English | Русский

quant-ai-lab

A research lab built around one question: is there a tradeable edge in crypto perpetuals that survives an honest check? Not "build a bot" — get a conclusion that cannot be faked by fooling yourself.

The answer is in. 10 pre-registered hypotheses × 3 venues (Binance, Bybit, Hyperliquid), time-series / cross-sectional / event-driven signals — all KILL. No predictive alpha is visible on public data. The one positive cash flow found is harvesting funding on an inefficient venue (~7%/yr realistic in backtest), and that is tail-risk management, not alpha. This is a real conclusion, not a dead end.

And — more important for trusting that conclusion — the method is measured: the judge issuing every verdict was run across the whole decision surface on synthetic data with known ground truth. False-positive rate on noise: 0%. Detection threshold: ~2.2 annualised Sharpe. The best net Sharpe this project ever observed (0.88) sits at less than half that threshold — so the KILL stream is evidence about the markets, not a stuck detector.

That detector — and the pre-registration discipline behind it — is what this repo sells: a mechanical, calibrated PASS/KILL verdict, as a CLI, a Python library, or a hosted API, for anyone who needs an independent check on a strategy's returns before committing capital. See What's for sale, below.

Navigation

Document	What's there
RESEARCH-CONCLUSION.md	Research-phase conclusion: the registry of all 10 hypotheses, what it means, the decision
quant-harness/	The tool — an honest backtest framework (pip-installable, 198 tests). Start with its README
quant-harness/docs/METHODOLOGY.md	Methodology: the seven ways a backtest lies, and the defence against each
quant-harness/reports/CALIBRATION.md	Judge calibration — the measured detection power (FPR, power curve, robustness)
quant-harness/reports/FINDINGS-CALIBRATION.md	Calibration write-up and why the KILL record can be trusted
quant-harness/api/README.md	Hosted verdict API — endpoints, auth, payment gates, environment variables
docs/specs/	Hypothesis pre-registrations (specs written before touching data)
docs/HARVEST-PATH-A.md	Funding-harvest: real basis risk, costs, venue tail — a research position, not a live product
CASE-STUDY-EN.md	Public write-up of the project: what was built, what was found, why it's trustworthy
site/index.html	Landing page for the tool and the case study
NEXT_SESSION.md	Current status and open decisions for the owner (internal working file — not part of the public snapshot)

Quick start

cd quant-harness
pip install -e ".[dev,api]"
python -m pytest                  # 198 tests (dev+api extras), including calibration guards
python examples/quickstart.py     # KILL on noise / PASS on a real edge, no data needed
qh-audit --returns examples/audit_sample_returns.csv --trials 20   # PASS/KILL verdict on a CSV

What's for sale

The output of this lab is not a trading signal — it's the verdict tool built to check one honestly. Two ways to use it:

• qh-audit — a CLI and Python library that turns a CSV of per-period returns into a mechanical PASS/KILL verdict (Combinatorial Purged CV, Deflated Sharpe, PBO, four pre-registered gates), plus a Markdown/JSON report. Runs locally, no network call.
• Hosted verdict API (quant-harness/api/) — the same judge behind a FastAPI service (POST /v1/verdict): upload a returns matrix, get the verdict, metrics, cost sensitivity, and a short report back over HTTP. Built so an AI agent can request an independent check as easily as a human can.

Who it's for: allocators, prop desks, bot marketplaces, and AI-agent builders who need an independent, calibrated verdict on a strategy's returns — their own, a vendor's, or a counterparty's — before capital is committed.

What it is not: not a signal generator, not trading advice, not a promise of returns. A KILL means "not visible on this data, at these costs, after correcting for how many configurations were tried" — not "no profit exists anywhere." See quant-harness/README.md for the full picture, including the boundaries of what a PASS does and doesn't mean.

Why this is worth trusting

1. Discipline carried through to a conclusion. 10 honest pre-registrations, none "pushed" to a PASS. Most retail backtests break on the first of seven common ways to fool yourself; all seven are closed here — structurally.
2. A mature tool. quant-harness is a reusable honest-backtest framework: leakage-safe WFA/CPCV, realistic costs, Deflated Sharpe, PBO, a mechanical verdict — with proven, measured detection power (see Navigation, above).
3. An honest negative result is still a result. "No edge visible on this data at these costs" is knowledge that saves capital — and it's stated as a KILL, not spun as a near-miss.
4. The funding-harvest research result is disclosed honestly. A live paper track (started 2026-07-22, cron/launchd-driven, no capital at risk) currently sits negative — about −1.5% over six weeks, roughly −13%/yr annualised — against a ~7%/yr realistic backtest estimate. Both numbers are shown; neither is advertised as a return you should expect.

---

Stack: Python 3.11+, numpy/scipy/pandas, FastAPI for the hosted API. Data: public Binance/Bybit dumps, the Hyperliquid API. Tool license: MIT. Repository: https://github.com/imadeptus/quant-harness.