A non-parametric approach to demand forecasting in revenue management

• A new mixed integer nonlinear programming model for estimating data from censored observations in revenue management systems.
• Simultaneous capture of customer behavior and seasonal variations.
• Design and implementation of a customized semi-global optimization algorithm, based on partial enumeration and cuts.
• Numerical tests.

In revenue management, the profitability of the inventory and pricing decisions rests on the accuracy of demand forecasts. However, whenever a product is no longer available, true demand may differ from registered bookings, thus inducing a negative bias in the estimation figures, as well as an artificial increase in demand for substitute products. In order to address these issues, we propose an original Mixed Integer Nonlinear Program (MINLP) to estimate product utilities as well as capturing seasonal effects. This behavioral model solely rests on daily registered bookings and product availabilities. Its outputs are the product utilities and daily potential demands, together with the expected demand of each product within any given time interval. Those are obtained via a tailored algorithm that outperforms two well-known generic software for global optimization.