HyperGrass v8 — the gate, the combinations, and the control that beats it
v8 is the version that runs live. It returned +1579% over 5.9 years in a backtest, and the honest reading of that number is narrower than it looks.
What v8 is, in two rules
- When to be in the market. A breadth-and-momentum reading called heat must be above 74, and bitcoin must trade above its own 200-day average. Both, not either. Otherwise the book sits in cash — which it did on 82% of days.
- What to buy. While the gate is open, hold the five or six strongest names by risk-adjusted momentum, equal weight, no more than 20% in one name, with a hysteresis band so a drifting holding is not churned.
Long only. No leverage — the model cannot spend money it does not have,
and an assertion stops the run if it tries. The universe is 34 coins
(majors-34-07090526).
Rule one does the work. Rule two does not.
We replaced the coin-picking with a coin flip — same days, same position sizes, same costs, only the names scrambled, 200 times. Then the lazier test: buy every coin the gate allows, with no ranking at all.
| Version | Total | Sharpe |
|---|---|---|
| v8 | +1579% | 1.41 |
| Buy everything the gate lets through | +1049% | 1.35 |
| Random coins, median | +716% | 1.17 |
| Random coins, the luckiest 5% | +1585% | 1.45 |
+1579% against a random p95 of +1585%, Sharpe 1.41 against 1.45. One coin-flip run in twenty does better than our selection. The momentum ranking is not carrying this result — the gate is. That makes v8 a market-timing rule, and we describe it as one.
A note on method, because it nearly caught us: the first pass ran 40 random draws and put v8 above the p95. At 200 draws the p95 moved from +1364% to +1585% and the verdict flipped. A null estimated from too few draws is not a null.
The combinations we tried
Each row changes exactly one thing about v8. The last column is the number that matters most: what is left after removing the twenty best days.
| What changed | Total | maxDD | Sharpe | Minus best 20 days |
|---|---|---|---|---|
| v8 — as published | +1579% | −23.5% | 1.41 | +50% |
| gate on heat alone, no BTC trend confirmation | +1221% | −41.1% | 1.22 | +16% |
| gate opens at heat 66 instead of 74 | +1751% | −32.2% | 1.38 | +58% |
| plain momentum instead of risk-adjusted | +1112% | −30.5% | 1.20 | −12% |
| inverse-volatility weights instead of equal | +1508% | −22.9% | 1.44 | +52% |
| no hysteresis band — exit the moment a name leaves the top | +2199% | −24.1% | 1.46 | +68% |
| + a short book while the gate is shut, capped at 50% | +1824% | −23.7% | 1.47 | +72% |
| the same short book, capped at 25% | +1675% | −24.0% | 1.44 | +59% |
Dropping the hysteresis band scores +2199% at Sharpe 1.46; opening the gate at heat 66 scores +1751% but at −32% drawdown instead of −23%. Both were measured on the whole history they are scored against, which is the same mistake as picking the best of a search and calling it a finding. v8's parameters were chosen blind, fold by fold, and they stay that way until a blind test says otherwise.
What the table does say clearly is where the value sits: remove the BTC trend confirmation and the drawdown nearly doubles, −23.5% → −41.1%, for less return. That one condition is the difference between a gate and a bear-market-bounce detector.
The bear leg
Adding a short book while the gate is shut improves every headline number — and it is not an improvement worth the name. The full argument is in the v9 report: about 2.5% a year, negative in 2022 and 2023, zero after removing five days, p≈0.22 once the threshold search is counted, and a funding rationale that is off by a factor of fifty. It is not traded live.
| Control | Total | maxDD | Sharpe |
|---|---|---|---|
| Hold the basket, never trade | +645% | −85.1% | 0.80 |
| BTC+ETH gated on a 140-day average | +1888% | −40.9% | 1.24 |
| The short fade on its own, gated | +12% | −30.5% | 0.19 |
| The same fade with no regime gate | −88% | −92.9% | -0.54 |
| Short everything liquid while risk-off | −36% | −62.5% | -0.02 |
The two ungated rows are why the gate exists at all: the same fade construction without a regime filter loses 88% of the account.
Twenty days
| Scenario | Total |
|---|---|
| Every day | +1579% |
| Minus the best 5 days | +655% |
| Minus the best 10 | +315% |
| Minus the best 20 | +50% |
| Just holding BTC | +517% |
This is a positive-skew bet: mostly nothing, occasionally a lot. Fewer than half the trades make money (42% win rate). Normal for the type, and the reason a bad year feels far worse than the average suggests.
Year by year
| Year | v8 | Drawdown that year | BTC buy-and-hold |
|---|---|---|---|
| 2020 | +15.1% | −17.7% | +177.0% |
| 2021 | +235.2% | −17.8% | +57.6% |
| 2022 | 0.0% | 0.0% | −65.3% |
| 2023 | +106.4% | −22.1% | +154.5% |
| 2024 | +122.6% | −23.5% | +111.8% |
| 2025 | −5.3% | −18.4% | −7.3% |
| 2026 | 0.0% | 0.0% | −27.4% |
The flat years are cash, not losses. That is the whole point of the gate, and it is also why v8 gives up most of 2023's and 2024's upside relative to holding.
What actually traded
| Coin | Trades | Avg | Win rate | Total |
|---|---|---|---|---|
| SOL | 18 | +22.9% | 50% | +413% |
| ADA | 12 | +31.1% | 33% | +373% |
| XRP | 9 | +35.4% | 67% | +318% |
| AVAX | 11 | +21.1% | 36% | +232% |
| XLM | 9 | +25.0% | 56% | +225% |
| HBAR | 9 | +21.0% | 44% | +189% |
| NEAR | 12 | +10.2% | 25% | +123% |
| BNB | 17 | +6.4% | 53% | +109% |
| DOT | 6 | +15.7% | 50% | +94% |
| ETH | 20 | +4.3% | 55% | +86% |
| TIA | 3 | +26.9% | 67% | +81% |
| BTC | 18 | +4.2% | 39% | +76% |
And the names that cost money:
| Coin | Trades | Avg | Win rate | Total |
|---|---|---|---|---|
| SEI | 9 | -1.1% | 44% | -10% |
| APT | 7 | -2.1% | 29% | -15% |
| EOS | 2 | -11.1% | 50% | -22% |
| ETC | 4 | -6.3% | 25% | -25% |
| VET | 5 | -5.2% | 20% | -26% |
| LTC | 11 | -4.9% | 18% | -54% |
What we could not break
- No look-ahead. The raw price and funding archives were truncated at two past dates and everything replayed: 0 of 1,382 trading days changed and the closed trades were identical.
- No leverage, one direction at a time. Both enforced by assertions that stop the run. Neither has ever fired.
- Reproducible byte for byte under four different random seeds.
Four adversarial audits found eleven defects across this programme, all of them accounting or description rather than timing, and all now fixed or written down. The most recent: book size was computed from today's coin list instead of the day's, and funding carry was credited one settlement too early.
Where this leaves it
The defensible sentence is narrow: a breadth-and-momentum regime gate, confirmed by BTC's own 200-day trend, kept the book in cash through 2022, 2025 and 2026 and out of an 85% drawdown. That is worth running. The coin selection on top of it is not distinguishable from chance on this evidence, and must not be sold as one.
The fine print
- Survivorship. The archive holds 34 coins still listed today. Coins that listed and died are absent. Every absolute number here is an upper bound.
- The gate is measured on the same hindsight list it is applied to. Recomputing it across the wider archive costs about 38% of the return and moves the drawdown from −24% to −40%.
- Costs. 0.30% round trip on spot, 0.20% on perps, adverse slippage on every fill.
- Effective sample. 271 trades is not 271 independent observations: the gate was open on 18% of days across roughly 40 episodes, five of which carry half the exposure.
- Multiple testing. Several hundred configurations were evaluated across this programme, and that budget accumulates across sessions rather than resetting.
Descriptive research on one venue's history. Not investment advice, not a performance claim, not a promise about anything forward. Back to the lab →