Computing the unscaled loss in two different ways yields different results?

Question

So I\'m training several deep learning for time series forecasting models in PyTorch. Now most of the models use data that is scaled (e.g. StandardScaler) so that t