Method for stability verification of air ticket demand forecasting models

Airline demand forecasting is a key step in revenue management and pricing systems. Current demand forecasting methods are predominantly based on machine learning and deep learning techniques, as they are able to take into account a wide range of patterns identified from historical data. Despite the good generalisability of such models, the approaches still require additional validation of performance stability. Applying the models to new data, such as a new destination or flight schedule, can severely distort demand forecasting. This paper presents a method to validate the stability of airline ticket demand forecasting models based on the sequential exclusion of groups of feature values from the training sample with subsequent evaluation of the model quality on a test period using the WMAPE metric. Experiments with a linear regression baseline and a gradient boosting model over decision trees trained on historical data from S7 Airlines demonstrate the effectiveness of the proposed approach for identifying sensitive attributes and preventing degradation of forecast quality, and show that gradient boosting achieves substantially lower forecast errors than linear regression while being more sensitive to exclusions of route-related features.