Dynamic ticket pricing of airlines using variant batch size interpretable multi-variable long short-term memory

Abstract

Research of airlines shows that seat inventory control and therefore, revenue management is based not on a systematic analysis but more on human judgement. Machine learning models have been developed and applied to support decisions for ticket pricing dynamically. However, conventional models and approaches yield low statistical evaluation scores. In this study, the features used in other studies were explored and the cost available seat kilometer (CASK) value and target revenue features were included for the first time to the best of our knowledge which are essential components of ticket price decision. Real data from a low-cost carrier airline in Turkey were collected and the observation data were splitted into two to study with the highest profit sale data. Then the outliers were filtered to let the models learn from and generate better price offerings businesswise. Observation datasets obtained in each step were recorded to be tested. 7 different model techniques were simulated and tested with 4 different datasets according to 6 different statistical evaluation criteria. A new approach to Interpretable Multi-Variable Long Short-Term Memory (IMV-LSTM) model was proposed by taking every flight and its sales as an independent series, that is to assign a dynamic batch size. Extensive experiments on real datasets reveal enhanced statistical evaluation scores by using the proposed approach and model. The proposed model can be used by the airlines to mitigate human judgement on ticket pricing, to manage their price offerings to reach their target revenues and to increase their profits. The model can be used by other business cases that have similar historical data as overlapping windows structure.