I am not sure what point the article is trying to make, Lag Llama was already shown unable to outperform even simple benchmarks and is a very bad model.Even in your article it clearly produces prediction intervals that are very bad in terms of the quality both visually and in terms of coverage/interval width.A recent paper have shown generally time series LLMs don’t work

https://arxiv.org/abs/2406.16964