| Presenter |
Title |
Affiliation |
Abstract |
|
Houman Goudarzi
|
Leveraging exogenous signals as leading indicators for unconstrained flight demand prediction |
ZYTLYN Technologies |
Traditional flight demand models rely heavily on booking data to forecast demand. However, in an era of rapid market shifts and evolving consumer behavior, true "unconstrained" demand—unfolds the latent interest that exists before capacity limits or pricing hurdles are applied, making unconstrained demand signals powerful, e.g. for price elasticity models. Houman Goudarzi will explore various aspects related to unconstrained demand signals, and furthermore the use of exogenous signals will be expanded on—such as events, weather, exchange rates, as second degree leading indicators used as features in predictive models of unconstrained demand. Houman will present a case study focused on the methodology of denoising unconstrained flight demand through cleaning but also through calibration against exogenous signals that have a causational relationship with demand. Concluding with an example of joint analytics of unconstrained vs. constrained flight signals. |
|
Muge Tekin
Kalyan Talluri
|
Estimation using marginal competitor sales information |
Turkish Technology |
An abiding challenge for firms is understanding how customers value their product relative to competitors. This is hard to quantify because, while prices are public, rival sales are not. In industries like aviation and hospitality, aggregated competitor sales can be obtained from third-party brokers, yet this data is rarely used in revenue management due to a lack of suitable models. We develop a constrained maximum likelihood method to address key challenges: (i) competitor data is aggregated across multiple lengths-of-stay; (ii) the no-purchase segment is unobservable; (iii) private group sales reduce competitor capacity and affect prices; and (iv) the partial-information likelihood is intractable. Monte Carlo simulations show our method recovers true parameters and outperforms existing approaches on real booking data. |
|
Jakub Figura
Stanislaw Robak
Maria Browarska
Maciej Pawelczyk
Hadar Sharvit
|
Quantifying External Shocks in Airline Demand: A Multi-Agent LLM Approach to Automated Event Intelligence |
Fetcherr |
Incorporating rare external events into airline demand forecasting remains a challenge. Manual curation and third-party feeds fail to quantify events at scale, lacking the responsiveness to address emerging shocks. We propose a multi-agent LLM-based pipeline to automate event discovery and quantification. Our framework extracts normalized attributes, including location, time, and severity, via web-grounded retrieval and evidence-backed validation. By converting complex global events into structured records, the system maintains model alignment with real-world conditions. Validated across historical scenarios, it demonstrates significant accuracy gains over baselines reliant on endogenous signals. Automating discovery improves disruption performance, scales efficiently, and supports robust forecasting, revenue management, and network planning under planned and unexpected shocks. |
|
Marc Nientker
|
Willingness-to-pay: A modern causal inference approach |
ADC |
The shift from class-based to continuous, dynamic pricing demands more accurate willingness-to-pay (WTP) estimates, yet standard methods yield biased results when pricing decisions are correlated with unobserved demand drivers. Addressing this endogeneity problem requires causal inference techniques capable of controlling for confounders that are neither directly observable nor fully understood. We introduce a modern econometric framework for WTP estimation that builds on recent advances in Poisson regression with direct and interactive fixed effects. By capturing confounders' influence across the full itinerary context rather than isolating individual variables, the method enables robust estimation without requiring prior knowledge of each confounder's exact nature. We present the theoretical foundations, empirical validation on real airline data, and implications for revenue management practice. |
|
Yacine Nabet
Mathias Lecuyer
Elie Lelouche
|
A Test-Set Paradigm for Elasticity Evaluation in Time-Series Revenue Management Models |
Wiremind |
The test-set paradigm, which makes data-driven decisions—from architecture to model selection—on held-out data, is key to deep learning’s success. Price optimization though requires predicting the unobservable demand’s elasticity to price: how do we then apply the test-set paradigm? Existing evaluations fall short: simulations are not data specific; A/B tests are only safe for already good models; and minimizing the deviation between optimal and observed prices reduces to predicting past observed prices. We propose two new approaches to evaluate a model’s elasticity on held-out data. First, we leverage orthogonal learning to assess a model’s performance under small deviations from typical prices. Second, we use Regression Discontinuity Designs to estimate the effect of observed price changes in a test set, to evaluate a model’s elasticity under a different assumption (continuity of potential outcomes) than that (observed confounders) of typical elastic models. |
|
Rutger Lit
Sebastian Andres Orellana Montini
|
A scalable blueprint for airline revenue management experiments: switchback design for seat pricing tests |
ADC
LATAM Airlines
|
We present a large-scale application of switchback experiments for seat ancillary pricing, drawing on joint work with LATAM Airlines using daily route-level data. We examine switchback designs from three perspectives: theoretical intuition, simulation studies calibrated to airline operations, and evidence from historical transactions. A two-way fixed effects framework explains why repeated within-route switching increases effective sample size and dampens demand shocks, confirmed through placebo experiments and power analyses on LATAM's production data. Across all three perspectives, switchback designs consistently outperform traditional fixed-route experiments, reducing standard errors by 30–70 percent while maintaining reliable inference. This allows revenue impact to be measured within weeks rather than months, offering a practical and scalable experimentation blueprint for airline revenue management teams. |
|
Laurie Garrow
Antonio Ramirez
|
Estimating Airline Price Sensitivity from Choice Set Data Using Multinomial Logit Models |
Georgia Institute of Technology |
We estimate price sensitivity using multinomial logit (MNL) models applied to PassengerSim choice set data. Choice sets reflect realistic, airline-specific offerings and include the airline’s own fares, the lowest competitor fare, and a no-fly option, preserving key tradeoffs and mimicking choice-based sampling. Estimated price coefficients are negative and decline in magnitude as departure approaches, indicating lower price sensitivity closer to departure. Results are robust to alternative definitions of early and late booking windows but may vary with additional trip characteristics. We use the MNL coefficients to compute price elasticities and translate them into FRAT5 curves. The resulting sensitivities are airline- and market-specific. We are evaluating the revenue impacts of using user-input versus estimated FRAT5 values. |
|
Pedro Sfriso
Marcio Rubio Martins Zacheo
|
From Forecasting Demand to Managing Booking Pressure |
AE Studio |
While traditional revenue management models forecast unconstrained demand effectively, they do not always explicitly capture the pressure that builds along the booking curve. We introduce the Demand Pressure Index, a practical way to guide how price and inventory should evolve over time. Instead of looking only at how full a flight is, it also considers how fast bookings are coming in and how close departure is.The framework also helps identify situations where sales begin to plateau after aggressive price moves, signaling that protection may have gone too far and adjustments are needed. Deployed at Azul Airlines as a thin intelligence layer above its RMS, the approach improved sell-out timing and delivered measurable revenue gains without disrupting existing workflows. It does not replace core systems; rather, it detects demand imbalances and converts them into bounded price or inventory adjustments with guardrails and stateful updates to ensure stability. |
|
Laleh Kardar
Ravi Kumar
|
Enhancing Price Elasticity Estimation for Airline Dynamic Pricing via KDE-Based Sampling and Clustering |
PROS |
We address the challenge of estimating price elasticities of passenger demand in airline dynamic pricing when no-purchase data is unavailable. Building on a two-stage Poisson semi-parametric framework that combines machine learning with causal inference techniques, we propose two enhancements to improve robustness in sparse data settings. First, we apply a KDE-based sampling strategy to densify sparse regions of the feature space without imposing strong parametric assumptions. This improves model stability in underrepresented segments. Second, we introduce a top-down clustering approach that replaces rigid partitioning with data-driven groupings that respect operational constraints while capturing behavioral variation across segments. Empirical results on real airline transaction data show that these enhancements lead to more stable and differentiated elasticity estimates, supporting more effective and interpretable pricing decisions. |
|
Daniel Fry
|
Airline Revenue Management and Welfare Economics: A First Look |
Alaska Airlines and Hawaiian Airlines |
In this presentation, I will explore some past research applying standard welfare economics tools to analyze the welfare impact of revenue management as applied in the airline industry. I will then begin to explore how the modeled representation of revenue management can be enriched and how standard welfare economics tools might be altered to improve the overall welfare model. |
|
Indrajit Chatterjee
Qun (Russel) Sui
Anubhav Jain
Joseph Mathews
|
DNA of Demand: Unlocking Travel Intent for Better Demand Forecasts |
United Airlines |
Understanding travel intent is critical for optimizing revenue management, as it directly dictates market seasonality and price elasticity. Business travelers typically exhibit lower price sensitivity and distinct seasonal patterns compared to Leisure passengers. Furthermore, within the leisure segment, passengers traveling for Vacation versus those Visiting Friends and Relatives (VFR) may also display divergent behaviors and sensitivities. In this presentation, we demonstrate how Machine Learning models can leverage trip attributes (without incorporating any personal information) to uncover these underlying travel intents. By identifying the intent behind each trip, we can establish a more robust foundation for demand forecasting and revenue management decisions. |
|
Aldair Alvarez Diaz
Philippe Gendreau
Dounia Lakhmiri
Tu-San Pham
Yossiri Adulyasak
Jean-François Cordeau
|
Reducing seat spoilage through no-show prediction and risk-aware optimization |
Ivado Labs |
Underutilization of seat capacity due to passenger no-shows remains a critical challenge for airline profitability. In this talk, we present a solution integrating machine learning and optimization developed for a major North American leisure carrier to mitigate seat spoilage through dynamic sellable capacity adjustments. The solution leverages a multi-level predictive architecture to translate granular booking data into robust flight-level no-show distributions. This stochastic input is then fed into a risk-aware optimization module that balances the trade-off between marginal revenue from additional seats sold and operational costs of denied boardings. The solution provides Revenue Management analysts with automated recommendations and transparency into risk metrics. This end-to-end solution has been successfully validated in production, demonstrating a significant reduction in unutilized capacity and a measurable uplift in both load factor and revenue. |
|
John Elder
John Bruer
James Graham
Randi Griffin
|
Calibrating RM Systems: A framework and practical lessons |
Boston Consulting Group |
Airline RM systems have become increasingly complex and automated, requiring real-world observation and calibration to sustain performance. We present a framework that evaluates pricing actions at sufficient granularity to enable rapid, targeted tuning during in-market tests. Implemented price changes are matched to unactioned controls via stratified sampling across decision-making drivers. Post-intervention uplifts in revenue, yield, and bookings are measured, with load forecasts incorporated to assess dilution and displacement risk. Pockets of underperformance are isolated, diagnosed, and translated into concrete adjustments. Applied at scale, the approach has delivered clear RASK improvements across multiple RM enhancement initiatives. We also discuss practical considerations: embedding calibration within RM organizations and leveraging automation, including AI-assisted pattern detection, to surface performance gaps and suggest specific system refinements. |
|
Arpit Ganeriwal
|
Leveraging Internal Ground-Truth to Classify Competitive Business and Leisure Demand |
United Airlines |
This study presents a robust methodology for classifying airline PNRs into business and leisure segments using machine learning. By leveraging internal "ground truth" (loyalty data and survey responses), we developed a high-accuracy XGBoost model using booking, customer, and destination attributes. This logic was then "mirrored" onto industry ticketing data (DDS), enabling granular classification of competitor traffic (AA/DL). Our validation confirms that the industry model's results for United align closely with internal data, proving its reliability for competitive benchmarking. The model is now a core commercial tracking tool, even inspiring enhancements in RM demand forecasting systems. The next phase explores sub-segmenting leisure into "Vacation" vs. "VFR" to quantify distinct price elasticities—a critical step for refining RM strategies in an evolving "bleisure" market. |