Portfolio Optimization Under Regime Switching and Transaction Costs: Combining Neural Networks and Dynamic Programs

Multiperiod financial models provide superior capabilities over single-period myopic approaches but, in general, suffer from the curse of dimensionality. Prominent features include transaction costs, rebalancing gains, intermediate cashflows, and short- versus long-term trade-offs. In this paper, we propose and test an algorithm combining dynamic programming with a recurrent neural network. The dynamic program provides advanced starts for the neural network. Empirical tests show the benefits of this novel strategy with optimizing a hidden Markov model in the presence of linear transaction costs. Test problems with 50–250 time steps and up to 11 risky assets are solved efficiently, relative to stand-alone dynamic programs or neural networks. The recurrent neural network addresses transaction costs within difficult multiperiod optimization models in polynomial run time.