It’s not clear to me even the theoretical motivation behind thinking a zero-shot time series model could ever work better than a model trained on the data. It seems like wishful thinking to make an analogy between this model and large language models which are built on a preexisting grammar and vocabulary. Time series just as tabular data can represent an infinite variety of natural (and unnatural) phenomena. Then again any time series can be represented by a few sine waves so what do I know haha.