Algorithms and anticipation: predictions of re-employment by unemployed workers, their caseworkers and machine learning
Abstract
Individual information about future idiosyncratic events may be used to improve AI-based predictions. Newly unemployed individuals as well as their caseworkers were asked for the perceived individual re-employment chances within 6 months. These are compared to algorithmic predictions from machine learning trained on large administrative data. Correcting the algorithm if the unemployed themselves predict long-term unemployment leads to a "super-predictor" with superior accuracy. The sources of this information advantage are obscured by anticipated future events not always being observable. We use individual risk aversion as indicator of incentives to collect information on future idiosyncratic events. We find that, indeed, the gain in accuracy of the super-predictor is concentrated among the risk-averse.
Country of study
Germany