mordecaiusm/kerno-microstructure

Kerno Crypto Microstructure Events

A daily, verifiable dataset of tick-level microstructure events for BTC and ETH on Binance (spot and USDT-M perpetual) and Bybit (perpetual). Each event comes with point-in-time features and the realized outcome 10 s and 30 s later.

It is built to be the dataset you would want for machine learning on market microstructure, without the usual traps:

• No look-ahead. Features use only trades at or before the event. Outcomes use only later trades.
• Realistic entry. The outcome is measured from the first trade 250 ms after the event, not from the print that triggered it.
• Deterministic and reproducible. Every file is regenerated bit-for-bit from the exchanges' public dumps by open-source code with the pinned dependency versions (requirements.txt). The manifest records the exact source SHA-256s, output SHA-256s, code version and parameters.

Files

Config	Path	One row per
events	events/exchange=*/symbol=*/date=*/events.parquet	detected event
summary	summary/date=*/summary.parquet	stream × day
basis	basis/pair=*/date=*/basis_1m.parquet	minute (Binance spot vs USDT-M perp)
—	manifest/date=*.json	day: provenance and parameters

events columns

Column	Meaning
exchange, symbol	binance/bybit; BTCUSDT = spot, BTCUSDT-PERP = perpetual
exchange_trade_id, event_time_ms, price	the trade that triggered the event
spike_bps, spike_dir	tick move that triggered it (bps, ±1)
bucket	size vs the trailing 5-minute distribution: MEDIUM (≥p75), LARGE (≥p90), EXTREME (≥p99)
status	RESOLVED, or NO_DATA if the tape had a gap > 5 s at an outcome time
price_entry	first trade ≥ event + 250 ms
price_10s, price_30s	first trade ≥ event + 10 s / 30 s
ret_10s_bps, ret_30s_bps	(exit − entry) / entry × 1e4
f_*	point-in-time features (spike z-score/percentile, Roll spread, realized volatility 5 s/1 m/5 m, aggressor-flow imbalance, burstiness, trade density, …)
feature_version, engine_version	definitions used

An event is a trade whose tick return is ≥ 1 bp and at least the 75th percentile of the non-zero tick returns of the previous 5 minutes, with at most one event per stream per second.

summary headline columns

continuation_rate_10s is the share of events whose 10 s move continued the spike direction. share_abs_ret_10s_gt_cost is the share that moved more than the reference round-trip cost (cost_bps).

Verify it yourself

git clone https://github.com/mordecaiusm922-create/Kerno && cd Kerno/trading-obs
pip install --require-hashes -r requirements.txt && pip install --no-deps -e .
kerno dataset build --date 2026-09-20 --out check/
# compare check/manifest/date=2026-09-20.json "outputs[*].sha256" with this repository's manifest

Sources and license

Derived from the public trade dumps published by Binance (data.binance.vision) and Bybit (public.bybit.com). Raw trades are not redistributed here.

The dataset is licensed CC BY-NC 4.0: free for research and non-commercial use with attribution. For commercial use, fresher data or an API, open an issue in the GitHub repository.

Not investment advice. This is descriptive market data. Past event outcomes do not predict future returns, and nothing here is a recommendation to trade.