Ensemble forecasting with IBM Granite time-series fashions

4 open Granite fashions, one extensible forecasting interface, and fewer
guesswork about which model to utilize.

Pre-trained time-series fashions can forecast with out teaching in your dataset, nonetheless you proceed to need to determine on one. A model that handles clear industrial indicators properly may battle with noisy retail demand or an prolonged seasonal horizon, and the best choice depends on the gathering and the obligation.

Granite Timeseries Ensemble reduces that choice to a smart default. One recipe combines 4 complementary Granite architectures and returns median and quantile forecasts, with no retraining and no fitted ensemble weights. Start with the Probabilistic Ensemble (uniform probability pooling), study the adaptive Inter Quartile Range (IQR) Ensemble, and lengthen the similar interface with appropriate forecasters. The ensemble recipe and its default Granite members are permissively licensed.

The idea will doubtless be acquainted from local weather forecasting. Ensemble forecasts uncover an expansion of attainable outcomes, and strategies such as a result of the North American Multi-Model Ensemble combine fully totally different fashions for seasonal prediction. Time-series forecasting benefits in the identical means: complementary fashions present fully totally different views of what comes subsequent, and an ensemble merges these views into one forecast.

Try Granite Timeseries Ensemble

Arrange Granite TSFM and cargo the actually helpful four-model recipe immediately from
Hugging Face:

pip arrange "granite-tsfm>=0.3.11"

import pandas as pd

from tsfm_public.fashions.ensemble.modeling_ensemble import QuantileEnsembleForecaster

df = pd.read_csv(
    "https://raw.githubusercontent.com/zhouhaoyi/ETDataset/most essential/ETT-small/ETTh1.csv",
    parse_dates=["date"],
)

ensemble = QuantileEnsembleForecaster.from_pretrained(
    "ibm-granite/granite-timeseries-ensemble-r1",
)

consequence = ensemble(
    df.tail(1024),
    timestamp_column="date",
    target_columns=["OT"],
    context_length=1024,
    prediction_length=24,
)

print(consequence.prediction_outputs) 
print(consequence.quantile_outputs)   

The recipe outlets the ensemble configuration fairly than merged model weights. It downloads the 4 member checkpoints at runtime, so allow extra time for the first run. For the whole configuration, member checkpoints, evaluation particulars, and limitations, see the model card.

consequence.prediction_outputs incorporates the aim forecast, whereas consequence.quantile_outputs incorporates the quantile predictions which by default range from 0.1 to 0.9. Functions can use the aim value as a central prediction and chosen quantiles to particular forecast ranges. For example, the 0.1 and 0.9 quantiles define a central forecast interval with nominal 80% safety.

The ensemble members

The default recipe combines 4 pre-trained Granite fashions with fully totally different architectures:

For background, see the PatchTST-FM-r1 paper,
PatchTST-FM-r2 construction overview,
FlowState paper, and
Tiny Time Mixers paper.

These architectures seize developments, seasonality, and native variation in a number of strategies, so their errors needn’t coincide. A 50-year evaluation of forecast
combos describes how combining forecasts can reduce the hazard of selecting the unsuitable single model, whereas stressing that member prime quality and selection nonetheless matter.

Two strategies to combine the forecasts

Each member returns quantile forecasts. The two most essential recipes use fully totally different pointers to indicate these forecasts into one predictive distribution and its median.

Probabilistic Ensemble: the actually helpful default

The Probabilistic Ensemble (uniform probability pooling) presents each member equal have an effect on, swimming swimming pools their quantile forecasts in probability home, and extracts the requested quantiles from the combined values. On account of the weights are mounted, there’s no turning into step and no weighting hyperparameters to tune for a model new assortment.

The forecast combination puzzle
affords useful motivation: in degree forecasting, straightforward equal-weight combos
have often held up properly in opposition to methods with estimated weights. This recipe applies the similar equal-influence idea to probabilistic forecasts. It is the default loaded inside the occasion above.

IQR Ensemble: an adaptive totally different

The IQR Ensemble weights members consistent with their forecast unfold, giving narrower intervals increased have an effect on. It approximates each member’s interquartile range from the closest accessible quantiles throughout the twenty fifth and seventy fifth percentiles. Weights differ by sample, aim, and horizon step, and identical to the Probabilistic Ensemble, nothing is fitted from historic forecast errors. Isotonic regression retains the combined quantiles from crossing.

A narrower interval would not basically suggest a further right or better-calibrated forecast, and IQR weighting can favor an overconfident member. If you happen to want to experiment, the recipe exposes a temperature setting (better values switch the weights in the direction of uniform) and an non-obligatory cap on any single member’s weight.

Load the evaluated IQR recipe by altering solely the subfolder:

ensemble = QuantileEnsembleForecaster.from_pretrained(
    "ibm-granite/granite-timeseries-ensemble-r1",
    subfolder="recipes/granite-iqr",
)

Which methodology should I exploit?

Start with the Probabilistic Ensemble. If forecast unfold appears to be informative to your utility, study the IQR Ensemble in opposition to it on marketing consultant info, since accuracy and calibration depend on the gathering and the horizon.

Does the ensemble improve on its members?

We evaluated every recipes on 97 GIFT-Eval dataset configurations. The fashions
forecast with out fine-tuning, and test targets weren’t used to select or match
the aggregation methodology. Proper right here we study the official ensemble submission outcomes
with the 4 constituent fashions, using printed leaderboard scores and ranks captured on 20 October 2026.

GIFT-Eval experiences MASE for degree forecasts and CRPS/MWQL (the benchmark’s
mean_weighted_sum_quantile_loss metric) for probabilistic forecasts. Scores
are normalized to Seasonal Naive and geometrically averaged all through configurations.
Lower is more healthy.

★ Best value among the many many forecasts confirmed. Lower is more healthy.

Rank columns widespread per-configuration ranks in opposition to the whole GIFT-Eval reference leaderboard, not merely these six forecasts, so rank values can shift as entries are added to the leaderboard.

Every ensembles outperform all 4 specific particular person members on mixture MASE and customary MASE Rank. The Probabilistic Ensemble achieves the underside MASE amongst these six forecasts: 0.683 in distinction with 0.685 for PatchTST-FM-r2. Lower widespread MASE Rank signifies stronger relative point-forecast effectivity all through GIFT-Eval’s quite a few dataset configurations. Collectively, these outcomes assist the ensemble as a wise begin line, with out requiring builders to select the strongest specific particular person construction upfront. The MASE margin over the right single member is small, so the precept revenue is robustness: you get comparable or increased mixture accuracy with out determining upfront which construction will suit your info.

PatchTST-FM-r2 retains the underside mixture CRPS/MWQL score. The IQR Ensemble achieves the underside widespread CRPS Rank amongst these six forecasts, whereas the Probabilistic Ensemble ranks behind PatchTST-FM-r2 on that measure. Mixture error and customary rank because of this truth favor fully totally different selections.

Inside the 20 October 2026 comparability, the ensemble technique ranks third amongst eligible strategies on MASE Rank and CRPS Rank. The decide beneath displays every recipes alongside the ten biggest eligible reference strategies per metric. Some strategies, along with entries excluded as fine-tuned or non-permissively licensed, score increased than each recipe, so the value of the ensemble is not that it tops the leaderboard. It is {{that a}} single default delivers sturdy, fixed outcomes all through many assortment with out requiring you to determine on a member.

Decide: GIFT-Eval average-rank standings on 20 October 2026, exhibiting every Granite ensembles alongside the ten biggest eligible reference strategies per metric. Eligible strategies have replication code and no declared test-data leakage; fine-tuned entries and strategies with recognized non-permissive licenses or member dependencies are excluded. Rank values are taken immediately from the printed complete leaderboard snapshot; filtering changes which strategies are displayed, not their rank values.

Lengthen the recipe to totally different forecasters

The default recipe makes use of 4 Granite fashions, nonetheless the widespread Forecaster interface moreover accepts appropriate exterior fashions. Chances are you’ll change the member mix or the aggregation methodology and keep the similar downstream forecast format, with no merging or retraining of member checkpoints. That makes it less complicated to swimsuit the ensemble into an present forecasting stack, and in order so as to add forecasters suited to your space as your needs evolve.

The repository moreover consists of expanded recipes that add IBM Evaluation checkpoints. These carry separate license phrases, along with non-commercial and research-use restrictions, so evaluation each member’s model card sooner than using them.

For any new combination, study forecast compatibility, accuracy, calibration, latency, and licensing.

Contemplate the recipe to your utility

GIFT-Eval is a broad benchmark, nonetheless mixture optimistic elements don’t suggest every assortment benefits, and customary ranks don’t arrange calibration. The way in which by which to learn the way the recipes behave to your utility is to backtest them on marketing consultant info:

• Match your use case. Use historic forecast house home windows that replicate the knowledge accessible at each forecast origin, the horizon you need, and the gathering you anticipate to serve.
• Consider like with like. Run the Probabilistic Ensemble, the IQR Ensemble, and a strong specific particular person member (similar to PatchTST-FM-r2) on the similar house home windows.
• Look previous degree error. Assess quantile loss alongside MASE, and study the empirical safety of the 0.1 to 0.9 interval in opposition to its nominal 80%.
• Measure the worth. File runtime and memory use to guage whether or not or not the advantage of combining 4 fashions justifies it to your workload.

Keep two limitations in ideas. The current recipe forecasts multivariate channels independently, with out using cross-channel information, and it a whole bunch 4 member checkpoints, which can enhance inference time and memory relative to a single member. Member context limits, horizons, preprocessing, and missing-value coping with moreover differ, and neither aggregation methodology ensures calibration or robustness beneath distribution shift, so keep monitoring forecasts after deployment.

Try it in your info

The actually helpful begin line is the default four-model recipe:

Load the default recipe to get median and quantile forecasts from all 4 Granite architectures in a single identify, and adapt the recipe as your needs evolve.

References