Quantitative Multi Factor Stock Selection Strategy Of Dynamic Optimization Machine Learning Algorithm

In the context of deepening financial science and technology,the recognition of quantitative investment is also increasing.In recent years,the proportion of quantitative funds and index funds in equity public funds continues to grow.In the whole year of 2019,the scale of quantitative fund increased from 0.12 trillion to 0.17 trillion,accounting for 5.4% from 5.2%;the scale of index fund increased from 0.49 trillion to 0.77 trillion,accounting for 23.8% from 22%.As an important AI technology,machine learning algorithm is more and more widely used in quantitative investment.A large number of fund companies have attracted talents in the field of machine learning.In this paper,we combine machine learning algorithm with traditional multi factor model,take China Securities 500 index as stock sample pool,and take2015-2019 as backtesting interval.We candidate seven dimensions including valuation,growth,quality,risk,emotion,momentum and technology,totally 29 factors.After factor validity test,we select 16 effective factors,and then combine them with support There are four machine learning algorithms: vector machine,random forest,k-nearest neighbor and xgboost.In the process of model building,the core content is the selection of model algorithm and the optimization of algorithm parameters.In this paper,the experimental idea of control variables is adopted,and only the model algorithm ischanged under the condition that other elements remain unchanged.By comparing the prediction effect of four algorithm models,the optimal algorithm is obtained.In a single algorithm model,the process of parameter optimization is also involved,which mainly depends on the 10 fold cross test and grid search.First,combined with the actual meaning of the parameters,the alternative values are given,and then the parameters with the best prediction effect are selected.In this paper,the dynamic optimization method is used to select the optimal algorithm and corresponding optimal parameters in each rolling training window.The traditional static model uses the fixed historical data to select the algorithm and parameters,and puts the selection process of the algorithm and parameters ahead of the back testing process,so the "optimal algorithm and parameters" is not optimal in every time period,but only for the fixed historical data.Compared with the ordinary static model,the dynamic adjustment model in this paper has a higher yield,and in the process of back testing,the optimal algorithm and parameters at each time point will change dynamically.Therefore,this paper initially determines the quantitative model,and then,in the process of model robustness test,this paper also adjusts the training window length,position number and label setting mode.Different training window length will cause different timeliness and data enrichment,which will affect the model effect;because the stock with the highest score of the model will not necessarily bring the highest return in the next period,The mechanism of quantitative stock selection model is to select stocks with relatively high scores,so this paper also adjusts the actual number of positions,and compares the actual effect of positions of 5,10,15 and 20 stocks;in addition,in the process of machine learning,labels also have an impact on the learning effect,so this paper also adjusts the way of setting labels.All the above adjustments are based on the idea of control variables.Finally,the importance of factors will be re screened,because it is one-sided to judge the actual effect of factors simply according to IC test.Finally,the mechanism of the optimal model is as follows: the number of positions is 10,the length of training window is 12 months,the first 10% of the stocks in the sample stock pool are used as positive classification,the last 10% of the stocksare used as negative classification,and all the initial candidate factors are used.From2015 to 2019,the model has achieved a cumulative return of 235.17% and an annual return of 47.03%,outperforming the public funds in the whole market in the same period.