TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
IBM Research has published Granite Timeseries Ensemble, an open recipe that combines forecasts from four pretrained Granite models. Its default method gives each model equal influence; an alternative adjusts weights using forecast spread. The supplied source describes an evaluation across 97 GIFT-Eval configurations but ends before giving results, so it does not establish whether the ensemble outperformed individual models.
IBM Research has published Granite Timeseries Ensemble, an open recipe for combining forecasts from four pretrained Granite time-series models without retraining them or fitting ensemble weights, building on IBM’s Granite time-series models. The tool offers a fixed, equal-influence default and an adaptive alternative, giving forecasting users a way to combine model outputs rather than choose a single model for every dataset.
The default, called the Probabilistic Ensemble, pools the members’ quantile forecasts with equal influence and produces a combined forecast distribution. The recipe has no fitted weights or weighting hyperparameters to tune for a new series. Users can retrieve a point forecast and quantiles; the defaults range from the 0.1 to the 0.9 quantile, which can be used to express forecast ranges.
The second option, the IQR Ensemble, changes each member’s influence based on its predicted spread. It gives greater weight to forecasts with narrower estimated interquartile ranges, with weights varying by sample, target and forecast horizon. The source says isotonic regression is used to prevent combined quantiles from crossing. IBM cautions that a narrower interval does not necessarily mean a more accurate or better-calibrated forecast.
The default recipe combines PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1 and TTM-r3, which use different model architectures, following an approach explored in combining Granite time-series models for forecasting. The Hugging Face example loads the ensemble through the Granite TSFM package and applies it to the latest 1,024 observations to forecast 24 steps. The configuration is stored separately from the model weights, so the four member checkpoints download at runtime and the first run may take longer.
A Default for Combining Forecasts
Forecast users often face a practical choice: select one pretrained model for a particular time series, or spend time testing alternatives. IBM’s recipe offers a ready-made combination of four models, with a default that requires no training or fitting of ensemble weights on the user’s data. That may make it easier to try a combined forecast in settings where series differ in noise, trend or seasonality.
The approach does not make the choice disappear entirely. Users still need to decide whether the equal-influence default or the spread-based alternative suits their use case, and should test results on representative data. In particular, the IQR method can favor a model that is confident but inaccurate. Forecast intervals and point predictions should be checked against observed outcomes before being relied on for operational decisions.
GPU workstation for time series forecasting machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Four Models, Two Pooling Rules
Pretrained time-series models can produce forecasts without being trained on a user’s particular dataset, but their performance may vary across series and forecast horizons. The report presents ensembling as a way to draw on models with different architectures and potentially different errors. It compares the idea to multi-model weather forecasting, while noting that the benefit depends on the quality and diversity of the members.
The two recipes pool probabilistic forecasts, rather than combining newly fitted model weights. Probabilistic pooling gives each member equal influence; IQR weighting adapts influence to forecast spread. The report describes both as requiring no fitting from historical forecast errors. It also says the recipe and default members use permissive licenses, and directs users to the model card for configuration, evaluation details and limitations.
““Start with the Probabilistic Ensemble.””
— IBM Research, in the Hugging Face report
NVIDIA GPU for running pretrained AI models locally
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmark Results Are Not Provided
The supplied report excerpt says IBM evaluated both recipes on 97 GIFT-Eval dataset configurations, but it cuts off before giving scores, comparisons with the individual members, or details of the findings. On the material available here, it is not possible to say whether either ensemble improved forecast accuracy or calibration, or by how much. The excerpt also does not specify the evaluation’s full methodology or results.
More broadly, the best recipe for a particular application remains uncertain until tested on the relevant series and forecast horizon. The IQR method’s spread-based weighting is not evidence by itself that its intervals are better calibrated. Users should treat the forecasts as model outputs to evaluate, not as guarantees of future outcomes.
external SSD for machine learning model checkpoints
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Test the Recipes on Local Data
Users can load the recommended recipe from the Hugging Face model repository using Granite TSFM version 0.3.11 or later, according to the report. They can then compare its predictions with the IQR recipe by changing the recipe subfolder. IBM says users interested in adaptive weighting should compare it with the default on representative data, checking both forecast accuracy and calibration across the horizons that matter to them.
The next useful evidence would be the complete evaluation results and methodology described in the model card, alongside independent tests on application-specific datasets. Until those details are available, the announcement establishes how the recipes work and how to run them, but not a general performance advantage over individual Granite models.
USB-C hub for AI development workstation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did IBM Research release?
IBM Research published Granite Timeseries Ensemble, a recipe and interface that combines forecasts from four pretrained Granite time-series models.
Does the ensemble retrain the models?
No. The report says the recipe combines pretrained model forecasts without retraining and does not fit ensemble weights from historical forecast errors.
What is the difference between the two recipes?
The Probabilistic Ensemble gives members equal influence. The IQR Ensemble varies their influence according to predicted forecast spread, but a narrower spread does not necessarily indicate greater accuracy.
Did the ensemble outperform the individual models?
The supplied report says IBM evaluated the recipes on 97 GIFT-Eval configurations, but the provided excerpt omits the results. It does not establish whether the ensemble outperformed its members.
What should users check before relying on its forecasts?
Users should compare the recipes with individual models on representative data, checking forecast accuracy and calibration at the horizons relevant to their application.
Source: rss
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
