关键词: 机器学习, 收益率预测, 资产定价, 研究综述

Abstract:

Accurately predicting asset returns is essential for informed investment decision-making and maintaining financial market stability. With the rapid advancements in artificial intelligence and computing technologies， machine learning （ML） has demonstrated notable advantages in handling high-dimensional data and modeling complex， nonlinear relationships. A comprehensive review of ML applications in asset return prediction， encompassing stocks， funds， cryptocurrencies， and bonds is presented. The existing research on algorithm selection， model construction， and performance evaluation is systematically sumarized. This review begins by examining the origins and significance of asset return prediction， challenging the efficient market hypothesis and contributing to behavioral finance by analyzing irrational investor behaviors and sentiments. A spectrum of ML methods is then explored， ranging from traditional linear approaches to advanced deep learning and large language models （LLMs）， highlighting their ability to address the complexities of financial markets. Techniques such as LASSO and Ridge regularization effectively manage high-dimensional datasets， while neural networks and recurrent neural networks （RNNs） capture long-term dependencies in time series data. Moreover， LLMs like BERT and GPT have shown promise in sentiment analysis and processing textual data， further improving predictive accuracy. The findings reveal that ML methods， particularly ensemble learning and deep learning models， consistently outperform conventional statistical models. For instance， Random Forests and Gradient Boosting Machines achieve superior out-of-sample accuracy， and integrating LLMs with financial text data opens new avenues for sentiment-based return prediction. The data sources employed， including historical prices， macroeconomic indicators， financial news， and social media sentiment， enable comprehensive model evaluations under diverse market conditions. By identifying current research gaps and future directions， this review underscores the importance of balancing predictive accuracy with model interpretability， as well as expanding the scope of asset classes examined. In summary， the analysis provides a holistic perspective on ML applications in asset return prediction， emphasizing their potential and challenges. This work informs investors， policymakers， and researchers， facilitating more effective strategies and decisions in the ever-evolving financial landscape.

Key words: machine learning, return prediction, asset pricing, research review