Exceptional events in a time series are observations which can be regarded as qualitatively significant anomalies. The detection of such events is a interesting problem in several domains, in particular for the generation of alarms in clinical microbiology. In this paper we propose an approach to the detection of exceptional events based on model selection. For each mathematical form of a model, we choose the parameters of the model by maximum likelihood techniques. Then we select, among the resulting instantiated models, the model which minimizes the mean square error. An exceptional event is detected with an assigned probability, if an observation lies outside the forecasting region defined by the selected model and a confidence interval.