Granite Timeseries Ensemble combines four models for time series forecasting
The Granite Timeseries Ensemble tool combines four pretrained models without fine-tuning or learning ensemble weights. Through the Granite TSFM library, it returns both point and quantile forecasts; the default method gives all models equal influence.
The Granite Timeseries Ensemble tool provides a common interface for time series forecasting using four pretrained models: Granite PatchTST-FM-r1, Granite PatchTST-FM-r2, Granite FlowState-r1.1 and Granite TTM-r3. Combining them requires neither fine-tuning on user data nor learning weights from historical forecast errors.
The example uses version 0.3.11 or later of the Granite TSFM library and the ibm-granite/granite-timeseries-ensemble-r1 configuration. It generates a 24-step forecast from the last 1024 observations. The interface returns a point forecast and quantiles, ranging from 0.1 to 0.9 by default. The outermost quantiles define an interval with nominal coverage of 80 %; actual coverage needs to be verified on the specific data.
The default Probabilistic Ensemble method gives all models equal influence. The alternative IQR Ensemble method assigns greater weight to models with narrower prediction intervals. The authors caution that a narrower interval does not necessarily mean a more accurate forecast, and that this method may favor an overconfident model. They therefore recommend comparing both variants on representative data.
The configuration contains ensemble settings, not merged model weights. The individual models are downloaded at runtime, so the first run takes more time. According to the article, both the configuration and the underlying models are available under permissive licenses.
Why it matters
The approach offers a concrete starting point for tasks where it is not clear in advance which time series model to choose. However, both accuracy and the reliability of prediction intervals depend on the data and the forecast horizon; combining models alone does not guarantee a better result.
Two audiences, two different impacts
What this means
For individuals
An analyst or developer can try combining four models without setting up fine-tuning or estimating weights for their own time series.
For a business
In demand planning, quantile forecasts can supplement a single expected value with a range of outcomes. For decision-making, it is essential to verify whether the intervals reflect the actual uncertainty in company data.
ProcessesCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.