Research On Stock Trend Prediction Based On Fundamental Indicators And Cost-sensitive Learning

The research on stock trend forecasting is of great significance to investment.Fundamental analysis is one of the important research methods in stock trend forecasting.Fundamental information or indicators(such as financial indicators)mainly reflect the quality or intrinsic value of stocks.In traditional stock trend forecasting research based on fundamental analysis,most stock features are selected from stock financial indicators and historical transaction data to predict the return rate of all stocks in the stock portfolio,and then a certain investment strategy is adopted to trade stocks.Such as long strategy,long-short strategy;but in the training process of most of traditional prediction models,the investment losses caused by the error of the stock return prediction result are not considered,that is the cost-sensitive problem;In addition,due to lots of features of stocks,there are a large number of studies on the feature selection of stocks to improve the prediction performance of the model.Commonly used stock feature selection methods include random forest,principal component analysis and so on,but these feature selection methods have nothing to do with the prediction model.The main work of this paper is as follows.(1)The traditional stock trend prediction task based on stock fundamentals is defined as an end-to-end supervised learning three-class sample balanced and unbalanced classification task,and a five-class sample balanced classification task to predict the monthly return rate of stocks rank and directly output trading signals.Stock features include 29 fundamental indicators and 11 technical indicators.(2)Aiming at the problem that feature selection has nothing to do with the prediction model,this paper proposes to use the Binary Grey Wolf Optimizer(BGWO)to select stock features,and the BGWO-LSTM model with LSTM as the classifier to predict stock trends.(3)A cost-sensitive learning method to deal with the investment loss problem caused by the classification error of the classifier is proposed.The key to the costsensitive learning lies in the construction of the cost-sensitive matrix combined with the stock trend prediction task.For this problem,different classification situations and investment strategies are combined.The design schemes of cost-sensitive matrices in six cases are discussed,and corresponding classifiers are constructed.(4)This paper uses data from the constituent stocks of the Shanghai and Shenzhen300 Index to predict the ranking of stock returns for the next month.The experimental results show that when the classification categories are 3 and 5,the prediction accuracy of the BGWO-LSTM model proposed in this paper is improved by at least 1.23% and1.31% compared with the benchmark model.The BGWO-LSTM model that introduces cost-sensitive information proposed in this paper has achieved good test results when simulating transactions in six cost-sensitive situations,which has increased the annualized rate of return by at least 35% compared with the model without costsensitive information,which are at least 6% and 23% higher than the average investment portfolio and the annualized rate of return of the Shanghai and Shenzhen300 Index,respectively,verifying the effectiveness of the BGWO-LSTM model that introduces cost-sensitive information in the stock trend prediction based on fundamental indicators.