Toward personalized risk-sensitive decision-making: A novel risk preference adaptive distributional reinforcement learning algorithm for stock trading

Due to the high uncertainty inherent in financial markets, returns (or rewards) in trading are fundamentally random variables. Traditional deep reinforcement learning (DRL) algorithms typically aim to maximize the expected return, while neglecting the full distribution of returns. This limitation hinders their ability to effectively manage tail risks and results in inadequate risk control. Moreover, investors’ varying risk preferences significantly influence portfolio selection and decision-making. To address these challenges, this paper adopts a distributional reinforcement learning (DiRL) framework for stock trading, which models and optimizes the entire return distribution rather than focusing solely on its expectation. This provides a more comprehensive representation of market risk characteristics. Building upon this foundation, we incorporate risk-sensitive decision theory via a probability distortion function to account for individual risk preferences. Based on this, we propose a novel algorithm—Risk Preference Adaptive Distributional Reinforcement Learning (RPADiRL)—which enhances both the robustness and practical applicability of trading strategies under uncertainty. Extensive experiments on eight representative stocks across various risk preferences demonstrate the effectiveness and superiority of the proposed approach compared to state-of-the-art baselines.