Performative Prediction

Performative prediction studies settings where a model’s deployment changes people’s actions or the data distribution, so future predictions can become self-influencing. Common approaches analyze stable fixed points and repeated risk minimization, and use gradient-based, online, randomized, or causal methods to find predictors that remain reliable under feedback. Applications include lending, recommendation, and markets, where strategic responses, selective labels, and external drift complicate evaluation. A central challenge is that stability depends on how people respond, yet those responses may be unknown or hard to identify from passive data alone.