Research On Alpha Quantitative Model Of Stock Based On Deep Learning

Recent years,quantitative trading has developed rapidly in China,and has become a potentially wonderful means of trading.However,as time goes on,traditional quantitative trading is facing great challenges such as lack of efficient stock factors,simple types of models and insufficient complex non-linear modelling ability.Thus,in this dissertation,stocks alpha quantitative model is constructed to solve these problems.And experiments have been carried out on constituents of SSE 50.Efficient stock factors mining and selecting is vital for traditional quantitative strategy.Traditional factors mining mainly depends on artificial violent search,which requires much time and effort and is inefficient.In this dissertation,autoencoder stock alpha factors mining model is presented,which combines sparse autoencoder and denoising autoencoder to construct mixed autoencoder and multilayer stacked autoencoder,to obtain deep feature representations of stocks and to achieve the aim of mining stocks factors automatically.The experimental results show that autoencoder stock alpha factors mining model is able to reduce noise in raw data and features,refine effective information and further obtain more effective factors features representations.To some extent,the model could replace the method depending on artificial violent search.Traditional quantitative prediction models mainly select investment path with high odds by statistical analysis.At the same time,these traditional methods often predict yield of single index or stock directly.This even increases the prediction difficulty and reduces accuracy.Moreover,methods based on statistical analysis lack enough non-linear ability of representation when faced with stock time series.In this dissertation,a novel stock alpha prediction model based on cost-limited LSTM is presented.The model predicts yields vector of all the constituents from stock index and then gets the best weights ratio of portfolio.Through active construction of portfolio,the model is capable of gaining alpha yields,which is irrelevant with market fluctuations.The experimental results show that the stock alpha prediction model based on cost-limited LSTM can gain more stable excess returns with lower trading frequency.The model lays a solid foundation for implementation of alpha hedge strategy.