Uncommon Factors for Bayesian Asset Clusters
63 Pages Posted:
Date Written: September 15, 2022
Asset returns exhibit grouped heterogeneity and a ``one-size-fits-all" model has been elusive empirically. We propose a novel Bayesian Clustering Model (BCM) combining Bayesian variable selection and tree growth guided by the marginal likelihood of factor models, with each leaf cluster fitting an endogenously selected model with potential uncommon factors and dynamic betas. BCM is applied to split the cross section of U.S. individual stock returns, identifying market, size, and short-term reversal as common factors, and several uncommon factors that lose exposure to some clusters during tree growth. Differential factor exposure and potential segmentation manifest primarily through differential stock variance, followed by market equity and earnings-to-price ratio. Built on leaf clusters, a tangency portfolio on cluster-selected factor models deliver exceptional in-sample and out-of-sample performance. We further discover that more skeptical beliefs about factor usefulness produce better interval coverages, and lag market equity positively drives market betas in most clusters. Beyond asset pricing, our framework applies to general problems of jointly clustering (unbalanced) panel data and estimating heterogeneous models for the clusters.
Keywords: Bayesian Inference, Cross Section, Factor Selection, Self-Supervised Clustering, Spike-and-Slab.
JEL Classification: C1, G11, G12
Suggested Citation: Suggested Citation