We propose that the formation of beliefs be treated as statistical hypothesis tests, and label such beliefs inferential expectations. If a belief is overturned through the build-up of evidence, we assume agents switch to the rational expectation. Thus, if the test size is unity, agents hold rational expectations. We solve a Dornbusch-style model of exchange rates under rational expectations and inferential expectations. Under the latter we prove that the regression tests of Uncovered Interest Parity and the rational expectations version of the term structure display a downward bias. The model also explains delayed overshooting and sharp changes in exchange rates.