Datasets:
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.
Heads up: a naive replay (only dropping a level when quantity hits 0)
accumulates "ghost levels" over time: price levels that fell out of scope
but were never explicitly zeroed out, so they just sit there looking real.
There's a nasty failure mode worth knowing about: a replay can be
completely gap-free and still end up wrong.
Oliver Zehentleitner ran a 25.10-hour BTCUSDT test on this exact problem (gap-free, sequence-validated, no synthetic gaps at all) and found that a naive, unpruned cache grew to 20,758 bid levels and 9,116 ask levels. At the final audit against REST, only 24.09% of those bid levels and 39.82% of the ask levels still matched. A pruned cache, checked the same way, held steady at 1,011 bids / 1,078 asks with 87.83% / 91.74% matching REST. So don't skip the pruning step, sequence continuity alone doesn't stop stale levels from piling up. Full writeup here.
The crypto-lob-stream
package (source: GitHub)
handles this for you and prunes automatically:
pip install crypto-lob-stream
from crypto_lob_stream import reconstruct
# point this at the folder containing trades/, depth/, snapshots/
book = reconstruct("./", exchange="binance", asset="BTCUSDT")
bids, asks = book.top(n=10) # best 10 levels per side, correctly pruned
By default it prunes Binance books to depth 1000 (that's the depth the
snapshot REST call requests. Binance itself doesn't cap the live diff
stream, so without pruning the book can genuinely drift past that over
time). Pass max_depth= to override, but I wouldn't disable it.
If you're not using Python, the manual version is: load the nearest
snapshot before your target window, discard diffs at or before its
last_update_id, apply the rest in ascending last_update_id order
(each diff's quantity is the new total, not a delta, 0.0 means
remove), and prune back to your intended depth after every single update.
That last step is the one that actually prevents the ghost-level problem
above.
Schemas
trades
| Field | Type | Notes |
|---|---|---|
| timestamp_ms | int64 | Event time (Unix ms) |
| exchange | string | binance |
| asset | string | e.g. BTCUSDT |
| trade_id | int64 | Exchange trade ID |
| price | float64 | |
| quantity | float64 | |
| buyer_maker | bool | True if the buyer was the maker |
depth (diff events)
| Field | Type | Notes |
|---|---|---|
| timestamp_ms | int64 | Event/receipt time (Unix ms) |
| exchange | string | |
| asset | string | |
| side | string | bid or ask |
| price | float64 | |
| quantity | float64 | 0.0 means the level was removed |
| first_update_id | int64 | Sequence start |
| last_update_id | int64 | Sequence number for replay ordering |
snapshots
| Field | Type | Notes |
|---|---|---|
| timestamp_ms | int64 | Snapshot time (Unix ms) |
| exchange | string | |
| asset | string | |
| side | string | bid or ask |
| price | float64 | |
| quantity | float64 | |
| last_update_id | int64 | Snapshot anchor for diff replay |
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. - Reconstruct each market on its own - never merge books across exchanges.
- Check the diff sequence for continuity before replaying across a boundary.
Known gaps
- 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 - treat the gap as a hard reset and do not replay a diff sequence across it.
Where this comes from
Collected with crypto-lob-stream
(GitHub, MIT licensed,
same as this dataset), a package I built for streaming L2 order book, trade,
and funding data off public exchange WebSocket feeds straight to Parquet.
Binance/BTC/ETH/SOL is what's published here, but the package also supports
Coinbase, OKX, Kraken, and Bybit (spot and perps) if you want to collect your
own. Issues and PRs welcome on the repo.