Machine Learning Portfolio Choice under Parameter Uncertainty

Abstract

In machine learning portfolio choice, uncertainty about how characteristics map to weights is economically valuable. It guides which positions to trade, how aggressively, and whether active investing is worth implementing. We quantify this uncertainty with a Bayesian neural network whose generalized posterior assigns a credible interval to each portfolio weight. The resulting trading rules substantially reduce turnover while preserving risk-adjusted returns, and the magnitude of the savings depends on the prior, which shapes the geometry of posterior uncertainty. Sparsity-inducing priors yield robust cost reductions; tight uniform-shrinkage priors do not. The key economic mechanism works through parameter uncertainty rather than forecast confidence.

Download

• 📄 SSRN Preprint

Citation

Bianchi, Daniele, and Xiaoyu Zheng. Machine Learning Portfolio Choice under Parameter Uncertainty. Working paper.

@article{bianchi2026bayesian,
  title={Machine Learning Portfolio Choice under Parameter Uncertainty},
  author={Bianchi, Daniele and Xiaoyu Zheng},
  journal={Available at SSRN 6359140},
  year={2026}
}