This article documents surprising learning patterns that can occur under model misspecification. An agent resorts to predictive regressions and fails to take into account autocorrelation in the dependent variable. Remarkably, when the dependent and independent variables are uncorrelated, we find cases for which the resulting out-of-sample is well above zero, which benefits the agent, in spite of the erroneous model. We refer to them as instances of unexpected opportunity. When both variables exhibit high levels of persistence, we reveal the existence of counter-intuitive configurations for which the increases when the absolute correlation between the series decreases. Our theoretical results are confirmed by extensive simulations and complemented by an empirical exercise of equity premium prediction for which we use 15 predictors commonly referenced in the economic literature.