Research On High-send-to-transfer Stock Prediction Based On Threestage Feature Selection And Enhanced-GBDT-logit

In recent years,as the domestic A-share market is keen to pursue the subject of high-send transfers,high-send transfers stock forecasting events have become an important research and practical issue in the field of financial investment for small and medium investors.In the past,research related to high-send stocks mainly focused on the selection of high-send stocks based on hypotheses related to stocks' high-send stocks.However,this will cause problems such as insufficient factor credibility and missing factors.In addition,most of the previous studies used simple linear models to predict high-send stocks,and did not consider the relationship between the combination of factors and the target variable,which caused problems such as the high-send stocks forecasting level that did not meet expectations.In response to these two major drawbacks,this article uses the third-quarter financial report data,basic data and daily trading data of corresponding stocks of listed companies in the domestic A-share market from 2013 to 2019 as the research object,and aims to pass the Relief-PIMP-Pearson/Mic The three-stage feature selection algorithm mines relevant factors from the data and builds a high-send stock factor pool,and then improves the cross-entropy loss function of the original GBDT algorithm from the perspective of difficult case identification when processing the two-class problem,and draws on the idea of GBDTLogistic.The Enhanced-GBDT-Logit algorithm is used to test the effectiveness of the algorithm improvement in this article by comparing the performance of the GBDT model before and after the improvement on the test set of each year.Finally,this article conducts a backtest analysis on the enhanced-GBDT-Logit predicted high-send stocks to confirm the announcement effect of the plan for the stock transfer event.The Relief-PIMP-Pearson/Mic feature selection algorithm not only selects the high delivery factors that have been confirmed in previous studies from the financial report data and stock transaction data,but also unearths new growth potentials such as year-on-year growth in earnings per share and year-on-year growth in company net profit.Factors and new sector factors such as whether it belongs to the small-cap growth sector and whether it belongs to the sub-new stock sector have further improved the high-send stocks factor pool.The Enhanced-GBDT-Logit algorithm in the training process through the idea of giving higher weight to the samples that are not perfect in the model training,the GBDT pays more attention to the difficult-to-learn samples during the training process.The improved GBDT is not only in the AUC and Precision of the test set.Accuracy indicators such as,Recall are stable higher than the original GBDT,and the training speed on the training set significantly exceeds the original GBDT.