|Continuous Pricing with Multiple Fare Quotes
|Continuous pricing enables fare quotation from a continuous range, as opposed to the current industry standard limiting airlines to a fixed set of published price points. Extending previous work in the PODS consortium, we introduce a framework for continuous pricing with multiple fare quotes through product differentiation and fare quote segmentation. Customers may be offered multiple continuously priced products to choose between (product differentiation), and their offers may differ based on segment identification at purchase request. PODS simulations are used to illustrate the potential for revenue gains from applying continuous pricing methods relative to traditional class-based RM and to investigate their competitive impacts.
|Airline offer optimization with the Markov chain choice model
|Airlines today distribute and price bundles of flights and ancillaries (i.e., Fare Families and Branded Fares), as well as a la carte ancillaries, using filed fares. New distribution technologies have enabled airlines to distribute an offer set containing multiple offers, each of which consists of a set of atomic products and a single offer price. The generation of prices for each offer – which depend on the other alternatives in the offer set – represents a new optimization problem in airline RM. In this talk, we discuss how the recently proposed Markov chain choice model (MCCM) could be used to jointly construct and price bundled offers. The MCCM is a flexible choice modeling framework that can be used to represent any random utility model and allows for efficient computation of optimal prices through a series of univariate optimizations. We show how MCCM prices differ from myopic, a la carte pricing (where each offer in the offer set is priced independently) and how MCCM can provide a revenue uplift. We also illustrate how the parameters for the MCCM can be estimated from historical data via maximum likelihood estimation.
Caroline Dietrich & Peter Wilson
|Crystal Ball 2.0 for passenger demand: Leveraging AI to power the calibration workflow
|Ivado Labs + 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.
Crystal Ball 2.0 for passenger demand: Evaluating the impact of the new end-to-end AI-driven analytics solution
|Ivado Labs + Air Canada
Airlines have been improving their already highly complex Revenue Management Systems for years, with both state-of-the-art models and enhanced workflow. Estimating the impact of these improvements on a key performance indicator of interest such as revenue is crucial, yet very challenging. It is indeed the difference between the generated value and the value that would have been generated keeping business as usual, which is not observable. We developed a methodology which predicts the counterfactual, not observed, revenue and compares it to the observed revenue subject to the impact. We provide a comprehensive overview of counterfactual prediction models. We use them in an extensive computational study, as a natural follow-up from the development of the end-to-end AI analytics solution of Air Canada and IVADO Labs. We present accurate linear and deep-learning counterfactual prediction models which achieve respectively 1.1% and 1% of error and allow us to estimate a simulated effect quite accurately.
Michael Wu and
|What Airlines can Learn from Retailers About Demand Forecasting
|2020 was a year of turbulence and disruption, especially for the travel industry. Drastic changes in travel behavior have greatly compromised airlines' ability to forecast future demand and therefore optimize revenue. Rather than relying on long histories over the past years, airlines must monitor the recent histories more carefully to understand the demand trajectory of the near future. This temporal squeeze significantly reduces the data that airlines can leverage and will increase the uncertainty of our demand forecast. To improve our forecast’s confidence bound, airlines must learn from the retailers and start leveraging a wider variety of data sources as predictors of future bookings. We constructed a model that leverages flight search data and found that this shopping data captures substantial information about people’s intent to book. This is illustrated by the fact that even a simple structured linear model is able to predict bookings with a high degree of accuracy (i.e. ~0.8 as measured by Pearson’s correlation coefficient). As more people are vaccinated, airlines that can augment the shorter relevant history creatively using shopping data will regain their capacity to forecast demand sooner and recover from the pandemic faster.
|Dynamic Pricing of Ancillaries using Reinforcement Learning
|Ancillaries have become a key driver for revenue growth for travel industries. Traditionally, pricing and offer generation for ancillary items have been managed using static business rules. In such scenarios, where historical prices show very little or no variation, typical methods of estimating purchase probabilities and then finding the optimal price are not applicable. In this study, we develop practical approaches for dynamic pricing of ancillaries based on reinforcement learning ideas. We propose a contextual bandit model for dynamic pricing of ancillaries considering the trip and customer features. The pricing setting presents significant challenges to the application of the multi-armed bandit framework since the arms are highly correlated. To capture correlation across arms, we consider a Bayesian logistic bandit framework and use Variational Bayes methodology to construct fast and scalable algorithms for this setting.
|Scenario-driven RM using simulation-based reinforcement learning
|Standard RM optimization methods such as dynamic programming make strong assumptions about the demand arrival process and are therefore not well-suited to deal with uncertainty about the demand distribution or correlation across the booking horizon. One way around this is a scenario-driven approach, but the corresponding optimization problem cannot be solved using dynamic programming. Instead, we use simulation-based reinforcement learning to approximate the value function and corresponding bid prices. Because standard RL methods such as Monte-Carlo or Temporal Difference (TD) learning do not lead to satisfying results in this context, we propose a modified TD learning update that aims to improve the quality of the bid price rather than the quality of the value function estimate.
|Fare level optimization with continuous demand-revenue curves and bid prices
|In an ideal revenue management world, airlines could always offer the optimal continuous price to customers to maximize revenue. Currently, however, most airlines can only offer a set of fixed fares which results in some lost revenue, therefore it is important to have optimal fare levels which minimize this lost revenue. We present a novel method of fare level optimization which does not use previous bookings, does not use an aggregated price-demand curve (without time dimension), does not need the minimum and/or maximum fare(s) to be set in advance manually, and does not assume that passengers with higher willingness-to-pay come after passengers with lower willingness-to-pay. Instead, it uses the daily forecasted continuous demand-revenue curve (obtained from the forecasted price-revenue curve) to first obtain the optimal continuous price to offer which is at the point where the tangent of the curve is equal to the bid price; this point also tells us what the expected daily optimal demand and optimal revenue would be at this optimal continuous price. However, we only have fixed fares therefore we cannot offer exactly this price. Instead, we can offer a "mix" (i.e. linear combination) of two fixed fares, one higher and one lower than the optimal continuous price, such that we get the same demand which we would get with the optimal continuous price; this way the demand i.e. booking intake would follow the same optimal path as with the optimal continuous price. However, the revenue that we get this way is slightly less than what the optimal revenue with the optimal continuous price would be. We can then find a set a fare levels which minimizes this lost revenue aggregated across all future departure dates and booking dates. It should be a subject of further research if such "mixing" of fares is justified in this optimization context.
Thomas Fiig and
|Demand Forecasting in Times of Change – Lessons Learned from a Year into the Pandemic
|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.
|How much to tell your customer? – A survey of three perspectives on selling strategies with incompletely specified products
|University of Duisburg-Essen
|Today’s technology facilitates new selling strategies. One increasingly popular strategy uses incompletely specified products (ICSPs). The seller retains the right to specify some details of the product or service after the sale. The selling strategies’ main advantages are an additional dimension for market segmentation and operational flexibility due to supply-side substitution possibilities. Since the strategy became popular with Priceline and Hotwire in the travel industry, it has increasingly been adopted by other industries with stochastic demand and limited capacity as well. It is actively researched from the perspectives of strategic operations management, empirics, and revenue management. This talk first describes the application of ICSPs in practice. Then, we introduce the different research communities that are active in this field and relate the terminology they use (e.g. opaque selling, flexible products, upgrades). The main part is an exhaustive review of the literature on selling ICSPs from the different perspectives. We see that strategic operations management has described advantages of ICSPs over other strategies in a variety of settings, but also identified countervailing effects. Today, empirical research is confined to hotels and airlines and largely disconnected from the other perspectives. Operational papers are ample, but mostly concerned with the availability of ICSPs. Research on operational (dynamic) pricing is surprisingly scarce.
|Tim Yuxuan Lu
Conditioned Sell-up Rate Estimation and Demand Forecasting under Unrestricted Fare Structures
Unlike in a fully restricted fare structure where passengers voluntarily purchase higher fare classes, passengers facing a fully unrestricted fare structure always choose the lowest available fare across airlines, creating spiral-down problems for sell-up estimation and strong competitive feedback effects for forecasting. This presentation examines the shortcomings of the current WTP-based Q-forecasting methodology under a fully unrestricted fare structure and describes an alternative approach for conditional forecasting and sell-up estimation. The proposed conditional forecaster generates forecasts explicitly conditioned on demand volumes and lowest available fare class information from all airlines in the market. Initial proof-of-concept experiments have shown promising accuracy improvements over Q-forecasts.
An Approach to Airline Offer Management Dynamic Bundling and Pricing
This talk will provide a practice-oriented, systematic framework for determining the composition and pricing of air ancillary dynamic bundles with customer trip-purpose segment and 1-to-1 personalization to maximize total sales. Extensions of this framework to more generalized applications such as air + ancillary and dynamic packaging (with non-air content) will also be described.
|Practical applications of fare search shopping data
|Sabre and Etihad
Traditional Revenue Management methods are heavily dependent on demand forecasts based on historic patterns and capacity optimization. As passengers increasingly rely on comparison shopping checking for different travel options before making the purchase, tapping into this shopping data and its influence on the customer choice will be beneficial and generate incremental revenue benefits for the airlines. In this presentation, we will describe three different use cases that leverage shopping data and how they generate incremental revenues. First use case is related to dynamic availability and pricing (i.e. how market context from shopping data can be used for improving an airline’s price competitiveness and yield in fare search). We will also show how this same information can be utilized for departure-date specific fare filings. Second use case is related to identifying the differences in prices for the same itinerary between operating carrier and marketing carriers (codeshare partners) – these price differences can cause revenue leaks, and addressing them will be beneficial. Third use case is related to using the shopping data to identify the demand for future flights instead of completely relying on the historic booking patterns – this will be beneficial in the post COVID world as travel schedules and patterns change.
Amrit Raj Misra
Yun Hsuan Tai
|Machine learning to dynamically price and personalize ancillaries
|Sabre and Etihad
Ancillary prices are filed in ATPCO and Merchandizing systems. These prices are generally determined by benchmarking, price testing, and historical data analysis. Ancillary pricing decision making is mostly manual, and analysts mine flown data to come up with complex pricing decisions that depend on various passenger, travel, aircraft, ancillary product and seasonal attributes.
Using airline ancillary and itinerary data, we built a platform that uses (ex: Gradient Boosting Machine (GBM) & Deep learning algorithm) that can understand intricate relations between numerous attributes and can make fast & accurate pricing decisions automatically. The analysts are relieved of manual work and have the flexibility to tweak the machine learning (ML) algorithm to suit business strategies. The ML algorithm learns the new trends and patterns as part of its training, and analysts can track its performance periodically. The output of ML algorithm seamlessly integrates with merchandising platforms to implement the Dynamic pricing of ancillaries and offers in the direct and NDC channels during booking or post-booking phases. The ML platform is extendable to all ancillary and ticket bundles. It can suggest an optimal mix of products and price points that have the highest propensity to purchase for a given customer and travel itinerary. The pattern recognition module of this ML algorithm can identify the best time to communicate with various passenger segments to improve the purchase conversion.
R K Amit
Shao Hung Goh
Baggage Pricing Model: An Ancillary Revenue Opportunity
Increasing ancillary revenue has become a significant contributor to airline profits in addition to supporting airline management to compensate for high operations costs. Further, offering ancillaries allows airlines to keep their ticket prices low while giving customers a personalized experience. Among all sources of ancillary revenue, baggage fee is identified as an essential component since this alone contributes 3.2% of the global airline revenue as per CarTrawler Global Estimate of Baggage Fee Revenue report-2018. However, there is no consensus in setting the baggage prices across the industry. A customer's excess baggage is charged based on either the number of pieces or the weight and is sometimes route-specific. This research considers the baggage linear-area and weight/linear area as an input to the model to generate recommended prices for extra pieces. We adopt a linear programming approach to maximize the piece-based revenue, while maintaining a revenue-equivalence between weight and piece-based pricing. The revenue-equivalence between weight and piece-based pricing will ensure a fairness on suggested prices (for baggage) and avoid any revenue dilution possibilities from baggage requests. The constraint of the model preserves the industry practice of imposing a higher penalty for each extra piece of baggage. Results depict that our model prices are generally aligned with current industry pricing policies, with a higher revenue reported from the model. This method simplifies the complex structure of baggage pricing and proposes a pricing alternative for the industry.