A Multi-factor Stock Selection Model Based On DenseNet

The paper establishs a multi-factor stock selection model based on DenseNet,a novel deep learning model which improves performance through the dense connection between model layers.The stock selection model takes 50 factors from fundamental analysis,market analysis and technical analysis as input and predicts stocks' future returns for the next quarter.The paper chooses the constituent stocks of CSI Small Cap 500 Index as the stock pool and sets the relocation period for one quarter.According to the model's prediction,a portfolio composed of 20 stocks from the stock pool is constructed.Also,a Fama-Fench 3-Factors model,an SVM model and a buy-and-hold strategy are built as a comparison.Factor data of stocks in stock pool from 2007 to 2017 are used as the training set for the model,and those from 2017 to 2019 as the test set.Before inputted to the model,samples with missing data are eliminated,and the data set is lagged,winsorized and normalized properly.At the same time,the effectiveness of the data set is tested on 6 different methods.Then the DenseNet model and the three comparison models are trained on the traning set and testified on the test set.In order to comprehensively evaluate the model's performance,11 performance indicators from 3 different aspects are calculated,and the Brinson model and factor exposure analysis are conducted to investigate performance attribution.The empirical research demonstrates an outstanding stock selection capability: during the backtest period,the DenseNet model has substantially outperformed its comparison models and achieved significant excess returns over the benchmark index.To further optimize the portfolio,index future short positions are added to the origin combination to hedge the market risk.After hedging,ther beta risk is separated from the portfolio.As a result,the portfolio achieves an annual alpha return of 24.24% with a beta value close to 0.Therefore,the feasibility to utilize DenseNet model in an alpha-hedge strategy has been confirmed.