Uncommon Factors for Bayesian Asset Clusters

Asset returns exhibit grouped heterogeneity, and a “one-size-fits-all” model has been elusive empirically. This paper proposes a Bayesian Clustering Model (BCM) combining Bayesian factor selection and panel tree for asset clustering. The Bayesian model marginal likelihood guides the tree growth for clustering assets, where each leaf cluster fits heterogeneous model selection and estimation. We apply BCM to split the cross section of U.S. individual stock returns, and find MktRF, SMB, and STR (short-term reversal) as common factors. We also identify several uncommon factors that are partially useful to some leaf clusters when splitting the cross section. The tree visualizes individual stock clustering with important splitting characteristics, such as stock variance and market equity. By considering different prior beliefs for factor usefulness, we further discover that factor models with more skeptic beliefs produce more accurate interval coverage. Beyond asset pricing, our framework generally applies to modeling grouped heterogeneity through jointly clustering panel data and variable selection.