This paper performs an out-of-sample comparison of linear factor asset pricing models from an economic perspective under predictability. I assess the economic value added of several factor models when a Bayesian investor is faced with a portfolio allocation problem whereby each model imposes cross-sectional restrictions on the parameters of a predictive stock return regression. The empirical framework explicitly accounts for investor skepticism about the model, i.e., mispricing uncertainty. Despite vast statistical work on in-sample model comparison for the new-generation asset-pricing models, their out-of-sample performance cannot beat a simple benchmark on a wide range of tests from an economic perspective. This is consistent with thus extends the conclusion for the first-generation of factor models. An exceptional case of factor models yielding significant economic gains is observed when evaluating industry portfolios at the shortest horizon (1-month). In this case, I find that the q5 model of Hou et al. (2021) and the behavioral factor model of Stambaugh and Yuan (2017) outperform the historical benchmark in several cases.