Unconstraining Methods In Airline Revenue Management With Incompleteness Of Demand Information

Along with the rapid economic development and further deepen reform, the service sector in China has played an increasingly important role in the national economy. As an effective tool, revenue management (RM) has helped many airlines to improve their revenues and profits through various techniques such as price differentiation, inventory control and overbooking. However, it is widely agreed that the success of these techniques largely depends on the accuracy and reliability of demand forecasting. The performance of an airline revenue management system (RMS) will be compromised without it in the process of conducting demand parameter estimation and optimization sequentially. The unconstrained historical demand data is the basis for demand forecasting. Due to the variety of RM inventory control technologies, the observed booking data in the RM systems often fail to fully reflect the true historical demand of passengers. In order to improve revenues for airlines, it is necessary to infer unconstrained demand data from censored demand data often caused by booking limits. This is a unique challenge of the demand forecasting in airline RM.Although demand unconstraining estimation is an "ancient" research topic, with the advent of the big data era and the increasingly fierce market competition both deepen the problem of incompleteness in airline passenger demand information, including the incompleteness of their demand distribution and choice behavior information. For the airlines, ignoring the incompleteness problems of passengers’demand information in the process of revenue management unconstraining estimation can lead to serious consequences. The true historical passenger demand would be misjudged, which caused by the increasing demand forecasting error. Besides that, the mistakes on optimization decisions would ultimately result in the loss of companies’revenue. Therefore, in the case of demand information incompleteness, which kind of unconstraining methods should the airlines use to correctly and effectively estimate the unconstrained demand of historical passenger? This is the central task of this research.This dissertation will first review unconstraining methods in the existing literatures. As most of these methods assume normality of the demand distribution, they cannot be applied to the situation where demand follows multi-distributions. This dissertation assumes various non-normal demand distributions and derives formulae of single-class EM algorithms and PD methods when nominal demand distribution information is unknown. Unconstraining estimation is a crucial step to the successful implementation of "unconstrained demand forecasting" in airline RMS. The difficulty to achieve its accuracy and effectiveness is to develop a simulation model that meets the practice of airline RM. The existing research only focuses on the comparative studies of single-class unconstraining methods based on normal distribution. For this purpose, a comparison model of unconstraining estimation for multi-distribution of demand in RMS is established using control theory and computer simulation. The single-class EM algorithms and PD methods based on Normal, Lognormal and Gamma distributions are compared under the joint-distribution of demand through simulation experiments with actual airline passenger demand data. Its accuracy, validity, and the impact on airline’s revenues is compared and evaluated. Simulation results show the effectiveness of the proposed model, and indicate that:1) the main factors that influence the performance of the unconstraining methods based on various distributions include skewness of the historical passenger demand data, the magnitude of variation and the level of the data being censored. The level of demand is not a crucial factor when choosing the distribution assumption form for single-class EM algorithm and PD method. Meanwhile, the parameter t plays a critical role for the single-class PD method.2) Improper choice of the distribution for unconstraining methods leads to significant loses of profit. Therefore, it is necessary to know the joint distribution of passenger demand in the process of demand unconstraining.In view of the incompleteness of both demand distribution and passengers’choice behavior information in historical sales data, this research developes multi-class spill models in which demand data is assumed to follow joint-distributions. Feasibility of proposed multi-class models is also discussed. Based on the practice of multi-class unconstraining estimation in airline RM, a numerical example is presented to compare the traditional single-class spill models with the proposed multi-class spill models in this dissertation. The results indicate that volatility of demand is the main factor that influences the accuracy of unconstraining process of the multi-class spill models. The proposed multi-class spill models are more effective in terms of preventing overestimating passenger demand among different fare classes in s same flight.In addition, traditional multi-flight unconstrianing methods in RM assume that passengers make one-time purchase decisions at their time of arrival. As a result, they cannot reflect the inter-temporal substitution behavior of strategic passengers. With the stochastic process theory and a nonparametric discrete choice model, this research consides strategic passenger behavior and develops a model based on the passenger rank-based preferences. For the incompleteness of passenger strategic choice behavior information in historical sales data, the EM algorithm is applied to jointly estimate the arrival rate of passengers and the probability mass function. The multi-flight unconstraining calculating method for primary demand considering historical strategic passenger behavior is also proposed. Another example is given to consider price changes in fare classes. Based on the practice of multi-flight unconstraining estimation in airline RM, it makes a comparative analysis between the proposed multi-flight EM algorithms and those in the existing literatures. The results show the proposed multi-flight method can capture the impact of firm’s dynamic pricing on strategic passenger behavior, and is more effective in preventing overestimating of future passenger demand among different fare classes.In the conclusion part, some suggestions regarding the relevant applications of this research are discussed, and their application prospect are predicted and evaluated.