Abstract
Performative Prediction addresses scenarios where deploying a model induces adistribution shift in the input data, such as individuals modifying theirfeatures and reapplying for a bank loan after rejection. Literature has had atheoretical perspective giving mathematical guarantees for convergence (eitherto the stable or optimal point). We believe that visualization of the losslandscape can complement this theoretical advances with practical insights.Therefore, (1) we introduce a simple decoupled risk visualization methodinspired in the two-step process that performative prediction is. Our approachvisualizes the risk landscape with respect to two parameter vectors: modelparameters and data parameters. We use this method to propose new properties ofthe interest points, to examine how existing algorithms traverse the risklandscape and perform under more realistic conditions, including strategicclassification with non-linear models. (2) Building on this decoupled riskvisualization, we introduce a novel setting - extended Performative Prediction- which captures scenarios where the distribution reacts to a model differentfrom the decision-making one, reflecting the reality that agents often lackfull access to the deployed model.