Prediction of stocks with high transfer based on ensemble learning

Abstract

The "high transfer"dividend policy implemented by listed companies is a special phenomenon in the Chinese stock market. The stocks with high transfer can generate significant positive excess returns before and after the announcement date, so it is important for investors to predict the high transfer stock as precisely as possible. Previous studies mainly focused on regression models to make predictions, such as logistic and probit models, the prediction results of which are average. This paper uses the data of the Chinese A-share market for the last eight consecutive years and screens out 13 important factors through stock dividend theory. We construct three ensemble models for prediction, namely XGBoost, LightGBM, and CatBoost. The results show that the three ensemble models are far superior to the traditional Logistic model. And the prediction effect of the three ensemble models have been improved after Bayesian parameter tuning. Furthermore, three models are fused by simple voting and rank-score, it turns out that almost all metrics, including accuracy, precision, F1 value and AUC value, perform better than the single model. In addition, the top 10 and top 20 hit rate reach 100%, suggesting that the fusion model can predict the top 20 stocks with high transfer accurately, which has certain practical significance for strengthening the security of investment.