Using novel nation-wide social security data combined with machine learning tools, we develop predictive models of income support receipt intensities for any payment enrolee in the Australian social security system between 2014 and 2018. We show that off-the-shelf machine learning algorithms can significantly improve predictive accuracy compared to simpler heuristic models or early warning systems currently in use. Specifically, the former predicts the proportion of time individuals will be on income support in the next four years with greater accuracy, by a magnitude of at least 22% (14 percentage points increase in the R2), compared to the latter. This gain can be achieved at little extra cost to practitioners since it uses data currently available to them. Consequently, our machine learning algorithms can improve the detection of long-term income support recipients accruing a welfare cost nearly AUD 1 billion higher than individuals identified in the current system.