Improved Forecast Accuracy in Airline Revenue Management by Unconstraining Demand Estimates from Censored Data

Abstract

Accurate forecasts are crucial to a revenue management system. Poor estimates of demand lead to inadequate inventory controls and sub-optimal revenue performance. Forecasting for airline revenue management systems is inherently difficult. Competitive actions, seasonal factors, the economic environment, and constant fare changes are a few of the hurdles that must be overcome. In addition, the fact that most of the historical demand data is censored further complicates the problem. This dissertation examines the challenge of forecasting for an airline revenue management system in the presence of censored demand data.

The number of seats an airline can sell on a flight is determined by the booking limits set by the revenue management system. An airline continues to accept reservations in a fare class until the booking limit is reached. At that point, the airline stops selling seats in that fare class-It also stops collecting valuable data. Demand for travel in that fare class may exceed the booking limit, but the data does not reflect this. So the data is censored or "constrained" at the booking limit.

While some models exist that produce unbiased forecasts from censored data, it is preferable to "unconstrain" the censored observations so that they represent true demand. Then, the forecasting model may be chosen based on the structure of the problem rather than the nature of the data. This dissertation analyzed the improvement in forecast accuracy that results from estimating demand by unconstraining the censored data.

Little research has been done on unconstraining censored data for revenue management systems. Airlines tend to either ignore the problem or use very simple ad hoc methods to deal with it. A literature review explores the current methods for unconstraining censored data. Also, practices borrowed from areas outside of revenue management are adapted to this application. For example, the Expectation-Maximization (EM) and other imputation methods were investigated. These methods are evaluated and tested using simulation and actual airline data. An extension to the EM algorithm that results in a 41% improvement in forecast accuracy is presented.