Lowest MASE is 0.683, third among eligible

The visible date on this IBM article at Hugging Face is Published October 22, 2026. The judgment first: this is a recipe that combines four Granite time-series models that were already trained. It is not a new training run. The page says there is no retraining and no fitted ensemble weights. The default probabilistic ensemble gives every member the same influence and pools quantile forecasts in probability space. The IQR ensemble assigns influence by how wide each forecast interval is, with a narrower interval counting for more. Neither recipe fits those weights from historical forecast errors.

The default members are four models with different architectures: PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1, and TTM-r3. The recipe stores an ensemble configuration, not one merged set of weights. The page says the recipe and its default members are permissively licensed. The same page prints an install command and a calling example. This story does not copy the usage, and it does not turn the example parameters into steps.

The evaluation uses 97 GIFT-Eval dataset configurations. The scores and ranks come from the public board captured on 20 October 2026. The page says lower is better, and a star marks the best value among these six forecasts. The probabilistic ensemble has MASE 0.683, average MASE rank 47.175, and CRPS/MWQL 0.471. The IQR ensemble has MASE 0.684, the same CRPS/MWQL of 0.471, and average CRPS rank 47.619, the best of the six in that column. PatchTST-FM-r2 has MASE 0.685 and CRPS/MWQL 0.467, the lowest of the six in that column. FlowState-r1.1, PatchTST-FM-r1, and TTM-r3 have MASE 0.701, 0.717, and 0.724. The page says both ensembles beat all four individual members on aggregate MASE and average MASE rank. The probabilistic ensemble has the lowest MASE of the six, 0.683 against 0.685. The page calls that margin small. The benefit it names is robustness: you do not have to know in advance which architecture will suit the data at hand.

On the probabilistic forecasts, aggregate error and average rank point to different choices. PatchTST-FM-r2 keeps the lowest aggregate CRPS/MWQL. The IQR ensemble has the best average CRPS rank of the six, while the probabilistic ensemble sits behind PatchTST-FM-r2 on that measure. The rank columns compare against the full board, not only these six. If later entries are added, the rank values can change.

In the 20 October 2026 comparison, this ensemble approach ranks third among eligible systems on both MASE rank and CRPS rank. Eligible means replication code is available and no test-data leakage was declared. Fine-tuned entries, and systems whose licenses are not permissive or that depend on such members, are not drawn on that figure. Filtering changes who appears, not their rank on the full board. The page also says some systems score better than either recipe, including ones left off the figure. The value of the ensemble is not that it tops the board. Aggregate gains do not mean every series benefits, and average ranks do not establish calibration.