Multi-dimensional Feature Stock Selection And Timing Combination Strategy Construction Based On Optimized Machine Learnin

Quantitative fund centered on quantitative trading strategies have become one of the most important investment methods in overseas financial markets.With the help of the powerful information processing capabilities of computers,they can comprehensively screen eligible investable objects to avoid investment blind spots.At the same time,it can prevent the interference of personal factors such as emotion and preference to the transaction,and realize rational investment to the greatest extent.The two most important aspects of quantitative investment: stock selection and timing,seeking excess returns from the perspectives of portfolio construction and transaction timing capture respectively.The combination of the two can buy and sell high-quality investment portfolios at the right time to control risks,then achieve the purpose of obtaining stable income.Therefore,based on the actual situation of the domestic stock market,this paper uses the combination of machine learning technology and optimization algorithm to build Extrme Gradient Boosting algorithm model(CS-XGBoost)and Long ShortTerm Memory neural network model(CS-LSTM)based on the Cuckoo Search algorithm.The combination strategy has achieved excellent performance in the backtest period.This paper takes the daily data of the CSI 300 constituent stocks from 2010/01/05 to2021/12/31 as the research object.This paper selects 30 factors from six angles of three categories: fundamentals,technology and emotional indicators to build a factor library.Through the information coefficient IC and information ratio IR tests,a total of 14 effective factors that are highly correlated with the stock’s next-period return are screened out,and the stacked autoencoder(SAE)is utilized to extract deep features and compress the factor dimension.Secondly,CS-XGBoost is adopted to predict the rising and falling trend of individual stocks and formulate a stock selection strategy,which selects the top 10 stocks according to the probability of rising to construct a portfolio,and conduct a real market backtesting during the period from 2020/01/01 to 2021/12/31.The results show that the proposed prediction model has obtained better trend classification results and backtesting performance than the traditional methods which select stocks by scoring method,logistic regression model(LR)and support vector machine(SVM).Then,according to the CS-LSTM model’s judgment on the overall market conditions,capture trading signals,formulate a stock selection timing combination strategy,and finally build a strategy that can stably outperform the benchmark income,and compared with a single stock selection strategy,it is largely Restrained the impact of the decline in the market index,controlled positions,increased returns and reduced risks.The research results show that the portfolio strategy has an improvement of 37.92% of the strategy return,the excess return is 69.23%,the Sharpe ratio reaches 1.712,and the maximum drawdown is only 12.75%,a decrease of 2.88%.