We study the power of state-dependent unemployment insurance (UI) in stabilizing short-run fluctuations in a heterogeneous-agent economy in which agents’ beliefs may deviate from rational expectations. Managing expectations is key in this application because higher UI generosity raises consumption, partly, by reducing precautionary saving. If the policy is implemented as a contingent rule that indexes benefits to the unemployment rate, households have to forecast it to anticipate the policy stance. First, we solve the model with a list of popular models of belief formation from the literature, including shallow reasoning, k-level thinking, cognitive discounting, adaptive learning and more. Second, we solve the model with a flexible semi-parametric model of beliefs that we estimate from systematic forecast errors of the unemployment rate in the Survey of Professional Forecasters. We find large efficiency loss, in terms of output stabilization, due to imperfect anticipation of higher benefits.