Evaluating machine learning techniques for predicting tourist occupancy: an experiment with pre- and post-pandemic COVID-19 data

Abstract

This paper analyses the prediction capacity of machine learning techniques under severe demand shocks. Specifically, three methods – Naive Bayes, Random Forest and Support Vector Machine – are tested in predicting rental occupancy for tourist accommodation in the city of Madrid. We compare two different scenarios: firstly, the predictive capacity in the years prior to COVID-19 and, secondly, the ability to anticipate demand behaviour once the pandemic started. The results demonstrate first that without market disturbances, the Random Forest model exhibits the best predictive capability. Second, the COVID-19 pandemic caused such major changes that none of the three tested models are entirely reliable, although the Random Forest and Naive Bayes models outperform the SVM model. As a methodological novelty, this paper includes occupancy quantiles to resolve problems with available data and temporal biases.