Contents
IBM Granite Timeseries Ensemble brings four pretrained models into one forecasting interface, reducing the need to choose an architecture before trying your data. Start with the equally weighted Probabilistic Ensemble and check whether its modest benchmark gains justify the cost of running four models.
Key Points#
- The default recipe returns median and quantile forecasts without retraining component models or learning ensemble weights.
- Across 97 GIFT-Eval configurations, Probabilistic Ensemble achieved a MASE of 0.683, beating the best individual model’s 0.685.
- IQR Ensemble adjusts weights based on forecast spread, but narrow intervals can also indicate overconfidence.
- Using four model checkpoints increases inference time and memory usage, so measure accuracy, interval coverage, and operating costs together.
What the Recipe Includes#
The default recipe includes four Granite models with different architectures.
| Model | Architecture | |---|---| | Granite PatchTST-FM-r1 | Patch-based Transformer | | Granite PatchTST-FM-r2 | Patch-based Conformer | | Granite FlowState-r1.1 | State-space model | | Granite TTM-r3 | MLP mixer |
These models can make different errors on the same time series. Combining their forecasts reduces reliance on choosing the best individual model in advance, but does not guarantee improvements on every dataset.
The recipe stores configuration, not merged weights. Component model checkpoints are downloaded at runtime, so the first run requires additional download time. The default recipe and its Granite component models use permissive licenses, but extended recipes that include other IBM Research checkpoints may carry noncommercial or research-only restrictions.
Choose an Aggregation Method#
Probabilistic Ensemble is the recommended default. It gives each component model equal influence, combines forecasts in probability space, and extracts the requested quantiles. There are no ensemble weights learned from historical data.
IQR Ensemble favors component models with smaller forecast spreads. Weights vary by sample, forecast target, and forecast step. It estimates the interquartile range from the available quantiles and uses isotonic regression to prevent quantile crossing after aggregation.
Compare IQR weighting on representative data before using it. Confident models can still be wrong. The recipe offers a temperature setting that makes weights more uniform as its value increases, along with an option to cap each component model’s weight. This alternative recipe, used in the evaluation, can be loaded with subfolder="recipes/granite-iqr".
Run the Default Recipe#
Install granite-tsfm>=0.3.11 and load the recipe from Hugging Face. This example uses 1,024 hourly observations to forecast the next 24 steps.
import pandas as pd
from tsfm_public.models.ensemble.modeling_ensemble import (
QuantileEnsembleForecaster,
)
df = pd.read_csv(
"https://raw.githubusercontent.com/zhouhaoyi/"
"ETDataset/main/ETT-small/ETTh1.csv",
parse_dates=["date"],
)
ensemble = QuantileEnsembleForecaster.from_pretrained(
"ibm-granite/granite-timeseries-ensemble-r1",
)
result = ensemble(
df.tail(1024),
timestamp_column="date",
target_columns=["OT"],
context_length=1024,
prediction_length=24,
)
print(result.prediction_outputs)
print(result.quantile_outputs)prediction_outputs contains median point forecasts. quantile_outputs contains quantile forecasts, ranging from 0.1 to 0.9 by default. The 0.1 and 0.9 quantiles define an interval with nominal coverage of 80%. Actual coverage must be checked against observed outcomes.
Interpret the Benchmarks Carefully#
The provided IBM article reports results across 97 GIFT-Eval configurations, based on leaderboard scores retrieved on October 20, 2026. Models ran without fine-tuning, and test ground-truth values were not used to select or train the aggregation method. Lower scores are better.
| Forecasting Method | MASE | CRPS/MWQL | |---|---:|---:| | Probabilistic Ensemble | 0.683 | 0.471 | | IQR Ensemble | 0.684 | 0.471 | | PatchTST-FM-r2 | 0.685 | 0.467 |
Both ensembles outperformed all four component models on aggregate MASE, but the gap over PatchTST-FM-r2 is small. PatchTST-FM-r2 performs better on the aggregate probabilistic forecast score. IQR has the best average probabilistic forecast rank among these six forecasting methods, showing that aggregate error and average rank can favor different options.
Treat these results as a reason to try the ensemble. They do not establish that it will improve performance on your time series or that its forecast intervals are well calibrated.
Check Accuracy and Operating Costs#
Backtest Probabilistic Ensemble, IQR Ensemble, and PatchTST-FM-r2 over the same historical forecast windows. Use only observations available at each forecast origin, and evaluate the forecast horizon your application requires.
- Measure point forecast accuracy with MASE and probabilistic forecast accuracy with quantile loss.
- Check whether roughly 80% of observations fall within the interval between the 0.1 and 0.9 quantiles.
- Record runtime and peak memory usage on representative workloads.
- Review results by time series and forecast horizon, since aggregate scores can hide weak cases.
The current recipe forecasts each multivariate channel independently, without using information across channels. Supported context lengths, forecast horizons, preprocessing, and missing-value handling also vary by component model. Neither aggregation method guarantees calibration or reliability when data patterns change.
Next Steps#
Run a backtest comparing the default recipe with PatchTST-FM-r2 over your production forecast horizon, and record MASE, quantile loss, 80% interval coverage, runtime, and peak memory usage.
Source: https://huggingface.co/blog/ibm-research/granite-timeseries-ensemble-r1
Contents
