You are on page 1of 15

MULTIPERIOD AIRLINE OVERBOOKING WITH A SINGLE FARE CLASS

RICHARD E. CHATWIN
Applied Decision Analysis, Inc., Menlo Park, California (Received March 1994; revision received April 1996; accepted February 1997) Consider a multiperiod airline overbooking problem that relates to a single-leg ight and a single service class. Passengers may cancel their reservations at any time, including being no-shows at ight-time. At that time, the airline bumps passengers in excess of ight capacity and pays a penalty for so doing. We give conditions on the fares, refunds, and distributions of passenger demand for reservations and cancellations in each period, and on the bumping penalty function, that ensure that a booking-limit policy is optimal, i.e., in each period the airline accepts reservation requests up to a booking limit if the number of initial reservations is less than that booking limit, and declines reservation requests otherwise. The optimal booking limits are easily computed. We give conditions under which the optimal booking limits are monotone in the time to ight departure. The model is applied to the discount allocation problem in which lower fare classes book prior to higher fare classes.

ffective yield management can save airlines hundreds of millions of dollars each year (Smith et al. 1992). Two of the more important among the airline yield management problems are the overbooking and the seat- or discount-allocation problems. In this paper we introduce two closely related models that address the airline overbooking problem for a single-leg ight with a single fare class. The second model can also be applied to the seatallocation problem in which the airline determines when to refuse reservation requests from lower fare class customers in order to protect seats for subsequent requests from higher fare classes. The two overbooking models presented in this paper are dynamic in nature and include customer reservation requests, cancellations, and no-shows explicitly. They are dynamic because in determining the booking rules the models consider not only the reservations currently on hand and the likelihood of such reservations canceling prior to ight time or being a no-show at ight time, but also the possibility of future customer reservation requests and subsequent cancellations. While dynamic models of airline (and the related hotel) overbooking have been presented previously, in this paper we extend the earlier work by developing the theory of the structure of the optimal solution. The major advance is an investigation of the circumstances under which a bookinglimit policy is optimal, i.e., in each period the airline accepts reservation requests up to a booking limit if the number of initial reservations is less than that booking limit, and declines reservation requests otherwise. By applying the theory of total positivity we are able to give sufcient conditions for the optimality of booking limit policies. These conditions are less restrictive than those imposed by previous researchers, and the results obtained are more powerful. For example, we allow the distribution of requests for reservations to depend on the number of

current bookings, and we allow for more general cancellation distributions than simply the binomial distribution that arises when one assumes that cancellations occur according to a Bernoulli process. We also point out some situations in which booking-limit policies are not optimal, contrary to what has been assumed in previous work. The second model is applied to the discount-allocation problem. Here multiple fare classes for a single-leg ight are considered, and cancellations are not allowed. This model extends previous work by allowing the number of reservation requests in a fare class to depend on the number of bookings in the earlier fare classes. No assumptions are made on the order in which the fare classes book. We also point out that this problem was previously solved in the context of inventory control by Topkis in 1968. The remainder of the paper is arranged as follows. In Section 1 we introduce the airline overbooking and discount-allocation problems and review the current literature. In Section 2 two multiperiod overbooking models (the Stationary-Fares and the Nonstationary-Fares Models) are introduced, and their distinctive characteristics are highlighted. The Stationary-Fares Model is the subject of Section 3, and the Nonstationary-Fares Model is the subject of Section 4. Under certain assumptions on the ighttime revenue function and on the potential reservations (the sum of the numbers of initial reservations plus reservation requests in a period) and cancellation distributions, we show that booking-limit policies are optimal in both cases, and that the optimal booking limits are easily computed. In Section 5 we show how these models may be adapted so as to explicitly account for no-shows, norecords, walk-ups, and stand-bys. In Section 6 monotonicity properties of the maximum revenue functions and the optimal booking limits are investigated. Under reasonable conditions on the fares and refunds, we show that the optimal booking limits fall toward ight-time. In Section 7

Subject classications: Dynamic programming, Markov, nite state: Markov decision model for airline overbooking. Inventory/production: uncertain yield and demand. Transportation: airline seat inventory control, yield management. Area of review: DISTRIBUTION. Operations Research Vol. 46, No. 6, NovemberDecember 1998

805

0030-364X/98/4606-0805 $05.00 1998 INFORMS

806

CHATWIN Most researchers have discovered that direct implementation of the obvious dynamic programming solution techniques leads to algorithms that are computationally intractable when applied to problems of realistic size (for example, see Alstrup et al. 1985, 1986a and b, 1989). In general, two approaches have been used to overcome this difculty: the rst is to make restrictive assumptions or to suppress certain elements of the problem, and the second is to use heuristics that lead to suboptimal, but easily implemented, rules. Both these approaches can provide approximate solutions to restricted or even realistic versions of the problem, but they do not provide insights into the structure of the optimal policy, nor do they suggest what the effects of changes in parameters of the problem might be. An alternative approach is to attempt to study the structure of the optimal solution in the hope that this might provide insight into how to easily implement the optimal policy. This approach has been employed by Liberman and Yechiali (1978) in the context of the hotel overbooking problem. The approach is also used by Wollmer (1992) in the context of the multifare-class discount-allocation problem for airlines, in which lower fare classes book rst and there are no cancellations or no-shows. This is the approach that will be employed in this paper. Its advantages are that it shows how to nd optimal policies that are easily implementable, it provides insight into the structure of the optimal solution, and it enables investigation of the effects of changes of parameters of the model. Specically, using this approach we show that a booking-limit policy is optimal, and that the optimal booking limits are easily calculated and are nondecreasing as a function of the time to ight departure. 1.2. Airline Seat or Discount Allocation In this paper we also address the discount-allocation problem in which there are multiple fare classes but no cancellations (and hence no overbooking). The discountallocation problem arises from the airlines practice of selling similar seats within the same cabin for a given ight leg at different prices and with different restrictions. The airline must protect some seats from lower revenue fare classes in order to satisfy later arriving demand in the higher revenue fare classes; the question is to what extent this protection should be employed. Early airline reservation systems simply allocated a certain number of seats to each fare class and did not allow higher fare classes access to the inventories of lower fare classes. Curry (1992) suggests the alternative approaches of overallocation and parallel fare-class nesting. A better approach is to allow for series fare-class nesting in which any higher fare class may take seats from the inventory of any lower fare class. This approach has been widely studied in the literature. Littlewood (1972) proposed a rule for the two-period, two-fare-class problem in which low-fare customers book prior to high-fare customers. His rule was shown to be optimal by Bhatia and Parekh (1973) of TWA, and later by Richter (1982) of Lufthansa. Belobaba (1987a

we show how our second model can be applied to the discount-allocation problem. Our model formulation is closely related to several previously published models and we obtain similar results: namely that booking-limit policies are optimal, and that the optimal booking limits rise towards ight time when lower fare classes book rst. In Section 8 we demonstrate that booking-limit policies need not be optimal when there is stochastic dependence between the demand of different fare classes, such as when there is diversion. This is contrary to the published literature, in which it is common practice to examine controls only in the form of booking-limit policies.

1. INTRODUCTION AND LITERATURE REVIEW 1.1. Airline Overbooking The airline overbooking problem arises from the propensity of airline customers, having made a reservation for a ight, to subsequently cancel that reservation or to be a no-show, i.e., to simply fail to show up at ight time. In anticipation that cancellations and no-shows will occur, the airline may overbook the ight, thereby already reselling a seat vacated by a customer who cancels or is a no-show. The potential extra revenue from overbooking a ight must be balanced against its costs. These arise because in overbooking the airline runs the risk of not having sufcient capacity to accommodate all its customers, in which case it must deny reservation requests or deny boarding to (bump) some of them, thereby incurring a cost measured both nancially and in loss of goodwill. Many authors have discussed the airline overbooking problem and the related hotel overbooking problem. A history of research in this area is given by Rothstein (1985) and is further discussed in Chatwin (1993, Chapter 1). Most early work on the overbooking problem, beginning with Beckmann (1958), employed a static one-period model with reservation requests, booking, and nally cancellations. American Airlines implemented (in 1976, with a major revision in 1987) such a model with additional constraints to ensure that the level of service was not overly degraded (Smith et al. 1992). More recent work on the static overbooking problem includes that of McGill (1989) and Bodily and Pfeifer (1992). A drawback of the aforementioned models is that the dynamic nature inherent in the reservations process is not considered. Several authors have developed dynamic approaches to the overbooking problem. Rothstein (1968, 1971) rst formulated the airline overbooking problem as a dynamic program, and he later did the same for the similar hotel overbooking problem (Rothstein 1974). Hersh and Ladany (1978) model ights with an intermediate stop using dynamic programming, and Ladany (1976, 1977) develops models for the hotel/motel industry that consider extensions to two or more fare classes. In this paper we employ a dynamic programming approach to the airline overbooking process.

CHATWIN and b) generalized Littlewoods rule to a multiperiod model to obtain allocations for more than two fare classes. Belobabas method (the Expected Marginal Seat Revenue method or EMSR) does not nd the optimal allocation but has been shown through simulations to perform very well. Belobaba (1992) has recently developed an enhanced version of the EMSR model, which may be computed more efciently than, and appears to outperform, the original model. McGill (1989), Curry (1990), Wollmer (1992), and Brumelle and McGill (1993) all prove that a booking-limit policy is indeed optimal under the assumption that lower fare classes book rst. Lee and Hersh (1993) and Robinson (1995) prove the same result without the assumption regarding the order in which the classes book. The model of Lee and Hersh allows for multiple seat bookings, which are a practical issue in airline seat inventory control. The model developed in this paper is most similar to that of Wollmer. We are able to show that booking-limit policies are optimal without assuming that lower fare classes book rst, and allowing the distribution of reservation requests to depend on the number of current bookings. Airline discount-allocation problems with two fare classes in which there is diversion, that is, customers who would be willing to pay the full fare divert to the discount fare class when such fares are available, have been studied by Pfeifer (1989), McGill (1989), Brumelle et al. (1990), Weatherford (1992), and Weatherford et al. (1993). This problem is discussed in Section 8, where it is demonstrated that booking-limit policies need not be optimal. 1.3. Related Problems in Airline Yield Management Factors that increase the complexity of the overbooking and discount-allocation problems include: (1) combined overbooking and discount-allocation: consideration of multiple fare classes and cancellations; (2) passengers whose itineraries involve multiple ights; and (3) interactions with other ights, both those of the same airline and those of competing airlines. The multifare-class overbooking problem is considered in Chatwin (1993, Chapter 4). This problem is addressed using techniques similar to those applied in this paper by Janakiram et al. (1994). A related overbooking model for the hotel industry that considers different types of hotel guests (6 p.m. hold reservations, walk-ins, and guaranteed reservations) is studied by Bitran and Gilbert (1996). Perhaps the hottest area of new research is the bid-pricing approach to the trafcmanagement (or origin-and-destination seat-allocation) problem. Here the airline must determine how to allocate seats on various ight legs between passengers with differing itineraries. Passengers whose itineraries include several ight legs will bring more net revenue than passengers who y only a single leg, although the revenue per mile will be less. Thus the airline must make trade-offs similar to, but more complex than, those encountered in the discountallocation problem. Very briey, using the bid price approach the airline computes bid prices for each ight leg. Then a reservation request for a multileg trip is accepted

807

only if the fare exceeds the sum of the bid prices for each of the ight legs. The literature on bid-pricing is somewhat sparse: papers addressing this issue include Colville and Phillips (1992), Smith (1993a and b), Phillips (1994), Swan (1994), and Talluri and van Ryzin (1995). Talluri and van Ryzin demonstrate that bid pricing is not an optimal form of control in general, but it is likely to be close to optimal when capacities and sales volumes are large. In this paper we restrict attention to a single-leg ight and consider the overbooking and discount-allocation problems separately; the other factors are not considered. 2. FORMULATION OF THE AIRLINE OVERBOOKING PROBLEM This section presents two closely related airline overbooking models that employ a discrete-time approach. Assume that the set of allowable reservations levels is of the form S {0, 1, . . . , R} where R is assumed to be nite, a minor restriction in practice. There is no theoretical reason why S should not be permitted to be the set {0, 1, . . . }, but since in practice an articial upper bound will always be placed upon the number of reservations on hand at any time, and some differentiation between the two cases would be necessary in the analysis, we restrict our attention to R nite. We call the two models the Stationary-Fares Model and the Nonstationary-Fares Model, respectively. The former places restrictive assumptions on the fares and refunds, since there is no penalty to customers who cancel or are no-shows and the fare value is constant over all time periods. In contrast, the Nonstationary-Fares Model makes no such assumptions. The fares and refunds may vary over time, but any refund received by a canceling customer is unrelated to the fare paid by that customer. This assumption might seem untenable in the airline industry, though it is well accepted in other commodities markets. (Janakiram et al. 1994 show how to assess all expected costs from cancellations and no-shows at the time a booking is made. They use a technique rst applied to the control of queues in Lippman and Stidham 1977.) However, there are several scenarios in which this more general fare structure is realistic and does prove to be advantageous. These will be described in the sequel. The penalty for the more general fare structure is that in the Nonstationary-Fares Model it is necessary to make somewhat more restrictive assumptions about the ight-time revenue function and the demand, cancellation, and noshow distributions. Thus the distinction between the two models lies in the different trade-off between the strength of the assumptions on the fare structure on the one hand, and on the ight-time revenue function and the demand, cancellation, and no-show distributions on the other. 3. STATIONARY FARES An airline seeks a multiperiod reservation policy for a ight with seat capacity C that maximizes its expected revenue from the ight. The model relates to a xed nonstop

808

CHATWIN This encompasses the case in which reservation requests and cancellations occur singly. Let prN be the (conditional) distribution of the number of potential reservations, that is, the sum of the number r of reservations at the beginning of the period N and the number of reservation requests in period N. (Here and in the sequel the word distribution is used to describe a probability distribution, or equivalently a probability mass function, e.g., as in the binomial distribution, and should not be interpreted as describing a cumulative distribution function.) Assume that this distribution depends only on N and r reservar. Thus, pN is the probability of there being t rt tion requests during period N given that there are r reservations at the beginning of the period. Similarly, assume N that the (conditional) cancellation distribution q s depends only on the period N and upon the number s of reservations immediately after bookings in the period. Thus, qN is sr the probability of there being r reservations after cancellations in period N given that there are s reservations after bookings in that period. In the sequel, reference to a potential reservations distribution p carries the implication 0 for all r t. Similarly, reference to a cancelthat prt 0 for all s r. lation distribution q implies that qsr Given the ight-time revenue function V0, dene inductively the maximum expected revenue functions immediately prior to reservation requests, bookings, and cancellations, respectively, in each period. Let VN be the r maximum expected revenue when there are r reservations immediately prior to reservation requests in period N; let UN be the maximum expected revenue when there are r rt reservations immediately prior to reservation requests in period N, and there are t r reservation requests during the period (of course the airline may accept at most R r of these reservation requests); and let vN be the maximum s expected revenue when there are s reservations immediately prior to cancellations in period N. Then the dynamic-programming recursions relating the revenue functions are, for each N 1, v sN
N U rt r N q sr V rN 1

ight and a single class of service, e.g., coach class. The number of reservations on hand at any given time prior to the departure of the ight is the state of the reservations system for the ight at that time. The airline permits a maximum of R reservations at any time. Then the state-space is S {0, 1, . . . , R}. The system changes state according to transition probabilities that depend on the distributions of customer reservation requests and cancellations, and on the airlines booking policy. Our assumptions imply that the transition probabilities are Markov and depend upon the time to ight departure. There are two assumptions that are central to the Stationary-Fares Model. The rst is that any customer who makes but subsequently cancels a reservation receives a full refund of the fare paid. The second is that the fare for a seat on the plane is independent of the period in which the reservation is made. Then the net revenue to the airline from a customer who makes, and later cancels, a reservation is zero. The only net revenue earned by the airline is from fares paid by passengers who occupy seats on the plane at ight time. The only net revenue losses incurred by the airline are for compensation of customers with valid reservations who are denied boarding at ight time because of overbooking. In this setting, since discounting is ignored (which is not unreasonable given the typically short time between reservations and ight time), there is no loss of generality in assuming that fares are paid at ight time. When a ight is overbooked the airline compensates customers with reservations who are denied boarding. The ight-time revenue V0 obtained by the airline at r ight time depends on the number r of customers with reservations who show up for the ight. Assume that V0 is r a quasi-concave function that attains its maximum at C, C, V0 is i.e., on r C, V0 is nondecreasing, and on r r r nonincreasing. To motivate this assumption, consider the following fare structure. Suppose that each passenger who boards the plane pays a xed fare f to the airline. If the airline denies boarding to x customers, then the cost to the airline is c( x), where c is an increasing function and c(0) 0. If an auction is held to determine which customers are to be denied boarding, then c is likely to be convex. Then the ight-time revenue function V0 will be V r0 rf Cf c r C if r if r C, C. (1)

(cancellations), (booking decisions), and (reservation requests).

(2) (3) (4)

r s t R

max v sN
N N p rt U rt

V rN

t r

C and is deNotice that V0 is linearly increasing on r r creasing on r C, and hence satises the required assumptions. Denote by N the number of periods prior to departure. The airlines decision during each period is whether or not to accept each reservation request during that period. Call this process booking. Suppose that in each period the order of events is reservation requests, bookings, and cancellations. Note that we assume the airline observes all the reservation requests in a period before making the decision as to how many to book and how many to decline.

Here the symbol represents the meet. To give a pren , cise denition: In a partially ordered set P, e.g., P the symbols and represent binary operations on P, where for t, u P, t u (respectively, t u) is the greatest lower (respectively, least upper) bound of t and u. Call a policy a booking-limit policy if in each period N there is a booking limit bN such that the decision in that period is to accept, i.e., book, reservation requests up to the level bN if the number of reservations immediately prior to reservation requests is less than bN, and to reject reservation requests otherwise. Our goal is to give conditions under which a booking-limit policy is optimal.

CHATWIN To this end, we present some known results that are particularly relevant to the development of the theory that follows. In the sequel assume that X and Y are open intervals of the real line or intervals of integers, i.e., sets of consecutive integers. Then a function K : X Y 3 is said to be totally positive of order r (abbreviated TPr) if for all x1 xi x2 xm, y1 X, y j Y; 1 m x1, x2, . . . , xm y1, y2, . . . , ym K x1, y1 K x2, y1 K xm, y1 K x1, y2 K x2, y2 K xm, y2 K x1, ym K x2, ym K xm, ym 0. y2 r, ym,

809

we have the inequalities K

straight) line at most twice and, in the case where there are two intersections, the arrangement of signs is , , . We now give a theorem on additivity whose two parts are specializations of results due to Karush and Vazsonyi (1957) and to Karush (1959), respectively. Let X be a subinterval of , f1, . . . , fn functions with domain X, and F(a, b) the maximum of in 1 fi( xi) subject to a x1 . . . xn b and xi X for all i. Recall that a real-valued function n if for all y, z g is said to be additive on a lattice L L, g( y z) g( y z) g( y) g( z). It is known (Lorentz 1953, Topkis 1978) that g is additive on a rectangle L in n if and only if g( y) jn 1 gj( yj) for all y ( yj) L for some functions g1, . . . , gn. Theorem 2. Additivity of the Maximum Function when Variables are Linearly Ordered. a. (Karush and Vazsonyi) The Quasi-Concave Case. If each fi is quasi-concave, then F is additive. Indeed, F(a, b) P(a) Q(b) on {(a, b) : a b}, where P is nonincreasing and Q is nondecreasing. b. (Karush) The Concave Case. If each fi is concave, then F is additive and concave. Indeed, F(a, b) P(a) Q(b) on {(a, b) : a b}, where P is nonincreasing and concave and Q is nondecreasing and concave. We will apply these results in the case n 1, i.e., when F(a, b) is the maximum of f( x) subject to a x b and x X. Let x* argmaxx X f( x). When f is quasi-concave it is easily seen that F(a, b) P(a) Q(b), where P(a) f(a x*), which is nonincreasing (and concave when f is concave), and Q(b) f(b x*) f( x*) which is nondecreasing (and concave when f is concave). Call a potential reservations distribution p in some period stochastically increasing if pr is stochastically increasing in r. We would expect the potential reservations distributions to be stochastically increasing in all practical situations because this implies that the more reservations we have, the more potential reservations we are likely to have. Call a cancellation distribution q in some period totally positive of order 3 (TP3) if the matrix (qsr) is TP3. Again, we would expect this property to hold in most practical situations. These assumptions on the potential reservations and cancellation distributions are discussed in more detail below. Theorem 3. Optimality of Booking-Limit Policies: Stationary Fares. Suppose that the ight-time revenue function is quasi-concave, and that in each period the potential reservations distribution is stochastically increasing and the cancellation distribution is TP3. Then a booking-limit policy is optimal, and in each period the maximum expected revenue functions immediately prior to reservation requests, bookings, and cancellations are, respectively, quasi-concave, additive, and quasi-concave. Furthermore, the set of optimal booking limits in period N is the interval of maximizers of the last function and is a subset of the interval of maximizers of the former function in that period.

The following theorem combines some results to be found in Karlin (1968, Proposition 3.1, p. 22; Proposition 3.2, p. 23; Theorem 1.5, p. 223; Theorem 3.1, p. 233). Theorem 1. (Karlin) Monotonicity, Quasi-Concavity and Concavity Preservation with TP2 and TP3 Densities. Let X and Y be sets of consecutive nonnegative integers and K : 1 for all x X. X Y 3 . Suppose that y Y K( x, y) Let f : Y 3 be a bounded function and consider the transformation g x
y Y

K x, y f y ,

X.

(i) If K is TP2 and f is nondecreasing, then g is nondecreasing. (ii) If K is TP3 and f is quasi-concave, then g is quasi-concave. ax b where a 0 and b are (iii) If y Y yK( x, y) real, K is TP3 and f is concave, then g is concave. These results follow from the variation-diminishing property of totally positive kernels. Without being too precise, this property can be described as follows. If K is TPr and f has at most r 1 sign changes, then g has no more sign changes than f. Further, if g has the same number of sign changes as f, then the values of f and g exhibit the same sequence of signs when their respective arguments traverse the domain of denition from left to right. The rst result of Theorem 1 follows on observing that f nondecreasing is equivalent to f having the property that for changes sign at most once, and if it does any , f( y) change sign the arrangement of signs is , . Then, be1 for all x X, g has cause K is TP2 and y Y K( x, y) changes sign at most the property that for any , g( x) once, and if it does change sign the arrangement of signs is , , which is equivalent to g nondecreasing. Similarly, the second and third results follow on observing that a characteristic property of quasiconcave (respectively, concave) functions is intersecting a horizontal (respectively,

810

CHATWIN In practice solving the model in real time is impractical due to the rapid response to reservation requests that is required. A practical application involves running the model off-line and using the results to develop simple decision rules by which booking decisions can be made. It is also important that the data requirements for storing the simple decision rules be not signicant due to the very large number of ights in the reservations system. The Stationary-Fares Overbooking Model, under the assumptions of Theorem 3, can be easily implemented as a simple decision rule. For each period (with a period being a day, for example) the only piece of information that need be stored is the optimal booking limit. In each period, reservation requests are accepted only if the number of current bookings is less than that periods booking limit. 3.2. Examples of Potential Reservations and Cancellation Distributions Here we discuss the assumptions on the potential reservations and cancellation distributions that are made in Theorem 3, and we give some examples of distributions that satisfy those assumptions. First, two examples of conditions on the potential reservations distribution p that are individually sufcient to ensure that it be stochastically increasing are: (1) the number of reservation requests in a period is independent of the number of reservations immediately prior to reservation requests in that period, i.e., pr is independent of r, in which case we say that p has independent increments; and (2) p is TP2. In the second of these examples the number of reservation requests in a period will depend on the number of reservations immediately prior to reservation requests in that period. Such a dependence could arise in a number of ways. For example, given a nite potential customer population, more reservations on hand would imply a smaller population left to generate reservation requests, and hence in probability a smaller number of such requests; scheduled events in a destination city would stimulate reservation requests in all periods prior to ight-time, and hence the more reservations on hand, the more reservation requests might be expected; and if customers whose requests are declined are more likely to make later requests (in the hope that cancellations have occurred) than general members of the population, then the more declined requests previously, the more future requests are to be expected: the number of reservations on hand can be used to estimate the number of previously declined requests. It might be argued, given the above examples, that the dependence between the number of reservation requests and the number of reservations on hand actually reects a more basic dependence between the number of reservation requests in different periods. This might indeed be true, but while the airline does know the number of reservations on hand, it cannot accurately measure the number of reservation requests in any period, since in general this number is censored due to the airline declining some such

Proof. The proof is by induction on N. By supposition V0 is quasi-concave. Suppose VN 1 is quasi-concave. Then, since qN is TP3, it follows from (2) and Karlins result that vN is quasi-concave. Hence the set m of maximizers of vN is an interval. Then it follows from (3) and Karush and m Vazsonyis result that UN is additive; indeed, for b
N U rt

v tN b

v rN b

N vb .

(5)

Thus the decision dened by the booking limit b is optimal in period N. Also it follows from (4) and (5) that V rN
t r N p rt v tN b

v rN b

v rN .

(6)

N Now vt b is nondecreasing in t, so since pN is stochastically increasing it follows from a well known result (Lehmann 1959, p. 73; Karlin 1960; Veinott 1965) that the rst term on the right-hand side of (6) is nondecreasing in r. Moreover that term is constant on r b. On the other hand the second term on the right-hand side of (6) vanishes on r b and is nonincreasing on r b. Consequently VN is quasiconcave and its maximizer set M is an interval containing b.

Remark 1. Dene m (respectively, m) to be the least (respectively, greatest) element of m, and similarly for M. If pN m 1, m 1 0, then M m, else M properly contains m. To see this, rstly observe that M m, for certainly M m so, on setting b m it follows from (6) that VN M vN M and VN m vN m. But M and m are both in VN m, whence vN M vN m, so M m. On M, so VN M the other hand, setting b m, it follows from (6) that if m, pN m 1, m 1 0 then VN m 1 VN m, whence M 0, then VN m 1 so that M m. But if pN m 1, m 1 m, so that M properly contains m. VN m, whence M In light of Theorem 3 and Remark 1, it is useful for future discussion to dene mN (respectively, MN) to be the interval of maximizers of vN (respectively, VN). Then the set of optimal booking limits in period N is precisely mN and is contained in MN. 3.1. Complexity, Implementation, and Storage Requirements The result of Theorem 3 is intuitively appealing. Furthermore, the proof of the theorem shows that the optimal booking limit bN in each period N is easy to calculate, being simply any maximizer of the quasi-concave function vN. Such a maximizer could be found efciently by performing a Fibonacci search. For each time period we can compute the expected revenue functions and the optimal booking limit in O(R2) time. Starting with VN 1 we compute vN from (2), then compute the optimal booking limit bN using a Fibonacci search. Finally we compute VN from (6) with the understanding that we suitably truncate the potential reservations distribution t R. Thus the running time of the algorithm for nding optimal booking policies for an N-period problem is O(NR2).

CHATWIN requests. Thus the model as described provides a reasonable method of accounting for any dependence between the number of reservation requests in different periods. Moreover, if this dependence is modeled directly, then in general booking-limit policies will not be optimal. This is discussed further, in the context of the NonstationaryFares Model, in Section 8. Examples demonstrating the nonoptimality of booking-limit policies are included there. The most important example of a cancellation distribution q that is TP3 is the binomial distribution. This example arises when each customer with a reservation immediately prior to cancellations in a period independently cancels during that period. the reservation with probability 1 )s r. The number of cancellations in Then qsr (s) r(1 r a period must depend on, and indeed never exceed, the number of reservations immediately prior to cancellations in that period. The forgetfulness property is the assumption that the number of cancellations depends only on the number of reservations on hand, and not, for example, on the length of time for which each reservation has been on record. That this assumption is reasonable has been statistically veried by a large-scale empirical study of customer cancellation behavior at Iberia Airlines, reported by Martinez and Sanchez (1970). Evidence that the binomial is a good model for the cancellation distributions has been provided by Thompson (1961), working with Tasman Empire Airways. 4. NONSTATIONARY FARES The Nonstationary-Fares Model is similar to the Stationary-Fares Model, with the essential difference being that in this model the fares and refunds are permitted to vary in a known manner with the time period. It is important to understand that this is not a pricing model in that the fares and refunds cannot be altered dynamically over time as information is acquired. Rather, the airline xes the fares and refunds at the beginning of the time horizon, although it may set different fares and/or refunds in each time period. There are natural fare/refund structures that have this property. Some examples are: (1) fares and refunds are stationary but the refund is smaller than the fare because of a cancellation penalty; (2) fares and refunds may differ, but are stationary except that cancellations in the nal period (no-shows) receive no refund; (3) fares are stationary but refunds are nonstationary, e.g., there is a cancellation penalty that gets larger toward ight time; (4) fares are nonstationary but refunds are stationary, e.g., fares get larger toward ight time but refunds are constant; (5) the airline has time preference for income, e.g., fares and refunds are stationary but the airlines preference for receiving fares earlier rather than later (and conversely for refunds) leads to a model in which fares and refunds decrease as ight time approaches; (6) the effects of ination would also cause stationary fares and refunds to be of different utility to the airline in different time periods. This model also assumes, as is the case in practice,

811

that the compensation paid to a customer who is denied boarding is independent of the fare paid. Note, however, that this still allows the airline to deny boarding to customers paying the lowest fares, a not uncommon practice. This model can also be interpreted as applying to the multifare-class problem. This application will be discussed in Section 7. 4.1. Optimality of Booking-Limit Policies Assume now that customers pay their fares when they make reservations and receive refunds at the times at which they cancel. Thus at ight time any passenger who boards the plane has already paid the airline for this privilege, and hence the airline receives no revenue from these passengers at this time, unlike in the Stationary-Fares Model. Therefore the ight-time revenue function V0 will consist only of the compensation paid by the airline to those customers who are denied boarding. As an example, suppose that the cost to the airline of denying boarding to x customers is c( x) where c is increasing and convex, with c(0) 0. When overbooking occurs, airlines, being concerned about the goodwill of their customers, will typically offer attractive compensation packages in order to encourage customers to voluntarily give up their seats. It is reasonable to expect that the cost of persuading a (k 1)st customer to give up a seat is greater than that of persuading a kth customer to give up a seat, which is why c is assumed to be convex. Thus the ight-time revenue function will be V r0 0 c r C if r if r C, C. (7)

Observe that V0 is nonincreasing and concave in r. These r will be the assumptions on V0. 0 and Denote the fare and refund in period N by fN N 0, respectively. Then the dynamic-programming rec cursions relating the revenue functions are, for each N 1, v sN
N U rt s r 0 N q sr V rN 1

s s

r cN r fN

(cancellations),

(8)

r s t R

max

v sN

(booking decisions), (9)

and V rN
t r N N p rt U rt

(reservation requests).

(10)

Our goal once again is to give conditions under which a booking-limit policy is optimal. We say a potential reservations distribution p has independent increments if the number of reservation requests in a period is independent of the number of reservations immediately prior to reservation requests in that period, i.e., prt pt r where p is the distribution of the number of reservation requests in the period and is independent of r. A distribution that both has independent increments and is TPr, is a Polya frequency function of order r (abbreviated

812

CHATWIN right-hand side of (12) vanishes on r b and is noninrfN is concave and creasing on r b. Consequently VN r its maximizer set M is an interval containing b. Given the additional assumptions, V0 is nonincreasing. Suppose VN 1 is nonincreasing. Then, since qN is TP3, it s follows from Karlins result and the facts that r qNr sr 0, that vN is nonincreasing. Since VN rfN and cN r attains its maximum at b, VN is certainly nonincreasing on r r b. Thus it remains to show that VN is nonincreasing on r r b. On this interval uN 0, so it follows from (12) that r V rN
t r N p rt v tN b

PFr). A (conditional) distribution p is said to have afne r for each r, where and are mean if t prtt 0, then p is said to have increasing afne constants. If 1, then p is said to have contracting afne mean; if also 0, replace afne by linear in mean. In the event that the preceding denitions. Suppose that a cancellation diss for each tribution q has afne mean. Then r qsrr s, where and are constants. Since q is a cancellation distribution it follows that q00 1, so that r q0rr 0; and s 1, so that 0 r qsrr s. Thus 0 and 0 r 0 qsr 1. We shall ignore the trivial case in which 0 and assume that q has contracting linear mean. Theorem 4. Optimality of Booking-Limit Policies: Nonstationary Fares. Suppose that the ight-time revenue function is concave, and that in each period, the potential reservations distribution either (1) has independent increments or (2) is TP3 with increasing afne mean, and the cancellation distribution is TP3 with contracting linear mean. Then a booking-limit policy is optimal, and in each period the maximum expected revenue functions immediately prior to reservation requests, bookings, and cancellations are all concave; the second of these is also additive. Furthermore, the set of optimal booking limits in period N is the interval of maximizers of vN sfN, and is a subset of the interval of s rfN. If, in addition, the ight-time maximizers of VN r revenue function is nonincreasing, and any potential reservations distributions of type (2) have contracting afne mean, then in each period the maximum expected revenue functions immediately prior to reservation requests and cancellations are nonincreasing. Proof. The proof is by induction on N. By supposition V0 is concave. Suppose VN 1 is concave. Then, since qN is TP3 and has afne mean, it follows from (8) and Karlins result that vN is concave. Hence the set m of maximizers of the sfN is an interval. Then it follows concave function vN s from (9), Karushs result, and the fact that nondecreasing concave functions of concave functions are concave, that m and r t UN is additive and concave; indeed, for b
N U rt

b fN

rf N

if r

b.

(13)

If pN has independent increments then (13) may be rewritten as V rN


x 0 N px v N r x b

fN

if r

b,

(14)

and the result follows since both ( x (b r)) and N v(r x) b are nonincreasing functions of r for each xed x, and fN 0. Alternatively, suppose pN is TP3 with contracting afne mean. Rewrite (13) as V rN
t r N p rt v tN b N p rt t

t r fN

b if r

t fN b. (15)

t r

N Now vt b and the term in braces on the right-hand side of 0, so the term in (15) are nonincreasing in t, and fN brackets on the right-hand side of (15) is nonincreasing in t. Thus, since pN is TP3, the rst term on the right-hand side of (15) is nonincreasing in r by Karlins result. Be0, the second term on the right-hand side of cause fN r ( (15) is also nonincreasing in r, since t r pNt rt for some constants 0 1 and 0. 1)r

Remark 2. As in the Stationary-Fares Model M m, and M m or M m (so that M m or M properly contains m), as pN m 1, m 1 is greater than or is equal to zero. The proof mirrors that of Remark 1. Remark 3. Observe that it is not necessary to assume that the ight-time revenue function is nonincreasing in order to prove that a booking-limit policy is optimal. However, since the ight-time revenue function incorporates only the overbooking costs we would expect it to be nonincreasing. If this is the case, then Theorem 4 shows that the maximum expected revenue functions at all earlier times are also nonincreasing. This is as expected, as a customer holding a reservation has already paid the fare, which thus represents a sunk benet to the airline. That customer is lling a reservation space, thereby preventing the airline selling it to another customer and obtaining additional revenue. Theorem 4 proves the equivalent result for the Nonstationary-Fares Model that Theorem 3 proved for the Stationary-Fares Model. As before booking-limit policies are optimal, and once again the calculation of a sequence

u tN uN t

u rN

rf N,
N

(11)

where (t b) f is nondecreasing and conN vr b (r b) fN vN bfN is cave in t; b nonincreasing and concave in r. Thus the decision dened by the booking limit b is optimal in period N. Also from (10) and (11) V rN rf N
t r N p rt u tN

N vt b and uN r

u rN.

(12)

rfN (and hence VN) is concave in r since the Then VN r r rst term on the right-hand side of (12) is concave in r when (1) pN has independent increments because then the N N x 0 pNur x; and sum may be written as t r pt ruN t x N when (2) p is TP3 with afne mean because Karlins result applies once again. Similar reasoning shows that on r b this rst term is nondecreasing. Further this term is constant on r b. On the other hand, the second term on the

CHATWIN of optimal booking limits {bN} may be easily accomplished since for each N, bN may be dened to be any reservation sfN attains its level at which the concave function vN s maximum. As before it is clear that the running time of the algorithm for nding optimal booking policies for an N-period problem is O(NR2). Remark 4. There have been very few attempts to take advantage of the properties that TP3 functions preserve quasi-concavity and concavity in order to prove results about the structure of optimal policies in inventory and related models. Karlin (1958a and b) uses these properties in both single-period and multiperiod inventory models, when assuming that the density of the demand for the stock in any period is PF , although his results were obtained under restrictive conditions. Porteus (1971) also assumes PF demand densities in each period of multiperiod inventory models, in order to prove some nice results regarding the optimality of generalized (s, S) policies when the ordering cost function is concave and increasing. Notzon (1970) uses these properties to produce a computational simplication for a minimax inventory model. 4.2. Discussion of the Assumptions on the Potential Reservations Distributions In Section 3.2 the assumptions on the potential reservations distributions made in Theorem 3 were discussed. These assumptions are strengthened in Theorem 4. The potential reservations distributions must either have independent increments or be TP3 with increasing afne mean. It is useful to give examples of potential reservations distributions of the second type. For example, suppose that the population of potential customers for a ight is nite, of size M say. Further suppose that in each period, those members of the population not already holding a reservation, independently request a reservation with probability . Then the potential reservations distribution is binomial, when there are r reservawith parameters M r and tions on hand, i.e., p rt M t r r
t r

813

4.3. Specialization to Stationary Fares When the fare is independent of the period in which the reservation is made, and any customer who makes, but subsequently cancels a reservation, or is a no-show, ref and cN f for all N 1, ceives a full refund, i.e., fN the optimal booking policy may also be found by using the Stationary-Fares Model. This section demonstrates a simple relationship between the maximum value functions obtained by the two models, and that (as expected) each gives the same optimal booking policy. To this end denote the maximum value functions obtained by the Stationary-Fares Model (Equations (1)(4)) v by V0, N, UN, and VN, respectively, and those obtained by the Nonstationary-Fares Model (Equations (7)(10)) by c V0, vN, UN, and VN, respectively. Let ( x) (respectively, c( x)) denote the cost to the airline of denying boarding to x customers in the Stationary-Fares Model (respectively, Nonstationary-Fares Model). In the Nonstationary-Fares Model any customer denied boarding has already paid a fare f and should have this refunded as well as receiving compensation for the inconvenience of being denied boarding. In the Stationary-Fares Model any customer denied boarding has yet to pay a fare so should receive only the denied boarding compensation. Therefore assume that c x x c xf. (16)

Then from (1) and (7) it follows that V r0 V r0 rf. (17)

Now it is easily seen that sN v N U rt and V rN V rN rf. (20) v sN


N U rt

sf, rf,

(18) (19)

Given Equations (16)(20) it is a trivial consequence of the results of the preceding sections that the optimal booking policies obtained by the two models are identical when the ight-time revenue function is concave. 5. INCORPORATING NO-SHOWS, NO-RECORDS, WALK-UPS, AND STAND-BYS This section discusses how the Stationary- and Nonstationary-Fares Models might be interpreted so as to incorporate no-shows, no-records, walk-ups, and stand-bys. Noshows are those customers who hold reservations but do not show up for the ight; no-records are customers who show up at ight time with reservations of which the airline has no record; walk-ups are customers without reservations who show up at ight time wishing to travel on the ight; and stand-bys are customers who buy tickets at (possibly reduced) rates with the restriction that they may travel on the next ight with available seats only after all reservations for that ight have been honored.

M t

This is TP3, and if Tr is distributed according to pr , then )r M , which is increasing and afne in r. ETr (1 3 0, while Notice that in the limit as M 3 and keeping M xed, this potential reservations distribution has independent increments, with the number of reservation requests being distributed according to a Poisson random variable with parameter M . Another example is to suppose that the number of reservation requests is distributed according to a Poisson random variable, where the parameter is an increasing afne function of the number of reservations on hand. Such a distribution might arise if the reservation requests in different periods are positively correlated, examples of which were suggested in Section 3.

814

CHATWIN Therefore (2) yields VN V N 1. each VN 1 r vN b r qN VN br r


1

These four customer types can be modeled in the last four periods before ight time. These four periods are effectively of zero time length, each occurring almost immediately prior to the time of ight. In period 4 there are no reservation requests, and the cancellations correspond to no-shows. The (conditional) no-show distribution must be TP3, and in addition in the Nonstationary-Fares Model must have increasing linear mean. In period 3 the reservation requests correspond to no-records. Since these customers have valid reservations the airline must accept them. This represents an important distinction between this and any other period. In light of this the no-record distribution must be TP3 and in addition in the Nonstationary-Fares Model must have increasing afne mean. There are no cancellations in this period, or indeed in periods 1 and 2. In periods 1 and 2 reservation requests correspond to stand-bys and walk-ups, respectively. In each of these two cases the airline may decide to accept or reject the reservation request. The walk-up and stand-by distributions must satisfy the conditions on the potential reservations distribution of Theorem 3 or 4 according to the model in question. 6. MONOTONICITY OF OPTIMAL BOOKING LIMITS AND MAXIMUM REVENUES In this section we examine the monotonicity of the optimal booking limits in the time N to ight departure. Theorem 5 shows that in the Stationary-Fares Model, the optimal booking limits fall toward ight-time. (Here, and in the sequel, fall should be taken to mean decreasing but not strictly. Similarly, rise should be taken to mean increasing but not strictly.) Theorem 7 gives conditions under which a similar result holds for the Nonstationary-Fares Model. Theorem 5. Booking Limits Fall Toward Flight-Time: Stationary Fares. Suppose that the ight-time revenue function is quasi-concave, that each potential reservations distribution is stochastically increasing, and that each cancellation distribution is TP3. Then the greatest optimal booking limits fall toward ight-time. Proof. We show that the sequence { mN} of greatest optimal booking limits is nondecreasing in N. It is shown in mN. Thus VN is nondecreasing Remark 1 that MN r N on r m . Given this and that the cancellation distribution qN 1 is TP3 it follows from (2) and Karlins result that mN. Thus since vN 1 is vN 1 is nondecreasing on s s s N m N 1. quasi-concave, it follows that m Theorem 6. Monotonicity of Maximum Revenue: Stationary Fares. Suppose that the ight-time revenue function is quasi-concave, that each potential reservations distribution is stochastically increasing, and that each cancellation distribution is TP3. Then maxr VN is nonincreasing in N. r Proof. Let b mN and VN Theorem 3, and hence VN maxr VN. Then b MN by r N N Vb . Now (6) gives Vb v N. b

VN

, since

In the Nonstationary-Fares Model a result similar to that of Theorem 6 is not really meaningful. However, there is an analog of Theorem 5, although some conditions on the fares and refunds are necessary. Recall that under the assumptions of Theorem 4, each cancellation distribution s N s for qN has contracting linear mean, so that r 0 qNr sr N N 1. some constant , 0 Theorem 7. Booking Limits Fall Toward Flight-Time: Nonstationary Fares. Suppose that the ight-time revenue function is concave, that each potential reservations distribution either (1) has independent increments or (2) is TP3 with increasing afne mean, and that each cancellation distribution is TP3 with contracting linear mean. Assume N N 1 N N f (1 )c 0 for each N 1. Then that fN the greatest optimal booking limits fall towards ight-time. Proof. We show that the sequence { mN} of greatest optimal booking limits is nondecreasing in N. From (8) v sN sf N
s r 0 s r 0 N q sr V rN 1

rc N rf N
N N 1 1

s fN

cN

N q sr V rN

s fN

cN .

(21)

Then by assumption, the second term on the right-hand side of (23) is nondecreasing in s. Also, the rst term is nondecreasing on s MN by Karlins result since VN 1 r N 1 rf is nondecreasing on r MN 1 and qN is TP3. MN 1. But it was Thus vN sfN is nondecreasing on s s N 1 mN 1, whence shown in Remark 2 that M N 1 N m . m The condition fN rewritten as
N
N N

(1 cN

)cN

0 may be

fN

fN

fN

0,

so will certainly be satised if fN fN 1 and fN cN, i.e., if the fares fall toward ight-time, and the fare exceeds the cN refund in each period. Observe that this is true if fN f for each N, as is the case in the Stationary-Fares Model (c.f. Theorem 5). In Section 7 we will show that if there are no cancellations (so we can assume that cN fN) and the fares rise toward ight-time, then the optimal booking limits rise toward ight-time. The following example shows that if the fares are stationary but are exceeded by the refunds, then it is not necessary for the booking limits to fall toward ight-time. Example 1. Need for Fares to Exceed Refunds. Let R 3, f2 10, c1 10, c2 20, and V0 (0 0 0 20)T. f1 Suppose that there are no cancellations in the rst period, V0. Now (v1 f1s) (0 10 20 10)T, so that so that v1 s 1 {2}. Assume that there is precisely one reservation m request in the rst period. This will be accepted in states 0

CHATWIN and 1, and declined in states 2 and 3. Then V1 (10 10 0 20)T. Suppose that the cancellation distribution in the second period is given by the matrix 1 q2
3 4 1 2 1 4

815

0
1 4 1 2 3 4

0 0 0 0

0 0 0 0 .

Then v2 (10 5 20 35)T, whence (v2 f2s) s T 2 {0}. Thus the optimal booking (10 5 0 5) and m limit in the second period is smaller than that in the rst period. To give some intuition behind this condition, suppose that in period N 1 the optimal booking limit is 1, and there is at least one reservation request with probability 1. If the optimal booking limits are to fall towards ight time then the optimal booking limit in period N must be at least 1. This will be the case if in period N it is optimal to accept a reservation request when there are no reservations on hand. To see when this will be so, consider the expected revenues associated with accepting and declining such a request. If the request is accepted then the airlines expected revenue is fN fN
N N 1 v1

1
N

cN fN
1

fN

N v1

cN

N v 1 1,

N ) is the probability that the customer canwhere (1 cels in period N. If the request is declined then the airvN 1. For it to be lines expected revenue is fN 1 1 worthwhile to accept the request we must have

period. Then Theorem 4 demonstrates the optimality of booking-limit policies for the discount-allocation problem. Note that we do not need to make any assumption on the order of the fare classes, i.e., on the relationships between the fi. The Nonstationary-Fares Model actually extends earlier work in this area by allowing the number of reservation requests to depend on the number of current bookings: all the previous models assume that the potential reservations distribution has independent increments. The models of Lee and Hersh and of Robinson, like the Nonstationary-Fares Model, make no assumption on the order in which the fare classes arrive. The models of Curry et al. assume that lower fare classes book prior to higher fare classes. This appears to be a realistic assumption for two reasons. First, it is mandated by the airlines pricing policies. These place early booking restrictions on the discount fares in order to protect the higher fare classes, and to allow for price discrimination among target customers; and second, it is a consequence of the character of the customers booking into each class: business travelers in full fare and leisure travelers in the discount classes. All of the authors demonstrate the optimality of booking-limit policies. With the additional assumption that fares are monotonically increasing over time, Curry et al. show that the optimal booking limits rise toward ighttime. This result is easily proved for the NonstationaryFares Model with no cancellations, and is presented here. The result should be contrasted with Theorem 8, in which conditions are given ensuring that the optimal booking limits fall toward ight-time for the general case in which cancellations are permitted. Theorem 8. Booking Limits Rise Towards Flight-Time with Rising Fares and No Cancellations. Suppose that the ight-time revenue function is concave, and that in each period, the potential reservations distribution either (1) has independent increments, or (2) is TP3 with increasing afne mean. Assume that there are no cancellations in any period, and that fares rise toward ight-time. Then the optimal booking limits rise toward ight-time. Proof. We show that the least and greatest optimal booking limits in period N ( mN and mN) are nonincreasing in N. Since there are no cancellations, vN 1 VN for all s, s s N. Then v sN
N vb 1

fN

cN

fN

N v1

fN

N v 1 1,

which reduces to the given condition. 7. APPLICATION OF THE NONSTATIONARY FARES MODEL TO THE DISCOUNTALLOCATION PROBLEM The Nonstationary-Fares Model can be interpreted as a multifare-class discount-allocation model as in McGill (1989), Curry (1990), Wollmer (1992), Brumelle and McGill (1993), Lee and Hersh (1993), and Robinson (1995). In this case, the airline sells tickets for the same type of seats at different prices (to different fare classes), and seeks to determine how many reservations to sell to customers in discount fare classes, and how many seats to protect in order to satisfy demand that may subsequently arise from customers in higher fare classes. To interpret the Nonstationary-Fares Model as a discount-allocation model we make two assumptions. First, we do not allow cancellations, thereby obviating the need to consider overbooking, cancellation penalties, or denied boarding costs. Second, we assume that the fare paid by a customer booking into the ith class is fi, and that all class i customers make reservation requests during the ith time

sf N

V sN M ,
N

sf N

s fN

fN .

(22)

Setting b
1 1 N vb N Vb 1 1

fN
N Vb

(23)
N

N 1

0,

since the rst term on the right-hand side is strictly negative by the denition of MN, and the second is nonposiMN. But as shown in Remark 2, tive. Thus mN 1 N N m , and hence mN 1 m N. M N N N sf is nonincreasing on By the denition of M , Vs fN, it follows from (21) that s MN and, since fN 1

816

CHATWIN demand. Let X be the low-fare demand and Y be the high-fare demand, i.e., X is the demand in period 2 and Y is the demand in period 1. There are no cancellations and f 1. f2 Example 2. High-Fare Demand is Stochastically Decreasing in the Low-Fare Demand. Suppose that the plane has one seat, i.e., C 1. Suppose that X is either 1 or 2. Suppose that if X 1, then Y 1, and that if X 2, then Y 0. Thus there are precisely two reservation requests in total, and it is plainly optimal to decline the rst, but to accept the second. So if X 1 then the airline should reject the request, but if X 2, then the airline should accept one request and decline one request. The reason that a booking-limit policy is not optimal here is that the more low-fare demand there is, the more requests the airline should accept, because they know that there will be less high-fare demand to ll any empty seats. Example 3. High-Fare Demand is Stochastically Increasing in the Low-Fare Demand. Let C 4. Suppose that X 2 and Y are each either 1 or 2. Now let P{Y 1 X 1} 3 1 2 P{Y 1X 2}, P{X 1} , and P{X 2} 3 3 1 . Let f1 2 and f2 1. Let V1( x) be the maximum r 3 expected revenue at the beginning of period 1 when there are r reservations and the demand in period 2 was x. Then V1 1 and V1 2
1 1 33 31 33 2 0 3

vN 1 sfN 1 is also nonincreasing on s MN. Hence s N 1 N N M m , where the second inequality m comes from Remark 2. It may seem unnecessary to allow fN fN 1. However if the potential reservations distribution p for a fare class can be expressed as the convolution of two potential reservations distributions p1 and p2 (i.e., the reservation request stream into a fare class can be represented as the sum of two independent reservation request streams) where p1 has independent increments and p2 is TP3 with contracting afne mean, then separating this class into two in the obvious manner allows the theory to be applied, and the same fare will be associated with each of the two new fare classes. The discount-allocation problem is, in fact, a special case of a class of inventory problems known as stock rationing problems. Models of this type have been developed by Evans (1968), Topkis (1968), and Kaplan (1969). As here, the motivation for rationing stock, i.e., declining to satisfy customer demand, is to conserve stock for possible use later to satisfy demand from a more important customer class. Topkis model can be simplied and reinterpreted so as to coincide precisely with the discountallocation problem. Topkis shows that the optimal rationing policy is determined by a set of critical rationing levels such that at a given time one satises demand of a given class only if no demand of a more important class remains unsatised and as long as the stock level does not fall below the critical rationing level for that class at that time. This is equivalent to showing that a booking-limit policy is optimal in the discount-allocation problem. Topkis also showed that the optimal booking limits rise toward ight-time when lower fare classes arrive rst.

23 23 23 2 0

8. NONOPTIMALITY OF BOOKING-LIMIT POLICIES WITH STOCHASTICALLY DEPENDENT RESERVATION REQUESTS As mentioned in Section 3.2, if dependence between the number of reservation requests in different periods is modeled directly, then booking-limit policies need not be optimal. McGill (1989) and Brumelle et al. (1990) have considered this problem in the context of the discountallocation problem. They considered allocating airline seats between two fare classes; when the low fare class books rst, there are no cancellations, and there is stochastic dependency between demands for low- and high-fare seats. They assume that a booking-limit policy is optimal, and assuming a weak condition on the dependency of the low- and high-fare demands, give a probabilistic formula for the optimal booking limit for the low-fare customers. However, they do not discuss whether or not booking-limit policies are optimal over the class of all allowable policies. It is shown here that they need not be. Three examples are given that demonstrate this fact under different assumptions on the dependence between the low- and high-fare

Restrict attention to the class of policies that are allowable when reservation requests occur sequentially, and a booking decision must be made as each request occurs. If a booking-limit policy is not optimal within this class, then no such policy can be optimal within the larger class of policies that are allowed when the airline is able to observe all requests within a period before making any booking decisions. Since X is either 1 or 2, split period 2 into two subperiods. Call these a and b. In subperiod a there is precisely one reservation request; in subperiod b there is one reservation request with probability 1/3, otherwise there are no reservation requests. Let V2a (respectively, r V2b) be the maximum expected revenue at the beginning r of subperiod a (respectively, b) when there are r reserva(2/3)V1(1) (1/3){V1(2) (1 tions. Then V2b r r r 1 Vr 1(2))}. Thus it is optimal to accept a reservation request if r 1 and to decline a request otherwise, which gives V 2b
2 2 8 39 39 29 2 0

2b Then (1 Vr 1). Thus it is optimal to accept a reservation request if r 2 and to decline a request otherwise. Hence, the optimal policy is not a booking-limit policy. For in both the cases when there are one or two reservations at the beginning of period 2, it is optimal to

V2a r

V2b r

CHATWIN accept the rst, but to decline the second, reservation request in period 2. The reason that a booking-limit policy is not optimal here is that the more low-fare demand there is, the fewer requests the airline should accept, because there will be more high-fare demand to ll any empty seats. Example 4. High-Fare Demand is Upgrades from LowFare Demand. Suppose that the high-fare demand comprises solely customers whose request for a low-fare seat was declined, and who decided to upgrade to the high-fare. Let C 2, suppose that X is either 1 or 2, and that a customer whose reservation request in period 2 was declined, makes a reservation request in period 1 with prob2 and f2 1. Let V1( x) be the ability 1/3. Let f1 r maximum expected revenue at the beginning of period 1 when there are r reservations and the number of declined reservation requests in period 2 was x. Then V1 0 and V1 2 1 1 3 1 1 9 0
T

817

discount-fare seats were made available after their full-fare purchase). The optimal policy of Example 4 applies even given the more restrictive assumptions of this model, which shows that even in this case a booking-limit policy is not optimal.

9. CONCLUSION In this paper we demonstrate how to successfully model the airline overbooking process so as to incorporate its inherent dynamic naturein particular, modeling customer cancellations, and no-shows in a dynamic framework. We introduce important new tools for analyzing optimal policies within this dynamic framework. This paper is the rst to take advantage of the properties that TP3 (totally positive of order 3) density functions preserve quasi-concavity and concavity in order to prove results regarding the structure of optimal policies. We give conditions that ensure the intuitive result that a booking-limit policy is optimal. We show how our results can be applied to the airline discount-allocation problem. Our model formulation is closely related to several previously published models, and we obtain similar results. Our model extends previous work in this area by allowing the number of reservation requests to be dependent on the number of current bookings. Many previous authors have restricted their attention to booking-limit policies when developing models of the airline reservations process. This has been particularly true when considering multiple-fare class seat allocation in which there is dependence between the demand of the different classes, such as when there is diversion. It is important to recognize that booking-limit policies are not always optimal. In this paper we give several examples that demonstrate the nonoptimality of booking-limit policies under different assumptions regarding the stochastic dependence between demand classes. There are two specic directions in which future research will be benecial. The rst involves generalizing the formulation of the models and weakening the assumptions. For example, as Robinson (1995) points out, in the discount-allocation problem we assume that the fare classes arrive sequentially, and not concurrently as is the case in practice. Modeling concurrent arrivals of fare classes involves generalizing the model to allow for arrivals from more than one fare class in the same period. This would require much more complicated forms of booking control. The second area of interest involves developing a model to handle both overbooking and discount allocation. Such a model would have to allow for bookings in multiple fare classes as well as cancellations and no-shows. Janakiram et al. (1994) have taken the rst steps in this direction, although their model requires the cancellation rates in the different fare classes to be identical.

000

V1 1

2 2 3 3

As in Example 2, restrict attention to the class of policies that are allowable when reservation requests occur sequentially, and a booking decision must be made as each request occurs. Since X is either 1 or 2, split period 2 into two subperiods. Call these a and b. In subperiod a there is precisely one reservation request; in subperiod b there is one reservation request with probability 4/5, otherwise there are no reservation requests. Let V2b( x) be the maxir mum expected revenue at the beginning of subperiod b when there are r reservations and the number of declined reservation requests in subperiod a was x. Then V r2b x
1 5

V r1 x

4 5

V r1 x

V r1

Now x is either 0 or 1. If x 0, then it is optimal to accept a reservation request only if r 1; if x 1, then it is optimal to accept a reservation request only if r 0. Thus V 2b 0
4 4 5 5

and V 2b 1

7 1 1 15 1 45 0

Let V2a be the maximum expected revenue at the beginr ning of subperiod a when there are r reservations. Then 2b V2a V2b(1) (1 Vr 1(0)). Thus it is optimal to accept r r a reservation request only if r 0. Hence, the optimal policy is not a booking-limit policy. For if there are no reservations at the beginning of period 2, it is optimal to accept both reservation requests in period 2; otherwise, it is optimal to decline such requests. A more realistic model for upgrades would be to suppose that any customer declined the discount fare immediately upgrades to the full fare with probability p. This model is examined by Pfeifer (1989) under the assumption that once one customer has been declined the discount fare, all subsequent customers are declined the discount fare (in Pfeifers words, in order to avoid the undesirable consequences that result if full-fare purchasers learn that

818

CHATWIN
CURRY, R. E. 1990. Optimum Seat Allocation with Fare Classes Nested by Origins and Destinations. Transp. Sci. 24, 193204. CURRY, R. E. 1992. Parallel Nesting and Overallocation. Abstract, ORSA/TIMS Joint National Meeting Bulletin, San Francisco, CA. EVANS, R. V. 1968. Sales and Restocking Policies in a Single Item Inventory System. Mgmt. Sci. 14, 463 472. HERSH, M. AND S. P. LADANY. 1978. Optimal Seat Allocation for Flights with Intermediate Stops. Comput. O. R. 5, 3137. JANAKIRAM, S., S. STIDHAM, JR., AND A. SHAYKEVICH. 1994. Airline Yield Management with Overbooking, Cancellations and No-Shows. Technical Report No. UNC/OR/TR 94-9, University of North Carolina at Chapel Hill, NC. KAPLAN, A. 1969. Stock Rationing. Mgmt. Sci. 15, 260 267. KARLIN, S. 1958a. One Stage Models with Uncertainty. Chapter 8 in Studies in the Mathematical Theory of Inventory and Production. K. Arrow, S. Karlin, and H. Scarf (eds.), Stanford University Press, Stanford, CA. KARLIN, S. 1958b. Optimal Inventory Policy for the ArrowHarris-Marschak Dynamic Model. Chapter 9 in Studies in the Mathematical Theory of Inventory and Production. K. Arrow, S. Karlin, and H. Scarf (eds.), Stanford University Press, Stanford, CA. KARLIN, S. 1960. Dynamic Inventory Policy with Varying Stochastic Demands. Mgmt. Sci. 6, 231258. KARLIN, S. 1968. Total Positivity. Stanford University Press, Stanford, CA. KARUSH, W. 1959. A Theorem in Convex Programming. Naval Res. Logist. 6, 245260. KARUSH, W. AND A. VAZSONYI. 1957. Mathematical Programming and Employment Scheduling. Naval Res. Logist. 4, 297320. LADANY, S. P. 1976. Dynamic Operating Rules for Motel Reservations. Dec. Sci. 7, 829 840. LADANY, S. P. 1977. Bayesian Dynamic Operating Rules for Optimal Hotel Reservations. Z. Opns. Res. 21, B165B176. LEE, T. C. AND M. HERSH. 1993. A Model for Dynamic Airline Seat Inventory Control with Multiple Seat Bookings. Transp. Sci. 27, 252265. LEHMANN, E. L. 1959. Testing Statistical Hypotheses. John Wiley & Sons, NY. LIBERMAN, V. AND U. YECHIALI. 1978. On the Hotel Overbooking ProblemAn Inventory System with Stochastic Cancellations. Mgmt. Sci. 24, 11171126. LIPPMAN, S. A. AND S. STIDHAM, JR. 1977. Individual Versus Social Optimization in Exponential Congestion Systems. Opns. Res. 25, 233247. LITTLEWOOD, K. 1972. Forecasting and Control of Passenger Bookings. In Proceedings 12th AGIFORS Symposium, 95 117, American Airlines, NY. LORENTZ, G. G. 1953. An Inequality for Rearrangements. Amer. Math. Monthly 60, 176 179. MARTINEZ, R. AND M. SANCHEZ. 1970. Automatic Booking Level Control. In Proceedings 10th AGIFORS Symposium, 120, American Airlines, NY. MCGILL, J. I. 1989. Optimization and Estimation Problems in Airline Yield Management. Ph.D. Dissertation, Commerce and Business Administration Program, University of British Columbia, Canada.

ACKNOWLEDGMENT The results of this paper are part of the authors Ph.D. thesis, completed in the Department of Operations Research at Stanford University under the direction of Professor Arthur F. Veinott, Jr. The author would like to express his most sincere thanks to Professor Veinott for his encouragement and support. The insight and perspective he provided during his close examination of the work led to improvements in, and the elimination of errors from, many of the results and proofs.

REFERENCES
ALSTRUP, J., S. BOAS, O. B. G. MADSEN, AND R. V. V. VIDAL. 1985. Booking Policy for Flights with Two Types of Passengers. Research Report No. 1A/1985, IMSOR, Technical University of Denmark, Lyngby, Denmark. ALSTRUP, J., S.-E. ANDERSSON, S. BOAS, O. B. G. MADSEN, AND R. V. V. VIDAL. 1986a. Experiences with an Airline Overbooking Model for Flights with Two Passenger Types. Research Report No. 2/1986, IMSOR, Technical University of Denmark, Lyngby, Denmark. ALSTRUP, J., S. BOAS, O. B. G. MADSEN, AND R. V. V. VIDAL. 1986b. Booking Policy for Flights with Two Types of Passengers. European J. Opns. Res. 27, 274 288. ALSTRUP, J., S.-E. ANDERSSON, S. BOAS, O. B. G. MADSEN, AND R. V. V. VIDAL. 1989. Booking Control Increases Prot at Scandinavian Airlines. Interfaces 19, 10 19. BECKMANN, M. J. 1958. Decision and Team Problems in Airline Reservations. Econometrica 26, 134 145. BELOBABA, P. P. 1987a. Air Travel Demand and Airline Seat Inventory Management. Report R87-7, Flight Transportation Laboratory, MIT, Cambridge, MA. BELOBABA, P. P. 1987b. Airline Yield Management: An Overview of Seat Inventory Control. Transp. Sci. 21, 6373. BELOBABA, P. P. 1992. Optimal vs. Heuristic Methods for Nested Seat Allocation. Abstract, ORSA/TIMS Joint National Meeting Bulletin, November, San Francisco, CA. BHATIA, A. V. AND S. C. PAREKH. 1973. Optimal Allocation of Seats by Fare. Presentation to AGIFORS Reservations Study Group. BITRAN, G. R. AND S. M. GILBERT. 1996. Managing Hotel Reservations with Uncertain Arrivals. Opns. Res. 44, 35 49. BODILY, S. E. AND P. E. PFEIFER. 1992. Overbooking Decision Rules. OMEGA 20, 129 133. BRUMELLE, S. L., J. I. MCGILL, T. H. OUM, K. SAWAKI, AND M. W. TRETHEWAY. 1990. Allocation of Airline Seats between Stochastically Dependent Demands. Transp. Sci. 24, 183192. BRUMELLE, S. L. AND J. I. MCGILL. 1993. Airline Seat Allocation with Multiple Nested Fare Classes. Opns. Res. 41, 127137. CHATWIN, R. E. 1993. Optimal Airline Overbooking. Ph.D. Dissertation, Department of Operations Research, Stanford University, CA. COLVILLE, G. AND R. L. PHILLIPS. 1992. Dynamic O&D Yield Management. Abstract, ORSA/TIMS Joint National Meeting Bulletin, San Francisco, CA.

CHATWIN
NOTZON, E. M. III. 1970. Minimax Inventory and Queueing Models. Ph.D. Dissertation, Department of Operations Research, Stanford University, CA. PHILLIPS, R. L. 1994. A Marginal Value Approach to Airline Origin and Destination Revenue Management. Proceedings of the 16th IFIP Conference on System Modeling and Optimization, J. Henry and J. P. Yvon (eds.), Springer Verlag, NY. PFEIFER, P. E. 1989. The Airline Discount Fare Allocation Problem. Dec. Sci. 20, 149 157. PORTEUS, E. L. 1971. On the Optimality of Generalized (s, S) Policies. Mgmt. Sci. 17, 411 426. RICHTER, H. 1982. The Differential Revenue Method to Determine Optimal Seat Allotments by Fare Type. In Proceedings 22nd AGIFORS Symposium, 339 362. ROBINSON, L. W. 1995. Optimal and Approximate Control Policies for Airline Booking With Sequential Nonmonotonic Fare Classes. Opns. Res. 43, 252263. ROTHSTEIN, M. 1968. Stochastic Models for Airline Booking Policies. Ph.D. Dissertation, Graduate School of Engineering and Science, New York University, NY. ROTHSTEIN, M. 1971. An Airline Overbooking Model. Transp. Sci. 5, 180 192. ROTHSTEIN, M. 1974. Hotel Overbooking as a Markovian Sequential Decision Process. Dec. Sci. 5, 389 404. ROTHSTEIN, M. 1985. O.R. and the Airline Overbooking Problem. Opns. Res. 33, 237248. SMITH, B. C. 1993a. Market Based Yield Management: Its Protable and Practical. Proceedings of the 1993 AGIFORS General Symposium. SMITH, B. C. 1993b. Continuous Nesting. IATAs 5th Revenue Management Conference and Technical Brieng (October). Montreal, Canada, 425 465.

819

SMITH, B. C., J. F. LEIMKUHLER, AND R. M. DARROW. 1992. Yield Management at American Airlines. Interfaces 22, 8 31. SWAN, B. 1994. Bid Price Methodology. IATAs 6th Revenue Management Conference and Technical Brieng (October), Barcelona, Spain. TALLURI, K. AND G. J. VAN RYZIN. 1995. An Analysis of Bid Price Control for Network Yield Management. Abstract, INFORMS National Meeting Bulletin, New Orleans, LA. THOMPSON, H. R. 1961. Statistical Problems in Airline Reservation Control. O. R. Quart. 12, 167185. TOPKIS, D. M. 1968. Optimal Ordering and Rationing Policies in a Nonstationary Dynamic Inventory Model with n Demand Classes. Mgmt. Sci. 15, 160 176. TOPKIS, D. M. 1978. Minimizing a Submodular Function on a Lattice. Opns. Res. 26, 305321. VEINOTT, A. F., JR., 1965. Optimal Policy in a Dynamic, Single Product, Nonstationary Inventory Model with Several Demand Classes. Opns. Res. 13, 761778. WEATHERFORD, L. R. 1992. Comparing Decision Rules in Perishable Asset Revenue Management Situations. Abstract, ORSA/TIMS Joint National Meeting Bulletin, San Francisco, CA. WEATHERFORD, L. R., S. E. BODILY, AND P. E. PFEIFER. 1993. Modeling the Customer Arrival Process and Comparing Decision Rules in Perishable Asset Revenue Management Situations. Transp. Sci. 27, 239 251. WOLLMER, R. D. 1992. An Airline Seat Management Model for a Single Leg Route when Lower Fare Classes Book First. Opns. Res. 40, 26 37.

You might also like