Strategy Tearsheet · Crypto Perpetuals

The TRIAD Book

A diversified systematic book on the largest crypto perpetual futures — five price signals, a crowd-positioning fade, and a perp-spot basis trade, blended and adversarially audited. Here is the full held-out tearsheet.

SuperThesis Research · frozen and forward-tracked · backtest through mid-2026

+2.9
Sharpe ratio in the held-out test
return per unit of risk — above 2 is excellent

Held out means the test window (Apr 2023–Jul 2026) was never used to build the signals. Robust across sampling: hourly +2.9 · daily +3.0 · weekly +2.8 · monthly +2.3 — quote the conservative +2.3 for monthly-reported returns.

Yearly return (test)+18%
Worst new loss (test)-2.8%
Volatility (test)6%/yr
Test split, thirds+3.9 / +2.6 / +2.2
Buy & hold Sharpe+0.3
Training Sharpe+1.4
Blocks / sleeves3 / 12
Universe14 majors

Growth of $1 — steadiness is the point

Buy & hold made more in raw total, riding the 2020–21 supercycle in the training era. But look at the held-out test (shaded): the strategy compounds quietly while buy & hold round-trips through brutal crashes. Same coins, a fraction of the risk. The curve below runs the strategy at its volatility-targeted leverage setting; the unlevered book returns 2.2×.

10×30×2020202120222023202420252026TRAININGHELD-OUT TESTstrategy 12×buy & hold 24×
Growth of $1, log scale. Dashed vertical line = the boundary between training and the held-out test; shaded region = held-out test.

A fair fight — matched, recent windows

The supercycle flatters buy & hold. Trim the start date to more recent, same-length windows and the picture inverts — the strategy compounds while buy & hold loses money through ~70% drawdowns:

Growth of $1 · worst drawdown, each window ending Jul 2026.
WindowThis strategyBuy & hold
Held-out test (3.2y)1.71× · -3%1.03× · -74%
Last 2 years1.33× · -3%0.57× · -74%
Last 12 months1.12× · -3%0.48× · -70%

Consistency and cost

The edge is not one lucky year: it is positive in six of seven calendar years (2022, a broad crypto bear, is the exception). And it survives realistic trading costs — even at market-order fees it clears +2.2, because the three blocks diversify fee damage the same way they diversify risk.

Sharpe by calendar year (2020–21 are the training era; 2026 is a half-year).
+1.82020+3.02021-1.02022+1.42023+3.82024+2.62025+2.32026
Held-out Sharpe vs trading cost per side (dashed = the 2.0 target).
target 2+2.90.02%/side+2.70.05%/side+2.20.10%/side

Capital policy — leverage is a dial, not an edge

Sharpe is invariant to leverage (~+2.9 at any setting) — leverage scales risk and reward together. The choice is a drawdown-appetite decision. The recommended policy, VT-20, smooths to a 20% volatility target weekly and caps at 4×; funding on leverage and liquidation gaps are not modeled, so read the levered rows as upper bounds.

Held-out test, compounded. Return/yr · volatility · max drawdown · growth of $1.
PolicyReturn/yrVolMax DD$1 →
1× fixed+18%6%-3%1.70×
2× fixed+38%11%-6%2.87×
3× fixed+62%17%-8%4.79×
VT-20 (vol-target)+71%19%-11%5.69×

Why it works: the diversification arithmetic

No single coin drives this. The best coin runs a Sharpe of only +1.8; the average coin barely +1.0. The portfolio Sharpe is far higher because the 14 coins' profit streams are nearly uncorrelated — averaging them keeps the return but cancels most of the risk:

avg coin +1.0 × diversification ×2.9portfolio +2.9

×2.9 = √( 14 / (1 + 13×0.05) ), with a measured average pairwise correlation of 0.05

Correlations are not constant, though — in a common shock they spike toward 1 and this multiplier collapses toward 1. Diversification is weakest exactly when it's needed most, which is why realized correlation is the first thing to watch forward. Every coin's own contribution, sorted:

Per-coin held-out stats — the strategy's own positions on each coin, net of fees.
CoinSharpe$1 →Max DDVolShare of profit
ADA+1.82.77×-12%19%15%
BTC+1.81.93×-7%12%9%
DOGE+1.52.02×-12%15%10%
SOL+1.41.89×-13%15%9%
BCH+1.42.25×-11%19%12%
AVAX+1.41.73×-11%13%8%
ETH+1.21.54×-11%12%6%
XRP+1.11.74×-19%18%8%
MATIC+0.91.26×-15%8%3%
1000SHIB+0.91.59×-12%18%7%
LINK+0.81.43×-32%14%5%
BNB+0.61.27×-21%13%4%
OP+0.61.46×-30%26%7%
LTC-0.70.73×-40%13%-4%

Profit is broad — ten of fourteen coins carry meaningful weight, and only one (LTC) is a net drag. There is no hidden single-name dependency behind the headline number.

What's inside

Twelve sleeves across three mechanically independent blocks — price behavior, crowd positioning, and the perp-spot basis — combined 60/20/20. Nothing here is exotic; the work is in the blending and the auditing.

SignalWhat it does
Weekday clocklong the weekday that's been strongest recently, short the weakest
Trendlong coins trending up over 1–3 months, short those trending down
Yearly-highlong coins nearest their yearly high, short those furthest below
Patient momentummomentum that ignores the latest week, so it doesn't chase spikes
Volume surgebuy unusual volume spikes that close strong; hold about 10 hours
Crowd fadefade a coin when its long/short account ratio hits a 90-day extreme
Basis convergencea two-leg perp-vs-spot dislocation trade, funding credited, patient

Know the limits

How it's built

For completeness, the frozen specification — enough for an engineer to reproduce the book from public data:

UNIVERSE 14 USDT-perps = top-40 by pre-2023 dollar-volume with ≥6y history (BTC ETH SOL XRP BNB ADA DOGE LINK SHIB AVAX LTC OP BCH MATIC). Size is near-optimal: the Sharpe-vs-size curve plateaus at ~10–20 coins. DATA 1h OHLCV; 8h funding; 5-min open-interest & long/short account ratio; spot close. Knowability rule: every input is the last value stamped AT or BEFORE the bar close; a feed staler than 2 days drops to flat. COMPOSITE pos = clip( 0.60·PENTA + 0.20·CROWD-FADE + 0.20·BASIS , −1, +1 ) fill one bar late; fee 2 bps/side per unit of turnover. PENTA equal-weight mean of 10 cross-sectional sleeves; each ranks a metric across the 14 coins, longs the top ~20% / shorts the bottom ~20%, and rebalances on a daily grid: trend, distance-from-yearly-high, skip-a-week momentum, volume-shock longs, day-of-week seasonality. CROWD fade a coin whose long/short account ratio sits at a trailing-90-day FADE extreme (≥95th pct → short; ≤5th → long); hold ~2 days; flat otherwise. BASIS basis = perp/spot − 1. Beyond ±0.30% enter the convergence trade (short the rich leg, long the cheap), exit inside ±0.05%. Two-leg P&L net of fees plus collected funding; execute maker-only, one bar late. CAPITAL fixed 1–3×, or VT-20: weekly set leverage = clip(0.20 / 30-day realized vol, 0.25, 4). Governor: cap leverage so worst drawdown ≤ appetite. GATES believe a change only if it clears a walk-forward hold-out (all folds > 0, in-sample > 0, ≥100 trades) and survives an adversarial look-ahead audit. Forward health check: rising per-coin correlation = the edge decaying.

This tearsheet is research and commentary for educational purposes only. It is not investment advice, an offer, or a solicitation, and it describes hypothetical backtested results that do not represent actual trading and were prepared with the benefit of hindsight. Simulated and past performance do not indicate future results. Trading leveraged crypto derivatives carries a substantial risk of loss.