Strategies Based On GBDT,RAF And Ensemble Models:Application On Stocks From CSI 300 Index

With the development of machine learning technology,quantitative investment presents some new development trends.Investors can use machine learning algorithm to deal with a large amount of data,mining the internal laws of the market.As one of the most important quantitative investment strategies,quantitative stock selection strategy plays an important role,and it has further developed with the help of machine learning.In this paper,we use data of stocks from the most representative CSI300 index from the beginning of 2004 to the end of 2019 to do the empirical research.We study the quantitative stock selection strategies based on random forests,gradient boosted trees and ensemble models.We first introduce the theoretical knowledge of the machine learning model that we used,and then build the basic model.We construct the training set and the trading set based on return characteristics,and train the model.We select the k stocks in the head to convert into the long positions and k stocks in the tail to convert into the short positions according to the prediction probability,and discard the stocks which are more uncertain in the middle part.In this way,we construct the daily portfolio.For k=105 prior to transaction cost,the strategy based on GBDT model produces out-of-sample returns of 0.3 7%per day,which is much better than the daily average return of 0.04%obtained by holding and trading the CSI300 index in the same way.For k=10,prior to transaction cost,the probability that the daily return of strategy based on ensemble model is greater than 0 is 69.62%,and that number still reaches 58.59%after deducting transaction cost.As a contrast,the number of hold and trade the CSI300 index is just 52.78%.We also improve basic strategies and form two new strategies,one is the weighting strategy,the other is the weekly strategy.For k=105 before transaction costs,the average daily return of the weighting strategy based on GBDT reaches 0.40%,which is a certain improvement compared with the basic model;And for k=10,before transaction costs,the weekly strategy based on the ensemble model produces average weekly returns of 0.71%,which is much greater than 0.18%produced by the same strategy of holding and trading CSI300 index.We also avoid tedious daily position adjustment using this strategy.Our findings have practical application value and may help investors to their decisions.