Demand change detection in airline revenue management

Journal of Revenue and Pricing Management https://doi.org/10.1057/s41272-022-00385-8 RESEARCH ARTICLE Demand change detection in airline revenue management Giovanni Gatti Pinheiro1,3 · Thomas Fiig2 · Michael D. Wittman2 · Michael Defoin‑Platel1 · Riccardo D. Jadanza4 Received: 10 September 2021 / Accepted: 1 May 2022 © The Author(s), under exclusive licence to Springer Nature Limited 2022 Abstract Demand shocks—unobservable, sudden changes in customer behavior—are a common source of forecast error in airline revenue management systems. The COVID-19 pandemic has been one example of a highly impactful macro-level shock that significantly affected demand patterns and required manual intervention from airline analysts. Smaller, micro-level shocks also frequently occur due to special events or changes in competition. Despite their importance, shock detection methods employed by airlines today are often quite rudimentary in practice. In this paper, we develop a science-based shock detection framework based on statistical hypothesis testing which enables fast detection of demand shocks. Under simplifying assumptions, we show how the properties of the shock detector can be expressed in analytical closed form and demonstrate that this expression is remarkably accurate even in more complex environments. Simulations are used to show how the shock detector can successfully be used to identify positive and negative shocks in both demand volume and willingnessto-pay. Finally, we discuss how the shock detector could be integrated into an airline revenue management system to allow for practical use by airline analysts. Keywords Change point detection · Demand forecast error · Airline revenue management · Demand shock · Forecasting · Markov decision process Introduction Motivation for shock detection in airline revenue management Airline revenue management (RM) analysts often spend a significant portion of their time searching for and correcting forecast errors in the airline’s revenue management system (RMS). These forecast errors can be costly to airlines—one study found that as little as a 10% error in an RMS demand forecast can be associated with a 1% decrease in airline revenue (Fiig et al. 2019). * Michael D. Wittman 1 Amadeus S.A.S., Avenue Jack Kilby, 06270 Villeneuve‑Loubet, France 2 Amadeus IT Group, Lufthavnsboulevarden 14, 2. tv, 2770 Kastrup, Denmark 3 University of Nice Sophia-Antipolis, Nice, France 4 Enerbrain SRL, Strada alla Villa d’Agliè 26, 10132 Turin, Italy Forecast errors fundamentally occur due to a mismatch between the demand model parameters assumed by the RMS forecaster and the customer behavior in the marketplace. Usually, shifts in customer behavior are automatically captured by the RMS during forecast parameter re-estimation, which typically uses a historical database consisting of departed flights. However, when customer behavior suddenly changes, the RMS can struggle to adapt quickly, since it takes time for the new behavior to enter the historical database and be detected by the parameter re-estimation. We refer to these sudden, abrupt changes in customer behavior as demand shocks. Demand shocks vary in intensity and can occur at the macro- or micro-level. The COVID-19 pandemic is one example of a highly impactful macro-level demand shock that affected demand across a wide range of flights and origin–destination (O&D) markets, while microlevel demand shocks affecting a handful of flights or markets frequently occur due to entry or exit of a competitor, special events such as conferences, concerts, or sporting competitions, changes in airline schedules, etc., that were not already anticipated and corrected by the airline analyst. Airline analysts typically identify and address demand shocks via relatively simplistic alerting mechanisms. For Vol.:(0123456789) G. Gatti Pinheiro et al. example, Weatherford (2019) describes how an analyst might set an alert to trigger if certain criteria for a flight departure date, such as current load factor (LF), falls above or behind a predefined threshold (e.g., greater than ±5 p.p. compared to the previous year) at a given time prior to departure. If an individual flight is alerted, the analyst would then apply a demand intervention to adjust the forecast for that flight. Vinod (2021) also describes a similar workflow where analysts define alerts by comparing key performance indicators (KPIs) from the RMS to predefined thresholds, conduct a root-cause analysis, and then apply interventions to forecasting or availability in response to a triggered alert. These methods for detecting demand shocks face several limitations. First, they are often quite rudimentary and rely on imprecise heuristics or rules of thumb. Since analysts are not provided with guidance on how to set the alert thresholds, they may either miss impactful shocks (false negatives, Type II error) or be overwhelmed with alerts that, after investigation, turn out to be normal behavior (false positives, Type I error). Additionally, these threshold-based approaches often evaluate flights one at a time, ignoring wider-scale demand shocks that affect multiple departure dates or markets at the same time. They also do not directly consider the effect of offered prices on demand behavior. For example, they may alert an analyst to a flight with a very high current load factor without considering whether the prices offered for that flight were higher or lower than the previous year. In contrast, our method considers the offered prices for each flight when determining whether or not a demand shock has occurred. Finally, traditional approaches to shock detection often consider KPIs taken at a single snapshot when the alerts were generated. Our method utilizes all accessible information—bookings and demand forecast given the control policy—across the entire booking horizon of each flight. Our approach also aggregates data across multiple active flights, allowing for faster and more accurate detection of shocks. Since analysts are often responsible for hundreds or thousands of flight departure dates at a time, this approach allows for greater efficiency and less time spent identifying demand shocks. Contributions In this paper, we introduce a science-based framework for demand shock detection that aims to improve airline analysts’ ability to identify sudden changes in demand. Our detector is based on well-known approaches for statistical hypothesis testing which we have adapted for the shock detection problem in airline revenue management. Given an observed set of booking activity for one or more active (non-departed) flights, we compute the log likelihood that those observations occurred given the offered prices and the RMS’s demand forecast. If the log likelihood—assuming no shock—deviates from a calculated acceptance range, this indicates a poor model description by the forecast parameters and leads to the conclusion that a demand shock has occur (...truncated)