https://github.com/nixtla logo
#neural-forecast
Title
# neural-forecast
s

Simon Weppe

04/14/2023, 5:29 AM
Hi all, quick question. In the TCN model, if I followed well there is an assumption that past values of the target are used as predictors correct ? I was wondering if it could adapt to a form where the TCN provide prediction of the target, based only on historic covariates for example. Thank you very much !
c

Cristian (Nixtla)

04/14/2023, 12:06 PM
Hi @Simon Weppe. It is not possible right now to only use information from the covariates. This is because during training the loss is computed with the values on the target variable
y
, but there is no way of disabling its information in the inputs.
You could modify the
forward
step of the models and remove the historic values of
y
(when they are concatenated with the historic covariates), or replace its values to 0.
k

Kin Gtz. Olivares

04/14/2023, 3:03 PM
Is there a reason why you may not want or have access to the autorregressive features? A suggestion in case you don't have access to the autorregressive feature is to run a simple regression model against the covariates. Note that not having autorregressives could potentially undermine the effectiveness of most forecasting methods.
s

Simon Weppe

04/16/2023, 10:47 PM
Thanks for the suggestions. my use case was feeding the TCN model with decomposed components of my target (using VMD) as covariates, rather than the target itself. I'll have a play to see if I can get it working.