深度学习能否让失效因子“起死回生”?——基于Transformer模型的中国A股市场月频动量效应挖掘研究  

Can Deep Learning Bring Failure Factors Back to Life?--Discovering Monthly Momentum Effect in China's A-share Market based on Transformer Algorithm

作　　者：李沛然 杨璐 Li Peiran;Yang Lu

机构地区：[1]中央财经大学金融学院金融科技系 [2]中央财经大学金融学院金融学系

出　　处：《投资研究》2023年第7期22-46,共25页Review of Investment Studies

摘　　要：本文使用基于Transformer模型的深度学习算法成功发现了中国A股市场的月频动量效应,并通过分析模型的信息挖掘机制解释了过去A股市场的“月频动量效应消失之谜”。具体而言本文证实Transformer模型能够利用高维嵌入算法与注意力机制成功甄别以彩票型股票偏好和非理性交易行为为代表的市场噪音,在去除相应的干扰因素后中国A股市场存在显著的月频动量效应,多空组合能够获得0.29%的平均月度收益。进一步地本文从行为金融的视角对A股市场动量效应的形成机制进行了相应分析,证实投资者对市场信息的反应不足是动量效应存在的重要原因。本文研究证实深度学习算法可以利用算力优势挖掘增量信息并增强传统因子的定价能力,对提升资本市场定价效率和投资实践具有一定的启示意义。We successfully discovered the monthly momentum effect in China's A-share market using the Transformer algorithm,and explained the"weak monthly momentum effect"phenomenon in the past.Specifically,we proved that Transformer can screen out market noise represented by lottery stock preference and irrational trading behavior,benefits from the high-dimension-al embedding algorithm and multi-head attention mechanism in the model.After removing the corresponding interference factors,there is a significant monthly momentum effect in China's A-share market,the long-short portfolio can achieve an average month-ly return of 0.29%.In addition,we further analyzed the mechanism of the momentum effect in A-share market from the perspec-tive of behavioral finance,and confirmed that the underreaction of investors is the main reason for the existence of momentum ef-fect.Our research proved that the deep learning algorithm can enhance the pricing ability of traditional factors,which has certain enlightenment significance for improving the pricing efficiency of the capital market and investment practice.

关 键 词：动量效应 Transformer模型 资产定价 金融科技 

分 类 号：F832.51[经济管理—金融学] F224