This paper examines the performance of forecast-based nonlinear Taylor rules in a class of simple microfunded models. The paper shows that even if the policy rule leads to a locally determinate (and stable) inflation target, there exist other learnable "global" equilibria such as cycles and sunspots. Moreover, under learning dynamics, the economy can fall into a liquidity trap. By contrast, more backward-looking and "active" Taylor rules guarantee that the unique learnable equilibrium is the inflation target. This result is robust to different specifications of the role of money, price stickiness, and the trading environment.