README.md

metadata

license: mit
language:
  - en
pretty_name: ResidualBench
size_categories:
  - 1K<n<10K
task_categories:
  - time-series-forecasting
tags:
  - benchmark
  - forecasting
  - residual-analysis
  - failure-modes
  - sparse-autoencoder
  - evaluation
  - neurips-2026-eandd

ResidualBench

A benchmark and evaluation protocol for cross-model forecast failure mode discovery, accompanying the NeurIPS 2026 Evaluations & Datasets Track submission "ResidualBench: A Benchmark and Evaluation Protocol for Cross-Model Forecast Failure Mode Discovery".

This repository hosts the small artifacts required for review and reproduction:

• croissant.json — Croissant 1.0 dataset metadata with the full RAI extension.
• residualbench-0.1.0.tar.gz — pip-installable Python package implementing the fit / encode / reconstruct method protocol, motif-level metrics, Hungarian alignment, and the lag-1 / learned selectors.
• results/*.json — pre-computed aggregated benchmark results across all 855 (dataset, forecaster, method, seed) cells.
• LICENSE, REPRODUCE.md — license and step-by-step reproduction guide.

The full per-window residual tensors (residuals.pt, ~123 GB) are not stored in this repository because of size. They can be regenerated from public source datasets in a few GPU-hours by following REPRODUCE.md Step 2 ("Train forecasters and collect residuals"). Optionally, a tar archive of the residuals can be uploaded as an additional release file via the upload_residuals_to_hf.py helper that ships with the source repository.

Provenance

ResidualBench builds on nine public time-series datasets (cited in croissant.json under prov:wasDerivedFrom):

Dataset	Source	License
ETTh1, ETTh2, ETTm1, ETTm2	https://github.com/zhouhaoyi/ETDataset	CC-BY-4.0
Weather	https://www.bgc-jena.mpg.de/wetter/	CC-BY-4.0
Electricity	UCI ML Repository (321)	CC-BY-4.0
Traffic (PEMS)	https://pems.dot.ca.gov/	Public domain
Exchange	https://github.com/laiguokun/multivariate-time-series-data	CC-BY-4.0
ILI	CDC FluView	Public domain

ResidualBench itself does not introduce a new raw dataset.

Headline numbers (re-derivable from results/all_results.json)

• 855 of 945 configurations (Spectral omitted on 6 high-dim datasets).
• Cross-forecaster Hungarian alignment 0.23–0.57, $4{-}11\times$ above shuffled null on most datasets.
• Lag-1 model selector beats best fixed forecaster by up to 17 % on temporal hold-out.
• Cohesion (TopK, mean of 5 forecasters $\times$ 3 seeds): ETTh1 0.481, ETTh2 0.712, ETTm1 0.656, ETTm2 0.615, Weather 0.652, Electricity 0.558, Traffic 0.304, Exchange 0.899, ILI 0.962.

Reproducing

pip install residualbench-0.1.0.tar.gz
python -c "from residualbench import ResidualBench; print(ResidualBench.__doc__)"

For end-to-end regeneration of the residual tensors and benchmark numbers, see REPRODUCE.md.

Citation

@inproceedings{residualbench2026,
  title={ResidualBench: A Benchmark and Evaluation Protocol for Cross-Model Forecast Failure Mode Discovery},
  author={Anonymous},
  booktitle={NeurIPS 2026 Evaluations and Datasets Track},
  year={2026}
}

License

MIT for the harness and metadata. Each upstream dataset retains its original license; see croissant.json for details.