Guide & data

The strategy. Donchian 55/20 long/flat on daily closes: go long when today's close makes a new 55-day high, sell to cash when it makes a new 20-day low. Never short, never leveraged, spot only. Signals use yesterday's completed window (no repaint) and act with a one-day lag; every backtest charges 0.10% per unit turnover.

Why the portfolio is the default view. Per-coin the signal is weak (78–91st percentile vs a scrambled null). Aggregated across 16 coins with inverse-vol weights it reached percentile 100 — diversification is the amplifier. Judging this system on one coin (HYPE, say) is exactly how you talk yourself out of a real edge or into a fake one.

What it buys you. Historically it wins by losing less in winters (2018: −7.6% vs −59.4%; 2022: −21.9% vs B&H's worse), and it lags every raging bull year. Time in market ~28%. The price of admission: stretches of up to ~26 months below the previous equity high. If you can't sit through that, the backtest is irrelevant.

Honesty features (always on). B&H is always plotted. The six canonical parameter neighbors are always shown — a real effect survives at 40/15 and 70/25; an overfit peak doesn't. Yearly windows are always shown. MAR (CAGR ÷ |maxDD|) is the headline, not the raw multiple. Longest-underwater is always visible.

The gauntlet. The Gauntlet tab reruns the frozen pre-registered test battery (SPEC.md, 2026-07-05): portfolio-vs-B&H bars, 6-variant robustness, yearly windows, the 2017 survivorship-resistant subset, a 500-rep joint 7-day-block scrambled null and a trade-level skew check. Verdict PASS needs all 6. Every run is logged permanently; the counter in the sidebar is a multiple-comparisons warning, not a decoration.

This is analysis tooling only. It places no orders, holds no keys, and knows nothing about your exchange accounts.