# Ensemble forecasting with IBM Granite time-series models

DevFeed: [Ensemble forecasting with IBM Granite time-series models](<https://devfeed.tech/articles/ensemble-forecasting-with-ibm-granite-time-series-models-95137.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/ibm-research/granite-timeseries-ensemble-r1>)

Author: Fearghal O'Donncha; Chandramouli Kamanchi; Guruprasad; Wesley M Gifford; Natalia Martinez Gil; Bradley Eck; Pankaj Dayama; Roman Vaculin; Ayhan Sebin

Published: 2026-10-22T07:51:31Z

Content type: article

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Time Series](<https://devfeed.tech/topics/time-series.md>), [Predictive AI](<https://devfeed.tech/topics/predictive-ai.md>), [AI Foundation Models](<https://devfeed.tech/topics/ai-foundation-models.md>)

Tags: [about](<https://devfeed.tech/tags/about.md>), [best](<https://devfeed.tech/tags/best.md>), [but](<https://devfeed.tech/tags/but.md>), [can](<https://devfeed.tech/tags/can.md>), [choice](<https://devfeed.tech/tags/choice.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [demand](<https://devfeed.tech/tags/demand.md>), [forecast](<https://devfeed.tech/tags/forecast.md>), [forecasting](<https://devfeed.tech/tags/forecasting.md>), [granite](<https://devfeed.tech/tags/granite.md>), [horizon](<https://devfeed.tech/tags/horizon.md>), [ibm](<https://devfeed.tech/tags/ibm.md>), [industrial](<https://devfeed.tech/tags/industrial.md>), [interface](<https://devfeed.tech/tags/interface.md>), [less](<https://devfeed.tech/tags/less.md>), [models](<https://devfeed.tech/tags/models.md>), [series](<https://devfeed.tech/tags/series.md>), [time](<https://devfeed.tech/tags/time.md>), [time-series](<https://devfeed.tech/tags/time-series.md>), [with](<https://devfeed.tech/tags/with.md>)

## AI overview

Granite Timeseries Ensemble combines four pretrained Granite forecasters without retraining or fitted ensemble weights, returning point and quantile forecasts through one interface. Its default Probabilistic Ensemble gives members equal influence; the optional IQR Ensemble weights narrower forecast spreads more heavily, which can favor overconfident members. On 97 GIFT-Eval configurations, both recipes beat the four members on aggregate MASE and average MASE rank; the default had the lowest MASE among those six forecasts, while PatchTST-FM-r2 retained the best aggregate CRPS/MWQL. Results vary by metric and do not establish calibration or gains on every series. The excerpt recommends backtesting both recipes against a strong member on representative forecast windows, checking point and quantile error, interval coverage, runtime, and memory. The recipes load four checkpoints, increasing inference cost; channels are forecast independently, and member preprocessing and limits differ. The default Granite members and recipe are permissively licensed, while expanded recipes with IBM Research checkpoints have separate restrictions.

## Source excerpt

*Four open Granite models, one extensible forecasting interface, and less guesswork about which model to use.* Pre-trained time-series models can forecast without training on your dataset, but you still have to choose one. A model that handles smooth industrial signals well may struggle with noisy retail demand or a long seasonal horizon, and the best choice depends on the series and the task. Granite Timeseries Ensemble reduces that choice to a sensible default.