IBM's Granite Timeseries Ensemble Combines Four Forecasting Models So Teams Need Not Pick One

What happened

IBM Research published an article on Hugging Face on October 22, 2026, introducing Granite Timeseries Ensemble. A time series is data recorded in time order, such as hourly electricity use or daily sales; an "ensemble" merges the forecasts of several models into one result. IBM notes that pretrained time-series models (models that arrive already trained) can forecast without training on the target dataset, but users still have to choose a model; one suited to smooth industrial signals may not suit noisy retail demand or long seasonal forecasts.

According to IBM Research, the recipe combines four complementary Granite architectures and returns median and quantile forecasts, with no retraining and no ensemble weights to fit. Quantile forecasts express a likely range: for example, the span between the 0.1 and 0.9 quantiles is a prediction interval nominally covering 80% of cases. The results and explanations below are IBM's claims.

Two ways to combine, and the four members

IBM lists the default members as: PatchTST-FM-r1 (Patched Transformer), PatchTST-FM-r2 (Patched Conformer), FlowState-r1.1 (state-space model) and TTM-r3 (Tiny Time Mixers, an MLP mixer). IBM says these architectures capture trends, seasonality and local variation in different ways, so their errors need not coincide.

IBM's recommended default is the Probabilistic Ensemble: each member has equal influence, and quantile forecasts are pooled in probability space. The weights are fixed, so there is no fitting step and no weighting settings (hyperparameters) to tune for new series. The other option, the IQR Ensemble, weights members by forecast spread, so members with narrower intervals have more influence. IBM cautions that a narrower interval is not necessarily more accurate or better calibrated (calibration means whether an interval's actual coverage matches its stated proportion), and it may favor overconfident members.

How to read the results

GIFT-Eval is a public benchmark and leaderboard for forecasting models. It uses MASE to measure error in single-value forecasts and CRPS/MWQL to measure error in range (probabilistic) forecasts; scores are relative to a simple seasonal baseline, and lower is better. IBM says the evaluation covered 97 GIFT-Eval dataset configurations, the models were not fine-tuned (given extra training for these datasets), and test targets were not used to select or fit the aggregation method. IBM says both ensembles beat the four single members on overall MASE and average MASE rank, but the MASE gap to the best single member is small, and the main benefit is robustness.

On the probabilistic metric CRPS/MWQL, PatchTST-FM-r2's 0.467 is still the lowest of the six. IBM notes the IQR Ensemble has the best average CRPS rank of the six, so overall error and average rank point to different choices. IBM also says the ensemble ranks 3rd among eligible systems on both MASE rank and CRPS rank, and some systems (including those excluded for fine-tuning or non-permissive licenses) score better, so the value is not in topping the leaderboard.

What it means for teams that forecast

For teams forecasting demand, usage or sensor data, this offers a "use first, compare later" default starting point, so they need not guess which architecture fits best. But IBM stresses that aggregate scores do not mean every series benefits. It suggests backtesting (checking forecasts against past data whose outcomes are known) on representative data. It also suggests comparing both ensembles and a strong single member over the same windows, and checking quantile loss and the actual coverage of the 0.1 to 0.9 interval against the nominal 80%. Teams should also record runtime and memory use.

On licensing, IBM says the ensemble recipe and its default Granite members are permissively licensed. The repository also has expanded recipes that add IBM Research checkpoints, which carry non-commercial and research-use restrictions. IBM also says the shared Forecaster interface can accept compatible external models while keeping the same downstream forecast format.

Browse the latest news in this topic

• AI news