how Darts Time Series forecasting library deals with covariates under the surface? why is the prediction not changing with the change of covariates?

Question

I am trying to understand why the model's prediction is independent of the changes of covariates. Given any random target series : y_train_trans and past_Covariate series : X_train_trans , i am finding that pred1 and pred 2 always end up exactly the same no matter how i change X_train_trans.

mymodel = XGBModel(lags=3, lags_past_covariates=2,
                 output_chunk_length=16,max_depth =5)

mymodel2 = XGBModel(lags=3, lags_past_covariates=2,
                 output_chunk_length=16,max_depth =5)

mymodel.fit(y_train_trans,past_covariates=X_train_trans)
mymodel2.fit(y_train_trans,past_covariates=X_train_trans*5)

pred1 = mymodel.predict(5,past_covariates=X_train_trans)
pred2 = mymodel2.predict(5,past_covariates= X_train_trans*5)

print(pred1-pred2) # always gives 0

Answer 1

This is probably because XGBoost is invariant to scaling features here. Try changing the actual shape of the covariates series (rather than simply scaling) and the results could be different.