Optimizing Portfolios with Neural Networks and Dynamic Programming: Tackling Regime Switching and Transaction Costs

Portfolio optimization is a central challenge in quantitative finance, engaging a wide range of stakeholders including individuals and institutions. While multi-period financial models are more intricate than single-period models, they also incur higher computational costs. Transaction costs play a significant role in portfolio optimization. Research examines dynamic programming algorithms for multi-period portfolio optimization without transaction costs in a regime-switching market. Additionally, the integration of dynamic programming with a recurrent neural network is investigated for scenarios involving linear transaction costs. Dynamic programming serves as a foundation for the neural network, facilitating the learning of the no-trade zone. Both algorithms are tested on a stock from the NIFTY50 index, and their performance is evaluated against various allocation strategies.