模型發布／時序預測
IBM Granite Time-Series Model Switches to Conformer and Overlapping Patches, Outputs 99-Quantile Forecasts
PatchTST-FM-r2 combines convolution-assisted attention, an 8,192-step context window, and probabilistic forecasting to target general-purpose, zero-fine-tuning time-series prediction. The model ranks second in GIFT-Eval’s reproducible zero-shot track, but the results were submitted by IBM and still await confirmation through integration into the benchmark project.
IBM Research has released Granite Time Series PatchTST-FM-r2, a roughly 385-million-parameter model, along with open weights, an inference pipeline, and architecture code. Users can choose between the Apache 2.0 and OpenMDW 1.0 licenses. Rather than simply training the existing PatchTST on more data, IBM replaced its standard Transformer blocks with Conformer blocks: two half-step feed-forward layers surround multi-head self-attention and temporal convolution, with convolution kernels alternating according to `{5,5,3,3}`. This lets convolution handle local variations while allowing attention to focus more on long-range dependencies. Other changes listed in the [model card](https://huggingface.co/ibm-granite/granite-timeseries-patchtst-fm-r2) include increasing the number of blocks from 20 to 30, adding pre-head layer normalization, and using patches of length 16 with stride 8, producing 50% overlap. Training applies Hamming-window weighting, while inference uses overlap-and-add to merge adjacent forecasts, with the goal of reducing discontinuities at patch boundaries.
The model supports contexts of up to 8,192 time steps. Its 99-quantile head can simultaneously produce point estimates, predictive distributions, and uncertainty intervals, and it can also perform missing-value imputation. The pretraining data combines public time series outside GIFT-Eval, data derived using KernelSynth and TSMixup, and approximately 500,000 CauKer synthetic sequences containing 4,096 steps each. This level of data disclosure helps assess potential test-set contamination, but it still cannot replace backtesting on domain-specific industry data.
IBM reports that, in GIFT-Eval’s zero-shot track—which excludes test leakage and requires reproducible code—the model achieved a CRPS geometric mean of 0.467 and a MASE of 0.6846, ranking second on both metrics. IBM also describes it as the highest-performing permissively licensed model in the track. The [release article](https://huggingface.co/blog/ibm-research/ibm-releases-sota-granite-time-series) also acknowledges that the results remain part of a leaderboard pull request awaiting integration, so they should currently be treated as publisher-reported evaluations. Engineering teams can reproduce the results using the public [Granite-TSFM repository](https://github.com/ibm-granite/granite-tsfm). More important areas to watch next include latency and calibration under irregular sampling, distribution drift, different forecast horizons, and online streaming workloads—not merely the model’s position on a single leaderboard.