Predicting re-employment: machine learning versus assessments by unemployed workers and by their caseworkers

We analyze unique data on three sources of information on the probability of re-employment within 6 months (RE6), for the same individuals sampled from the inflow into unemployment. First, they were asked for their perceived probability of RE6. Second, their caseworkers revealed whether they expected RE6. Third, random-forest machine learning methods are trained on administrative data on the full inflow, to predict individual RE6. We compare the redictive performance of these measures and consider how combinations improve this performance. We show that self-reported (and to a lesser extent caseworker) assessments sometimes contain information not captured by the machine learning algorithm.