Confounding-Valid Conformal Inference for Counterfactual KPIs in Wireless Networks

Assessing "what-if" operational decisions in cellular networks typically relies on observational telemetry that suffers from unobserved confounding variables. While purely randomized trials eliminate confounding, they are too disruptive to run at scale. The authors introduce Confounding-Valid Counterfactual Conformal Inference (CV-CCI), combining abundant confounded logs with minimal randomized experiments through the General Synthetic-Powered Inference (GESPI) framework. Tested across two radio access network (RAN) control settings, the approach maintains finite-sample validity under arbitrary hidden confounding while delivering substantially sharper prediction sets than existing baselines.