earino/noaa-tide-flooding

Dataset Factory · v2026.09

Dataset website · GitHub release · Hugging Face · Catalogue

Artifact: fb49f932c7b4ea5ab532c20f7dce59a7d733de925d1e172054ff1d38abe1eeed

Suitability and evaluation requirements

Retained for retrospective model-development research with documented timing limitations. Later evaluation rows expose earlier outcomes; predictions require restricted access.

Supported uses: Retrospective train/validation model-development experiments under the documented timing and selection limitations; Final evaluation only under a protocol that isolates holdout data from development and prevents later rows influencing earlier predictions

Not supported: Claims of verified real-time or operational forecasting validity; Treating development/validation scores as untouched final-test results; Forecasting scores from unrestricted cross-row access to this artifact

Evaluation access: chronological. The labeled eval split is development/validation data. Keep the final holdout outside agent development; prevent later rows from influencing earlier predictions.

Historical availability: assumed. Training-label timing: unverified. Vintage: current snapshot. Selection: unverified.

Cross-row audit: recovery demonstrated. Both split designs (temporal, station-disjoint) tested; temporal 88,915/89,038 covered AUC 0.99929 with 423 binary mismatches; station-disjoint 29,889/29,930 AUC 0.99922 with 191; a D+2 margin_ft variant still ranks ~0.9993, so deleting the obvious column is not a fix Holdout recovery was not tested by this audit.

Structural qualification and baseline scores do not certify operational forecasting validity.

• Near-perfect future-outcome ranking in both split designs (see cross_row_audit scope).
• Dictionary/build day-definition conflict: release/noaa-tide-flooding/code/build.py assigns the official minFlag to date+1 while DATA_DICTIONARY.md describes the target-day value.
• One calendar week straddles train/eval in each design (244 temporal, 82 station-disjoint eval rows).
• Historical feature/label availability is assumed from current snapshots, not verified against release vintages.
• The current scorer passes the complete feature frame; isolation is not mechanically enforced.

Corrections to the historical documentation:

• For row date D, observed_max_ft is the measurement on D; margin_ft and trailing summaries use observations preceding D (through D-1 when consecutive). The label is NOAA’s official minor-flood flag for D+1. It is not the same-day flag and is not defined by simply thresholding observed_max_ft. Historical verification/publication timing remains unverified.
• The original dictionary says carry columns are dropped. The bundled template excludes only target and id_columns; observed_max_ft and threshold_ft are therefore available to that baseline. This does not by itself expose its own D+1 label, but the documentation must not claim those columns were excluded.

Review: 2026-10-24; applies to v2026.09 / fb49f932c7b4ea5ab532c20f7dce59a7d733de925d1e172054ff1d38abe1eeed. Frozen data and tags are unchanged.

NOAA Tide Flooding Prediction (v2026.09)

Status: public. Released after explicit operator authorisation; publication changed no data file, tag or checksum.

Will a tide gauge's daily maximum observed water level exceed its published minor flood threshold tomorrow? 122 NOAA NOS stations, 2006-2025, 2,440 station-years, one row per station-day.

• Artifact version: fb49f932c7b4ea5ab532c20f7dce59a7d733de925d1e172054ff1d38abe1eeed
• Version tag: v2026.09 (use this revision for reproducibility)
• GitHub counterpart: https://github.com/earino/noaa-tide-flooding
• Rows: 1,396,692 across two levels

This is a derived dataset and not an official NOAA, NOS or NWS product. The source is NOAA CO-OPS public-domain data.

Loading

Version-pinned. revision is the tag, not a branch, so the bytes cannot change under you. There are two configs, one per level, because each is a complete task instance.

from datasets import load_dataset

ds = load_dataset("earino/noaa-tide-flooding", "temporal", revision="v2026.09")
print(ds)
# DatasetDict({
#     train: Dataset({features: [...], num_rows: ...}),
#     eval: Dataset({features: [...], num_rows: ...}),
#     holdout: Dataset({features: [...], num_rows: ...})
# })

# the label is `minor_flood`: 1 when the target day exceeded the station's threshold
print(ds["train"].features["late"])

The repository is public, so loading needs no credential:

ds = load_dataset("earino/noaa-tide-flooding", "station_disjoint", revision="v2026.09")

The files are also plain CSVs, so they can be used without datasets:

hf download earino/noaa-tide-flooding --repo-type dataset --revision v2026.09 \
    --include "temporal/*" --local-dir ./noaa

temporal/ and station_disjoint/ mirror the GitHub release layout exactly: public/train.csv, public/eval.csv, private/holdout.csv, meta.json, quality.json.

The two levels

config	train	eval	holdout	what it asks
temporal	2006-2021, all 122 stations	2022-2023	2024-2025	the benchmark-typical panel split
station_disjoint	2006-2021, 81 stations	2022-2023, 41 unseen stations	2024-2025, the same 41	whether a model learned something transferable rather than one station's local behaviour

The station-disjoint level is why this dataset exists. NOAA publishes the observations, the verified maxima and the annual counts, so access is not the contribution. A frozen multi-station panel, an observations-only feature contract, a station-disjoint generalisation split and a leakage gate around all of it are not things NOAA's products package.

The label

minor_flood is the binary target: 1 when the station's daily maximum observed water level exceeded its published nos_minor threshold on the target day, 0 otherwise.

• The label comes from NOAA's own minor-flood days (the HTF daily product's minFlag). Our own reduction of the 6-minute maxima overcounted flood days by about 20% against NOAA's counts on a year with real signal (113-114 against 94 over four stations, 2024), while NOAA's flags matched exactly, station by station. A label that disagrees with the publisher's own product has to be justified, and there is no reason to ship one.
• The threshold is the physical published value on station datum (STND). It is not a quantile chosen to balance the classes. Comparing MLLW-referenced heights against it silently yields zero positives at every station - a trap this dataset documents in meta.json and measures.
• Prediction time is the end of the previous local day. Features are observations only, with trailing windows, so the future cannot leak into a feature. NOAA tide predictions and model guidance are excluded: they are model output, and including them would hand the task over.

Only about 2% of station-days are positive. A constant "no flood" answer agrees with the label roughly 96% of the time. Score AUC, not accuracy.

Results, including the uncomfortable one

Baseline through the benchmark's own train.py/validate.py, copied verbatim into the GitHub counterpart:

level	recorded	independent reproduction
temporal	0.8633	0.8633
station_disjoint	0.8690	0.8690

The independent clean-room reproduction on the published bytes returned the recorded pair to the digit (job noaa-consumer-003, release 396698120). The runner's training is not bit-reproducible (xgboost, default threading, identical bytes), so the last decimal moves between runs - the earlier 2026-09-20 pass, against the superseded artifact, gave 0.8651 and 0.8687. The digests do not move.

The station-disjoint level scores the same on stations the model never saw, so the signal is not station-specific memorisation.

And the floor, measured without training: a single stock feature - yesterday's maximum against the station's own threshold - already reaches 0.8265 (temporal) and 0.8454 (station-disjoint). The trained model's real headroom is about +0.0368 and +0.0236. The headroom is narrow, and it is stated here rather than discovered by a reviewer. The floor was measured on the superseded artifact's eval split, which carries the same rows and positives as the released one, so the headroom is approximate. The full measurement is in the GitHub counterpart's measurements.json.

Licence

• This compilation (frozen station list, row selection, derived label, both partitions, feature construction): CC0-1.0.
• Code and documentation: MIT.
• NOAA's data: LicenseRef-Public-Domain (US Government work), preserved unchanged.

NOAA National Ocean Service, Center for Operational Oceanographic Products and Services (CO-OPS), verified water level observations and flood thresholds, https://tidesandcurrents.noaa.gov/, accessed 2026-09-18

NOAA's constraint - modified content must not be presented as official government material - is honoured: this dataset is labelled as derived everywhere it appears.