# Machine Learning based Framework for Robust Price-Sensitivity Estimation with Application to Airline Pricing

**Authors:** Ravi Kumar, Shahin Boluki, Karl Isler, Jonas Rauch, Darius Walczak  
**arXiv:** [2205.01875](https://arxiv.org/abs/2205.01875)  
**Date:** 2022-05-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.01875v1/

## Key points

- The paper proposes a machine learning-based framework for robust price-sensitivity estimation in the context of dynamic pricing of products with feature-dependent price sensitivity.
- A demand model is constructed using a Poisson semi-parametric approach, where the price-related part is parametric and the remaining (nuisance) part is non-parametric and modeled using sophisticated ML techniques.
- To address potential biases in nuisance parameters, a two-stage estimation methodology is proposed that makes use of deep neural networks or other sophisticated ML estimators to construct estimators of observed purchases and prices given feature vectors.
- In the second stage, a Bayesian dynamic generalized linear model is leveraged to estimate the price-sensitivity parameters using these estimators.
- The proposed framework is tested on simulated and real sales transaction data from the airline industry, demonstrating its superiority over direct one-stage approaches in terms of accuracy of estimates.
- It has several advantages over existing methods such as flexibility and interpretability due to its semi-parametric nature as well as ability to handle complex interactions between features and prices while being robust to biases in nuisance parameters.
- This paper contributes to improving pricing strategies by providing an accurate method for estimating price-sensitivity parameters in dynamic pricing models which could be applied beyond airline pricing to other industries where dynamic pricing with feature-dependent price sensitivity is relevant.

## Summary

The paper proposes a machine learning-based framework for robust price-sensitivity estimation in the context of dynamic pricing of products with feature-dependent price sensitivity. The authors construct a demand model using a Poisson semi-parametric approach, where the price-related part is parametric and the remaining (nuisance) part is non-parametric and modeled using sophisticated ML techniques. To address potential biases in nuisance parameters, they propose a two-stage estimation methodology that makes use of deep neural networks or other sophisticated ML estimators to construct estimators of observed purchases and prices given feature vectors. In the second stage, they leverage a Bayesian dynamic generalized linear model to estimate the price-sensitivity parameters using these estimators. The proposed framework is tested on simulated and real sales transaction data from the airline industry, demonstrating its superiority over direct one-stage approaches in terms of accuracy of estimates. Additionally, it has several advantages over existing methods such as flexibility and interpretability due to its semi-parametric nature as well as ability to handle complex interactions between features and prices while being robust to biases in nuisance parameters. This paper contributes to improving pricing strategies by providing an accurate method for estimating price-sensitivity parameters in dynamic pricing models which could be applied beyond airline pricing to other industries where dynamic pricing with feature-dependent price sensitivity is relevant.