Unlocking Time Series Power with LLMs: A New Approach

Imagine you have a powerful language model that's great at understanding text, but you want it to handle time series data. That's where this new method, called TEST, comes in. TEST doesn't try to change the language model itself. Instead, it changes how time series data is fed into the model. It first breaks the time series into smaller pieces, like words in a sentence. Then, it uses a special encoding method to make the time series data look more like text to the model. This way, the language model doesn't need to be retrained or fine-tuned.

The method even creates soft prompts to help the model understand the new data better. Experiments showed that this approach worked well with different language models. It could sometimes outperform or match the best current time series models. Plus, it worked well even when there wasn't much data. This method treats the language model like a pattern machine, letting it process time series data without losing its language skills.

Actions