A Machine Learning Model for Algorithmic Optimization of Superannuation Schemes

This paper sought to address the challenge of designing a superannuation pension scheme by developing a machine learning-based recommendation model for optimal asset portfolio selection and allocation. Pension schemes face challenges in dealing with the uncertainties associated with financial markets, especially in selecting an appropriate assets portfolio that can optimize the Return-on-Investment. This study used various machine learning algorithms to build optimal portfolios, which were evaluated based on the portfolio’s return and Sharpe ratio. Data used was obtained from the annual financial reports on pension assets’ cost and market value as well as asset income for the period of July 2013 to June 2023. The Genetic Algorithm, Particle Swarm Optimization, K-means clustering, and Mean-Variance Optimization techniques were employed to construct optimal portfolios. Evaluation based on portfolio return and Sharpe ratio revealed the K-means cluster focused on government securities as a high-performing, low-risk option with 11.16% return and 7.11 Sharpe ratio. Conversely, the genetic algorithm and particle swarm optimized portfolio demonstrated a more diversified conservative asset allocation, leading to a mean return of 7.63% with a Sharpe ratio of 3.91, and a mean return of 6.61% with a 3.16 Sharpe ratio respectively. Comparing these constructed portfolios with OECD (2022) which reported that Kenya achieved a real investment return of 2.9% in 2021, signifies that 7/10 of the constructed optimal superannuation portfolios would have resulted in better performance.