Bayesian Parametric Portfolio Policies

Battle Count: 57

Directly addresses portfolio construction from predictive signals in a high-dimensional setting (242 signals, 6 factors). Provides a principled correction for estimation risk that improves Sharpe ratios, reduces turnover, and lowers tail risk. The framework is implementable with standard optimization tools (L-BFGS-B) and requires no return distribution modeling. Particularly relevant for factor timing strategies and institutional portfolio management where risk aversion is high and signal sets are large.

Requires implementing: (1) CRRA utility maximization with portfolio constraints via L-BFGS-B, (2) diagonal Laplace approximation for posterior variance, (3) Monte Carlo averaging over posterior draws, (4) expanding-window estimation with dynamic prior mean, (5) signal standardization without look-ahead bias. The horseshoe variant adds coordinate-ascent over local/global scales. Computational cost is moderate due to Laplace approximation avoiding full MCMC, but the high-dimensional parameter space (K×L = 6×242 = 1,452 parameters) requires careful numerical implementation.

Paper provides detailed algorithmic specifications in Appendix B, including L-BFGS-B optimization, Laplace approximation, gradient formulas, and prior calibration. Data sources are publicly available (French's Data Library, Chen and Zimmermann's Open Source Asset Pricing database). Expanding-window out-of-sample protocol is clearly specified. However, no code repository is explicitly linked, and some hyperparameter choices (δ=0.35, p₀) require judgment.