0xarchive-backtest

A deterministic backtester for Hyperliquid perps, on 0xArchive data. The engine is Rust; strategies are Python. The same strategy, data and parameters give the same result hash on every machine, so a result is a claim anyone can check.

It is the engine behind 0xarchive-quant, a quant research skill for coding agents: hand Claude Code or Codex a wallet, an idea or a script, and it comes back with a verdict you can rerun.

Install

pip install "0xarchive-backtest[curated]"

Prebuilt for macOS, Linux and Windows; Python 3.12 or later (3.11 without [curated], the pinned numpy, pandas and scipy that strategies may import).

Quickstart

Pull a minute dataset with your 0xArchive API key. Months are inclusive. Free keys read only the last 30 days, so this September window needs a paid plan:

export OXARCHIVE_API_KEY="0xa_your_api_key"
0xbt data pull --markets BTC,ETH,SOL,HYPE --from 2026-09 --to 2026-09

Write a strategy:

# momentum.py: each day, long the strongest of four perps over 3 days and short the weakest
import oxarchive_backtest as oxb
from oxarchive_backtest import x


def strategy(lookback="3d"):
    return oxb.Strategy(
        universe=["BTC", "ETH", "SOL", "HYPE"],
        clock=x.daily(),
        signal=lambda h: h.close.ret(lookback),
        weights=x.long_short(0.25),
    )

Backtest it, then put it through the gauntlet:

0xbt backtest momentum.py --start 2026-09-01 --end 2026-10-01
0xbt gauntlet momentum.py --start 2026-09-01 --end 2026-10-01 \
  -g lookback=1d,3d --card card.md

The backtest prints the return, Sharpe, drawdown, fees and funding, and a result hash. The gauntlet runs every check on the best setting and reads the verdict off a fixed ladder. This one stops at fails nulls: one in ten random books did as well, over the 5% bar, so a month of momentum on four markets is not an edge.

The gauntlet

0xbt gauntlet runs, in order: the grid as trials; costs and a benchmark; nulls (the flipped signal, random books and the strategy's own signal at random lags, redrawn at its holding period); a split and walk-forward; the deflated Sharpe over every try; and robustness. The verdict is the first rung whose rule fails:

0xbt study init starts a trial ledger: every backtest below it is recorded, the deflated Sharpe counts all of them, and the last months of data stay sealed until one candidate is committed, once.

What the engine models

Minute bars with taker flow, funding, open interest and liquidations; taker and maker fees; funding payments; leverage limits and liquidation; point-in-time universes (x.top_n, x.rule) so delisted markets are included; stops, take-profits, trailing and time exits checked every minute; and deterministic math, so results don't depend on the platform.

Links

• Source, the agent skill and examples
• The strategy API
• 0xArchive

MIT licensed.