A reinforcement learning approach for hotel revenue management with evidence from field experiments

Abstract We consider a budget hotel chain's revenue management problem of deciding how to dynamically allocate capacity to multiple segments of customers. Our work solves an industrial‐sized problem faced by practitioners, with the reality of implementation motivating us to develop a tailored reinforcement learning approach. Our approach proceeds in two steps. First, a recommended average discount is computed with a reinforcement learning algorithm. Then, the recommended average discount is turned into a capacity allocation through a linear program. This approach overcomes the challenges of characterizing demand and estimating cancellations, and it facilitates hotel managers' acceptance of the revenue management system. We implement this approach in the hotel chain in a pilot study and assess its effectiveness using synthetic control methods. Our approach improves the key operational performance measure—revenue per available room—by 11.80%. There is heterogeneity in how the pilot hotels improve their revenue per available room. Some mainly increase their occupancy rate, some mainly increase the average daily room rate, while others experience significant increases in both. Further analysis shows that our approach uncovers the individual sources of suboptimal performance in pilot hotels and correspondingly improves decision‐making. Our work demonstrates that a reinforcement learning approach for hotel revenue management is promising.