Google secretly opened sourced TimesFM 2.5, a time series prediction base, with parameters reduced to 200M and context skyrocketed to 16K. By throwing in historical values, zero shot predictions can be made directly, and family members can make cryptocurrency predictions, haha~ TimesFM is a pre trained time series prediction base model. If you feed in the historical numerical sequence, it can first run a zero shot prediction for you, and the future trend and quantile range can be spit out together. And the latest 2.5 version is quite tough: the model has been reduced from 500M to 200M, the context window has been pulled up to 16K, and even the original frequency indicator has been removed, making the interface more user-friendly than before. More importantly, this repository is no longer just "paper supporting code": the official has also added XReg covariate support, Transformers versions, LoRA fine-tuning examples and tests, and its practicality is clearly moving towards engineering. I think it is particularly suitable for creating a baseline for tasks such as business sales, website traffic, server load, and demand forecasting. Warehouse comment section