Research On Stock Forecasting Model And Trading Strategy Based On Deep Learning

Quantitative trading replaces people's subjective judgment with advanced mathematical models,which can effectively avoid making irrational investment decisions in extreme situations.It has a wide range of academic research and application value in the field of stock,futures and foreign exchange,and has become a research hotspot in the field of investment.With the rapid rise of deep learning in recent years,the investment community has paid more and more attention to the application of artificial intelligence in the field of quantitative trading.This paper proposes a stock prediction model based on deep learning,and establishes trading strategies based on the prediction results to obtain excess returns.The main research contents of this paper are as follows:1)Considering that news and market sentiment are both important factors influencing stock price trends,the social heat and sentiment indicator factors are made and added to the existing factor library,and at the same time,the boruta method is used to select all the characteristic factors related to the dependent variable.Experimental results show that the introduction of alternative factors can further improve the quality of the factor library,and the use of feature selection methods can effectively improve the quality of feature selection and improve the final prediction accuracy.2)Aiming at the problem of financial time series prediction,this paper proposes a prediction model based on the extraction and fusion of different scale information.Firstly,considering the particularity of stock prediction,the traditional Inception structure is modified to extract and fuse information at different scales to obtain better characterization.In this structure,one-way one-dimensional convolution is used to replace the traditional convolution method.One-directional convolution ensures that only forward information can be obtained,which conforms to the actual situation that information after the current trading day cannot be known in the stock market.At the same time,because there is no need to extract information from the overall arrangement of sequences,more attention is paid to the change rule of the same sequence under different cycles,and one-dimensional convolution kernel is used instead of the traditional two-dimensional convolution kernel.In addition,by introducing residual connection into the structure,it can not only avoid the difficulty of model training caused by the deep network,but also help to integrate the original information with the extracted information.Secondly,based on the improved Inception structure,a complete network structure is constructed by combining LSTM and attention mechanism to further improve the overall model performance.Finally,the validity of the model is proved by a complete experiment.3)Based on the prediction results,a "dynamic capital allocation strategy" is established.This strategy dynamically allocates funds and stock positions in proportion to the predicted rise and fall,and comprehensively considers investment income and risks.Compared with the traditional binary rise and fall forecast,this strategy uses capital allocation Further increase the excess return rate.In addition,the professional backtest framework is used for backtest,which can simulate the real transaction fees,commissioned orders,etc.,to provide a more comprehensive and reliable analysis means for strategy evaluation.Finally,we compare the strategy with the benchmark strategy and the market return to verify the effectiveness of the strategy.Finally,this paper formed a complete research framework of"data set construction and preprocessing-deep learning model construction-model training and prediction-strategy formulation and backtest".