconsis-
tencies with respect to fare product restrictions such as advance purchase require-
ments. It might seem counter-intuitive for the pricing staff to categorize a fare
product with a longer advance purchase requirement into a higher fare class than a
lower-priced fare product with a shorter or no advance purchase requirement. Along
the same lines, those responsible for pricing and fare product categorization might
feel compelled to replicate the fare class categories of competitors to reduce con-
fusion on the part of travel agents. For a seat inventory control system to have a
positive impact on total revenues, however, the pricing function of the airline must
adhere to the criterion of relative revenue levels in determining fare class categories.
Fare product categorization is particularly important at the top end of the fare
class hierarchy, that is, the "Y" fare class intended to accommodate all "full coach
fare" bookings. At most airlines, the Y fare class is also used to book various other
fare products at revenue levels far below the full coach fare. The result can be
extreme "pollution" of the Y fare class demand data gathered from historical flights
and a prorated Y-class average revenue level that falls below one or more of the
supposedly lower fare classes.
Currently, the Y fare class inventory is used by most airlines to book not only
the full coach fare product, but some or all of the following reduced or zero-fare
products as well:
167
" airline and travel industry reduced fares (up to 90 percent off the full coach
fare);
" frequent flyer award ticket travelers, paying reduced or zero fares;
e "positive space" airline employees paying zero fare.
The rationale behind booking these fare products into the Y fare class involves
issues of seat availability. Being at the top of the nested inventory hierarchy, Y
seats will be available as long as any seats remain unsold. Access to the Y class
inventory for a fare product thus ensures maximum seat availability. Tradition,
as well as competitive considerations, have acted to keep the fare products listed
above in the Y fare class at most airlines. Children's and senior citizen's fares are
unrestricted equivalents to the full coach fare, and pricing departments will argue
that they should have access to any available seats as a result. "Visit USA" fares are
booked into the Y fare class because many foreign carriers have only the capability
to request coach cabin space for their clients.
The three remaining fare product groups listed have unrestricted access to the
Y class inventory at many airlines because of corporate policies and competitive
posturing. Airline and travel industry reduced fares "have always been" booked in
Y class, and a departure from this practice by one airline could alienate, for example,
travel agents. Similarly, discounted and free frequent flyer program ticket awards
are booked in Y class because those responsible for these programs in the airline
insist that:
1. If such awards were restricted to a less available fare class, it would reduce the
value of these awards relative to those offered by competing carriers' frequent
flyer programs.
Both arguments are valid, given the competitive nature of airline marketing, and
they illustrate clearly how the objectives of seat inventory control can conflict with
those of other airline functions. The objective of seat inventory control is to save
seats for the highest revenue passengers, while the frequent flyer program awards
these valuable seats for no tangible revenue.
The impacts on seat inventory control, and the EMSR framework in particular,
are twofold. First, the demand data for the Y fare class can be inflated by reduced-
fare or even zero-fare passengers, causing the model to protect more seats for Y
168
class than would be optimal. If the dilution of the Y-class revenues is so severe that
its average prorated revenue falls below that of other classes, on the other hand, the
EMSR model will not protect any seats for the Y fare class, regardless of the demand
levels. It can be difficult for those responsible for seat inventory control to explain
to upper management why no seats are being protected for the "highest" fare class.
Overall, the use of the Y fare class to book other reduced-fare products simply makes
it impossible for the EMSR model to determine a revenue-maximizing protection
level without making arbitrary assumptions about the fare products being booked
in Y class and their revenues.
169
functions and their potential revenue impacts, they simply cannot be left separate
if an effective system is to result.
170
of losing business and goodwill. The same arguments apply to cases in which check-
in agents are presented either with tickets issued for travel on another carrier, or
with tickets issued for travel on a restricted fare product. Short term gains may be
realized when check-in agents allow virtually any ticketed passenger to "stand-by" if
empty seats are available, particularly if the passenger holds another carrier's ticket.
Each stand-by passenger accepted in contravention of the restrictions of the fare
product purchased (e.g., Saturday night stayover required), however, undermines the
longer term revenue-maximizing objectives of differential pricing and seat inventory
control. Several carriers have in the recent past strengthened their policies with
respect to this issue, but experiences to the contrary suggest that such policies must
be understood by check-in agents and enforced more consistently.
The preceding discussion provides an overview of the critical functions with the
greatest direct impact on seat inventory control system development. There could
well be numerous others, depending on the policies and procedures of the specific
airline involved. The intent of this discussion has been to illustrate the extent to
which the objectives of functions throughout the airline can affect those of seat in-
ventory control. Under ideal conditions, the policies and procedures of the airline
should all correspond exactly to the revenue maximization objective of seat inven-
tory control. In practical terms, however, those responsible for system development
and implementation must at least be aware of the functions that might undermine
the objectives of this development. While efforts to coordinate policies should be
undertaken, seat inventory control system development can also require a review of
these functions, given their impacts on system performance and data availability. As
outlined in this section, seat inventory control system development must be coordi-
nated with the airline's pricing, fare product development, marketing, promotional
programs, overbooking, reservations, sales and passenger check-in functions.
171
7.1.2 Data Availability and Estimation Methods
The EMSR decision framework developed in Chapter Five requires two types
of input data for a particular future flight leg departure - estimates of expected
demand and its variation by fare class, and estimates of the average prorated leg
revenue associated with a passenger booking, again by fare class. The sensitivity
analysis presented in Section 6.2 demonstrated the importance of accurate estimates
of both demand and revenue for input to the EMSR decision model. This sub-
section considers the data requirements of the EMSR model and the extent to which
computer capabilities and operating procedures can affect the data that is actually
available. Data availability will in turn determine the methods that may be used to
transform the data into the estimates required as inputs.
The demand inputs required by the basic EMSR model are estimates of the total
number of requests expected for a future flight leg departure, by fare class. Because
the model takes into account stochastic variation in demand, an estimate of the
variance in total requests around the expected value is also required. For dynamic
applications of the EMSR framework, estimates of partial demand by fare class in
the form of requests still to come from day t to departure are required, as well as
estimates of the variation of this partial demand. Furthermore, the number of actual
bookings already accepted for the particular future flight leg being considered are
also necessary demand inputs.
Most airline reservations systems do not approach this level of data sophistica-
tion, and few are likely to in the near future. Currently, most reservations systems
log total bookings by fare class for a future flight leg. Database management systems
have been developed by many airlines to generate extracts of the current booking
levels and booking limits by fare class on a daily basis. These extracts become part
of an evolving historical database of total bookings by day before departure for all
flight legs for which reservations are being accepted. Once the flight has departed,
172
Figure 7.1: Historical Booking Build-Up for a Departed Flight
04 1 10
05 99
06 97
07 94
08 96
09 92
10 90
11 87
12 101
13 98
14 98
15 97
16 97
17 89
18 85
19 68
20 66
21 66
22 66
23 66
24 63
25 63
26 61
27 57
28 57
29 57
30 57
31 61
32 57
33 56
34 48
35 46
36 46
37 45
38 43
39 42
40 44
41 39
42 39
43 38
44 42
45 21
173
a complete booking history for that flight leg can be retrieved from the database.
Figure 7.1 shows an example of such a historical build-up of reservations from one
airline's database.
From this type of database, total bookings by fare class for a departed flight
can be extracted, as can partial bookings for any selected range of days prior to
departure (e.g., day 21 to departure). The difference in total bookings between
any two observations in the build-up history represents the net change in booking
levels in that fare class. Thus, in Figure 7.1, we can determine that a net gain
of 10 bookings in the Q class was realized between days 35 and 14 prior to the
departure of the flight. We know nothing, however, about the actual number of
bookings accepted and the number cancelled over the same period. No existing
airline reservations system currently has the capacity to log both bookings and
cancellations as separate data items.
For the purposes of the EMSR model, then, existing reservations systems, even
with the "sophisticated" database management software developed or purchased
by many airlines, are able only to provide a historical record of booking levels by
fare class by day prior to departure, as well as total bookings by fare class, net
of intervening cancellations for any period during the reservations process. These
reservations systems are also able to provide counts of the actual bookings on hand
for future flight departures.
The EMSR model, for initial and dynamic applications, requires estimates of
the mean and standard deviation of request by fare class. These estimates may be
derived from a sample of past operations of the same flight, or similar flights on the
same leg. If this historical sample includes only departed flights for which no fare
class booking limits were reached during the reservations process (implying that no
requests were refused), net booking levels may be used directly in the estimation
of requests by fare class. A sample of observations from the recent past might be
selected, with simple means and standard deviations of the fare class booking totals
in the sample used as reasonable proxies for total requests for each fare class. The
same estimation procedure can be used for partial booking levels, as required for
input to the EMSR dynamic adjustment routine.
It is not likely that all the flights in the historical sample will have booked up
without reaching one or more of the fare class booking limits. In such cases, one or
more requests for a particular fare class and flight are likely to have been refused
by the airline, and the disposition of these refused requests cannot be determined
from the available data. The net booking levels in the reservations system database
thus represent a constrainedestimate of total requests for a particular flight leg and
fare class. It is possible to use statistical methods to derive unconstrained estimates
174
of requests by fare class, given the booking levels for each observation in the sam-
ple and knowledge of whether these booking levels were in fact constrained by a
fare class booking limit. The statistical methods used can be as simple as graphi-
cal extrapolations on "normal probability paper", or can involve more complicated
maximum likelihood estimation techniques [65].
The premise of either method is to estimate what the mean and standard devi-
ation of requests would have been in the absence of booking limit constraints, based
on the assumption of normally distributed request totals. As a general rule, these
procedures can only be applied validly when fewer than one half of the observations
in the sample were subject to a constraint. For a flight leg and fare class that reached
its booking limit consistently during the period sampled, arbitrary approximations
of unconstrained request totals by fare class may be required. Such a situation may
signal a need to increase seat availability, which in turn would allow better data to
be collected.
For both the initial and dynamic applications of the basic EMSR framework,
then, the necessary demand estimates can be derived from the existing data avail-
able to most airlines from their reservations systems. The EMSR model extensions
to incorporate overbooking factors and fare class upgrade probabilities, however,
require demand data inputs that are not readily available to most airlines. In the
development of a seat inventory management system that includes these extensions,
the potential to improve data availability should be reviewed, as should the proce-
dures that generate the data.
With respect to the overbooking factors required by one extension to the basic
EMSR decision model, a simple approach of comparing actual boardings by fare
class with bookings just prior to departure to estimate no-show rates might be
considered. Issues that can arise in making data available for even this simple
approach to no-show estimation, however, include:
2. To what extent does the existing check-in process affect the relationship be-
tween "final bookings" and actual boardings by fare class?
175
and total cancellations on each day before departure rather than net bookings by
fare class would be required by models that determine optimal overbooking limits by
fare class as a function of time left to departure. Efforts to improve data availability
for such models should not only focus on the development of computer capabilities
to log, manipulate and report greater quantities of data, they should also examine
the meaning and validity of these data, and the impacts of the procedures that
generate them.
The extension to the EMSR model that incorporates passenger choice shift prob-
abilities poses even greater data requirements than that of overbooking factors. Ma-
jor re-design of existing reservations systems would be required to provide accurate
flight-by-flight and class-by-class estimates of upgrade behavior and total passenger
recapture. Without the ability to collect this type of data automatically through its
reservations system, the airline must derive estimates based on surveys of passenger
behavior or comparisons of historical flight build-ups for which booking limits were
reached.
The demand data available from most existing airline reservations can be exten-
sive when extracted to a database on a daily basis. For the purposes of seat inventory
control and the EMSR framework in particular, the type of booking data available
currently can be adequate if valid estimation methods are used. Improvements are
possible, of course, to the amount and detail of data collected by the reservations
system and then extracted to a historical database. In all cases, the airline pro-
cedures that generate the data should be a primary consideration in making such
improvements.
In contrast to the volume of historical booking data that airlines can extract
from their reservations systems, the availability of detailed revenue data is extremely
limited at many airlines. Few carriers maintain revenue databases that can provide
revenue information by flight leg and fare class. In fact, several major airlines
have no idea of the revenue they collect for specific flights, and must resort to
periodic samples of collected ticket coupons to obtain detailed revenue information.
Developed before deregulation, the revenue accounting systems at most airlines
were designed to report total system revenues, perhaps broken down by source of
sale (e.g., travel agencies, airline ticket offices). Little effort was made to gather and
store revenue data by flight and fare class, let alone by passenger itinerary.
With the evolution of seat inventory control practices, the need for better rev-
enue data has become critical. In the EMSR framework, the revenue associated
with bookings in different fare classes is just as important an input to the decision
model as the estimates of fare class demand. For the model to be used effectively,
average prorated leg revenues by fare class and city-pair are required at a minimum.
176
Historic revenue data by day of week and flight number might provide more accurate
estimates of the average revenues that can be expected for each booking accepted
for a future flight departure. As the development of inventory systems based on 0-
D ticket values progresses, much more detailed revenue data by passenger itinerary
will be required.
When the basic EMSR framework is applied to the existing leg-based manage-
ment of seat inventories by fare class, some estimate of the relative revenue value
of passengers booked in each fare class for the future flight under consideration is
required. A simple estimate may be provided by an assessment of the relative price
levels being charged for fare products in each of the fare classes that share the coach
cabin inventory. The shortcoming of this approach is that is does not reflect the
actual mix of fare products and O-D itineraries that are booked into each fare class,
because relative price levels among fare products can differ substantially from one
O-D itinerary to another on the same flight leg. Furthermore, the passengers ac-
tually being booked into each fare class might be using fare products that are far
different from the non-discounted fares published for each fare class. The dilution
of Y-class revenue (and demand) data by numerous reduced-fare products provides
an example of this problem. Fare product uage in each fare class is an important
consideration.
Another approach to estimating revenues by flight leg and fare class involves the
airline actually sampling ticket coupons collected at departure time and/or tickets
reported sold to build a revenue database. Total O-D ticket values can then be
prorated by flight leg and categorized by fare class, according to the fare basis
code on each ticket coupon. At least one major airline has developed this type of
revenue database, which can provide the EMSR framework with direct estimates of
average fare class revenues prorated by flight leg. Adjustments to these estimates
might be required to account for expected or observed changes to relative revenue
levels of different fare classes that are not reflected in the most recent historical
information in the database. Similarly, substantial changes in the O-D itinerary
mix of passengers traveling on a flight leg might require adjustment of the historical
revenue averages. In either case, however, the aggregation of many itineraries and
fare products into a single flight leg and fare class revenue average means that minor
changes to fare product prices and their usage in a small number of markets will
not have a significant impact on the aggregate estimates.
177
O-D markets. The O-D nesting of virtual inventory classes, described in Section 6.3,
is an attempt to increase the differentiation in revenue estimates and, in turn, to
give the airline greater ability to discriminate between requests of different revenue
value to the airline.
The amount of revenue data required for applications of the EMSR approach
to O-D virtual nesting far exceeds that being collected by most airlines. American
Airlines, a leader in the development of virtual seat inventory control, has also devel-
oped a revenue database that is updated to include tickets collected from departed
flights within days of departure. This revenue database can be used to generate
passenger counts and average ticket revenues by passenger itinerary, fare class, and
fare product. Control of virtual inventories nested by O-D ticket revenues does not
require proration of on-line revenues by flight leg.
In the longer run, seat inventory control system development should be related
to, and even direct, the development of data collection and retrieval systems. Sig-
nificant changes might be pursued in terms of the type of demand and revenue
data being collected, as well as policies and procedures that can affect the purity
of the data. At the same time, the capabilities of the reservations and database
management systems should also be reviewed.
The discussion in this section has focused on the system development and im-
plementation issues which are likely to arise when an airline decides to apply what
appears to be a relatively straightforward mathematical approach to its seat inven-
tory control process. While the EMSR model is not overly complex in theoretical
terms, incorporating it into a seat inventory control system within an existing airline
operating environment can be more complicated. The system development process
provides an opportunity to evaluate the impacts of existing policies and procedures
not only on the planned system, but on the overall objectives of seat inventory con-
trol. Consideration of the input data required by the system presents an opportunity
178
to assess the usefulness of the data being collected by the airline and to develop a
more comprehensive database management system. While it might be necessary to
implement a seat inventory control system around the constraints posed by existing
conditions, the potential for reducing or eliminating these constraints should not be
overlooked.
Many of the issues discussed in the previous section had to be addressed dur-
ing 1986 by Western Airlines, as it prepared to implement an Automated Booking
Limit System (ABLS) for seat inventory management. These procedural and data
availability constraints were in fact reviewed by Western management, and action
was initiated with respect to several of the problems identified over the course of
the year-long development effort. Still, there remained numerous policy issues and
data availability limitations that could not be resolved, meaning that the system
developed at Western was, to a certain extent, built around existing procedures and
capabilities.
The ABLS developed at Western incorporated the EMSR model in its static
and dynamic forms, as described in Sections 5.2 and 5.3 of this dissertation. The
objective of developing ABLS was to make the process of setting and adjusting fare
class booking limits for future flight departures more systematic and to automate it
as much as possible. It was hoped that the implementation of an automated system
would reduce the manual effort required on the part of a relatively small staff of seat
control analysts, allowing them to increase their analysis efforts for the small pro-
portion of flights requiring closer attention. Inclusion of the EMSR decision model
in the system was intended to provide the analysts with specific recommendations
of what the fare class booking limits should be, based on a systematic evaluation of
the input data.
179
class booking limits for all future flight leg departures, based on booking data from a
sample of recent departures of the same flight leg on the same day of week. Estimates
of the mean and variance of the total and partial requests by fare class were derived
from the booking data, and adjusted when necessary to account for "constrained"
demand due to booking limits having been reached on the sample flights. Revenue
averages extracted from Western's prorated revenue database for the most recent
available sample period were input directly into the EMSR formulations.
Fare class booking limits were calculated for a future flight leg 90 days before
departure, using the initial EMSR model. These fare class limits were revised weekly
thereafter, up to and including 6 days before departure. The batch routine operated
on a day-of-week rotation, handling all flight legs scheduled to depart on all future
Tuesdays up to 90 days out, for example, on Wednesday mornings. Each weekly
revision run between day 90 and day 41 made use of the initial (static) EMSR
model, incorporating the additional data provided by the most recent (i.e., the
previous day's) departure of the same flight leg into the demand density estimates,
as well as any changes to the relative revenue averages.
The "on-line" routine allowed the analysts to intervene and to run the EMSR
decision model for an individual flight leg and day of week. This routine was designed
to display the input data used by the EMSR model to the analysts. It allowed the
analysts to make exclusions of specific past demand data or to change the input
data assumptions outright. The EMSR model calculations could then be performed
on-line for the revised inputs and, if the results were judged to be reasonable by
the analysts, the resultant booking limits would be loaded into the reservations
system instead of those recommended initially by the batch routine. The on-line
180
routine also enabled the analysts to compare the batch routine recommendations
for booking limits with current limits and actual bookings already on hand for
each future departure of the flight leg on the same day of the week (up to 90 days
out). The EMSR recommended protection levels and booking limits could simply
be overridden manually before being loaded into the reservations system.
ABLS was thus designed to allow user (seat control analyst) intervention in
the application of the EMSR decision model to future flights. This capability was
especially important in light of the many "imperfections" in the system. As of
the end of 1986, ABLS did not include a demand forecasting model or an ability
to make seasonal adjustments to demand and revenue estimates. Furthermore, no
upgrade probabilities were included in the derivation of optimal fare class booking
limits. These limitations required that analysts have the capability to override
the recommended limits on flight legs where upgrade potential was thought to be
significant and in markets or during periods for which the recent historical data did
not provide a valid estimate of the revenue or demand conditions expected for future
departures of the same flight leg.
The EMSR dynamic adjustment routine was implemented into the Automated
Booking Limit System in December 1986. Further system development was not
possible at Western Airlines because of the announced merger with Delta Air Lines.
Fortunately, this pause in system development activities allowed the revenue impacts
of ABLS, and of the EMSR dynamic adjustment routine in particular, to be tested
for a sample of actual Western flight legs, as described below.
This final section of Chapter Seven describes a performance evaluation of the Au-
tomated Booking Limit System (ABLS) developed at Western Airlines, conducted
during the first three months of 1987. The testing methodology is described first and
its limitations are discussed. The revenue impacts of ABLS seat inventory control
relative to the established manual method are then presented, and their implications
for further system development are assessed.
181
limits set automatically by ABLS, which included a dynamic adjustment routine
based on the EMSR decision model. The automated system (or "batch" routine)
began setting limits for a single day of the week for each of 21 flight legs. Both the
flight legs and days of week involved in the test were known to the analysts because
they had to refrain from applying their manual seat inventory control methods
to these flights. The group of 21 flight leg/day of week combinations for which
ABLS would be allowed to set and revise fare class booking limits without analyst
intervention comprised the BATCH test group.
To give the analysts an opportunity to become familiar with ABLS and to in-
teract with it, they were told to use the system for each of the 21 flight legs in the
test, but on a specified day of week for each leg, different from the BATCH day of
week. The seat control analysts could run the on-line version of ABLS for these
flight legs/days of week and then adjust the EMSR recommended booking limits
on the basis of their own judgement before loading them into the reservations sys-
tem. Ultimately, interactive use of ABLS was to be the norm, with the combination
of automation and human intervention resulting in what should have been higher
flight leg revenues than what the automated system alone could generate. For the
purposes of this test, however, this ONLINE test group was not intended to provide
a measure of the joint performance of the system and the analysts. The objective of
this test was to compare the completely automated system of seat inventory control
used for BATCH test flights with the existing manual methods used for CONTROL
test flights.
For each of the 21 flight legs in the test sample, then, five days of the week
remained, for which booking limits would be set manually by the analysts, based
entirely on their own judgement, as they had been for years. One of these remaining
days of the week was selected to represent a CONTROL flight for each of the BATCH
flights in the test. The specific day of week to be used as a basis for comparison
with the BATCH day of week was unknown to the analysts. In fact, for the first
two months of the test, the existence of any CONTROL flights at all was not known
to the analysts.
The city-pairs and flight legs to be used in the test were selected on the basis
of expected demand levels high enough to cause fare class booking limits to be
reached in one or both of the test groups (i.e., BATCH or CONTROL). If fare class
booking limits were never reached for the flight leg departures in the test because
of very low demand in all fare classes, it would not be possible to determine the
impacts of the different seat inventory control methods. The primary consideration
in selecting flight legs for the test was therefore an expectation of relatively high
demand, preferably in more than one fare class.
182
Historical traffic data by flight leg from the fall of 1986 (the most recent data
available at the time) and for January through March 1986 (corresponding to the
three months of the planned test in 1987) were examined and the highest load factor
flight legs identified. Actual bookings already on hand as of December 15 for each
of these flight legs were then reviewed for the test period. All flight legs for which
one or more fare classes had already booked out on specific days of the week for
departures during the test period were eliminated from further consideration. This
step effectively removed all winter vacation markets from the test sample, since
many of the flight legs considered already had one or more low fare classes sold out
for the February and March peak period departures.
For each of the city-pairs/flight legs remaining in the sample, a more detailed
analysis of historical booking patterns, load factors, and fare class mix of passengers
was performed, including a comparison of days of the week. For a flight leg to be
selected for the test, it had to exhibit stable and relatively high historical load
factors, as well as a similar fare class mix of passengers for three days of the week
(i.e., BATCH, ONLINE, CONTROL). The final test sample included 21 flight legs
operating between 19 city-pairs, as listed in Table 7.1. These city-pairs include a
variety of stage lengths and involve primarily mixed business/pleasure markets on
flight legs feeding Western's Salt Lake City or Los Angeles connecting hubs.
The three days of the week associated with each flight leg in the sample were
assigned to the three test groups randomly. When minor differences in historical
load factors between the days were apparent, this assignment was performed on
the basis of the load factor rankings of the three days. For example, the highest
ranked day of week was assigned to the BATCH group for one flight leg, to the
CONTROL group for the next flight leg, and so on. This was an attempt to reduce
the potential of systematic bias in demand levels by day of week favoring one test
group over another.
The test sample thus consisted of 21 flight legs. For each flight leg, all departures
on one day of the week would have fare class booking limits set automatically by
the ABLS "batch" routine, while all departures on another day of the week would
be managed manually by seat inventory control analysts, thereby providing the
"control" group for comparison purposes. A third day of the week for each flight leg
would be handled through an "online" process in which analysts could use ABLS
to assist in setting fare class limits. The primary focus of this evaluation of ABLS
performance, however, was on the traffic and revenue levels of the BATCH group
relative to the CONTROL group.
Several problems with interpreting the results of this test were identified even
before the test began. The format of the test was constrained by the design of
183
Table 7.1: ABLS Flight Legs in Test Sample
184
the system being tested, while comparisons of traffic and revenue impacts would be
limited by the uncertainty associated with the demand for future flights and by the
nature of any post departure assessment of the impact of different fare class booking
limits on flight performance.
From the outset, the day of week rotation programmed into ABLS dictated that
a day of week approach to comparative testing be used. The BATCH flights for a
particular flight leg would be those departing on Tuesdays, for example, throughout
the duration of the test, while all CONTROL flights might be Thursday departures.
Although the days of the week for each flight leg were selected originally on the basis
of similar historical loads and booking patterns, unexpected consistent differences
in demand for one day in the test could give a systematic advantage for one test
group over the other on that flight leg. Valid comparisons of the two test groups
for the leg under such conditions would be difficult to make. It would have been
preferable to assign different days of the week to each test group for different weeks
during the test period.
A second interpretation problem involved the previously mentioned fact that the
effects of different fare class booking limits can only be evaluated for flights that
actually reached one or more of these limits in at least one of the test groups. If no
fare class limits are ever reached for a set of flight legs, any difference in loads or
revenues will be attributable solely to variations in demand between the departed
flights. Even though the flight legs in the test sample were selected on the basis
of high load factors for historical periods similar to the test period, there was no
guarantee that unexpectedly low demand would not be observed on some of the test
flights due to changing market conditions.
Finally, the inability to subject the same flight departure to different seat inven-
tory control methods under exactly the same demand conditions meant that much
of the analysis of the test results would be speculative. Even with a complete history
of the fare class limits applied and the booking levels by day before departure, we
could only speculate about what might have occurred in the booking process for
the same flight departure had different fare class limits been applied. A compar-
ative analysis of similar flights thus would not provide conclusions with certainty,
although it was expected that the evidence of revenue impacts of different limits
would allow reasonable conclusions to be reached. This is the major problem faced
by airline managers hoping to measure the effects of seat inventory control practices
- there is no way to determine exactly what revenues or loads would have been
realized in the absence of seat inventory control or under different methods.
185
booking limits likely affected actual bookings in each fare class, as well as total
flight revenues. It was decided from the outset that aggregate revenue and traffic
totals by test group would not be used to compare the performance of ABLS relative
to the manual method of seat inventory control. Furthermore, all flight departures
for which no fare class booking limits were ever reached would be excluded from the
comparisons. The objective in interpreting the test results was therefore to identify
BATCH and CONTROL flight departures that would have had similar booking
patterns in the absence of different fare class limits.
Only flights departing after January 18 were included in the evaluation to ensure
that they had been exposed to differences in booking limits for at least a month
prior to departure. For the purpose of making direct comparisons between flights
with as similar a demand patterns as possible, BATCH-CONTROL flight pairs
were identified from the same week of the test period. It was felt that comparing a
BATCH flight from one week with a CONTROL flight from another might introduce
time-dependent variations in both demand and test group performance.
The results of the revenue impact test were evaluated in terms of the differences
in fare class mixes of passengers carried, load factors, and total flight revenues
between the BATCH and CONTROL test groups. Actual boarding data by fare class
were provided by Western's reservations and traffic database for departed flights, as
were detailed pre-departure booking and fare class limit histories. Revenue levels
associated with each boarded passenger on a test flight were the prorated leg revenue
averages by fare class used as inputs to the EMSR calculations, extracted from
Western's revenue database. The assessment of revenue impacts was based on a
comparison of flight pairs (i.e., BATCH and CONTROL) that departed during the
same week on the same flight leg.
Each flight pair presented a potential basis for comparison of the revenue and
traffic impacts of ABLS relative to the manual method of seat inventory control used
at Western. Post-departure results and booking histories were collected for a total
of 210 flight pairs (21 flight legs, 10 test weeks). The major task in analysing the
results, as introduced above, involved identifying the flight pairs that demonstrated
differences in loads and/or revenues due to the application of different fare class
booking limits, with all else being equal.
186
booking histories for each flight pair were examined. From the outset, flight pairs
for which no fare class booking limits were reached for either flight were eliminated
from further consideration. In spite of efforts to select flight legs and days of the
week for which high demand was expected during the test period, more than half of
the flight pairs were eliminated for this reason.
Identification of flight pairs for which at least one fare class booking limit was
reached during the reservations process did not ensure a valid comparison of similar
flights. Each remaining flight pair had to be judged as providing a "valid" compari-
son of the effects of different fare class limits or not, based on the historical booking
build-ups of the flights involved. Acceptance of a flight pair as a valid comparison
was based on the criteria described below.
A difference in load factors between the two flights in a valid flight pair should
be attributable to differences in one or more fare class booking limits. For example,
the lower load factor flight might have had lower fare class limits applied to a lower
fare class, constraining total bookings for that flight relative to the total bookings
accepted for the other flight with a higher limit on the same fare class. Significant
differences in load factors that could not be explained by differences in fare class
booking limits indicated a higher level of total demand for one flight over the other,
resulting in rejection of that flight pair as a valid comparison.
In cases where booking limits on the same fare classes were reached for both
flights, bookings in the next higher class should reflect a presence or lack of "up-
grade" behavior, rather than a difference in total demand for the two flights. For
example, bookings on one flight in a pair might have reached the booking limit on
the lowest fare class (Q) of, say, 40, while bookings in the next highest class for the
same flight reached a total of 30. For this flight to provide a valid comparison to
the other flight in the pair, which had a higher unreached limit of 60 on its Q-class
bookings, total bookings in B and Q classes for the second flight must be greater
than or equal to the 70 accepted for the first flight. If the total demand patterns
are the same for the two flights, a constraint on Q demand on the first flight should
result in fewer total B and Q bookings than that accepted for the unconstrained B
and Q classes on the second flight.
Even if the two lowest classes were judged to provide a valid comparison of
similar demand processes, differences in booking totals for higher fare classes would
result in rejection of the flight pair. The general rule used in selecting valid flight
pairs was that any difference in load factors and fare class mix of passengers between
the flights must be explainable by differences in the fare class limits applied during
the booking process.
187
Determining the similarity of the demand process for the two flights in each flight
pair considered also required a comparison of the pre-departure booking build-up
process of each flight. For the two flights to be judged similar, the rate of booking
level growth in the various fare classes had to be similar at various points in time
prior to departure, in the absence of booking limit constraints. This task was made
all the more difficult by the dynamic revision of booking limits performed on the
BATCH flights. While the CONTROL fare class limits exhibited relative stability
over the course of the booking process for each flight, perhaps changing once or twice,
BATCH fare class limits were revised weekly for all fare classes. It was thus possible
for a fare class to be closed down at one point during the booking process, only to
be re-opened later due to ABLS calculating a higher revised fare class booking limit.
The best way to illustrate the characteristics of a valid flight pair comparison is
to present specific examples from the test. Figure 7.2 shows the historical booking
build-up for a BATCH and CONTROL flight pair from the BOI/SLC flight leg.
The asterisks indicate that the fare class had reached its booking limit and was
therefore unavailable on that particular day prior to departure. For the BATCH
flight, no limits were reached during the booking process, whereas the CONTROL
flight reached its Q limit 10 days prior to departure. On day 10, total bookings
for the CONTROL flight actually exceeded those for the BATCH flight in all fare
classes. The BATCH flight accepted a total of 71 more bookings, 32 of which were in
Q class. The CONTROL flight accepted only 35 more bookings from day 10, none
of which were in Q class. The net result was 20 more bookings for BATCH than
CONTROL, and a load factor of 83 percent relative to 75 percent, respectively. The
revenue totals for the two flights sh6wed a 5.3 percent advantage for the BATCH
flight.
The final result for this flight pair again showed a substantially different fare class
mix of passengers booked and boarded. The BATCH flight sold out on the day of
departure and went out full (some passengers might have been denied boarding),
188
Figure 7.2: Valid Flight Pair Comparison - Example 1
125 65 * 119
110 60* 101
103 61 * 96
86 59* 85
73 58* 84
59* 84
59* 85
53 79
52 77
43 65
38 59
37 58
36 57
36 57
33 53
26 44
25 41
24 38
24 38
23 37
20 35
21 36
15 28
13 23
8 18
8 18
8 18
9 16
9 17
8 14
0 0 2 25 27 1 2 2 9 14
0 0 2 25 27 1 2 1 5 9
189
Figure 7.3: Valid Flight Pair Comparison - Example 2
**ACTUAL 8 37 7 55 107 57 i8 12 93
00 7* 50 9 60* 126 * 60 20 12 * 97
01 4 42 11 60* 117 61 20 13 * 100
02 2 39 10 61* 112 56 24 13 * 95
03 2 35 8 59 * 104 55 24 13 * 93
04 2 34 8 57* 101 52 23 12 * 88
05 2 34 9 60 * 105 52 25 12 * 90
06 1 25 9 61 * 96 52 25 1.3 * 92
07 1 22 7 59 89 41 25 12 * 79
08 1 21 6 56 84 33 23 13* 69
09 1 15 4 43. 63 27 21 13 * 61
10 1 12 4 41 58 22 17 13* 52
11 1 12 4 41 58 18 15 13* 46
12 1 12 4 40 57 18 15 11 44
13 1 10 4 37 52 18 15 11 44
14 1 8 4 41 54 18 14 11 43
15 1 8 3 38 50 12 14 12 38
16 1 7 3 26 37 10 14 14* 38
17 1 7 3 25 36 9 14 12 35
18 1 7 3 27 38 6 14 12 32
19 1 7 3 22 33 6 14 12 32
20 1 9 3 18 31 6 14 12 32
21 1 8 3 17 29 7 12 33
22 1 7 3 15 26 7 11 13* 31
23 1 4 3 14 22 5 9 13* 27
24 1 4 4 14 23 5 9 13* 27
25 1 4 4 14 23 3 7 13* 23
26 1 4 4 13 22 3 7 13* 23
27 1 4 4 11 20 3 7 13* 23
28 1 4 4 11 20 3 8 9 20
29 1 4 4 11 20 3 8 13* 24
30 1 6 2 11 20 2 7 12* 21
31 1 5 1 12 19 4 5 13* 22
32 1 5 12 19 3 4 13 * 20
33 1 5 12 19 3 4 13* 20
34 1 5 12 19 3 4 13 * 20
35 1 5 12 19 3 2 14* 19
36 1 5 14 21 4 2 14* 20
37 1 5 10 17 3 2 13* 18
38 1 5 9 16 3 1 11 15
39 5 1 9 16 0 3 1 11 15
190
whereas the CONTROL flight realized an 87 percent load factor. In this case, the
differences in fare class mix on the two flights resulted in a 5.1 percent higher total
revenue for the CONTROL flight, even though the BATCH flight carried more
passengers.
Even with a relatively liberal application of the criteria described above, the final
set of valid flight pair comparisons proved to be relatively small. Out of the 210
possible flight pairs in the test, approximately two-thirds were eliminated because of
unexpectedly low demand. Winter flight cancellations and a change in aircraft type
on one of the flight legs in the test also contributed to the number of flight pairs
eliminated from the outset. Systematic day of week demand differences or simply
different booking build-up processes eliminated about one-half of the remaining
flight pairs, leaving 36 valid flight pair comparisons (72 flight departures) for an
assessment of ABLS impacts.
The flight pairs showing positive and negative revenue impacts of ABLS are listed
in Tables 7.2 and 7.3, respectively. For each flight pair listed, the fare class mix of
passengers as well as the flight load factors and the percentage difference in flight
revenues (BATCH over CONTROL) are shown. The passenger mix is the number
of passengers actually boarded by fare class, and the load factor is the percentage
of seats on the aircraft filled for that specific flight departure. The total revenues
by flight have been omitted for reasons of data confidentiality.
The asterisks in both figures indicate that the associated fare class limit was
reached by bookings for that flight during its reservations process, and that the
closing of the fare class had a significant impact on how the flight proceeded to
book up relative to the other flight in the pair. It is therefore possible, especially for
flights that booked out very close to departure, that a 100 percent load factor might
be indicated without asterisks being shown for one or more of the fare classes. The
asterisks imply a significant constraint on the booking process for the relevant fare
classes.
Table 7.2 details the 25 flight pairs in which ABLS was judged to have a positive
impact on total flight revenues under similar demand conditions. The average
191
Table 7.2: ABLS Positive Revenue Impact
192
Table 7.2: ABLS Positive Revenue Impact (cont'd)
BOI/SLC B 69 83 + 5.5 %
C 56* 75
SLC/PHX B 127 * 100 + 7.1 %
C 49 * 85
LAS/LAX B 62 63 + 22.6%
C 41* 49
LAX/PHX 64* 88 + 35.7%
49* 64
LAX/PHX 38* 100 + 15.1 %
62* 100
IAD/SLC B 50 77 * 100 + 14.7%
C 41* 68* 87
SLC/PHX B 13 118 * 100 + 4.0 %
C 54 44* 90
MSP/SLC B 19 110* 96 + 29.8 %
C 5 71* 71
IAD/SLC B 61 61* 100 + 8.0 %
C 48* 66* 95
LAS/LAX 3 4 1 69 64 + 23.6 %
8 7 1 43* 49
193
Table 7.2: ABLS Positive Revenue Impact (cont'd)
194
impact per flight pair in this group amounted to a 14.3 percent higher revenue
for the BATCH flight over the CONTROL flight. When weighted by the revenue
levels and loads associated with each flight pair, the aggregate positive impact on
total revenues amounts to a 12 percent advantage for the BATCH flights. The
average flight load factors for all flights in this positive impact group were also 12.3
percentage points higher for the BATCH flights than for the CONTROL flights, due
to the application of ABLS-generated fare class booking limits.
A scan of the fare class mixes and load factors in this group reveals that, in
the majority of the flight pairs, the BATCH flights carried a far greater number
of passengers in the lowest fare classes than did the CONTROL flights, with a
higher total revenue. The flight pairs in this group illustrate cases in which the
revenue benefit of selling more low-priced seats outweighed that of closing down
the lowest fare classes in the hope that denied requests would result in fare class
upgrades. Note, however, that in many of the flight pairs, lower Q-class limits on
the CONTROL flight did in fact increase the number of B passengers carried.
The relatively high booking limits on the lower fare classes set by ABLS for the
BATCH flights can be attributed to at least two factors. First, the prorated leg
revenue averages used as inputs by ABLS did not differ radically among the fare
classes for many of the flight legs in the test. This was the case for flight legs in
highly competitive, short-haul markets served by Western, in which the price levels
of the fare products being sold in adjacent fare classes can differ by as little as
$10 (e.g., $49 for a Q-class fare versus $59 for a B-class fare). When such similar
price levels are aggregated and prorated over flight legs, the fare class revenue ratios
used by the EMSR formulation contribute to relatively low protection levels for the
higher fare classes, leaving more seats available to the lowest fare class, Q.
The second factor is related to the first, and involves the lack of any measure
of upgrade potential in the EMSR formulation used in this test. Whereas the seat
control analysts at Western could could set fare class limits on CONTROL flights
with some expectation of upgrade potential, the BATCH limits reflected an assump-
tion of zero upgrade potential. As a result, there is evidence in many of the flight
pairs listed in the positive impact group of upgrade behavior for the CONTROL
flights with constrained fare class demands, as indicated by the asterisks. For the
flight legs operating in markets in which the price levels of fare products in adjacent
fare classes were similar, or in which Western faced little effective competition, the
actual upgrade potential appeared to be significant. The flight pairs for which the
inability of ABLS to take into account fare class upgrade potential led to a negative
revenue impact are listed in Table 7.3, and will be discussed later.
The higher total revenues for the BATCH flights in Table 7.2 can be explained
in the majority of the flight pairs listed by higher load factors stemming from more
195
liberal booking limits on lower fare classes. This effect is readily apparent from the
fare class mixes and asterisks shown for these flight pairs. Several of the flight pairs
on the list, however, merit further explanation:
" Week 4, LAX/PHX: In the only flight pair for which the BATCH flight showed
a lower load factor than the CONTROL flight, higher total revenues were
realized on the former because of lower Q limits and substantial upgrades
from Q to B. This result was also observed in Week 8, for the same flight leg.
" Week 5, SLC/BOI: Both B and Q classes were constrained on the CONTROL
flight, with little upgrade activity, relative to the BATCH flight.
" Week 7, LAS/LAX: For short-haul, highly competitive markets like Las Vegas
- Los Angeles, very little upgrade activity was observed when Q class was
constrained, especially for relatively low load factor flights like this one.
" Week 7, LAX/PHX: In this case, the higher number of B passengers for the
BATCH flight is due to a relatively low Q-class limit having been applied by
ABLS early in the booking process. The Q-class limit on the BATCH flight
was then increased late in the process, when expected bookings in the higher
classes did not materialize. The lower Q-class limit on the CONTROL flight
applied throughout the booking process, keeping total bookings lower.
" Week 9, LAX/ABQ: This pair deserves special mention because it provides an
example of how the seat control analysts began to adjust their own practices
given exposure to the ABLS recommended limits. In this case, the CONTROL
flight actually carried 83 Q-class passengers, a number much higher than the
Q-class limits applied for the same CONTROL flight earlier in the test period.
" Week 9, PDX/SLC: The pair demonstrates how the limits set for fare classes
higher than Q (in this case, B) can be just as important to total revenues as
the Q-class limits.
196
To summarize the results of the positive impact flight pairs, the BATCH limits on
lower fare classes were substantially higher than those for CONTROL flights, with a
few exceptions. The CONTROL limits on the lowest fare classes were generally too
low to maximize total flight revenues. On the other hand, there is some evidence
that the BATCH limits on the lower fare classes might have been too high, giving
passengers that would have been willing to purchase higher-price seats access to
lower fare classes. This finding is supported by the results for the 11 flight pairs
presented in Table 7.3, which demonstrate the negative impact on revenues of ABLS.
In Table 7.3, all of the flight pairs once again show a BATCH flight load factor
greater than or equal to that of the CONTROL flight, but a total flight revenue
that was lower for the BATCH flight in each case. The revenue advantage for
the CONTROL flights stemmed from lower booking limits on Q class and upgrade
movements to higher revenue fare classes. The asterisks in this figure demonstrate
this phenomenon clearly. Flight pairs that merit further explanation include:
" Week 1, LAX/SMF: The upgrade movement here involved a vertical shift
from Q to M class, likely because the B-class fare product offered in the local
market served by this flight leg was not as attractive an upgrade alternative
at the M-class product (i.e., same price, more heavily restricted). The same
observation applies to the mix of bookings for Week 7, SLC/BOI.
" Week 7, IAD/SLC: The BATCH flight reached its limits in both B and Q
classes, but these limits were higher than those for the CONTROL flight in
the same classes.
For the 11 flight pairs in this group showing negative revenue impact, the av-
erage revenue difference per flight pair amounted to a 6 percent shortfall for the
BATCH flight relative to the CONTROL flight. The weighted difference in aggre-
gate revenues for this group amounted to a 5.9 percent shortfall for the BATCH
flights. The difference in average flight load factors was 5.7 percentage points in
favor of the BATCH flights, many of which in fact departed with no empty coach
seats.
The overall conclusion that can be drawn from the flight pairs shown in Table 7.3
is that the BATCH flights definitely did have booking limits on the lower fare classes
that were too high relative to the CONTROL flights. Upgrade activity contributed to
higher total revenues for the CONTROL flights in this group, in spite of higher load
197
Table 7.3: ABLS Negative Revenue Impact
LAX/SMF 85 98 -4.0%
54* 95
LAX/ABQ 90 97 -0.5%
49* 85
SLC/PHX 114 99 -4.0%
43 * 90
4 SLC/PHX 18* 107* 100 - 10.7 %
67* 44* 100
LAX/ABQ 9* 89* 100 - 10.7 %
39* 43* 93
SLC/PHX 4 4 11 109* 94 -2.7%
2 12 59 46* 88
SLC/BOI B 55 * 100 -5.1%
C 12* 87
7 IAD/SLC B 3 53* 73* 100 -2.8 %
C 23 38* 61* 99
8 SLC/BOI B 36 21* 46* 99 - 15.3 %
C 52* 22* 10* 99
9 SLC/PHX B 7 12 116 * 100 -6.0%
C 19 63* 46 * 96
10 SLC/PHX 4 13 116 100 -5.0%
23 60* 35 * 93
198
factors for the BATCH flights. Most of the flight pairs listed, however, belonged
to one of three flight legs in the test, suggesting that these flight legs experience
a significant upgrade potential that might not be realized on other flight legs in
different markets.
The aggregate revenue and load factor results for all 36 flight pairs considered
to reflect a valid comparison of ABLS versus the manual method of seat inventory
control are summarized in Table 7.4. An overall positive revenue impact of 6.2
percent was realized for the BATCH flights over the CONTROL flights in this
comparison. The average flight load factor advantage for BATCH flights amounted
to 10.3 percentage points. This discrepancy between the magnitude of the increase
in load factor and total revenues is explained by the lower overall yield realized on
the BATCH flights than on the CONTROL flights. The shortfall in overall yield
was outweighed, however, by the increase in total revenues, providing evidence that
yield maximization does not necessarily mean revenue maximization.
The potential positive impact on total flight revenues of a more systematic ap-
proach to seat inventory control than the application of booking limits based on
analysts' judgements was clearly demonstrated by this initial test of a relatively ba-
sic version of Western's Automated Booking Limit System (ABLS). This conclusion
is supported by the many other examples of individual flights for which dynamic
revision of fare class limits by ABLS had a significant positive impact on total flight
revenues, but for which the corresponding CONTROL flight did not reach its limits
or did not exhibit a similar demand pattern.
While the final results of this test show a positive revenue impact due to the
application of the quantitative decision approach of the EMSR framework, there
are several caveats with respect to the test results that should be emphasized. First,
this test compared the impact on loads and revenues of an automated system for seat
inventory control relative to the manual methods being used at Western Airlines.
We can only assume that the judgement and expertise of the analysts involved in
this test is representative of that of analysts at other airlines. It is possible that
analysts with better training and skills, or even a larger number of analysts paying
closer attention to individual flights could have reduced the the observed differences
between the BATCH and CONTROL test flights.
On the other hand, given that many airlines currently rely on the judgement
of relatively few seat control analysts, as discussed in Chapter Three, this test suc-
ceeded in demonstrating that there are advantages associated with a more systematic
decision approach. It should also be noted that ABLS was designed to enable ana-
lysts to apply the EMSR-derived booking limits or to intervene in instances where
factors not accounted for in the ABLS routine might affect the optimal booking
199
Table 7.4: Summary of Test Results
200
limits -for a future flight. This test compared a strictly automated approach with
a strictly manual approach, meaning the joint potential of automation with human
intervention was not assessed.
While the performance of flights for which analysts set fare class limits on the
basis of the ABLS recommended limits (i.e., the ONLINE test group) was not evalu-
ated in detail, the results for the interactive group seem to be mixed. The analysts at
Western were not trained formally with respect to the use of ABLS in an interactive
environment. The result was substantial variation between analysts, across flight
legs, and over time in the extent to which the recommended limits were actually
applied. As the test progressed, there was in fact evidence of a learning curve effect
on the part of the analysts, which also affected the CONTROL flights in some cases.
Given repeated exposure to the ABLS recommended limits for ONLINE flights, the
analysts began applying remarkably similar limits to some CONTROL flights.
The second caveat involves the version of ABLS that was used for this test. The
system tested included only the basic EMSR formulations for calculating initial fare
class limits and revising them periodically in light of changes to the input data and
actual bookings for a future flight. The EMSR model extensions to include upgrade
probabilities and to take into account overbooking factors were not incorporated
into ABLS. Furthermore, the data inputs were limited to those available from West-
ern's reservations and revenue databases. The prorated revenue averages reflected
substantial data pollution in many cases, while the demand inputs used were sim-
ple historical averages from recent operations of the same flight leg. No seasonal
adjustment or growth trend forecasting was performed.
The period between ABLS revision runs for each future flight was chosen to be
7 days, which proved to be too long for many of the flight pairs examined in this
test, particularly within the last two weeks of the booking process. More frequent
revisions would have reduced the BATCH limits on low fare classes substantially in
many of the flight pairs. In spite of these shortcomings of the system tested, analysis
of the performance of ABLS with respect to specific flight pairs clearly illustrated
the benefits of an automated seat inventory control system that can adjust booking
limits dynamically.
The proportion of flights for which an impact was observed presents the third
major caveat with respect to the test results. It was known from the outset that
different methods of seat inventory control will only have an observable impact when
demand for at least one of the fare classes available on a future flight is relatively
high. The greater the demand for a flight, the greater the impacts of effective
seat inventory control methods. The test results described above generally involved
flights with load factors above 80 percent. In general, about one-quarter of an
201
airline's flight legs depart with such high load factors, meaning that the positive
revenue impacts observed in this test would be reduced when spread over all of
the flight legs operated by an airline. However, high load factor flights generate a
disproportionately high percentage of total airline revenues, and it is the high load
factor flights that require the greatest amount of seat inventory control attention.
The shortcomings of the system tested and of the test itself highlight the dif-
ficulty of measuring the impacts of seat inventory control on airline traffic and
revenues. The performance evaluation of ABLS completed at Western Airlines as
part of this research nonetheless demonstrated that there exist potential benefits
in automated seat inventory control over manual methods. These benefits could
be even greater with further improvements to the estimation methods and decision
models employed. The EMSR decision framework developed earlier in this disserta-
tion provided the quantitative approach used by the automated system to derive the
fare class booking limits applied to the BATCH flights. The specific formulations
used in ABLS, however, did not incorporate fare class upgrade probabilities. The
test results provided numerous examples of the importance of upgrade potential to
revenue maximization. Furthermore, the demand and revenue inputs to the EMSR
framework had shortcomings that could be overcome through the development of
better estimation techniques.
202
Chapter 8
The preceding chapters have dealt with the many interrelated aspects of air-
line seat inventory management, from the theoretical underpinnings of the problem
through the development of a quantitative decision framework and its implemen-
tation. The focus of this work has been on the existing constraints of the seat
inventory control problem faced by airlines and their impacts on practical solution
approaches. References have also been made to the ways in which some of these
existing constraints might be overcome and solution methods improved through
further seat inventory control system development.
This chapter concludes this dissertation by relating the characteristics of the seat
inventory control problem, current and future, to those of airline systems designed
to manage seat inventories and maximize revenues. The findings and contributions
of this research are summarized first. Lessons from the development and implemen-
tation of the EMSR decision framework into an automated booking limit system
are also discussed, with an emphasis on the most immediate needs for system de-
velopment. Finally, some of the most important unanswered theoretical questions
surrounding the practice of differential pricing of airline fare products and limiting
the availability of seats sold at different price levels are presented as directions for
further research.
Much of the discussion in this dissertation has emphasized the way in which the
methods employed to manage seat inventories are related to the marketing policies,
203
operating procedures, and technical capabilities of an airline. The argument has
been made that improvements to the seat inventory control process must be made
in the context of the practical constraints posed by these elements. At the same time,
the development of better seat inventory control methods presents an opportunity
to eliminate some of these constraints.
From the outset, the need for airlines to practice seat inventory control came
about because of the differential pricing and market demand segmentation practices
that have evolved in the air travel marketplace. The importance of effective demand
segmentation through the definition of fare product attributes was described in
Chapter One. The price sensitivity of the consumer and the time sensitivity of the
trip being considered were the two criteria used to segment demand in the model
described in this thesis. These criteria were incorporated into a qualitative model of
the consumer choice process for air travel, and extended to reflect the reservations
process as viewed by airlines. This descriptive framework, based on the concepts
of travel disutility minimization, provides a foundation for the development and
estimation of quantitative consumer choice models.
Further work in this area will require significant improvements to airline data
availability. Data collection practices in the airline industry are surprisingly primi-
tive, yet the potential for the retrieval of useful data from airline reservations systems
is enormous. Millions of transactions in the form of flight availability inquiries, book-
ings, cancellations, and ticket purchases occur every day. Currently, even the most
advanced airline reservations systems can only provide running totals of cumulative
bookings by fare class for a future flight, net of all booking and cancellation activity
that occurs between the times when data is extracted to historical databases. Ac-
tivities at flight check-in can be as important to the accuracy of the data collected
as the booking process itself, yet few airlines' systems have the capability to provide
detailed information about no-shows, "go-shows", misticketed passengers, and class
of service upgrades.
Even more surprising than the lack of detailed reservations and traffic data is
the virtual non-existence of detailed revenue data at several major airlines. While
a few carriers have developed revenue databases, most others must rely on sporadic
samples of ticket coupons collected from passengers to estimate revenues and yields
204
by flight leg, O-D market, and fare class. Having focused on demand data while
taking revenue data for granted, airlines are now beginning to realize that making
optimal seat inventory control decisions requires revenue data that corresponds in
the level of detail to the demand data being collected.
The decision approach of the Expected Marginal Seat Revenue (EMSR) model
developed in this dissertation can be applied with little modification to most existing
reservations systems with multiple nested fare classes. The theoretical basis of the
EMSR model is a focus on expected future requests by fare class and the derivation
of optimal protection levels for upper fare classes, a requirement of nested fare class
inventories not addressed by previous work. The introduction of fare class overbook-
ing factors and upgrade probabilities into the derivation of these protection levels
represent further theoretical contributions to seat inventory control research.
By the turn of the decade, several of the largest airlines will have completed
the transition to seat inventory control on the basis of a "virtual nesting" approach
that limits seat availability by flight leg and passenger itinerary/ticket revenue,
not simply by fare class. The data requirements of this "next generation" of seat
inventory control systems will be substantially different from those discussed above,
and extensions to the mathematical methods already developed will be required. As
discussed in Chapter Six, the basic decision approach of the EMSR framework can
be extended to reservations systems that employ virtual nesting.
Practical contributions of this work include the application of the EMSR deci-
sion model to a dynamic revision process that incorporates actual bookings received
for a future flight into the adjustment of fare class protection levels and booking
limits. The revenue impacts of using the EMSR dynamic adjustment routine were
demonstrated in the evaluation of automated seat inventory control at Western Air-
lines. Given that most airlines still rely almost exclusively on human judgement in
making seat inventory control decisions, there is great potential for using mathemat-
ical models to assist in this decision process. Airlines must recognize, however, that
205
even the most sophisticated mathematical approaches will require the intervention
of skilled analysts.
206
and reservations systems. Both of these tasks represent medium-term objectives for
airlines that hope to improve their revenue management systems. The most complex
and longest term objective involves the concurrent reconfiguration of reservations
systems to manage seat inventories by passenger itinerary and ticket revenue, and
the development of mathematical decision models to forecast future demand and to
maximize system revenues in the context of these reservations systems.
The need for further theoretical and empirical research related to airline seat
inventory control and revenue management depends to a large extent on the future
evolution of the "problem", as defined by the practical constraints discussed in this
dissertation. Most important of these constraints is the specific technical approach
to seat inventory control that is adopted by airlines, the existing approach being fare
class/flight leg control. The current status of seat inventory control practices and
the planned development of more sophisticated approaches to revenue management
suggest two distinct yet interrelated areas for further research.
The mathematical models required to determine optimal booking limits for fu-
ture flights represent the first research area of significance to the future development
of seat inventory control systems. In this thesis, the problem presented by the nested
fare class inventories currently used by most airlines to control bookings by flight
leg was addressed. The EMSR framework was developed for application specifically
to the problem as defined by this approach to seat inventory control.
The network formulations reviewed in Chapter Four provide directions for the
development of optimization models that can take into account the true impacts on
airline system revenues of requests for different fare products and flight itineraries.
207
Further work on these approaches might reduce the size of the formulations involved
and the computing resources required. Concurrent improvements to computing
capabilities could lead to the implementation of system-wide revenue maximization
models, in which case the basic approach to controlling seat inventories might have
to be changed once again.
The estimation of future demand for flights and/or O-D itineraries requires the
development of forecasting models to predict forthcoming requests at various points
in time prior to departure. Extensive empirical analysis of airline reservations and
boarding data is necessary to develop such models and to answer the following
questions:
2. How can seasonal variation from past years as well as current traffic growth
trends be incorporated into estimates of demand by fare class for future flights?
This last question will require analysis of conditional probabilities and densities of
demand for homogeneous subsets of the total demand in a market. As mentioned
earlier, there is a practical need for generalizable forecasting models that can be
applied to a variety of markets and time periods. Given the limitations of existing
208
airline demand data, deriving generalizable conclusions from empirical analysis will
be difficult.
Estimating fare class upgrade potential will require empirical analysis of histor-
ical reservations data, to determine aggregate upgrade behavior. Once again, the
limitations of the data collected from existing reservations systems will make any
such estimates subject to numerous qualifications. Disaggregate data on fare class
upgrades would have to be collected from surveys of individual passengers and/or
reservations transactions. The key to deriving accurate estimates will be analysis
of passenger behavior under different conditions and in different markets, with the
objective of developing a usable model of passenger choice.
Estimates of horizontal choice shifts will become more critical as the scope of
the seat inventory control problem addressed by optimization models is expanded.
The time dimension of the seat inventory control problem has yet to be addressed
in the context of system revenue maximization. With the growth of "mega-airlines"
that offer a wide variety of departure times each day in a single market on flight
itineraries through several connecting hubs, the revenue and traffic implications
of "horizontal" choice shifts by passengers to alternate flights and times can be
substantial. To further complicate the problem, the times, fare products offerings,
and seat availability of competitors' flights in the same O-D market all affect the
revenue maximization objectives of each individual airline.
209
different rates, particularly those associated with non-refundable and partially re-
fundable low-priced fare products must be assessed. This information should be
incorporated into the revenue-based decision approach of the EMSR framework,
such that overbooking limits are determined in conjunction with the optimal fare
class booking limits.
At this point in time, the need for better demand modelling is greater than
the need for more sophisticated optimization methods. Any further development
of optimization methods should be directed toward the immediate needs of existing
leg-based seat inventory control systems and the imminent needs of virtually nested
systems. Concurrent development of improved optimization models and demand
estimation techniques would be desirable, so that the requirements of the former
can be addressed by the latter. Above all, the emphasis in further research must
continue to be on the limitations posed by reservations system capabilities, data
availability, and the rapidly changing nature of airline competitive practices.
Recent developments in airline seat inventory control have focused on the per-
sonnel and database management systems required. Potentially the largest gains
from more sophisticated seat inventory control techniques are yet to be realized,
as resprvations systems are revised and linked together, innovations to allow O-D
seat inventory control are implemented, and the use of decision models is expanded.
The potential benefits to both the airline industry and the traveling public are great.
Airlines can look forward to better management of their available capacity, higher
load factors and, higher total revenues. The traveling public, in turn, can look
forward to greater access to available seats and a wide variety of fare products, as
airlines gain greater control over their pricing and distribution practices in a highly
competitive marketplace.
210
References
211
15. 'ibid, pp 101-125.
16. Joseph P. Schweiterman, "Fare is Fair in Airline Deregulation", AEI Journal
on Government and Society, July-August 1985, pp 32-38.
17. ibid, p. 38.
18. R.W. Dodd, op cit.
19. Joseph P. Schweiterman, op cit.
20. R.W. Dodd, op cit.
21. Peter P. Belobaba, "TWA Reservations Analysis: Demand Distribution Pat-
terns", Flight Transportation Laboratory, MIT, Cambridge MA, April 1985.
22. Fares available with no advance purchase for travel on Trans World Airlines,
August 30, 1986.
23. Walter Nicholson, Microeconomic Theory: Basic Principles and Eztensions,
2nd Edition, The Dryden Press, 1978, Chapter 3.
24. Moshe Ben-Akiva and Steven R. Lerman, Travel Behavior: Theories, Models
and Prediction Methods, Chapter 5.
25. H. Simon, "A Behavioral Model of Rational Choice", in Models of Man, John
Wiley Publishers, New York, 1957.
26. R.D. Luce and P. Suppes, "Preference, Utility, and Subjective Probability",
in Handbook of Mathematical Psychology, Vol. 3, Wiley, New York, 1965,
Chapter 19.
27. Marvin L. Manheim, Fundamentals of TransportationSystems Analysis: Basic
Concepts, MIT Press, Cambridge MA, 1979.
28. K.J. Lancaster, "A New Approach to Consumer Theory", Journal of Political
Economy, Vol. 84, 1966, pp 132-157.
29. Michael D. Meyer and Eric J. Miller, Urban Transportation Planning: A
Decision-OrientedApproach, McGraw-Hill, New York, 1984, pp 257-258.
30. Barry C. Smith, "Reservation Recapture Rate Analysis", American Airlines,
Presentation to AGIFORS Reservations and Traffic Control Study Group,
New York City, April 14, 1986.
31. ibid, p. 2.
32. H.R. Thompson, "Statistical Problems in Airline Reservation Control", Oper-
ational Research Quarterly, Vol. 12, pp 167-185.
33. Marvin Rothstein, "OR and the Airline Overbooking Problem", Operations
Research, Vol. 33, No. 2, March-April 1985, pp 237-248.
34. Marvin Rothstein, "Stochastic Models for Airline Booking Policies", Ph.D.
Thesis in the Graduate School of Engineering and Science, New York Univer-
sity, 1968.
35. Robert Crandall, Chairman - American Airlines, quoted in "Life Under
Deregulation The American Way", Travel Weekly, May 31, 1985.
212
36. ibid.
37. T.G. Plaskett, Senior Vice-President Marketing - American Airlines, Pre-
sentation to Shearson, Lehman Brothers Airline Industry Seminar, Key Largo
FL, February 14, 1986.
38. "Airline Observer", Aviation Week and Space Technology, April 1, 1985, p.
35.
39. "Yield Managers Now Control Tactical Marketing", Lloyd's Aviation Econ-
omist, May 1985, pp 1 2 - 1 3 .
40. "The Optimal Partitioning of an Airplane's Seating Capacity", Sales Tech-
nology Dept., Boeing Commercial Airplane Company, Seattle, WA, February
1982.
41. Oli B.G. Madsen, "Booking Policy for Flights with Two Passenger Types",
European Journal of Operation. Research, forthcoming.
42. "The Optimal Partitioning of an Airplane's Seating Capacity", op cit.
43. Dennis F.X. Mathaisel and Herve Duchesne de Lamotte, "Optimization of Air
Transport Yields and Capacity", Memorandum M83-1, Flight Transportation
Laboratory, MIT, Cambridge MA, April 1983.
44. Dennis F.X. Mathaisel, "Optimization of Air Transport Yields and Capacity:
A Tactical Model", Memorandum M83-3, Flight Transportation Laboratory,
MIT, Cambridge MA, May 1983.
45. Fred Glover, Randy Glover, Joe Lorenzo and Claude McMillan, "The Passen-
ger Mix Problem in the Scheduled Airlines", Interfaces, Vol. 12, No. 3, June
1982, pp 73-79.
46. "The Optimal Partitioning of An Airplane's Seating Capacity", op cit.
47. Douglas Aircraft Company, "Electronic Bulkhead Seat Allocation",
McDonnell-Douglas Corporation, Long Beach CA.
48. Ewalds D'Sylva, "0 and D Seat Assignment to Maximize Expected Revenue",
Sales and Technology Dept., Boeing Commercial Airplane Company, 1982.
49. Richard D. Wollmer, "An Airline Reservation Model for Opening and Closing
Fare Classes", Unpublished Internal Report, McDonnell-Douglas Corporation,
Long Beach CA, 1985.
50. Richard D. Wollmer, "A Hub-Spoke Seat Management Model", Douglas Air-
craft Company Unpublished Report, McDonnell-Douglas Corporation, Long
Beach CA, 1986.
51. Kenneth Littlewood,"Forecasting and Control of Passenger Bookings", AGI-
FORS Proceedings 1972, pp 95-117.
52. A.V. Bhatia and S.C. Parekh, "Optimal Allocation of Seats by Fare", Pre-
sentation by Trans World Airlines to AGIFORS Reservations Study Group,
1973.
53. Helmut Richter, "The Differential Revenue Method To Determine Optimal
Seat Allotments By Fare Type", AGIFORS Symposium Proceedings 1982, pp
339-362.
213
54. Michael Mayer, "Seat Allocation, or A Simple Model of Seat Allocation Via
Sophisticated Ones", AGIFORS Symposium Proceedings 1976, pp 103-125.
55. Bernard Titze and Raimund Griesshaber, "Realistic Passenger Booking Be-
havior and the Simple Low-Fare/High-Fare Seat Allotment Model", AGIFORS
Symposium Proceedings 1983, pp 197-223.
56. Jorn Buhr, "Optimal Sales Limits for 2-Sector Flights", AGIFORS Symposium
Proceedings 1982, pp 291-303.
57. Ken Wang, "Optimum Seat Allocation for Multi-Leg Flights With Multiple
Fare Types", AGIFORS Symposium Proceedings 1983, pp 225-237.
58. "Boeing Promotional Fare Management System: Analysis and Research Find-
ings", Sales Technique Department, Boeing Commercial Airplane Company,
Seattle, WA July 1982.
59. J.J. Beckman, "Decision and Team Problems in Airline Reservations", Econo-
metrica, Vol. 26, 1958, pp 134-145.
60. M. Rothstein, "An Airline Overbooking Model", TransportationScience, Vol.
5, 1971, pp 180-192.
61. "Boeing Promotional Fare Management System", op cit.
62. Peter P. Belobaba, "TWA Reservations Analysis: Demand Distribution Pat-
terns", op cit.
63. Joao Sa, "Reservations Forecasting in Airline Yield Management", Master's
Thesis in Flight Transportation, Massachusetts Institute of Technology, Febru-
ary 1987.
64. Tables of Z values of the cumulative normal distribution can be found in most
statistical textbooks, including, A.M. Mood, Introduction to the Theory of
Statistics, McGraw-Hill, 1950.
65. John M. Chambers, William S. Cleveland, Beat Kleiner, and Paul A. Tukey,
Graphical Methods for Data Analysis, Chapter 6: Assessing Distributional
Assumptions About Data, Wadsworth International, Belmont CA, 1983.
214