Predicting unexpected retirement benefits withdrawals using machine learning algorithms

Abstract

Access to Retirement Benefits is normally through the attainment of the statutory retirement age, which we term as expected exits from a scheme. There are circumstances where the benefits may be accessed before attaining the retirement age i.e., upon separation from an employer in an occupational scheme where the employer is contributing on behalf of the member. The timing of these unexpected withdrawals is variable and therefore challenging to predict. This research, therefore, sought to develop a machine-learning model that was able to accurately predict the unexpected withdrawals from the scheme and understand the factors that contributed to the withdrawal event. The research applied secondary data from a Defined Contribution multi-employer retirement scheme in Kenya with over 63,000 participating members for a 1-year period between 2022 and 2023. The research followed the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology, which is a widely used methodology for data science solutions. The data was cleaned and pre-processed in preparation for the model training stage. Further, the data was split into separate sets for training and testing to evaluate the performance of the model and check its generalisability. The Classification algorithms applied in this research were Logistic Regression, Support Vector Machines (SVM), Random Forests, and eXtreme Gradient boosting (XGB). The performance of the models was evaluated using accuracy, precision, recall, and F1-score. The members were categorised into continuing members, expected exits, and unexpected exits, then the algorithms were trained and it was found that the Random Forests and XGB, consistently outperformed the Logistic Regression and SVM algorithms in classifying this data. The logistic Regression and SVM algorithms’ performance improved when the outliers were treated, then improved significantly when the Synthetic Minority Over-sampling Technique (SMOTE) technique was applied to treat the imbalanced data. The XGB and Random forests accuracy did not change but the precision improved slightly but recall marginally deteriorated on the SMOTE balanced data. The selected model was the XGB as it had superior performance across all the metrics with 97% accuracy, 87% precision, 92% recall, and 89% F1-score and it was noted to generalise well on unseen data. The research further explored the explainability of the model to enhance transparency and conducted both global and local explainability. The global explainability was conducted through the XGBoost’s feature importance and showed the top 3 features were Balance, Sponsor Code, and Age. The local explainability was explored using the Local Interpretable Model-agnostic Explanations (LIME) which showed the importance of the features in predicting specific instances The prediction of unexpected exits is expected to help the retirement schemes in their planning for investments, make adequate budgetary provisions for these unpredictable exits, and contribute to interventions to reduce withdrawals.

KEY WORDS: Retirement Benefits, Machine learning, algorithms, Big Data, Unexpected withdrawal Prediction, Classification Techniques.