In order to study the expectation formation of financial institutions in the foreign exchange market we develop and apply a recursive selection and estimation algorithm to a dataset of surveyed foreign exchange market expectations. Responses are classified into two groups and forecasting models are endogenously determined within the groups. Estimation results reveal that a fundamentalist–chartist model is capable of explaining a large portion of foreign exchange market expectations. Fundamentalists are found to have mean-reverting expectations whereas chartists have contrarian expectations. Allowing panelists to switch between models significantly improves the fit of the model, especially at the relatively shorter forecast horizons. We find that the fundamentalist model is increasingly used as the forecast horizon extends. Finally, results indicate that model choice is based on a combination of period-specific and individual-specific determinants.