Predictions in the social world generally influence the target of prediction, a phenomenon known as performativity. Self-fulfilling and self-negating predictions are examples of performativity. Of fundamental importance to economics, finance, and the social sciences, the notion has long been absent from the development of machine learning. In machine learning applications, performativity surfaces as a distribution shift. A predictive model deployed on a digital platform, for example, influences consumption and thereby changes the data-generating distribution.

Learning under performativity

In the presence of performativity, prediction is no longer a static problem. Instead it is a dynamic endeavor where the environment responds to the deployment of a machine learning model. This adds a new degree of complexity to machine learning. Building on the formal framework of performative prediction the group has recently contributed novel algorithms for anticipating performativity by learning from interactions [], and measuring performativity using observational [] and experimental [] approaches from causal inference.

Power in digital economies

One consequence of performative prediction is a distinction between learning and steering. The fact that predictions can influence the population implies that an institution can not only achieve small risk by fitting patterns in existing data but it can also steer towards data that benefits its objective. The ability to steer in turn is closely related to questions of power in digital markets. Based on this intuition we developed a novel notion of power tailored to specifics of digital markets [] and demonstrated how to measure and instrumentalize it by executing a large scale online experiment on digital platforms [].

Contesting algorithmic systems

Predictive systems can steer outcomes. But the objectives of institutions are not always aligned with those of comsumers. We investigate how the same tools that help firms optimize their systems can also be used by consumers to steer predictive systems towards their own objectives. In particular, we develop a formal framework to study data as a lever for coordinated efforts [] and maintain a repository of practical cases documented in the gig economy. We also explore concrete strategies applied to music streaming platforms [] and ride hailing services []. Performative prediction has seen a plethora of follow-up works, among others establishing connections between performative power and the effectiveness of collective action [].