Glonce

IBM releases Granite Timeseries Ensemble of four forecasters

modelsThe Glonce editorsHugging Face🔥56Glonce rating56Impact45Novelty50Relevance90Surprise20Players55Credibility80
IBM releases Granite Timeseries Ensemble of four forecasters

IBM Research has published Granite Timeseries Ensemble on Hugging Face. It is a recipe that combines four pre-trained Granite time-series models, each with a different architecture, into one forecaster. It returns median and quantile forecasts, with no retraining and no fitted ensemble weights. The pitch is that pre-trained time-series models can forecast without training on your data, but you still have to pick one, and a model that suits smooth industrial signals may struggle with noisy retail demand or a long seasonal horizon. The ensemble turns that choice into a default. IBM compares the idea to weather forecasting, where systems such as the North American Multi-Model Ensemble combine different models. The recipe and its default Granite members are permissively licensed.

Usage is short. After pip install "granite-tsfm>=0.3.11", you load QuantileEnsembleForecaster from tsfm_public.models.ensemble.modeling_ensemble with from_pretrained("ibm-granite/granite-timeseries-ensemble-r1") and call it on a pandas DataFrame, giving the timestamp column, target columns, context length and prediction length. The IBM example uses the ETTh1 dataset with a context of 1024 and a horizon of 24. The result holds prediction_outputs (the point forecast) and quantile_outputs, which by default span 0.1 to 0.9; the 0.1 and 0.9 quantiles give a central interval with nominal 80% coverage. The recipe stores the ensemble configuration rather than merged weights, and it downloads the four member checkpoints at runtime, so the first run takes longer.

There are two ways to combine the members. The Probabilistic Ensemble (uniform probability pooling) is the recommended default. It gives each member equal influence, pools their quantile forecasts in probability space and extracts the requested quantiles from the combined values. With fixed weights there is nothing to fit and no weighting hyperparameters to tune for a new series. IBM links this to the forecast combination puzzle: in point forecasting, simple equal-weight combinations have often held up well against methods with estimated weights. The alternative is the IQR Ensemble, which weights members by forecast spread so that narrower intervals get more influence. It approximates each member's interquartile range from the quantiles nearest the 25th and 75th percentiles, lets weights vary by sample, target and horizon step, fits nothing from past errors, and uses isotonic regression to stop the combined quantiles from crossing. It has a temperature setting (higher values move weights toward uniform) and an optional cap on any one member's weight. IBM warns that a narrower interval is not necessarily a more accurate or better-calibrated one, and that IQR weighting can favor an overconfident member. It is loaded by adding subfolder="recipes/granite-iqr".

For evidence, IBM evaluated both recipes on 97 GIFT-Eval dataset configurations, without fine-tuning and without using test targets to select or fit the aggregation method. The comparison uses published leaderboard scores and ranks captured on 20 October 2026. MASE covers point forecasts and CRPS/MWQL covers probabilistic ones; scores are normalized to Seasonal Naive and geometrically averaged across configurations, and lower is better. Both ensembles beat all four individual members on aggregate MASE and average MASE Rank. The Probabilistic Ensemble has the lowest MASE of the six forecasts shown: 0.683 against 0.685 for PatchTST-FM-r2, the best single member. IBM itself calls that margin small and says the main benefit is robustness, meaning comparable or better aggregate accuracy without knowing in advance which architecture suits your data.

The probabilistic picture is mixed. PatchTST-FM-r2 keeps the lowest aggregate CRPS/MWQL score. The IQR Ensemble has the lowest average CRPS Rank among the six, while the Probabilistic Ensemble ranks behind PatchTST-FM-r2 on that measure. On the full GIFT-Eval leaderboard on 20 October 2026, the ensemble approach ranks 3rd among eligible systems on MASE Rank and CRPS Rank. Eligible systems have replication code and no declared test-data leakage. Some systems, including entries excluded as fine-tuned or non-permissively licensed, score better than either recipe, so IBM says the ensemble's value is a single default that gives strong, consistent results across many series.

The interface is extensible. The common Forecaster interface also accepts compatible external models, so the member mix or aggregation method can change without merging or retraining checkpoints. The repository also has expanded recipes that add IBM Research checkpoints under separate license terms, including non-commercial and research-use restrictions. IBM recommends backtesting on representative data: compare the two ensembles with a strong single member such as PatchTST-FM-r2 on the same windows, look at quantile loss as well as MASE, check empirical coverage of the 0.1 to 0.9 interval against its nominal 80%, and record runtime and memory. Stated limitations: multivariate channels are forecast independently, without cross-channel information; four checkpoints raise inference time and memory over a single member; member context limits, horizons, preprocessing and missing-value handling differ; and neither method guarantees calibration or robustness under distribution shift.

Key facts

  • Granite Timeseries Ensemble combines four pre-trained Granite time-series models into one forecaster that returns median and quantile forecasts, with no retraining and no fitted ensemble weights.
  • The Probabilistic Ensemble (uniform probability pooling) is the recommended default; the IQR Ensemble, which weights members by forecast spread, is the adaptive alternative.
  • On 97 GIFT-Eval configurations, the Probabilistic Ensemble reaches an aggregate MASE of 0.683 against 0.685 for the best single member, PatchTST-FM-r2; IBM calls the margin small and the benefit mainly robustness.
  • PatchTST-FM-r2 keeps the lowest aggregate CRPS/MWQL score; as of 20 October 2026 the ensemble approach ranks 3rd among eligible systems on MASE Rank and CRPS Rank, behind some other systems.
  • Install with pip install "granite-tsfm>=0.3.11"; the recipe and default members are permissively licensed, but expanded recipes add IBM Research checkpoints with non-commercial and research-use restrictions.

Why it matters

Choosing among pre-trained time-series models is a real friction point, because no single model suits every series or horizon. IBM's answer is to ship a default that pools four Granite architectures, so practitioners do not have to pick the strongest one in advance. The idea borrows from weather forecasting and from the forecast combination literature, where equal-weight combinations have often held up against estimated weights. The result here is incremental: the MASE gain over the best member is 0.683 versus 0.685.

Who it affects

Data scientists and engineers who forecast demand, sensor or other time series and want a zero-shot starting point without per-dataset training. It also suits teams with an existing forecasting stack, since the common Forecaster interface accepts compatible external models and keeps the same downstream forecast format.

How to use it

Run pip install "granite-tsfm>=0.3.11", then load QuantileEnsembleForecaster.from_pretrained("ibm-granite/granite-timeseries-ensemble-r1") and call it with a DataFrame, timestamp column, target columns, context length and prediction length. The output has a point forecast and quantiles from 0.1 to 0.9. To try the IQR variant, add subfolder="recipes/granite-iqr". The first run downloads the four member checkpoints. The recipe and its default Granite members are permissively licensed. A model card, a getting-started notebook and the Granite TSFM repository are linked from the post. Backtest on your own windows against a strong single member before relying on it.

How solid is it

The numbers come from IBM Research's own blog post, using published GIFT-Eval leaderboard scores and ranks captured on 20 October 2026, across 97 dataset configurations, without fine-tuning and without using test targets to choose the aggregation. Both ensembles beat all four members on aggregate MASE and average MASE Rank, but not on every measure: PatchTST-FM-r2 keeps the best aggregate CRPS/MWQL. IBM notes that rank values can shift as leaderboard entries are added, and that aggregate gains do not mean every series benefits.

Risks and caveats

The current recipe forecasts multivariate channels independently, ignoring cross-channel information. Loading four checkpoints raises inference time and memory over a single member, and IBM gives no figures for either. Member context limits, horizons, preprocessing and missing-value handling differ. Neither aggregation method guarantees calibration or robustness under distribution shift, and IQR weighting can favor an overconfident member. Some systems outscore both recipes on the leaderboard. Expanded recipes that add IBM Research checkpoints carry separate licence terms, including non-commercial and research-use restrictions, so check each member's model card.

“the value of the ensemble is not that it tops the leaderboard”

— IBM Research blog post on Hugging Face