When policy forecasts are based on the policymaker's present and past actions, current policy affects expectations of future policy, contrary to what happens when forecasters can replicate policymaking perfectly. We show that when forecasts are generated through any linear combination of present and past policy functions that produces expectations consistent with the implemented policy, the optimal discretionary policy exploiting learning converges toward the optimal commitment plan as we approach a situation where people do not discount the future. Since influencing expectations permits improving policy, successful policymakers need to know how policy expectations are formed and how they can affect these expectations.