Rational expectations models fail to explain the disconnect between the exchange rate and macroeconomic fundamentals. In line with survey evidence on the behaviour of foreign exchange traders, we introduce model misspecification and learning into a standard monetary model. Agents use simple forecasting rules based on a restricted information set. They learn about the parameters and performance of different models and can switch between forecasting rules. We compute the implied US-UK post-Bretton Woods exchange rate and show that the excess volatility of the exchange rate return can be reproduced with low values of the learning gain. Both assumptions, misspecification and learning, are necessary to generate this result. However, the implied correlations with the fundamentals are higher than in the data. Including more lags in the model tilts the balance of our findings slightly towards rational expectations and away from the learning hypothesis.