Time-LLM

Why we included this project

Most forecasting tooling assumes you train a purpose-built network. Time-LLM takes a different route: it keeps a large language model frozen and turns time series patches into text-like tokens the model can read and project forward, borrowing the LLM's pretrained reasoning for the forecast. This is the official ICLR 2024 implementation, and the code is straightforward enough that a team can benchmark it against classical baselines without digging through undocumented internals. The paper's ablations also make it fairly clear where the gains come from, which helps when you are deciding whether the added complexity is worth it. Since the method has already been folded into libraries like NeuralForecast and PyPOTS, you can try it inside a maintained forecasting stack instead of running the standalone repo.