Bayesian Portfolio Optimization by Predictive Synthesis

Battle Count: 263

This paper is highly relevant to quantitative trading as it directly addresses portfolio optimization under distributional uncertainty, which is a core challenge in quantitative finance. The BPS framework provides a principled way to combine multiple prediction models, which is common in quant trading. The mean-variance, quantile-based (VaR/CVaR), and risk parity approaches are all standard in institutional portfolio management. The empirical results on US and Japanese equities demonstrate practical applicability. The robustness to poor-performing expert models is particularly valuable for trading systems.

The methodology involves multiple complex components: Bayesian Predictive Synthesis with dynamic linear models, MCMC posterior sampling, state space model estimation with discount factors, and three different portfolio optimization formulations (constrained mean-variance, quantile-based with penalty functions, and risk parity). The MCMC implementation and proper specification of the dynamic linear model (including beta-gamma volatility model) require significant statistical expertise. However, the expert models themselves (AR, sample means) are straightforward.

The paper provides detailed mathematical formulations of BPS, dynamic linear models, and portfolio construction methods. Expert models (sample means, AR(1)-AR(3)) are clearly specified. However, no code repository is provided, specific hyperparameter values for discount factors (e, delta) are not fully detailed, and MCMC implementation details are referenced to prior work [11] rather than fully described. Stock data is publicly available but exact data sources are not specified.