We observe an increasing amount of sequential data, for instance open data sources provide real-time information. In order to apply classical learning algorithms, sequential data are often modelled in an attribute-value setting using a sliding window. In this paper, we propose a relational approach. A first advantage is to let the relational algorithm choose the length of the window. A second advantage is to allow to consider conditions based on the existential quantifier and aggregates. A third advantage is to be able to consider several granularities at the same time.