filter_out_training_samples is executed during the analysis (training, prediction and evaluation) and it filters out samples that have already been used for training. This function was designed assuming that all the training data for a sample is trained during the same run. This is fine for models whose time-splitting methods know the number of ranges beforehand. In such cases the samples are not used for training until the end of the analysable. Examples of these targets are quarters or quarters accumulative.

In 3.7 we introduce time-splitting methods whose number of ranges is not known before hand. An example of this is a weekly time-splitting method. In these cases we have new training data every week.

We need to add the rangeindex field to mdl_analytics_train_samples and we need an upgrade step so that all the current records are repeated for each of the rangeindexes available in the time-splitting method used by the model.