The prediction is simply to repeat the last value

[korosig]

I'm trying to implement the package you created. I have to say it is very transparently written and easy to read.

However, I have a strange observation. The model has not been able to learn anything on any of the datasets I have tested (the sample examples on the internet* have the same problem ), it only repeats the last known information. So if you want to predict at time t, the model returns the value at time t-1.

My question is, did it work differently for you? Was it really able to learn anything from the time series?

In my test, I want to predict the traffic speed of the sample task https://keras.io/examples/timeseries/timeseries_traffic_forecasting/ per NODE using the NBeast model.

*- https://medium.com/@jonathanbechtel/kerasbeats-an-easy-way-to-use-n-beats-in-keras-395b24c5cc28

[JonathanBechtel]

@korosig Thank you for writing. I'll look into it.