Datasets:

EvalData
/

residualbench

This guide reproduces the four headline findings reported in the paper "ResidualBench: A Benchmark and Evaluation Protocol for Cross-Model Forecast Failure Mode Discovery" (NeurIPS 2026 Evaluations & Datasets Track).

Hardware

1 GPU with >=12 GB VRAM (we used NVIDIA A100/3090; CPU works but is slow on N-BEATS / TimesNet / TopK-SAE).
~50 GB free disk for residual artifacts at H=96.

Environment

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

This installs the residualbench package (pyproject.toml) and dev dependencies (pytest, ruff, mypy).

Sanity check:

pytest tests/test_residualbench.py -q

Step 1 — Datasets (~5 min)

python scripts/download_data.py --include-large

Downloads ETTh1/h2, ETTm1/m2, Weather, Electricity, Traffic, Exchange, ILI from their public hosts to data/. Total ~3 GB.

Step 2 — Train forecasters and collect residuals (~6-10 GPU-hours)

Trains all 5 forecasters (DLinear, PatchTST, iTransformer, N-BEATS, TimesNet) with seed 42 on all 9 datasets at H=96, dumping residuals to results/benchmark/<dataset>/<forecaster>/residuals.pt.

python scripts/regimes/run_full_benchmark.py --phase forecasters --device cuda

For the seed-robustness study (Appendix K, 90 additional residuals):

python scripts/regimes/run_multi_seed_forecasters.py --gpu 0
python scripts/regimes/run_multi_seed_forecasters.py --gpu 1   # if available

Step 3 — Run the decomposition harness (~30 min CPU + ~30 min GPU)

python scripts/regimes/run_full_benchmark.py --phase methods --device cuda

Runs the 7 decomposition methods (PCA, Dense AE-16, Dense AE-64, k-means, TopK SAE, ICA, Spectral) under the proper train/test protocol on each (dataset, forecaster) pair, totalling 855 of 945 configurations (Spectral is omitted on 6 high-dim datasets; see paper Section 3.4).

Step 4 — Reproduce headline findings

Finding 1 (no single method dominates) and Finding 3 (proper protocol)

python scripts/regimes/evaluate_regime_metrics.py
python scripts/regimes/generate_figures.py        # paper Tables 3, 4 + Figs 2, 3

Finding 2 (cross-forecaster sharing)

python scripts/regimes/evaluate_regime_stability.py --alignment hungarian
python scripts/regimes/analyze_trivial_similarity.py            # Appendix F
python scripts/regimes/compute_cross_seed_all.py                # Appendix K
python scripts/regimes/plot_alignment_sources.py                # Figure 4

Finding 4 (lag-1 selector + learned selector)

python scripts/regimes/evaluate_selectors.py --base-dir results/benchmark

Statistical tables (Appendix G, J)

python scripts/regimes/compute_bootstrap_wilcoxon.py

Seed-robustness study (Appendix K)

python scripts/regimes/compute_forecaster_seed_robustness.py

Multi-horizon stability (Appendix I)

python scripts/regimes/run_multi_horizon.py --datasets ETTh1 Weather \
    --horizons 48 96 192 336

Expected outputs

After Step 4 you should have:

results/benchmark/analysis.json — main numbers cited in Sections 4.1-4.2
results/benchmark/selector_comparison.json — selector / lag-1 numbers
results/benchmark/bootstrap_ci.json, wilcoxon_full.json — Appendix J/G
results/benchmark/forecaster_seed_robustness.json — Appendix K
paper/figures/*.pdf — all paper figures regenerated

End-to-end smoke test (~20 min on a single GPU)

For reviewers who want to verify the pipeline end-to-end on a single small dataset:

python scripts/regimes/run_full_benchmark.py \
    --datasets ETTh1 --forecasters dlinear patchtst --device cuda
python scripts/regimes/evaluate_regime_metrics.py --datasets ETTh1

This runs only ETTh1 with DLinear and PatchTST (the two cheapest forecasters) and produces a partial analysis.json that should match the ETTh1 row of paper Table 3 within seed noise.

Troubleshooting

CUDA OOM on Electricity/Traffic. Reduce N-BEATS/TimesNet model sizes via --model-scale 0.5 (paper Section 3.1 documents this).
Spectral clustering hangs on high-dim datasets. Expected; we omit Spectral on Electricity/Traffic/ILI/Exchange/Weather/ETTm1 and report 855 rather than 945 configurations.
ILI Hungarian alignment is heterogeneous. Expected stress case (3 test windows at H=96); paper Section 4.2 reports 4-11x null on the other 8 datasets and discusses the ILI exception.