Can’t really recall the reference for this point (if any). But you’ll discover the point yourself once you start implementing any multivariate time series model.

Consider this scenario — your model requires multiple input to make it’s prediction. Since the model is not auto-regressive in nature (ie, automatically uses it’s past value (ie t-1) to make future (t+1) prediction), it will essentially need to rely on whatever input that it has at time t.

To circumvent this problem, you can make multiple forecast models, one for each variable that you have. With the predicted values of each of the models, you can then create another model that leverages all of the predicted values from each individual input to continue predicting at t+1 continuously into the future.