Machine learning for risk profiling: An analysis of pension fund participants

Abstract

This study examines the use of machine learning (ML) techniques for profiling the risk of pension fund participants. We analyze a dataset of 81,563 individual investors in a major Turkish pension fund company (2018–2022), comparing various ML models to the regulatory benchmark. Using recursive feature elimination, we identify self-reported risk attitudes and age – with a nonlinear relationship – as the most important predictors of actual portfolio risk. Our cross-validation results indicate that boosting methods yield modest improvements in predictive accuracy relative to the regulatory risk score. Notably, the performance from using just four variables is comparable to that from using the full questionnaire. Although the overall explanatory power remains modest across all models ($R^2$ of 0.13–0.17), the findings suggest that ML can enhance risk profiling by identifying informative variables and capturing nonlinear relationships. These results have practical implications for designing more efficient risk assessment tools in pension fund settings, potentially simplifying questionnaires without sacrificing predictive accuracy.