Economic Growth and the In-Work Poverty Gap Between Temporary and Permanent Employees in the European Union: An Explainable Machine Learning Approach

Abstract

Economic expansion does not necessarily narrow poverty inequalities within employment, yet the predictive value of growth for such disparities remains insufficiently examined. This study develops a contract-specific forecasting framework to assess whether real GDP growth improves predictions of the in-work poverty gap between temporary and permanent employees and whether its predictive value depends on labour-market segmentation and social protection. The framework is demonstrated using a simulated panel structured around Eurostat indicator definitions for 27 European Union countries over 2005–2024, comprising 513 lag-aligned country-year observations. Persistence and Elastic Net benchmarks are compared with Random Forest, XGBoost, and a multilayer perceptron using expanding-window out-of-time evaluations for 2015–2024. Matched predictor-ablation and interaction tests identify incremental forecasting gains, while SHAP explains model predictions. Including GDP growth reduces held-out root mean squared error from 0.817 to 0.721 percentage points for Elastic Net and from 0.807 to 0.733 for XGBoost. The multilayer perceptron achieves the lowest pooled error (0.693), although historical poverty disparities dominate feature attribution, and institutional interactions deliver negligible and inconsistent improvements. The study introduces a transparent framework for distinguishing the predictive contribution of economic growth from its presumed distributional consequences.

Keywords

In-work poverty; Economic growth; Labour-market dualisation; Explainable machine learning; European Union

JEL codes

I32; J41; O47; C45