METHOD · DATA AS OF 6 OCTOBER 2026

How traders compare with Guard’s defaults: the method

We replayed real Hyperliquid accounts through Guard’s own engine and counted how many of their entries would have fit Guard’s default rules. This page says how, what the numbers can and cannot carry, and every measurement taken. It shows no data about any single account.

The question

Did long-term winners on Hyperliquid trade within Guard’s defaults more often than accounts that were liquidated? The answer was to be published whichever way it came out. The protocol was written down before any per-account result was pulled.

What “fit” means

• Strict (D1): the entry would have passed with no change at all. The median was 0% for winners and 0% for liquidated accounts: almost nobody traded with a resting stop.
• With Guard’s stop (D2, the number on the homepage): the entry passes unchanged once Guard attaches its 2% stop. The size limits are not loosened for it.
• Both are shares of an account’s entries that could be audited. They are counted per account, for accounts with at least 20 audited entries.

The rules

Guard’s defaults:

• leverage 5×;
• 2% loss at the stop;
• a 2% stop attached where an order has none;
• liquidation at least 10% away;
• position at most 200% of equity;
• open risk 6%;
• daily stop 6%;
• drawdown halt 25%.

Every verdict comes from Guard’s engine (zunder-risk 0.1.0 (8fedd4d)), through the same replay this site’s backtest uses.

Who was measured

• Long-term winners:
  • From Hyperliquid’s leaderboard (47,471 accounts), those with positive PnL and at least $10,000 today (16,142). The top 3,000 by PnL were screened.
  • They had to have at least 12 months of history, perp PnL above zero, and profits not concentrated (the three best periods made under half of it). 49 passed.
  • After the replay, accounts with no fills, market makers and accounts with under 100 entries were left out. That gives 20 accounts, 15 of them measurable.
• Liquidated accounts:
  • Every account liquidated on Hyperliquid in September 2026 (61,635); 24,623 of them had at least $1,000 liquidated.
  • They were shuffled with a fixed seed and replayed in order until 80 passed. 57 were measurable.
• Public vaults were tried as a third group. Too few could be audited for stops, so they are not shown.

Per account

• The replay: the last 180 days, or less where the window holds more than the 10,000 fills the replay reads, its most recent (11 of 20 winners; until 7 Oct 2026 the replay kept the oldest 10,000 of such a window instead, so those accounts' figures are to be rerun).
• The verdicts: every entry is judged as it was traded.
• Two curves: the account as it traded, and the account behind Guard.
• Markets: main-dex perps only; HIP-3, spot and outcome markets are left out.

Statistics

• Intervals: medians per account, with bootstrap 95% intervals (10,000 resamples, seed 20261006).
• Shares: Wilson intervals.
• Comparing the groups: Mann-Whitney U.

Results

• The difference in fit: median 83 points (95% interval 19–100), mean 58 points (interval 36–77).
• Without the halt: with the halted entries taken out, the difference is 57 points (interval 4–78). The ordering does not rest on the halt alone.

The blow-ups

For each liquidated account we looked at the entries that built the position that was finally liquidated.

• Overall: Guard’s defaults would have refused, resized or halted at least one of them in 78 of 80.
• By a sizing rule: leverage, open risk, loss per trade or the position cap acted in only 8 of 80 (10%, interval 5%–19%).
• By the halt alone: in 70 of 80, only the drawdown halt acted, because the account had already lost 25%.

The halt is measured without deposits. In 35 of the 80 accounts, money was added along the way. So the halt-only count is an upper bound, and the homepage does not claim that Guard “would have stopped” these blow-ups.

Limits

• Survivors. The winners are today’s leaderboard. Traders with the same habits who failed are not among them, and some of them are among the liquidated. The numbers say what surviving winners did, not what makes traders win.
• Small and filtered. 15 measurable winners out of 3,000 screened.
• Different worlds. The median winner started with $396,000; the median liquidated account with $228. Part of the gap is size and experience, not rules.
• Stops are known only for an account’s last 2,000 orders. Older entries cannot be audited for a stop.
• Main-dex perps only. Entries in HIP-3, spot and outcome markets are not judged.
• A what-if. Skipping or shrinking a trade could have changed the trader’s next decisions; the replay keeps them. Its approximations are stated in its output, and several still favour Guard.
• The liquidated sample was chosen for having lost everything. Guard’s drawdown halt stops trading after a 25% loss by design, so part of their improvement is mechanical.
• No evidence that the winners’ risk fell. Their drawdown and worst day did not change significantly with Guard. Nor is there evidence that winners who stayed within Guard’s limits did better.
• Past behaviour, not a forecast and not a promise of returns. Guard is a risk tool. It gives no signals and no investment advice.

Every measurement taken

Nothing else was tried: no filter, threshold or definition was tried and dropped.

Deviations from the protocol

1. The winners’ screening pool was widened from 1,000 to 3,000 before any replay ran, because only 24 of the first 1,000 passed the unchanged filters. The first 1,000 are reported on their own too.
2. For accounts with more than 10,000 fills in the window, the replay kept the window's oldest 10,000 fills, not its most recent (a data-path bug fixed on 7 Oct 2026); their figures describe that earlier stretch and are to be rerun.
3. Engine defects found while running (same-millisecond ordering, no marking at the window start, settlements, outcome markets) were fixed on main independently of the results, and the same cohorts were rerun. Both versions are reported; the page uses the corrected one, and a claim whose direction differs between them is not used.

What we do not claim

• Not that top traders follow Guard’s defaults. On the strict reading almost none did.
• Not that Guard would have prevented the liquidations. The liquidated accounts’ positions behind Guard were smaller, but they are not shown to have survived.
• Not that staying within Guard’s limits makes traders do better. That is not supported.

Data and upkeep

• Data: as of 6 October 2026, aggregates only. No address or per-account figure is published.
• Upkeep: the study is rerun before a number older than three months is shown. If a rerun reverses a direction, the section comes down until it has been reviewed.