Multi-Factor Stock Selection Strategy Based On Gated Graph Convolution Model

Multi-factor stock selection is the most representative strategy model in the field of quantitative investment and is widely used in the investment field.The core of multifactor stock model is Factor Mining.Any quantitative index that is highly correlated with stock returns and can help predict stock returns can be called factors.Traditional factor mining mostly focuses on the fundamental data.Such factors usually have clear economic meaning,and the prediction cycle is long,but the profit space is limited.In recent years,the short period technical factors based on the combination of stock price,trading volume,turnover rate and other trading data began to attract the attention of quantitative researchers.Technical factors are characterized by low logic,but strong ability to predict stock returns,especially short-term changes of returns.How to construct technical factors is a popular topic of quantitative investment.Factor mining usually relies on researchers' understanding of macroeconomic data and fundamental information of companies,while technical factors need to rely on complex algorithms.Selecting the right algorithm is the key to mining technical factors.,Deep learning,as the most powerful algorithm in artificial intelligence field,has a strong ability to extract information from big data.Therefore,it is the focus of this paper to study whether the deep learning model has ability to mine technical factors,and to construct trading strategies with stable profits based on the mined factors.This paper integrates two deep learning models: gated cyclic neural network and graph convolution neural network,and constructs a gated graph convolution model based on attention mechanism.Different from the general stock prediction model which directly outputs the predicted value of stock data,the unique feature of the gated graph convolution model is that the main function of the model is to carry out feature engineering on the original features of the stock,and the final output can be connected with the multi-factor model.All trading days from January 1,2017 to December 31,2021 are taken as the total sampling interval,and the interval is divided into model training interval,factor test interval and strategy back test interval.The stock pool selects 1813 selected stocks from all A shares.The trained gated graph convolution model can generate 100 technical factors(called alpha factors)for each stock,and then carry out factor test and screening for these factors under the framework of multi-factor stock selection theory.Finally,21 alpha factors are screened to construct stock selection strategy.From January 2021 to December 2021,after excluding all transaction costs and taking shanghai Securities Composite Index as the benchmark,the strategy achieved 30.00% return,30.99 annual return,24.05% excess return,12.86% maximum retreat and 1.089 information ratio.In addition,this paper also proves the feasibility and superiority of the factor construction method used in this paper through the benefit attribution analysis and the introduction of other outstanding technical factors for strategic comparison.