· Oct 22, 2026
IBM Research authors present a four-model Granite forecasting ensemble
Markdown · JSON · Original source
IBM Research authors describe Granite Timeseries Ensemble, a recipe that combines four models and returns median and quantile forecasts without retraining or fitted ensemble weights. The recommended Probabilistic Ensemble gives each member equal influence; the alternative IQR Ensemble weights members according to forecast spread, which may favor an overconfident model. In the authors’ GIFT-Eval comparison, both ensembles outperform the four individual members on aggregate MASE, while PatchTST-FM-r2 retains the lowest aggregate CRPS/MWQL score. The authors say the MASE margin over the best single member is small and present the ensemble as a practical starting point, not a leaderboard leader. The recipe can also accept compatible external models, but combining four members increases inference and memory relative to using one. The authors caution that aggregate results do not mean every series benefits, and neither aggregation method guarantees calibration or robustness under . How do the two recipes compare on representative data for a specific application, including quantile loss, empirical coverage, runtime, and memory use?
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Oct 22, 2026 (2026-10-22T07:51:31.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-06T16:31:13.600Z. First observed: 2026-10-22T07:52:11.849Z. Retrieved: 2026-10-22T08:00:27.945Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported evaluation is on GIFT-Eval; the article cautions that aggregate gains do not mean every series benefits and average ranks do not establish calibration. The recipe forecasts multivariate channels independently, without using cross-channel information, and loads four member checkpoints, increasing inference and memory relative to a single member. The article says neither aggregation method guarantees calibration or robustness under ; performance for a particular application remains to be tested.
Open question
On representative data for a specific application, how do the two recipes compare in accuracy, empirical coverage, runtime, and memory use?
Source-stated
IBM Research authors describe a recipe combining four pre-trained Granite models to produce median and quantile forecasts without retraining or fitted ensemble weights.
One recipe combines four complementary Granite architectures and returns median and quantile forecasts, with no retraining and no fitted ensemble weights.
Source-stated
The Probabilistic Ensemble gives each model equal influence; the IQR Ensemble weights models based on forecast spread and can favor an overconfident member.
The IQR Ensemble weights members according to their forecast spread, giving narrower intervals greater influence.
Source-stated
In the reported GIFT-Eval comparison, both ensembles outperform the four individual members on aggregate MASE, but PatchTST-FM-r2 has the lowest aggregate CRPS/MWQL score.
Both ensembles outperform all four individual members on aggregate MASE and average MASE Rank. The Probabilistic Ensemble achieves the lowest MASE among these six forecasts
Source-stated
The article says combining four models increases inference time and memory compared with a single member, and that aggregate benchmark gains do not mean every series benefits.
The current recipe forecasts multivariate channels independently, without using cross-channel information, and it loads four member checkpoints, which increases inference time and memory relative to a single member.
Source-stated
The article recommends evaluating both recipes on representative data and checking accuracy, calibration, runtime, and memory for the intended application.
The way to find out how the recipes behave for your application is to backtest them on representative data:
Source and contributor context
Hugging Face · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Pre-training
A model's initial training on a large dataset, before further adaptation for particular uses.
Time-series
Observations arranged over time, such as monthly sales; forecasting uses their patterns to estimate later values.
Distribution shift
A change between the data a model was trained on and the data it meets in use, which can lower its accuracy.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: b6b9ad3d-f0ff-4d64-841d-4cf3901505e5.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-06T16:31:13.600Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 985a70f91ae142c4b4bee4ba81a90f3bdd965e80bfd6a71efe87791a2a487540. Source text hash: c8a2938c6e09618f46092f9aed553537015e2695683429ddacd02842426154b2.
Revision and source-check history
Version 1 · 2026-11-06T16:31:13.600Z
IBM Research authors present a four-model Granite forecasting ensemble
IBM Research authors describe Granite Timeseries Ensemble, a recipe that combines four models and returns median and quantile forecasts without retraining or fitted ensemble weights. The recommended Probabilistic Ensemble gives each member equal influence; the alternative IQR Ensemble weights members according to forecast spread, which may favor an overconfident model. In the authors’ GIFT-Eval comparison, both ensembles outperform the four individual members on aggregate MASE, while PatchTST-FM-r2 retains the lowest aggregate CRPS/MWQL score. The authors say the MASE margin over the best single member is small and present the ensemble as a practical starting point, not a leaderboard leader. The recipe can also accept compatible external models, but combining four members increases inference and memory relative to using one. The authors caution that aggregate results do not mean every series benefits, and neither aggregation method guarantees calibration or robustness under . How do the two recipes compare on representative data for a specific application, including quantile loss, empirical coverage, runtime, and memory use?
Original source
· Article
Source details
Original source date: Oct 22, 2026 (2026-10-22T07:51:31.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-06T16:31:13.600Z. First observed: 2026-10-22T07:52:11.849Z. Retrieved: 2026-10-22T08:00:27.945Z.
Limits
The reported evaluation is on GIFT-Eval; the article cautions that aggregate gains do not mean every series benefits and average ranks do not establish calibration. The recipe forecasts multivariate channels independently, without using cross-channel information, and loads four member checkpoints, increasing inference and memory relative to a single member. The article says neither aggregation method guarantees calibration or robustness under ; performance for a particular application remains to be tested.
Open question
On representative data for a specific application, how do the two recipes compare in accuracy, empirical coverage, runtime, and memory use?
Source-stated
IBM Research authors describe a recipe combining four pre-trained Granite models to produce median and quantile forecasts without retraining or fitted ensemble weights.
One recipe combines four complementary Granite architectures and returns median and quantile forecasts, with no retraining and no fitted ensemble weights.
Source-stated
The Probabilistic Ensemble gives each model equal influence; the IQR Ensemble weights models based on forecast spread and can favor an overconfident member.
The IQR Ensemble weights members according to their forecast spread, giving narrower intervals greater influence.
Source-stated
In the reported GIFT-Eval comparison, both ensembles outperform the four individual members on aggregate MASE, but PatchTST-FM-r2 has the lowest aggregate CRPS/MWQL score.
Both ensembles outperform all four individual members on aggregate MASE and average MASE Rank. The Probabilistic Ensemble achieves the lowest MASE among these six forecasts
Source-stated
The article says combining four models increases inference time and memory compared with a single member, and that aggregate benchmark gains do not mean every series benefits.
The current recipe forecasts multivariate channels independently, without using cross-channel information, and it loads four member checkpoints, which increases inference time and memory relative to a single member.
Source-stated
The article recommends evaluating both recipes on representative data and checking accuracy, calibration, runtime, and memory for the intended application.
The way to find out how the recipes behave for your application is to backtest them on representative data:
Source and contributor context
Hugging Face · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Pre-training
A model's initial training on a large dataset, before further adaptation for particular uses.
Time-series
Observations arranged over time, such as monthly sales; forecasting uses their patterns to estimate later values.
Distribution shift
A change between the data a model was trained on and the data it meets in use, which can lower its accuracy.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: b6b9ad3d-f0ff-4d64-841d-4cf3901505e5.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-06T16:31:13.600Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 985a70f91ae142c4b4bee4ba81a90f3bdd965e80bfd6a71efe87791a2a487540. Source text hash: c8a2938c6e09618f46092f9aed553537015e2695683429ddacd02842426154b2.
uncertain · 2026-11-06T16:31:13.600Z · anthropic/claude-opus-5.5
Publication: b6b9ad3d-f0ff-4d64-841d-4cf3901505e5
attribution — pass: The draft correctly attributes the work and claims to IBM Research authors writing on Hugging Face. Benchmark results stay framed as the authors' comparison, not as independently verified facts.
We evaluated both recipes on 97 GIFT-Eval dataset configurations.
summary — uncertain: The summary is faithful but hard to follow for a general reader. MASE and CRPS/MWQL are central to the results paragraph, yet the draft never says the first measures point-forecast error and the second measures probabilistic forecasts. Without that, the split between the ensembles and PatchTST-FM-r2 is opaque. The draft also drops the authors' main rationale: robustness without choosing a model in advance. Smallest repair: one plain-language clause on what each metric measures, plus the robustness point. The body also repeats the question verbatim.
The MASE margin over the best single member is small, so the main benefit is robustness: you get comparable or better aggregate accuracy without knowing in advance which architecture will suit your data.
date — pass: The originalPublished value of 2026-10-22 matches the article's stated publication date of October 22, 2026.
Published October 22, 2026
support — pass: The source supports all claims and limitations: the four-member recipe, both weighting methods, the overconfidence caveat, the MASE and CRPS results, the not-top-of-leaderboard framing, the inference cost, and the calibration and distribution-shift caveats. The question follows from the authors' own backtesting advice.
Some systems, including entries excluded as fine-tuned or non-permissively licensed, score better than either recipe, so the value of the ensemble is not that it tops the leaderboard.
media — pass: The selected media is the verified link to the original Hugging Face article, which is the most direct way to reach the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceA model's initial training on a large dataset, before further adaptation for particular uses.
Definition referenceObservations arranged over time, such as monthly sales; forecasting uses their patterns to estimate later values.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA change between the data a model was trained on and the data it meets in use, which can lower its accuracy.
Definition referenceHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA change between the data a model was trained on and the data it meets in use, which can lower its accuracy.
Definition referenceA model's initial training on a large dataset, before further adaptation for particular uses.
Definition referenceObservations arranged over time, such as monthly sales; forecasting uses their patterns to estimate later values.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA change between the data a model was trained on and the data it meets in use, which can lower its accuracy.
Definition referenceHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA change between the data a model was trained on and the data it meets in use, which can lower its accuracy.
Definition reference