Log in

Abstract submissions for this year's conference are now closed.

Contact(s) Title Institution Abstract
Tiago Gonçalves

Bernardo Almada-Lobo
Enhancing Robustness to Forecast Errors in Availability Control for Airline Revenue Management Lufthansa Traditional RM systems are built under the assumption of independent demand per fare. The fare adjustment theory is a methodology to adjust fares that allows for the continued use of optimization algorithms and seat inventory control methods, even with the shift towards dependent demand. Since accurate demand forecasts are a key input to this methodology, it is reasonable to assume that for a scenario with uncertainties it may deliver suboptimal performance. This study demonstrates, firstly, the theoretical dominance of the fare adjustment theory under perfect conditions. Secondly, it lacks robustness to forecast errors. A Monte Carlo simulation indicates that a forecast error of 20% can prompt a necessity to adjust the margin employed in the fare adjustment theory by -10%. Moreover, a tree-based machine learning model highlights the forecast error as the predominant factor, with bias playing an even more pivotal role than variance.
Aviv Cohen

Uri Yerushalmi 
Generative AI & LMM: Transforming Airline RM Fetcherr This presentation explores the application of Large Market Models (LMM) informed by Generative AI within the sphere of airline revenue management. It focuses on advanced market dynamics prediction techniques powered by Generative AI, forming the cornerstone of LMM’s superior forecasting capabilities for passenger demand and competitive actions. Furthermore, the discourse elaborates on the strategic advantages of AI-enhanced pricing and inventory control, showcasing how nuanced seat allocation and pricing strategies contribute to revenue maximization for a broad spectrum of airlines, ranging from traditional bucket-based carriers to those employing dynamic pricing models. Through real-world case studies, the presentation highlights the tangible benefits of LMM in revenue management emphasizing both direct and indirect value creation via efficiency improvements and revenue growth.
Christian Nauerz

Paolo Gorgi
Enhancing Airline RM Measurement Capabilities Through Rigorous and Disciplined Experimentation  acmetric This presentation introduces a novel method for analyzing the impact of price and capacity changes on airline revenue, moving beyond basic historical comparisons. We emphasize the importance of experimentation in revenue management, critiquing simple approaches and advocating for controlled experiments for their reliability and reduced variability. We critique traditional A/B testing for its flawed IID assumption, proposing stratification and CUPED to refine experiments and improve accuracy. These techniques address data correlation, enhancing analysis speed and reducing pre-treatment biases. Additionally, we'll guide on designing experiments, covering sample size determination and data correlation adjustments to enhance experiments power. Our method offers deeper insights into pricing and capacity strategies, making statistical concepts actionable for airline revenue management.
Prabhupad Bharadwaj

RK Amit

Atul Malik
Baggage Allowance Optimization: Balancing Passenger Demand and Fare Policies in Airline Strategies  IIT Madras Baggage allowance is an important decision for airlines. Although this plays a crucial role in passenger demand, the literature does not capture the optimal baggage allowance decision and its effect on passenger demand and ticket prices. Employing a causal loop diagram, we capture interactions and effects among model components, establishing links between baggage allowance, baggage, and passenger demand. Utilizing a pooled demand effect, we develop a mathematical framework considering the stochastic nature of baggage demand and demand for passengers induced by the allowance decision through first-order stochastic dominance. Considering a price response curve for the expected average market price, the solution to the model determines both the average price and the optimal baggage allowance for airlines. The extension of the model generates ticket fares and allowances for two different cabin classes, suggesting higher allowances for the higher-paying class.
 Soheil Sibdari Optimizing Airline Operations through Integrated Flight Scheduling and Revenue Management University of Massachusetts, Dartmouth This study examines the implementation of dynamic flight scheduling and aircraft deployment in response to demand and price variation among different flights. We address the importance of adjustments in flight schedules and aircraft deployment based on fluctuating demand, seat occupancy, and pricing strategies. We use a mixed integer programming to determine optimal flight schedules and fleet allocation, taking into account demand elasticity and operational constraints. We use a simulation model to illustrate our results.

Laurie Garrow

Calibration of networks for competitive RM simulation tools  Georgia Tech As part of ongoing activities of ATL@GT, we are developing PassengerSim, a competitive RM simulator that we are developing to test different RM strategies. Competitive RM simulation tools such as PassengerSim require multiple inputs. One of the key inputs is a network that has been calibrated to reflect user-defined metrics such as load factor distributions, fare ratios and fare class mix, percent of local and connecting passengers, percent of nonstop and connecting ODs, competition structure (e.g., number and mix of LCC, hybrid and traditional carriers in an OD), etc. Given the number of metrics, combined with the fact that networks tend to vary geographically across the world, it is desirable to semi-automate the process of calibrating network for competitive RM simulation analysis. This presentation focuses on a process we are developing using a U.S.-based market as a case study.

 Kishor Jha

 Ashutosh Rawat

Transforming Revenue Management: Air India's Journey with Data Science and Advanced Analytics  Air India This presentation sheds light on the evolution of Air India's revenue management strategies, achieved through the strategic adoption of data science and advanced analytics. Having developed a suite of dashboards to monitor all RM KPIs, Air India in a short span of time have also built a proprietary PAX and Revenue Forecasting algorithms and implemented ML based continuous flight monitoring system. This presentation also showcases how Air India is leveraging generative AI to propel the analytics at RM. We also present the challenges faced by Air India in its quest for data-driven RM excellence, and highlight the strategies employed to overcome these hurdles. By showcasing Air India's remarkable success story, this presentation seeks to inspire and empower industry peers to embrace data science and advanced analytics as powerful drivers of transformation, fostering a new era of innovation and excellence in revenue management in the aviation industry.
 Ezgi Eren An Adaptive Data-Driven Approach to Air Cargo Revenue Management  PROS It is a well-recognized fact that air cargo revenue management is quite different from its passenger airline counterpart. Inherent demand volatility due to short booking horizon and lumpy shipments, multi-dimensionality and uncertainty of capacity as well as the flexibility in routing are a few of the challenges to be handled in air cargo revenue management. In this talk, we present a machine learning based data-driven revenue management approach which is well-designed to handle the challenges associated with the air cargo industry, primarily the demand volatility. We present multiple frameworks to handle the weight and volume modeling and show that a simple approach of applying the maximum of weight and volume bid prices works best in simulations tailored to air cargo setting.
Alex Matson  When good is good enough: A discrete approach to forecast calibration Amadeus  Conditional-parametric approaches to forecasting have been a boon for robustness and effectiveness of demand forecasting over the last decade. However, in practice, they require significant effort and close attention to calibrate well. This presentation proposes a new approach where each forecast component is selected from a preexisting discrete set of observed behaviors, rather than independently calibrated from data. We demonstrate that this approach performs just as well out-of-sample as fully calibrated parametric models, especially in situations with limited relevant historical data. We also discuss future opportunities available with this approach to enhance the breadth of features included in the forecast and the way RMS users work with forecasting in practice.

Thomas Fiig

Simon Nanty

Application of generative AI for airline demand forecasting  Amadeus At last year's AGIFORS conference, we presented a hybrid demand forecasting approach that blends traditional knowledge-based analytical models with Machine Learning (ML) techniques for improved accuracy. The flexibility of the ML model components unlocks opportunities to refine the demand forecasting model by incorporating previously unexplored data sources. In this presentation, we investigate the application of generative AI to hybrid demand forecasting. Specifically, we construct a dataset comprising of descriptions of the world’s cities from a travel experience perspective using an LLM. These descriptions are used to generate pre-trained embeddings that replace the trained embeddings in the ML component. Remarkably, this approach leads to improved forecast accuracy. We discuss how this is possible and how our findings could unlock other opportunities to apply generative AI in demand forecasting.

Karim Pérez

Balaji Narasimhan

Vilmar de Sousa 

Unlocking Revenue Potential in Complex Networks using Hybrid Forecasting: Collaborative Study with Amtrak

ExPretio Technologies


Rail operators face complex network challenges when optimizing inventory controls, and Amtrak's northeast corridor (NEC) is a prime example. With Acela and Northeast Regional (NER) trains driving over 50% of Amtrak’s revenue, some inventory control challenges arise, particularly for NER trains crossing NEC and state-supported markets. Hybrid forecasting solutions can overcome this challenges by applying choice-based demand models to markets with rich customer behavior data, such NEC markets, and  independent models  to other markets with limited data. This effectively optimizes inventory controls to maximize revenue. This approach extends beyond railroads. Airlines operating multi-leg flights connecting emerging cities to major hubs can benefit. It also addresses the data gap in revenue management, enhancing industry-wide efficiency.
Fedor Nikitin Shadow prices of demand and applications  Finnair  Network revenue management optimization problem with dependent demand can be approximated by linear optimization problem with two types of constraints. One type of constraints are the ones imposed by capacities of legs in airline network. Another type of constraints arise due to limited demand for each O&D pair. Values of dual variables or shadow prices associated with capacities are used e.g. in calculation of bid-price controls. However, to the best of author knowledge shadow prices for demand levels did not get attention neither in the RM research nor in practice. This report fills this gap. First, duality theory of linear optimization problem in the context of RM is discussed. Second, it is demonstrated how to calculate useful characteristic of price-demand curves called NRV (Network Response to Volume change) values. Lastly, possible applications of these values are presented.

Andrei Furtuna

Georgios Sarlas

Ravi Kumar

Shahin Boluki

Darius Walczak

AI-driven Pricing of Airline Seats: Evidence From a Large-scale Online Field Experiment

Lufthansa Group 


Estimating price elasticity is crucial for implementing effective dynamic pricing systems, yet it remains a complex task due to challenges such as data sparsity and price endogeneity. This talk presents a novel modeling paradigm for dynamic pricing of airline seats that integrates modularly in the existing Revenue Management System. The methodology employs a two-stage process combining causal inference and advanced machine learning techniques, such as Deep Neural Networks, to generate robust price elasticity estimates. We evaluated the performance of the approach through a large-scale online field experiment, employing automated A/B testing, to demonstrate its potential for increasing airline revenues. This approach also paves the way for the calculation of tailored offers, by supporting the inclusion of more booking attributes in the optimal price determination.

Akhil Gupta

Naman Shukla

Generating optimal bid prices via reinforcement learning with batch and shape constraints  FLYR Airline revenue management relies on an optimal bid price (BP) policy for revenue uplift. In this study, we propose a novel framework for producing optimal BP policies which jointly learns (1) the optimal next sellable BP via offline reinforcement learning, and (2) the shape of the full BP curve around the next sellable BP via supervised learning. The optimal next sellable BP model utilizes a batch-constraining approach which is an enhancement of Q-learning to mitigate extrapolation error from out-of-sample state-action pairs. The full BP curve model incorporates shape constraints on the neural network's output based on domain-specific knowledge, ensuring that the learned policies adhere to real-world considerations. Our experiments demonstrate the effectiveness of our framework in generating BP policies. Finally, we discuss the challenges associated with our approach, and outline potential solutions, paving the way for future research in this domain.

Abhinav Garg

Naman Shukla

Talal Mufti

Maarten Wormer

Mid-term decision making in airline cargo using machine learning  FLYR Effective mid-term revenue management is crucial for airline cargo carriers to maximize the revenue from capacity allocation and pricing of longer-term allotment contracts. This study proposes a machine learning approach to forecast expected cargo demand in terms of weight, volume, and expected revenue at the origin-destination (OD) level for each week over the length of the upcoming season. Our approach leverages predictive modeling to forecast demand at the granular OD-departure date level, unlike traditional methods that use higher-level aggregations. This framework solves the mid-term cargo forecasting problem as time-series regression and utilizes a mixture of experts (MoE) technique with models ranging from traditional statistical to advanced deep learning to effectively predict the demand. The proposed framework can help carriers optimize allotment contract pricing and capacity allocation, enhancing profitability over the mid-term planning horizon.
 Ben Shaw Load factor optimisation through closed user groups at scale Flyla GmbH Globally more than 20% of plane seats stay empty during take-off. With predictive analytics airlines know on which routes and at what times those seats stay empty. They are maximising yield on those routes so those seats are empty by design. With advanced technology and distribution capabilities however airlines don't need to choose between yield and load factor, they can do both simply by creating a parallel price curve and using distribution and authentication partners to get more cheaply priced seats to opaque booking closed user groups. This talk will focus on the full technology chain to get from empty seat prediction to at scale distribution and real load factor optimisation without compromising pricing.

Julien Bruno

Antoine Winckels

Sara Ghamloush 

Dynamic pricing for airline ancillary products with machine learning  Air France Ancillary products have become an increasingly important source of revenue for airlines. In this context, a pricing strategy that is based upon an accurate customer's willingness to pay estimation, holds great potential for optimizing revenue. In this talk, we dive into the realm of dynamic pricing for ancillaries, with a particular focus on baggage fees. We introduce a two-step approach, with a willingness-to-pay estimation model based on machine learning, followed by a revenue optimization that aims at recommending optimal prices. The model utilizes historical ancillary purchase data to capture customer's behaviors based on their booking characteristics. This approach also unveils new customer segmentation features that can enhance the pricing process, offering valuable insights to ancillary pricing managers. The findings and methodologies open the way for a more refined and data-driven ancillary pricing strategy

Karolina Macielak

Marcin Gorczyca

Marcin Łukaszewski

Patryk Radoń

Daniel Śliwiński

Revenue forecast: machine learning model vs. analytical approach LOT Polish Airlines Airlines typically utilize systems for managing revenue that are based on analytical models. In parallel, an increasing number of AI and ML-based tools and applications—including those for demand management—are being developed for the aviation sector. The purpose of this presentation is to compare the revenue forecast produced by the analytical system with the forecast produced using machine learning techniques. We will examine the benefits and drawbacks of both strategies in addition to the forecast's accuracy. We will also present our suggestions for revenue management applications of the ML forecast.

Cagla (Chala) Keceli

Anubhav Jain

Independent optimization or not  United Airlines  Many airline revenue management systems forecast leg demand and optimize bid prices independently of each other. This is different from how passengers typically decide, who consider the price of all flights before booking their flight. To better align with passenger choice, there is a desire to jointly optimize all flight prices in hope to get better revenue. Before trying that, we investigate the independence assumption through simulations on a small network and examine the interaction of bid prices of different flights under different scenarios.

Richard Ratliff

Brandon Rundquist 

Automated Benefits Estimation for Personalized Recommendations with Experimentation Engines Sabre  This talk describes a benefits estimation methodology for personalized offer recommendations with multi-armed bandit experimentation. It provides automated metrics for ongoing comparison of clickthrough rate performance of offers at three different levels of aggregation (which are commonly seen in practice) including: 1) personalized, segment-specific offers, 2) generally popular offers (non-segment specific) and 3) equally weighted offers (non-segment specific). Such benefits metrics are helpful in answering the business question: “What are the quantifiable benefits of using personalized recommendations for my offer application?” In this talk, actual results of ancillary offers from a leading hotel retail application will be discussed to illustrate the application of these metrics for benefits estimation, for visually tracking recommender model convergence patterns and as a simple alternative to dedicated A/B testing.
Stacey Mumbower  Price Endogeneity in Demand Models for Revenue Management  Embry-Riddle Aeronautical University  In revenue management systems, prices are heavily influenced by demand signals. This creates a common estimation challenge in demand models called price endogeneity. Using an airline dataset to model flight-level daily demand, we demonstrate two modeling approaches where price endogeneity can be corrected with instrumental variables. More specifically, we compare a linear demand modeling methodology (using a two-stage least squares regression) to a discrete choice modeling methodology (using a two-stage control function multinomial logit model) and compare price elasticity estimates from both approaches. We highlight several key differences between the two approaches, discuss challenges that researchers may encounter in model estimation, and summarize diagnostic tests that can be used. We also briefly discuss alternative approaches that have recently been used in the literature and provide ideas for future research directions.
Stefan Druzdzel  Using Simulations to Improve Revenue Management Systems Breeze Airways  In order to correctly handle the concept of spoilage in airline Revenue Management, it can be very helpful to build simulations. In my presentation I use two examples to illustrate this point; a novel approach to pricing connecting flights, and a simulation environment used to design and test methods for setting “allocated units” (availability by fare class / RFD (Reservation Booking Designator). Side note: Rather than presenting truly novel research, I intend for this presentation to be very practical and for other airlines to be able to easily replicate it. I will have code snippets in the presentation (perhaps in an appendix).

Abhijeet Bhardwaj

Anubhav Jain

Cross Elasticity in Airline Revenue Management  United Airlines  Elasticity measures the drop in demand for unit increase in price. As many airline revenue management systems assume demand independence across different itineraries in a market, the elasticity changes on one travel option due to price change of other alternatives is ignored (cross elasticity). In reality, demand change might not just depend on price across alternatives, but also departure time and quality of alternatives. Further, elasticity estimates should also account for passengers’ budget. We discuss various choice models that can help capture the cross elasticity which depends on these factors, especially the departure time and support our discussion with the help of simulation study. Further we also discuss how existing “opt-out” models can be used to model willingness-to-pay in a market with multiple alternatives.

Powered by Wild Apricot Membership Software