The validation framework
This is exactly how a strategy earns its place on QuantRhino — every test it must pass, and the kinds we deliberately let it fail. It is versioned and dated; material changes produce a new dated version. Nothing here is a promise about the future: every step below is running in production right now.
- Discovery
- Stress test
- Unseen data
- Forward live
- Certified
1 · Data
Binance spot OHLCV, May 2018 through today, across eight timeframes (5m to 1M) — strategy discovery runs on the seven from 15m up, with 5m data fetched for context. Bars are fetched complete, deduplicated, and indexed in UTC. Every published result is independently reproducible on standard charting platforms. Sub-2018 history exists on other venues and informs research context, but no published result depends on it.
2 · The five gates
complete history, clean bars, warm-up period excluded from scoring.
cheap rejection of structures with no plausible edge; roughly 95% of candidates die here.
tuned on past data, then required to stay profitable on data it never saw, net of fees.
performance is broken down by calendar year — bull, bear and chop, including 2022.
must survive thousands of random-scenario stress tests and a live forward run before it ships.
3 · Execution model
Simulation reproduces a realistic broker: bar-by-bar fills, partial take-profits, trailing and break-even stops, no same-bar re-entry, the terminal position left open. Real fees apply on every trade. Two metric families are kept — a per-fill view and the stricter per-entry view that reflects what a member’s account actually experiences. Gates run on the stricter one.
4 · Failure cases (kept on purpose)
The framework's value is what it rejects. Recent, real examples:
A volume-confirmation filter looked profitable in-sample. Out-of-sample it was negative in four of six configurations — the filter was noise. The block family was rejected.
Entries matched the venue exactly (191 = 191 trades) but the win rate drifted −1.5 percentage points through exit-path ambiguity on three bars. Reproducible, therefore rejected. The standard is exact.
At real taker fees, 40 of 40 candidate structures died at screening. There is no discoverable edge on that timeframe at those costs, so we don't sell one.
A learned filter on top of validated entries was tested on three timeframes (AUC 0.503 / 0.407 / insufficient data) and made results worse. The mission was closed and documented. The engine publishes what survives; it does not predict.
5 · What we do not claim
- —We do not promise profit, win rates, or any future performance.
- —Backtested results are hypothetical and labeled as such; they assume the stated fees and fills.
- —Forward results are published as they accumulate, including losing periods.
- —QuantRhino provides research tooling and information, not personalized investment advice.
6 · On the mechanism
QuantRhino uses modern computational methods, including machine learning, inside a validation framework designed to be indifferent to how a hypothesis was generated. We talk about the framework because the framework is what protects you. Where a learned method failed to add value, that result is published in section 4 and the method is not used.
Trading involves substantial risk of loss. Past performance does not guarantee future results.