IBM’s Granite time-series ensemble edges its best member by 0.002 MASE

IBM’s four-model recipe slightly improves aggregate point forecasting on GIFT-Eval, but adds runtime cost and does not lead every metric.

2 min read

Analytics charts displayed on a computer monitor

Analytics charts displayed on a computer monitor

IBM Research has published a Granite Time Series Ensemble R1 recipe that combines four existing Granite forecasting models rather than introducing a single merged checkpoint. The configuration downloads two PatchTST variants, FlowState and Tiny Time Mixers, then aggregates their forecasts to reduce the risk of choosing the wrong architecture for an unfamiliar dataset.

On 97 configurations from GIFT-Eval, IBM reports an aggregate MASE of 0.683 for the ensemble, compared with 0.685 for the strongest individual Granite model. The 0.002 improvement is real within the reported table but small, and the best individual model still led the ensemble on aggregate probabilistic loss. Some excluded or non-permissively licensed systems also score better, so this is not evidence that Granite is the overall forecasting leader.

The practical argument is robustness: different architectures win on different datasets, while a fixed ensemble can offer a reasonable default before task-specific selection or tuning. The tradeoff is equally concrete—running four checkpoints increases memory use and inference latency, and performance can still depend on context length, channel handling and preprocessing. The results are publisher-reported and may change as GIFT-Eval evolves.

The canonical Granite TSFM repository, maintained by IBM, is Apache-2.0 licensed and had 914 total GitHub stars, 283 forks and 2,588 commits when checked October 22, 2026. Those totals indicate an established project, but no historical baseline was available to measure recent star growth. Users should also inspect optional recipe dependencies separately, because expanded configurations can add components with different research-use terms even though the default four-model Granite recipe is permissively licensed.

Sources: IBM Research announcement · Canonical repository · GIFT-Eval

ForecastingTime seriesGIFT-EvalIBMGranite