Chendi
Ni,
PhD
candidate
David
R.
Cheriton
School
of
Computer
Science
We propose a data-driven Neural Network (NN) approach to solve the optimal asset allocation problem with stochastic benchmark targets during a phase of a defined contribution pension scheme. Compared to previous work, we include path-dependent benchmark targets as part of the objective function to reflect the investors’ pursuit of outperforming the benchmark indices in real life. The NN strategy is directly learned from the market return sample paths generated using bootstrap resampling of historical market data without assuming a particular parametric form.
We show that the NN strategy has consistently higher terminal wealth than the benchmark strategy. Despite that the objective function only includes the terminal wealth explicitly, we show that the NN strategy consistently outperforms the benchmark strategy during the entire phase of the pension scheme. To illustrate the robustness of our approach, we test the NN strategy learned from both synthetic market data sets and bootstrap resampling data sets on bootstrap resampling data sets generated with different block sizes and show that the NN strategy consistently outperforms the benchmark strategy.