datasets.csv

id,category,verdict,checked,measurement
DS-01,cross-platform resolved prediction markets,CANDIDATE,2026-07-24,"987,250 resolved markets across a prediction market + an event market, 7 parquet shards, 406 MB download / 3.2 GB uncompressed, no auth. Grade B because it is TERMINAL-STATE, not a time series: outcome_prices is the settled vector, so it answers base rates and calibration at scale but cannot backtest entry timing. Last modified 2025-11. Earlier v1-v4 cuts: skip."
DS-02,regulated event-market executed trades,CANDIDATE,2026-07-29,"5.3 GB / 20 parquet, cc-by-4.0. an event market executed trades — a venue whose own history is a known ghost. Caveat that caps it: last updated 2026-01-30, six months silent, so it is a frozen sample rather than a feed."
DS-03,perp-venue node fills by block (frozen copy),CANDIDATE,2026-07-29,"106.0 GB / 1,078 parquet of a perp venue node fills by block, last updated 2025-12-08. Interesting precisely BECAUSE it is frozen: the live bucket a perp venue-mainnet-node-data keeps node_fills as a 64-day WINDOW (2025-05-25..07-27 as of 2026-07-29, see our internal tooling), so a frozen copy may hold fills that have already aged out upstream and are otherwise unrecoverable. Verify the actual date range inside before relying on it."
DS-04,perp-venue node trades (frozen copy),CANDIDATE,2026-07-29,"34.4 GB / 540 parquet, frozen 2025-12-08. Same author and same shape as the verified mirror: file names carry the upstream lz4 key (batch_upto_20250324_14.lz4_...), i.e. node_trades out of a perp venue-mainnet-node-data. Value is the frozen window again — the live bucket rolls, this copy does not. Reconcile with verify_hf_hl_mirror.py before use; no license on the card."
DS-05,"prediction-market order book, single asset",CANDIDATE,2026-07-29,"255.3 GB / 27,838 files (zip + jsonl) of a prediction market BTC order book; sibling poly-sol-orderbook for SOL. No license on the card, last updated 2026-06-02 — a snapshot, not a stream. Take a SLICE against a specific probe (execution calibration on PM), never wholesale: 255 GB would push the object storage bill to ~$<dataset> "
DS-06,live prediction-market snapshot collector,CANDIDATE,2026-07-29,"LIVE collector, not a dump: three products (market_snapshots / order_book_depth / top_holders) partitioned date=YYYY-MM-DD, a new run every ~6 min (1,684 runs over 2026-07-23..07-29), cc-by-4.0, 938 MB so far => ~4 <dataset> Read one snapshot: best_bid, best_ask, FULL <dataset> (median 31 bid levels, max 194), source_timestamp + source_latency_ms, 100 markets x 2 outcomes per run. This is the thing data-source-quality.md calls impossible — a retained PIT <dataset> series for a prediction market (previously only <dataset> Jan-Mar 2026 5-min slice). Open before it can carry a backtest: WHICH 100 markets (the run sampled <dataset> alongside politics, so the universe is the collector's choice, not the venue), and provenance is a third party's collector (polylake-0.1.0). Window starts 2026-07-23 and only grows forward — mirroring early costs pennies, starting late cannot be undone."
DS-07,prediction-market wallet-level user data,CANDIDATE,2026-07-29,"110.5 GB / 14,687 parquet, cc-by-4.0, updated 2026-07-06. Wallet-level a prediction market user data — the smart-money substrate for PM, the analogue of what our internal notes does with a perp venue positions. Composition not inspected yet."
DS-08,perp-venue replica commands (sample only),REJECT,2026-07-29,"2.0 MB / 21 parquet — a sample, not the replica_cmds archive (the upstream product is orders of magnitude larger). Same for <dataset> a perp venue-l4-data (2.5 KB, metadata only) and a perp venue-data (n<1K). Nothing to use."
DS-09,retail-terminal working set (FX/metals),REJECT,2026-07-29,"2.7 GB but it is the author's <dataset> working set, not a market archive: XAUUSD_H1_training.parquet, raw_ticks_7d.parquet (306 MB, one week), native_rates_M1/<dataset> , plus 16 .py and a .mqh. Seven days of gold ticks with no provenance does not compete with a commercial data vendor for the one thing we need (free <dataset> on non-crypto)."
DS-10,DEX pair OHLCV,REJECT,2026-07-28,"an L1 chain pair OHLCV, last updated December 2024. Dead source — rejected on staleness alone."
DS-11,CEX futures order-book re-publication,REJECT,2026-07-29,"apache-2.0, 3,530 downloads, 10M-100M rows of a large CEX futures order book. Rejected as redundant, not as bad: we already ingest a large CEX Vision directly (lake_ingest.py), which is the primary source and needs no provenance check. Same for <dataset> "
DS-12,L1 chain archive (accounts/tokens/txs),REJECT,2026-07-28,"CC-BY-4.0 an L1 chain archive. Corrected twice: it DOES carry txs/ with <dataset> token balance changes (pool reserves per block, i.e. T1 without a commercial data vendor), and the 2024-11..2025-01 season we care about is a legitimate target. Still rejected on measured coverage: txs/ is 477 daily partitions spread over 17 months, and the hole falls exactly on what we need."
DS-13,"US equity daily prices, 1962-2026",TRAP,2026-07-28,"Advertises 7000+ US stocks, 1962-2026, and contains NOT ONE delisted name — 1,935 tickers in the RELI-ZZZ shard, all nine probed dead names absent (TWTR, TIF, XLNX, UTX, WORK, VIAB, TCO, TSS, ZAYO). The card says nothing. This is the most direct survivorship-bias source we have found. Sibling Bonds-Daily-Price / ETFs-Daily-Price are gated:manual — not worth the time."
DS-14,prediction-market 5-min order-book slice,USE,2026-07-20,"5-min orderbooks + crypto_prices + markets-with-outcomes, Jan-Mar 2026. The only historical PM order-book DEPTH slice found free. Narrow window, so it calibrates depth assumptions rather than backing a full backtest."
DS-15,prediction-market dump (markets/trades/users),USE,2026-07-20,"a prediction market <dataset> parquet; markets.parquet carries outcome_prices, closed, end_date, neg_risk. Grade B: reconstruct price series from trades (executed prints, not quotes) and volume-filter — illiquid markets have the same last-print disease as an event market. Fixed cut date, stale between refreshes. Already wired into our internal tooling"
DS-16,perp-venue L1 misc events by block (frozen copy),USE,2026-07-29,"54.9 GB / 698 parquet, frozen 2025-12-08, mirror of misc_events_by_block (_src column carries the upstream key). INSPECTED one 257 MB shard (1,261,816 rows), correcting my own earlier claim: there are NO liquidation events in it. Event types are LedgerUpdate (6,139 per 60k rows), ValidatorRewards, CWithdrawal, CDeposit, Delegation and Funding. Liquidations on a perp venue are fills flagged in node_fills, so that is the dataset to use for them, not this one. What this dataset DOES carry is bigger than what I went looking for: each hourly Funding event holds ~221,000 deltas covering ~76,800 UNIQUE USERS across ~180 coins, each with user, coin, szi (position size), funding_amount and funding_rate. That is an hourly position snapshot of the ENTIRE venue, historically — where our own P0 collector takes top-N accounts forward-only. Caveat: szi and funding, no entry or liquidation price."
DS-17,mirror of a paid perp-venue archive + own book capture,USE,2026-07-29,"TWO classes under one name, and only one is the archive. (1) <dataset> (2.15 TB): parquet containers of 1000 rows with key = the original object storage key and blob = that object's bytes. RECONCILED 2026-07-29: 6/6 sha256 identical to object storage://a perp venue-archive (verification cost $0.0001 of paid egress); rerun with our internal tooling, which journals to source_health.jsonl. (2) <dataset> , l4_*, candles/, trades/ (519 GB + change): the author's OWN normalized websocket capture (coin, bid_px_1..N, ask_px_1..N) — nothing upstream to diff against, provenance unstated, so never call it 'the archive'. Total 76,990 files / 2.7 TB, updated 2026-07-26. Value: the hub egress is free where object storage charges the reader $0.09/GB, so order-book sampling stops costing money. No licence on the card => research substrate, not republishable."
DS-18,"US equity minute bars, 1992-2026",USE,2026-07-28,"US minute bars 1992-2026, 413 monthly parquet, 81.7 GB measured 2026-07-29 via the tree API (our internal notes estimated ~120 GB by multiplying shard sizes; the real number is 47% lower). Does NOT fit container disk — slice or to-object storage. Survivorship VERIFIED by name (<dataset> FRC present in 2022-06; <dataset> in 2015-06). Prices as-traded, NOT split-adjusted. Includes <dataset> (08:00-23:27 UTC) — session filter is on us. No license on the card, provenance unstated: research substrate only, and one day must be reconciled against our a commercial data vendor grouped panel before anything promotes."