The New Development: A Ready-to-Use Recipe for Time Series
The Hugging Face blog presents Granite Timeseries Ensemble, a recipe that brings together four pretrained Granite models with different architectures to produce time-series forecasts without retraining on the user's dataset. According to the source, the starting problem is choosing a single model: a network that performs well on regular industrial signals can disappoint on noisy retail demand or on long seasonal horizons, and the best solution depends on the series and the task.
The ensemble preserves the configuration, not the models' fused weights: the four checkpoints are downloaded at runtime, so the first run takes longer. The Forecaster interface is extensible and also accepts compatible external models, while keeping the downstream output format unchanged. The recipe and the default Granite members have permissive licenses.
How the Two Variants Work
The default recipe, Probabilistic Ensemble, assigns equal weight to each member and combines quantile forecasts in probability space, with no fitting steps or hyperparameters to tune. It is the option suggested as a starting point. The alternative, IQR Ensemble, weights members according to the width of their forecast interval, giving more influence to those that produce narrower intervals; the weights vary by sample, target, and horizon step, and isotonic regression prevents quantile crossing.
The source warns that a narrower interval does not necessarily imply a more accurate or well-calibrated forecast and that IQR weighting can favor an overconfident member. For this reason, the recipe exposes a temperature parameter, which brings the weights closer to a uniform distribution, and an optional cap on the weight of a single member. The output includes a point forecast and quantiles, by default from 0.1 to 0.9.
The Stated Results and Limitations
In the evaluation on 97 configurations of the GIFT-Eval benchmark, with leaderboard scores retrieved on October 20, 2026, both ensembles outperform the four individual members on aggregate MASE and mean MASE rank. The Probabilistic Ensemble records the lowest MASE among the six models compared, 0.683 versus 0.685 for PatchTST-FM-r2, but the margin is described as small by the source itself: the main stated benefit is robustness, not first place. On aggregate CRPS/MWQL, PatchTST-FM-r2 remains ahead, while the IQR Ensemble has the best mean CRPS rank.
The ensembles rank third among eligible systems by MASE rank and by CRPS rank, and some systems excluded because they were fine-tuned or had non-permissive licenses achieve better scores. Accuracy, calibration, latency, and licenses remain to be verified in one's own domain through backtests on representative data. The recipe also provides independent forecasts for multivariate channels, without exploiting cross-channel information, and loading four checkpoints increases inference time and memory compared with a single model.
