Revenue management forecasting in times of change – Lessons learned from a year into the pandemic
Thomas Fiig and Mike Wittman - Amadeus
The COVID-19 pandemic has significantly disrupted the traditional paradigm of how revenue management systems (RMS) use historical data to forecast future demand. To avoid polluting the historical database with unreliable or irrelevant observations, many airlines froze their demand forecasts and relied solely on manual interventions to steer flights. In this talk, we discuss the lessons learned from adapting an RMS to an environment where historical data was no longer reliable. We describe the methodology of a newly-developed forecasting concept that rapidly adjusts forecasts based on as little as a few months of live sales data, and discuss how separating forecast components into two categories – resilient and volatile – allowed us to ensure forecast stability while enabling adaptivity to the latest trends. We demonstrate how our method reduces forecast error using actual airline data, and discuss learnings from deploying this concept into production. Finally, we discuss how we see the future of demand forecasting in light of this changing business environment.
Demand Shock Detection in Airline Revenue Management
Michael Wittman (Amadeus), Thomas Fiig (Amadeus)
Demand shocks – unobservable, sudden changes in customer behavior – are a common source of forecast error in airline revenue management systems. The COVID-19 pandemic was one example of a highly impactful macro-level demand shock, while smaller micro-level shocks often occur due to special events or changes in competition. Airline analysts currently have limited and simplistic tools to detect demand shocks affecting their flights or markets. In this talk, we introduce a science-based framework for shock detection to quickly identify sudden changes in demand. The framework allows us to directly compute the statistical relationship between time to detection and various criteria, including shock magnitude, statistical power, and sample size. This allows analysts to easily configure the detector based on their specific needs. We show through simulation how the detector can be used to identify shocks in volume or willingness-to-pay affecting one or more flights.
FLAI: Reinforcement Learning Virtual Platform for Travel
Naman Shukla (Deepair Solutions), Arinbjorn Kolbeinsson (Imperial College), Benedikt Kolbeinsson (Imperial College) and Kartik Yellepeddi (Deepair Solutions)
Recent achievements in reinforcement learning are fueled by agents learning to interact in virtual environments such as video games, where algorithms can be evaluated in a safe and reproducible manner. We introduce FLAI, a novel set of reinforcement learning environments which simulate the marketplace interaction in travel. Using different environments in FLAI, agents can be trained to learn various pricing strategies in realistic settings. FLAI is designed to be challenging yet learnable for an agent, easy to use and customizable for the user. Additionally, it will be available under a permissive open-source license. Here, we preview different market scenarios with varying difficulties, benchmark performances of standard agents, and baseline results for one of the environments - SeatSmart: paid seat pricing.
Optimizing flight and truck network for cargo at American Airlines
Ravi Suman (American Airlines), Ronald Chu (American Airlines), Na Deng (American Airlines), Cumhur Gelogullari (American Airlines)
In this work, we optimize the flight and truck network of American Airlines (AA) using a proposed Linear Programming formulation. In the linear programming, we maximize the profit of AA’s cargo network by making decisions on optimal cargo flight allocation plans and trucking strategies. We identify new freight cargo opportunities by looking at the industry market share vs current AA market. We additionally develop a network constructor to determine all the feasible routes for our flights and truck legs. Preliminary estimation of benefits is $2M in incremental revenue from optimization.
Enhancing day-ahead airline planning with data-driven flight delay predictions
Sebastian Birolini (University of Bergamo), Alexandre Jacquillat (MIT Sloan School of Management), Stephanie Franklin (MIT Sloan School of Management), and Gabrielle Rappaport (MIT Sloan School of Management)
Flight delays are the major drivers of disruptions and unexpected costs in airline operations. It is therefore of paramount importance to get visibility into flights’ delays as early as possible and as accurately as possible, in order to minimize their overall impact. In this paper, we collaborate with Vueling Airlines to build predictive models of flight delays and enhance day-ahead planning decisions accordingly. We first assemble a large-scale database of flight-level observations, using airline-specific features, system-wide features, and environmental features. Using a quantile regression model, we estimate minimum turnaround times for each pair of flights and reconstruct each flight’s primary (as opposed to propagated) delay. We then develop machine learning models to predict primary delays. Our best model, based on extreme gradient boosting, achieves a mean absolute error of 7–8 minutes—a significant improvement as compared to baseline models using simpler machine learning methods or simpler sets of predictors. Finally, we embed our data-driven delay predictions into a tail assignment model to support day-ahead planning. Out-of-sample results demonstrate that leveraging the proposed predictive model can reduce overall delay costs by 3–5%. Ultimately, this paper shows the potential of combining advanced predictive and prescriptive analytics methods to enhance airline planning and operations decisions.
Organizational Business Risk Management Based on Human Factor Concept
Dejan Devic (Air Serbia), Velimir Isakovic (Air Serbia)
HF (human factor) in civil aviation accounts for more than 80% of the organizational business risks. Technical and technological risks can be easily found from the reliability of the most vulnerable segment. HF cannot be directly calculated as people are not reliable outside the preventive and proactive (vector) management system, implying the use of non-technical knowledge and methods. The goal is to establish a reliable RM (risk management) system with the use of advanced HF new RM methods. These include proper employees’ training, increasing and clarification of the responsibilities, matching talents to jobs. Operating cost savings of up to 25% and efficiency increase by 30% can be achieved in all segments of work-related processes. The difference between the increasing profits and the bankruptcy is often contained in these small changes which ultimately lead to the survival of a low-profit transport sector such as aviation.
Estimation of price elasticities on passenger itinerary choice situations
Rodrigo Acuna-Agost (Amadeus); Eoin Thomas (Amadeus); Alix Lheritier (Amadeus)
During this session we will present our recently published paper “Price elasticity estimation for deep learning-based choice models: an application to air itinerary choices” (Journal of Revenue and Pricing Management, 2021). In this work, we address the dilemma of choosing an approach to predict traveller choices: should we focus on high accuracy (e.g., using deep learning) or on high business interpretability (e.g., using MNL models)? We present a novel methodology to estimate price elasticity (gaining interpretability) from deep learning-based choice models (highly accurate). The methodology is flexible enough to allow price elasticity analysis at a single shopping session level. All these insights are of interest of airlines to better adapt their offer, allowing them to take into account detailed attributes of the travel solution such as day of week, trip purchase anticipation, type of trip – domestic/international and many others.
Dealing with distribution shifts in customer choice due to COVID - 19
Abhinav Garg (University of Illinois at Urbana-Champaign), Naman Shukla (Deepair solutions), Lavanya Marla (University of Illinois at Urbana-Champaign), Sriram Somanchi (University of Notre Dame)
Traditional AI approaches assume that the data distribution at test-time is similar to training. However, this assumption may be violated in practice because of the dynamic nature of customer buying patterns due to unanticipated events such as COVID-19. This work presents a study on how the customer choice of ancillary purchase for one of the major airlines changed during the COVID-19 pandemic. We frame this problem as a covariate shift detection problem to identify which customers changed their purchase behavior and the attributes affecting that change through a series of experiments employing different shift detection methods from literature. The results indicate that the customer choice changed during the pandemic with notable changes in increased advanced bookings and preference for one-way trips. We also recommend different approaches that can be used to mitigate such changes to increase the overall performance of the pricing model.
Crystal Ball 2.0 for passenger demand: Leveraging AI to power the calibration workflow
Caroline Dietrich (Air Canada), Neda Etabarialamdari (Ivado Labs), Peter Wilson (Air Canada)
A well-calibrated demand forecast is an essential driver for revenue management systems. Traditional forecasting solutions rightfully focus on very granular forecasts based on bookings and constraints adapted to different optimization methods. To calibrate passenger demand forecasts more quickly and accurately, Air Canada teamed up with IVADO Labs to develop an AI-driven analytics solution that provides key insights to Revenue Management (RM) practitioners freeing up time for strategic decisions. Named Crystal.AI, this innovative module integrates the demand analyst workflow end-to-end and recommends adjustments to existing systems by prioritizing those with the greatest impact. The methodology combines machine learning models and mathematical programming. First, a Recurrent Neural Network (RNN) model predicts demand. Second, these predictions, along with initial RM forecasts, are used as inputs to a mixed integer programming formulation whose solution is a suggested list of adjustments to the RM forecasts having the highest impact. The RNN were trained on historical data as well as augmented data to deal with changed demand patterns due to the COVID-19 pandemic. Designed, tested, implemented, extended, and improved in several phases and with close collaboration between the teams, the new tool is being tailored to add even more value in the challenging environment the industry is currently navigating.
Airline Recovery using Machine Learning and Optimization
Ahmet Esat Hizir (Massachusetts Institute of Technology), Cynthia Barnhart (Massachusetts Institute of Technology), Vikrant Vaze (Dartmouth College)
Due to the irregular nature of flight operations, airlines need to take a range of actions to recover their aircraft and crew schedules. The limited time frames prevent airlines from using a full-scale optimization approach. Consequently, airlines usually apply recovery solutions that are far from being optimal. This study proposes a practical method that combines machine learning and optimization to find better solutions for the recovery problems than alternative approaches. The developed procedure is based on the idea that the most effective constraints (cuts) to add to the recovery models without sacrificing the solution quality can be determined in advance by leveraging the similarities between disruptions. Conducted experiments show that this approach can accelerate the optimization significantly while keeping the solution quality close to the optimal.
Mission success and survivability in Flight Planning
Fanruiqi Zeng (Georgia Institute of Technology), John-Paul Clarke (University of Texas at Austin)
In this work, we propose a receding horizon control strategy with novel trajectory planning policies that enable dynamic updating of the planned trajectories of autonomous vehicles operating in environments where potential conflicts are, from a statistical perspective, either partially known or completely unknown. The proposed policies utilize two metrics: (1) the number of feasible trajectories; and (2) the robustness of the feasible trajectories. We measure the effectiveness of the suggested policies in terms of mission survivability. We demonstrate that a linear combination of both metrics is an effective objective function when there is a mix of partially known and unknown uncertainties. When the operating environment is dominated by unknown disturbances, maximizing the number of feasible trajectories results in the highest mission survivability.
Multi-Modal Economic Analysis of COVID-19 Pandemic Exogenous Shock to Domestic Commercial Aviation
Max Litvack-Winkler (USDOT Volpe Center), Jacob Wishart (USDOT Volpe Center), David Pace (USDOT Volpe Center), Seamus McGovern (USDOT Volpe Center)
A COVID-19 impact study was conducted to research and quantify the pandemic’s effect on different transportation sectors, with a focus on domestic commercial aviation. The study focused on two categories of affected data sets—passenger and cargo—for several modes with an objective of seeking to determine any possible patterns or relationships, as well as to describe how each is recovering. The multi-disciplinary, data-based analysis is intended for use in determining possible mitigations and provide for future resiliency. This presentation will provide a description of the study task and objectives, examine the data that were considered, review the data selected for use in the analysis and why, and discuss the comparative results between transportation modes and other economic indicators before and after the shock, as well as general comments and observations on those results to include identification of pairs having strong statistical correlations.
Identification and Prediction of Disruptions in Airline Networks
Xiyitao Zhu (University of Illinois at Urbana-Champaign), Max Z. Li (Massachusetts Institute of Technology), Karthik Gopalakrishnan (Massachusetts Institute of Technology), Hamsa Balakrishnan (Massachusetts Institute of Technology), Aritro Nandi, Lavanya Marla (University of Illinois at Urbana-Champaign)
Air transportation disruptions can lead to and are caused by demand-capacity imbalances, resulting in flight delays and cancellations due to traffic management and system recovery actions. To better predict the impact of disruptions and provide targeted system recovery actions, it is critical to identify characteristics such as: (1) When did a disruption begin and end, (2) Where and with what intensity did a disruption occur, and (3) how will an ongoing disruption evolve. We formalize the notion of disruption-recovery trajectories (DRTs), which capture information regarding both the magnitude and spatial impact of disruptions. We use DRTs to represents network performance metrics as transitions between discrete states, and develop two prediction models: The first identifies whether or not the system will recover in the next hour; the second seeks to predict the trend of key performance metrics. We report prediction results for four major US airlines.
The adjustment of Littlewood's Rule formula to reflect airlines risk aversion in crisis conditions
Yassine El charkaoui, (IATA employee, speaking in a private capacity)
The computation of the protection level u- number of seats to be protected for higher fare customers against lower fare ones using Littlewood Rule (LR), focuses on revenue maximization, and implicitly assumes that airlines are risk neutral, which is not always true , especially in cases of crisis such as COVID 19. Under crisis conditions, airlines risk appetite will naturally shift from a risk neutral or risk seeker status to a risk averse one. In difficult times, the utility value from getting less revenue amount with certainty exceeds the utility value of getting higher revenue amount with uncertainty, in other words airlines in dire financial situation will focus more on utility maximization rather than revenue maximization. As result of this, the computation of u using (LR), for a risk averse airline, would be overstated. This working paper will answer "how (LR) can be adjusted to come up with u values which reflect the risk averseness of airlines?"
Know your worth: valuing new pricing policies with reinforcement learning
Jon Ham, Flyr Labs
Revenue management software tools often claim superior pricing model performance. But given two new pricing models, how can we know which one yields better revenue performance without experimenting with real-world airline revenue? We look to Reinforcement Learning (RL), the branch of machine learning that trains models to interact optimally with a changing environment. However, many RL techniques require the model to interact with the real-world, which can make applications to airline revenue management somewhat financially risky. New RL approaches have been developed called Offline Policy Evaluation (OPE), which use historical data on past decisions to learn the value of any new decision-making policy, without having to deploy the new policy in the real world. We explore OPE techniques as a method to determine which of our RM pricing models would yield the best revenue performance prior to implementation, and compare this to how they actually performed.
The fundamental weak point of quantum computers
Maxim Andreyev (Rosterize), Dr. Alexey Tarasov (Rosterize)
Quantum computing (QC) is near. QC researchers are promising that they will be helpful for several areas including combinatorial optimization, which is very important for crew scheduling problems. This presentation will show that QC has limitations that are unlikely to allow a revolution in combinatorial optimization. Instead of technical limitations which will disappear with time, I am looking into the intrinsic properties of QC. Main such property – limitation of solving problems with several global minima. It is a big problem for all types of QC: classical and adiabatic. A QC cannot choose from two identical optimal results. QC can be super-efficient if in the problem with a unique answer, even if this answer is hidden. But for combinatorial optimization, some problems will be simple for QC and some difficult. And most disappointing, that most complex combinatorial optimization problems are also tricky to QC as well.