A Review of Quantitative Methods for Stock Prediction from Traditional Models to Machine Learning Integration

Predicting expected stock returns is a core issue in financial market research. Traditional quantitative methods, represented by the CAPM model, struggle to capture complex dynamics. Machine learning has become a necessary technical means to break through the limitations of traditional models. This paper adopts a literature review method, combing through classic literature on traditional financial pricing models (such as CAPM, APT, VAR, etc.) since the 1960s, as well as research on the application of machine learning (such as SVR, CNN, LSTM, ensemble learning, LLM, etc.) since the 2010s. It summarises research trends by comparing different methods' theoretical foundations and technical characteristics. It is found that machine learning can specifically address the four limitations of CAPM through dynamic , multi-factor mining, non-linear pricing, and heterogeneous expectation analysis. Meanwhile, the risk-return balance idea of CAPM provides a theoretical constraint and an interpretable framework for machine learning. Current quantitative investment research requires breakthroughs in model interpretability, domain-adaptive transfer learning, and multi-modal data fusion to advance intelligence and refinement.