What transplants from the trading literature, and what doesn't

Published strategies are the cheapest hypotheses available. Someone else did the work of inventing the rule and arguing for the mechanism; you get to skip straight to the only question that matters, which is whether it still holds somewhere you'd actually trade it.

Three sweeps took well-known published rules, froze them with zero parameter changes, and pointed them at instruments their authors did not test. Some survived. The pattern in which ones survived is more useful than any individual result.

The chronological versions are in the research log.

The rule for this kind of test

A frozen spec is the whole point. The moment you re-tune a published rule for a new instrument, you have stopped testing the literature and started curve-fitting — and you've given up the one advantage the transplant had, which is that the parameters were chosen by someone with no knowledge of your test data.

So: parameters exactly as published, run on new instruments, across two independent windows, against the same gates as everything else in this library — net positive in both windows, survives deleting the single best trade, max drawdown ≤ 20%.

What survived

RSI(2) dip-buying, moved to sectors and international equity

The base rule — buy when RSI(2) is deeply oversold above the 200-day average, exit on a short-term strength close — is genuinely robust on broad US index ETFs, and ships as entry #2. Run verbatim on five sector SPDRs and two international ETFs, 5 of 7 passed.

The family extends. It also thins: three of the five passers only clear the bar before trading costs are stressed, which is worth knowing before you run seven versions of the same idea and call it diversification.

Double-7s, frozen at its 2009 parameters

Connors' Double-7s — buy when the close makes a 7-day low above the 200-day average, sell on a 7-day high — was run with its original parameters untouched. QQQ and XLK pass the full bar, including the 2000–2010 stress decade on all three US instruments tested (QQQ +75.7%, PF 2.77, surviving strip-best). Fee stress at 0.1% per side leaves profit factors of 1.6–2.4.

SPY is a near-miss: every check passes except the second window's drawdown, 22.6 against the ≤20 bar — and on an account basis rather than a sum-of-percentages basis that same drawdown is 19.1. The log attributes that breach to a single uncleared covid hold with no stop, not to a structural drawdown. It is not in the library because the bar is the bar. EEM fails on drawdown outright.

Updated by sweep #11. The bar gained a sixth check in September 2026 — beat a random-entry null — and this entry does not clear it in its modern window: 76.3rd percentile against a null p95 of +65.3%. It keeps its place on the shelf with the result attached rather than being delisted. See can a strategy pass every check and still be luck?.

This ships as entry #4 with a prominent correlation disclosure, because it is the same mean-reversion family as entry #2. Two strategies from one family is one sleeve, not two positions.

SMA-200 trend timing on US equity indices

The Faber-style monthly-ish cross — hold when price is above its 200-day average, stand aside when it isn't — passes on SPY and QQQ and ships as entry #3. EFA fails, at −4.3% in the first window.

The entry publishes its own worst case: the 2000–2010 decade fails strip-best. Trend timing buys you drawdown protection and charges you whipsaws, and the stress decade is where you see the invoice.

Updated by sweep #11. This entry clears neither window of the random-entry null added in September 2026: 89th percentile in 2010–18 and 77.3rd in 2018–26, against null p95s of +71.0% and +93.9%. The fair reading, which its own page makes, is that a return-based null is the wrong instrument for a strategy whose published claim was never excess return but a drawdown a third of holding's. What the test adds is that its return is what the exposure bought, not what the signal earned. See can a strategy pass every check and still be luck?.

What didn't

The same trend rule on metals

Killed. GLD and IAU flip between windows — the 2011–2015 gold bear delivers −18% at a 0.47 profit factor in the first window, then +131–135% in the second. SLV is positive in both windows but fails strip-best in the first, at a 41% drawdown.

Same rule, same gates, same test that SPY and QQQ passed. The instrument decided the outcome. Trend timing on metals is regime-hostage: it is a bet on the existence of a multi-year metals trend, dressed as a systematic rule.

Mean reversion moved to crypto

Killed — but not for the reason the numbers first suggest. Entry #2's RSI(2) spec on crypto 1h returns profit factors of 0.51–1.19 and −40% to −140% per window across roughly 750 trades each.

Sweep #4 went back and took that result apart, and the signal turns out not to be the problem. At ~750 trades a window paying a 0.2% round trip, the fee bill alone is ~150% — more than the loss. Gross, the same cell is positive at roughly +95%. The execution model killed it, not the mechanism.

The same sweep then tested the obvious fix: rest a limit order instead of crossing the spread, at the maker rate. Passive fills do capture the gross edge — BNB +179.5%, ETH +112.1%, BTC +59.0% in the first window, all surviving strip-best. Then the second window flips BTC and ETH negative and BNB breaches the drawdown bar. So the standing verdict is regime-fragile, not cost-dead: the reversion is real and reachable at maker fills, and it belongs to the high-volatility era it was born in.

What does transplant badly is the execution assumption, not the idea. A daily dip-buy on a diversified index pays its costs once and has hundreds of businesses behind the bounce; an hourly dip-buy on one crypto asset pays 750 times for a smaller move each time. That is a market-structure difference, not evidence that the underlying can go to zero.

A crypto daily trend-pullback

Killed. Buy above the 200-day average when RSI(14) drops under 40, trail 3×ATR — a sensible-sounding transplant of equity trend-pullback logic. It flips window to window everywhere: BTC and ETH negative in the first window and positive in the second, BNB the mirror image. No neighbours were run; the hypothesis was closed.

The pattern

Read together, the six results say something more useful than any one of them:

What transplants is a mechanism moved to a similar market structure. RSI(2) dip-buying survives the move from broad index to sector and international index, because the thing that makes it work — a diversified basket over-reacting and reverting — is still present.

What doesn't transplant is a mechanism moved to a different market structure. The same rule fails on single crypto assets, where the diversification that powered it is gone. Trend timing fails on metals, where the multi-decade equity drift it implicitly relies on is gone.

Before you port a published strategy, the question worth answering is not "will the numbers hold" but "is the reason it worked still present here." The numbers are downstream of that, and they will tell you the answer eventually — usually after you've traded it.

One more thing the sweeps measured: fills

The same runs quantified how much of each result depended on the assumption of filling at the bar close.

The dip-buy family is genuinely flattered by close fills: SPY's second window drops from +36.4% to +19.4% when entries are moved to the next open. Still profitable, and now honestly priced. The trend-timing system is fill-indifferent — next-open is marginally better.

That difference is mechanical rather than incidental. A strategy that buys weakness is buying at a moment when the close is, by construction, a good price; a strategy that follows trend is not. If you are evaluating any dip-buying rule and the backtest fills at the close, assume some of the result belongs to the fill model.

Related

• Does any low-timeframe crypto strategy actually work?
• Funding-rate cash and carry: the honest numbers
• The strategy library — the six entries that survived, with full evidence.