Comment by Jgrubb
14 hours ago
Could you elaborate? because that sentence made my brow wrinkle with confusion. I have thought to myself before that all business data problems eventually become time series problems. I'd like to understand your point of view on how LLMs fit into that.
Time series just means that the order of features matter. Feature 1 occurs before feature 2.
E.g, fitting a model to house prices, you don’t care if feature 1 is square meters and feature 2 is time on market, or vice versa, but in a time series, your model changes if you reverse the order of features.
With text, the meaning of word 2 is dependent on the meaning of word 1. With stock prices, you expect the price at time 2 to be dependent on time 1.
Text can be modeled as a time series.
A language model tells you the next character/token/word depending on the previous input.
Language models are time series.
It’s not an audacious claim.
Any student of nlp should have met a paper modeling text as time series before writing their thesis. How could you not meet that?
As a data structure it is an ordered list of integers but no LLM needs to accès it in a database, it's way to slow for anything serious.
RAG and vector Approximate Nearest Neighbour (ANN) is the the go to use case.
[1] https://towardsdatascience.com/llm-powered-time-series-analy...
[2] https://arxiv.org/abs/2506.02389
[3] https://arxiv.org/html/2402.10835v3
Some links from the top of Google search.
Take a look here, also, it's an important law: https://en.wikipedia.org/wiki/Benford%27s_law
It is possible for LLMs to learn Bernford's law, implicitly. So they will be non-null predictors of time series data, because time series data is also Bernford-law-distributed [4].
[4] https://ui.adsabs.harvard.edu/abs/2017EGUGA..19.2950T/abstra...