IBM's Granite Timeseries Ensemble pools four open models and edges each member on GIFT-Eval point-forecast MASE

IBM Research published Granite Timeseries Ensemble R1 on October 22, 2026. It is a recipe that combines four pre-trained Granite time-series models into a single forecast, returning median and quantile predictions with no retraining and no fitted ensemble weights. On 97 GIFT-Eval dataset configurations, the Probabilistic Ensemble scores a 0.683 MASE, narrowly ahead of the best single member at 0.685.

What IBM released

The ensemble is published as ibm-granite/granite-timeseries-ensemble-r1 on Hugging Face. The recipe stores the ensemble configuration, not merged weights. It downloads the four member checkpoints at runtime, so the first run takes longer. IBM says the recipe and its default Granite members are permissively licensed. Loading it requires granite-tsfm>=0.3.11.

The four default members:

Model	Architecture
Granite PatchTST-FM-r1	Patched Transformer
Granite PatchTST-FM-r2	Patched Conformer
Granite FlowState-r1.1	State-space model
Granite TTM-r3	Tiny Time Mixers (MLP mixer)

By default the ensemble returns quantiles from 0.1 to 0.9. The 0.1 and 0.9 quantiles give a central interval with nominal 80% coverage. The interface is extensible to other compatible forecasters.

Two aggregation methods

• Probabilistic Ensemble (recommended default): Gives each member equal influence and pools their quantile forecasts in probability space. There is no fitting step and no weighting hyperparameters.
• IQR Ensemble: Weights members by forecast spread, favoring narrower intervals. Weights vary by sample, target and horizon step. Isotonic regression prevents quantile crossing. It exposes a temperature setting and an optional cap on any single member's weight. IBM warns that IQR weighting can favor an overconfident member.

The IQR recipe loads with subfolder="recipes/granite-iqr".

Benchmark results

IBM evaluated both recipes on 97 GIFT-Eval configurations with no fine-tuning. It says test targets were not used to select or fit the aggregation method. The scores below are published leaderboard values captured October 20, 2026, normalized to Seasonal Naive and geometrically averaged. Lower is better. These are IBM's own submissions and reporting.

Forecast	MASE	MASE Rank	CRPS/MWQL	CRPS Rank
Ensemble R1 (Prob)	0.683	47.175	0.471	50.948
Ensemble R1 (IQR)	0.684	48.773	0.471	47.619
PatchTST-FM-r2	0.685	51.371	0.467	48.918
FlowState-r1.1	0.701	66.474	0.490	69.485
PatchTST-FM-r1	0.717	74.794	0.488	68.134
TTM-r3	0.724	83.062	0.520	88.959

Both ensembles beat all four members on aggregate MASE and average MASE Rank. The probabilistic results are less clear-cut. PatchTST-FM-r2 has the lowest CRPS/MWQL at 0.467, versus 0.471 for both ensembles. The IQR Ensemble has the best CRPS rank at 47.619. Rank columns are averaged against the full GIFT-Eval leaderboard, not just these six entries, so they can shift as new submissions arrive.

The source text was truncated before IBM's discussion of the MASE margin, and the model card's limitations were not part of the provided material. Neither is reflected here.

What this means

The headline MASE gain is small: 0.683 against 0.685 for PatchTST-FM-r2. The more credible case for the recipe is that it removes the need to pick a model in advance. Single-model rankings shift across datasets and horizons, and the ensemble stays near the top of both rankings without per-series tuning.

PatchTST-FM-r2 matching or beating the ensemble on CRPS/MWQL suggests that teams who care mainly about calibrated intervals should benchmark it directly. Runtime cost is also a factor. The recipe loads four checkpoints, so inference compute and latency will exceed a single model's. IBM has not disclosed latency figures in the provided material. The equal-weight default also follows the forecast combination literature, where simple averaging often holds up against estimated weights. That makes the recipe a sensible baseline to beat on your own data.