Traditionally, airlines relied on historical information combined with time-series forecasting methods to predict future demand. Customer-choice modeling, a more reliable, accurate approach, examines factors affecting customer behavior and uses this data in addition to historical trends to achieve a more precise forecast.
The fundamental task of the airline-planning process is to create a schedule that maximizes profit while satisfying all operational and business restrictions. This objective, in particular, requires matching available capacity with demand and carefully evaluating both cost and revenue components of an airline’s operations.
Over the years, advances in availability of accurate and timely information enabled airlines to calculate precisely all cost components. The demand, however, has an inherent random element that is impossible to eliminate. Therefore, correct demand forecasting is vital for the successful performance of an airline.
Predicting demand for passenger air transportation is required in multiple stages of the planning and operational process. Long-term travel demand is required for fleet and network planning that happens five to 10 years in advance. More detailed passenger forecasting is used in scheduling and capacity allocation prior to each season.
Finally, a very low-level forecast operating with demands for specific routes, fare classes, points of sale and booking periods is necessary for a successful revenue management practice. In addition to passenger demand, bags and cargo forecasts are generated to supplement passenger revenue; originating and connecting traffic forecast is used in planning airport operations; flight on-time performance predictions are required for recovery procedures, etc.
Traditionally, the main technique for generating demand forecasts was based on analysis of historical information and application of some type of time-series forecasting techniques to predict future demand. However, a more dynamic nature of airline operations and the need to adjust to rapid changes in the business environment in conjunction with increased customer focus led to development of new forecasting methodology based on customer-choice models.
This approach focuses on understanding factors affecting customer behavior and predicting it under conditions of future operations rather than simply following historical trends.
Basic customer-choice models assume that a customer’s behavior is driven by attractiveness of available alternatives (competitors). The larger this attractiveness is, the greater the chance of the customer selecting an alternative.
Typically, attractiveness is characterized by utility value that is defined through a set of attributes. In terms of air travel, attributes include price, total travel time, number of connections, departure time and many other factors. Production systems currently used in the industry can apply up to 20 different attributes. Each attribute is assigned a weight coefficient that defines its relative importance compared to the others.
Figure 1 presents an example of a market A-B that has three travel options to choose from: a non-stop flight, a thru flight and an interline connection. The non-stop flight is operated by a national carrier at a convenient departure time, therefore, it has a very high value. The thru flight does not increase elapsed time much and uses the same aircraft for the second leg, hence minimizing chances of being disrupted. Thus, utility of this option is still high. Finally, the interline option has long connection times and departs in the middle of the day, so chances of it being selected are very low. Market shares for each option are calculated as ratios of utility of an alternative and sum of utilities of all alternatives available in the market.
Definition of attributes weights is a complex computational process called models calibration. It is performed by collecting information on travel options that were available in the past and determining weights that make models prediction match actual shares as close as possible.
This is a crucial step for achieving high-forecasting accuracy. For a large airline operating many different markets and serving multiple customer segments, it typically takes several weeks. The process employs sophisticated optimization algorithms acting on massive datasets that reflect historic customer behavior over 12 to 18 months.
One of the main benefits of customer-choice models is the ability to adjust easily to the changes in market conditions. If characteristics of available alternatives in a market change, customer-choice models do not require recalibration or collection of additional historical information to adjust the forecast.
Figure 2 demonstrates how market shares would change if connection time for the third itinerary would decrease. As a result, this option becomes more attractive and “steals” some demand from the other two.
Moreover, customer-choice models are capable of quickly adjusting to a more dramatic change in the market place such as disappearance of one of the alternatives that might occur due to a schedule change or if it is no longer available per revenue management practices.
For example, in Figure 3, the third itinerary was removed and market demand was simply redistributed among the remaining options. Demand adjustments here are proportional to attractiveness of the remaining options.
Finally, Figure 4 illustrates importance of customer-choice models for airline planning processes. Fleet assignment problems deal with allocating optimal equipment type to each flight an airline is planning to operate.
From Constrained Options To Unconstrained Options
Customer-choice modeling also helps estimate redistribution of the demand among options existing in the market in case some of them are constrained with allocated capacity. In this case, demand from constrained options is redistributed to unconstrained options proportionally to their attractiveness.
This process requires an accurate estimation of spilled demand due to restricted capacity and recaptures effect caused by it. Customer-choice models provide an efficient mechanism of redistribution of passengers who could not be accommodated within their original choice to other available alternatives.
This process requires an accurate estimation of spilled demand — all passengers who could not be accommodated within their original choice due to restricted capacity. It also requires estimation of recapture — a process of redistribution of spilled demand to other available alternatives. Customer-choice models provide an efficient mechanism for this estimation that does not require additional calibration or collecting of historic information. Spilled demand is simply split proportionally based on the attractiveness of the options that have open seats. This process requires accurate estimation of the number of people who could not be accommodated within their original choice. Then an accurate estimation is necessary to determine how these people would be distributed to other available alternatives.
Today, customer-choice models are actively used in airlines’ scheduling and revenue management decision-support systems. In scheduling, they generate long-term demand forecasts to provide a base for developing a seasonal schedule. These models are typically calibrated using historical data available in MIDT enhanced with additional information obtained from airports, government agencies and tourist bureaus.
These sources contain information not only about the host airline but also about other competing airlines. Thus, it is possible to forecast demand and market share for all itineraries in each market and predict how changes that one airline makes will affect the others.
This information, combined with fare and cost data, lies in the core of profitability forecasting systems — a backbone of all processes involved in preparing a seasonal schedule such as network development, flight scheduling, codeshare planning and fleet assignment. Decision-support systems that provide optimal solutions for these problems have customer-choice models embedded in their algorithms and, therefore, are able to approximate revenue with high accuracy.
Customer-choice models used in revenue management, although similar in nature, have several important differences. First, scheduling models either do not use fare as one of the attributes or they use it on a very aggregate level only, so demand’s elasticity is typically not captured. In revenue management, however, fare becomes the single most important driver of customer behavior since products offered to customers are differentiated not only by routes, but also by fare classes.
Second, scheduling allows customer-choice models to be calibrated and used during a fairly long schedule development process that happens once a season. On the contrary, revenue management requires much more intense calculations as a control strategy for every flight for up to a year into the future is re-optimized every night. Therefore, revenue management customer-choice models typically contain less attributes and rely on a more-complex defaulting logic. For the same reason, these models usually consider travel options available within the host airline’s network only and represent competition as a single alternative with a given market-share behavior.
As more airlines’ processes focus on customers, customer-choice models will play an even larger role in the future. Areas such as merchandizing, ancillary revenues, loyalty programs, promotions, distribution and many others can benefit from predicting customer behavior and offering travelers the most attractive options. For example, in online distribution, customer-choice models are already applied for display optimization that helps improve convergence ratios. Another example is customization of in-the-air service offers for the specific customer segments travelling on a given flight.
Traditionally, future demand was predicted using direct forecasting that employs time series algorithms. In this approach, forecasts for a specific flight is based on the historical demand for that flight only.
For example, a flight departing from London to Chicago at 10 a.m. on Monday would be forecasted using information for flights from London to Chicago at 10 a.m. on several previous Mondays and potentially on Mondays in the same period of the previous year. However, airlines rarely operate in a stable environment that allows selection of a time period that is long enough to collect sufficient performance data for a specific flight.
Schedule changes might affect departure time of the flight and, therefore, customer preferences for that flight. Departure times of other flights might be affected as well, so some historically available connections would not exist anymore, but some new ones would be created. Even if the airline operates a stable schedule, its competitors might enter or exit the market or change their offer, affecting market shares.
Finally, additional factors such as equipment used on a specific flight, fares offered in the market, strength of the airline’s presence in a particular airport, type of agreements an airline has with its partners and many other factors are extremely dynamic as well. All of these possible changes can affect customer behavior and, therefore, demand for the future flight might be different than the observed history. In the event of major schedule alterations, it might require two to three months for the direct forecast to catch up to these changes and become accurate.
In revenue management, the situation is even more complex since, in addition to schedule characteristics, travel options are differentiated by booking class as well. Throughout the history, a set of available options is constantly changing as some classes become unavailable. Moreover, demand for a specific class depends on availability of other classes and, hence, it cannot be derived by analyzing historic booking behavior for that class only.
For instance, in extreme cases of very elastic demand, there might not be any bookings observed in a class unless it is the lowest class available. Even if this simplification of 100 percent sell-down behavior is assumed to be true, passengers’ booking patterns would still depend on lower classes being available on earlier or later flights in the same market.
These difficulties are eliminated if a customer-choice model framework is used for demand forecasting. Once a customer-choice model is calibrated, it can be applied to any future schedule and availability scenario and instantaneously adjust demand distribution. Forecasting in this case consists of capturing trend and seasonality on an aggregated level that predicts the total demand for the entire market. Then a customer-choice model distributes this demand among travel options that are available at a future departure date.
Even though the customer-choice model approach is well established and proved to generate significant benefits, both academic and industry communities continue working on improving and adjusting it to increase forecast accuracy and better match to changing business requirements.
Over the years, customer-choice models evolved from simple binary choice models to generalized extreme value models that capture additional properties of customer behavior.
For example, nested models combine available alternatives that have similar characteristics into clusters, so customers first choose among clusters and then they select a specific alternative within a chosen cluster.
For air transportation, typical clusters might be defined by departure time of day, so preferences of originating morning travelers are separated from preferences of returning passengers that travel in the afternoon. Alternatively, clusters can combine products offered by each airline operating in a market. This approach captures carrier-specific preferences that are associated with frequent flyer programs, service levels, marketing campaigns, etc.
Until recently, the only information available for understanding customer preferences in air transportation was through observed bookings and operated schedules. Airlines used this information to approximately reconstruct the alternatives from which a customer was choosing.
Now, it is possible to analyze shopping data that includes the exact requests made by potential travelers, the options presented to them (including their display order and all revealed attributes, history of prior searches and bookings for those customers) and, finally, the decisions they made during that particular session.
This information helps identify the time of day customers would like to travel at un-served or underserved markets with high demand. It also helps understand customers’ perception on competing and complimentary services. A combination of shopping data with data available through social networks can also be used for designing better travel offers and discovering new factors that affect customer choice for air travel.
Customer-choice modeling isn’t a brand new concept, and many airlines already benefit from it. However, the potential of this concept is only partially realized. Customer centricity has become one of the major trends in the air transportation industry. Airlines are exploring new revenue sources, discovering new communications channels, collecting much more detailed information about their customers and introducing new customer-management practices.
Customer-choice modeling that helps accurately identify customer preferences is one of the key enablers for these initiatives. With new technological advances in this area and wider application of this concept, customer-choice modeling will continue to contribute to the success of airlines’ planning, marketing and operational processes.