Abstract
In this study, we have developed a dynamic asset allocation investmentstrategy using reinforcement learning techniques. To begin with, we haveaddressed the crucial issue of incorporating non-stationarity of financial timeseries data into reinforcement learning algorithms, which is a significantimplementation in the application of reinforcement learning in investmentstrategies. Our findings highlight the significance of introducing certainvariables such as regime change in the environment setting to enhance theprediction accuracy. Furthermore, the application of reinforcement learning ininvestment strategies provides a remarkable advantage of setting theoptimization problem flexibly. This enables the integration of practicalconstraints faced by investors into the algorithm, resulting in efficientoptimization. Our study has categorized the investment strategy formulationconditions into three main categories, including performance measurementindicators, portfolio management rules, and other constraints. We haveevaluated the impact of incorporating these conditions into the environment andrewards in a reinforcement learning framework and examined how they influenceinvestment behavior.