PREDICTION INTERVALS IN BAYESIAN MODEL STACKING

Sammanfattning

Model averaging is commonly used to account for model uncertainty in practical applications. However, combining models makes statistical analysis more complex in various ways. This thesis considers one of them: the construction of prediction intervals. Intervals obtained from Bayesian model stacking, introduced by Yao et al. [2018], are investigated through the scope of nested linear regression models. By simulation, it is shown that stacked prediction intervals are well-behaved in the settings considered. Generally, the intervals provide appropriate coverage without sacrificing precision, regardless of sample size. Furthermore, the results indicate that Bayesian stacking asymptotically produces valid prediction intervals when correctly specified models are among the candidates.