For October's journal club Andrew Smith will be exploring the idea of foundation models for time series forecasting. Foundation models have been hugely successful for Natural Language Processing (NLP), but do they make sense for forecasting across the different domains where time series forecasting is applied? What would they learn that is applicable across domains when you have 100 billion time series data points to train from?

Motivating this discussion is evaluating if these are forecasting tools that we need to pay "attention" to (pun intended).

The paper (most) under discussion is:

A decoder-only foundation model for time-series forecasting, Abhimanyu Das, Weihao Kong, Rajat Sen, Yichen Zhou,  International Conference on Machine Learning 2024.
https://arxiv.org/abs/2310.10688

with attached blog:

https://research.google/blog/a-decoder-only-foundation-model-for-time-series-forecasting/

Please remember that as Monday is a public holiday that journal club is on the first work day of the week. I hope to see you all there, and apologies for the late notice.