Chess_champion

Me too, hoping the original creator or anyone has any input on this. I am not sure if there is a problem with TFT algorithm or the implementation. Have you...

Can you please elaborate on what you mean by including more categories? Are you talking about static categories? You mean more predictors? I can try it and update it here.

I have done the same and tried everything, but the forecast is still flat. If I find out the solution, I will let you know (post it here or StackOverflow)

I tried adding other predictors or adding group variables as static variables, playing with some hyper-parameters but still nothing, constant prediction.

Since it does not capture any volatility, then the accuracy or MAPE would not be good, but regardless of error, we need to capture the volatility in data.

Hi everyone, I think I fixed this issue, the problem is coming from decoder side and relative_idx is not enough to learn any volatility. Mine now has a acceptable prediction...

Sure, yes the day of week, month, year, etc help a lot in capturing the volatility.

@valentinfrlch, no by default TFT, NHITS, DeepAR, etc are normalizing the data sequence by sequence, you can configure it such that it is done by time series by time series.

@meteoDaniel Any update on your findings? I do not think it is the package.

To be honest I got some good prediction out of TFT. Try adding static variables, temporal variables, etc. However it still looks very suboptimal. For example, if you use NHiTS,...