Not the line, the touch
Notes · BTC · 2026-05-06A long-EMA signal“felt useful” across timeframes. Six rounds of walk-forward; one form survived.
Sharpe by round, in-sample vs walk-forward / holdout. Surviving forms highlighted.
The signal was the 400-period exponential moving average — a line this desk has looked at for years. Round one tested dual-EMA crossovers, four timeframes, forty-nine parameter combinations. The in-sample winner reached a Sharpe of 1.10. Out of sample, with monthly re-selection, it fell to 0.16.
The state of being above or below a line is regime-dependent. A rule that wins in one regime is the wrong rule in the next, and a walk-forward selector thrashes between them. Three rounds of variants — single-line versions, daily timeframe, four mechanism splits — produced the same shape.
Round four reframed the question. Stop measuring state. Measure events. A specific fifteen-minute pattern — price tags the EMA, then closes back through within the same bar — produced a walk-forward Sharpe of 1.87. The true holdout produced thirteen trades at Sharpe 2.251 — too few to validate, kept as a provisional observation, not a result. Every filter added on top of it hurt.
The line is not the signal. What happens at the line is.
- 1Thirteen trades is too few to validate. We keep the holdout result as a provisional observation, not a result; every filter added on top of it hurt. ↩