How to upset your customers: a revenue management story

Daniel Hopman (Vrije Universiteit Amsterdam)

From an academic perspective, there is a lot of research into maximizing revenue from an airline's perspective. And, while it is naturally to do so, this can generate potential scenarios that may be confusing, or even upsetting to the customer, which in turn may decrease goodwill. In this talk, we will discuss a few examples of how traditional revenue management and pricing could be perceived as unfair, confusing, or even upsetting - should airlines be more open to their revenue management practices and educate their customers, or is it time to let go?

Recommending bundles

Kalyan Talluri (Imperial College Business School)

We present a model of customer purchases for an assortment of products that is especially suited for the problem of recommending bundles-of-ancillaries that airlines face.   Based on this model we give a fast heuristic that recommends the top-k bundles to each customer segment, taking into consideration product (or bundle) features, assortment-effects, and customer-specific parameters.

How should we measure forecast accuracy?

Jonas Rauch (PROS)

In this presentation we revisit a question that RM researchers and practitioners have struggled with for decades: how to best measure forecast accuracy? We start with the premise that a good error metric is one that is highly correlated with the quantity we are really interested in, expected revenue. With that in mind, we perform a simulation study introducing various levels of forecast error with and without error correlation across forecast entities. We then compute multiple commonly used error metrics on different levels of aggregation. Lastly, to understand which error measures are most predictive of revenue we do exactly that: Train simple ML models to predict expected revenue using each of the error measures as input. We hope that our somewhat surprising results will help RM researchers and managers make better decisions when designing their next forecast accuracy study.

The Ups and Downs of City-Based Seasonality for Revenue Management 

Gizem Burr and Qingleng Tan (United)

In creating a Revenue Management system's forecast which uses normalized conditions for booking history, seasonality factors need to be applied to account for a departure date’s relative strength. In order to identify robust seasonality patterns, similar markets are grouped together into clusters to generate the appropriate factors. However, when certain cities in the cluster have special events or holidays that create vastly different peaks compared to other markets, the forecast quality reduces for both markets to/from the impacted city, as well as the others in the cluster which don’t have anything special going on. A new city-based approach to seasonality is presented to address this issue.

Combining Analytical and Machine Learning Models to build Interpretable Demand Forecasts with Superior Accuracy

Thomas Fiig, Simon Nanty, Ludovic Zannier, Michael Defoin-Platel (Amadeus)

Analytical models (AM) and machine learning (ML) models are often considered to be at opposite ends of the modelling spectrum. AMs are closed form expression based on first principles which require deep domain knowledge, are difficult to construct but can extrapolate to unseen data and are data-efficient and interpretable. At the other end, ML models require little or no domain knowledge to construct, are flexible, and can provide superior accuracy in data-rich environments, but cannot extrapolate, are data-inefficient and are black boxes. We investigate how to consolidate these opposite views to obtain the best of both worlds in the context of airline demand forecasting. We leverage on an existing AM baseline and employ a deep learning-based ML model as a correctional multiplicative term. This approach provides a transparent, interpretable hybrid model with a forecast accuracy outperforming both pure AM and pure ML models.

Who Needs a Good Forecast Anyways?

Ross Winegar and Tom Gorin (PROS)

Who spends all day worried about forecasts?  We’re all here to make money.  Maybe that bad forecast is good?  What would happen if forecasts were perfect?  What does it even mean to have a perfect forecast?  Since users intentionally hurt forecast accuracy all the time for ulterior motives, could I help them be more effective in that? Or should I just ditch the forecast altogether? In this presentation we will briefly review a quick history of airline forecasting and explore what makes a good forecast and how to evaluate the forecast efficacy.

Fare proration versus bid price subtraction in dynamic programming

Bertalan Juhasz (Finnair)

One method of solving the network optimization problem in airline RM is to run a dynamic program (DP) to calculate a bid price vector on each leg. During the DP, the fare of a multi-leg itinerary is converted into a leg fare either by subtracting the displacement costs (i.e. bid prices) of the other legs or prorated based the leg bid price ratios. With proration, no part of the itinerary fare will be double counted on multiple legs and thus the total expected revenue over the whole network will be not be inflated, which makes proration an attractive choice. We will argue, however, that an inflated total network revenue is actually an unavoidable side effect when the bid prices are correctly calculated. Moreover, we will show that the DP calculation results in the correct (approximate) bid prices if the leg fare is obtained by bid price subtraction and not proration. The same holds when the itinerary fares and demands come from marginal revenue transformation.

A Peek Inside the Black Box: Deep Learning Model Explainability with Integrated Gradients

Jon S. Ham and Zoey Zheng (FLYR)

We plan to discuss the use of integrated gradients, a powerful technique for understanding the inner workings of deep learning models, to gain insight into the decisions made by FLYR’s deep learning-based airline pricing models. We employ integrated gradients to reveal the impact each feature has on pricing decisions, enabling Data Scientists and RM Analysts to better comprehend how the model arrives at specific prices based on factors like holidays, events, bookings, capacity, etc. The talk will cover the fundamentals of integrated gradients, showcasing their ability to provide insights into deep learning models, and how we can utilize the technique to understand both global feature importance and individual pricing decisions. Additionally, we will explore their application at FLYR for model evaluation and in the user interface to assist analysts in understanding the model's decision-making process.

Exploring Deep-Q-Learning in the context of Airline pricing

Sharath Nataraj, Jeswin Varghese, Adarsh R, Aparna Muralidhar, Ebin Joseph, Jeswin Varghese, Ranjith Menon (IBS)

Optimizing ticket prices is a critical component of airline operations and a challenging task. Traditional revenue management methods are continuously being refined through machine learning research. Deep reinforcement learning (DRL), a branch of machine learning, has shown promise in various domains, including airline revenue management. In this paper, we explore the effectiveness of DQN (Deep Q Networks) for ticket pricing. Specifically, we train DQN models for two simulated routes, which are somewhat similar and aim to dynamically price tickets to maximise revenue from each route. We present our findings and insights gained from the study, which can offer valuable information to airlines to enhance their revenue maximisation strategies. Additionally, we discuss the feasibility of adopting Deep RL on a broader scale across the airline.

Challenges of Pricing in Air Cargo

Rudolf Zivcic (RMSP)

As passenger airlines also provide air cargo service, and there are many analogies in selling seats and selling cargo space, airlines naturally apply similar strategies for revenue management and pricing in both areas. However, cargo has its own specifics, and on few examples will be demonstrated where adoption from passenger area is beneficial (similarity between booking of cargo space and group booking), where it can be seen as almost illogical (price curve shaped upwards closer to the departure is not reflecting booking urgency and market size), and where new pricing strategies could be developed (using factors/dimensions which are not present in passenger area, such as shipment density). Purpose of this presentation would be to encourage development/ideation of methods and techniques specifically applicable for area of air cargo, as knowledge from passenger side provides perfect base for that, and potential revenue gain in cargo can be significant.

Aumann-Shapley Pricing for Passenger Baggage

Prabhupad Bharadwaj and R.K. Amit (IIT Madras)

Airlines derive a substantial portion of their ancillary revenue from baggage fees. Although there are similarities between cargo and baggage transport, there is no standardized pricing approach for baggage. To address that, we consider baggage transport costs as a joint cost of infinitesimally divisible goods- weight and volume and propose a modified Aumann-Shapley pricing model. The players in this axiomatic method are a continuum of consumable products, and the pricing mechanism is based on the overall cost allocation to the chargeable capacity. We prove that this mechanism satisfies linear functional characteristics in a measurable space. This inspires us to suggest a weighted value strategy for providing advance purchase discounts for pre-booked baggage. The results indicate that the free allowance granted by airlines affects unit prices. This method encompasses all aspects of baggage price and permits airlines to offer “fair engineering prices.”

Balancing Displacement Cost and Willingness to Pay: A Dual Optimization Framework for Airline RM

Cole Wrightson and Grzegorz Goryl (FLYR)

Airline RM and Pricing practices must balance the displacement cost of the next available seat against the willingness to pay (WTP) of the next customer. The variability of WTP across different itineraries on the same leg adds complexity to this balancing act. Traditional RM systems often optimize for a single leg-level control, resulting in the risk of buy-down when WTP is high but demand is low. To address this issue, airlines use techniques such as manual pricing adjustments, rule-based real-time availability adjustments, and dual bid prices. Our work formulates the RM problem as a dual optimization challenge, using two deep neural networks to predict optimal leg bid-price and O&D WTP estimation. Our talk will cover the benefits of this approach and real-world performance, including pricing effectiveness and revenue uplift, and its extensibility to ancillaries and continuous pricing.

Application of quantum computing in the field of airline Revenue Management

Thomas Fiig, Antoine Boulanger, Alexander Papen (Amadeus)

Mogens Dalgaard, Janus Halleløv Wesenberg, Rune Thinggaard Hansen (Kvantify)

Application of quantum computing in Revenue Management is evaluated, both considering the quantum devices that exist today as well as assessing the potential in the future with an increasing number of qubits and reduced error rates. We develop and implement a quantum algorithm for solving the single flight leg optimization problem by mapping the Bellman equation to an Ising model (spin model used to describe magnetism). This allows execution on existing hardware from D-Wave Systems. The results are limited to toy scenarios, but still demonstrate the state-of-the-art of quantum computing. Further, we develop a quantum algorithm for solving a Choice-Based Deterministic Linear Program (CDLP) for network optimization with customer choice. CDLP, although expected to be superior to existing approaches, is far too complex for existing classical computers. Utilizing a quantum computer, we obtain a quadratic speed up, which allow for more realistic problem sizes.

Revenue optimization for airline branded fares with machine learning

Aldair Alvarez; Adam Bockelie; Teodora Dan; Tianjiao Liu; Carl Perreault-Lafleur; Alan Regis; Yury Sambale; Sajad Aliakbari Sani; Cindy Yao; Emma Frejinger; Andrea Lodi; Guillaume Rabusseau (Air Canada and IVADO Labs)

Airlines commonly offer their products in bundles, named branded fares, containing incremental ancillaries. The markup of each branded fare (relative to the other ones) should take into consideration the different needs and willingness to pay of the airlines’ customers due to the significant revenue impact that this can have. However, despite the rich body of literature on revenue management in the airline industry, we have observed a gap in branded fare pricing. In addition, commercial tools are not available to aid decision-makers in this context. In this talk we introduce machine learning and optimization models to capture willingness to pay and maximize the expected revenue from branded fares. The models developed are deployed in a decision support tool for pricing managers.

Recursions for Excursions

Darius Walczak (PROS)

We revisit the single-leg dynamic program (‘DP’), a popular optimization method in revenue management and pricing (‘RM’), that underlies the modern network bid price control in use at many airline carriers.    We focus on the concept of recursion (both backward and forward) to show how to generate several important business metrics and quantities that are crucial in RM practice in the airline world.  We show how to generate several business metrics that are key in practical applications such as expected booked demand, load factor, probabilities of certain events (e.g., inventory stock out), confidence intervals, as well as the distribution of inventory left-on-hand or the booking curve forecast.  The latter, for instance, is used to monitor the pace of accepted bookings. While the DP derives the optimal policy by means of backward recursion once the policy is obtained and stored, forward recursion provides a natural way to generate the aforementioned metrics.

Airfare asymmetry in African countries: evidence from the European destinations

Soheil Sibdari, Professor of Operations Management (University of Massachusetts Dartmouth)

This study compares the average airfare in different African countries that vary in airport infrastructure, competition level, customers’ willingness-to-pay, and governmental regulations. We illustrate our results using non-stop flights between African airports and European markets. The data were collected from various sources, including public data sources, airline websites, third-party booking platforms, and industry reports. This study provides valuable insights into the airfare trends in Africa and Europe, highlighting the differences and similarities between  different African countries

Optimizing Airline Revenue with Dynamic Bundle Pricing: A Collaborative Solution by Airnguru and Piano

Javier Jimenez (Airnguru)

Airnguru, in partnership with Piano, has developed a dynamic pricing engine for bundle discounts that enables airlines to optimize total revenue from potential passengers. The tool harnesses the power of automated A/B testing technology to test various bundle discounts and extract conversion probabilities and price sensitivity for each bundle in different passenger segments. This information is then used to optimize the expected revenue from potential passengers browsing the airline’s webpage. The solution proposes an ongoing process of automated price experimentation and optimization that continuously updates price sensitivity and adjusts bundle discounts accordingly. We have partnered with SKY, a regional LCC, which will start a live Proof of Concept (POC) in May 2023.

Overview of AirSim, a Competitive RM Simulation Tool

Laurie Garrow (Georgia Tech)

In January of 2023, a new center called ATL@GT was established at Georgia Tech under the direction of Laurie Garrow.  The mission of ATL@GT is to lead research and education activities related to airline revenue management. The center is currently focused on conducting research related to offer management, ancillary pricing, and continuous pricing based on recent advancements in new distribution capability, cloud computing, and machine learning. In this presentation, we will give an overview of AirSim, a competitive RM simulator that can be used to test different RM strategies related to offer management, continuous pricing, ancillary pricing, and other areas. To that end, the competitive RM simulator is being designed with a highly flexible architecture that will enable students, researchers, and industry partners to write customized code to test out their own ideas while interfacing with the core simulator.