Issue with in-sample forecasts for RNN type models

[tomeklapinski]

Hi !
I run a in-sample forecast for AirPassanger dataset from tutorial on RNN, GRU, LSTM models with step_size=1. The models where fitted with forecast horizon h=1 with default parameters.. For all the model the results are far from expected. The usual forecast gives reasonable results. See the picture blow with usual forecast at the bottom

The models like NHITS, TFT with the same setting give reasonable results.
Could you please let me know what is happening ? Maybe some additional tuning is needed? Maybe a bug for RNN type networks ?