Measuring the Predictability of Life Outcomes with a Scientific Mass Collaboration

A mass collaboration of 160 teams using machine learning on rich longitudinal data (Fragile Families) found that life outcomes remain stubbornly unpredictable, with the best models barely exceeding simple benchmarks. Demonstrates the value of the "common task method" for social science while revealing practical limits to prediction that should concern policymakers using such models.

Despite comprehensive feature spaces and optimized machine learning, life outcomes remain stubbornly unpredictable, and prediction error was strongly associated with which family was being predicted rather than which technique was used. This reveals a fundamental limit of representation: even rich encodings cannot capture the complexity that determines individual trajectories, challenging the assumption that more data yields proportionally more predictive power.