HyperGrass v9 — what it does
v9 turned +2016% over 5.9 years in a backtest.
Nothing is leveraged: the model cannot spend money it does not have.
The short book runs only while the gate is shut, and no long book is open then.
What the thing actually does
Two rules, and only one of them matters.
- When to be in the market. A breadth-and-momentum reading called heat has to be above 74, and bitcoin has to be trading above its own 200-day average. Both, not either. If that is not true, the book sits in cash. Over six years it was in the market 18% of days.
- What to buy. When the gate is open, hold the 5-6 strongest coins by risk-adjusted momentum, equal weight, no more than 20% in any one name.
A third rule was added and then argued about all day: while the gate is shut, run a small short book — never more than half the capital — against coins paying abnormally high funding.
Rule two does nothing
This is the finding that matters, so it goes before the pretty numbers. We replaced the coin-picking with a coin flip — same days, same position sizes, same costs, only the names scrambled — and ran it many times. Then we tried the laziest possible version: buy every coin the gate allows, no ranking at all.
| Version | Total | Sharpe |
|---|---|---|
| v9 | +2016% | 1.52 |
| Buy everything the gate lets through | +1706% | 1.54 |
| Random coins, median | +1360% | 1.42 |
| Random coins, the lucky 5% | +3129% | 1.72 |
And one random draw in twenty beats v9 outright. The momentum ranking, the risk adjustment, the hysteresis band — none of it survives its own control. What is working is rule one: knowing when to be out of the market. That is a market-timing rule, and we will describe it as one.
The bear leg: two and a half per cent a year, and we cannot prove it is real
The short book multiplies the final result by +15.9% over 5.9 years. Spread out, that is about 2.5% a year. Here is what it did each year:
| Year | Bear leg |
|---|---|
| 2020 | 0.00% |
| 2021 | +0.34% |
| 2022 | −2.57% |
| 2023 | −4.00% |
| 2024 | +2.93% |
| 2025 | +17.21% |
| 2026 | +2.39% |
2022 and 2026 are the years a short book exists for; it was negative in 2022 and it cut 2023. Remove five days from six years and the whole contribution goes to zero. A single day — 6 April 2025 — is worth more than a third of it.
We also checked the story we were telling ourselves. The idea was that shorting coins with sky-high funding pays you to wait. It does not: by the time the position exists, the funding has decayed to about 0.004 percentage points, against 0.20 points of trading costs. That is fifty times too small. Whatever the short book earns, it is not carry.
Statistically it is a coin flip: p ≈ 0.11 against random coins, p ≈ 0.22 once you account for the fact that we tried several thresholds before settling on one. The honest range for its contribution is roughly −4% to +42%.
Twenty days
Six years, 2137 trading days. Take out the twenty best ones:
| Scenario | Total |
|---|---|
| Every day | +2016% |
| Minus the best 5 days | +836% |
| Minus the best 10 | +416% |
| Minus the best 20 | +92% |
| Just holding BTC | +482% |
This is not a steady strategy and must never be described as one. It is a positive-skew bet: mostly nothing, occasionally a lot. Fewer than half its trades make money (49% win rate). That is normal for this kind of system and it is also the reason a bad year feels much worse than the average suggests.
Year by year
| Year | Long only | With the bear leg |
|---|---|---|
| 2020 | +15.1% | +15.1% |
| 2021 | +267.7% | +269.0% |
| 2022 | 0.0% | −2.6% |
| 2023 | +120.5% | +111.7% |
| 2024 | +98.7% | +104.5% |
| 2025 | −1.6% | +15.3% |
| 2026 | 0.0% | +2.4% |
The two flat years are cash, not losses. In 2025 the basket fell 67.0% and bitcoin fell 7.3%; the model sat out most of it. A year with no opportunity is not a defect to engineer away.
What we could not break
Four independent adversarial audits ran against this engine, the last one specifically against v9. They were asked to falsify it, not to confirm it.
- No look-ahead. We cut the raw price and funding archives off at two past dates and replayed everything. Zero of 1,382 trading days changed, and the closed trades were identical. The model never reads tomorrow's data.
- No leverage, no double-siding. Both are enforced by assertions that stop the run. Neither ever fired.
- Reproducible. Same input, same output, byte for byte, under four different random seeds.
What the audits did find were accounting and description errors — eleven of them across the programme, every one now fixed or written down. The most recent: the book size was computed from the coin list as it exists today rather than as it existed on the day, the take-profit we advertised on the short book turned out to cause 12 of 128 exits and removing it improved the result, and funding carry was credited from one settlement too early.
Where this leaves it
Our own house rule blocks promotion when a backtest looks anomalously good until a timing audit clears it. v9 clears the timing audit and fails a more basic test: it cannot beat a coin flip at the thing it claims to do. The defensible sentence is narrow — a regime gate that kept the book in cash through 2022, 2025 and 2026 and avoided an 82% drawdown — and that is a risk control, not an edge.
The fine print, for anyone who wants to argue with it
- Survivorship. The archive is 31 coins that are still listed today. Coins that listed and died are simply absent. Every absolute number here is inflated by that, and by an unknown amount.
- The gate is measured on the same hindsight list. Recomputing it across the wider archive costs 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. The short book's edge is 0.63 points a trade against 0.20 points of assumed friction — it halves at 0.20% slippage and inverts at 0.50%.
- Effective sample. 356 trades is not 356 independent observations. The market was open to us on 18% of days, in about 40 separate episodes; five of them carry half the exposure.
- Multiple testing. Several hundred configurations were evaluated across this programme. Some of them look good by chance, and that budget accumulates across sessions rather than resetting.
Descriptive research on one venue's history. Not investment advice, not a performance claim, and not a promise about anything forward. Back to the lab →