When (and What) Matters in the Late-Career Path? A Diagnostic Sequence Approach Using Transformers to Locate Poverty Risk in Late-Career Histories

Linda Vecgaile , Max Planck Institute for Demographic Research

European welfare states increasingly prioritise extended working lives and recognise the need to support individuals in late career—often from age 55—to prevent hardship as workers near pension age. A practical question remains: when and where to intervene to reduce near-term poverty risk? We study a late-career support horizon (ages 55–59) and its association with poverty at ages 60–62—a pre-retirement period when shocks can compound and shape income into retirement. Treating a Transformer model as a measurement device, we use CPF-harmonised panels from Germany and the United Kingdom to encode yearly labour-market states (employment, hours, sector, industry, occupation, disability) and predict poverty at 60–62. We identify when within 55–59 information is most informative, which features matter most, and run calibrated “what-if” edits that flip disability to no disability to yield changes in predicted risk. We also examine differences by country and gender. Three results stand out. First, among those who become poor, late-window disability is the principal predictor—especially in the UK and for women. Second, among those who avoid poverty, stable working time is the dominant protective margin. Third, signal concentrates at ages 58–59. Consistent with this, flipping disability from “yes” to “no” lowers predicted risk in the UK. We make no causal claims. Instead, we offer a transparent diagnostic framework that indicates when support is most informative (ages 58–59) and what to prioritise (late disability risk vs. working-time stability), to complement causal evaluations in European social policy.

See paper

 Presented in Session P7. Education, Labor Market, and Economic Issues