Forecasting with xreg=snaive

It is just that ambiguity that a minimal reproducible example reprex is designed to clarify. Come back with a fully specified problem that includes all objects, starting with training. Less reverse engineering of the problem offers the possibilities of more attempts at an answer.