IBM Granite Timeseries Ensemble Beats Single Models on GIFT Eval
IBM Granite Timeseries Ensemble combines four open models, beating each on GIFT-Eval point forecasts while offering probabilistic and adaptive recipes.
Summary
IBM’s October 22, 2026 Granite Timeseries Ensemble R1 combines four permissively licensed, pretrained forecasters without retraining or fitted weights: Granite PatchTST-FM-r1, a Patched Transformer; Granite PatchTST-FM-r2, a Patched Conformer; Granite FlowState-r1.1, a state-space model; and Granite TTM-r3, a Tiny Time Mixers MLP mixer. Granite TSFM 0.3.11 or later loads the Hugging Face recipe through QuantileEnsembleForecaster, then downloads four checkpoints. One call returns a median forecast and default 0.1 to 0.9 quantiles, whose endpoints form a nominal 80% interval.
The recommended Probabilistic Ensemble pools members equally in probability space, requiring no fitting or weighting hyperparameters. The alternative IQR Ensemble changes weights by sample, target and horizon step, favoring narrower estimated interquartile ranges; isotonic regression prevents quantile crossing, while temperature pushes weights toward uniform and an optional cap limits one member. Because narrow intervals can be overconfident, IBM advises testing both recipes on representative data.
On 97 GIFT-Eval configurations, without fine-tuning or target-based selection, the October 20, 2026 leaderboard snapshot gave Probabilistic and IQR MASE of 0.683 and 0.684, beating PatchTST-FM-r2 at 0.685, FlowState-r1.1 at 0.701, PatchTST-FM-r1 at 0.717 and TTM-r3 at 0.724. Scores are Seasonal Naive normalized, geometrically averaged, and lower is better. PatchTST-FM-r2 kept the best CRPS/MWQL, 0.467 versus 0.471 for both ensembles, but IQR led the six forecasts’ average CRPS rank at 47.619; Probabilistic led MASE rank at 47.175. The ensembles ranked third among eligible systems on both rank measures, not overall, as stronger fine-tuned or nonpermissive entries were excluded. The interface accepts compatible external models, but expanded IBM Research recipes may restrict commercial use. Users should backtest accuracy, calibration, interval coverage, latency and memory: four checkpoints cost more than one, channels are forecast independently, member limits vary, and neither method guarantees resilience to distribution shift.
Positives
- 0.683 MASE made the Probabilistic Ensemble the strongest point forecaster among the six Granite entries evaluated.
- Both ensemble recipes beat all four individual members on aggregate MASE and average MASE rank across 97 GIFT-Eval configurations.
- Granite TSFM 0.3.11 or later delivers median and quantile forecasts from four architectures through one interface without retraining.
- 47.619 average CRPS rank gave the IQR Ensemble the best probabilistic ranking among the six compared forecasts.
- Compatible external forecasters can replace or supplement Granite members without merging or retraining checkpoints.
Risks & concerns
- 0.467 CRPS/MWQL allowed PatchTST-FM-r2 to outperform both ensembles, which each scored 0.471.
- Four downloaded checkpoints increase first-run time, inference latency and memory use compared with one model.
- IQR weighting can favor an overconfident model because narrower forecast intervals do not guarantee accuracy or calibration.
- Multivariate channels are forecast independently, so the current recipe cannot exploit relationships between channels.
- Neither aggregation method guarantees calibration or robustness under distribution shift, requiring representative backtesting and continued monitoring.
- Expanded IBM Research recipes may impose noncommercial or research-only licensing restrictions.