KANStock: A KAN-Integrated Reinforcement Learning Approach for Predictive Portfolio Management with Dynamic Risk Control

Deep reinforcement learning (DRL) has emerged as a promising methodology for enabling agents to derive effective portfolio management strategies through dynamic interactions with the financial market. However, most existing RL-based methods do not consider risk control explicitly, limiting their ability to balance risk and return well. In this paper, we propose KANStock, a novel DRL method that integrates an independent risk control module to predict the future risk for decision making and enforce liquidation when the short-term risk exceeds a predefined threshold. Empirical evaluations on two benchmark stock indices verify the efficacy of KANStock in harmonizing risk mitigation with return optimization.