This paper implements a method to identify and estimate treatment effects in a dynamic setting where treatments may occur at any point in time. By relating the standard matching approach to the timing-of-events approach, it demonstrates that effects of the treatment on the treated at a given date can be identified even though non-treated may be treated later in time. The approach builds on a "no anticipation" assumption and the assumption of conditional independence between the duration until treatment and the counterfactual durations until exit. To illustrate the approach, the paper studies the effect of training for unemployed workers in France, using a rich register data set. Training has little impact on unemployment duration. The contamination of the standard matching estimator due to later entries into treatment is large if the treatment probability is high.