A CTA Trend-Following System on Binance Perpetuals

This repository turns a public seven-component trend-following construction into a complete Binance USDT-perpetual backtest. The source is the November 2022 thread. The exercise checks both the published construction and its empirical claims.

Sections 1-8 freeze the rule set before any result is computed. Sections 9-12 report measurements. SOURCE.md states what the thread does and does not publish; every number below that the thread does not supply is marked [choice].

---

1. Scope

The source names seven components and gives a conventional design for each. All seven appear here, but the test venue differs from the cross-asset setting in the source.

#	Component	Thread's prescription
1	Universe selection	as many diversifying markets as possible
2	Trend detection	fast MA minus slow MA of log price, normalized by volatility
3	Signal to position	saturating response: sigmoid, or x*exp(-x^2)
4	Sector weights	equalize long-run risk across asset-class groups
5	Portfolio risk targeting	scale the book to a volatility target
6	Trading rules	buffering, to trade only when far from target
7	Execution	model commissions and slippage honestly

The thread's setting is 60-400 cross-asset futures. This project has 788 crypto perpetuals, which is a large nominal universe inside a single asset class. The diversification measurements separate nominal market count from independent breadth throughout.

2. Data

Universe	788 USDT-quoted perpetuals, every contract with published archives, delisted included, stablecoin bases excluded
Bars	hourly klines 2020-01-01 to 2026-07-22, resampled to daily UTC
Funding	monthly fundingRate archives, complete published history
Source	data.binance.vision, mirrored locally; see TREND_MIRROR_DIR

Daily bars are built from hourly bars: open of the first hour, close of the last, max high, min low, summed volume and quote volume. A day with fewer than 20 of 24 hourly bars is marked incomplete and is not tradeable [choice]; its return is still carried so a position held through it is marked correctly.

Delisted contracts remain in the panel until their archives stop. When the next open is missing, the position is marked to the latest observable close, then frozen at that mark until observations resume. Archive exhaustion alone does not imply that the prior close was an executable liquidation: the position remains frozen unless an open resumes. Only positions executable at the predeclared backtest endpoint are liquidated at that day's close.

3. Frozen rule set

All parameters below are **[choice]**s unless quoted from the thread. Rationale is given where a defensible alternative exists.

3.1 Timing convention

Daily rebalance at 00:00 UTC. Signals use bars strictly through the previous day's close. Fills are at the current day's open. A position set at open t earns open(t+1)/open(t) - 1. No information used at time t is unavailable at 00:00 UTC on day t.

3.2 Point-in-time universe filter

A symbol is tradeable on day t when, using data through t-1:

• at least 120 complete daily bars of history;
• trailing 30-day median daily quote volume at least $5,000,000;
• a complete bar on t-1 and a bar present on t.

The volume floor is the one component with no analogue in the thread (futures CTAs pick markets by hand). It exists because a $50k/day perpetual cannot absorb a vol-targeted position at any account size.

This rolling selection, with delisted contracts kept to their last archive, also follows the December 2022 "backtest errors" series. Its first entry covers universe look-ahead (S3 in docs/source-supplements.md).

3.3 Volatility forecast (component 3 sizing input)

Thread: "a blend of a steady-state long term volatility (estimated over 10 years) and a short-term realized volatility (eg 30-60 days)".

$$ \hat\sigma_{i,t}=\max!\left(;0.70,\sigma^{\mathrm{ST}}_{i,t}+0.30,\sigma^{\mathrm{LT}}_{i,t},;;\sigma_{\min}\right) $$

• $\sigma^{\mathrm{ST}}$: standard deviation of the last 60 daily log returns (thread's "30-60 days"; 60 chosen as the more stable end).
• $\sigma^{\mathrm{LT}}$: expanding-window standard deviation over all history available at t, capped at the last 10 years. The archive is 6.5 years deep, so this is an expanding estimate in practice, a labeled deviation forced by the venue's age.
• Weights 0.70/0.30: the standard practitioner blend; the thread gives none.
• $\sigma_{\min}$ = 20% annualized, a floor that stops pegged or stalled contracts from demanding unbounded leverage.
• Annualization uses $\sqrt{365}$; perpetuals trade every day.

3.4 Trend signal (component 2)

Thread: "a fast moving average of (log) price minus a slow moving average, probably normalized by volatility", on horizons "from 1 month to 1 year, with an average around 3-6 months".

Three exponentially-weighted crossover speeds, spans in days:

$$ (S_k,L_k)\in{(16,48),;(32,96),;(64,192)} $$

The slow spans are 1.6, 3.2 and 6.3 months, centering the ensemble on the thread's stated 3-6 month average. A commonly published three-speed set is one octave faster; the shift here is deliberate and labeled.

For each speed, with $p=\log(\text{close})$ and $\sigma^d$ the daily volatility forecast of 3.3:

$$ y_{k,i,t}=\frac{\mathrm{EWMA}_{S_k}(p)-\mathrm{EWMA}_{L_k}(p)}{\sigma^d_{i,t}\sqrt{L_k}}, \qquad z_{k,i,t}=\frac{y_{k,i,t}}{\mathrm{sd}_{365}(y_{k,i,\cdot})} $$

The first division is the thread's "normalized by volatility" and makes the signal dimensionless. The second is a trailing 365-day (minimum 60 observations) rescaling to unit variance so the three speeds enter the average with equal risk weight rather than equal nominal weight. Both use only past data. The combined signal is

$$ z_{i,t}=\frac{1}{|K_{i,t}|}\sum_{k\in K_{i,t}}z_{k,i,t}, $$

clipped to $[-6,6]$ for numerical safety only; the response function saturates far below that. A speed whose slow span is not yet filled drops out of the average, so a young symbol trades on its faster speeds alone until the slow ones become available.

3.5 Response function (component 3)

Thread: "a sigmoid function (left) or a function like x * exp(-x^2) which reduces the target position for very large values of the trend signal".

Baseline [choice]: $f(z)=\tanh(z)$, the sigmoid branch, capped at $\pm1$.

First-class alternatives, reported side by side rather than selected on:

Name	$f(z)$
tanh (baseline)	$\tanh z$
overextension	$z e^{-z^2/4},/,(\sqrt2 e^{-1/2})$, peak 1 at $z=\sqrt2$
linear	$\mathrm{clip}(z,-1,1)$
binary	$\mathrm{sign}(z)$

The thread also mentions adding a long bias for markets expected to rise on average but does not favor it. It is not implemented: there is no crypto analogue of an equity risk premium that is knowable ex ante.

3.6 Unscaled target exposure (components 3, 4)

$$ u_{i,t}=f(z_{i,t})\cdot\frac{g_{c(i),t}}{\hat\sigma_{i,t}} $$

so that, before sector weights, every market at full signal contributes the same risk, matching the thread's "inversely proportional to a market-specific volatility forecast ... each market contributes equally to risk".

Sector weights $g$ implement "bucket up the universe into broad asset class groups, and try and have roughly equal volatility from each asset class in the long run". Crypto has no exogenous sector taxonomy that is stable and point-in-time, so groups are discovered from returns:

• every 7 days [choice], take the trailing 180-day return matrix of tradeable symbols;
• correlations centered and standardized separately on each pair's shared observations, shrunk 30% toward the constant-correlation matrix, then eigenvalue-projected and renormalized to a positive-semidefinite correlation matrix;
• average-linkage hierarchical clustering on $d=\sqrt{2(1-\rho)}$, cut into $K=8$ clusters [choice];
• with $q_i=1/\hat\sigma_i$ at full signal, compute each active cluster's standalone pre-sector risk $R_c=\sqrt{q_c^\top\Sigma q_c}$; set $g_c\propto1/R_c$, normalized to mean 1 across active symbols. Thus every cluster has equal standalone full-signal risk regardless of its market count or internal correlation.

Ablation no_sector fixes $g\equiv1$ and isolates what component 4 adds.

Only symbols carried by the current correlation estimate may hold exposure, so the ex-ante volatility of section 3.7 covers every position in the book rather than a subset of it. A newly eligible symbol waits at most seven days before it can be traded.

3.7 Portfolio risk targeting (component 5)

$$ m_t=\frac{\sigma^{\mathrm{target}}}{\sqrt{u_t^{\top}\Sigma_t u_t}}, \qquad N^{*}_{i,t}=m_t,u_{i,t},E_t $$

with $\sigma^{\mathrm{target}}=20%$ annualized (the number the thread uses in its own example), $\Sigma_t$ the same shrunk weekly covariance, and $E_t$ current equity. Gross notional is then capped at $2\times E_t$ [choice], scaling all positions pro rata if breached. That cap is a perpetual-futures margin reality with no counterpart in the thread.

$m_t$ is unbounded because $u$ is a notional-per-dollar weight whose scale depends on the number of live markets. The gross leverage cap is the hard portfolio bound.

An account whose equity reaches zero is recorded as ruined: it stops trading and every later day is reported flat. No run is floored at a positive equity to keep a return series alive.

Ablation no_target replaces $m_t$ with a constant calibrated so its average post-buffer ex-ante risk matches the targeted run during the design period ending 2023-12-31. The constant is then frozen for the held-out period. Its daily post-buffer risk divided by $\sigma^{\mathrm{target}}$ is the direct test of claim C10 ("0.3x to 3x").

3.8 Trading rule (component 6)

Buffering. With $\tau_{i,t}=b\cdot m_t,\frac{E_t}{\hat\sigma_{i,t}}$ the notional value of $b$ units of the market's own risk allocation:

• if $|N_{i,t}-N^{*}_{i,t}|\le\tau_{i,t}$, do not trade;
• otherwise trade to the near edge, $N^{*}_{i,t}\pm\tau_{i,t}$.

Baseline $b=0.10$ [choice]. Swept over ${0,0.05,0.10,0.20,0.40}$ to produce the turnover-versus-net-return curve that claim C12 asserts exists. The thread's refinements (widening the buffer when spreads are wide, and partial fills toward target for large books) are not implemented; the archive has no spread data and the capacity question is out of scope.

3.9 Costs (component 7)

• Trading: 10 bps of traded notional [choice], being 5 bps taker fee plus 5 bps slippage. Swept over ${2,5,10,20}$ bps.
• Funding: measured, not assumed. Each funding timestamp charges $r_{\text{funding}}\times N_{i}$ against a long and credits a short, using the contract's own published rate and interval. Payments are bucketed by the calendar day they stamp, so the 00:00 UTC payment lands on the position set at that day's open rather than the one held into it, a one-payment boundary approximation. Symbols or months with no published archive are charged zero funding, which is optimistic and is reported as a coverage fraction.
• No borrow, margin-interest or liquidation modeling. Positions are notional; the 2x gross cap is the only leverage constraint.

3.10 Accounting

One compounding equity account starting at $100{,}000$. Daily P&L is $\sum_i N_{i,t},r_{i,t}$ minus trading cost and funding, where $r_{i,t}=\mathrm{open}{i,t+1}/\mathrm{open}{i,t}-1$. Sharpe ratios are annualized from daily returns with $\sqrt{365}$ and are computed on arithmetic daily returns of the equity curve.

When the next open is missing, $r_{i,t}$ first marks the position from the current open to the current close without inventing a liquidation. That close is carried during the missing-open interval, and the remaining cumulative return is realized when an open resumes. Only positions with an executable price on the predeclared final backtest day are closed and charged closing turnover. Drawdown includes the initial account value as its first high-water mark.

4. Claims under test

Restated from SOURCE.md with the statistic each maps to.

#	Claim	Statistic
C1	single-market trend Sharpe $\approx$ 0.2	distribution of per-symbol standalone Sharpe
C2	Sharpe scales like sqrt(independent breadth)	effective breadth $1/(w^\top Rw)$ and point-in-time N-market diagnostics
C3	universe expansion worth 1.5-2x Sharpe	ratio of large-N to small-N Sharpe
C4	correlated additions add little	correlation and breadth within/across clusters
C5	practitioners use 1-12 month horizons, averaging 3-6	performance diagnostic per crossover speed
C6	saturating responses are common	implementation check and response sensitivity
C7	inverse-vol sizing is common	implementation check and equal-notional sensitivity
C8	simple volatility forecasts capture most benefit	comparison among simple forecasts; no oracle claim
C9	sectors can be equalized for diversification	cluster-risk invariant and no_sector sensitivity
C10	untargeted risk spans 0.3x-3x	quantiles of ex-ante vol / target
C11	risk targeting improves Sharpe	full system vs no_target
C12	buffering keeps positions near target and cuts turnover	turnover, target tracking error and net Sharpe vs $b$

5. Evaluation protocol

• Full sample 2020-01-01 to 2026-07-22, split into a design period through 2023-12-31 and a disjoint held-out period beginning 2024-01-01. The no-target control is calibrated on the design period only; every other parameter is fixed before either period is measured.
• Every variant runs on identical dates, universe, costs and timing. Only the named component changes.
• Reported per run: gross and net annualized return, annualized volatility, Sharpe, max drawdown, turnover, average gross exposure, cost decomposition, and long/short attribution.
• Sharpe differences between variants are reported with a stationary bootstrap on the daily return difference (mean block 21 days, 2000 resamples), because the variants share the same market paths and their differences are strongly autocorrelated.

6. Pre-analysis invariants

• Signal, volatility, correlation, cluster and universe membership use data strictly before the fill.
• Missing bars never become favorable fills; a symbol without a bar is not tradeable that day.
• Delisted contracts stay in the panel to their last archive.
• The correlation and cluster estimates are rebuilt on a fixed weekly cadence, never conditioned on outcomes.
• Sweeps in section 10 are diagnostics of the claims. The baseline reported in section 9 is the one frozen here, whether or not a swept cell beats it.
• The freeze-before-measuring discipline also appears in the March 2021 supplement, which builds a convincing backtest on pure random noise to show what in-sample parameter choice does (S1 in docs/source-supplements.md).
• Funding coverage is reported as a fraction. Uncovered exposure is charged zero, which flatters the result.

7. Known limitations before running

• Open-to-open execution with a flat cost model; no book, queue or impact.
• Daily rebalancing cannot see an intraday short squeeze: a position is marked open-to-open, so a pump-and-retrace inside the day is invisible and the liquidation it would have forced is not modeled. The supplementary evidence describes a switch to 5-minute rebalancing after a 20x intraday pump (S4 in docs/source-supplements.md). Section 13 measures the associated position-sizing defense.
• No capacity analysis. A 2x-gross vol-targeted book across 500 alt perpetuals is not executable at institutional size, and no claim is made that it is.
• Single venue, single quote currency, single asset class.
• No margin, liquidation or exchange-risk modeling.
• The volatility floor, gross cap and volume floor are structural choices that a different implementer would set differently.

8. Running it

python -m pip install -r requirements.txt
make test        # 38 synthetic unit tests, no archive needed
make funding     # one-off: fetch the full fundingRate history into the mirror
make panel       # daily panel + funding panel from the mirror
make backtest    # baseline, ablations, sweeps -> results/runs.csv
make claims      # C1-C12 measurement -> results/claims.json
make report      # figures and results/report.md

make all runs panel, backtest, claims and report in order. The run is deterministic: repeating it reproduces the baseline Sharpe to four decimals.

TREND_MIRROR_DIR points at the archive mirror and defaults to ./data/mirror, so a fresh clone runs with no path editing. make funding is the only step that writes into the mirror; nothing else writes outside the project directory.

---

9. Baseline result

The frozen rule set runs from 2020-01-01 to 2026-07-22. The first trade is on 2020-05-08, once the earliest contracts clear the 120-day history gate and the risk model has a correlation estimate.

	Full sample	Design (through 2023)	Held out (2024-)
Net annualized Sharpe	1.04	0.68	1.54
Gross Sharpe	1.15
Total return	+242.1%
CAGR	+20.6%
Realized volatility	19.91%		(20% target)
Max drawdown	-30.1%

Average tradeable markets	143 (15 in 2020, 238 in 2025)
Average held ex-ante volatility	19.49%
Average gross exposure	0.42x equity
Peak gross exposure	1.22x equity (below the 2x cap)
Annual turnover	17.1x capital
Trading cost drag	1.71%/yr at 10 bps
Funding cost drag	0.01%/yr, measured on 98.7% of tradeable symbol-days
Long / short P&L split	+$189k / +$75k

The risk target works in aggregate, but performance is regime-dependent: the held-out Sharpe is materially higher than the design-period Sharpe and the full-sample drawdown reaches 30%. The run shows that the construction operates as specified on this venue. It does not establish a stationary Sharpe-1 process.

Removing trading costs and funding lifts Sharpe from 1.04 to 1.15. Trading cost is the larger measured drag in this run.

10. Claim-by-claim measurements

Full tables in results/report.md.

#	Source statement	Current measurement	Interpretation
C1	single-market Sharpe around 0.2	median +0.17 gross, +0.10 net over 268 established markets	consistent on a gross basis
C2	Sharpe scales with $\sqrt{\text{independent breadth}}$	equal-weight strategy-return breadth 3.81 across 268 streams	conditional identity; nominal count is not breadth
C3	broader cross-asset universe can add 1.5-2x Sharpe	gross Sharpe changes 1.01x from 64 to 290 eligible crypto markets	venue does not supply comparable independent breadth
C4	uncorrelated additions help more	8 cross-cluster markets provide breadth 6.22 vs 2.75 in-sample and 3.23 vs 2.34 out-of-sample	supported, with weaker cluster separation out of sample
C5	typical horizons are 1-12 months, averaging 3-6	crypto diagnostic peaks at a 48-day slow span	practice statement; performance optimum is venue-specific
C6	saturating responses are common	tanh 1.04; overextension 0.87; linear 1.03; sign 1.15	implementation sensitivity, not a test of industry practice
C7	inverse-volatility sizing is common	1.04 vs 0.92 equal-notional; CI for difference excludes zero	inverse-vol helps this implementation
C8	simple volatility forecasts capture most benefit	blend 1.04; short 1.10; long 0.88; EWMA 1.07	simple variants measured; no complex-oracle comparison
C9	sectors can be risk-equalized	exact cluster-risk invariant; +0.09 Sharpe vs no-sector, wide CI	portfolio construction works; return effect is undetected
C10	untargeted risk can span 0.3x-3x target	0.05x-2.92x, with 5.7% of days outside that band	similar upper scale, materially lower downside
C11	risk targeting improves Sharpe	1.04 vs 0.82 at design-risk-matched scale; CI [-0.05,+0.51]	positive estimate, not statistically resolved
C12	buffering reduces turnover while staying near target	20.2x to 15.0x turnover; tracking error 0.4%-10.1% of equity	supported trade-off

Independent breadth, not listing count

Across 268 perpetuals with at least two calendar years after first eligibility, the standalone strategy has median Sharpe +0.17 gross and +0.10 net. The net cross-section is broad: its 10th and 90th percentiles are -0.45 and +0.54, and 60% of markets are positive.

The breadth curve reports gross returns from point-in-time portfolios. Each random-priority draw assigns symbols a fixed random priority and holds the first N currently available markets. No future listing or survival information selects a constituent. No unmodeled constituent-turnover cost is presented as net performance. Gross Sharpe reaches about 0.58 by eight slots and is essentially flat beyond 32; moving from 64 to the maximum 290 eligible markets changes it by less than 1%. Square-root scaling applies to independent breadth, not nominal slots.

For equal weights $w$, the variance-equivalent independent breadth is $B=1/(w^\top Rw)$. The 268 established strategy-return streams have mean pairwise correlation 0.26 and $B=3.81$. The weekly price risk models average 153 markets, mean correlation 0.57 and breadth 1.82. Those are the quantities that govern diversification; the 788-contract listing count does not.

Clustering on the full record illustrates C4 mechanically: eight markets drawn across clusters have average correlation 0.04 and breadth 6.22, against correlation 0.28 and breadth 2.75 inside one cluster. A stricter test clusters only on pre-2023 returns and scores post-2023 correlations. There the eight-market breadth advantage is 0.88 and the 16-market advantage is 0.20. Correlation diversification persists, although the historical cluster map separates future returns much less strongly than it separates the sample on which it is built.

The liquidity sweep shows the capacity trade-off: raising the volume floor from $5M to $100M/day reduces Sharpe from 1.04 to 0.58 by concentrating the book in the most liquid, correlated contracts.

Horizons run faster than the thread's

This crypto performance diagnostic peaks at a 48-day slow span (1.6 months) with portfolio Sharpe 0.32. The 96- and 192-day cells score 0.16 and -0.11; the 384- and 768-day cells score -0.40 and -0.47. The source's 3-6 month statement describes practitioners' average horizons, not a universal optimum. The sweep is reported as sensitivity analysis and does not select a replacement baseline.

Component sensitivity and risk control

Paired stationary-bootstrap intervals on Sharpe(variant) - Sharpe(baseline):

Component removed	Sharpe	$\Delta$ vs baseline	95% CI
Risk targeting (C11)	0.82	-0.22	[-0.51, +0.05]
Inverse-vol sizing (C7)	0.92	-0.12	[-0.21, -0.04]
Buffering (C12)	1.02	-0.03	[-0.11, +0.06]
Sector weights (C9)	0.95	-0.09	[-0.82, +0.64]
Costs and funding	1.15	+0.11	[+0.02, +0.21]

The no-target multiplier is calibrated only on 2020-2023 to match the baseline's average post-buffer ex-ante risk, then frozen. Over the full sample it realizes 27.3% volatility against 19.9% for the targeted book and draws down 43.1% against 30.1%. Its Sharpe is lower by 0.22, but the paired interval crosses zero, so C11 is directionally supported rather than statistically resolved.

C10 uses the fixed book's actual post-buffer ex-ante risk divided by the 20% target. The ratio spans 0.05x-2.92x; its 5th, median and 95th percentiles are 0.26x, 1.18x and 2.24x. Only 5.7% of days lie outside the source's 0.3x-3x illustrative range, all on the low side.

Inverse-volatility sizing is the only component ablation whose paired interval excludes zero. Response sensitivity is larger than before sector risk targeting: tanh and clipped-linear are near 1.04, overextension is 0.87, and binary is 1.15. These are performance diagnostics; the source's C6 and C7 statements concern common industry construction. The simple volatility variants range from 0.88 to 1.10 and do not include a complex or oracle forecast, so C8 remains only partially measurable here.

C12 is visible directly. Increasing the buffer from 0 to 0.4 lowers annual turnover from 20.2x to 15.0x capital while average absolute target tracking error rises from 0.4% to 10.1% of equity. Net Sharpe ranges from 1.02 to 1.10; the differences are not statistically distinguishable. The frozen $b=0.10$ produces 17.1x turnover and 4.8% average tracking error.

11. Reproduction verdict

The seven-component construction is implementable on Binance perpetuals and produces a full-sample net Sharpe of 1.04 at 19.9% realized volatility. Its design and held-out Sharpes differ substantially, and its worst drawdown is 30%, so the full-sample number should not be read as a stable expectation.

The clearest source-consistent result is structural: single-market trend is weak, and diversification depends on independent return streams rather than contract count. This venue supplies only 3.8 variance-equivalent independent strategy streams among 268 established contracts, so nominal-universe expansion saturates quickly. That observation is consistent with C2 and C4; it does not falsify a square-root rule whose premise is independence.

Within this implementation, inverse-volatility sizing has a detectable Sharpe contribution. Risk targeting has the expected sign and controls the risk distribution, but its Sharpe contribution is not statistically resolved. Sector risk equalization is exact by construction, while its return effect is unstable across periods. Buffering provides the stated turnover/tracking trade-off. Descriptive statements about typical industry horizons and response functions are preserved as construction guidance rather than treated as backtest hypotheses.

12. Limitations

Everything in section 7 still applies. The measurements add these limits:

• The venue is 6.5 years old, so the long-run volatility estimate is an expanding window rather than the thread's ten years. After warm-up, the design and held-out measurements cover about 3.7 and 2.6 years.
• Effective breadth is the equal-weight variance equivalent $1/(w^\top Rw)$. It is exact for the homogeneous-return square-root argument but does not account for heterogeneous expected Sharpes.
• Funding is charged from published archives on 98.7% of tradeable symbol-days; the remainder is charged zero, which flatters the result slightly.
• The C1 cross-section requires two calendar years after first eligibility, so it describes established contracts. The breadth curve instead uses a point-in-time random priority and never selects on future availability.
• No capacity analysis. A 0.42x-gross book spread over 145 alt perpetuals is not an institutional vehicle at any size the thread has in mind.

13. Post-freeze extension: the asymmetric short cap

In July 2026 a supplementary post, written in response to an account of memecoin shorts caught in an intraday pump, stated a rule the original thread does not contain: size shorts more tightly than longs, because a long can lose at most its notional while a short can lose an unbounded multiple of it. The sizing is explicit: holding a fraction p of gross exposure as margin and requiring survival of an m-times pump, no short should exceed roughly $p/(2m)$ of gross; the worked example ($p=0.2$, $m=10$) gives 2%. The posts are summarized as S4 and S5 in docs/source-supplements.md.

It is implemented as a labeled extension. max_short_frac_gross caps each short at the stated fraction of the day's leverage-capped target gross, applied to both the target and post-buffer books. The cap is one pass, so a trimmed short can exceed the same fraction of the final smaller gross. The baseline remains uncapped.

Cap per short	Net Sharpe	$\Delta$ vs baseline	95% CI	Max DD	Design	Held out
none (baseline)	1.04			-30.1%	0.68	1.54
10%	1.09	+0.05	[-0.03, +0.12]	-28.7%	0.74	1.57
5%	1.09	+0.05	[-0.13, +0.22]	-27.7%	0.82	1.46
2% (source example)	1.04	-0.00	[-0.31, +0.31]	-32.0%	0.83	1.34
1%	1.00	-0.04	[-0.40, +0.33]	-34.0%	0.84	1.25

No tested cap has a statistically resolved Sharpe effect. The 2% rule changes maximum drawdown only slightly and has no resolved Sharpe effect; tighter caps increasingly tilt the book away from short exposure. Funding drag rises from 0.01% to 0.84% per year at 2%, because shorts collect funding on average in this sample.

This daily backtest still cannot observe the intraday pump-and-retrace for which the rule exists, and it has no liquidation model. The table measures the ordinary-path cost of the insurance, not its tail-event payoff.

14. Disclaimer

Research artifact, not investment advice.