Research On Timing And Stock Selection From The Perspective Of Structural Divergence In The Stock Market

Timing and stock selection are important parts of the portfolio in intelligent finance,and also two critical issues in financial investment.In this paper,we apply machine learning,such as clustering,classification,and deep learning,to timing and stock selection in financial investment.This research not only enriches the system of intelligent finance,but also improves the return on investment and market efficiency,which has outstanding theoretical and practical value.The main contents and contributions are as follows:1.We propose the concept of structure disagreement and explore its impact on market returns.Different from previous studies of the individual stock disagreement,for the stock market,we propose structure disagreement based on clustering and Gini impurity,and analyze the effect of structure disagreement on market returns.The results of experiments for CSI 100 stock markets show that from Apr.2015,to Aug.2015,where it is the peak,and rapid decline period,the market has experienced extreme stock aggregation,and the structure disagreement is abnormally small.Besides,the structure disagreement has a significant impact on market returns.The higher the structure disagreement,the greater the market returns in the next two weeks.2.We analyze the predictability of China's stock market,explore the impact of structure disagreement on market predictability,and explain the influencing factors of structure disagreement from technical indicators.Considering the excellent performance of machine learning,we choose five classification algorithms,including Linear Regression(LR),Ada Boost(AB),Gradient Boosting(GB),XGBoost(XGB)and Random Forest(RF),which are all classic classification algorithms and have good performance in many fields.Then,we analyze the predictability of China's stock market from a rigorous perspective.According to the experimental results for CSI 100 stock markets from Dec.26,2016,to Jul.1,2018,it is found that the market exhibits short-term unpredictable and long-term predictable characteristics.In particular,the structure disagreement proposed can further improve the prediction accuracy(an average increase of one percentage point),with an accuracy of 59%.Finally,we find that the structure disagreement mainly comes from the technical indicators,trading volume and net capital inflows.3.We propose a stock selection strategy from the perspective of structure disagreement,and explain and verify the strategy effectiveness by factor models.Considering the impact of structure disagreement,we propose a stock selection strategy from structure disagreement.The experimental results for CSI 100 stock markets show that from Jan.3,2014,to Aug.24,2018,the cumulative return of this strategy reaches1.5,while the index for the same period is only 0.6.In other words,the proposed strategy is superior to the passive investment strategy.At the same time,for SSE 180 and CSI 300 constituent stock markets,experimental results are robust.Besides,based on the CAPM model,the Fama three-factor model,the Carhart four-factor model,and the Fama five-factor model,we find that the return of this strategy comes from market factors,market-value factors,book-to-market factors,and momentum factors.Especially,after factor interpretation,there are still significant ? returns,which further prove the effectiveness of the proposed strategy.4.We analyze the effect of dimensionality reduction on stock selection in different market situations,and propose a new rotation stock selection strategy based on this effect.Considering the importance of dimensionality reduction for cluster analysis with high-dimensional data in the stock selection strategy,and the complex relation among market situations,dimensionality reduction,and noise trading,we explore the effect of dimensionality reduction(principal component analysis,stacked autoencoder,and stacked restricted Boltzmann machine)on stock selection in different market situations for CSI 100 and Nikkei 225 constituent stocks.Experimental results show that dimensionality reduction can significantly improve the performance of stock selection in trend situations,but whether it's in up or down depends on the market analyzed.Further,we propose a rotation strategy with and without dimensionality reduction.The results of experiments show that the proposed rotation strategy outperforms the stock market indices as well as the stock selection strategies based on dimensionality reduction and cluster analysis for CSI 100,SSE 180,Nikkei 225,and S&P 500 constituent stock markets.Based on clustering,classification,and deep learning,we explore the two major issues,timing and stock selection,in financial investment.It enriches the system of intelligent finance,provides advice for investment and regulatory,and has significant theoretical and practical value.