The following datasets are included in the supplied hf_lob.h5 file: 1) features: an array with (453975, 15, 144) values for each time step. For each time step the most recent 15 feature vectors are extracted (144-dimensional features vectors). 2) targets: an array with (453975, 5) values. Five different predictions targets are extracted (only targets 0 (next time step), 3 (next 5 time steps) and 4 (next 10 timesteps) are used in the paper for the evaluation). For each time step there are three different classes: 0 down, 1 stay, 2 up 3) day_splits: a list with 9 items (tuple of two lists with evaluation indices for each evaluation setup) 4) target_splits: a list with 5 items (tuple of two lists with evaluation indices for each evaluation setup) Each split is composed of a two-valued tuple, with the fist value referring to the training indices, while the second one referring to the testing indices.

2019