IBM ships a four-model time-series ensemble that skips weight fitting, edging the best single model on GIFT-Eval

Time-series forecasting gains an out-of-the-box default: four pre-trained Granite models return point and quantile forecasts in one call, so users need not first decide which architecture fits their data. Pre-trained time-series models can forecast without training on a dataset, but the choice of model still falls to the user, and a model that handles smooth industrial signals may struggle with noisy retail demand or a long seasonal horizon. Across 97 GIFT-Eval dataset configurations, the probabilistic ensemble scores a MASE of 0.683 and the IQR ensemble 0.684, both below 0.685 for PatchTST-FM-r2, the best of the four members.

Two combination rules are offered, and neither is fitted from historical forecast errors. The probabilistic ensemble uses uniform probability pooling, giving each member equal influence and merging quantile forecasts in probability space before extracting the requested quantiles; the IQR ensemble weights members by their forecast spread, giving narrower intervals greater influence, with weights varying by sample, target and horizon step. Both recipes were evaluated without fine-tuning on the 97 GIFT-Eval configurations, with test targets excluded from selecting or fitting the aggregation; the IQR ensemble takes the lowest average CRPS rank among these six forecasts, while PatchTST-FM-r2 keeps the lowest aggregate CRPS/MWQL score at 0.467.

The recipe targets developers forecasting many series over several horizons who would rather not commit to one architecture in advance, and its evaluated scope is zero-shot forecasting on 97 GIFT-Eval configurations; the combined forecast is not guaranteed to beat individual members on every series, and the stated path is to backtest the probabilistic ensemble, the IQR ensemble and a strong member such as PatchTST-FM-r2 on representative data.

The MASE lead over the strongest member is small, PatchTST-FM-r2 still holds the lowest aggregate CRPS/MWQL, and aggregate error and average rank therefore point to different choices; the current recipe forecasts multivariate channels independently without cross-channel information and loads four member checkpoints, raising inference time and memory over a single model; neither aggregation method guarantees calibration or robustness under distribution shift.