The Impact of COVID-19 on Airfares—A Machine Learning Counterfactual Analysis

Florian Wozny

doi:10.3390/econometrics10010008

Abstract

This paper studies the performance of machine learning predictions for the counterfactual analysis of air transport. It is motivated by the dynamic and universally regulated international air transport market, where ex post policy evaluations usually lack counterfactual control scenarios. As an empirical example, this paper studies the impact of the COVID-19 pandemic on airfares in 2020 as the difference between predicted and actual airfares. Airfares are important from a policy makers’ perspective, as air transport is crucial for mobility. From a methodological point of view, airfares are also of particular interest given their dynamic character, which makes them challenging for prediction. This paper adopts a novel multi-step prediction technique with walk-forward validation to increase the transparency of the model’s predictive quality. For the analysis, the universe of worldwide airline bookings is combined with detailed airline information. The results show that machine learning with walk-forward validation is powerful for the counterfactual analysis of airfares.

Highlights

The International Civil Aviation Organization (ICAO) establishes universal aviation standards for the international airline industry
The results show that machine learning with walk-forward validation is powerful for the counterfactual analysis of airfares
ICAO Data Plus is high-dimensional with heterogeneous features

Summary

Introduction

The International Civil Aviation Organization (ICAO) establishes universal aviation standards for the international airline industry. This paper develops a universal prediction model with as few assumptions about the structure of the data as possible, to ensure usability for different reform scenarios. The prediction error of the ML model was neglectable over the whole observation period, revealing its suitability for predicting counterfactual control scenarios for the evaluation of universal policies. The other strand of the literature uses ML predictions as counterfactual control scenarios to evaluate universal reforms (Abrell et al 2019; Burlig et al 2020). This paper’s further contribution is a multi-step forecasting approach with walkforward validation that has not been used in the corresponding literature so far This procedure is of importance by revealing that the deviation between post-reform observed and predicted values is not caused by a measurement error.

The Universal COVID-19 Policy Response

Airfares and Airline Information

Imputation and Feature Selection

Counterfactual Prediction and Walk-Forward Validation

Results

Conclusions