Learning customer preferences and dynamic pricing for perishable products

This research proposes a revenue management framework for perishable products when customer preferences are unknown before the selling season begins. In this research, customer preferences are measured by the distribution of customer willingness to pay (WTP). When the WTP distribution is initially unknown, a long short-term memory (LSTM) neural network is adopted to quickly learn the distribution by using limited selling data in early periods of the selling season. The average LSTM estimation error is less than 5% in the fifth period of the selling horizon and approximately 1% in the 25th period when the WTP follows a Gaussian distribution with an unknown mean. The estimation of WTP distribution is then used by a dynamic pricing model to generate optimal price decisions. To reduce the calculation burden of the pricing model, we present the existence of a lower bound on the optimal price, provided that the coefficient of variation of WTP distribution is bounded by 80%. In our numerical study, the proposed pricing framework is benchmarked against the optimal pricing strategy under a known customer WTP distribution, and the revenue difference caused by the unknown WTP distributions is less than 2% in most cases. This small revenue difference represents the costs to learn the unknown customer preferences. For perishable products without inventory replenishment, such as airline tickets or hotel rooms, the small preference learning costs make the proposed framework especially valuable.