From the Real Towards the Ideal: Risk Prediction in a Better World

Prediction algorithms assign scores in [0,1] to individuals, often interpreted as “probabilities” of a positive outcome, for example, of repaying a loan or succeeding in a job. Success, however, rarely depends only on the individual: it is a function of the individual's interaction with the environment, past and present. Environments do not treat all demographic groups equally. We initiate the study of corrective transformations τ that map predictors of success in the real world to predictors in a better world. In the language of algorithmic fairness, letting p^∗ denote the true probabilities of success in the real, unfair, world, we characterize the transformations τ for which it is feasible to find a predictor q̃ that is indistinguishable from τ(p^∗). The problem is challenging because we do not have access to probabilities or even outcomes in a better world. Nor do we have access to probabilities p^∗ in the real world. The only data available for training are outcomes from the real world. We obtain a complete characterization of when it is possible to learn predictors that are indistinguishable from τ(p^∗), in the form of a simple-to-state criterion describing necessary and sufficient conditions for doing so. This criterion is inextricably bound with the very existence of uncertainty.