我国金融市场日内状态特征聚类与优化交易执行

Clustering Analysis of China’s Stock Market Intraday States and Application in Optimal Execution

为探究我国金融市场各日内时段市场状态的规律性与相关性,文章采用一种无监督的聚类方法识别各时段的市场状态特征,并将其应用于优化交易执行问题.文章以中证100全样本股票逐笔成交数据为例,分别在15分钟和5分钟采样频率下构造了反映交易日内不同时段整体市场状态集.通过聚合超顺磁聚类算法,文章实现对训练数据集进行聚类分析,提取市场状态特征向量,并实现实时市场状态聚类.在Almgren-Chriss最优交易执行框架文章构建了最优交易执行强化学习模型,将市场状态特征向量纳入输入变量,通过强化学习中深度确定性策略梯度(DDPG)进行求解.不同时间维度下的实证结果均表明我国股市均呈现出明显的日内效应特征,各交易日相同或相近时段的市场状态特征具有一定相似性,且基于聚类结果的市场状态特征向量能够有效地提升交易策略的表现.

To explore the regularity and relevance of intraday market states in China’s stock market,we apply an unsupervised clustering method to identify the market states characteristics of each intraday period and incorporate it in optimizing trading execution strategies.Specifically,we use the limit order book data of CSI 100 stocks to construct a dataset that reflects the overall market state of every 15 min and 5 min of each trading day.Agglomerative Super-Paramagnetic Clustering is applied to analyze the inter structure of the dataset.Then we detect market state signature vectors,which are further used in online market state detection.Under the framework of Almgren-Chriss optimal trade execution model,we build an optimal transaction execution reinforcement learning model,in which market state signature vectors are included as input variables.Empirical results show that,stock market shows obvious intra-day effect characteristics,that is,market state characteristics of the same or similar time periods on each trading day are pretty similar.Market state feature vectors extracted from the clustering results are effective in optimizing execution strategies.