Research On Multi-Factor Stock Risk Prediction Method Based On Deep Learning

In recent years,the widespread application of deep learning methods in the financial field has greatly promoted the development of stock-related forecasting technologies.As an important factor in measuring the value of stock investment,stock risk can effectively help investors improve investment risk assessment and stabilize returns.In order to further improve the sensitivity and prediction precision of the prediction model to stock risk,this essay combine the multi-factor model with some predictive factors on stock risk.At the same time,it also improved the problems in current research and applications.The main content and contributions of this essay are as follows.(1)For the problem that a large of prediction models rely too much on existing structured data in feature selection,this essay builds s set of multi-factor libraries with institutions' research report as the main source.These reports,as a professional analysis report issued by securities institutions for stocks,can bring a series of fluctuations and impacts on investor behavior and the stock market.The unstructured data of these reports such as investment recommendation levels and profit forecasts can be used as important factors for stock risk prediction.At the same time,this essay also designs and implements a singer-factor testing framework,which completes the testing and selection of features based on domain knowledge.(2)Aiming at the problem of general long short-term memory network(LSTM)model processing multi-factor input and long-short term information combination,the single-layer LSTM prediction model in this essay is adapted to the multi-factor input from by changing the mapping relationship between its input layer,hidden layer and output layer.On the other hand,we use bidirectional LSTM as its internal structure to consider the impact of historical and future data.Based on this,a hierarchical LSTM prediction model with better use of long-term information is proposed.Meanwhile,in order to effectively extract the multi-dimensional feature information of stock data in the input,we propose to combine CNN and LSTM to build a hybrid prediction model,and use the Attention mechanism to improve the scalability and prediction precision of the model.Finally,the experimental results show that the prediction results of hybrid prediction model has been significantly improved.At present,the relevant factors from unstructured data in this essay have been launched on the Zhiyutouyan platform.And the risk results obtained through the prediction model can be used for the related quantitative strategy research.