https://github.com/nixtla logo
a

Afiq Johari

08/07/2023, 1:52 AM
Are there recommended feature selection packages that work well with neuralforecast? I'd like to use NHITS with exogenous variables, although the challenge is more about figuring out the top variables that help increase the NHITS accuracy.
v

Viet Yen Nguyen

08/07/2023, 6:27 AM
We’ve been using granger forecastability tests to that end.
a

Afiq Johari

08/07/2023, 6:41 AM
@Viet Yen Nguyen any specific library you found useful? or is it the one implemented in
statsmodel
library?
v

Viet Yen Nguyen

08/07/2023, 6:43 AM
We use the one in statsmodel too. But it’s also not so difficult to implement one yourself to suit specific needs.
🙏 1
@Afiq Johari just out of curiosity: what kind of data are you working on?
a

Afiq Johari

08/07/2023, 6:48 AM
I'm working on monthly time interval, mostly internal business data as my target and economics data as my features.
👍 2
m

Mark

08/07/2023, 12:12 PM
me too. You working with very sparse/intermittent data?
p

Phil

08/07/2023, 4:42 PM
Have you ever worked with Tsfresh? https://tsfresh.readthedocs.io/en/latest/ I've used it before for automatic feature extraction
a

Afiq Johari

08/07/2023, 4:42 PM
@Mark for the time being, my data is not really sparse/intermittent
m

Mark

08/07/2023, 4:43 PM
ah ok. @Phil good results with it?
p

Phil

08/07/2023, 4:44 PM
It worked really well for my use case. I was training a metalearning model which predicted which timeseries forecasting model would give me the best MAPE forecast based on feature of the input timeseries. https://robjhyndman.com/papers/fforms.pdf
👍 3
m

Mark

08/07/2023, 5:04 PM
@Monica Scott