Stock Analysis Features Selection And Concept Drift Mining Based On Causality Discovery Method

In big data era,the amount of financial data is increasing quickly.Discovering meaningful relation,especially causal relationship,from financial big data by using data mining techniques is an important topic for stock market analysis.Comparing correlational relationship,causal relationship draws a distinction between cause and result,has an irreplaceable value for market regulations,market and economic policies making,corporate governance and market prediction.However,most of popular data mining techniques used in stock analysis can only discover the correlational relationship between variables.And the most popular causal discovery technique in stock analysis,Granger Causal Relation Test,is only suitable for time-series analysis in lower-dimensional environment.Fama–French framework is widely used approach for stock markets cross-sectional analysis.But this approach highly dependent on expert knowledge,and has exposed clear limitations in analyzing financial data which is high-dimensional and suffering from concept drift.Concept drift is a common phenomenon in stock analysis.It means that because of the change of market environment,or some other factors have not been considered,knowledge learned from previous financial data devalue in the future.To discover causal relations between financial factors and stock returns in highdimensional environment with concept drift,this paper introduce a new causal discovery technique called modified Additive Noise Model with Conditional Probability Table(ANMCPT).In empirical test,ANMCPT is compared with Fama–French framework in stock market cross-sectional analysis.Results show that the features selected by ANMCPT outperform the features adopted in most previous researches use Fama–French framework.To the best of our knowledge,this paper is the first to compare intelligence causal discovery technique with traditional cross-sectional analysis approach in stock analysis features selection.Our method provides a new technique to causal knowledge discovery in stock market.