Zipeng365/TSDecompose-Benchmark

TSDecompose Benchmark

This repository is the release bundle for Time-Series Decomposition as a Standalone Task: A Mechanism-Identifiable Benchmark (ICML 2026 camera-ready artifact).

It combines:

• a paper-aligned benchmark data release;
• a post-rebuttal second expansion of the real-data evidence;
• a source snapshot of TSDecompose / tsdecomp;
• machine-readable result tables for the synthetic benchmark, real-data proxy track, and semi-synthetic transfer checks.

This is a repository-style artifact, not a PyPI package release. The code is included so users can inspect and run the benchmark from source.

Public Hugging Face links:

• Dataset repository: Zipeng365/TSDecompose-Benchmark
• Web leaderboard: Zipeng365/TSDecompose-Benchmark-Leaderboard

What Is Included

data/paper_tables/
  Exact CSV versions of the camera-ready main and appendix tables, plus the
  raw 6-scenario x 50-draw six-family leaderboard used to regenerate Table 2.
  Appendix robustness/transfer tables are paper snapshots; not all of them
  are regenerated by the lightweight paper-core runner.

data/paper_figures/
  Paper figure assets used by the leaderboard Space for visual alignment checks.

data/synthetic_full22_extension/
  Benchmark-only 6-scenario synthetic extension with a 22-method roster.

data/real_proxy22/
  Real-data companion track with canonicalized public time series and proxy diagnostics.

data/semisynth_transfer/
  Semi-synthetic transfer summary and raw metric exports, excluding large downloaded source files.

data/post_rebuttal_second_expansion/
  Later real-data expansion from the rebuttal stage, separated from the paper-aligned tables.

code/TSDecompose/
  Source snapshot for the decomposition benchmark API and CLI. The `tsdecomp`
  Python package source is included at `code/TSDecompose/src/tsdecomp/`.
  The paper core benchmark runner is
  `code/TSDecompose/scripts/run_paper_benchmark.py`.

site_data/v1.0.0/
  Lightweight data files for a future leaderboard website or Hugging Face Space,
  regenerated from the current paper-aligned outputs.

metadata/
  Release manifest, schema notes, table provenance notes, file inventory, and
  checksums.

Paper-Aligned Scope

The primary paper studies standalone decomposition as component recovery under controlled synthetic mechanisms. The main leaderboard and paper tables should be read as diagnostic capability profiles, not as a universal single-score ranking.

The 22-method files are included as benchmark-only expansion and transfer tracks. Several rows correspond to additional method prototypes or benchmark-side mechanism proxies, not camera-ready paper claims. These rows use the same 6-scenario, 50-draw synthetic protocol where applicable, but they should not replace the primary six-family Table 2 / Figure 3 interpretation.

The one-command reproduction path in this repository regenerates the primary paper-core synthetic leaderboard. Appendix robustness and transfer checks are released as exact camera-ready CSV summaries, with source summaries and run metadata where available. They should be used to audit the paper tables, but they are not all exposed through the same lightweight runner.

Versioned Evidence Layout

The release separates the frozen camera-ready paper snapshot from benchmark extensions:

• Camera-ready paper snapshot: data/paper_tables/ and data/paper_figures/.
• Living benchmark extensions: data/synthetic_full22_extension/, data/real_proxy22/, and data/semisynth_transfer/.
• Post-rebuttal second expansion: data/post_rebuttal_second_expansion/.

The second expansion includes the real-data additions used in the rebuttal: mechanism-aware checks on CO2 and tides, plus a broader six-dataset proxy and stability panel. It is included for transparency and follow-up analysis, not to replace the camera-ready paper tables.

Core Task

Given an observed time series y, a decomposition method returns components:

(trend, seasonal, residual)

The synthetic benchmark evaluates recovery against known ground-truth components using:

• trend R2;
• trend DTW;
• seasonal R2;
• seasonal spectral correlation;
• seasonal max-lag correlation.

These five metrics are the paper's core component-recovery metrics. They are computed per generated draw after method outputs are aligned to (trend, seasonal, residual), then averaged by scenario, tier, or regime. Coverage reports the fraction of successful runs with valid metrics.

The camera-ready Table 2 view intentionally displays only two of the five core metrics: Trend R2 and Seasonal spectral correlation, each split over stationary regimes (Tiers 1-2) and non-stationary regimes (Tier 3). Its displayed values use a tier-balanced aggregation: first compute method means separately for Tier 1, Tier 2, and Tier 3 over valid metric values; stationary columns are the equal-weight average of Tier 1 and Tier 2 means, while non-stationary columns are Tier 3 means. Seasonal metrics are undefined for the trend-only scenario, so the by-tier file records metric-specific valid-row counts. Figure 3 should be read as the five-metric capability profile. The expanded 22-method files may include mean-rank convenience columns, but those are extension summaries and are not the primary paper definition.

Machine-readable metric definitions are in:

site_data/v1.0.0/evaluation_metrics.json

Data Tracks

paper_tables

Small CSV files matching the camera-ready manuscript tables. These are the safest files to cite directly when checking paper consistency. The paper-core reproduction files are:

• global_performance_summary.csv: the rounded Table 2 values used by the manuscript and leaderboard.
• paper_core_50draw_leaderboard.csv: raw 1,800-row paper-core output (6 scenarios x 50 draws x 6 methods).
• paper_core_50draw_by_tier.csv: tier-level means and metric-specific valid-row counts used to derive the tier-balanced Table 2 columns.

Appendix table snapshots include bounded tuning, period robustness, alignment robustness, boundary sensitivity, real-data proxy, semi-synthetic transfer, and MSSA pilot summaries. Their audit status is summarized in metadata/paper_table_provenance.md.

synthetic_full22_extension

This benchmark-only track uses the same 6-scenario synthetic generator and five-metric evaluation protocol with length 512, 50 draws per scenario, true-period-given evaluation, and a 22-method roster. It includes raw rows, overall summaries, by-scenario summaries, by-tier summaries, coverage, protocol matrix, and backend-selection metadata. It is an expanded roster view, while the camera-ready Table 2 / Figure 3 source is the six-family paper table under data/paper_tables/.

real_proxy22

This track uses public real time series with known periods, known mechanisms, or motivated trend expectations. Because real data rarely expose exact component ground truth, the metrics are proxy diagnostics: band plausibility, resampling stability, spectral overlap, residual autocorrelation, trend smoothness, and reconstruction error.

semisynth_transfer

This benchmark-only transfer track injects known mechanisms into six real monthly backgrounds using a 22-method roster, three mechanisms, two background scales, and eight windows per setting. The released five-metric ranking is data/semisynth_transfer/results/summary/ranking_paper_5metric_overall.csv. It is a living benchmark extension rather than a camera-ready paper table. Large downloaded source files are intentionally excluded from this release; the canonical small CSV backgrounds and metric tables are included.

The camera-ready paper's semi-synthetic paragraph is the frozen six-family snapshot in data/paper_tables/semisynthetic_transfer_summary.csv; the 22-method transfer files in this directory are a later benchmark extension and may have different method-level ordering.

post_rebuttal_second_expansion

This later companion track contains rebuttal-stage real-data evidence. real_physics_track_b covers CO2 and tides with mechanism-informed approximate structure. real_proxy_track_c covers CH4, GPCC precipitation, NDVI, QBO, Arctic sea ice, and sunspots using proxy and stability diagnostics.

Source Snapshot

The source tree is under:

code/TSDecompose/src/

The standalone tsdecomp package source is included directly in this Hugging Face repository:

code/TSDecompose/src/tsdecomp/

Compiled development-machine binaries are intentionally excluded; the package uses the pure-Python fallback path unless users build their own native extension.

To run from source:

cd code/TSDecompose
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
export PYTHONPATH="$PWD/src"
python scripts/run_paper_benchmark.py

On Windows PowerShell:

cd code/TSDecompose
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
$env:PYTHONPATH = "$PWD\src"
python scripts/run_paper_benchmark.py

This command runs the camera-ready core synthetic benchmark: 6 scenarios, 50 deterministic draws per scenario, and therefore 300 generated synthetic series. The default method set is the six-family Table 2 roster (ma_baseline,stl,ssa,emd,vmd,wavelet). Because each generated series is evaluated by each selected decomposition method, the raw leaderboard.csv has one row per scenario, draw, seed, and method. The public data/paper_tables/paper_core_50draw_leaderboard.csv file is this raw output after stable row sorting; paper_core_50draw_by_tier.csv and global_performance_summary.csv are deterministic aggregations of it. The script pins imports to the bundled source snapshot, so local editable installs of other tsdecomp versions will not change the run.

Minimal smoke run:

python scripts/run_paper_benchmark.py --smoke

Equivalent direct CLI call for the full paper core run:

python -m tsdecomp run_leaderboard --suite core --methods ma_baseline,stl,ssa,emd,vmd,wavelet --seeds 0 --n_samples 50 --length 512 --dt 1.0 --out artifacts/paper_core_benchmark --aggregate

The script entrypoint is recommended for reproduction; the direct module CLI is included for clean environments where this source tree is the resolved tsdecomp package.

The native extension binary from the development machine is not included. The release uses the pure-Python fallback path by default. To test a separately built native extension, set TSDECOMP_ALLOW_EXTERNAL_NATIVE=1.

Suggested Citation

@inproceedings{wu2026tsdecompose,
  title = {Time-Series Decomposition as a Standalone Task: A Mechanism-Identifiable Benchmark},
  author = {Wu, Zipeng and Wei, Jiani and Zhou, Shiqiao and Chen, Jiajun and Spill, Fabian and Andrews, James W.},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year = {2026}
}

License

The release bundle is provided under the MIT License. External real-data sources retain their original terms; see data/real_proxy22/results/summary/dataset_discovery_log.md and the registry files for provenance notes.