IBM Releases SOTA Time Series Forecasting Model

IBM has published a new model called Granite Time Series PatchTST-FM-r2 on Hugging Face — a state-of-the-art (SOTA) time series model released under a commercially friendly license. According to the developers, it's built on the PatchTST architecture and delivers forecasting quality on par with leading industry models.

Models like this are known as time series foundation models: a single pretrained network can forecast very different metrics — sales, demand, inventory levels, cash flow — without retraining from scratch for each task. The open commercial license matters just as much: companies and developers can embed the model into their own products without negotiating usage terms with IBM.

Check which time series your business already collects — daily revenue in your accounting system, warehouse stock levels, incoming payments — and whether anyone actually uses them for forecasting today. Period comparison and early anomaly detection in 1Cifer run on similar logic: the agent spots unusual swings in revenue or payments on its own, catching what's easy to miss manually.