Pick a dataset, pick a model, run. Every forecast and every anomaly score on this page was produced by the actual Granite Time Series models, zero-shot, on a CPU.
Public datasets under open licenses, and seeded synthetic ones with labeled events so detection can be scored.
Choose the task and the model, then run. The values after the origin were never shown to the model; the metrics come from that holdout.
The chart fills in once you choose a stream.
MASE is scaled by the seasonal naive error over the same context: 1.00 means as good as repeating last cycle, lower is better. Coverage is the share of actuals inside the 10 to 90 band.
The models you just ran are the ones Confluent hosts. One statement of Flink SQL, per event, no training loop. Early Access.
Granite TTM r3, FlowState r1.1, PatchTST-FM r2 and TSPulse r1, the open weights on Hugging Face, run through the granite-tsfm library on a CPU runner. No training and no tuning on these datasets.
Each forecast origin sees only the past. The actual values drawn after the origin were held out. MASE compares the error with a seasonal naive forecast over the same context; coverage checks the band against what really happened.
Public datasets are CC BY 4.0 or public domain and are credited below. Synthetic datasets are generated with a fixed seed and carry labeled events, so detection can be scored against ground truth.