Research Of Quantitative Stock Selection Based On Transfer Learning

With the rapid development of Machine Learning and Artificial Intelligence,many scholars have tried to use data mining methods such as Support Vector Machine and Neural Network to find the relationship between factors and stock returns by combining machine learning with quantitative investment.However,there are two main problems when the above methods are applied into financial field: 1.Although there is a lot of data available in the financial market for research,the effective sample data in the model training process is very limited.It is worth to study how to use the limited data to construct a machine learning model with less generalization error.2.Financial market is constantly changing,which means that training samples and test samples collected at different time are hardly to maintain the same distribution.To deal with the above problems,this paper attempts to introduce Transfer Learning to the multi-factor stock selection model.To avoid the problem that the training sample and the test sample in the conventional machine learning model need to be at the same distribution,the proposed model constructs training data by learning and selecting samples which has similar distribution with the same-distribution training data from the diff-distribution training data.In this paper,a factor library construction system including factor screening and data dimension reduction and a complete migration learning quantitative stock selection model framework is given.By using the real market transaction data from the begin of 2013 to the end of 2018,the proposed model is empirically analyzed compared with the Adaboost multi-factor stock selection model with same parameters.The backtest results of these strategies prove that the Transfer Learning Multi-factor stock selection model is effective and the idea of Transfer Learning can improve the performance of the traditional machine learning model when the diffdistribution training data and the same-distribution training data are rationally designed.