Policy Learning with Weak Signals
Abstract
Policy learning in digital experimentation faces three challenges: weak signal-to-noise ratios, rich covariate spaces, and massive data volumes. We formalize this regime by modeling treatment-effect estimates from increasingly fine covariate partitions as Gaussian observations with bounded signal-to-noise ratios. We establish that, in general, the optimal treatment policy is not learnable in this setting. Even learning the optimal policy value suffers from impractically slow rates. However, when treatment effects vary smoothly, we derive minimax-adaptive policies based on linear smoothers that achieve vanishing welfare regret. We demonstrate the practical value of our framework by applying it to large-scale real-world experiments at Netflix, showing that personalized linear-smoothing policies can dominate unpersonalized policies even in this challenging empirical setting.
Related
- Recovering Weak Signals with Normalizing Flows
- Weak Convexity and Proximal Bundle Methods for Nonsmooth Policy Optimization in Robust Control
- Combating Organized Platform Abuse: Amplifying Weak Risk Signals with Structural Information
- A hybrid approach for steady-state production optimization of a real oil and gas platform: Integrating physics-based models, machine learning techniques, and field monitoring signals
- Reinforcement Learning Under Runtime Confounding: Policy Learning and Off-Policy Evaluation
BibTeX
@misc{koch2026policy,
title = {{Policy Learning with Weak Signals}},
author = {Benedikt Koch and Winston Chou and Aurélien Bibaut and Nathan Kallus},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.10167},
eprint = {2610.10167},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.10167},
note = {CC BY-SA 4.0},
}
RIS
TY - UNPB
TI - Policy Learning with Weak Signals
AU - Benedikt Koch
AU - Winston Chou
AU - Aurélien Bibaut
AU - Nathan Kallus
PY - 2026
DA - 2026/10/07
PB - arXiv
DO - 10.48550/arxiv.2610.10167
UR - https://arxiv.org/abs/2610.10167
AB - Policy learning in digital experimentation faces three challenges: weak signal-to-noise ratios, rich covariate spaces, and massive data volumes. We formalize this regime by modeling treatment-effect estimates from increasingly fine covariate partitions as Gaussian observations with bounded signal-to-noise ratios. We establish that, in general, the optimal treatment policy is not learnable in this setting. Even learning the optimal policy value suffers from impractically slow rates. However, when treatment effects vary smoothly, we derive minimax-adaptive policies based on linear smoothers that achieve vanishing welfare regret. We demonstrate the practical value of our framework by applying it to large-scale real-world experiments at Netflix, showing that personalized linear-smoothing policies can dominate unpersonalized policies even in this challenging empirical setting.
ER -