ModelsArchitectures & capability
IBM Research releases Granite Timeseries Ensemble R1, combining four open forecasting models without retraining
IBM Research's Granite Timeseries Ensemble R1 pools four open forecasting models behind one interface without retraining. On GIFT-Eval it edges out the best single member on point error, but that member still leads on probabilistic error.
From the Hugging Face article of October 22, 2026 and cited sources: The ensemble combines four architecturally different Granite members through one interface: PatchTST-FM-r1 (Transformer), PatchTST-FM-r2 (Conformer), FlowState-r1.1 (state-space model) and TTM-r3 (MLP mixer). [1] [6]
SOURCES[1] Ensemble forecasting with IBM Granite time-series models huggingface.co[6] Model card: granite-timeseries-ensemble-r1 huggingface.co
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- The ensemble combines four architecturally different Granite members through one interface: PatchTST-FM-r1 (Transformer), PatchTST-FM-r2 (Conformer), FlowState-r1.1 (state-space model) and TTM-r3 (MLP mixer). [1] [6]
- Original source
- Hugging Face article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 22, 2026 article published
- Covered
- Oct 22 edition
- Evidence
- 9 sources across 7 sites
- 1Ensemble forecasting with IBM Granite time-series modelshuggingface.co
- 2GIFT-Eval (NeurIPS workshop 2024)mlanthology.org
- 3Salesforce: GIFT-Eval benchmarksalesforce.com
- 4IBM Granite time-series docsibm.com
- 5IBM blog: PatchTST-FM-r2 releasehuggingface.co
- 6Model card: granite-timeseries-ensemble-r1huggingface.co
Reach: not measured yet
The ensemble combines four architecturally different Granite members through one interface: PatchTST-FM-r1 (Transformer), PatchTST-FM-r2 (Conformer), FlowState-r1.1 (state-space model) and TTM-r3 (MLP mixer). The default recipe gives each member equal weight and pools their quantile forecasts with nothing fitted. An alternative recipe weights each member by how narrow its interquartile range is, which IBM acknowledges can favour an overconfident member. The recipe is Apache-2.0, but each member keeps its own license, and the expanded variants add non-commercial terms. [1] [6]
In IBM's zero-shot GIFT-Eval results across 97 configurations, the two recipes score 0.683 and 0.684 MASE, against 0.685 for PatchTST-FM-r2, the strongest member. On CRPS, PatchTST-FM-r2 still leads at 0.467 versus 0.471 for both ensembles. All of these figures are vendor-reported. [1] [5]
IBM's own documents disagree on the leaderboard standing. The blog places the ensemble third among eligible systems, while the model card says top four and lists different average rank values. Ranks are relative to the full leaderboard and shift as new entries are added. [1] [6]
IBM's case for the ensemble is robustness rather than higher accuracy: users don't have to guess which architecture suits their data. The trade-off is the latency and memory of running four models, and multivariate series are forecast one channel at a time, so links between channels go unused. [1] [6]
The evidence shows accuracy almost identical to the best single model. The practical upside is that teams can skip choosing a model for each dataset under commercial terms.
Read the full assessment
In exchange, they pay for four models' inference and should check interval coverage on their own data.
Executive brief
IBM Research has released Granite Timeseries Ensemble R1. It combines four open Granite forecasting models through one interface, with no retraining. On the benchmark it reports, the gain over IBM's best single model is very small: the default ensemble scores 0.683 MASE (a point-forecast error metric) against 0.685 for PatchTST-FM-r2, the strongest member. On probabilistic error (CRPS), that single model still has the best aggregate score (IBM blog). IBM argues the value is robustness: you don't have to guess which architecture suits your data. The trade-off is running four models, which costs more inference time and memory. No independent evaluation of the ensemble was found.
What changed and event timeline
GIFT-Eval benchmark published
Salesforce researchers released a zero-shot forecasting benchmark covering 28 datasets and more than 144,000 series. It is now the reference leaderboard for time-series foundation models (;).
FlowState places #3 on GIFT-Eval
IBM's state-space model, under 10M parameters, was ranked #3 on the leaderboard and described as the best zero-shot entry at the time ().
PatchTST-FM-r2 released
This is a Conformer model with about 385M parameters. IBM reports CRPS of 0.467 and ranks it top among permissively licensed zero-shot models ().
Leaderboard snapshot captured
IBM takes the GIFT-Eval scores and ranks used to compare the ensemble with its members from this date ().
Ensemble R1 published
The release includes the recipe, model card and two aggregation methods, loadable with
QuantileEnsembleForecaster.from_pretrained(;).
Capabilities and access
- Repo:
ibm-granite/granite-timeseries-ensemble-r1. The repo stores the ensemble configuration only; the four member checkpoints download at runtime. - Members: PatchTST-FM-r1 (Transformer), PatchTST-FM-r2 (Conformer), FlowState-r1.1 (state-space model) and TTM-r3 (MLP mixer) (model card).
- Output: a point forecast plus quantiles from 0.1 to 0.9 by default.
Read the full section
- Repo:
ibm-granite/granite-timeseries-ensemble-r1. The repo stores the ensemble configuration only; the four member checkpoints download at runtime. - Members: PatchTST-FM-r1 (Transformer), PatchTST-FM-r2 (Conformer), FlowState-r1.1 (state-space model) and TTM-r3 (MLP mixer) (model card).
- Output: a point forecast plus quantiles from 0.1 to 0.9 by default.
- Recipes:
granite-linear(the default) andgranite-iqr. Theexpanded-*variants add IBM Research checkpoints that carry non-commercial or research-only terms. - License: Apache-2.0 for the recipe; each member checkpoint keeps its own license.
- Version conflict: the blog says to install
granite-tsfm>=0.3.11, while the model card says ≥0.3.10.
Technical analysis for researchers and developers
- Probabilistic Ensemble (default): each member gets equal weight. Nothing is fitted.
- IQR Ensemble: weights each member by how narrow its predicted interquartile range (25th to 75th percentile) is, with weights varying by sample, target and horizon step.
- Evaluation: 97 GIFT-Eval configurations, zero-shot. Scores are normalised to a Seasonal Naive baseline and geometrically averaged.
Read the full section
- Probabilistic Ensemble (default): each member gets equal weight. Their quantile forecasts are pooled in probability space and the requested quantiles are read from the pooled result. Nothing is fitted.
- IQR Ensemble: weights each member by how narrow its predicted interquartile range (25th to 75th percentile) is, with weights varying by sample, target and horizon step. Isotonic regression stops the combined quantiles from crossing. A temperature setting pulls weights toward uniform, and an optional cap limits any one member's weight. IBM notes this can favour an overconfident member (IBM blog).
- Evaluation: 97 GIFT-Eval configurations, zero-shot. Scores are normalised to a Seasonal Naive baseline and geometrically averaged. Test targets were not used to choose the aggregation method.
- Multivariate data: each channel is forecast independently, so information across channels is not used.
Claims and evidence
All figures are IBM-reported, using IBM's own eligibility filter (it excludes fine-tuned entries and non-permissively licensed systems). Coverage on daily.dev restates the blog; it says the two ensembles rank first and second on MASE rank, which differs again from the blog's wording.
Read the full section
| Claim | Evidence |
| Both ensembles beat all four members on MASE and on average MASE rank | IBM table: 0.683 / 0.684 vs 0.685–0.724 (blog) |
| PatchTST-FM-r2 remains best on CRPS | 0.467 vs 0.471 for both ensembles (same source) |
| The ensemble ranks "3rd among eligible systems" | Blog. The model card instead says "top four", with ranks of 37.19 (MASE) and 36.90 (CRPS) (model card), versus 47.175 and 47.619 in the blog. The two documents conflict. |
All figures are IBM-reported, using IBM's own eligibility filter (it excludes fine-tuned entries and non-permissively licensed systems). Coverage on daily.dev restates the blog; it says the two ensembles rank first and second on MASE rank, which differs again from the blog's wording.
Context and prior work
- Combining forecasts has a long track record.
- The idea mirrors multi-model ensembles in weather forecasting.
- The members come from IBM's run of lightweight forecasting models: TTM, FlowState (IBM Research blog) and the larger PatchTST-FM-r2.
Read the full section
- Combining forecasts has a long track record. IBM cites a 50-year review of forecast combinations and the "forecast combination puzzle" (Smith & Wallis, 2009): simple equal weights often match or beat estimated weights (IBM blog).
- The idea mirrors multi-model ensembles in weather forecasting.
- The members come from IBM's run of lightweight forecasting models: TTM, FlowState (IBM Research blog) and the larger PatchTST-FM-r2.
- The original GIFT-Eval study found PatchTST to be a strong generalist (Moonlight review).
Limitations, safety and contested findings
- Small, mixed gains: the MASE margin over the best member is 0.002, and on CRPS the best member still wins.
- Ranks can drift: rank values are relative to the full leaderboard, so they shift as new entries are added.
- Stronger systems exist: some fine-tuned or non-permissively licensed entries outscore both recipes.
Read the full section
- Small, mixed gains: the MASE margin over the best member is 0.002, and on CRPS the best member still wins.
- Ranks can drift: rank values are relative to the full leaderboard, so they shift as new entries are added. IBM's own documents already report different rank numbers.
- Stronger systems exist: some fine-tuned or non-permissively licensed entries outscore both recipes.
- No guarantees: neither method guarantees calibration or robustness under distribution shift, and members differ in context limits and missing-value handling (model card).
Business and practitioner implications
- A reasonable default for zero-shot forecasting under a commercial license, especially where teams can't afford to pick a model for every series.
- Cost: budget for four models' latency and memory. A single PatchTST-FM-r2 may be a better deal when CRPS is the target metric.
- Backtest on your own data: compare both recipes and PatchTST-FM-r2 on the same historical windows, and check that the 0.1–0.9 interval actually covers about 80% of outcomes.
Read the full section
- A reasonable default for zero-shot forecasting under a commercial license, especially where teams can't afford to pick a model for every series.
- Cost: budget for four models' latency and memory. A single PatchTST-FM-r2 may be a better deal when CRPS is the target metric.
- Backtest on your own data: compare both recipes and PatchTST-FM-r2 on the same historical windows, and check that the 0.1–0.9 interval actually covers about 80% of outcomes.
- Check licenses before switching recipes: the
expanded-*recipes add non-commercial terms. - Extensible: the Forecaster interface accepts compatible external models without changing the output format.
The source trail.
Sources (9)
Ensemble forecasting with IBM Granite time-series models
Article text retrieved; extracted text may omit tables or interactive elements.
huggingface.co