On Bayesian Mechanics: An Economics of and by Beliefs

We develop Bayesian Mechanics Portfolio Theory (BMPT), which represents an investor as an agent that infers hidden market states, encodes preferences, evaluates policies by expected free energy, and acts. We first prove a boundary theorem showing that if actions change neither the admissible observation channel nor the transition law of the latent state, expected information gain is policy invariant, so policy rankings are determined by pragmatic value, implementation terms, and any nonuniform policy prior. Under a mean–variance-compatible pragmatic cost, BMPT recovers the subjective tangency portfolio at this boundary. We instantiate BMPT as Causal Hierarchical Active Inference (CHAI), a hierarchy of coupled generative-model layers around a hierarchical risk parity base. A prospectively registered walk-forward exercise on twelve exchange-traded funds serves as an implementation demonstration, not as a test of trading-strategy outperformance. CHAI is classified as a tie against its matched base; the Sharpe ratio difference is −0.023 (90% interval [−0.037,−0.009]), with the interval inside the prespecified equivalence margin. A constructed action-gated environment restores policy-specific epistemic value. BMPT therefore provides a broader, explicit, auditable, and causally disciplined model of adaptive portfolio choice; the paper does not claim superior realized returns.