Forecasting demand—both during COVID-19 and after the recovery period—is another fundamental challenge that airlines are facing. According to Ruhlin (2020) of United, “The actual revenue management forecast that we relied on…has become, at least temporarily, less relevant.” Stomph (2020) of KLM elaborated that “We need to reinvent the way we look at bookings or revenue management. The demand curves we used in the past [are not relevant today] and our machine learning algorithms [are not forecasting well] because the data they were trained on are no longer valid because [our historic data are] from a different era.”
Traditional demand forecasting approaches are struggling to adapt to an environment that is characterized by high schedule volatility and ever-changing travel restrictions. As Lange (2020) of Airbus described, “We have no body of data, we have no body of past experience which gives us the full ability to predict what might happen.” Consequently, multiple airlines have stopped using historical data to predict demand and are relying more on manual forecasts and new modeling approaches.
For example, Air Canada set up a task force to predict what the recovery profile would look like at the market level. Air Canada brought in their artificial intelligence (AI) team to forecast what the virus evolution curve could look like in the coming months and how government restrictions could impact demand. Cleaz-Savoyen (2020) explained:
We started with a very simple model based on how demand recovered from previous crises like 2008 and SARS and then progressively started adding more dimensions to it. The demand segmentation is a very important component because we’re expecting the recovery profile to be very different if you are talking about VFR versus business versus leisure versus group demand.
Dennis Buitendijk at Qatar also confirmed that they are now starting to model how COVID-19 will progress, when cases will spike, and how to jointly model interactions between individuals’ desire to travel and travel restrictions.
Consistent with the observation by Richard Cleaz-Savoyen at Air Canada, Ruhlin (2020) of United noted that during COVID-19, being able to separately model leisure and business travelers instead of forecasting demand by class of service has been critical, “because it’s pretty clear so far that the return to willingness to fly on the part of passengers has come a little bit earlier for leisure passengers than for business passengers.” However, Ruhlin added that demand forecasting during the pandemic has been further complicated by the fact that “everything is changing so fast” and with all of the schedule volatility occurring, the ability to first forecast market-level demand and then use a passenger choice model to allocate that demand across flights is “something we really wish we had but that is hard to implement.”
During this period, many airlines are relying more on revenue management users to manually adjust forecasts. As Arjan Westerhof of Air France–KLM shared, “For this moment, we believe our analysts can do a better job than the system in this area because there are a lot of variables that we don’t capture in our systems, like the number of COVID-19 cases,…different travel restrictions, and schedule uncertainty.” The increased reliance on manual intervention was further explained by Ruhlin at United who noted that “even as different waves of resurgence of the Coronavirus come, it’s a different behavior that we’re seeing among passengers and among governments and their restrictions than it was last time. So things are changing so fast that the systems are having a difficult time keeping up” (Ruhlin 2020).
As border restrictions cause pent-up demand, it has become more important for airlines to identify, uncover, and track pockets of demand so that when the travel restrictions are lifted, the airline is ready and has the right capacity to offer customers (Cleaz-Savoyen 2020). Shopping data is one source that have been used to help identify outliers that may be an indication of demand increasing (Cleaz-Savoyen 2020).
Given the shift to more manual forecasting, combined with the fact that “the data you are collecting now is not going to be helpful after the recovery” (Buitendijk 2020), many are exploring ways of blending historical data with manual overrides to support forecasting during and after the pandemic. For example, Belobaba (2020) at MIT has been looking at ways to forecast demand with no historical data and exploring ways to shorten the historical period used in forecasting, pick up on trends more quickly, and incorporate demand adjustments made by human revenue management users.
Cleaz-Savoyen (2020) described how Air Canada has been transitioning back to historical forecasts in some markets using a “deconfining process.” They used this process initially to bring historical data back into their no-show forecasts. First, they worked with their flight model managers to establish criteria for stability and consistency. Once the forecaster was producing forecasts within the “consistent and stable” range, they would start using historical data for forecasting. The process is reviewed on a market-by-market and weekly basis. Air Canada is looking to apply a similar approach to incorporate historical information back into its demand forecaster at a future date, but noted that “it’s going to require a lot more care” because demand levels and booking curves are much different now than they were pre-COVID-19 (Cleaz-Savoyen 2020). Ruhlin (2020) of United also viewed the transition to using historical data for demand forecasting to be a market-by-market decision and one in which the forecast “gets turned back on, but there’s still a certain amount of overrides applied to the forecast.”