I suspect the weakness of time-series models (transformers, n-hits, etc.) is less their architecture, but more the way we present data to them. We too often assume the obstacle to time-series forecasting is trend analysis, and make it a technical problem rather than our own knowledge of what factors influence the present observations.