On a large scale of 42 time series datasets, Chronos demonstrates impressive empirical performance. In the zero-shot setting, it matches or even outperforms many baselines which are trained on the dataset.

this is an amazing contribution for the community!

TL;DR:

* Scale the time series data and quantize the floating point values into B bins.

* Each bin becomes a corresponding token id in a vocabulary of B embeddings.

* Train a small LLM to predict the next token id given a sequence of token ids.

* At each time step, the LLM gives you a probability distribution over B bins.