Keywords: performative prediction, causal features
Abstract: Predictive models affect the world through inducing a strategic response or reshaping the environment in which they are deployed---a property called performativity. This results in the need to constantly adapt and re-design the model. We formalize one possible mechanism through which performativity can arise using the language of causal modeling. We show that using features which form a Markov blanket of the target variable for prediction closes the feedback loop in this setting. Thus, a predictive model that takes as input such causal features might not require any further adaptation after deployment even if it changes the environment.