基于机器学习的事件选股方法研究

his paper constructs an intelligent stock selection system that integrates machine learning with event-driven strategies, focusing on quantitative modeling of analyst events and news events. In modeling analyst events, it evaluates market impact by constructing three types of features: initial coverage intensity, rating transition degree, and high yield target premium rate, along with a dynamic weighting mechanism, and employs the XGBoost model for prediction. For news events, a FINBERT-BiLSTM hybrid model is proposed: utilizing the pre-trained FINBERT model in the financial domain to extract sentiment and event type features, while capturing temporal dynamic evolution patterns through bidirectional LSTM. The experiments are based on A-share data from 2016 to 2025, benchmarking against the CSI 300, and comparing the performance of linear models, XGBoost, and FINBERT-BiLSTM. The experimental results indicate that both event models effectively capture market signals, with the XGBoost model based on analyst events demonstrating superior profitability and the news event model excelling in risk control, validating the application value of event-driven strategies in quantitative stock selection.