When algorithmic predictions inform social decision-making, these predictions don’t just forecast the world around them: they actively shape it. Building models that influence the world is, in fact, often the primary goal of prediction. For example, in medicine, we predict the risk of a person developing a disease to hopefully minimize the likelihood that it occurs. In elections, we predict voting preferences with the goal of targeting information campaigns that are explicitly designed to influence people’s political beliefs. If done properly, social predictions are performative. They directly interact with the world around them and change it.

In this thesis, we will first introduce a learning-theoretic framework, performative prediction, that places these problems on formal mathematical grounds. We will illustrate how the framework can be used to analyze common social prediction dynamics, such as repeated retraining in response to strategic effects. Furthermore, we will discuss how it can be used to design decision rules that embrace the distinction between forecasting future outcomes accurately and steering them towards socially desirable targets.

In the second part of the thesis, we will use these theoretical ideas as a guiding lens to study early warning systems, a popular class of risk prediction tools used in over half of US public high schools. We will present the results of a collaboration with the Wisconsin Department of Public Instruction in which we performed the first large-scale evaluation of the long-term impacts of early warning systems on graduation rates. At the end, we will close with a discussion regarding the policy implications of our work.