| Presenter |
Title |
Institution |
Abstract |
| Jonas Rauch |
A Practical Perspective on “Disentangling Capacity Control from Price Optimization"
|
PROS |
In 2017, we (J. Rauch, K. Isler, S. Pölt) published a paper titled “Disentangling Capacity Control from Price Optimization,” demonstrating that Revenue Management (RM) can be mathematically divided into two distinct problems: pricing and capacity control. We showed that it is theoretically possible to solve the capacity control problem without extensive knowledge of the pricing mechanism or detailed price data. However, our technical approach in the paper and conference presentations did not effectively argue why one would want to adopt this method. In this presentation, we revisit this question from a practical and intuitive standpoint, highlighting the circumstances under which a disentangled approach is advantageous. |
|
Kalyan Talluri
Hanzhao Wang
Xiaocheng Li
|
Gen AI architectures for RM estimation |
Imperial College London Business School |
We outline our experience with using GenAI models to capture a number of sequential decision-making tasks such as dynamic pricing, inventory management, resource allocation, and queueing control. Under this framework, all these tasks can be viewed as a sequential prediction task where the goal is to predict the optimal future action given all the history information. When adequately trained this approach can offer distinct advantages over any existing models. We present our experiments and experience specifically on dynamic pricing and RM estimation. |
| Clay Youngblood |
Outside-of-RMS Machine Learning: Boost RM Decision Making by Thinking Outside the Box |
Southwest Airlines |
With the rise of open-source machine learning, airlines have greater opportunities to build in-house models that can complement or extend their revenue management systems. Such models can accurately predict important KPIs that facilitate smarter decision making. However, the booking data that practitioners will use to build said models is ripe with bias. In this presentation, we explore the consequences of constructing a ML model that learns a spurious correlation between bid prices and flown load factors – two variables that are correlated, but ultimately influenced by broader “demand” factors. We illustrate problems with the constructed model and propose, as a solution, a method of omitting inventory metrics all together to avoid the creation of the problematic relationships. Therefore, creating models only guided by a phenomenon we call implied revenue management strategy. We then dive into empirical examples of this methodology working at Southwest. |
|
Müge Tekin
Kalyan Talluri
|
Estimation using marginal competitor sales information |
Erasmus University Rotterdam School of Management |
An abiding concern for firms is how customers value their product compared to the competitor’s. Estimating this is challenging as, even though prices are public, competitors’ sales are typically unobservable. However, in hotel industry, marginal aggregated competitor sales data can be obtained through STR reports. Hotels participate by reporting their sales and in turn receive aggregated competitor data across groups and LOS. Such data is rarely used in RM estimation due to lack of robust methodologies. This paper tackles this problem under a market-share model for hotels, addressing key challenges: (i) competitor data is aggregated across LOS with distinct demands, (ii) no-purchasers are unobserved, and (iii) competitor group sales and capacity remain private. Using Monte Carlo simulations, our method recovers true parameters from synthetic data. We then apply it to real-bookings. Our method surpass alternate estimation methods from NT and RM literature. |
| Saar Teboul |
Generative AI & Large Market Models: Transforming Airline Revenue Management through Multi-Objective Optimization |
Fetcherr |
This study leverages generative AI and large market models to revolutionize airline revenue management by applying multi-objective optimization to the complex dynamics of pricing. Focusing on the balance between revenue, load factor, and competitive positioning, we introduce a framework to showcase the trade-offs between these key performance indicators, we demonstrate how airlines can utilize advanced AI-driven models to adjust fares strategically, optimizing seat occupancy, profitability, and market competitiveness in an ever-evolving landscape. |
|
Charles Pierre
Mathias Lecuyer
|
Training and Evaluating Causal Forecasting Models for Revenue Management Time-Series |
WIREMIND |
We leverage deep learning time-series models to make forecasts that inform downstream decisions. Since these decisions can differ from those in the training set, there is an implicit requirement that time-series models will generalize outside of their training distribution. In reality this requirement is really hard to attain as the historical data only contain a very small subset of all possible prices. Thus naive model will have difficulties infering price elasticity and generalized outside the historical distribution of price. We extend orthogonal statistical learning framework to train causal time-series models that generalize better when forecasting the effect of actions outside of their training distribution. We leverage Regression Discontinuity Designs popular in economics to construct a test set of causal treatment effects. |
| Gopal Ranganathan |
A cutting-edge AI design using a Quantitative Large Learning Model Q-LLM) for Enterprise Revenue Management |
QuadOptima |
A cutting-edge AI design using a Quantitative Large Learning Model Q-LLM) for Enterprise Level Revenue Management is presented, to address emerging challenges like NDC protocols. These challenges require classless retailing, bundling and continuous dynamic pricing. The design advances RM by blending the analytical approach of operations research with machine learning based transformer approach to handle large multidimensional data. The transformer features a DNA Analyzer , forecast and optimization engines. It also features a micro-segment data model to facilitate handling the transformer architecture. The design drives better tradeoffs in Global vs Local optima and increases profits and long-term value. An example application of dynamic pricing using this new modern RM design is presented. |
|
Vladimir Antsibor
Marc Nientker
|
Estimating willingness-to-pay under unobserved confounders |
ADC (acmetric) |
The airline industry is shifting from class-based to dynamic pricing, requiring accurate willingness-to-pay (WTP) estimation. This demands causal inference techniques to address endogeneity from unobserved confounders. Traditional methods, such as instrumental variables, rely on strong, non-verifiable assumptions, while newer debiased machine learning approaches attempt to account for all relevant influences. However, real-world complexities make it difficult to anticipate and incorporate all confounders. We introduce a modern econometric method for WTP estimation that models confounders' broader impact on itineraries, enabling robust estimation without prior knowledge of their exact nature. Our validated approach helps airlines refine pricing strategies while addressing real-world complexities. We will discuss the theory, empirical validation, and revenue management implications. |
| Burak Ozdaryal |
Rethinking Price Elasticity: The Competitive Dimension |
Sabre |
Traditional approaches to price elasticity of demand in airline revenue management often operate under the assumption of independence of demand across different offerings. This presentation challenges this assumption by examining the impact of competitive product offerings. We introduce a slight modification to the traditional q-forecasting framework that incorporates these competitive effects to illustrate the necessity to re-evaluate traditional pricing strategies and marginal revenue transformation. Our analysis suggests that neglecting the competitive landscape can lead to suboptimal pricing decisions. This talk will explore the limitations of conventional models to stimulate thought on developing better revenue management techniques for today's dynamic marketplace. |
|
Richard Ratliff
Helder Inacio
Xiaoyun Niu
Keji Wei
|
A Stochastic SBLP Optimizer for Network RM with Dependent Demands |
Sabre
|
We will present a new model innovation for solving stochastic, class-based, airline network revenue management problems under dependent demands in a computationally efficient manner that optimizes item allocations. Our approach uses simulation-based, linear-programming optimization (with dependent demand, spill, and recapture effects) based on sample average approximation (SAA). Our overview in this presentation includes the model formulation along with applied examples and computational results. |
| David Foster |
A survey of cargo revenue management problems |
Delta Air Lines
|
Cargo Revenue Management differs in substantial ways from Passenger RM, however it is poorly understood and studied. This paper illustrates some of the differences between the two areas and proposes several avenues of future investigation. |
|
Stanisław Robak
Maciej Pawełczyk
Konrad Kubzdela
Maria Browarska
|
Beyond Traditional Forecasting: Uncovering New Demand Signals for Airline RM |
Fetcherr
|
Airline revenue management teams traditionally forecast demand using historical booking data and static seasonal assumptions. However, research indicates that companies typically leverage only about 40% of available data, leaving critical market intelligence—such as competitor actions, local events, weather changes, and macroeconomic shifts—largely untapped. This data gap means traditional forecasting models frequently miss sudden or unconventional demand fluctuations, leading to missed revenue opportunities. We address this challenge by integrating external market data into our Large Market Model. By systematically analyzing real-time external influences alongside traditional data, our approach identifies demand shifts earlier, enhances forecast accuracy, and supports more proactive revenue management decisions. This presentation explores how expanding the RM data horizon beyond historical patterns significantly improves airline forecasting and profitability. |
| Awais Farrukh Shamim |
Value of Bid Prices: A Monte Carlo Simulation for Cargo Revenue Management |
Qatar Airways
|
Air Cargo Revenue Management concerns with the practice of maximizing revenue generated on Air Cargo Network through capacity management and pricing decisions. While the field of Cargo Revenue Management has a similar objective as the field of Passenger Revenue Management, there are a host of differences between the two fields. Cargo Revenue Management is made complex due to the multi-dimensional nature of shipments, significant proportion of business done through mid-term contracts or allotments, short booking window and volatility in demand. Moreover, the implementation of RM techniques in the domain of Air Cargo are quite limited. In this study, we utilize a Monte Carlo Simulation on a sample network to demonstrate that Cargo Revenue Management Practices such as Bid Prices can add value to carrier’s bottom-lines. We conclude our study by presenting avenues for further research to bring Cargo Revenue Management up to speed with Passenger Revenue Management. |
|
Thomas Fiig
Simon Nanty
|
From Black Box Models to Analytical Models - Using ML to extract Analytical Models for Airline Demand Forecasting |
Amadeus
|
"This paper explores whether machine learning (ML) can extract analytical models (AM) for airline demand forecasting, combining the interpretability of AM with the accuracy of ML. As a test case, we examine the pick-up curve (PUC). We test two ML approaches to generate analytical models. Symbolic regression fits data well but produces overly complex, impractical models. Leveraging the Kolmogorov-Arnold Representation Theorem, we use Kolmogorov-Arnold Networks (KAN) to uncover underlying analytical structures. Our findings show that PUC is well-described by a mixture of Weibull distributions and reveals day-of-week sales variations and distinct business and leisure demand segments. The KAN model requires only six parameters per market—compared to 100 million in large neural networks—while maintaining comparable accuracy. This approach ensures efficiency, interpretability, and robustness, offering a scalable solution for airline revenue management." |
|
Bartłomiej Wójcik
Karolina Macielak
Magdalena Kollar
|
Sales increment evaluation using machine learning models |
LOT Polish Airlines
|
Evaluating sales increment in the airline industry is uniquely complex due to two distinct time dimensions: the date of sale and the date of departure. Promotions may only accelerate purchases rather than increase total demand, but sparse historical data for unique campaigns makes it difficult to quantify true incremental gains. This paper proposes a novel evaluation framework to isolate promotion-driven increments using machine learning models. The method offers airlines a practical tool to optimize promotional ROI while addressing the inherent uncertainty of demand acceleration versus genuine demand creation. |
|
Luciano Zavala
Ana Lucia Carrizo
|
Revenue Management AI Assistant |
Copa Airlines
|
Revenue management (RM) is vital in industries like airlines but generates vast, complex data. This paper presents an RM AI Assistant using a Retrieval-Augmented Generation (RAG) approach to enhance knowledge access. It integrates a large language model (LLM) with an enterprise RM knowledge base, leveraging hierarchical retrieval (RAPTOR method), multi-query generation, and reciprocal rank fusion for accuracy. The system, built on a cloud platform with open source tools, supports text2SQL for structured data queries. Initial results show efficient, precise information retrieval, improving decision-making. We discuss challenges like scalability and data quality, along with future enhancements. |
|
Domenico Fabrizi
Adam Thorsteinson
Daniel Fry
|
The Value of Information in Airline Overbooking |
Alaska Air Group
|
This paper demonstrates that leveraging traveler-level data significantly enhances the prediction of cancellations and no-shows, a core function of overbooking. By improving the quality of revenue management (RM) decisions in overbooking, value is created for both the airline and the consumer through a reduction in excessive overbooking and excessive seat spoilage (seats for which there was demand but which flew empty). First, using proprietary data from Alaska Airlines, we show that booking-level characteristics improve the accuracy of cancellation and no-show forecasts. Second, we evaluate the impact of integrating these predictions into RM optimization of both pricing and virtual capacity via counterfactual simulations. We illustrate how airlines can adjust overbooking and pricing strategies to maximize revenues while minimizing denied boarding utilizing additional information. |
| Clement Zhang |
Enhancing Airline Revenue Management Intelligence through Large Language Models and Machine Learning |
FlightBI LLC
|
This study explores the integration of large language models (LLMs) and machine learning techniques into aviation business intelligence to support airline revenue management. By combining natural language interfaces with large-scale proprietary airline data—including external market sources such as schedules, ticketing, and market size, and internal data such as bookings, fares, and inventory—the project examines how LLMs can generate SQL queries and return interpretable insights with automated visualizations in response to unstructured user input. A machine learning component enhances demand forecasting at the route level, particularly in volatile markets. The research assesses the feasibility and effectiveness of this hybrid approach in delivering timely, accurate insights for RM professionals, with the goal of streamlining analysis and reducing reliance on technical querying and manual charting. |
| Tiziano Parriani |
An Application of Swarm Optimization to Market Adaptive Dynamic Pricing Problem |
Sabre Corporation
|
Market Adaptive Pricing (MAP) in airline revenue management utilizes same-day shopping session data to optimize fare positioning. We propose a heuristic approach to MAP based on Particle Swarm Optimization (PSO). Given a customer choice model that associates choice probabilities with combinations of displayed fares, we define a constrained fare proposition for all itineraries of the host airline, maximizing a predefined revenue function. The proposed approach accommodates highly nonlinear objective functions typical of customer choice models. Preliminary results from 229 realistic sessions demonstrate significant gains compared to traditional approaches, though at the expense of increased computational cost. As a next step, ad hoc modifications to the general-purpose PSO algorithm may substantially reduce computational costs and/or enhance result quality. |
|
Thomas Fiig
Mike Wittman
|
Will RMS survive the Offer/Order transformation? |
Amadeus
|
For over 40 years, Revenue Management Systems (RMS) have been the foundation of how airlines price their products. However, the Offer/Order transformation will enable airlines to perform more real-time, session-specific optimization through Dynamic Pricing (DP). As offers continue to become more dynamic in both product and price, will RMS even survive in this new world? If so, what role should it play? We discuss different possibilities for how RMS might interact with DP, from a minimalistic RMS that computes only the bid price to an all-seeing RMS that is responsible for pricing all products. We argue that the best design lies somewhere in the middle, where RMS computes a continuous reference price for each product that is further refined by DP. This configuration allows DP to maintain its flexibility while still ensuring price and product consistency aligned with the airline’s strategy. |
| Darius Walczak |
Revenue and Load Factor Trade-off in Revenue Management |
PROS Inc.
|
The primary objective of revenue management (RM) is to maximize expected revenue and sophisticated software systems have been deployed to achieve it. In practice though, users of RM systems need to know what expected bookings and thus load factors are achievable under those controls. These key business metrics rival revenue in importance (e.g., for financial reporting) but have historically received less attention from the modeling perspective. We focus on the single-leg dynamic setting for which there are results available to generate the so-called efficient frontier between revenue and expected bookings/load factor. The efficient frontier makes the tradeoff between the two key RM metrics precise and interpretable. We also extend our analysis to RM models with price-sensitive demand by broadening the concept of the fare class efficiency (‘concave envelope’) and the associated transformed fares/fare adjustments to account for the load factor objective. |
| Emanuele Concas |
Pricing Bundles for Airline Revenue Management |
ENPC
|
Airlines have transformed into retailers, selling not only tickets but also ancillaries and bundles, which are essential for revenue. The next frontier is to prioritize personalized experiences to increase customer satisfaction and better capture their willingness to pay. The IATA’s New Distribution Capability increases flexibility, enabling airlines to optimize profits while offering tailored choices. However, revenue management still relies on static pricing rules that overlook key factors like the substitution effect. Ancillary pricing is often determined without considering the flight context. Airlines need more advanced customer choice models, particularly for joint pricing and assortment optimization. Academic research has largely neglected these developments, with limited studies on bundles and ancillaries. This presentation addresses these gaps by developing practical models that integrate new capabilities, enhancing revenue management strategies. |
|
Mikhail Andriyanov
Damien Lopez
Andrew Nestor
Sofoklis Kyriakopolous
Konstantin Prokopchik
|
Comparative Performance of Reinforcement Learning and Large Language Models for Dynamic Pricing |
Andriyanov & Partners Mathematicians and Economists PartG, Decision Lab Ltd
|
We compare the performance of reinforcement learning (RL) and large language model (LLM) agents to make dynamic pricing decisions in an airline pricing simulator. Agents sell a limited number of seats over a finite booking horizon with increasing demand and decreasing price sensitivity as the flight day approaches. A genetic algorithm (GA) is used to derive the pricing strategy with baseline performance. Both RL and LLM agents converged to intuitive price curves with results comparable to GA. RL converged with Proximal Policy Optimization (PPO) after failing with simpler methods. The LLM agent was built on a low-code basis by using a commercially available network. Only a few prompted episodes were required to achieve a result close to RL. This makes LLM agents promising for complex dynamic optimization tasks. Minimal ML expertise is enough to quick start the approach, with more research needed to explore the complexity of fine-tuning for stable production use. |
|
Ravina More
Manuj Kumar Jain
|
Machine Learning for Funnel-Based Conversion Prediction and Offer Targeting |
Air India
|
We present a novel machine learning framework for optimizing airline website offers. By analyzing customer behavior in the booking funnel, our dual-model approach first predicts conversion probability, then determines optimal incentives. The system strategically withholds offers from likely bookers while providing appropriate incentives to hesitant customers. Using features from time-spent metrics, navigation patterns, and customer attributes, our adaptive ML models predict booking completion with high accuracy. This approach increases conversion rates while reducing overall discounting compared to traditional methods. The framework continuously refines through a feedback loop capturing seasonal and market changes. Our approach demonstrates how behavioral analytics can create economically efficient offer strategies that maximize revenue by deploying incentives only when necessary and proportional to conversion probability. |
|
Tino Quadranti
Mathias Frank
|
Scenario Based Automated Capacity Control |
Swiss International Airlines AG
|
This talk presents the concept and simulation results of an automated capacity control system (RM Autopilot) at LHG. The traditional bid price control is created by using the most likely demand scenario received from the forecaster. Whereas these bid price vectors do a good job in reacting to the stochastic nature of demand, they struggle when the forecast is off. RM Autopilot would create an alternate Bid Price vector, calculated by incorporating different plausible demand scenarios based on the forecast. For each scenario autopilot assesses the probability of it to be true, and the final bid price vector is a weighted average of each of the scenario bid price vectors. The current state of a flight (remaining time & seat index) is used to update the probabilities of the different scenarios. As such, RM Autopilot capacity control can adapt automatically to the stochastic nature of demand as well as adapting the forecast assumption based on booking income. |
|
Laurie Garrow
Jeffrey Newman
Alan Walker
Maitreyee Talnikar
|
Comparing Q-Forecasting Methods and Leveraging ML to Enhance RM Performance |
Georgia Institute of Technology
|
As part of ongoing activities of ATL@GT, we are developing PassengerSim, a competitive RM simulator to evaluate different RM strategies. Within PassengerSim, we have implemented two versions of Q-forecasting: a PODS-based approach and a conditional version that accounts for product availability. In this presentation, we will show conditions under which the two methods are similar, and explain the crucial role weighting plays in an underlying linear regression formulation. Results based on simulated networks highlight differences between the two methods, and opportunities for future research that will likely be of practical importance to airlines. We complete the presentation by highlighting how we are using ML methods to analyze simulation results and identify opportunities to improve RM performance through better understanding the relationship between forecast accuracy and RM outcomes. |