Alpha

Status: Alpha is computed and stored by PerpForge but not displayed on any user-facing surface as of 2026-07-01. Three structural failures make it misleading on a mixed-direction, mixed-leverage leaderboard. See ADR-0010 for the full rationale and the condition under which alpha display may be reconsidered.

A strategy's return minus the buy-and-hold return over the same window — "how much it beat, or trailed, just holding the coin." Positive = the strategy added value; negative = holding won.

In plain English

Alpha is the part of a strategy's return that isn't explained by the market simply going up. If ETH rose 392% and your strategy made 420%, your alpha is +28 points — the extra you earned for your trouble. If you made 200%, your alpha is −192 points: you underperformed the lazy default and the trading actively cost you.

The opposite of alpha is beta — return you got just by being exposed to the market. In a raging bull, beta is enormous and easy; alpha is what's left after you subtract the free beta. Beating buy-and-hold is hard precisely because buy-and-hold already harvests all the beta with zero skill.

Formula / mechanic

alpha = strategyReturn − buyAndHoldReturn
strategyReturn = finalEquity / initialEquity − 1

Stored per strategy as alpha_pct on backtest_summaries (Phase 600). "Beat buy-and-hold" (beat_buy_and_hold in the dossier CSV) is exactly alpha > 0.

Three honesty caveats — read alpha with all three (they are part of the metric, not garnish)

1. In-sample, per-window — never a forward claim. Alpha is a fact about this historical window only. On run 83's bull window, only 6 of 210 rows had positive alpha — but in a bear or sideways window a disciplined trend-follower could beat holding. See out of sample testing, market regime.
2. Strategy return is at the strategy's configured leverage; buy-and-hold is unleveraged (1×). The engine deliberately does not normalize the strategy to 1×. The consequence: a positive alpha on a leveraged row can be amplification, not edge — leverage magnified a run past a 1× benchmark. To judge genuine edge, read sharpe significance and edge_significance, not alpha. See leverage.
3. Weakly meaningful for shorts. A short (profits when price falls) compared against a buy-and-hold that profits when price rises is a weak comparison. The number is still stored, but its headline use is for longs.

Why it matters for this fleet

Alpha is honesty question #4 made into a number. But it is the least risk-aware metric on the board: it compares two raw returns and ignores that buy-and-hold earned its return by surviving a near-wipeout drawdown (see risk adjusted return, drawdown). On run 83:

• 6 of 210 rows beat hold (positive alpha) — every one a 50× long. Beating the market here was a leverage trick, not an edge.
• Of the 5 edge-significant rows, only 1 beat hold — and only at 50×.
• 0 of the fair-leverage edge-significant signals beat hold — their alpha runs −1,384 to −2,052 percentage points. The honest trend-followers all lost to holding, by a lot.

So the 6 positive-alpha rows are the most suspect on the board (selection bias + leverage amplification), not the most trustworthy — verify them on edge-significance, never on alpha.

Examples from the live fleet

• id522 (ETH 4h 50× long, EMA 21/50) — alpha +8,123 pp vs the 1× hold; the only row that is both a hold-beater and edge-significant. But its CAGR of 116.9% (vs ETH's 31.9%/yr) is 50× amplified beta, paid for with a 30–61% drawdown, and its win-rate confidence interval (±7 pp) is far fuzzier than the trustworthy losers. Strip the leverage and it loses to simply holding ETH.
• The fair-leverage edge-significant signals — real, well-sampled edges (e.g. id523, the SOL 1h 21/50 long at low leverage) all carry negative alpha (−1,384 to −2,052 pp). They are genuine trend-followers that still lost to a bull-market hold. Alpha sign alone would dismiss them; their edge-significance says the edge is real, just smaller than free beta — exactly the Q1-vs-Q2 distinction.

Related

• buy and hold
• risk adjusted return
• leverage
• sharpe significance
• selection bias
• out of sample testing
• edge

Refined 2026-06-28

Alpha has units — it is not unitless. Alpha is measured in the same units as return: percentage points (pp). "Strategy returned 12%, benchmark returned 8% → alpha = +4 pp." This matters for cross-timeframe comparison: 4 pp alpha on a 6-month window is not the same achievement as 4 pp alpha on a 2-year window, because the benchmark return over those windows differs.

Contrast with Sharpe ratio, which IS unitless (return ÷ volatility — units cancel). Sharpe is comparable across strategies and timeframes without adjustment; alpha is not.

Cross-symbol interpretation (with a fixed window). When all strategies run the same backtest dates, alpha % is calendar-comparable across timeframes (a 4h strategy and a 1h strategy both played the same market). But across symbols, alpha is "local" — each strategy's benchmark is buy-and-hold of its own symbol. A strategy with +10% alpha on ETH beat holding ETH; one with +10% alpha on BTC beat holding BTC. These are different achievements if the two symbols had different buy-and-hold returns that window. Alpha ranks cleanly within a symbol; cross-symbol alpha comparison requires normalization.

Alpha needs a CI. Like all backtest metrics, alpha is a sample estimate with uncertainty. A strategy with 3 trades showing +50% alpha and one with 300 trades showing +50% alpha look identical on raw alpha — but the first has a confidence interval (CI) of roughly ±200% and the second ±5%. The number is meaningless without the CI. See GAP-003 in .planning/threads/product-gaps.md.

Sources

• wiki/qa-sessions/2026-06-22-session.md#q1 (first formal entry)
• wiki/qa-sessions/2026-06-28-session.md#q3 (units clarification)
• wiki/qa-sessions/2026-06-28-session.md#q4 (cross-symbol / cross-timeframe interpretation)
• growth/content/dossiers/ema-cross/1-analysis.md §4 + §6 (the buy-and-hold beaters, run 83)
• docs/analytics.md §"Stored per-strategy buy-and-hold + alpha"

Refined 2026-06-28 — what alpha actually measures in perp context

Alpha answers "did you beat a passive long exposure?" — not "did you beat doing nothing?"

In spot trading, "doing nothing" = hold the asset = buy-and-hold. The two questions are the same.

In perpetual futures (perps), "doing nothing" = hold cash (0% return). Opening any position — long or short — is active. So in perp context:

• A profitable perp strategy always beats "doing nothing" (0%).
• The meaningful question is whether it beat passive long (B&H).
• Alpha measures the second question, not the first.

Why show alpha at all? Beta removal. Without it, a long strategy in a bull market looks great even if the market did all the work. Alpha strips out the "free ride" — what's left is what the strategy earned on its own.

The leaderboard confusion risk: A user seeing alpha may read it as "excess over doing nothing" when it's "excess over passive long." For short and bidirectional strategies, this actively misleads. See buy and hold and GAP-007 in .planning/threads/product-gaps.md.