Sharpe Ratio

Mean per-trade return divided by the volatility (the standard deviation, or spread) of those returns. It measures edge per unit of noise — how much return you earned for the bumpiness you endured. Higher is better; here it is computed per trade, not annualized.

In plain English

Two strategies can earn the same average return. The one whose trade-to-trade results are smoother — less wild swinging between big wins and big losses — has the higher Sharpe ratio. It is the "quality-adjusted" return: reward divided by variability.

• High Sharpe → a steady edge, returns clustered tightly around a positive mean.
• Sharpe near 0 → the average return is buried in the noise; you can't tell skill from luck.
• Negative Sharpe → losing, on average.

Because Sharpe divides by volatility (the spread of outcomes), it is the natural input to the question "is this edge real?" A small Sharpe needs far more trades to prove out than a large one. That test lives in sharpe significance.

(Note: this fleet reports a per-trade Sharpe — return per trade ÷ spread per trade. It is not annualized, so the numbers look small next to the "Sharpe above 1" figures quoted for funds.)

Why it matters for this fleet

Across the 210 EMA-cross variants, the per-trade Sharpe ranges from −2.68 to 0.557, with a median of −0.058 — 135 of 210 (64%) are negative. The typical EMA-cross config has, on average, a slightly-losing, noisy edge. The positive Sharpes are small: even the best survivors sit around 0.1 to 0.26.

Small Sharpes are the whole story for trend-following. The edge per trade is thin; it only adds up over a large sample size — and only if the sample is big enough to distinguish that thin Sharpe from zero.

Examples from the live fleet

• id523 (EMA 21/50 · SOL · 1h · 2× · long) — Sharpe 0.110 on 436 trades. Thin, but on enough trades that the edge test clears: a Sharpe this size needs about 317 trades to prove out, and it has 436.
• id511 (EMA 21/50 · BTC · 1h · 2× · long) — Sharpe 0.020 on 469 trades. Five times smaller. To prove a Sharpe this thin you would need N ≥ 9,213 trades; with 469 it is statistically indistinguishable from zero. Same family, same window — the Sharpe alone decides whether the sample is enough.
• id478 (EMA 50/200 · BTC · 1d · 2× · long) — Sharpe 0.557, the fleet's highest — on 3 trades. Like its profit factor of 20.8 (see profit factor), it is a tiny-sample illusion, not a quality signal.
• id628 (EMA 9/21 · BTC · 1m · 2× · short) — Sharpe −0.71 on 10,574 trades, with a −98% drawdown. A large, confidently negative Sharpe: this one is reliably bad.

How to read it honestly

1. A high Sharpe on few trades is noise — id478's 0.557 on 3 trades ranks first in the fleet and means nothing.
2. Sharpe sets the sample bar. The smaller the Sharpe, the more trades you need to prove it: N ≈ (1.96 ÷ Sharpe)². See sharpe significance.
3. Compare Sharpe at a fair leverage. [[Leverage]] scales returns and their spread together, so it barely moves Sharpe — but it does move the drawdowns Sharpe ignores. Read Sharpe at 2×, read survival separately (drawdown, risk of ruin).

Related

• sharpe significance
• profit factor
• confidence interval
• drawdown
• edge
• sample size

Refined 2026-06-28 — the badge mismatch (GAP-006)

The authenticated leaderboard's edge badge creates a three-way mismatch between the number displayed and the numbers that drive verdict and sort:

1. Displayed number = Sharpe point estimate (e.g. ▲ 1.82 Real edge). This is what EdgeCell.tsx renders via <span className="ev">{value}</span>.
2. Verdict word = driven by Sharpe CI. "Real edge" requires CI lower bound > 0; "No edge" requires CI upper bound ≤ 0.05. The displayed Sharpe point estimate has no direct relationship to which verdict fires — a high Sharpe with few trades can still be "Inconclusive" if the CI is wide.
3. Sort order = CI lower bound. A strategy with Sharpe 2.0 but a wide CI can rank below one with Sharpe 1.2 and a tight CI.

A user reading ◦ 1.82 Inconclusive cannot tell why 1.82 is inconclusive. The CI width — which depends on trade count, not Sharpe magnitude — is invisible. The badge shows the least useful of the three numbers.

Documented as GAP-006 in .planning/threads/product-gaps.md. Resolution options include: (a) show CI lower bound instead of point estimate; (b) show both point estimate + CI range; (c) redesign the badge around CI as the primary signal.

Related

• sharpe significance
• profit factor
• confidence interval
• drawdown
• edge
• sample size

Sources

• growth/content/dossiers/ema-cross/1-analysis.md (run 83 analysis)
• growth/content/dossiers/ema-cross/1-dataset.csv (the 210-row result set)
• wiki/qa-sessions/2026-06-28-session.md#q7 (badge verdict logic)
• wiki/qa-sessions/2026-06-28-session.md#q8 (displayed number = Sharpe point estimate)
• packages/shared/src/public-edge-verdict.ts:129 (classifyEdge)
• apps/frontend/src/components/EdgeCell.tsx:149 (.ev span)