Cristian Challu

### Description Tune's library `Tuner` class can stop hyperparameter search based on execution time (using `time_budget_s`) instead of the number of configurations (currently specified with `num_samples`). See https://docs.ray.io/en/latest/tune/key-concepts.html for more...

### Description Raise a Warning in `fit` and `cross_validation` methods when `n_timestamps-val_size-test_size < input_size+h`. ### Use case _No response_

### Description As mentioned in #434, the output generated during the training of `auto` models can be overwhelming. We believe the interaction between PyTorch-Lightning and Tune libraries causes this issue....

### Description Current models learn to forecast the entire window of size `h`. This new feature will allow weighting individual timestamps of the forecasting window. A new parameter (for eg....

### Description The original NBEATSx model (https://www.sciencedirect.com/science/article/pii/S0169207022000413) proposed an exogenous block with a Temporal Convolution Network encoder. The current `nbeatsx` implementation lacks this exogenous block. ### Use case _No response_

### Description Currently, `base_recurrent` models (RNN, LSTM, GRU, TCN, etc.) scale time series using the global statistics for each time series (on the train set). This might reduce the performance...

### Description Add a tutorial for `cross_validation` function of the `core` class. Explain inputs and how it works. ### Link _No response_

### Description Create several default config search spaces for auto models varying by dataset size. ### Use case _No response_

### Description Add tutorial for `StemGNN` model on traffic data. ### Link _No response_

[BUG] Recurrent models do not update scales in cross_validation and predict_isample

1

Recurrent models compute scaling parameters one for the whole training data. Scales are not updated during cross-validation or predict_insample with the lastest information, degrading the performance.