Double normalization

[Yiwen233]

Thanks for your work. I read the source code of Informer with RevIN, and I found that the data is normalized by standard normalization, and then sent to the RevIn layer. So there are actually two normalization steps by default. So is that the way to enhance performance and get the result in your paper?

[PlanckChang]

I also found this issue. Do you have any new ideas recently?

[khchul]

Yeah I also think this is a big issue that the authors should answer. Because RevIN is casually used in many recent Time Series papers, the double normalization problem appears very frequently. I'm also wondering if the result of the paper could change if the scale variable in the dataloader is set to false.

[Yiwen233]

I also found this issue. Do you have any new ideas recently?

Not yet

[SamKnightGit]

I would also be interested in the Author's weighing in on this. Table 3 in the paper seems to suggest that RevIN replaces standard normalization techniques but if standardization is necessary for the RevIN layer to reproduce the results, this should probably be made explicit. I plan on replicating a subset of the results with NBEATS on the ETTh1 dataset, with and without pre-standardization of the input data, I'll update here with my results.