The method
Every test, with its receipts.
The depth behind the listed rules: how they were searched, the one look each got at prices it had never seen, the checks we ran against ourselves, and what the research desk's gates do when a rule stops earning its place.
01In sample and out of sample
Tested through a slump, then through a rally.
We searched 39 trend rules on prices before 1 Mar 2026, a stretch where holding all 32 markets lost 11.1%. 38 of them cleared the bar; the frozen rule made +56.8%. Then we ran the 38 survivors once, on prices from 1 Dec 2025, where holding made +44.8%.
Every one of the 38 fell less than holding: worst falls between −25.1% and −12.6%, against −41.8% for holding. The frozen rule kept most of the rally, +41.5%, at a third of the fall.
← deeper worst fallreturn ↑
← deeper worst fallreturn ↑
Bright dots are listed rules; the shaded corner is "beat holding on both". The walk-forward bar here is the
catalogue's: on the unseen months 1 of 38 cleared it again. A rule longer than 205 candles starts, and seeds its holding line, at its own
length, so a few holding points differ slightly; the marker shows the frozen rule's. The two slices overlap from
1 Dec 2025 to 1 Mar 2026 for warm-up. Source: docs/evidence/walk-forward-2026-10-23.json, 22 Oct 2026.
02Market by market
The frozen rule on each market, unseen months.
Spread across many markets, the rule does not need to be right about any one of them. In the unseen months it finished ahead of holding in 24 of 32 markets. Most of that came from losing less: of the 24 markets that fell, it was ahead in 20.
A trend rule buys once a move has proved itself and sells once it turns. So in 3 of the five biggest risers (VVV, HYPE, NEAR), holding outright kept the first and last stretch of the run that the rule stepped out for.
Each market from its own first decision, 5 bps per side, before funding. Source:
docs/evidence/walk-forward-2026-10-23.json, 22 Oct 2026.
03For the technical reader
The checks we ran against ourselves.
Survivorship: drop the markets that listed late
The 32 markets are the ones liquid today. To check that late arrivals do not carry the result, we removed the 5 that listed after 1 Jun 2025 (CASHCAT, LIT, PUMP, XMR, XPL). The picture holds: every rule still falls far less than holding, and 2 of 9 make more than holding, against 3 of 9 on all markets.
| All 32 markets, unseen months | 27 early markets, unseen months | |||||
|---|---|---|---|---|---|---|
| Rule | Return | Worst fall | Bar | Return | Worst fall | Bar |
| Breadth Trend 100 | +30.0% | −12.0% | fail | +35.9% | −12.8% | fail |
| Breadth Trend 150 | +30.6% | −11.3% | fail | +30.5% | −13.1% | fail |
| Breadth Trend | +42.3% | −13.9% | pass | +42.9% | −14.6% | fail |
| Golden Cross 50/200 | +30.4% | −17.1% | fail | +32.9% | −18.0% | fail |
| Slow Cross 100/200 | +34.9% | −19.9% | fail | +31.8% | −20.4% | fail |
| Momentum 150 | +43.9% | −18.1% | fail | +46.1% | −20.3% | pass |
| Momentum 200 | +24.5% | −21.3% | fail | +25.2% | −24.2% | fail |
| Momentum 320 | +50.4% | −22.8% | pass | +54.9% | −21.8% | pass |
| Above the 200 | +32.8% | −19.5% | fail | +37.1% | −22.7% | fail |
| Holding | +41.8% | −32.7% | +44.5% | −34.0% | ||
Common start bar 320 for every rule, so these differ slightly from the walk-forward
table. It removes late listings only; a market that died before the freeze is in neither set. Source:
docs/evidence/survivorship-2026-10-23.json, 23 Oct 2026.
Concentration: fewer, larger positions
Sizing the frozen rule into fewer slots raises return and worst fall together. Past 30% the account halt would have stopped the run, which is why the default is one slot per market.
| Before 1 Mar 2026 | From 1 Dec 2025 | ||||
|---|---|---|---|---|---|
| Breadth Trend sized as | Return | Worst fall | Return | Worst fall | Bar |
| 1/32, one slot per market | +59.0% | −23.9% | +42.3% | −13.9% | pass |
| 16 slots | +193.6% | −28.0% | +56.3% | −18.2% | pass |
| 8 slots | +215.7% | −46.0% | +89.1% | −25.5% | pass |
| 4 slots | +835.1% | −55.0% | +102.8% | −39.5% | fail |
| Holding | −8.0% | −65.9% | +41.8% | −32.7% | |
Falls in amber are past the 30% halt. Slot books chase whichever markets ran, so they inherit
more of the survivorship above. Source: docs/evidence/construction-2026-10-23.json, 23 Oct 2026.
Other candle sizes
The same families on daily and 2-hour candles, out of sample, against the same bar, with one slot per market:
daily, 4 of 65 pass; 2-hour, 6 of 73 pass. Across every sizing, 9 of 260 and 2 of 292
rows would have improved on the best 4h results, found after the fact, often with neighbouring
settings that failed. They are leads for research, not listings. Source: docs/evidence/timeframes-2026-10-23.json, 23 Oct 2026.
04Forward test
The frozen rule's universe and settings were fixed on 15 Sep 2026 and a forward window was set before any of it was collected. It closes on 14 Nov 2026. The tooling refuses to report early, on purpose. We will publish the result here whether it passes or fails.
The fine print
Selection
The listed rules were chosen on the same history the full-history numbers come from. Those numbers describe the rules; they do not predict them.
Funding is left out
Every replay here is before funding. Longs on Hyperliquid usually pay it. The one rate the research observed was BTC at 10.95% a year, charged only while a position is held. Source: docs/REFACTOR.md, 15 Sep 2026.
No slippage
Fees of 5 bps a side are charged on every trade, but fills are at the close. Live orders cross up to 1% to fill and pay what the book offers.
Survivorship
The markets are those liquid on the day of the freeze. Markets that faded before then are not in any test.
Short history
Hyperliquid serves at most 5,000 candles per market, a little over two years at 4h. The unseen slice is under a year, in one kind of market.
The bar is relative
Beating holding is easy when holding loses and hard when it wins. Read the holding columns before the pass/fail marks.
05Retired by the gates
The bar is applied every day, not once.
Listing is not a lifetime pass. The research desk rechecks every listed rule against the listing bar on the newest prices, and a rule that no longer clears it makes room. On 28 Oct 2026 the gates retired 6 rules. Accounts already running one keep it until they switch; nobody new is offered it. That is the system doing its job.
- Above the 200retired 28 Oct 2026
- Momentum 150retired 28 Oct 2026
- Momentum 200retired 28 Oct 2026
- Golden Cross 50/200retired 28 Oct 2026
- Slow Cross 100/200retired 28 Oct 2026
- Momentum 320retired 28 Oct 2026
Decision records: docs/evidence/desk/demotions/ and docs/evidence/desk/registry/.