Anyone has looked into using methods like

SHAP

or

LIME

with models like `NHITS`to explain the model? Or, are there any new, advanced techniques that help us understand how `exogenous variables`(

future

,

historical

,

static

) affect a

unique_id

time series forecast? As we aim for better accuracy, it's increasingly important to explain how these

exogenous variables

and a

unique_id

past values impact the future forecasts. Also, a way to quantify how much of the forecasts could be attributed to unexplained/random errors?

Hi @Afiq Johari, we dont have this methods available in the library yet. If you want to understand the effect of exogenous variables I recommend using the

NBEATSx

method.