A Study of Stock Price Prediction Models Based on Kernel Clustering Localized Sliced Inverse Regression and Bayesian Model Averaging

Multifactor prediction models are pivotal in quantitative finance research, yet face issues such as the curse of dimensionality, complex interconnections, and model overfitting. To address these challenges, this study introduces a machine learning predictive model grounded in sufficient dimension reduction and model averaging principles, tailored for stock price forecasting. This method innovatively employs kernel function-based clustering and weighting to refine classic Sliced Inverse Regression, thereby mitigating the curse of dimensionality while maximally preserving the efficacy of predictive factors on stock prices. Furthermore, this approach utilizes Bayesian Model Averaging to navigate the intricate relationships between factors and stock prices, alleviating the risks of overfitting and underfitting. Empirical analysis demonstrates that, compared to traditional quantitative prediction models, this approach produces lower mean squared error, absolute error, and relative error in stock price forecasting, thereby confirming its accuracy and robustness.