This paper investigates a model which aims at predicting the order of events; each event is an english sentence. While previous methods relied on a graph representation to infer the right order, the proposed model is made of two stages. The first stage use a continuous representation of a verb frame, where the predicate and its arguments are represented by their word embeddings. A neural network is used to derive this continuous representation in order to capture the compositionality within the verb frame. The second stage uses a large margin extension of PRank. The learning scheme is very interesting: the error made by the ranker is used to update the ranker parameters, but is also back-propagated to update the NN parameters.