Actions Have Consequences: Detecting Outcome Performativity using Intervention Testing

This paper studies outcome performativity, where predictions causally influence the outcomes they are intended to predict. It introduces Outcome Performativity A/B Detection (OPAB), which compares outcome distributions generated under different prediction interventions. The authors derive sample-complexity bounds for several performativity assumption classes and report empirical validation of those bounds. They also identify regions of indistinguishability in which the available interventions are insufficient for detection. A case study uses the Open Bandits dataset, with implications for settings where interventions are scarce, expensive, or ethically constrained.