Ensemble forecasting with IBM Granite time-series models

IBM is bundling four pre-trained Granite time-series models into a single forecasting recipe, so users don’t have to guess which architecture fits a given series. The pitch is simple: one interface, no retraining, and a sensible default that can return both point forecasts and quantiles straight away.

That matters because time-series models are picky. A setup that handles smooth industrial data well can stumble on noisy retail demand or longer seasonal patterns. The new Granite Timeseries Ensemble tries to smooth over that problem by combining four complementary architectures instead of betting on one. The default recipe uses uniform probability pooling, while a second option, the IQR Ensemble, gives more influence to members with narrower forecast spreads.

The ensemble doesn’t merge model weights. It stores the configuration and loads the four checkpoints at runtime, which means the first run takes longer. The public example uses a 1,024-step context and forecasts 24 steps ahead, and the quantiles default from 0.1 to 0.9, so users can turn the output into a central forecast and an interval. The package is installed through granite-tsfm, and the recommended recipe is available on Hugging Face under ibm-granite/granite-timeseries-ensemble-r1.

IBM says the four members bring different strengths: PatchTST-FM-r1, PatchTST-FM-r2, FlowState, and Tiny Time Mixers. The point is diversity. Different models catch trends, seasonality, and local variation in different ways, so their errors don’t line up neatly. That’s the old ensemble idea from weather forecasting, just applied to time series.

On 97 GIFT-Eval configurations, both ensemble recipes beat all four individual members on aggregate MASE and average MASE rank. The Probabilistic Ensemble posted the lowest MASE among the six forecasts shown at 0.683, just ahead of PatchTST-FM-r2 at 0.685. PatchTST-FM-r2 still led on aggregate CRPS/MWQL, while the IQR Ensemble took the best average CRPS rank. In other words, the ensemble is not a trophy hunter. It’s a practical default for people who want solid results without first becoming an amateur model selector.

IBM also lets the same interface accept compatible external forecasters, which makes the setup easier to extend without retraining or checkpoint merging. But there are limits: the current recipe forecasts multivariate channels independently, uses four checkpoints, and brings extra runtime and memory cost. The sensible move is still the boring one — backtest it on representative data and see whether the extra machinery earns its keep.