PDF Ebook Stock Return Predictability and Model Uncertainty
We use Bayesian model averaging to analyze stock return predictability from a perspective of an investor who faces model uncertainty, or uncertainty about which economic variables should appear in the return forecasting model. Model uncertainty could be more important than the within-model parameter uncertainty, especially when economic variables are at their recently observed levels. The Bayesian approach to model uncertainty is consistent with the existence of out-of-sample predictability in contrast to the classic based analysis, which detects no such predictability. Moreover, the odds in favor of predictability are substantially higher for small-versus-large and high-versus-low book-to-market stocks.
Financial economists have identified economic variables that predict aggregate stock returns through time. Such variables include the dividend-price ratio, expected and unexpected inflation, lagged returns, and the differences between yields on long-term and short-term government bonds and between low grade and high grade corporate bonds (e.g., Campbell (2000)). For several reasons, the “correct” specification of the regression of stock returns on predictive variables has remained uncertain. First, asset pricing theories are not explicit about the true predictors, raising doubts about the external validity of the empirical evidence. In particular, several recent studies (e.g., Bossaerts and Hillion (1999)) confirm in-sample predictability but fail to detect out-of-sample predictability. Second, the multiplicity of potential predictors raises difficulties in interpreting the empirical evidence. For example, one may find that an economic variable is significant based on a particular collection of regressors, but becomes insignificant when an alternative pecification is examined. Whether such a variable is a robust predictor or not is ambiguous.
The uncertainty about the true set of predictive variables, commonly termed “model uncertainty,” is a small sample phenomenon. In sufficiently large samples all potential predictors can be included in an all-inclusive specification. In this regression, irrelevant variables will have slope-coefficient estimates converging to zero, their true value. However, in practice there are many possible predictive variables, but only a limited number of observations. The classical regression paradigm thus offers little help.
Download
PDF Ebook Stock Return Predictability and Model Uncertainty
Posted in :
