darts
darts copied to clipboard
Support for zero-shot forecasting using LLM
A recent paper titled "Large Language Models Are Zero-Shot Time Series Forecasters" https://arxiv.org/abs/2310.07820 showed that LLMs have on par if not better forecasting performance compared to SOTA models while performing zero-shot inference. The key is really some simple pre-processing and post processing. It would be cool if DARTs enabled a feature to allow plugging in a LLM for zero-shot forecasting.
Related to #1715.
WDYT @dennisbader ?
Hi @kyleke1998. Indeed related to #1715. I still believe there are other areas which have higher priority to improve first. We are definitely not against it, but this is something where we'd need more capacity/help from the community.
Happy to help to integrate this into darts.
To centralize conversation in a single issue, closing this one, duplicate of #2359.
As mentioned by Dennis, we would prefer to focus our capacity on other features but PRs from the Darts community are welcomed :) And before implementing it, agreeing on a nice architecture would be probably necessary.