As avid readers of our blog may have noticed in past publications, we often write about the complexity of airline networks (as in the series about the curse of IROPs part 1 and part 2), and the consequential importance of their careful and meticulous design. While the number of variables that impact the cost-efficiency and on-time performance of an airline network continues to grow every year, accurate travel demand forecasts have become increasingly important in airline network design processes. And as if demand forecasting wasn’t already difficult in its own right, another variable entered the mix in 2020: COVID-19.
The COVID-19 pandemic and its impact on the airline industry has been, as stated again and again, unprecedented. Just as Artificial Intelligence (AI) models was starting to find its place in airline planning processes, with airlines moving away from more traditional methodologies based on “gut feeling” and “rules of thumb” to more scientific approaches rooted in data, COVID-19 struck and changed the playing field. AI models that often rely on large amounts of data to yield accurate forecasts have suddenly been cut off from their most valuable fuel: historical data. Since the number of executed flights has decreased by more than 90% in some regions of the world, a drop in demand unseen in the history of civil aviation, models have not been able to match the environments of their predictions with the historical data available to them.
In AI, an environment is considered uncertain “if it is not fully observable or deterministic” (Artificial Intelligence: A modern approach. S. J. Russell, P. Norvig). While one of the main objectives of AI is, indeed, to help humans “make better decisions under uncertainty”, the question now is: has COVID-19 introduced too high of levels of uncertainty in aviation for accurate travel demand forecasts? After all, historical data is not an accurate representation of this new world. Moreover, with the new events and restrictions that impact air travel demand happening daily, what can we realistically expect in terms of forecast accuracy and horizon airlines from AI models?
In the pre-COVID era, airlines steadily started to get a hang of the value of data and digitalisation. Network planning, scheduling and fleet assignment processes were becoming more reliant on historical travel patterns to predict future travel demand. By feeding cost data back into airline network planning in real time, significant improvements in profitability were being achieved.
The decisions regarding launching new origin-destination (OD) pairs used to be based on external indicators such as economic activity. However, even before the disruption of COVID-19, such indicators were noted as insufficient in accurately predicting travel demand—more so since demand should be estimated across all available transport options for an OD pair (air, rail, etc.) and accompanied by passenger profiling so as to better understand their modal choices. Upon realising this, the aviation industry started riding the wave of data science, just in time to be stifled by the pandemic.
The need in the aviation industry to use various data sources to more accurately capture future demand, both in its short and long term changes, is maybe even more relevant now than before. AI models for demand forecasting should make use of data sources such as searches on online travel agents, social media chatter, event planning and professional networks, etc. Moreover, to better understand travelers’ choices across different modes, they should be enriched with anonymised data from sources such as credit card data, social media, etc.
Specifically, in order for demand forecasting models to be able to properly react and adjust to shocks such as the COVID-19 pandemic, they need to be dynamically infused with data on government travel regulations, current events, data related to working from home, or maybe even COVID-19 spread forecasting models; all said sources could be infused in addition to a number of other data sources that should be inspected through correlational analysis.
Developing indicators that have the potential to capture the influence of the events across the globe is also extremely important so that airlines can adjust their schedules accordingly. This is even more relevant in upcoming times in which we can expect to see those events being constantly and rapidly postponed, reorganised or cancelled.
An example of such an indicator that assigns the risk of an event impacting air travel has been developed by PredictHQ. They designed an index by which they measure the risk of an event impacting air travel on the logarithmic scale from 0 to 100, and their solution can be plugged into custom AI models. More of such solutions could prove extremely useful to airlines in their network planning.
Demand forecast models have likewise been shown to benefit from clustering markets by certain characteristics and then feeding that information back into models. As the pandemic evolves at different scales across the globe, the models could be adapted using the information obtained by additionally clustering markets according to various pandemic indicators. Relying on unsupervised machine learning techniques, airlines could group markets into clusters of similar environments in which the travel demand is expected to behave similarly, thereby reducing the need for excessive amounts of historical data.
As we are still learning the range of impact of this outbreak (not only medium or long-term, but even short-term), we will see many swings in demand across various OD pairs. It is of utmost importance for airlines to stay innovative and adapt novel, dynamic and flexible approaches to generating demand forecasts.
Image reflecting real data, data source: Eurocontrol