Iterative Retraining for Performative Stability

18 concepts12 papers2 proposals

In classical machine learning, we treat data as a static resource. We assume that the distribution of features and labels exists independently of the model we train. However, in social and economic systems—such as credit scoring, hiring, or recommendation engines—the act of deploying a model often changes the underlying data distribution. This phenomenon, termed performative prediction by Perdomo et al., creates a feedback loop where the model’s predictions influence the very outcomes it aims to forecast.

The Feedback Loop of Prediction

When a model is deployed, agents in the environment may change their behavior to adapt to the model's logic. For example, if a bank uses a specific set of features to grant loans, applicants may strategically alter those features to improve their chances. As Mendler-Dünner notes, this transforms machine learning from a passive observation task into a dynamic interaction.

To maintain accuracy in these reactive environments, practitioners often use Repeated Risk Minimization (RRM). The process is intuitive:

  1. Train a model on current data.
  2. Deploy the model and allow the data distribution to shift in response.
  3. Collect new data from this shifted distribution.
  4. Retrain the model on the new data and repeat.

The goal of this iterative retraining is to reach a performatively stable point—a state where the model is the optimal predictor for the specific distribution that its own presence induces.

Reaching Performative Stability

The central question in this research direction is whether RRM actually converges to a stable equilibrium or if it leads to unstable oscillations. Perdomo et al. proved that RRM converges to a unique stable point provided two conditions are met: the loss function is smooth and strongly convex, and the "sensitivity" of the distribution—how much the data shifts in response to model changes—is sufficiently low.

While early theory focused on simple convex models, Mitliagkas et al. extended these guarantees to deep neural networks. They showed that by focusing on convexity with respect to the model's predictions rather than its parameters, RRM can achieve linear convergence even in complex, high-dimensional settings.

Scaling to Complex Systems

Real-world performativity is rarely as simple as a single model interacting with a static population. Research has expanded the RRM framework to handle several layers of complexity:

  • Stateful Environments: In many cases, the population's response depends not just on the current model, but on history. Brown et al. demonstrated that iterative retraining remains robust even in "stateful worlds" where the environment has memory, converging to near-optimal outcomes without requiring the complexity of reinforcement learning.
  • Multi-player Competition: When multiple models coexist or compete, the distribution shifts become a product of all their decisions. Narang et al. and Yuan and Gao have mapped these interactions to game theory, defining Nash equilibria for decision-dependent distributions and providing algorithmic frameworks for agents to find stable points even when they don't know how their competitors influence the data.
  • Sequential Fairness: Performativity has significant implications for ethics. Zhang et al. argue that "fair" models can become unfair over time due to feedback loops. They propose using time-lagged causal graphs within the RRM process to ensure that long-term fairness guarantees remain stable as the population evolves.

Managing the Change-Learn Trade-off

A fundamental tension exists in performative learning: the more a model shapes its environment, the harder it becomes to learn from the resulting data. Rodemann et al. describe this as the change-learn trade-off. If a model causes massive shifts in the distribution, the historical data quickly becomes obsolete, leading to high generalization error.

To mitigate this, researchers have introduced specialized optimization tools:

  • Regularization: Cyffers et al. proved that the optimal level of regularization should scale with the strength of performative effects. In some over-parameterized settings, performativity can actually be harnessed to improve performance if managed with the right penalty.
  • Performative Scoring Rules: To prevent models from entering "limit cycles" (where they flip-flop between different states), Boeken et al. suggest incorporating performative-aware scoring rules into the loss function. This aligns the optimization objective with the reactive nature of the environment, guaranteeing convergence to a stable optimum.
  • Historical Anchoring: In the context of Large Language Models, Wang et al. found that iterative self-play can be stabilized by comparing current outputs to historical "anchor" responses, preventing the performance crashes that often occur when models are retrained solely on their most recent (and potentially degraded) synthetic data.

Key Foundational Work

Performative Prediction (2020) — contributed 2 concept elements to this direction.
Performative Prediction in a Stateful World (2022) — contributed 2 concept elements to this direction.
Multiplayer Performative Prediction: Learning in Decision-Dependent Games (2022) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_repeated_risk_minimization_d3b840ac,
  title = {repeated risk minimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/repeated-risk-minimization/rdr_d3b840ac4e5a42e1b99be441e4eff688},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_repeated_risk_minimization_d3b840ac_v1,
  title = {repeated risk minimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/repeated-risk-minimization/rdr_d3b840ac4e5a42e1b99be441e4eff688/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake