Combining Economic and Search-Request Variables to Predict Local Airline Market Shares: A Comparison of Forecasting Methods

Abstract

Our aim in this article is to predict local airline market shares in the United States at the company level combining traditional economic indicators (at the national and local levels) with Google search engine requests. We resort both to simple econometric models and to more sophisticated machine learning tools (random forests, neural networks and support vector machines) and compare their respective predictive power. Using data from the American market for the period 2004-2018, our study bears three key findings. First, we highlight the usefulness of combining search-engine requests with other traditional economic indicators as explanatory variables to predict local airline market shares. Second, the comparison of the different forecasting techniques reveals that tree methods consistently outperform the alternative forecasting tools. Third, in line with the growing literature dedicated to frugal forecasting, we show that no advanced model is able to beat our heuristic benchmark, which consists in rolling increments of annual variations, such that variations in market shares are best predicted by past variations.