IBM's Granite ensemble combines four forecasters, MASE 0.683

IBM Research published Granite Timeseries Ensemble r1 on 22 October 2026. IBM says it is a single recipe that combines four pre-trained Granite time-series models and returns median and quantile forecasts without retraining on your data. IBM says the recipe and its default Granite members are permissively licensed.

What is in it, per IBM

The default recipe pools PatchTST-FM-r1 (a patched transformer), PatchTST-FM-r2 (a patched conformer), FlowState-r1.1 (a state-space model) and TTM-r3 (a tiny time mixer, an MLP-based model). IBM's argument is that different architectures make different errors, so combining them reduces the risk of picking the wrong single model for a given series.

There are two ways to combine them. The recommended default, the Probabilistic Ensemble, gives each member equal influence and pools their quantile forecasts, so there are no weights to tune. The alternative, the IQR Ensemble, gives narrower forecast intervals more influence, which IBM warns can favour an overconfident member. The recipe stores the configuration, not merged weights, and downloads the four checkpoints at runtime.

What IBM reports

IBM evaluated both recipes on 97 GIFT-Eval dataset configurations without fine-tuning, using leaderboard scores it captured on 20 October 2026. On MASE, where lower is better, the Probabilistic Ensemble scores 0.683 and the IQR Ensemble 0.684, against 0.685 for PatchTST-FM-r2, 0.701 for FlowState-r1.1, 0.717 for PatchTST-FM-r1 and 0.724 for TTM-r3.

The probabilistic measure tells a different story. PatchTST-FM-r2 has the lowest aggregate CRPS/MWQL, 0.467 against 0.471 for both ensembles. The IQR Ensemble has the best average CRPS rank of the six forecasts, and the Probabilistic Ensemble ranks behind PatchTST-FM-r2 on that rank. IBM says the MASE margin over the best single member is small, so the benefit is robustness rather than accuracy.

IBM also says the ensemble ranks third among eligible systems on both MASE rank and CRPS rank, and that some systems, including fine-tuned and non-permissively licensed entries, score better than either recipe.

What the post leaves out

The authors are IBM Research staff evaluating IBM models, and I found no independent replication. IBM lists its own limits: the recipe forecasts multivariate channels independently, loads four checkpoints (more inference time and memory than one model), and neither aggregation method guarantees calibration under distribution shift. It gives no latency or memory figures. Expanded recipes that add other IBM Research checkpoints carry non-commercial or research-use terms, so check each member's model card.

What to do

• If you forecast demand, load or capacity, backtest the Probabilistic Ensemble, the IQR Ensemble and PatchTST-FM-r2 on the same historical windows and forecast origins you will serve.
• Compare quantile loss as well as MASE, and check that the 0.1 to 0.9 interval actually covers about 80% of outcomes, as IBM suggests.
• Record runtime and memory for the four-model recipe before deciding that a gain of a few thousandths of MASE is worth it.