Forecasting New Employment Using Nonrepresentative Online Job Advertisements With an Application to the Italian and EU Labor Market

Abstract

Using online job advertisement data improves the timeliness and granularity depth of analysis in the labor market in domains not covered by official data. Specifically, its variation over time may be used as an anticipator of official employment variations. However, online job advertisements may not be representative in terms of key labor market variables. The paper presents a methodology that forecasts publicly available official employment LFS recent job starters exploiting its relationship with a bias-corrected version of online job postings, obtained by predictions of a bivariate sample selection model, which jointly estimates the number of vacancies within job profiles and the probability of endogenous selection for nonzero vacancies. LFS new hires (obtained from LFS microdata) were used as benchmark data to measure the bias of online data and adjusted predicted counts. The proposed framework is illustrated using a dataset of Italian online job advertisements spanning from the period 2013-Q2 to 2018-Q2 to forecast quarterly LFS recent job starters 1 year ahead and the Cedefop's Skills-OVATE data using Italy, France, Germany, and Spain in 2022. Results demonstrated that raw vacancies present a strong bias level with respect to benchmark data, whereas sample selection models reduced this bias by half, unlike multilevel estimates. Moreover, LFS forecasts using a VECM that leverages cointegration between LFS recent job starters and adjusted online vacancy series offer a valuable alternative to traditional univariate forecasting methods.