Performative Prediction, explained.
Performative prediction occurs when deploying a prediction changes the outcomes or data distribution the model tries to predict. The model becomes part of the environment instead of only observing it.
Why it happens
A promoted product can receive more demand because it was forecast to be popular. Retraining on that new demand then absorbs the influence of the deployment. With a negative response, an announcement can also discourage the very activity it predicted. Learning and influencing must be distinguished.
Performative prediction studies situations where deploying a predictive model changes the distribution of the outcomes it is meant to predict. Learning from the changed data and steering the system are distinct processes.
Read the result
Deploy and retrain one cycle at a time. The solid line follows the changing prediction, while the dashed line keeps the original forecast fixed. Frozen-model demand is shown separately because deploying even that unchanged model affects demand. Change the response to restart from the same baseline and compare reinforcement with alternating corrections.
A worked example
A popularity forecast with an invented response rule
Begin at prediction 20 and response 0.5. Deploying produces demand 20+0.5×20=30.
Retrain to 30, then deploy again: the same response rule produces demand 35.
Further iterations approach 40 in this bounded teaching example. Reaching a stable number does not prove that the original demand forecast was accurate or that the feedback benefits users.
OPTIONAL DEEPER DETAILGo deeper: inside the model
Inside this model
Baseline demand is 20 units. Deployment creates demand d=clamp(20 + r×prediction, 0, 100), where response r ranges from −1.5 to 1.5. Retraining sets the next prediction equal to d. The initial prediction is 20. One click is one deploy-and-retrain cycle; positive response can reinforce forecasts, negative response can cause alternating corrections.
Where this idea is useful
A practical use
A popularity forecast can affect which products receive promotion, changing the demand later recorded as training data.
A common misconception
“A forecast cannot affect outcomes because it only describes them.”
When forecasts guide allocation, recommendations, or incentives, those actions can change the subsequent outcome. A predictive model can then participate in generating its own training data.
What this explanation leaves out
- The response equation is invented for teaching, not fitted to a platform. Its scalar fixed point is not the full optimisation problem or a proof of the research paper's convergence guarantees.
Does a stable loop establish a desirable deployment?
No. Stability means repeated updating stops changing much under the assumed response. Accuracy, social consequences, incentives, and the objective being optimised are separate questions.
Which measurement in your work changes because people can see or act on the prediction about it?
Associated thinkers
Associations marked provisional are awaiting source review.
Further reading
Explore the original research or the teaching reference behind this experiment.