Altarino1/dataset-quality-notes

Dataset Quality Notes

A worked method for deciding whether a public time-series dataset can carry a backtest — with the measurements that came out of applying it. 18 datasets assessed, updated 2026-07-30.

Datasets are listed by CATEGORY, not by name. The point here is the method and the failure modes, not a directory of links: a name invites "is mine on the list?", while a category invites you to run the checks on whatever you are about to use.

The four questions, and why each one earned its place

1. How much does it actually weigh? The declared size bucket is the uploader's claim. Summing the file tree gives the real number — one dataset estimated at ~120 GB by multiplying shard sizes measured 81.7 GB, a 47% error in the direction that makes you skip a usable source.
2. Is it still alive? A last-modified date is a rejection criterion on its own. Two datasets in this table were dropped for nothing but staleness.
3. Is it complete, or only complete-looking? For equities this is survivorship. One popular daily-price dataset advertises 7000+ US stocks over 1962-2026 and contains zero delisted tickers: ten dead names probed, ten absent, while two living names return 11,481 and 10,156 rows. Nothing on its card says so. Anything cross-sectional built on it measures the returns of companies that happened to survive.
4. Where did the bytes come from? One dataset mirrors a requester-pays archive. Six objects were hashed on both sides: 6/6 sha256 identical, so the mirror is genuine — but only for that one product. Other directories in the SAME repo are the author's own capture with no upstream to diff against, and the card does not distinguish them. Provenance is per-product, not per-repo.

Files

File	What
datasets.csv	one row per assessed dataset: category, verdict, date checked, and the measurement behind it
survivorship_probes.csv	per-name presence results — the check that separates a real universe from a survivor list
provenance_checks.csv	byte-level checks: object size on both sides and the hash verdict

Verdicts

• USE — measured, usable, with the caveats in the note
• CANDIDATE — measured and promising, one blocking check still open
• TRAP — looks right, is wrong; the note says how it lies
• REJECT — measured, and it does not solve the problem it appears to solve. Often it means redundant (the primary source is already available), not bad

Method notes worth stealing

• Search matches repo NAMES, not descriptions. A long natural-language query returns nothing and reads as "this data does not exist"; the one-word query returns rows. An empty result is evidence about the search, not about the world.
• The dataset viewer is not a fact source. It errors on repos whose files do not share one format, and its filter endpoint needs ~40-60s of index warm-up before it answers at all. An empty answer from it says nothing about the data.
• A frozen copy can beat a live one. Some upstream archives keep only a rolling window (one keeps ~64 days), so a third-party copy that stopped updating months ago may hold data the source itself no longer has.
• Reading costs requests, not bytes. Measured through the rate-limit headers: ~13 range reads for a single-file query on a 376 MB parquet, ~25 per file across a 12-file scan. Against a 5,000 requests / 5 minutes ceiling, one full pass over a 413-file dataset is ~10,300 requests — it fails halfway. Repeatedly scanned data belongs in storage that bills requests instead of capping them.
• Check the contents, not the product name. A dataset named for one event type turned out to contain none of it — and something more useful instead. One shard, ten minutes.

Limitations

Verdicts are dated measurements, not endorsements: recheck before relying on one. The notes describe what was measured for one specific use, and nothing here is advice about trading anything.