For a while, our crypto research backtest produced an eye-watering long-only multiple versus BTC buy-and-hold. That number was real as a spreadsheet output. It was not a credible promise of what live trading would deliver.
So we stopped leading with it.
What that number actually was
It came from a full-history re-sim of a regime filter on daily closes: long BTC when the model was risk-on, cash otherwise. No leverage. Still, a multi-decade-style multiple over a few years is almost always a warning label, not a headline.
Path dependence, one crypto cycle, optimistic costs, and marketing that freezes a lucky equity curve will destroy trust the first time the live book has a dull quarter.
What we publish instead
- Research backtest — full re-sim of the production algorithm, with friction, labeled as research.
- Live paper ledger — equity from scores that were actually published to the database, lagged one session, with friction and a size-scaled path.
- Tradable output — LONG / FLAT / NO_TRADE, a size cue from 0 to 1, and a reliability grade.
If the live paper ledger underperforms a sexy re-sim, that is the point. Honesty compounds better than vanity multiples.
How to use the crypto model
Treat it as a permission layer, not an autopilot:
- Risk-on with solid reliability: you may size up within your own rules.
- Neutral or NO_TRADE: sit out or cut size.
- Never ignore your own risk limits because a chart from 2020 looked heroic.
The honest caveat
Past research results do not guarantee future results. Crypto is path dependent. Friction, gaps, and weekends are real. The product is daily context with a graded signal, not a guaranteed alpha factory.
Full numbers, including the live paper ledger when enough published days exist: macro-bias.com/crypto/track-record.