Revenue management with minimal demand information

Research by Itir Karaesmen and Michael Ball

Hotels, cruise lines and car rental companies all use similarly complicated models for revenue management. These models require information about how many people are going to want those airline seats in the future—they need to forecast demand for their product into the future. Unfortunately, forecasting is notoriously difficult and demand estimates are inaccurate. The state of the economy, seasonal variations in travel demand and even the competitive actions of other airlines make finding reliable information even more difficult.Airlines want to sell their seats to the right customers at the right time for the right price. The more fare classes or products they have, the more difficult it is to optimize the revenue for each product. That is, airlines have turned to complex mathematical models for revenue management.

The inherent challenges in creating accurate demand forecasts has been a key factor limiting the realistic application of revenue management within new industries, and can cause problems for new products within these industries, such as new airline routes or new properties in a hotel chain.

Side-stepping the problem of demand forecasting can provide a robust solution to these difficulties, according to research by Michael Ball, Orkand Corporation Professor of Management Science, and Itir Karaesmen, assistant professor of management science, with doctoral students Yingjie Lan and Huina Gao, who have developed models that use minimal demand information for revenue management.

Rather than try to create a better forecasting system, the authors worked with the least amount of information to provide robust methods that match demand with supply effectively. They created a model that maximizes revenue under the worst-case scenario for an airline without extensive demand information. In fact, the model does not need any information about demand or the arrival process beyond a simple upper and lower bounds that represent the aggregate demand; this type of information is easier to obtain from expert judgment when there is no historical sales data. The approach relies on competitive analysis of online algorithms, and focuses on “regret” which measures the loss in performance due to lack of information: it compares what the seller can do with limited information versus what he could have done if he had perfect information on-demand.

The policies derived from the model are robust, since they guarantee a certain performance level under all possible demand scenarios. And the average revenues from these policies are comparable to other well-known procedures, even though they use less information.

It’s counterintuitive to think that you could have less information and yet get a result that is as reliable as a result gained from a plethora of data.

“Using less information, we can still get very good results that will allow airlines to make very good decisions,” says Karaesmen. “In some cases you will never have good data. Can you actually make decisions without good data? This paper shows that you can, and in a way that is much faster than existing methods.”

New and smaller airlines, such as Virgin America, or airlines that are opening new flights, might find this model useful, because it allows for reliable revenue management without requiring a stream of pre-existing data to fuel demand-forecasting models.

“Forecasting is a challenge because airlines and hotels only have data on what they have sold. They don’t know what they could have sold,” says Karaesmen. “This is a very safe way of starting a business, until you get to a point where you have lots of data and can use more sophisticated methods.”

But it could also be used as a quality control method by airlines that have made significant investments in demand forecasting. Ball and Karaesmen’s model could run in the background as a type of quality control; if the actual performance differed greatly from the results given by the model, managers could identify errors before they became problematic.

This work builds on prior research by Ball (with Maurice Queyranne of University of British Columbia), who used competitive analysis of online algorithms to produce policies that did not rely on any demand information. Future papers will add a layer of complexity to the problem by considering the problem of overbooking and the effects of competition on this model.

“Revenue Management with Limited Demand Information” will be published in Management Science. This research was funded in part by a grant from the National Science Foundation. For more information about this research, contact mball@rhsmith.umd.edu.