AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Ensemble forecasting with IBM Granite time-series models

Collected Oct 22, 2026

IBM has published Granite Timeseries Ensemble, an open recipe that pools four pre-trained IBM Granite time-series models into a single forecasting call, announced on the Hugging Face Blog. The recipe is designed to answer a recurring problem: customers can use pre-trained forecasters without training on their own data, but still have to choose which one. A model suited to smooth industrial signals may perform poorly on noisy retail demand or long seasonal horizons, so the default recipe replaces that choice with a fixed combination of complementary architectures and returns both a point forecast and quantiles with no additional training.

The recommended four-member mix includes PatchTST-FM-r1, PatchTST-FM-r2, FlowState and a Tiny Time Mixers model. In practice, mixer- and patch-based forecasters process the same series differently, capturing trends, seasonality and local variation through distinct mechanisms, so their errors need not coincide. The default path installs the granite-tsfm package at version 0.3.11 or newer and loads the QuantileEnsembleForecaster with from_pretrained on the ibm-granite/granite-timeseries-ensemble-r1 repository. The snippet in the announcement uses a 1024-step context and a 24-step horizon on the ETTh1 dataset, passing the target column OT and a timestamp column, then reads result.prediction_outputs for the point forecast and result.quantile_outputs for quantile predictions. Quantiles run from 0.1 to 0.9 by default, so 0.1 and 0.9 define a central interval with nominal 80 percent coverage.

The published recipe stores the ensemble configuration rather than merged model weights. At runtime it downloads the four member checkpoints, which means the first run takes longer and inference carries the cost of four models. The default aggregation is the Probabilistic Ensemble, a uniform probability pooling method that gives each member equal influence, pools their quantile forecasts in probability space, and then extracts the requested quantiles from the combined distribution. Because the weights are fixed, there is no fitting step and no weighting hyperparameters to tune for a new series.

The second option, the IQR Ensemble, adapts weights per sample, target and horizon step, assigning more influence to members whose forecast spread is narrower. It approximates each member's interquartile range from the quantiles closest to the 25th and 75th percentiles, applies isotonic regression so the combined quantiles do not cross, and is loaded by setting subfolder to recipes/granite-iqr. The recipe exposes a temperature setting, where higher values push the weights toward uniform, and an optional cap on any single member's weight. Narrower intervals are not proof of accuracy or calibration, and the IQR approach can favor an overconfident member.

IBM evaluated both recipes on 97 GIFT-Eval dataset configurations without fine-tuning, and test targets were not used to select or fit the aggregation method. The comparison cites leaderboard scores and ranks captured on 20 October 2026, with MASE reported for point forecasts and CRPS/MWQL for probabilistic forecasts, normalized to Seasonal Naive and geometrically averaged. On aggregate MASE, both ensembles beat all four individual members. The Probabilistic Ensemble scored 0.683 against 0.685 for PatchTST-FM-r2, the best single member, a small difference. The announcement argues that this makes the benefit one of robustness rather than raw top accuracy: teams get comparable or better aggregate error without knowing in advance which architecture fits their data.

On probabilistic metrics the picture is mixed. PatchTST-FM-r2 retains the lowest aggregate CRPS/MWQL, while the IQR Ensemble holds the best average CRPS Rank among the six forecasts shown, and the Probabilistic Ensemble ranks behind PatchTST-FM-r2 on that measure. The ensemble approach ranks third among eligible systems on both MASE Rank and CRPS Rank, and some excluded entries, including fine-tuned or non-permissively licensed systems, score better. That framing matters: the ensemble's value is consistency from one default, not leaderboard leadership.

The interface is extensible. The Forecaster interface accepts compatible external models, so teams can swap the member mix or the aggregation method and keep the same downstream forecast format without merging or retraining checkpoints. The repository also includes expanded recipes that add IBM Research checkpoints, which carry separate license terms including non-commercial and research-use restrictions. The published ensemble recipe and its four default members are permissively licensed, and the announcement advises checking each member's model card before use.

The announcement is candid about limits. The current recipe forecasts multivariate channels independently and does not use cross-channel information. Loading four checkpoints raises inference time and memory relative to a single member. Member context limits, horizons, preprocessing and missing-value handling differ, and neither aggregation method guarantees calibration or robustness under distribution shift. To evaluate on their own data, teams are advised to backtest on windows that mirror their forecast origins and horizons, compare the Probabilistic Ensemble, the IQR Ensemble and a strong single member such as PatchTST-FM-r2 on the same windows, examine quantile loss and interval coverage alongside MASE, and record runtime and memory before committing to a four-model stack.

Why it matters: developers building forecasting features now have a maintained open default that removes architecture selection, and they can adopt it with a single from_pretrained call. The cost is four times the checkpoint load and inference work, plus independent channel treatment for multivariate data. A likely trade-off, in my reading, is that teams gain a safe baseline quickly, but those with tight latency budgets or strongly correlated channels may still prefer one member or a custom combination. The permission split between the default members and the optional IBM Research checkpoints means licensing review is part of the integration, and anyone relying on quantiles should validate empirical coverage against the nominal 80 percent interval on their own data.

Read at Hugging Face Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt