Forecasting Excess Returns and Analyzing Rolling Beta Dynamics: A Comparative Performance Analysis of CAPM and Machine Learning Models in Vietnam's Banking and Securities Sector Across Covid-19 Phases

This study examines the prediction of excess stock returns within a CAPM framework in Vietnam's banking and securities sectors from 2016 to 2024 by contrasting the standard Capital Asset Pricing Model (CAPM) with two machine learning algorithms: Extreme Gradient Boosting (XGBoost) and Random Forest (RF). A rolling beta approach is used to track time-varying exposure to systematic risk across three periods: before the COVID-19 pandemic, during the pandemic, and after it. Based on popular performance metrics, including coefficient of determination, mean absolute error, root mean square error, mean absolute scaled error, and symmetric mean absolute percentage error, the models are assessed under a variety of validation procedures. Overall, in models with only one explanatory variable, the CAPM is the most stable and efficient, especially in the post-COVID-19 period. RF and XGBoost are generally competitive and sometimes outperform CAPM during periods of high market volatility but do not systematically outperform when restricted to a single-factor framework like CAPM. This study shall be limited to its focus on one sector and the use of only three models. Suggested future research is to expand the analysis to more industries, including deep learning and multifactor frameworks, and to incorporate macroeconomic or sentiment-based variables in the robustness and generalizability exercise.