The time required for a bus to finish its route and the inter-stops time are two relevant dimensions to predict in the context of bus route planning. If the full route is just the aggregation of the inter-stops, a solution is to train a chained model for each inter-stop and sum the individuals predictions to get the full route time. An eÿcient time prediction will use sequential information as input (i.e. the bus previous delay). However, model chaining fails to take into account the fact that each individual model use corrupted data (as prediction themselves) for its predictions while it was trained using observed data. We introduced a new chained model that correct this issue by chaining not only the prediction but also the training step. Experiments show that our method improve full route prediction in most cases. Moreover, the comparative study of full route and inter-stops time predictions shows that improving the latter does not necessarily improve the former. These observations opens interesting perspectives involving multi-objective optimization procedures.