In the aftermath of the recent financial crisis there has been considerable focus on methods for predicting macroeconomic variables when their behaviour is subject to abrupt changes, associated for example with crisis periods. We propose similarity-based approaches as a way to handle parameter instability and apply them to macroeconomic forecasting. The rationale is that clusters of past data that match the current economic conditions can be more informative for forecasting than the entire past behaviour of the variable of interest. We apply our methods to predict both simulated data in a set of Monte Carlo experiments, and a broad set of key US macroeconomic indicators. The forecast evaluation exercises indicate that similarity-based approaches perform well, in general, in comparison with other common time-varying forecasting methods, and particularly well during crisis episodes.