Description

Administrative employment records and online job postings together constitute one of the richest sources of information for understanding labor market dynamics, with direct implications for forecasting labor turnover, anticipating wage pressures, and informing monetary policy. Recent advances in natural language processing (NLP) make it possible to process this large volume of unstructured text at scale and with high granularity, opening the door to systematic analysis of phenomena that have traditionally been hard to measure. In this paper, we develop a deep learning classifier based on BETO (a Spanish BERT model) that maps approximately two million online job postings and six million job contracts into the International Standard Classification of Occupations (ISCO-08), using its Chilean adaptation (CIUO 08.CL). The classifier achieves an F1-score of 0.85 at the most granular (Unit Group) level. We combine this classifier with a keyword-based indicator of working from home (WFH) to examine the impact of the COVID-19 pandemic on WFH job postings across occupations and its medium-term effects. Using job posting data at the firm level and a difference-in-differences (DiD) approach, we find that COVID-19 increased WFH job offers in industries more likely to offer WFH by 2.0 percentage points relative to les exposed industries (a 78% increase compared to the pre-pandemic period). The effect is particularly pronounced in occupations amenable to remote work, such as managers, professionals, and administrative assistants. However, we observe no significant effects in occupations less likely to offer WFH options. Our findings, supported by a pre-trend analysis, demonstrate the value of machine learning in understanding labor market trends, especially in the context of a global crisis like the COVID-19 pandemic.