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

ABSTRACTThis 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.KEYWORDS: Tourist occupancyAirbnbpredictiontourist demandmachine learning Disclosure statementNo potential conflict of interest was reported by the author(s).Additional informationFundingThis study has been carried out in the framework of the research project ‘Digital Transition and Innovation in the Labour Market and Mature Sectors. Taking Advantage of AI and Platform Economy (DILATO)’, funded by the Spanish Ministry of Science and Innovation as a 2021 Green and Digital Transition Project, with reference [grant number TED2021-129600A-I00].