I introduce a new learning-to-forecast experimental design, where subjects in a virtual New-Keynesian macroeconomy based on Woodford (2013) need to forecast individual instead of aggregate outcomes. This approach is motivated by the critique of Preston (2005) and Woodford (2013) that substituting arbitrary forms of expectations into the reduced-form New-Keynesian model (consisting of the “DIS” equation, the “Phillips curve” and the “Taylor” rule) is inconsistent with its microfoundations. Using this design, I analyze the impact of different interest rate rules on expectation formation and expectation-driven fluctuations. Even if the Taylor principle is fulfilled, instead of quickly converging to the REE, the experimental economy exhibits persistent purely expectation-driven fluctuations not necessarily around the REE. Only a particularly aggressive monetary authority achieves the elimination of these fluctuations and quick convergence to the REE. To explain the aggregate behavior in the experiment, I develop a “noisy” adaptive learning approach, introducing endogenous shocks into a simple adaptive learning model. However, I find that for some monetary policy regimes a reinforcement learning model, applied to different forecasting rules, provides a better fit to the data.