Using Machine Learning to Create an Early Warning System for Welfare Recipients

Summary

IZA Discussion Paper presenting ensemble ML models (LASSO, SVR, Boosting) trained on approximately 1,800 features from DOMINO administrative data covering 50,615 individuals aged 15-66 in the Australian Centrelink system. Predicts proportion of time on income support over four years. Reports at least 22% improvement (14 pp increase in R-squared) over existing early warning benchmarks. Estimates potential fiscal savings of approximately AUD 0.99 billion through improved identification of long-term recipients.