We develop a portfolio allocation framework that recasts monthly ETF rebalancing as a contextual bandit problem. At each decision date, a Bayesian linear model with shrinkage priors – Gaussian ridge and regularized horseshoe – produces a full posterior distribution over one-month-ahead expected returns conditioned on macroeconomic and factor predictors. The contribution is a portfolio learning system designed for realistic conditions where uncertainty, risk control, and execution frictions interact rather than being treated as separate modules.