| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "cr": "http://mlcommons.org/croissant/", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "data": { |
| "@id": "cr:data", |
| "@type": "@json" |
| }, |
| "dataType": { |
| "@id": "cr:dataType", |
| "@type": "@vocab" |
| }, |
| "dct": "http://purl.org/dc/terms/", |
| "equivalentProperty": "cr:equivalentProperty", |
| "examples": { |
| "@id": "cr:examples", |
| "@type": "@json" |
| }, |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileProperty": "cr:fileProperty", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "format": "cr:format", |
| "includes": "cr:includes", |
| "isLiveDataset": "cr:isLiveDataset", |
| "jsonPath": "cr:jsonPath", |
| "key": "cr:key", |
| "md5": "cr:md5", |
| "parentField": "cr:parentField", |
| "path": "cr:path", |
| "recordSet": "cr:recordSet", |
| "references": "cr:references", |
| "regex": "cr:regex", |
| "repeated": "cr:repeated", |
| "samplingRate": "cr:samplingRate", |
| "replace": "cr:replace", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform" |
| }, |
| "@type": "sc:Dataset", |
| "conformsTo": "http://mlcommons.org/croissant/1.0", |
| "name": "ResidualBench", |
| "description": "ResidualBench is a benchmark for systematically comparing methods that discover forecast failure modes from residual windows. It includes pre-computed residuals from five forecasters (DLinear, PatchTST, iTransformer, N-BEATS, TimesNet) across nine named datasets (ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity, Traffic, Exchange, ILI) and a Python harness implementing seven decomposition methods (PCA, Dense AE-16, Dense AE-64, k-means, TopK SAE, ICA, Spectral) under a proper train/test evaluation protocol with motif-level metrics validated against synthetic ground truth.", |
| "url": "https://huggingface.co/datasets/EvalData/residualbench", |
| "version": "0.1.0", |
| "license": "https://opensource.org/licenses/MIT", |
| "citeAs": "@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}}", |
| "datePublished": "2026-05-06", |
| "creator": [{"@type": "sc:Person", "name": "Anonymous Author(s)"}], |
|
|
| "rai:dataCollection": "Pre-computed forecast residuals (r = y - y_hat) were generated by training five forecasters (DLinear, PatchTST, iTransformer, N-BEATS, TimesNet) with input window 336 and forecast horizon 96 on the temporal train splits of nine canonical long-horizon forecasting datasets, then dumping (input, target, prediction, residual) tuples on the held-out validation and test windows. Three independent training seeds (42, 11, 22) are released for every (dataset, forecaster) pair, supporting the seed-robustness analyses in Appendix K of the paper.", |
| "rai:dataPreprocessingProtocol": "Per-channel standardization computed on the training split only, applied identically to validation and test. Temporal 6:2:2 split for ETT datasets, 7:1:2 for Weather, Electricity, Traffic, Exchange and ILI (matching prior long-horizon work). No shuffling. AdamW optimizer, one-cycle learning-rate schedule, gradient clipping at norm 1.0, mixed-precision training, early stopping on validation MSE with patience 10. N-BEATS and TimesNet hidden sizes are reduced on Electricity (321 channels) and Traffic (862 channels) to fit a 24 GB GPU; all other forecaster hyperparameters are held constant across datasets.", |
| "rai:dataAnnotationProtocol": "No human annotation. Pre-computed residuals are deterministic functions of the publicly available source datasets and the trained forecasters. The benchmark also exposes per-window motif-assignment labels produced by each decomposition method; these labels are computed automatically by the released code and are not human-curated.", |
| "rai:dataUseCases": "(i) Evaluating new residual decomposition methods (sparse coding, autoencoders, clustering, regime-switching) under a shared cross-forecaster harness with proper train/test discipline. (ii) Studying cross-architecture failure mode similarity for time-series forecasting research. (iii) Designing portfolio / model-selection strategies that switch between forecasters based on recent error feedback. (iv) Diagnostic analysis of deployed forecasters for failure modes that aggregate metrics conceal.", |
| "rai:dataLimitations": "The four ETT datasets share the power-transformer domain, so cross-domain generalization claims are weakest there; Appendix L of the paper reproduces the four headline findings on the five non-ETT datasets. The ILI dataset has only three held-out windows at horizon 96 and is treated as a small-sample stress case throughout the paper. Spectral clustering is omitted on six high-dimensional datasets where it exceeded a 30-minute compute budget, yielding 855 of a possible 945 (dataset, forecaster, method, seed) configurations. Forecasters are trained at a single horizon (96) for the main results; cross-horizon stability on H in {48, 96, 192, 336} is reported on ETTh1 and Weather only (Appendix I).", |
| "rai:dataBiases": "All nine source datasets are widely used long-horizon forecasting benchmarks and inherit any selection biases of the original collection processes (sensor placement for ETT, station coverage for Weather, customer subset for Electricity, fixed loop detectors for Traffic, country basket for Exchange, US-only for ILI). Cross-forecaster failure-mode sharing is therefore measured on these specific datasets and may not extend to deployments outside the energy / weather / electricity / traffic / finance / medical mix studied here.", |
| "rai:personalSensitiveInformation": "ResidualBench contains no personally identifying information. None of the source datasets release individual-level personal records: ETT records industrial transformer measurements; Weather and Electricity report aggregated sensor or substation-level series; Traffic reports highway loop counts; Exchange reports country-level FX rates; ILI reports CDC-level weekly counts of influenza-like illness with no patient identifiers.", |
| "rai:socialImpact": "ResidualBench is a diagnostic and benchmarking artifact for time-series research and is not intended for direct deployment. The released lag-1 model selector and learned selectors are illustrations of how failure modes can drive portfolio strategies; we explicitly recommend domain-specific validation and human oversight before any safety- or finance-critical deployment, and the paper's broader-impact statement makes this constraint explicit.", |
| "rai:dataSocialImpact": "We anticipate that systematic failure-mode benchmarks lower the cost of detecting silent failures in production forecasters (energy load, traffic, electricity demand) and enable better targeted retraining; the same machinery could in principle be used to optimize forecasters for adversarial financial trading, which we do not condone.", |
| "rai:isSyntheticData": false, |
| "rai:hasSyntheticData": "Optional synthetic motif templates (5 canonical residual patterns with controlled noise) are bundled for metric validation; they are clearly marked, are not used for any of the four headline findings, and exist only to ground the cohesion / VR metrics against ground truth. Synthetic windows live under the path 'synthetic/' inside the artifact and are never mixed with real-world residuals.", |
|
|
| "prov:wasDerivedFrom": [ |
| {"@type": "sc:Dataset", "name": "ETT (Electricity Transformer Temperature)", "url": "https://github.com/zhouhaoyi/ETDataset", "license": "CC-BY-4.0"}, |
| {"@type": "sc:Dataset", "name": "Weather", "url": "https://www.bgc-jena.mpg.de/wetter/", "license": "CC-BY-4.0"}, |
| {"@type": "sc:Dataset", "name": "Electricity", "url": "https://archive.ics.uci.edu/dataset/321/electricityloaddiagrams20112014", "license": "CC-BY-4.0"}, |
| {"@type": "sc:Dataset", "name": "Traffic (PEMS)", "url": "https://pems.dot.ca.gov/", "license": "Public domain (US government)"}, |
| {"@type": "sc:Dataset", "name": "Exchange Rate", "url": "https://github.com/laiguokun/multivariate-time-series-data", "license": "CC-BY-4.0"}, |
| {"@type": "sc:Dataset", "name": "ILI (Influenza-like Illness)", "url": "https://gis.cdc.gov/grasp/fluview/fluportaldashboard.html", "license": "Public domain (US government)"} |
| ], |
| "prov:wasGeneratedBy": { |
| "@type": "prov:Activity", |
| "rdfs:label": "ResidualBench residual generation pipeline", |
| "description": "Forecasters trained with the released configs in configs/ via 'python scripts/regimes/run_full_benchmark.py --phase forecasters', then residuals dumped via the same script's collection phase. Multi-seed extension produced via scripts/regimes/run_multi_seed_forecasters.py." |
| }, |
|
|
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "residualbench-package", |
| "name": "residualbench-package", |
| "description": "Pip-installable Python package implementing the fit/encode/reconstruct method protocol, motif-level metrics (cohesion, VR), Hungarian alignment, lag-1 / learned selectors, and the proper train/test harness.", |
| "contentUrl": "https://huggingface.co/datasets/EvalData/residualbench/resolve/main/residualbench-0.1.0.tar.gz", |
| "encodingFormat": "application/gzip", |
| "md5": "861fa4ba9c0fa8a7bd3027eb2f8c9792" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "residual-tensors", |
| "name": "residual-tensors", |
| "description": "Pre-computed forecaster residuals as PyTorch tensors. One file per (dataset, forecaster, seed) tuple, totalling 9 datasets x 5 forecasters x 3 seeds = 135 files. Each file contains keys 'train_input', 'train_target', 'train_pred', 'train_residual' and the corresponding 'test_*' splits.", |
| "containedIn": {"@id": "residualbench-package"}, |
| "encodingFormat": "application/x-pytorch", |
| "includes": "results/benchmark*/{ETTh1,ETTh2,ETTm1,ETTm2,Weather,Electricity,Traffic,Exchange,ILI}/{dlinear,patchtst,itransformer,nbeats,timesnet}/seed_{42,11,22}/residuals.pt" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "results-summary", |
| "name": "results-summary", |
| "description": "Aggregated benchmark results JSON: one row per (dataset, forecaster, method, seed) cell with cohesion, VR, and reconstruction MSE for the 855 reported configurations.", |
| "contentUrl": "https://huggingface.co/datasets/EvalData/residualbench/resolve/main/results/all_results.json", |
| "encodingFormat": "application/json", |
| "md5": "05b15308f0652c013aea85e9ac7d49e8" |
| } |
| ], |
|
|
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "@id": "benchmark-cells", |
| "name": "benchmark-cells", |
| "description": "One record per (dataset, forecaster, method, seed) benchmark cell.", |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "benchmark-cells/dataset", |
| "name": "dataset", |
| "description": "Source dataset identifier (one of ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity, Traffic, Exchange, ILI).", |
| "dataType": "sc:Text", |
| "source": {"fileObject": {"@id": "results-summary"}, "extract": {"jsonPath": "$[*].dataset"}} |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "benchmark-cells/forecaster", |
| "name": "forecaster", |
| "description": "Base forecaster identifier (dlinear, patchtst, itransformer, nbeats, timesnet).", |
| "dataType": "sc:Text", |
| "source": {"fileObject": {"@id": "results-summary"}, "extract": {"jsonPath": "$[*].forecaster"}} |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "benchmark-cells/method", |
| "name": "method", |
| "description": "Decomposition method identifier (PCA, Dense_AE_16, Dense_AE_64, k-means, TopK, ICA, Spectral).", |
| "dataType": "sc:Text", |
| "source": {"fileObject": {"@id": "results-summary"}, "extract": {"jsonPath": "$[*].method"}} |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "benchmark-cells/seed", |
| "name": "seed", |
| "description": "Random seed for the decomposition fit (42, 11, or 22).", |
| "dataType": "sc:Integer", |
| "source": {"fileObject": {"@id": "results-summary"}, "extract": {"jsonPath": "$[*].seed"}} |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "benchmark-cells/cohesion", |
| "name": "cohesion", |
| "description": "Motif cohesion metric (paper Section 3.2). Higher is better; bounded in [0, 1].", |
| "dataType": "sc:Float", |
| "source": {"fileObject": {"@id": "results-summary"}, "extract": {"jsonPath": "$[*].cohesion"}} |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "benchmark-cells/vr", |
| "name": "vr", |
| "description": "Variance ratio (within-motif vs across-motif residual variance). Higher (more positive) is better.", |
| "dataType": "sc:Float", |
| "source": {"fileObject": {"@id": "results-summary"}, "extract": {"jsonPath": "$[*].vr"}} |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "benchmark-cells/recon_mse", |
| "name": "recon_mse", |
| "description": "Reconstruction mean squared error of the decomposition method on test residuals.", |
| "dataType": "sc:Float", |
| "source": {"fileObject": {"@id": "results-summary"}, "extract": {"jsonPath": "$[*].recon_mse"}} |
| } |
| ] |
| } |
| ] |
| } |
|
|
