README.md

metadata

language:
  - en
license: mit
pretty_name: BTCUSDT Microbar v2
tags:
  - finance
  - cryptocurrency
  - market-microstructure
  - binance
  - futures
  - btcusdt
size_categories:
  - 10M<n<100M

BTCUSDT Microbar v2

Sub-candle microstructure data for Binance USD-M Futures BTCUSDT, collected continuously over six WebSocket streams. Successor to Torch-Trade/btcusdt-microbar.

A standard OHLCV candle compresses thousands of trades into 6 numbers. This dataset preserves the raw event-level data — every individual trade, every best bid/ask change, every depth snapshot — so the underlying microstructure features can be reconstructed at any timeframe.

Why v2?

In April 2026 we discovered that the v1 collector had a silent routing bug. Binance USD-M Futures split market data across two WebSocket routing paths (/public and /market) — the legacy unified endpoint silently delivered streams that mapped to /public and dropped the rest without error.

Concretely, in v1:

Stream	v1 status	v2 status
trade	working	working
bookTicker	working	working
depth5@500ms	partial (missing during routing migration)	working
markPrice	broken (subscriptions silently dropped)	fixed
miniTicker	broken (subscriptions silently dropped)	fixed
forceOrder	broken (subscriptions silently dropped)	fixed

If you only need trades and book_ticks, the v1 dataset remains usable. If you need funding rate, mark price, 24h stats, liquidations, or full depth — use v2. We renamed the repo (rather than appending) so consumers get a clean discontinuity instead of a silent quality jump.

The fix is in binance-microbar collector commit f88a52a (group streams by routing path and open one WebSocket per group).

Layout

data/btcusdt/
├── trades/<YYYY-MM-DD>/<HHMMSS>.parquet          # individual trades
├── book_ticks/<YYYY-MM-DD>/<HHMMSS>.parquet      # best bid/ask updates
├── depth/<YYYY-MM-DD>/<HHMMSS>.parquet           # top-5 order book snapshots
├── liquidations/<YYYY-MM-DD>/<HHMMSS>.parquet    # forced liquidations
├── mark_price/<YYYY-MM-DD>/<HHMMSS>.parquet      # mark price + funding
└── mini_ticker/<YYYY-MM-DD>/<HHMMSS>.parquet     # 24h rolling stats

Files are flushed to disk every 60 seconds. Each row carries a timestamp_ms (exchange clock, UTC) which is the only safe key for joins — wall-clock arrival times are not preserved.

Schemas

trades/

column	type	description
timestamp_ms	int64	exchange trade time, ms since epoch UTC
price	float64	trade price (USDT)
quantity	float64	trade quantity (BTC)
is_buyer_maker	bool	True = aggressive sell, False = aggressive buy

Volume: ~50 trades/sec → ~4M rows/day.

book_ticks/

column	type	description
timestamp_ms	int64	exchange event time
bid_price	float64	best bid price
bid_qty	float64	best bid quantity
ask_price	float64	best ask price
ask_qty	float64	best ask quantity

Volume: ~100 updates/sec → ~8M rows/day.

depth/

Top 5 order book levels. Columns: timestamp_ms, bid_price_0..4, bid_qty_0..4, ask_price_0..4, ask_qty_0..4. Snapshots arrive every 500 ms → ~170k rows/day.

liquidations/

column	type	description
timestamp_ms	int64	liquidation time
price	float64	average fill price
quantity	float64	liquidated size
side	string	"BUY" = short squeezed (bullish), "SELL" = long liquidated (bearish)

Sparse — bursts during volatile moves.

mark_price/

Updates every 3 seconds. Columns: timestamp_ms, mark_price, index_price, funding_rate, next_funding_time_ms. Funding rate is the most predictive macro signal in crypto futures.

mini_ticker/

Updates every second. Columns: timestamp_ms, open_24h, high_24h, low_24h, close, volume_24h, quote_volume_24h. Distance from 24h high/low is a support/resistance signal.

Loading

from huggingface_hub import snapshot_download
import pandas as pd
from pathlib import Path

local = snapshot_download(
    "Torch-Trade/btcusdt-microbar-v2",
    repo_type="dataset",
    allow_patterns=["trades/2026-04-29/*.parquet"],
)

trades = pd.concat(
    pd.read_parquet(f) for f in sorted(Path(local, "trades/2026-04-29").glob("*.parquet"))
)

To compute aggregated microstructure features over arbitrary timeframes, use the binance-microbar library — examples/build_feature_dataset.py rebuilds 54-feature ML-ready datasets directly from these raw streams.

Collection

• Source: Binance USD-M Futures public WebSocket streams (no auth required)
• Collector: binance-microbar at commit f88a52a or later
• Host: continuous collection on a Raspberry Pi 5 (colony1), uploads daily at 03:00 UTC
• Coverage: starts 2026-04-29; the immediate prior period (2026-04-28 → 2026-04-29) is missing because the collector was offline for the routing-fix deployment

License

MIT. Market data is sourced from Binance's public WebSocket streams and is provided as-is. No financial advice; not affiliated with Binance.