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

Aug 3, 2026

Using high-quality nationwide social security administrative data for Australia between 2014 and 2018, the study develops machine learning predictive models for income support receipt intensities across payment enrollees. The results show that machine learning substantially improves predictive accuracy over simpler heuristic models and existing early warning systems. The improvement is measured as a meaningful rise in explanatory power, achieved without additional costs to practitioners because models rely on data already available to caseworkers. Enhanced detection of long-term welfare receipt can support timely interventions for at-risk individuals.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages34

Reading Time1 h 26 min

AccessFree to download

Abstract

Introduction

Long-term income support and policy goals

Early intervention and target-group identification

Research aim and machine learning approach

Motivation and uncertainty about performance

OXFORD BULLETIN OF ECONOMICS AND STATISTICS, 0305-9049 doi: 10.1111/obes.12550 Using Machine Learning to Create an Early Warning System for Welfare Recipients * DARIO SANSONE†,‡ and ANNA ZHU‡,§ † Department of Economics, University of Exeter Business School, University of Exeter, Rennes Drive, Exeter EX4 4PU, UK ([e-mail: d.sansone@exeter.ac.uk](e-mail: d.sansone@exeter.ac.uk)) ‡IZA, Bonn, Germany (e-mail: [anna.zhu@rmit.edu.au](anna.zhu@rmit.edu.au)) § RMIT University, Melbourne, Victoria, Australia Abstract Using high-quality nationwide 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 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 are on income support in the subsequent 4 years with greater accuracy, by a magnitude of at least 22%(14 percentage points increase in the R-squared), compared to the latter. This gain can be achieved at no extra cost to practitioners since the algorithms use administrative data currently available to caseworkers. Consequently, our machine learning algorithms can improve the detection of long-term income support recipients, which can potentially enable governments andinstitutions to offer timely support to these at-risk individuals. I. Introduction Long-term income support (welfare) receipt is an issue many governments around the world aim to prevent (HM Government, 2010; Welfare Working Group, 2011; Reddel, 2018; Scoppetta and Buckenleib, 2018; Hanna, 2019) . In basic security and/or targeted welfare systems, income support payments are designed to provide a minimum standard of living to households who are unable to meet essential consumptions needs with income from private sources (Korpi and Palme, 1998) . Individuals who regularly receive income support – over an extended period of time – are most likely to suffer JEL Classiﬁcation numbers: C53, H53, I38, J68 . *We thank the Editor, two anonymous referees, Bruce Bradbury, Simon Feeny, David McKenzie, Tim Reddel, and Tim Robinson, as well as the participants of seminars delivered at the Social Policy Research Centre, University of New South Wales, the Australian National University, University of Exeter, and IMT Lucca for their helpful comments. Yin King Fok provided excellent research assistance. Zhu acknowledges the support of the Australian Research Council (ARC) Linkage Project (LP170100472) . This paper uses unit record data from the Centrelink administrative records from the Department of Social Services (DSS) . The ﬁndings and views reported in this paper are those of the authors and should not be attributed to the ARC or DSS. All errors are our own. 1 © 2023 The Authors. Oxford Bulletin of Economics and Statistics published by Oxford University and John Wiley & Sons Ltd. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. 2 Bulletin long-term economic disadvantage and social exclusion (Whiteford, 2010) . Entrenched reliance on income support also imposes signiﬁcant demands on government budgets, reduces economy-wide market output, and leads to the intergenerational transmission of welfare cultures (Dahl, Kostøl, and Mogstad, 2014; Dahl and Gielen, 2021; Cobb-Clark et al., 2022) . Individuals remaining unemployed for extended periods is an ongoing issue given the signiﬁcant changes to the economy from automation, technological change, and the continuing impacts ofthe COVID-19 pandemic.1 To prevent such entrenched reliance, policymakers try to intervene early in the welfare careers of high-risk registrants with labo

FAQs

What data and time period are used to build the early warning system?

The models are developed using nationwide Australian social security administrative data covering the period from 2014 to 2018.

How do the machine learning models compare with simpler early warning methods?

Machine learning algorithms significantly improve predictive accuracy compared with simpler heuristic models and early warning systems currently in use.

Why can the proposed approach be implemented without extra cost for caseworkers?

Because the algorithms use administrative data that is already available to caseworkers, so the implementation does not require additional data collection.

Aug 3, 2026

Upload by Evangeline