Better Than Blume: An Approach to Forecasting Future Stock Betas with Linear Bayesian Models and Sample-Wide Priors

This paper investigates the predictive accuracy of stock beta forecasts by comparing the widely used Blume adjusted beta to four Bayesian regression models I developed. Stock beta, an essential measure of systematic risk, plays a crucial role in portfolio management and valuation; however, raw betas lack forecasting capabilities. While various attempts have been made to develop alternative beta forecasting models, the Blume adjustment remains the most commonly used. To evaluate this, I created four Bayesian models utilizing a single, sample-wide set of priors, simplifying the computational requirements compared to stock-specific priors. The analysis is based on daily return data for stocks with market capitalizations exceeding $1 billion, tested over three distinct periods: 2004–2009, 2010–2015, and 2016–2021. Performance is assessed using standard regression metrics and a set of Diebold-Mariano tests. Although the Blume adjustment consistently outperforms my Bayesian models in forecast accuracy, some of my Bayesian models show potential through their improvement in R-squared over the Blume model. These findings emphasize the challenges of enhancing beta forecasts while maintaining simplicity but also indicate that my Bayesian approach could be further refined and readily applied.