Deep Learning Model Ensemble Based On Portfolio Methods

Out-of-sample predictability of financial asset is the core of financial economics^[1],we conduct empirical research in SSE 50 Index and CSI 300 Index return.We use linear regression,stepwise regression,Lasso,elastic net,principal component regression and fully connected neural networks,the result shows that the deep learning method can produce more useful result than standard methods,moreover,not the deepest neural networks perform the best.In order to improve the out-of-sample predicting and application ablity of models,we take models'extreme loss into account and propose a model ensemble method based on portfolio methods.We compare the proposed method with Mallows model averaging(MMA)method by conducting an empirical research in predicting CSI 300 Index component stocks movements.The result indicates that our proposed method has a stronger head predicting ablity than MMA method,and through backtesting we find out our proposed model ensemble method has the highest Sharpe ratio and lowest max drawdown compared with single LSTM models and MMA averaging model.