Machine Learning for Binary Classification of Daily Stock Price Movements in the Chinese A-Share Market: A Narrative Review

Abstract

The stock market exhibits characteristics such as high volatility, nonlinearity, and strong noise. Precisely predicting stock price trends has long been a challenging issue in quantitative finance. Current research often conflates the two prediction tasks: stock trend classification and price return regression. Common issues include inappropriate model selection and unclear applicability to the scenario. In response to the research challenges mentioned, this paper focuses on stock trend prediction, identifies the drawbacks of mainstream models, and provides references and guidance for subsequent research. This article adopts the narrative review method. It examines the empirical literature published from 2020 to 2025, comparing the representative methods such as SVM, XGBoost, LSTM, CNN-LSTM, and Transformer from four aspects: model principles, experimental design, prediction performance, and application scenarios. As a result, different models have their own applicable scenarios based on data characteristics, market environment, and prediction periods. The cost-benefit ratio of XGBoost in the low-frequency prediction scenario is relatively high. LSTM is the current mainstream model in this field. The transformer only has a slight performance advantage in high-frequency and large-sample scenarios. The out-of-sample accuracy of daily stock trend prediction has remained stable between 52% and 60% over the long term. Currently, several common problems in this field include sample category imbalance, non-standard time-series experimental designs that lead to artificially high prediction accuracy, misalignment between classification accuracy and real trading returns, and a lack of interpretability in deep learning models. Based on the research findings, this paper proposes optimization paths across four dimensions: experimental norms, evaluation system, model interpretability, and multimodal feature construction, to advance the intelligent stock trend prediction model towards standardization, practicality, and interpretability.