Application of machine learning to retirement fund preservation : identifying significant variables in retirement fund preservation decisions

Abstract

This study aims to understand the retirement fund preservation field and determine which factors lead to low preservation of retirement funds. In addition, the study aims to build a machine learning model that classifies the retirement fund preservation data. The study applied feature engineering to the preservation of retirement fund data from a large insurer in South Africa. The three feature-engineering methods applied were Ordinal Encoding, Dummy Encoding and Target Encoding. These methods were applied to build the three models: Logistic Regression, Random Forest and a Support Vector Machine (SVM). All three models can accurately predict whether an individual will preserve or not. The random forest overall performed best but had the lowest precision. The SVM produces the highest precision of the three models. The results from the logistic regression and the random forest showed that individuals who preserve part of the amount paid to them and take the other part in cash have better preservation rate than those who preserved their full amount or did not preserve at all. This is a strong indicator because it shows that if individuals can preserve more and still take a part of their funds in cash the overall preservation of their retirement funds is good. This study could benefit the industry through identifying variables to focus on to improve the individual’s preservation of their retirement funds.