An artificial neural network approach to origin-destination demand forecasting in airline revenue management

After the deregulation of the airline industry in the United States in 1978, increased competition caused by new low cost entrants put well established legacy carries into a difficult position. One of these legacy carries, namely American Airlines, realized that it is not possible to compete with the low cost carriers by simply reducing the prices due to its very high cost structure compared to its low cost new entrant rival. As a response to this severe low cost competition, American Airlines formulated a counter attack by publishing promotional fares with restrictions and limiting the number of promotional fares that can be sold in each flight. This was the origin of Revenue Management practice and American's success encouraged other airlines to use revenue management as well, and Revenue Management has become common practice among all airlines. In order to maximize revenue by selling the right seat, to the right customer, at the right price at the right time, Revenue management is concerned with establishing different fare classes in each and every Origin and Destination (OD) market with imposed restrictions on the lower fares in order to prevent passengers with high maximum willingness pay from buying the lower fares, and limiting the number of seats that can be sold by the lower fare classes in order not to reject the late coming higher fare class demand due to capacity limits. Airlines use revenue management systems in order to determine the maximum number of seats that can be sold in each flight/fare class or OD/fare class combination. Leg based revenue management systems are utilized for maximizing each flights revenue separately whereas, advanced airlines with high transit passenger ratio are using OD revenue management systems which strive for maximizing the revenue of the whole network totally. Zaki reported airlines could obtain up to 10% incremental revenues by applying revenue management. Yet this percentage of incremental revenue can only be achieved by having accurate demand forecasts in the detailed level required by the utilized revenue management system. In order to maximize the whole network`s revenue totally, OD revenue management systems require demand forecast in the finest detail possible. Demand observations in this fine detailed level are so small that they are usually much less than one, in the order of one tenth or one hundredth. Hence, the dataset is highly nonlinear and extremely noisy. Apart form that, OD demand data, unlike time series are not measured within equally spaced time intervals. This fact dramatically undermines the applicability of time series methods in OD demand forecasting. In this study, an Artificial Neural Network (ANN) model is designed specifically for the OD demand forecasting in the detailed level required by OD revenue management systems, and its performance is tested by utilizing a major European carrier`s data and compared to the currently utilized revenue management system`s forecasting performance. Obtained results suggest that the proposed ANN approach`s OD demand forecasting performance is satisfactory and its forecasts are more accurate compared to the currently utilized revenue management system. These finding reveal that the proposed ANN approach can be successfully utilized for the OD demand forecasting required by OD revenue management systems. Even though they are not used as inputs to the revenue management systems, load factor (LF) forecasts are of extreme importance since load factor has been considered as one of the most commonly used and important performance indicators for passenger airlines. Significant under or overestimated forecasts of a route`s load factor can cause pricing, marketing and sales, scheduling and advertisement teams to take inappropriate actions. In this study, the historical load factors of a major flight of the considered airline has been utilized to in order to develop a robust model for forecasting load factors. Unlike B-V data, load factors are obtained in equally spaced time intervals. Hence, ARIMA method has been used to forecast load factors since it is considered as one of the most successful statistical times series analysis technique. Yet, load factor data set also represent high fluctuations and nonlinearity not only caused by seasonal variability. Therefore, in order to forecast load factor another ANN model has been designed and its forecasting performance has been compared with ARIMA and the airline`s current system. Obtained results indicate the proposed ANN approach for load factor forecasting produce more accurate forecasts compared to ARIMA and the system in place utilized by the considered airline.