IBM Research has combined four of its open Granite forecasting models into one recipe that pools their predictions instead of picking a single winner. The Granite time series ensemble was published on Hugging Face on 22 October 2026. It needs no retraining, and on the GIFT-Eval benchmark its point forecasts beat every one of its four members.
Key takeaways
- The release is
granite-timeseries-ensemble-r1, a recipe that runs four pre-trained Granite models and merges their forecasts. - You can merge them two ways: an equal-weight “Probabilistic” mix, which is the default, or an “IQR” mix that gives more weight to the more confident models.
- On GIFT-Eval (leaderboard snapshot of 20 October 2026), both versions scored a better aggregate MASE than any single member.
- The trade-off: running four checkpoints costs more time and memory than running one.
What is the Granite time series ensemble?
A time-series model forecasts the next values in a sequence of numbers, such as hourly power demand, server load or weekly sales. IBM’s Granite family contains several of these “foundation models”. They are pre-trained on huge amounts of data, so they can forecast a series they have never seen. That ability is called zero-shot forecasting.
The ensemble adds no new neural network. According to IBM Research’s announcement, the recipe “stores the ensemble configuration rather than merged model weights” and downloads the four member checkpoints when it runs. Think of it as asking four weather forecasters for their predictions and blending them, instead of trusting whoever was right last week.
Which four models are inside?
The four members are deliberately different. A large transformer, a mid-sized state-space model and a tiny mixer tend to make different mistakes, and an ensemble works best when its members’ errors don’t line up.
| Model | Architecture | Parameters |
|---|---|---|
| PatchTST-FM-r1 | Patched Transformer | 0.3B |
| PatchTST-FM-r2 | Patched Conformer | 0.4B |
| FlowState-r1.1 | State-space model | 9.07M |
| TTM-r3 | Tiny Time Mixer | 1.41M |
Three of the four arrived together. IBM announced FlowState-r1.1, PatchTST-FM-r1 and TTM-r3 in a research blog post on 31 March 2026, and it says the TTM family had passed 37 million Hugging Face downloads by then. PatchTST-FM-r2, the largest member at about 385 million parameters, followed on 9 September 2026, as Unite.ai reported.
How does it combine four forecasts?
Each member returns quantiles, not just a single number. A quantile is a probability cut-off: the 0.9 quantile is the value the model expects the real outcome to fall below 90% of the time. The recipe offers two ways to merge them.
- Probabilistic Ensemble (default): every member counts equally. Their forecast distributions are pooled into one combined distribution, and the requested quantiles are read off it. Nothing has to be fitted.
- IQR Ensemble: members are weighted by how tight their forecast range is, so a member with a narrower interquartile range gets more say. The weights change for each series, target and forecast step. A clean-up step called isotonic regression keeps the quantiles in order, and you can cap the weight any single member gets.
Does the ensemble actually beat its members?
Mostly, yes. GIFT-Eval is a zero-shot forecasting benchmark from Salesforce AI Research, and IBM ran it across 97 dataset configurations. MASE (mean absolute scaled error) measures how far the point forecast misses. CRPS measures how good the whole probability range is. For both, lower is better, and 1.0 means you only matched a simple seasonal baseline.
| Forecast | MASE | CRPS |
|---|---|---|
| Ensemble r1, Probabilistic | 0.683 | 0.471 |
| Ensemble r1, IQR | 0.684 | 0.471 |
| PatchTST-FM-r2 | 0.685 | 0.467 |
| FlowState-r1.1 | 0.701 | 0.490 |
| PatchTST-FM-r1 | 0.717 | 0.488 |
| TTM-r3 | 0.724 | 0.520 |
Look closely and the margin over PatchTST-FM-r2 is tiny, and r2 still wins on raw CRPS. IBM says so itself. Its main claim is consistency: the ensemble ranked 3rd among eligible systems on both MASE rank and CRPS rank, because it does well across many datasets rather than brilliantly on a few.
In my opinion, that is the honest selling point. If you forecast hundreds of different series and can’t hand-pick a model for each one, a steady performer is worth more than a sometimes-champion.
How do you run it?
Install the toolkit with pip install "granite-tsfm>=0.3.11", then load the recipe with QuantileEnsembleForecaster.from_pretrained("ibm-granite/granite-timeseries-ensemble-r1"). IBM’s example uses 1,024 past steps of the classic ETTh1 electricity dataset to forecast the next 24.
You get a point forecast plus quantiles from 0.1 to 0.9 by default. The gap between 0.1 and 0.9 is your nominal 80% range. To switch to the IQR method, you change one setting: the subfolder becomes recipes/granite-iqr.
What are the catches?
- Cost: four checkpoints mean more inference time and memory than one.
- No cross-talk: each column of a multivariate dataset is forecast on its own, so the ensemble can’t learn that temperature drives power demand.
- Mismatched members: context limits, horizons and missing-value handling differ from model to model.
- Overconfidence: the IQR method can reward a member that is confidently wrong. A narrow range is not the same as a correct one.
- Licences: the default recipe and members are permissively licensed, but the expanded recipes that add IBM Research checkpoints carry non-commercial or research-only terms.
IBM’s own advice is to backtest on data that looks like yours and monitor forecasts after you ship. That is the same discipline we recommend in our pre-launch AI evaluation checklist.
What happens next
The leaderboard keeps moving. Rivals such as Google’s TimesFM-3 post better individual scores, and the version number on this recipe suggests IBM plans further releases. Grid operators, IT teams and supply-chain planners are the obvious first users: IBM pitches these models at industrial monitoring and electrical grid disruptions. Grid forecasting matters more as demand shifts, as our look at the US gas-plant rush showed.
FAQ
Do I need to train the Granite time series ensemble?
No. It is zero-shot and has no fitted ensemble weights. You load the recipe and it downloads the four pre-trained members.
Which method should I pick, Probabilistic or IQR?
Start with Probabilistic, IBM’s recommended default. Try IQR if your backtests show that one member is reliably sharper on your data.
Can I use it commercially?
IBM says the default recipe and its Granite members are permissively licensed. Expanded recipes with research checkpoints are not, so check each model card.
Next step: grab a year of your own data, hold back the last month, and see whether the ensemble beats the model you use today.
Leave a comment