Research And Application Of Multi-factor Stock Selection Model Based On RNN-ACT Algorithm

The traditional multi-factor stock selection model is mostly constructed by scoring method or linear regression,which is difficult to excavate some complex nonlinear relations.An intuitive method is to use the neural network to construct the model,for the method of mining the nonlinear relation between variables.But in this process,the neural network model will face the problem of gradient vanishing and the potential systemic risk of financial market,which will eventually lead to the expected effect of neural network model is not ideal.Aiming at the above problems,this paper mainly studies how to combine the traditional multi factor model and neural network algorithm,constructing the multi factor model based on RNN-ACT algorithm.What's more,aiming at the risk of potential problems in system model,put forward the relative concept of singular points.Based on position management and risk control,by the relative singular point motif detecting,which can optimize and expand RNN-ACT multi factor model.The model solves the gradient disappeared,and also has the cumulative rate of return far beyond the benchmark(CSI 300 index)performance.With the comparing of the traditional multi factor model,LSTM model and off-line learning model,found that the cumulative abnormal return trend is large,which prove that this model is based on online learning can be better on the stock yield prediction,and can adapt to the change of market environment.And in avoiding the potential system risk,we use the relative singular point motif detection method to manage the positions reasonably and avoid certain potential risks,thus confirming the validity and superiority of the model.