A list of “the best traders on Hyperliquid” sounds like a shortcut: find accounts that are good, then pay attention to what they do. The difficulty is in the word “good”. Any list ranked by recent profit is a filter on outcomes, and outcomes mix skill, risk-taking and luck in proportions the list cannot reveal. A scorecard that measures what tracked accounts do after they were selected is more useful than the ranking itself, but it has limits of its own.
This part of the series explains how SuperQuant’s tracker selects accounts, what its trader scorecards measure, what they leave out, and how to read a hit rate without being misled by a small sample. Part one introduced the selection rules; here they become the main subject.
How the list is selected
SuperQuant’s guide lists the selection rules. Vault accounts (managed pools rather than individual traders) are excluded. The rest need at least $5M of 30-day volume, $1M of 7-day volume, a $100k account and turnover under 1,500 times account value, and are ranked by 30-day profit. The top 100 are re-ranked every six hours and checked about once a minute.
Every step of that pipeline is reasonable, and the last one builds in a bias. Accounts enter the list because their last 30 days went well. An account that took a large, risky position and happened to win will rank highly; an account with the same strategy that happened to lose will not appear at all. That is survivorship in its simplest form: the list contains the survivors of a recent contest, not a random sample of traders with a given approach.
Strategy research has the same problem. The paper that introduced the probability of backtest overfitting (PBO) proposes a way to estimate how often the strategy that looks best in-sample, on the data used to choose it, underperforms the median candidate out of sample, on data it has not seen. The backtest overfitting guide works through one example. A ranked list of traders applies a similar selection to accounts instead of strategy variants.
What the scorecard measures
Because the ranking flatters itself, the scorecard looks at what accounts did after they joined the list. The guide says each top-trader move is measured at +1h, +24h and +7d from the price when it was detected.
- Entries only. Opens, increases and flips count as entries. Closes and reductions are not scored.
- Hit rate. The 24-hour hit rate is the share of measured entries where price moved in the trader’s direction over the 24 hours after detection. The guide says it is shown once 10 moves have been measured; until then the scorecard shows “n/a”.
- Averages and style tags. It also shows the average 24-hour and seven-day moves, the median holding time, and style tags such as scalper (a short median hold), swing (a long one), basket (many positions open at once) and high leverage. The guide says detected moves are kept for 90 days.
What the scorecard leaves out
Scoring only what happens after selection keeps the ranking window out of the score, but it does not remove selection entirely, and it leaves gaps.
- Accounts that drop off. The list is re-ranked every six hours by 30-day profit, so an account whose detected entries go badly is more likely to fall out of the top 100. Scorecards shown next to the current list lean towards accounts that kept doing well.
- The trader’s own fill. Returns start from the detection price. Tracked accounts are checked about once a minute, so detection can come a minute or more after the trader’s fill. For a fast-moving coin, that gap can account for much of the move.
- The trader’s exit. A 24-hour return assumes the position was held for 24 hours. A scalper who exited after 20 minutes is scored on a window they were not in.
- Costs. Fees, funding and slippage are not deducted.
- Size. Averages are per entry, not weighted by position size. A small winning entry counts as much as a large losing one.
- History limits. The scorecard covers only moves detected while an account was tracked, within that 90-day window. If you rebuild a longer record yourself, Hyperliquid’s info endpoint documentation says only the 10,000 most recent fills are available, so a very active account’s history can be incomplete.
These are reasons to treat a scorecard as a description of detected entries, not as an account’s real profit and loss.
Reading a hit rate without fooling yourself
Small samples produce extreme-looking numbers. Suppose a fictional account has 12 measured entries and 7 of them moved in its direction over 24 hours: a 58 percent hit rate. A 95 percent Wilson interval, a standard way to put a range of plausible values around a proportion, runs from roughly 32 percent to 81 percent for 7 successes in 12 trials. In other words, that record is consistent with an account that is right less than half the time, and with one that is right most of the time. The same 58 percent rate on 50 entries narrows the interval to roughly 44 to 71 percent, and on 200 entries to roughly 51 to 65 percent.
Three habits help:
- Look at the count first. A hit rate with fewer than a few dozen measured entries says very little. The scorecard’s 10-move minimum is where a number appears, not where it becomes reliable.
- Compare with a base rate. In a rising market, long entries “hit” more often regardless of skill. Compare the account’s hit rate with the share of all 24-hour windows that rose in the same coin and period.
- Watch for streaks. Runs of wins and losses are normal even for random outcomes. The guide to losing streaks and risk of ruin shows how long they can be.
Avoiding hindsight when you research whale data
If you test an idea such as “follow the top-10 accounts”, rebuild the list as it stood at each date. Using today’s list to evaluate past months tests accounts chosen because those months went well. That is a form of data leakage, which scikit-learn’s documentation describes as information that would not be available at prediction time being used to build a model; it warns that leakage produces overly optimistic performance estimates. The look-ahead bias guide describes the same problem for indicators, and the on-chain data guide describes it for wallet labels applied after the fact.
A checklist for any trader scorecard
- How was the account selected, and over which window?
- How many entries are measured, and from which price?
- Are fees, funding and the trader’s real holding period included?
- What is the base rate for the same coin and period?
- Would the conclusion survive another 20 entries going the other way?
The next part, paper-copying a whale, asks a related question: even if a trader is good, how much of their result would reach a follower? For the wider due-diligence questions about following someone else’s record, read evaluating crypto copy trading.
This is educational research context, not investment advice.