The authors incorporate a Bayesian learning model into a fairly genera
l model of exchange rate determination in discrete time. The model is
applied to the period following the widening of the French-German ERM
bands in August 1993, in which a systematic underprediction of the fra
nc can be observed until February 1994. A (substantial) part of these
forecast errors can be mimicked by a Bayesian learning process. Simula
tions with our model show that, after the widening of the bands, agent
s, contrary to their initial expectations, gradually learned that the
true process driving monetary conditions in France had not changed not
ably.