Research On The Prediction Of Stock Suspension Based On Machine Learning

Stock suspension is a mechanism adopted for the purpose of improving the degree of information disclosure,ensuring the stability of stock fluctuations,and ensuring orderly trading of securities.It has played an important role in stabilizing the market and has been widely used in major securities markets.In the development process of the securities market,China has gradually introduced the suspension mechanism and adjusted it for many times to adapt to the actual situation,but the problems such as complex types and many times of suspension have not been well solved.The irregular suspension and long-term suspension of shares often cause great trouble and loss to investors.Therefore,it is very meaningful to explore the law of the suspension of the company's stocks and then realize the prediction of suspension.In view of the lack of research on the prediction of stock suspension at home and abroad,this paper studies the prediction of stock suspension,explores the use of machine learning technology to achieve the prediction of stock suspension,and then completes the construction of the prediction model of stock suspension.Based on the research objectives of this paper,the following work has been completed.The first is the construction of prediction index system.There are many factors that may cause the suspension of trading,such as abnormal fluctuations in the stock market,changes in management,etc,Combined with literature analysis on merger and acquisition forecasts,stock price forecasts,etc.,this paper selects high-dimensional features of both financial and stock data,and forms a data set after preprocessing.The second is the construction of stock suspension prediction model.Considering the imbalance between the data sets in this paper,which will seriously affect the classification effect of the model,leading to the final identification will be more inclined to a large number of categories.In this paper,SMOTE+Tomek joint sampling is used to process the unbalanced samples before the model training to form the balanced data set.On this basis,the training of each single algorithm prediction model is completed.According to the evaluation effect of each single algorithm model on the test set,the base classification algorithm is selected,And the learning of the sub model is completed on multiple training subsets divided based on the feature subspace.Based on this,the base classifiers added to the integrated system are screened by diversity measures,and then a certain combination strategy is used to complete the integration model construction based on feature selection.In order to further improve the reliability of the integrated model prediction,this paper also introduces the confidence mechanism with rejection option.In the classification judgment,the confidence judgment is realized by setting a threshold value,and the samples in the rejection domain are further processed by other machine learning methods.In addition,in order to compare and analyze the integrated model based on feature selection constructed in this paper,a heterogeneous integration model based on the above single models and a random forest model based on random features are constructed.Finally,the data of some listed companies in China are selected to evaluate the prediction effect of the model.The experimental results show that: The selective integration model proposed in this paper has achieved relatively good prediction effect,especially the introduction of confidence mechanism,which makes the model achieve higher credibility and better prediction effect.It is superior to single model and other integration models in each evaluation index.At the same time,from the specific situation of each category prediction,the selective integrated model based on the confidence mechanism not only improves the prediction accuracy on the suspension samples,but also makes the prediction results of the two categories more reasonable,and also obtains more obvious advantages than other models...