README.md

metadata

tags:
  - ml-intern
  - crypto
  - finance
  - backtesting
  - binance
  - time-series

crypto-quant-eval-data

Curated market data for the Avi-L/crypto-quant-eval strategy evaluation harness.

File	Rows	Freq	Content
bars_BTCUSDT_1h.parquet	59,654	1h	BTCUSDT OHLCV + funding_rate_last + funding_tick, 2020-01 → 2026-10
bars_BTCUSDT_1d.parquet	2,487	1d	same, daily
funding_BTCUSDT.parquet	5,295	8h	Binance USDT-M perp funding rate
funding_ETHUSDT.parquet	5,295	8h	same, ETH
onchain_btc_daily.parquet	~4,800	1d	BTC on-chain metrics (active addresses, market cap, fees, issuance, tx count, NVT)
manifest.json	—	—	sha256 of every output + source attribution

Sources

• Bars: vaquum/binance_btcusdt_1m_klines, resampled 1m → 1h/1d
• Funding: Torch-Trade/btcusdt_perp_funding_8h_05_2021_to_02_2026 and ETH twin (MIT)
• On-chain: coinmetrics-community btc.csv (no usable on-chain repo exists on the Hub)

Convention: timestamps are bar START times (UTC). A bar starting at t closes at t+freq; a signal computed at t may only use bars strictly before it. The harness enforces this with purge/embargo splits.

Known gaps: ETH spot bars not included (source is 34.6 GB); funding coverage starts 2021-05 (bars start 2020-01, so funding_rate_last is null before that — harmless for the harness, which forward-fills).

Usage:

from crypto_eval.data import load_local
data = load_local("<dir with these files>", "BTCUSDT", "1h")
data.assert_sane()