Win rate across our tracked cohort ran from 20.8% to 78.8% in September, so a discovery surface is worth building. The list a venue hands you is a population it defined on rules it may not publish.
Centaur · Research
October 17, 2026 · 10 min read
Every trading app eventually ships a top traders surface. It answers the one question a funded user asks before they place anything, which is who to follow next, and it is the cheapest retention feature available because the alternative is your user finding the answer in someone else's app.
It is also the surface you will be asked to justify. When a user loses money following a name you ranked, the question becomes how that name got there. That answer lives in two decisions most teams inherit rather than make: who was eligible to appear, and what the list was sorted on.
Traders do differ, measurably, and the spread is wide enough to build on.
In our record as pulled on 9 September 2026, across the traders with at least 20 settled 30-day calls, 84 of them over entries from 11 June to 10 August 2026, win rate ran from 20.8% to 78.8%. That is a 58-point spread. The names at the top were not obscure: Peter Brandt sat in that cohort on 22 settled calls, and James Chanos appeared on the seven-day board on 30. These are people your users already follow somewhere else.
That spread is the whole basis for a discovery surface. If every record looked the same there would be nothing to rank, and if none of it were measurable there would be nothing to defend. The problem is never that ranking traders is impossible. Research on the complete transaction records of every day trader on the Taiwan Stock Exchange from 1992 to 2006 found the top-ranked group went on to earn 37.9 basis points a day after fees against -28.9 for the bottom, published in the Journal of Financial Markets in 2014. Past performance carries information when you hold the complete population, every trade and the fees.
The trouble is that a public leaderboard is none of those things.
Venues publish the rules that decide who reaches their boards, and the rules are worth reading before you embed one.
Bitget's display criteria for elite traders, last updated 30 December 2025, require that a futures trader's account profit, ROI and follower profit all sit at or above zero. Its wording: "If any of these metrics are negative, your account will not be displayed on the leaderboard during that period." The list refreshes hourly, so a trader drops off during a losing stretch and returns during a winning one, while existing followers keep copying throughout. Only the shop window is filtered.
The same page settles what kind of choice that is. For spot portfolios the rule is simply that "Anyone who creates a spot elite portfolio can be displayed." One venue, one product family, two boards, and only one of them is gated on performance.
Bybit is the most explicit about the two stages. Its help centre article on how master traders rank in the different categories, updated 7 September 2026, describes a "Recall Stage" that forms the candidate set and a "Ranking Stage" that orders it. Entry to the ROI category requires positive returns over 7 and 90 days, positive follower profit and a drawdown cap, all before the sort runs. Its Lowest Drawdown board, which reads like a risk-adjusted view, requires a minimum 30-day ROI of 10% to qualify, so a board about controlling losses is open only to traders who already made money. Gate's leaderboard rules exclude providers "whose return rate is too low", without defining too low.
This is not universal, and the exceptions matter for how you read it. OKX gates its lead trader list on capital rather than performance, at 100,000 USDT in lead assets. Binance publishes no eligibility criteria at all. Hyperliquid publishes everybody, and 49.3% of its listed accounts carry a negative all-time PnL.
| Board | Who reaches it before the sort | Gate type |
|---|---|---|
| Bitget futures | Account profit, ROI and follower profit all at or above zero | Performance |
| Bitget spot | Anyone who creates a spot elite portfolio, with at least 100 USDT in spot assets | Capital |
| Bybit ROI | Positive 7- and 90-day returns, positive follower profit, drawdown cap | Performance |
| Bybit Lowest Drawdown | Minimum 30-day ROI of 10% | Performance |
| Gate | Excludes providers whose return rate is too low | Performance, threshold undefined |
| OKX | 100,000 USDT in lead assets | Capital |
| Binance | No eligibility criteria published | Not published |
| Hyperliquid | Every account, 49.3% with negative all-time PnL | None |
So the honest generalisation is not that every board hides its losers. It is that each board is a population somebody else defined, on rules that are sometimes about winning, sometimes about capital, and often not published. Filtering a list you were handed cannot recover rows that were never in it.
Anyone presenting performance to professional buyers already works the other way round: the GIPS standards require firms to include every fee-paying discretionary account in a composite, specifically to "prevent firms from cherry-picking their best performance." That is a voluntary standard for investment managers and it does not bind an exchange, which is the point. The discipline exists, and the boards you can license were not built to it.
Even with a clean population, the ranking column does more work than the data.
Take that 9 September cohort of 84 traders, one window, and rank it four defensible ways. The win-rate top ten and the average-return top ten shared one name. Median return shared six, and call count shared one. All four lists are correct, and each answers a different question. A board that shows only its output has chosen one of those questions for your user.
Win rate is the usual default and the one to be most careful with. The sixth name in that cohort closed 59.4% of its calls positive and averaged -6.56% across 32 of them. Sorted on win rate, a record like that outranks traders who made money.
The window matters as much. The same 84 traders ranked on a seven-day horizon instead of a 30-day one shared two names in the top ten. You can check the effect without our data. Hyperliquid's public leaderboard is readable without an account, and the feed behind it returned 47,228 accounts when we pulled it on 17 October 2026. Its 30-day PnL and 30-day ROI top hundreds share three names.
Median account value in the PnL hundred is $14 million. In the ROI hundred it is $9 thousand. The two sorts sit more than a thousandfold apart in account size, so they return two different classes of account rather than two orderings of the same traders. A venue offering both under one heading is offering two products, and a team embedding that board inherits the choice without being told it was made.
Rank is a compression of a history. The history is the durable asset, and it is what a defensible surface is built from.
Take one asset. On 17 October 2026, CoreWeave had 21 tracked positions in our record over the rolling 90-day window, 16 long and 5 short, four of them open. Two calls on 14 October sat on opposite sides. At 14:20 UTC Diligent Plane, an X account with about 79,000 followers, posted that it had bought some CoreWeave.
The post is public. One call returns it as a long on CoreWeave at 67.72, timestamped to the minute with a link back to the source. That account's settled 30-day record over entries from 19 July to 17 September 2026 is 30 calls at a 70.0% win rate, averaging +7.88%. Nine hours later another tracked account posted that it had shorted the same name, at 65.78.
No rank carries that. Two records, both real, both timestamped, on opposite sides of one name inside ten hours. Positions, not vibes. CoreWeave is also an equity, which is worth saying because this is not a crypto-only dataset.
The reason the losing calls are still in that record is mechanical rather than editorial. Centaur reads public posts on X and Telegram and writes the trade down at the moment it appears. Delete the post next week and the record of what was said survives, because it was captured while the post was there. That is why the bottom of our 9 September distribution sat at 20.8%, and a buyer running diligence should want it to. A dataset where every record looks good cannot tell you who to pick.
Five properties make a ranked trader surface defensible, and each is something a user or a regulator can check. We publish a ranked view too, so this applies to us.
The Social Signals Integration Playbook covers how teams wire that into a live surface, including the field set and three integration patterns. It sits behind a short form asking for a name, work email and website, and it is written for product and engineering teams, so it is the wrong document for a research desk. The business case for putting discovery inside your app at all is in your users don't know what to trade.
The venue rules above are what those venues publish, not an audit of what their systems do. A venue that publishes no gate may still apply one.
Our record carries no position size, so it ranks decisions rather than portfolios. A 30-day board also lags by a month by construction, because a call needs 30 days to settle. The cohort figures above come from our record as pulled on 9 September 2026, and the record covers a rolling 90 days, so a pull run today returns a different cohort.
Centaur measures trades that traders disclosed publicly. We do not verify broker or on-chain activity, so a posted call is a disclosure rather than a fill.
The spread between a 20.8% record and a 78.8% one is worth surfacing to your users. The list somebody hands you is not, because the eligibility rule, the sort and the window were all decided before you saw it, and none of them is recoverable from a ranking. All three are recoverable from the calls underneath.
Start with what social trading data is for the shape of the dataset, and the field level breakdown of what the API returns for a single call. To run the same cut on today's record, the partner documentation covers the endpoints, the settled outcome statuses and the classification flags.
