For industries such as airlines and hotels, finding the best pricing model is the difference between survival and certain death. If an airline charges more than the market will bear and flies a row of empty seats from Dulles to Los Angeles, it can never regain that lost revenue. Finding the right mix of fares to maximize profits, while still filling the plane, is the science of revenue management.
“There’s always a sweet spot between how many seats you want to sell or target to each segment, and that becomes a complicated decision,” said Assistant Professor Itir Karaesmen Aydin, whose expertise is in pricing optimization and revenue management in the travel, hospitality, and service industries.
Airlines, hotels, and similar businesses have finite inventory, making them profitable only to the extent that they maximize revenue for each unit. At the same time, they must be careful not to set prices so high that no one buys.
Air travel is unique in that passengers have varying degrees of need: a business traveler meeting an important client in Chicago may be willing to pay handsomely for a seat leaving Atlanta at 6:35 a.m. on Friday, while the leisure traveler spontaneously visiting the Windy City may go only if she gets a great deal.
Airlines traditionally rely on historical passenger data, feeding these into mathematical algorithms that determine the best mix of fares for a particular flight and how many seats to overbook in case passengers cancel. But this method isn’t feasible for a new airline or an existing carrier opening a new route. Flying blindly, as it were, is too risky for an industry already on shaky financial ground, so airlines need a reliable way to make those decisions.
Enter Karaesmen. She has helped create a new decision-making model that would work even if you didn’t have historical data. The research relies on judgmental forecasts and is based on the concept of nesting, in which an algorithm determines the best mix of fares. Nesting might indicate, for example, that an airline will sell no more than 10 tickets for $100 or less, 15 tickets for $200 or less, and 20 tickets for $300 or less.
As random as fare fluctuations may appear to consumers, they are actually far from it. If an airline’s nesting algorithm determines that it can sell no more than 10 tickets for $100, the 11th person will be out of luck (though still able to buy a pricier ticket). Then again, suppose the next day’s weather forecast calls for an ugly weekend in Chicago, causing leisure travelers with flexibility to postpone their visit. An airline analyst might reopen the lower fare so that buyers can once again buy a $100 ticket.
“There’s a constant interaction between the algorithm and the system and the revenue, and the analyst can actually override a decision as the flight time gets closer,” Karaesmen said.
According to the Federal Aviation Administration, the airline industry’s financial woes are unlikely to improve in the foreseeable future. Its forecast predicts a slow recovery from setbacks ranging from the 9/11 attacks to the recent recession, and says airlines will need new strategies to ensure profitability.
Maximizing revenue on each flight is a critical component of that profitability. Karaesmen explained that small improvements to decision-making data can have exponential benefits, considering that a major carrier may have hundreds of thousands of flights in a single year. Delta Air Lines, for example, had a fleet of 740 planes in 2009 and carried 108.6 million passengers, according to the trade group Airlines for America.
On the flip side, poor data can be disastrous to decision making. Researchers who work on mathematical algorithms have a term for this, Karaesmen said: “Trash in, trash out.”
Before joining the faculty at American University, Karaesmen co-led a group of doctoral students studying revenue management with limited demand information. While the traditional practice of revenue management assumes that airlines will use historical data to create probability distributions about passenger demand, the study provided a distribution free mathematical model for cases where that data did not exist. Instead, the researchers set upper and lower bounds on the range of possible scenarios by considering worst-case and best-case outcomes.
The mathematical models that airlines traditionally use to make decisions require heaps of information, Karaesmen said, and data must be of high quality or the algorithm results will be skewed. Trash in, trash out.
“The argument in our paper is, especially for new routes, that figuring out these probabilities is difficult and you will get inaccurate numbers,” she said. “The question is, ‘Can I build an alternate model that would work even if you didn’t specify any probabilities?’” The new model is so different from previous approaches that it poses a fundamental question about the validity of existing decision-making processes.
Research succeeded, confirming that airlines can make effective fare decisions and retain the nesting structure even without probability data.
|The number of passengers who rode US airlines in 2011, on 9,455,092 domestic and international flights|
|1,150||Complaints about cancellations, delays, and misconnections, for comparison||3||The number of formal compliments passengers bestowed on US airlines during the same time period|
|6,306||Jets and other planes that made up the US passenger airline fleet in 2011||3.8||The percentage of formal complaints to US airlines that concerned overbooking in the first five months of 2012: 192 out of 5,071 complaints|
Sources: Airlines for America, Bureau of Transportation Statistics, US Department of Transportation
Next, Karaesmen analyzed whether the model established in the previous paper could apply to overbooking.
The practice affects more travelers than you might think. In the second quarter of 2012, 13,236 passengers on 16 airlines were bumped involuntarily as a result of overbooking, according to the US Department of Transportation (DOT). Another 155,051 passengers were bumped voluntarily, meaning they accepted travel vouchers or other compensation to give up their seats.
From the airlines’ perspective, overbooking is crucial. If they don’t overbook, they run the risk of flying with empty seats, which means lost revenue. In her paper, Karaesmen noted that in 2006, US Airways would have lost at least $1 billion had it not engaged in overbooking. The trick, of course, is deciding how far to push it. Airlines need a scientific method for making those decisions to ensure they don’t leave anything to chance—the financial stakes are simply too high, Karaesmen said.
The practice hasn’t always gone smoothly, however; in exchange for letting airlines overbook, the DOT mandates certain protections for passengers. Between 2007 and 2011, the DOT assessed fines against eight airlines for violations related to oversales practices. Alleged violations by Delta, including failing to properly compensate passengers for their missed flights, triggered a $375,000 fine. In April 2011, the DOT amended overbooking rules to increase passenger compensation and set stricter requirements for airlines.
In addition to studying whether they could apply their mathematical model to overbooking, the researchers removed the standard assumption that the decision maker is risk-neutral—that is, someone who takes the long view so that as long as the carrier performs well on average, it isn’t too concerned about short-term lows. A new airline carrier likely will feel differently, Karaesmen argued. It will be more concerned about the worst-case scenario because if the carrier doesn’t break even soon, it may go out of business. Improving its worst-case scenario would enable a carrier to improve all possible scenarios.
“We’re saying that in a new business, you don’t have too much time,” she added. “You want to make sure you do well from the start.”
They also took into account the regret criterion, which measures the effectiveness of a policy by benchmarking it to the performance of the ideal action, as determined by 20-20 hindsight.
When the researchers put their new model to the test, it produced “effective, consistent, and robust decisions” for overbooking and fare allocation and proved that nested booking limits are the best way for airlines to maximize revenue.
The third article in the series built on the findings of the first two studies, but introduced a curveball: What happens when there is competition on the route?
The study applies not only to competitor carriers, but also to multiple routes by the same carrier. In a sense, US Airways is competing against itself by offering six DC-to-New York flights on the same day. The researchers used game-theory analysis, a study of strategic decision making, to examine the effect of competition on fare decisions in their mathematical model, with limited data on consumer demand in a worst-case scenario.
They found, Karaesmen said, that the nesting algorithms hold up, despite using less data.
“We did a lot of simulations and analysis afterward to show that the quality of the decisions is as good as if I had more information such as probabilities,” she explained. “The quality of the decision is still as good if you use our framework versus the old framework, so the framework we brought is extremely useful for new flights, new routes, new hotels.”
This research garnered a good bit of attention from airline employees when presented at conferences, according to Karaesmen. Transitioning to a new model would be difficult for airlines, however, because their current systems are so complicated.
“If you wanted to run my algorithm as opposed to theirs, it would require a significant investment,” she said. “Airline employees understand its value, they definitely see potential in that, but implementation [throughout the whole network] is difficult.”
The next step in her research may be the toughest one yet: understanding consumer behavior. The explosion of information available to travelers has made them significantly more sophisticated than travelers of 20 years ago, Karaesmen said, which means airlines must adapt.
“So far in the airlines, the algorithms will treat everyone like it’s a herd: there’s a herd of passengers and 10 will pay $100, 20 will pay $200, and so on. But for the individual passenger, the decision making is more complicated. So then the question becomes, is there a way to incorporate individual decision making and choice behavior to model demand? Will worst-case analysis or models that use the regret criterion be effective in this more complicated setting?”
Y. Lan, H. Gao, M. Ball and I. Karaesmen, “Revenue Management With Limited Demand Information.” Management Science (2008).
H. Gao, M. Ball and I. Karaesmen, “Competitive Seat Inventory Control Decisions Under the Regret Criterion.” Journal of Revenue and Pricing Management (2010).
Y. Lan, M. Ball and I. Karaesmen, “Regret in Overbooking and Fare Class Allocation on a Single Leg.” Manufacturing & Service Operations Management (2011).