Portfolio Allocation with Dynamic Risk Preference Via Reinforcement Learning : Evidence from the Taiwan 50 Index

In reinforcement learning, an agent interacts with its environment to act according to a policy in accordance with the observation state; these are executed based on a class of neural networks. In the interaction between states and actions, the series of decision sequences are often described using Markovian decision processes, and the corresponding reward is calculated before the appropriate parameters of the strategy are identified by maximizing the advantage function. In this study, we construct the characteristics of an investment environment based on market data in an enhanced learning framework and identified portfolio allocation conditions corresponding to different risk preferences by using proximal policy optimization with the long short-term memory framework. For a given risk preference, a portfolio weighting that maximizes expected returns or minimizes risk can be identified on the efficiency frontier in conjunction with Markowitz portfolio theory. In our empirical study, we construct portfolios from specific stocks in the Taiwan stock market and then train and test models with a moving window to compare the performance and investment risk of the reinforcement learning dynamic portfolio allocation model against other models under various market conditions