jpms5/imbalance-ts-hub

imbalance-hub

A curated, versioned catalog of imbalanced time series — "OpenML for imbalanced time series." Every series here has been scored with imbalance_eval (the relevance-function methodology from Moniz, Branco & Torgo, 2017) and kept only if it actually has a rare regime worth studying.

The catalog metadata and client library live in the companion GitHub repo, jpmsilva1/imbalance-ts-hub — start there for the schema, the scoring methodology, and how the catalog is built. This dataset repo holds the actual data: one Parquet file per accepted series.

Quick start

pip install git+https://github.com/jpmsilva1/imbalance-ts-hub.git

from imbalance_hub import load_catalog, pull

catalog = load_catalog()
severe_hourly = catalog[(catalog.imbalance_level == "severe") & (catalog.granularity == "H")]

series = pull(severe_hourly.id.iloc[0])   # -> pd.Series, values ready to use

# Filter by collection + severity + length before pulling anything.
candidates = catalog[
    (catalog.collection == "m4_monthly")
    & (catalog.imbalance_level.isin(["severe", "extreme"]))
    & (catalog.length >= 200)
]

load_catalog() fetches the metadata CSV from GitHub (not from this repo) and caches it locally, re-checking with a conditional request on the default version="latest" so an unchanged catalog isn't re-downloaded; pull() downloads only the one Parquet blob you asked for from here, verified against a content_hash recorded in the catalog. Browsing the catalog's 59k+ rows never downloads more than the metadata.

Dataset structure

One Parquet file per accepted series, at {source}/{collection}/{shard}/{key}.parquet (shard is a 2-hex-char directory keeping every folder under Hugging Face's per-directory file limit — an implementation detail, not part of the catalog id). Each file has:

Column	Type	Notes
value	float64	Raw values, in original order. NaNs preserved where the source series has missing points.
timestamp	datetime64	Present only when the source series carries real timestamps (TSLib); absent for GluonTS series, which are index-only.

The catalog id (source:collection:key, e.g. gluonts:m4_hourly:h1) is what maps an id to its blob path — see blob_path_for() in scripts/upload_blobs.py in the GitHub repo, or just use the catalog CSV's blob_path column directly.

Filterable columns

The columns you'll actually filter load_catalog()'s DataFrame on. Full column list/types and the complete numeric-distribution table live in the GitHub README's Catalog schema section.

Column	Values
source	gluonts (58,354), tslib (1,135)
license	cc-by-4.0 (28,868), unlicensed (29,486), unknown (1,135) — see Licensing
imbalance_level	moderate (32,867), severe (13,256), mild (12,194), extreme (1,172)
granularity	12 canonical values (normalized from source-inconsistent notation), from min up to Y — see the GitHub README's Filterable values for the full duration-ordered table
collection	55 distinct, from m4_monthly (16,998) down to single-series collections

Numeric ranges (min / median / max) for the columns most people filter on:

Column	Min	Median	Max
length	12	330	526,980
%Rare	0.01	8.70	100.00
IR	0	0.10	∞ (all-rare series, n_normal=0)

Dataset summary

• 59,489 accepted series out of 136,673 scanned (~43.5%) — the rest were excluded by the inclusion gate (no rare regime) or failed to embed (too short).
• 55 source collections, from GluonTS (58,354 series — M4, M3, electricity, traffic, tourism, and more) and TSLib (1,135 series — ETT, weather, exchange rate, national illness).
• Severity distribution: 1,172 extreme, 13,256 severe, 32,867 moderate, 12,194 mild (see the GitHub README for how imbalance_level is computed from %Rare).

Licensing

No single license applies — this catalog spans sources with different terms, checked against the original loaders rather than assumed. The curation layer (catalog metadata, imbalance-scoring methodology, pipeline code) is separate from the underlying values: the code is MIT-licensed, which does not extend to the data — see DATA_LICENSES.md for that scope split. For the raw values themselves:

• Most GluonTS collections (electricity, traffic, tourism_*, nn5_*, weather, wind_farms_*, m1_*, and most others fetched via GluonTS's _tsf_datasets.py) come from the Monash Time Series Forecasting Archive on Zenodo, which licenses its datasets under CC BY 4.0 — verified directly against the NN5 Daily record.
• M4 (m4_* — GluonTS's largest single contributor by series count) is fetched from M4Competition/M4-methods on GitHub. That repository publishes no license (confirmed via the GitHub API): no LICENSE file, no license declared anywhere. Under default copyright, that means no redistribution rights have been granted for this data, by imbalance-hub or by anyone else who mirrors it. It's rehosted here in line with common practice in the forecasting field, not because the rights question is resolved. If the M4 organizers ever ask, these series will be removed from the Hugging Face mirror on request.
• TSLib (ETT-small, electricity, exchange_rate, illness, traffic, weather) — the thuml/Time-Series-Library repo itself is MIT-licensed, but that covers the code; the bundled benchmark CSVs' original licenses aren't stated in that repo.

If you plan to use a specific series commercially or redistribute it standalone, verify that series' actual source collection rather than relying on this summary.

Citation

To cite the catalog itself:

@software{silva2026imbalancehub,
  author  = {Silva, João P. M.},
  title   = {imbalance-hub},
  year    = {2026},
  url     = {https://github.com/jpmsilva1/imbalance-ts-hub}
}

If you use this catalog's imbalance scoring, also cite the methodology it implements:

@article{moniz2017resampling,
  title={Resampling strategies for imbalanced time series forecasting},
  author={Moniz, Nuno and Branco, Paula and Torgo, Lu{\'\i}s},
  journal={International Journal of Data Science and Analytics},
  volume={3},
  pages={161--181},
  year={2017},
  publisher={Springer}
}