Can algorithms reliably predict long-term unemployment in times of crisis? Evidence from the COVID-19 pandemic

In this paper, we compare two popular statistical learning techniques, logistic regression and random forest, with respect to their ability to classify jobseekers by their likelihood to become long-term unemployed. We study the performance of the two methods before the COVID-19 pandemic as well as the impact of the pandemic and its associated containment measures on their prediction performance. Our results show that random forest consistently out-performs logistic regression in terms of prediction performance, both, before and after the beginning of the pandemic. During the lockdowns of the first wave, the number of unemployment entries and the fraction of individuals that become long-term unemployed strongly increases, and the prediction performance of both methods declines. Finally, while the composition of the (long-term) unemployed changed at the beginning of the COVID-19 pandemic, we do not find systematic patterns across groups with different levels of labor market attachment or across different sectors of previous employment in terms of declines in prediction performance.