Datasets:

Goooddy
/

crypto-lob-stream

license: mit
task_categories:
  - time-series-forecasting
tags:
  - finance
  - cryptocurrency
  - order-book
  - market-microstructure
  - high-frequency
pretty_name: Binance BTC/ETH/SOL L2 Order Book & Trades
configs:
  - config_name: depth
    data_files: depth/binance/*/*.parquet
  - config_name: trades
    data_files: trades/binance/*/*.parquet
  - config_name: snapshots
    data_files: snapshots/binance/*/*.parquet

Binance BTC/ETH/SOL - L2 Order Book, Trades & Snapshots

Continuous Level-2 order book (depth diffs), trades, and periodic full-book snapshots for BTCUSDT, ETHUSDT, and SOLUSDT, captured from Binance's public WebSocket feeds with crypto-lob-stream and published monthly as a contribution to open market-microstructure research.

Full depth-of-book data is paywalled or licensed for most asset classes; crypto is the exception, where the raw feeds are genuinely public. This dataset exists to lower that data barrier for independent researchers and students.

Layout

depth/binance/{ASSET}/YYYY-MM.parquet        # order book diff events
trades/binance/{ASSET}/YYYY-MM.parquet       # executed trades
snapshots/binance/{ASSET}/*.parquet          # full-book anchors for replay

Files are Snappy-compressed Parquet, compacted to one file per month per asset, and carry an explicit exchange column - i.e. the native schema of the crypto-lob-stream package, so its reconstruction helper works directly.

Reconstructing the order book

Depth rows are diffs, not a standing book. Replay a snapshot plus every subsequent diff, pruning to your intended depth after each update, to rebuild the book at any instant. The package does this for you, ghost-level-safe:

pip install crypto-lob-stream

from crypto_lob_stream import reconstruct
book = reconstruct("./depth", exchange="binance", asset="BTCUSDT")
bids, asks = book.top(n=10)

Coverage & limitations

Reconstructable from 2026-06-03 onward. Every depth file here carries first_update_id / last_update_id, so the diff sequence can be replayed against the snapshots. Earlier data lacks sequence ids and is not published.
Known gap — July 2026: capture paused ~~2026-07-05 20:56 to ~21:39 UTC (~~43 min) for all three symbols across depth and trades, due to a host restart. A fresh snapshot was written on reconnect, so books before and after the gap reconstruct cleanly; do not replay a diff sequence across it.
Reconstruct each market on its own - never merge books across exchanges.
Check the diff sequence for continuity before replaying across a boundary.