口径与边界
这一页写我们怎么算的，以及哪些话我们不说。
做这行太容易把「测过」说成「有效」，所以把边界摆在明处。
我们不说什么
不说「我们会挑人」。三个维度的前瞻检验都不显著，
按收益率选人的点估计甚至比随机差。真正起作用的是「这一笔他压得多重」，不是「他是谁」。
不说「跟单一定赚钱」。把消失的账户计入之后，
纸面十个月只是打平——同期买 BTC 亏 25%，我们比它少亏，但不是会赚。
不预测行情、不发买卖信号、不承诺收益。
站上所有百分比都是历史同类下注的分布，不是对你这一笔的预测。
会说坏消息：池子按盈利筛选过、存在幸存偏差、
417 天里只有 77 天是涨市、牛市会跑输 BTC。
会说你要付多少：成交额 0.1%，不抽利润分成、
不托管、授权随时可在你自己的钱包里解除。
数据怎么来的
数据源是 Hyperliquid 官方 S3 的逐笔成交，自 2025-07-27 起逐日重建，
目前 418 个日分区。我们把每一笔成交还原成逐地址、逐币、逐日的盯市账本：
1,133 万行、9,733 个地址、440 个币。
盈亏一律按当前市价算：手里还没平的浮盈浮亏、付出去或收到的资金费、
交易手续费，全都算进去。
这和交易所自带的「已实现盈亏」不是一回事 —— 后者只在平仓那一刻才记账，
于是一个「赚了就平、亏了死扛」的人，在它眼里胜率能到 100%。
按市价算就没法这样藏：账面上亏着，就是亏着。
跟仓的赔率是怎么算的
分母是你自己投入的钱，不是鲸鱼的权益。这一条很关键——
我们试过用鲸的期初权益、平均权益、峰值权益、加仓前仓位当分母，每一个都会骗人
（充提污染、分母趋零、被比较的动作本身改变分母）。换成「你投多少、拿回多少」就全干净了。
| 项 | 口径 |
|---|
| 收益 | 方向 × (出场价 ÷ 进场价 − 1) × 事件杠杆 |
| 进场 | 事件次日开盘（滞后一天，不用当日收盘，避免前视） |
| 持有 | 30 个交易日 |
| 成本 | 28.63 bps 往返（2026-09-21 真钱实测：HL 8.64 + builder 20.00） |
| 事件杠杆 | 该笔名义仓位 ÷ 他当时的账户权益 |
⚠️ 晚进场会吃掉赔率。实测：次日进场中位 6.08%，
晚 7 天只剩 2.08%（被吃掉 66%）；
但「翻倍」概率只从 17.4% 掉到 15.9%。
稳定边际集中在事件后头几天，右尾靠的是大行情、七天内跑不完。
所以信号页按新鲜度分组，每组标各自的赔率，超过 7 天的不再标注。
已知的偏差
幸存偏差。池子里 81.5% 是盈利地址——
它是按盈利筛进来的。同样打法没扛过来的那批人，不在我们的数据里。
只有一个 regime。417 天里只有 77 天是涨市，
而涨市里有些结论会翻转（比如「跟加仓优于跟开仓」在涨市中是反的）。
进场延迟没测过。所有数字按「次日开盘」算，
真实产品是分钟级跟进去，差多少目前不知道。
没设爆仓限制。高杠杆档的最差 5% 超过 −100%，
意思是本金亏光还倒欠——真实账户会在那之前被强平。
非托管：交易授权走 Hyperliquid 协议层的 approveAgent，
代理只能下单、转不走钱，你随时可以在自己的钱包里解除。
Method & limits
This page says how we compute things, and what we refuse
to claim. In this business it is far too easy to turn "we tested it" into
"it works", so the limits go in plain sight.
What we do not claim
We do not claim "we can pick people". All three
forward tests came back insignificant, and picking by return rate scores worse
than random on the point estimate. What actually carries signal is
how hard they staked this one trade, not who they are.
We do not claim "following makes money". Once the
accounts that disappeared are counted back in, ten months on paper come out
flat — buying BTC over the same period lost 25%, so we lost less,
but that is not the same as making money.
We do not forecast the market, do not issue buy or sell
calls, and promise no returns. Every percentage on this site is the
distribution of comparable past bets, not a prediction about yours.
We do report the bad news: the pool was filtered
for profitability, survivorship bias is present, only 77 of 417 days were
rising days, and we underperform BTC in a bull run.
We do state what you pay: 0.1% of volume.
No cut of your profit, no custody, and the authorization can be revoked from
your own wallet at any time.
Where the data comes from
The source is Hyperliquid's official S3 trade-by-trade feed, rebuilt
day by day since 2025-07-27 — 418 daily partitions so far. Every fill is
replayed into a mark-to-market ledger per address, per coin, per day:
11.33M rows, 9,733 addresses,
440 coins.
P&L is always marked to the current price: unrealized moves on open
positions, funding paid or received, and trading fees are all included.
This is not the same as an exchange's "realized P&L", which only books at
the moment a position closes — so someone who takes profits quickly and
sits on losers can show a 100% win rate there. Marking to market leaves nowhere
to hide it: if the account is down, it is down.
How the odds are computed
The denominator is the money you put in, not the whale's equity. This
matters more than it sounds — we tried their starting equity, average equity,
peak equity, and pre-add position size as denominators, and every one of them
lies (deposits and withdrawals contaminate it, it trends to zero, or the very
action being measured changes it). "What you put in, what you get back" is clean.
| Item | Definition |
|---|
| Return | side × (exit ÷ entry − 1) × event leverage |
| Entry | next day's open after the event (lagged one day,
never the same day's close — that would be look-ahead) |
| Holding | 30 trading days |
| Cost | 28.63 bps round trip (measured live 2026-09-21: HL 8.64 + builder 20.00) |
| Event leverage | notional of that trade ÷ their account equity at the time |
⚠️ Entering late eats the odds. Measured: entering the next
day gives a median of 6.08%; waiting seven days leaves
2.08% (66% of it gone). But the chance of a
"double" only falls from 17.4% to
15.9%. The steady edge sits in the first few days, while
the right tail depends on a real move that takes longer than a week to play out.
That is why the signals page groups by freshness and quotes each group's
own odds — and stops quoting them past 7 days.
Known biases
Survivorship. 81.5% of the pool are profitable
addresses — because the pool was filtered for profitability. The people who
played the same way and did not make it are not in our data.
Only one regime. Just 77 of 417 days were rising
days, and some conclusions flip in a rising market (for instance "following an
add beats following an open" reverses).
Entry delay is untested. Every number assumes "next
day's open". The real product follows within minutes; how much that differs is
something we do not currently know.
No liquidation floor. In the high-leverage buckets
the worst 5% goes past −100%, meaning the capital is gone and then some — a real
account would have been liquidated before that point.
Non-custodial: trading authorization uses Hyperliquid's protocol-level
approveAgent. The agent can place orders and cannot move your funds, and you
can revoke it from your own wallet whenever you want.