Machine Learning for Investment Models with Return of Premium Clauses Under Different Assumptions (Academic research)

Summary

• The study develops a continuous-time machine learning framework for optimal dynamic asset allocation in Defined Contribution pension schemes with return-of-premium clauses.p.1
• The trained policy reproduces an analytical benchmark with mean absolute error below 0.0038 across 100,000 simulation paths.p.6
• The Deep RL agent achieves a higher expected terminal wealth (5.214) and lower variance (1.118).p.9
• Over a 20-year horizon, an annual balance charge of 2.0% erodes terminal retirement wealth by over 33% relative to the zero-fee benchmark.p.9
• The Deep RL strategy achieved a lower maximum drawdown (15.8%) than static approaches (28.4%).p.10
• During a variance shock the agent reduces equity exposure from 55% to 24%.p.10
• In early career years the agent allocates approximately 65% of the portfolio to equity.p.8

Key insights

1. Expected terminal wealth is 6.368 with a zero balance fee and 4.281 with a high balance fee.p.7
2. The classical unconstrained formula demands an equity allocation exceeding 140%, which pension regulations prohibit.p.8
3. A Dirichlet policy distribution keeps portfolio weights on the simplex without post-clipping.p.8
4. The agent is trained in an environment with 7 continuous state variables discretized over N = 200 time steps.p.5
5. Variance surging from a long-term mean of 0.04 to a peak of 0.11 triggers an immediate reduction in equity exposure.p.10
6. Charge on balance causes exponential decay of expected terminal wealth, while charge on flow causes linear decay.p.9