In classical machine learning, we treat the world as a static observer. We assume that the data distribution $D$ is fixed, and our goal is simply to find a model $\theta$ that minimizes loss on that distribution. However, in social systems—like credit scoring, traffic routing, or recommendation engines—the act of deploying a model changes the very data it aims to predict.
Juan Carlos Perdomo, Tijana Zrnic, Celestine Mendler-Dünner, and Moritz Hardt formalize this phenomenon as Performative Prediction in their foundational paper, Performative Prediction. When a model's predictions influence the underlying distribution, the distribution becomes a function of the model's parameters: $D(\theta)$.
The Performative Feedback Loop
The core challenge of performative prediction is a feedback loop. When a bank deploys a new credit scoring model $\theta_1$, individuals may change their financial behavior to improve their scores. This shifts the population from the initial distribution $D(\theta_0)$ to a new distribution $D(\theta_1)$. If the bank then retrains its model on this new data to find $\theta_2$, the distribution shifts again to $D(\theta_2)$.
This cycle is known as Repeated Risk Minimization (RRM). In this regime, the learner is not just chasing a moving target; they are actively pushing the target away. The central question of this research direction is whether this iterative process ever stops. Does the model-data loop settle into a steady state, or does it oscillate or diverge indefinitely?
Finding the Retraining Equilibrium
A central concept in this field is the performative stable point. As defined in Performative Prediction (arxiv 2020), a parameter set $\theta^*$ is performatively stable if it is the optimal predictor for the distribution it creates:
$$\theta^* = \text{arg min}_{\theta} \mathbb{E}_{z \sim D(\theta^*)} [\ell(z, \theta)]$$
At this fixed point, the feedback loop closes. If you deploy $\theta^*$, the resulting data distribution $D(\theta^*)$ makes $\theta^*$ the best possible model. Further retraining will not change the parameters.
Perdomo and colleagues proved that if the loss function is smooth and strongly convex, and the distribution $D(\theta)$ is not too sensitive to changes in $\theta$, RRM converges to this unique stable point at a linear rate. This provides a theoretical justification for the common industry practice of "continual retraining"—it isn't just a heuristic; it is a dynamic system seeking an equilibrium.
Sensitivity and the Change-Learn Trade-off
The stability of these systems depends entirely on distribution sensitivity. If a small change in the model causes a massive shift in user behavior, the system may become unstable. Julian Rodemann and colleagues describe this in Performative Learning Theory as a "change-learn" trade-off: the more a model shapes its environment, the harder it becomes to reliably learn from the resulting data.
To manage this, researchers have explored various optimization constraints:
- Regularization: Edwige Cyffers and Marco Mondelli show in Optimal Regularization for Performative Learning that ridge regularization can mitigate the volatility of performative shifts. Surprisingly, in over-parameterized settings, performativity can sometimes improve risk by reinforcing helpful data trends.
- Robustness: Zhuangzhuang Jia and others propose Distributionally Robust Performative Optimization, which optimizes for the worst-case distribution shift within a neighborhood, ensuring stability even when the exact response of the world is unknown.
Long-Term Consequences and Fairness
Performative effects are not just a matter of accuracy; they have profound social implications. Lu Zhang and Yaowei Hu argue in Achieving Long-Term Fairness in Sequential Decision Making that "fair" models can become unfair over time if they don't account for how decisions (like loan denials) affect future population statistics.
Similarly, Liam Peet-Pare and Nidhi Hegde demonstrate in Long Term Fairness for Minority Groups via Performative Distributionally Robust Optimization that by seeking performative stability through robust objectives, we can protect minority groups from the negative feedback loops that standard retraining might exacerbate.
Decentralized and Multi-Player Dynamics
The framework also extends to multi-agent environments. In Performative Federated Learning, Kun Jin and Xueru Zhang show that in decentralized settings, the way users react to a global model is often more critical for convergence than the initial differences in their local data.
When multiple models compete or coexist, the system moves toward a game-theoretic equilibrium. Zhixian Zhang and Linjun Zhang provide a Unified Inference Framework for these multi-player settings, proving that even with multiple actors, iterative retraining can reach a stable state where every model is optimized for the environment they collectively create.