2026-10-22 · ← News
IBM folds four Granite models into one forecast. It beats the best of them by two thousandths, and the real win is not having to choose
IBM Research has released Granite Timeseries Ensemble, a recipe that combines four pre-trained time-series forecasting models without any extra training. On the GIFT-Eval benchmark it beat each of its members on point accuracy, but the best one by only two thousandths. The real value is that a data team no longer has to guess in advance which model suits its data.
The image could not be loaded.
On 22 October IBM Research published Granite Timeseries Ensemble on Hugging Face. The recipe combines four of its own time-series forecasting models and returns a single point forecast plus quantiles, with no retraining and no fitted weights.
IBM stacks four different architectures into one forecast
The members are Granite PatchTST-FM-r1 (patched transformer), PatchTST-FM-r2 (patched conformer), FlowState-r1.1 (state-space model) and TTM-r3 (Tiny Time Mixers). Each captures trend, seasonality and short-term swings differently, so their errors need not coincide. Their sizes differ by orders of magnitude: Hugging Face lists 1.41 million parameters for TTM-r3, 9.07 million for FlowState, and 0.3 and 0.4 billion for the two PatchTST models.
The default Probabilistic Ensemble gives every member equal weight and pools their quantile forecasts into one distribution. The alternative IQR Ensemble favours members with narrower intervals. IBM itself warns that a narrower interval does not necessarily mean a more accurate forecast and that this method can favour an overconfident member. The recipe and its default members are permissively licensed and install through the granite-tsfm package from version 0.3.11.
On GIFT-Eval the ensemble wins by two thousandths and loses on quantiles
IBM tested both recipes on 97 GIFT-Eval dataset configurations, without fine-tuning. The Probabilistic Ensemble has the lowest MASE of the six forecasts compared, 0.683. The best single member, PatchTST-FM-r2, scores 0.685. On CRPS, which grades the whole probabilistic distribution, PatchTST-FM-r2 keeps the lead at 0.467 against 0.471 for both ensembles. On the 20 October 2026 leaderboard, the ensemble ranked third among the systems IBM counts as eligible, on both average MASE rank and average CRPS rank.
For a data team in energy, logistics or e-commerce, the average rank matters more than the absolute score. The ensemble lands reliably near the top across varied data, even if another model wins on a particular series. IBM says so openly: the main benefit is robustness. In a company where forecasting serves hundreds of series, from demand to electricity load, the argument over which model to pick goes away. What remains is one interface that also accepts compatible forecasters from other authors.
Radar has previously covered PatchTST-FM-r2 as the strongest permissively licensed model on GIFT-Eval. The ensemble narrowly overtakes it on point accuracy and stays behind on probabilistic accuracy. For a team that is happy with one good model, little changes.
Four checkpoints cost more memory, and the expanded recipes carry other licences
The recipe downloads all four checkpoints at runtime, so inference takes longer and uses more memory than a single model. It forecasts multivariate series channel by channel and ignores the relationships between them, which can hurt with data such as temperature and load. According to IBM, neither method guarantees calibration or robustness under distribution shift. And the expanded recipes with additional IBM Research checkpoints carry their own licence terms, including non-commercial and research-use restrictions. Anyone who grabs an expanded variant from the repository without reading the model card may pull a legal problem into their product.
A backtest on your own data will decide it, not the leaderboard
IBM itself recommends comparing both ensembles with PatchTST-FM-r2 on the same historical windows, checking whether the interval between the 0.1 and 0.9 quantiles really covers 80% of cases, and measuring runtime and memory. If teams find that a gain measured in thousandths costs noticeably more latency, they will stick with one model. The second signal is whether outside forecasters start appearing in the interface. Then IBM's recipe could become a neutral way to combine forecasts from models by different authors.
Lilith's verdict
Meteorologists stopped arguing long ago about which model is right and read the weather from the whole set of runs. IBM now offers the same to warehouse planners and grid operators: a forecast two thousandths more accurate and, above all, the end of half-hour meetings spent fighting over which model to trust.