深度学习视角下限价指令簿信息含量研究

Research on the Information Content of Limit Order Book from the Perspective of Deep Learning

指令驱动市场中,限价指令簿是集中反映交易者意图和市场流动性的重要信息载体.量化评估限价指令簿微观特征的信息含量,是深入探究市场动态的基础,对于股票价格精准预测、市场效率有效评估等方面有重要意义.本文利用我国股票逐笔指令流数据重构了实时演化的限价指令簿,并基于深度学习模型,在多档位价格以及成交信息的基础上,同时纳入从横向时间维度衡量的指令流增量信息以及从纵向空间维度衡量的指令流存量信息,通过预测多个微观结构变量综合探讨了限价指令簿特征的信息含量.实证结果表明,相比于成交数据,未成交的指令流数据蕴含了更为丰富的有效信息,在预测微观指标方面展现出显著优势.随着预测窗口期的延长,这种信息优势变得更为突出,凸显了指令流数据在市场分析与预测中的重要地位.然而,成交数据信息传递效率更高,可以为市场指标的预测提供互补信息,两个特征的组合使用可以有效提升预测性能,该结论在对于股票自身特征的异质性分析中依然稳健.另外,价格档位与跨资产效应对于特征信息含量具有显著影响,因此在指令流数据挖掘过程中需要审慎选择市场深度以及充分考虑市场环境因素.

In order-driven markets,limit order book serves as a crucial information carrier that centrally reflects traders’intentions and market liquidity.Quantitatively assessing the information content of the micro-characteristics of the limit order book lays the foundation for in-depth exploration of market dynamics and is significant for precise stock price prediction and effective evaluation of market efficiency.This paper rebuilds limit order book using tick-by-tick order flow data.Based on deep learning model,it incorporates incremental information of order flow measured from the horizontal time dimension and stock information of order flow measured from the vertical space dimension,in addition to multi-level price and transaction information.The information content of limit order book characteristics is comprehensively explored through predicting multiple microstructure variables.Empirical results indicate that compared to transaction data,unfilled order flow data contains richer and more effective information,demonstrating significant advantages in predicting micro indicators.As the prediction window extends,this information advantage becomes more prominent,highlighting the important role of order flow data in market analysis and prediction.However,transaction data has higher information transmission efficiency and can provide complementary information for predicting market indicators.The combined use of these two features can effectively improve prediction performance.This conclusion remains robust in the analysis of heterogeneity regarding stocks’own characteristics.Additionally,price levels and cross-asset effects significantly impact the information content of characteristics.Therefore,careful selection of market depth and thorough consideration of market environmental factors are necessary during the process of order flow data mining.