README.md

metadata

license: other
license_name: research-use
license_link: LICENSE
pretty_name: pm_ws — anonymised multi-venue crypto microstructure capture

pm_ws — anonymised multi-venue crypto microstructure capture

High-resolution market-microstructure research corpus: multi-day, multi-venue capture of crypto limit-order-book state, event streams, and settlement-label observations for short-horizon (5m/15m) binary-option markets and their underlying spot feeds.

All identifying values are anonymised at source. Instruments are stable c0..cN labels, venues are v0..vN labels, and identifiers (market ids, token ids, session/connection ids, hosts) are salted digests. The salt and the label map are not published and will not be. Prices, sizes, depth ladders, clocks, sequence numbers, and lifecycle structure are preserved exactly — the corpus is analysis-ready without the identity map.

Layout

path	stream	semantics
ws_l2/	feed_a_*	Event-driven websocket demux across venues: top-of-book updates (bookTicker, best_bid_ask, price_change), trades (aggTrade, last_trade_price), parent frame envelopes, market-lifecycle events, connection lifecycle markers. Silence between rows inside an alive epoch is data (no event occurred), not a gap.
pm_l2/	feed_b_*	Polled binary-market L2 book snapshots: 10-level bids/asks ladders, best bid/ask, depth aggregates, OFI inputs, trade prints ([ts, price, size, signed_aggressor]), reciprocal-side identity via salted token ids.
oracle/	feed_c_*	Settlement/label observation stream: availability clocks, authentication flags, outcome observation rows. Rows state explicitly what is and is not authenticated.

Files are gzip JSONL, one JSON object per row, named feed_<x>_<YYYYMMDD>[T<HH>].anon.jsonl.gz (whole-day or hour chunk, UTC).

Parquet derivative (recommended for analysis)

Every archive also exists as Parquet under parquet/<stream>/<name>.parquet (zstd, nested ladders as list columns, per-kind union schema with sparse nulls; fields whose wire type varies across rows are strings). The gz JSONL archives remain the canonical bytes; Parquet is the streaming-friendly derivative.

Stream without downloading:

import duckdb
duckdb.sql("INSTALL httpfs; LOAD httpfs;")
df = duckdb.sql("""
  SELECT kind, available_at_ts, bb, ba
  FROM 'hf://datasets/billyyiu747/pm_ws/parquet/ws_l2/feed_a_20260725.anon.parquet'
  WHERE kind = 'bookTicker' LIMIT 100000
""").df()

or with datasets/pyarrow: load_dataset("parquet", data_files="hf://datasets/billyyiu747/pm_ws/parquet/pm_l2/*.parquet", streaming=True). Column projection and predicate pushdown make partial reads orders of magnitude cheaper than fetching the gz archives.

Row model

• Dispatch on kind (ws_l2) / row schema (schema_version field where present).
• Clocks: source_event_ts (venue), recv_ts (capture), available_at_ts (earliest usable time — use THIS for causal feature construction), recv_mono_ns (monotonic).
• Continuity: session_id / connection_id / epoch_generation / epoch_id / sequence define continuity spans. Reconnects open a new epoch and are marked by explicit connection-close/open rows. Never compute features across epoch boundaries; missing intervals are unknown, not zero activity.
• Venue-prefixed enum values carry the venue label prefix (e.g. v0_connection_close).

Caveats

• Coverage is not uniform across days; some days have polled-stream coverage gaps (marked by lifecycle rows, never silently).
• oracle/ rows with authenticated=false are observations, not certified labels.
• This is raw research capture, not a curated benchmark: validate row-level invariants before use.

Provenance

Produced by an allowlist-based, fail-closed anonymisation pipeline: any unclassified field aborts the export rather than shipping it. Output is scanned against the identity map before publication; a scan hit refuses the file.