IBM Releases Granite Timeseries Ensemble R1 for Zero-Shot Forecasting

IBM shipped one recipe across four forecasting architectures

IBM Research published Granite Timeseries Ensemble R1 on October 22, 2026. The public, ungated Hugging Face repository is ibm-granite/granite-timeseries-ensemble-r1. It stores reusable configuration recipes rather than a merged checkpoint, so the runtime downloads each selected member from its own repository. IBM’s launch example installs granite-tsfm 0.3.11 or newer and loads the default with QuantileEnsembleForecaster.from_pretrained.

The pinned default config names four members: ibm-granite/granite-timeseries-patchtst-fm-r1, ibm-granite/granite-timeseries-patchtst-fm-r2, ibm-granite/granite-timeseries-flowstate-r1 at revision r1.1, and ibm-granite/granite-timeseries-ttm-r3. They represent patched Transformer, patched Conformer, state-space and lightweight mixer approaches. The product decision is therefore a four-model local inference stack, not a new monolithic foundation model or hosted endpoint.

The default returns a median and forecast ranges

Each member produces quantile forecasts. The default linear_pool recipe gives every member equal influence in probability space, returns quantiles from 0.1 through 0.9 and uses the 0.5 quantile as its point forecast. IBM calls this the Probabilistic Ensemble. It requires no training, historical validation set or fitted weight vector, which makes its aggregation behavior easier to reproduce than a learned combiner.

The alternative IQR recipe varies influence by the width of each member’s predicted interquartile range, then uses isotonic regression to prevent crossed output quantiles. Narrower is not necessarily more accurate or better calibrated: IBM explicitly warns that this method can favor an overconfident member. Its temperature and maximum-weight controls are tuning surfaces, not evidence that the adaptive recipe will fit a particular demand, capacity or sensor series.

The reported point-error gain is real but narrow

IBM evaluated the methods without fine-tuning on 97 GIFT-Eval dataset configurations. In the October 20 comparison published with the launch, the Probabilistic Ensemble reported 0.683 MASE, the IQR Ensemble 0.684 and PatchTST-FM-r2 0.685; lower is better. Both ensembles also beat the four members shown on average MASE rank. These are IBM-reported results derived from a public benchmark, not AccessAllGPT reproductions.

The margin over PatchTST-FM-r2 is only 0.002 MASE at the displayed precision. The single PatchTST-FM-r2 checkpoint also retained the lowest aggregate probabilistic error shown: 0.467 CRPS/MWQL versus 0.471 for both ensembles. The IQR method had the best average CRPS rank among the six displayed forecasts, illustrating that aggregate error and per-configuration rank answer different questions. No one row wins every decision.

Two official result snapshots do not have identical ranks

The launch article says its leaderboard scores and ranks were captured on October 20, 2026. The pinned ensemble card labels its rank standings as September 28 and shows different average-rank values, while some displayed aggregate scores also differ slightly in precision or value. A changing reference leaderboard can move ranks as entries are added, and IBM says filtering changes the displayed peer set rather than the underlying full-leaderboard ranks.

Treat the date, recipe revision, member revisions, GIFT-Eval data and leaderboard snapshot as part of the benchmark record. Do not cite “third” or any other placement without the snapshot date and eligibility rules. More importantly, neither snapshot establishes accuracy, interval coverage or economic value on a buyer’s own series.

Availability is open, while runtime cost remains local

The recipe, default member repositories and Granite TSFM implementation were publicly reachable during review. IBM labels the ensemble recipe and implementation Apache 2.0, and the repository describes the four default members as permissively licensed. Optional expanded recipes introduce separately licensed IBM Research checkpoints, including non-commercial or research-use restrictions, so the root recipe’s license summary must not be transferred to every custom combination.

There is no per-forecast API price because the reviewed release is a downloadable package and model-repository workflow. The current project configuration marks granite-tsfm as beta, requires Python 3.11 through 3.13 and pins substantial runtime ranges including PyTorch. Teams must price checkpoint storage and transfer, accelerator or CPU time, four-model memory, cold starts, preprocessing, backtests, monitoring and operator review.

The operational limits matter more than the leaderboard

IBM says multivariate channels are forecast independently, so the recipe does not learn cross-channel relationships. Member context limits, supported horizons, preprocessing and missing-value behavior also differ. Loading four checkpoints increases memory and latency relative to one member, and neither aggregation method guarantees calibrated intervals or robustness after a distribution shift.

A central interval from the 0.1 and 0.9 outputs has nominal 80% coverage, but that label is not observed coverage. Backtest interval coverage by series, horizon and regime before using it for inventory, staffing, energy or financial decisions. Add explicit fallback behavior for missing data, unsupported lengths, download failures, memory pressure and forecasts that violate physical or business constraints.

What forecasting teams should do next

Freeze one decision and one forecast origin protocol. Run the equal-influence ensemble, the IQR recipe and PatchTST-FM-r2 on identical rolling windows. Measure a point metric appropriate to the business loss, quantile loss, empirical 0.1–0.9 coverage, interval width, bias by horizon and failure slices. Record checkpoint commits, package version, preprocessing, context and horizon so later leaderboard or repository changes cannot silently rewrite the comparison.

Adopt the default only when its avoided error or model-selection effort pays for four-model latency and memory. Constrain it to series and horizons that pass coverage and stability gates. Wait when cross-channel structure, managed-service terms or a fixed support contract is mandatory. Reject it when one checkpoint matches quality at materially lower operating cost or when forecast errors can trigger irreversible action without review.

Disclosures

AccessAllGPT received no IBM or Hugging Face account, checkpoints, API access, credits, hardware, benchmark logs, training data, private briefing, review, payment or compensation for this article. IBM and Hugging Face did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank the recipe or its members. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with IBM, Hugging Face or Salesforce. Publication-wide relationships are listed on the disclosures page.