D. Thomakos
A very quick, but I think useful, draft of some questions that I asked ChatGPT about foundational issues for time series analysis and forecasting. Have a look at the pdf below and comments are much welcomed. The gist is this, to my humble opinion:
(a) LLMs are useful but their defining characteristic is the second L=language so their immediate application (with or without transformers etc etc.) to time series forecasting will not provide immediate forecasting performance gains unless specific conditions on the data generating process are met (repeated patterns).
(b) The application of tokenization to time series analysis is embedding and the application of rolling windows in ttime series modeling and as such is not a new idea.
(c) Attention-based models for forecasting are a particular way to apply time-varying parameters in modeling, by no means the only one available.
(d) No model can, obviously, capture the data generating process and in particular when this process changes over time - this includes LLMs all their derivatives.
