Market makers provide financial market liquidity by continuously offering buy and sell orders at publicly quoted prices, while simultaneously earning profits from the bid–ask spread in the process. Various deep reinforcement learning algorithms have been proposed to address such high-frequency sequential decision-making application. However, identifying and resolving the traditional concept drift problem of machine learning system in highly dynamic and complex financial environments has always been a very challenging task. In this paper, a novel reinforcement learning framework with environmental sentiment awareness incorporating curriculum learning and knowledge distillation is proposed. With the aid of a sudden concept drift detector based on market sentiment analysis, our trading model will restructure itself during significant market changes. Additionally, a novel curriculum learning method has been designed to enhance learning efficiency in diverse time segments comprising extensive learning environments. Furthermore, knowledge distillation is adopted to refine the agent’s adaptive capabilities for handling daily gradual concept drift. Experiments with TAIEX Options (TXO) data demonstrate that our method outperforms traditional models, achieving a 38.17% increase in PnL-MAP and a 0.07 increase in Sharpe ratio, while maintaining comparable inventory risk. During testing, sudden concept drift events were detected approximately once every five market-making trading days (i.e., about once per week). This also validates that our proposed market-making strategy based on a sentiment-aware reinforcement learning framework effectively enhances trading performance by modeling sudden and gradual concept drifts.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号