Forecasting Early Retirement in Older Adults: Integrating CASP-Based Quality-of-Life Indicators from ELSA

As statutory retirement ages rise, predicting early retirement becomes crucial for workforce planning and policy. This study applies machine learning techniques to predict early retirement using Wave 7 of the English Longitudinal Study of Ageing (ELSA), incorporating demographic, financial, social, and psychosocial predictors, including the CASP-19 quality-of-life scale. Four tree-based algorithms (CatBoost, XGBoost, LightGBM, Random Forest) were compared, with ensemble methods explored to optimize performance. XGBoost achieved optimal performance (ROC-AUC = 0.674, accuracy = 65%), while ensembles provided marginal improvements. Feature importance analysis revealed age as the dominant predictor, followed by aging attitudes, social connections, and financial stress. An ablation study showed CASP-19 adds minor incremental predictive value, with minimal performance decline when excluded. Application to working populations achieved 54% accuracy with self-reported retirement expectations, highlighting gaps between algorithmic predictions and individual intentions. These findings show comprehensive psychosocial assessment enhances prediction beyond traditional economic indicators, informing targeted workforce planning.