Performative Prediction

R07 said theory is an engine, not a camera. R01 said a ranking remakes the reality it claims to measure. In 2020 four machine learning theorists (Perdomo, Zrnic, Mendler-Dünner and Hardt) turned this into definitions, theorems and convergence conditions. "Measurement shapes reality" thereby stopped being a sociological metaphor and became something computable: you can compute where repeated recalibration ends up, and whether the place it ends up is any good.

Performative prediction means that the prediction itself changes the world it predicts. The original definition is one sentence: once a prediction is used to support decisions, it influences the outcome it aims to predict.

Start with the concrete setting. A bank trains a model to predict who will default, then lends according to the prediction. Applicants are not rocks: as soon as "which features lower the predicted default risk" is published or worked out, they adjust their behaviour, so future application data no longer comes from the world the model was originally fitted to. The bank retrains on the new data, the world responds again, and round it goes. Standard machine learning assumes a fixed data distribution that the model fits; performative prediction removes that assumption and makes the distribution itself a function of the model: whichever model you deploy is the shape the world grows into.

The core of the formalisation is a dual role. A model's risk (average loss) has to be computed on the distribution the model itself induces: the model parameters appear twice in the expression, once in the loss and once in the data distribution. That second appearance is where reactivity sits in the mathematics.

The AI evaluation ecosystem is one of the most performative settings there is. Publishing a benchmark is a prediction deployment: it declares what counts as a good model, developers tune their training to it, and the distribution of the next batch of models changes because of it. Miller et al. gave the situation an accurate name in 2021: the echo chamber. You only ever see the distribution you induced, and everything you hear is your own voice coming back.

The two concepts in the module deserve one more pass in plain words. Stability: the model is already optimal for the world it itself created, and retraining will not move it. Optimality: the genuinely best model once "I will change the world" is factored in. The theorem says the two generally do not coincide, and the stable point can be arbitrarily far from the optimal one. Repeated retraining arrives at the former. In the bank's terms: the stable point is "my model is optimal for the current batch of applicants, who have learned to dress up their applications"; the optimal point is "the globally best model, given what kind of applicants my model will produce". The stable point takes this batch of dressed-up behaviour, the batch the model itself taught, and accepts it wholesale as normal data. The optimal point counts the fact that the model taught that behaviour as part of the cost, and will not accept it wholesale. The gap between them is the part of the world your own measurement spoiled.

It and the strategic classification of Y10 are two brothers of the same lineage, and the division of labour has to be kept clear. Strategic classification is a one-shot game: you publish a classification rule, the other side sees it and remakes itself, one round and done. Performative prediction handles the dynamics of repeated deployment and repeated retraining, and asks where that loop finally stops. Later stateful extensions are closer to reality: the world's response has inertia and lag, so today's distribution depends not only on the current model but also drags yesterday's tail, and the convergence conditions get correspondingly stricter.

The premise of the convergence theorem is the part most often skipped. "Repeated retraining converges to a stable point" depends on the world responding gently enough: the data distribution has to change smoothly with the model, by less than a threshold set by the shape of the loss function. When the response is violent or discontinuous, the loop may not converge, or may converge to a very bad equilibrium. Now substitute "benchmark" for "model" and reread the premise: vendor responses to a leaderboard are not continuous fine-tuning but launch-event jumps, swapping models and strategies wholesale and privately testing variants. Smoothness is precisely the assumption the evaluation ecosystem satisfies least. Whether the ecosystem eventually stabilises or oscillates forever depends on a quantity nobody has ever measured: how strongly the world responds to measurement.

In 2025 this line acquired a five-year review and a real-world test. Perdomo et al. re-examined a decade of data from Wisconsin's dropout early warning system and found that the prediction system was to a large extent simply restating the pre-existing inequality at the school level: the prediction is "accurate" because it took part in maintaining the structure it predicted. The mechanism is identical to the self-fulfilling prophecy in R02, except that this time the protagonist is an algorithm retrained every year.

One open question: convergence theory for performative prediction rests on the world responding gently, while the evaluation ecosystem's mode of response (jump-like releases, private testing, withdrawn scores) is precisely violent and discontinuous. Can that response strength be measured? If not, then "the evaluation ecosystem will eventually stabilise" and "the evaluation ecosystem oscillates permanently" are both articles of faith. research/deep/D6 §2 lists it as a formalisation frontier.

The one-line takeaway: a prediction that changes the world, recalibrated over and over, does not converge on the truth but on a world it made itself and that holds only for itself.