Both the private and public sectors are shifting away from defined benefit (DB) pension plans to defined contribution (DC) plans. This growing trend was recently highlighted in a Federal Reserve Economics Research Note, and motivated University of Waterloo researchers to develop a new framework.
The data-driven approach, which is enhanced by machine learning, significantly outperforms traditional methods used by investors and people saving for retirement. In fact, it increases the probability of meeting your investment target by threefold.
If you participate in a DC pension plan as an employee, the framework gives you a strategy for making decisions on how to allocate your assets to ensure a reasonable level of salary replacement upon retirement.
“If one considers that the DC fund will be managed by the employee for over 30 years of employment, it is clear that asset allocation will be of crucial importance,” said Peter Forsyth, co-author of the paper A Data-driven Neural Network Approach to Optimal Asset Allocation for Target Based Defined Contribution Pension Plans.
“A standard strategy that financial advisors use when guiding clients on how to invest their retirement savings, is to put 60 per cent in stocks and 40 per cent in bonds, with periodic rebalancing,” added Forsyth, a Distinguished Professor Emeritus in Waterloo’s David R. Cheriton School of Computer Science. “It’s not a bad strategy but we can do significantly better. Our strategy reduces the probability of not reaching your savings goal from 50 per cent to 15 per cent.”
In developing the multi-period neural network optimization framework, the researchers directly imputed historical data into the machine learning algorithm, and in doing so bypassed the traditional method of constructing a parametric model with a fixed set of parameters.
As a test, they generated some scenarios from a parametric model and asked the machine-learning algorithm to provide the best possible strategy without revealing to it what the model was, where it came from, or any other detail. The machine learning algorithm was able to produce the correct strategy despite having only been provided with the scenarios.
“We found that the machine learning can actually learn the decision-making function very accurately which allows us to bypass the error-prone assumption making steps used in traditional methods,” said Yuying Li, co-author and professor in the Cheriton School of Computer Science.
“We then looked at real market data and found that our approach offers great potential for incorporating much more information into decision making. It’s also better able to solve high-dimensional problems where there are many more uncertain factors, while the classical approach could not solve these problems.”
