| In recent years,quantitative investment as a new investment method in the field of fintech has attracted much attention,and its use of popular artificial intelligence to build models instead of man-made operation,avoiding human subjectivity.China’s various funds have also expanded nearly threefold in five years from nearly 20 trillion yuan at the end of 2014 to 54 trillion at the end of 2019.Quantitative investment,as a mature overseas development and relatively late domestic start ingress investment vehicles,is becoming more and more popular with investors and institutions with its high yield and low volatility characteristics.In this paper,we try to apply the high-performing machine learning algorithms such as GBDT to the multi-factor stock-picking model,break through the limitations of the traditional multifactor linear stock-picking model,and improve the algorithm according to the characteristics of financial market.Firstly,it summarizes the research and application of machine learning in the field of quantitative investment at home and abroad,as well as the basic statistical theory method of machine learning,then introduces the pool of factors selected in this paper,including consistent expectation factors,valuation factors,growth factors,etc.,to select more comprehensive factors as far as possible,and then introduces in detail the feature selection principle and algorithm design based on feature engineering to achieve the optimal subset of data selected input models,and model design selects a sliding window method that is more in line with investment decision-making.and grid search and parameters to find the optimal parameters,and finally establish the model.Based on the above ideas,this paper uses gradient lift tree,random forest,support vector machine,multi-logical regression algorithm to verify the effectiveness of stock selection model,empirical results show that GDBT performs better,followed by random forest,which also shows that integrated learning is better than individual models.Based on the characteristics of gradient lift tree,this paper further studies the importance of factors,and finds that the more important factors are mostly technical factors,which have strong predictive ability for stock prices,and in the stability analysis of the model,the accuracy of the GDBT model will change with the number of different periods,and the accuracy of the model is concentrated near 0.68,which generally performs well.In the historical performance backtest,this paper uses the above four models to predict the results of the backtest,the probability value of each period prediction as the dominant stock is sorted,each period long probability ranking in the top 20 stocks,in this way to build the portfolio,the back test results show that GBDT performance is the best,based on GBDT strategy long back measurement income of 167.79%,long year to 43.14%,and the benchmark of the benchmark 500 years.It shows that the strategy has some reference value to investors,and shows that the risk return of the strategy is better.Finally,the paper tries to adjust the investment strategy by using the hybrid model,and finds that the hybrid model investment strategy based on gradient lift tree,random forest and logical regression has some improvement in risk and return.Based on the multifactor stock selection model based on machine learning,this paper aims to explore the complex relationship between financial data,filter the effective information,and build the model to have some reference significance for investment decision-making. |
