基于深度强化学习的高频交易优化算法

High-frequency trading optimization algorithm based on deep reinforcement learning

为解决高频交易的高额交易费用问题,该文提出了一种融合长短期记忆(LSTM)网络细胞结构的深度确定性策略梯度交易算法。该算法利用细胞结构对当前信息和历史特征进行环境特征提取和保存,用于指导交易决策。通过深度确定性策略梯度算法实现在线自动交易,并考虑了交易费率和收盘价格对奖励函数的影响。在上证50指数基金的分钟级数据上进行实验,结果表明,该算法能有效捕获稍纵即逝的交易机会,是一种低风险高收益的稳健型投资策略;LSTM细胞结构和所设的奖励函数能大幅减少交易次数,不仅增加了算法对交易费率的包容性,还提升了收益的稳定性。

In order to solve the problem of high transaction costs in high-frequency trading,a deep deterministic policy gradient trading algorithm is proposed here integrating long short term memory(LSTM)cell structure.Cell structure is used to extract and save environmental features from current information and historical features,which can guide trading decisions.Online automatic trading is realized through deep deterministic policy gradient algorithm,and the impact of transaction costs and closing prices on the reward function is considered.Conducting experiments on the minute level data of Shanghai Stock Exchange(SSE)50 index fund,the results show that the algorithm can effectively capture fleeting trading opportunities,which is a robust investment strategy with low risk and high return.LSTM cell structure and the setting of reward function can greatly reduce the number of transactions,which not only increase the tolerance of the algorithm to transaction costs,but also improve the stability of profits.