Winners from Winners: A Tale of Risk Factors

Abstract

Starting from twelve distinct factors from the recent literature, plus twelve principal components (PCs) of anomalies unexplained by the initial factors, a Bayesian comparison of approximately seventeen million models in terms of marginal likelihoods and posterior model probabilities shows that {Mkt, MOM, IA, ROE, MGMT, PERF, PEAD, FIN}, plus the nonconsecutive principal components, {[Formula: see text]} are the best supported risk factors. Pricing tests and annualized out-of-sample Sharpe ratios for tangency portfolios suggest that this asset pricing model should be used for computing expected returns, assessing investment strategies and building portfolios. This paper was accepted by Lukas Schmid, finance. Supplemental Material: The data files are available at https://doi.org/10.1287/mnsc.2023.4668 .

US efficient factors in a Bayesian model scan framework

“…We use the model scan approach building on Chib et al . (2020), and Chib et al . (2022) to examine the performance of different models that can be formed from 13 investment factors and their efficient counterparts over the sample period is between July 1972 and December 2022.…”

“…Second, we complement the Bayesian model scan studies of Barillas and Shanken (2018), Chib et al (2021a, b), andChib et al (2022) in U.S. stock returns by conducting a model scan on an extended group of factors. Third, as well as conducting a model scan across the whole sample period, we also consider whether the best factor model in more recent data is different from the best factor model using all data using the approach of Chib et al (2021b). We find that the optimal change point in the sample period is December 1997.…”

“…The attraction of their approach is that the ML can be solved analytically as relies on multivariate normality. Chib et al . (2022) use this approach to run a model scan for comparing models in a set of 12 U.S factor returns.…”

“…(2022) to examine the performance of different models that can be formed from 13 investment factors and their efficient counterparts over the sample period is between July 1972 and December 2022. We use the first 10% of the sample as a training sample as in Chib et al. (2022) to estimate the hyperparameters for the prior distribution, and we use the remaining 90% to conduct the model comparison tests.…”

Integrating Factor Models

“…To our knowledge, Avramov and Chao (2006) is the first study to formally compare asset pricing models, both nested and nonnested, using posterior probabilities. Subsequent studies include Anderson and Cheng (2016), Stambaugh and Yuan (2017), Barillas and Shanken (2018), Chib and Zeng (2019), Chib, Zeng, and Zhao (2020), Bryzgalova, Huang, and Julliard (2023), and Chib, Zhao, and Zhou (2023). Our study differs from these in four major respects.…”