Machine Learning Portfolio Choice under Parameter Uncertainty

Abstract

Parameter uncertainty and nonlinearity interact to shape portfolio construction. We estimate Bayesian neural networks (BNNs) that map firm characteristics directly to portfolio weights by maximising expected utility under the posterior distribution. BNNs significantly outperform linear models with the same priors. The posterior over network parameters induces a distribution over portfolio weights, enabling a no-trade boundary and a posterior-sparse portfolio that substantially reduce turnover while preserving risk-adjusted performance. The prior determines how much these rules save: sparsity-inducing priors produce cost savings robust to the utility-prior trade-off, while uniform-shrinkage priors require precise calibration. The mechanism operates through parameter uncertainty, not return prediction.

Download

• Working Paper

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}
}