This paper develops an end-to-end Bayesian portfolio construction framework for US-listed ETFs spanning equity, fixed income, real estate, and commodity asset classes. We propose three nested return models of increasing structural complexity: a Bayesian shrinkage mean, a named macro-and-style factor model estimated via Bayesian ridge regression, and a hybrid specification that augments the factor structure with block-level residual PCA. Expected returns are estimated separately within asset-class blocks to prevent equity-dominated covariance from distorting exposures in smaller segments of the cross-section. Portfolio weights are selected by optimising a mean-minus-CVaR objective over posterior predictive scenarios, embedding estimation uncertainty directly into the allocation decision. The framework is evaluated through a monthly walk-forward backtest with expanding estimation windows, using a three-layer diagnostic protocol that spans regression fit, Bayesian calibration, and realised portfolio performance.