TY - JOUR
T1 - A Bayesian analysis of a variance decomposition for stock returns
AU - Hollifield, B.
AU - Koop, G.M.
AU - Li, K.
PY - 2003/12
Y1 - 2003/12
N2 - We apply Bayesian methods to study a common vector autoregression (VAR)-based approach for decomposing the variance of excess stock returns into components reflecting news about future excess stock returns, future real interest rates, and future dividends. We develop a new prior elicitation strategy, which involves expressing beliefs about the components of the variance decomposition. Previous Bayesian work elicited priors from the difficult-to-interpret parameters of the VAR. With a commonly used data set, we find that the posterior standard deviations for the variance decomposition based on these previously used priors, including ''non-informative'' limiting cases, are much larger than classical standard errors based on asymptotic approximations. Therefore, the non-informative researcher remains relatively uninformed about the variance decomposition after observing the data. We show the large posterior standard deviations arise because the ''non-informative'' prior is implicitly very informative in a highly undesirable way. However, reasonably informative priors using our elicitation method allow for much more precise inference about components of the variance decomposition.
AB - We apply Bayesian methods to study a common vector autoregression (VAR)-based approach for decomposing the variance of excess stock returns into components reflecting news about future excess stock returns, future real interest rates, and future dividends. We develop a new prior elicitation strategy, which involves expressing beliefs about the components of the variance decomposition. Previous Bayesian work elicited priors from the difficult-to-interpret parameters of the VAR. With a commonly used data set, we find that the posterior standard deviations for the variance decomposition based on these previously used priors, including ''non-informative'' limiting cases, are much larger than classical standard errors based on asymptotic approximations. Therefore, the non-informative researcher remains relatively uninformed about the variance decomposition after observing the data. We show the large posterior standard deviations arise because the ''non-informative'' prior is implicitly very informative in a highly undesirable way. However, reasonably informative priors using our elicitation method allow for much more precise inference about components of the variance decomposition.
KW - vector autoregression
KW - priors
KW - nonlinear functions
KW - statistics
KW - stock
KW - shares
KW - econometrics
UR - http://finance.sauder.ubc.ca/~kaili/BVAR_JEF.pdf
UR - http://dx.doi.org/10.1016/S0927-5398(03)00006-9
U2 - 10.1016/S0927-5398(03)00006-9
DO - 10.1016/S0927-5398(03)00006-9
M3 - Article
VL - 10
SP - 583
EP - 601
JO - Journal of Empirical Finance
JF - Journal of Empirical Finance
SN - 0927-5398
IS - 5
ER -