IBM Granite Time-Series Ensemble R1 Release Notes

Overview

The IBM Granite Timeseries Ensemble provides a unified interface for time-series forecasting that aggregates predictions from four complementary model architectures. By combining these models, the system offers a reliable default for forecasting tasks, reducing the need to manually select a single architecture for specific datasets or horizons.

Ensemble Architecture

The default recipe utilizes four pre-trained Granite models, each employing a different architecture to capture varying trends, seasonality, and local variations:

  • Granite PatchTST-FM-r1: A Patched Transformer model.
  • Granite PatchTST-FM-r2: A Patched Conformer model.
  • Granite FlowState-r1.1: A state-space model.
  • Granite TTM-r3: A Tiny Time Mixer (MLP mixer) model.

Because these models process data differently, their error profiles are non-coincident, allowing the ensemble to produce more stable and accurate results than any single member on average.

Aggregation Methods

The ensemble supports two primary methods for combining forecasts, neither of which requires fitting or training:

  • Probabilistic Ensemble (Default): This method uses uniform probability pooling, giving each member equal influence. It is recommended as the starting point for most applications due to its simplicity and lack of tunable hyperparameters.
  • IQR Ensemble: This adaptive method weights members based on their forecast spread, favoring models with narrower intervals. It is useful for applications where the predicted uncertainty is considered informative.

Performance and Evaluation

Evaluations on 97 GIFT-Eval dataset configurations indicate that the ensemble approach provides superior robustness compared to individual models. While individual models like PatchTST-FM-r2 may occasionally outperform the ensemble on specific metrics (such as aggregate CRPS), the ensemble consistently achieves higher rankings across diverse datasets. The primary benefit is improved reliability across varied series without the need to identify the optimal architecture in advance.

Implementation

The ensemble is accessible via the granite-tsfm library. The system is designed to be extensible, allowing users to swap or add compatible forecasters to the existing interface. 

from tsfm_public.models.ensemble.modeling_ensemble import QuantileEnsembleForecaster

# Load the default ensemble
ensemble = QuantileEnsembleForecaster.from_pretrained(
    "ibm-granite/granite-timeseries-ensemble-r1",
)

# Generate forecasts
result = ensemble(
    df.tail(1024),
    timestamp_column="date",
    target_columns=["OT"],
    context_length=1024,
    prediction_length=24,
)

Limitations and Considerations

Users should be aware of several operational constraints:

  • Resource Usage: The ensemble downloads and runs four checkpoints at runtime, which increases memory consumption and inference latency compared to single-model deployments.
  • Independent Channels: The current implementation processes multivariate channels independently and does not leverage cross-channel correlations.
  • Calibration: While the ensemble improves aggregate accuracy, it does not guarantee calibration or robustness against significant distribution shifts. Backtesting on representative data is recommended to verify performance for specific use cases.

Sources