[논문 리뷰] The Gibbs Posterior and Parametric Portfolio Choice

Christopher G. Lamoureux|arXiv (Cornell University)|2026. 03. 02.

Financial Markets and Investment Strategies인용 수 0

한 줄 요약

본 논문은 매개변수 포트폴리오 선택을 위한 일반화된 베이지안(Gibbs posterior) 프레임워크를 개발하여 priors를 업데이트해 tilts와 out-of-sample 수익에 대한 포스터리를 얻고, 학습률 lambda를 인-샘플 식별 프런티어(KNEEDLE)로 최적화하며, 미국 주식(1955–2024)을 분석해 higher-moment 효과와 리스크 회피가 정규화에 미치는 영향을 평가한다.

ABSTRACT

Parametric portfolio policies may experience estimation risk. I develop a generalized Bayesian framework that updates priors, delivering a posterior distribution over characteristic tilts and out-of-sample returns that is the unique belief-updating rule consistent with the investor's utility function, requiring no model for the return generating process. The Gibbs posterior is the closest distribution to the prior in Kullback-Leibler divergence subject to utility maximization. The posterior's scaling parameter $λ$ controls the weight placed on data relative to the prior. I develop a KNEEDLE algorithm to select optimal $λ^*$ in-sample by trading off posterior precision against numerical fragility, eliminating the need for out-of-sample validation. I apply this to U.S. equities (1955-2024), and confirm characteristic-based gains concentrate pre-2000. I find that $λ^*$ varies meaningfully with risk aversion and depends on higher-order moments.

연구 동기 및 목표

• Provide a coherent Bayesian decision framework for parametric portfolio choice that remains model-free like PPP.
• Deliver a posterior distribution over portfolio policies and out-of-sample returns for decision-making.
• Regularize in-sample without requiring held-out validation by optimizing the learning-rate lambda.
• Quantify uncertainty on economically meaningful objects such as certainty equivalents and Sharpe ratios.
• Examine how higher-order moments and risk aversion affect regularization and posterior decisions.

제안 방법

• Use a Gibbs posterior p_lambda(theta|data) ∝ exp{λ U(data, theta)} π(theta) with a prior θ ~ N(0, I).
• Implement Metropolis-within-Gibbs to draw θ components conditional on 240 months of data.
• Calibrate a learning-rate λ via an identification frontier that trades off posterior precision (log det Σ) against numerical fragility (condition number of Σ).
• Define information deceleration as {-m}/κ^2 from a projection of -log det Σ onto log κ to locate λ* via a KNEEDLE elbow.
• Approximate posterior near the mode with Laplace expansion to relate Σ and the Hessian H_γ of the utility, showing τ = γλ governs the posterior.

실험 결과

연구 질문

• RQ1 Can a likelihood-free, utility-based updating rule coherently update priors for PPP without a return-generating model?
• RQ2 How should the learning-rate λ be chosen to balance precision and fragility in a Gibbs posterior for portfolio decisions?
• RQ3 What is the role of higher-order moments and risk aversion in determining λ* and posterior characteristics?
• RQ4 Do characteristic tilts provide utility gains, and how do these gains evolve over time and across utility specifications?

주요 결과

• Characteristc tilts yield large utility gains prior to 2001; gains diminish in the 21st century.
• λ* varies meaningfully with risk aversion (γ) and departs from the quadratic benchmark due to higher-order moments.
• Posterior geometry (Σ, λ) explains regularization without out-of-sample validation; confirmation that posterior-based decision rules improve certainty-equivalent returns.
• The mean-variance approximation links λ* to γ via τ = γλ, with higher moments affecting the Hessian and thus λ*.
• The identification frontier (KNEEDLE) pinpoints the elbow where marginal learning gains are offset by fragility, guiding λ* in-sample.
• The posterior mean-based, decision-theoretic portfolio yields out-of-sample certainty-equivalent returns that exceed those implied by the posterior mean.