Posts get deleted, accounts go quiet, and traders start publishing once they have something worth publishing. Six questions to ask any source of trader performance.
Centaur · Research
October 28, 2026 · 15 min read
Every public trading record has been filtered before you see it. Posts get deleted. Accounts go quiet after a bad run. Traders start publishing once they have something worth publishing. None of that requires bad faith, and all of it bends the record in the same direction.
If you are choosing which traders to surface in a product, feed into an agent, or carry in a research process, you are making a selection decision on evidence assembled by the person being selected. This sets out a test you can apply to any source of trader performance, including ours.
Manager diligence has a mature apparatus. Due diligence questionnaires, limited partner reports, attribution letters from prior funds, audited composites. The Alternative Investment Management Association has been publishing standardised questionnaires since 1997. All of it assumes a firm that volunteers to be examined and keeps records a third party can inspect.
That apparatus also does less than most people assume. Under the GIPS Guidance Statement on Verification, verification tests whether a firm's policies and procedures for constructing composites comply with the standards. It does not assure the accuracy of any specific composite's input data or calculations. The strongest badge in institutional asset management certifies a process rather than a number.
A trader posting calls on X or Telegram has none of this. No composite, no auditor, no data room, no former colleague to call. The institutional apparatus does not weaken here, it stops applying.
The products that sell track-record verification do not close the gap either. Broker-linked services connect to a trading account and display its history, equity curve, drawdown and returns, which is genuine evidence against a fabricated screenshot. What they establish is that a connected account produced those results. They do not establish that the account holds the person's complete trading activity, that no other accounts exist, that the account predates the connection, or that the public calls match what was actually traded. The distinction a buyer needs is between a claim that a connected account produced certain results and a claim that a record is complete. The first is commercially available. The second is not.
So the test has to be rebuilt from what public activity actually leaves behind, which is the subject of what social trading data is and where its boundaries sit.
Hedge fund databases ran into these problems first, and the research measured them. The names transfer cleanly onto public posts.
Backfill. Funds report to a database after the fact, and disproportionately when early results were good. In an analysis of the TASS database, Saha (2006) found that backfilled returns averaged more than 700 basis points higher than returns that had been reported contemporaneously. The public analogue is the call that was posted and later removed.
Survivorship. The same analysis puts the gap at 442 basis points, 13.74% for surviving funds against 9.32% across all funds from 1996 to 2003. One detail transfers with unusual precision: funds that stop reporting underperform in the months before they go quiet. The public analogue is the account that posts less and then stops.
Selection. Nobody is obliged to publish a call. A record that exists because its owner chose to create it is a sample that owner selected.
The consequence runs further than inflated averages. Brown et al. (1992) showed that truncating a sample on performance manufactures the appearance of skill where none exists. Simulating 600 managers across 20,000 replications with no true persistence built in, they found that once managers are dropped for poor performance, apparent persistence shows up among the survivors. Even mild truncation produced an unacceptably high chance of concluding that past performance predicts future performance.
Applied to a public feed, that is the resting state rather than an edge case, and the disappearance itself has been measured. A 2025 study of two large Twitter corpora found that 23% to 29% of posts were unavailable roughly a year after posting, and that the posts which went missing carried more negative sentiment than those that remained (Khan et al., 2025). At the extreme, an analysis of more than a billion deletions across more than eleven million accounts found that deletion is used strategically rather than incidentally (Torres-Lugo et al., 2022). What disappears from a public feed is not a random sample of what was posted.
One boundary belongs here. Those studies count posts that became unavailable, a category that includes suspensions, account deletions and privacy changes as well as deliberate removal, and none of them tests whether a financial call is more likely to be deleted when it turned out wrong. No published study measures that. So the mechanism is established and the disappearance is measured, while the specific rate at which losing calls vanish is not, and the fund-database magnitudes above remain the closest thing to a size estimate.
Copy-trading venues make this concrete, and it has been documented rather than alleged. In a study of cryptocurrency copy-trading platforms, Kawai et al. (2024) report that leaders can close losing portfolios and open new ones at their discretion, which lets them assemble a better history by clearing the underperformers. On TraderWagon the median portfolio life was 2.9 days. Bybit restricted leaders to a single portfolio, and fewer than 20% of its portfolios were active against more than half on TraderWagon.
The same paper shows the interface driving the selection. Copiers predominantly follow whichever ranking the platform puts in front of them, and those rankings track leader popularity more closely than future returns. When TraderWagon removed its return-based front-page ranking in March 2023, the number of copiers fell measurably.
Ranking on raw past return fails on its own terms too. Dorfleitner et al. (2018) tested strategies that follow social trading platform leaders and found that selecting the traders with the highest accumulated returns produced heavy losses, that adding risk information improved the outcome, and that no strategy they tested delivered a positive abnormal return once transaction costs were counted. They also found no positive relationship between a leader’s follower count and their returns.
If you are working out how to put trader activity in front of users without inheriting these problems, the Social Signals Integration Playbook covers the three integration shapes. It asks for a first name, a work email and a website.
The correction to leaderboard optimism is not that past performance carries no information. It does, and the best measurement of it is large.
Barber et al. (2014) examined every day trader on the Taiwan Stock Exchange from 1992 to 2006. Sorting traders on one year’s returns and measuring them the next, the top 500 went on to earn 61.3 basis points a day before fees and 37.9 after, while the lowest-ranked lost 11.5 before fees and 28.9 after. The authors state plainly that performance persistence is present.
The scale is the part that changes a product decision. Of roughly 450,000 people who day-trade in an average year, fewer than 1% predictably outperform after costs, and about a fifth finish a typical year profitable, which is close to what chance delivers. Their method doubles as a specification worth copying: they rank on the Sharpe ratio of past net returns, and they require a minimum of ten trading days before a trader is ranked at all.
The regulated fund world produces the same shape. In the U.S. Persistence Scorecard Year-End 2024, S&P Dow Jones Indices found that among top-quartile active domestic equity funds at December 2020, not one remained in the top quartile across the following four years, and that 4.21% of US equity funds in the top half in the first year stayed in the top half over the four that followed, against the 6.25% a chain of coin flips produces.
Carhart (1997) had already shown that most measurable persistence in mutual funds is explained by common return factors and expenses, with the one reliable pattern being continued underperformance by the worst funds.
Three conclusions follow for anyone selecting traders. A single good window is close to meaningless. A rank on raw accumulated return is the wrong sort order. And the most dependable signal in the data sits at the bottom of the distribution rather than the top.
What a reliable prediction record looks like is not an open question. Forecasting tournaments settled it: questions are logged before they resolve, forecasts are timestamped, revisions are retained, scoring covers the complete set of resolved questions, and a participant cannot choose which forecasts enter the record. One difference matters more than the rest. A tournament participant who stops submitting is visible as missing participation, while a trader who stops posting is not.
The six questions below are that design applied to a source you do not control. They are worth forwarding to whoever owns the integration, and they apply to a venue leaderboard, a data vendor, an internal build, or Centaur.
1. Was the call recorded when it was made? A record reconstructed later inherits every revision made in between. A good answer: each event carries the timestamp of the message that disclosed it and the price at that moment, captured at that moment.
2. Can you reach the primary evidence? You should be able to move from any row back to the original message and read it yourself. A good answer is a link, per event, that resolves today.
3. Are the losing calls still in it? Ask specifically what happens to the record of an account that stopped posting, and to a call the trader later deleted. A good answer describes retention as a property of how capture works rather than as a policy that could change.
4. Is there a stated minimum sample? Serious data providers publish floors. Preqin's Private Capital Performance Data Guide requires at least four funds to report a median and eight for quartiles. Barber and co-authors required ten trading days. A good answer names a threshold below which no ranking is published.
5. Is there a clear evaluation metric? The same trader reads differently depending on how trades are closed and measured, and a win rate quoted without measure is unreproducible. A good answer attaches a window and a date range to every number, and prefers the longer sample when two disagree.
6. Is there a source for how the trade appeared? A provider that has thought about this will tell you where its evidence stops before you ask. A source with no stated limits has not been examined closely enough.
Centaur's record is built from public posts across crypto, equities and commodities, which puts it in a specific position against these six questions rather than a flattering one.
Capture is the answer to questions 1 and 2. Each event is timestamped to the message that disclosed it, carries the price at that moment, and links back to that message. Attribution is measurable: adds and trims trace to the trader's own message 100% of the time and opens 78.9% of the time, in a 2 September 2026 read of Centaur's API, with the wider extraction figures set out in what social trading data adds to trading agents.
Question 3 is structural here. A record assembled from public posts sits outside any single venue, so a trader who goes quiet or deletes a call does not remove what was already captured. The traders whose calls went wrong are still in the data, which is what makes a win rate worth computing at all.
A live example, for question 2. Between 20 and 27 October 2026, seven distinct traders posted adds to Nebius longs, at 170.86, 170.86, 172.09, 171.57, 153.09, 159.05 and 170.15 USD, as the stock dipped and recovered. Each one resolves to the message that disclosed it. That is an equity, which is the ordinary case rather than an exception.
On question 6, the boundary is worth stating precisely. Centaur measures trades disclosed publicly by traders and does not verify broker or onchain activity. A posted call is a disclosure rather than a filled order, and nothing in the record asserts that anyone executed at the price shown. That distinction is the reason the data supports research and product decisions rather than attestation, and it is the same reason the gap between a claimed record and a checkable one matters more in this category than in regulated funds.
The reason to build the test is that the alternative is inheriting somebody else's filtering. A venue leaderboard that hides losing traders, a ranking sorted on accumulated return, or a feed whose deleted calls quietly disappear will each produce a shortlist that looks better than the population it came from. Brown and co-authors showed that this manufactures the appearance of skill on its own, with no help from anyone acting in bad faith.
Apply the six questions to whatever you use now, including a leaderboard you built yourself. The answers you cannot get are the more useful half of the exercise.
The API and MCP surface, the event schema and the field definitions are at docs.centaur.io. If the immediate question is how trader activity should appear inside a product, start with the Social Signals Integration Playbook, which is gated on a first name, work email and website.
Saha, A. (2006). Caveat Emptor: Analyzing Hedge Fund Returns. Analysis Group. https://www.aei.org/wp-content/uploads/2011/10/20060515_SahaNPSTFinal.pdf
Barber, B., Lee, Y., Liu, Y. and Odean, T. (2014). The Cross-Section of Speculator Skill. Journal of Financial Markets, 18, 1–24. https://faculty.haas.berkeley.edu/odean/papers/day%20traders/The%20Cross-Section%20of%20Speculator%20Skill.pdf
Brown, S., Goetzmann, W., Ibbotson, R. and Ross, S. (1992). Survivorship Bias in Performance Studies. Review of Financial Studies, 5(4), 553–580. https://terpconnect.umd.edu/~wermers/ftpsite/FAME/Brown_Goetzmann_Ibbotson_Ross.pdf
Carhart, M. (1997). On Persistence in Mutual Fund Performance. The Journal of Finance, 52(1), 57–82. https://doi.org/10.1111/j.1540-6261.1997.tb03808.x
Kawai, Soska, Routledge, Zetlin-Jones and Christin (2024). Stranger Danger? Investor Behavior and Incentives on Cryptocurrency Copy-Trading Platforms. https://www.andrew.cmu.edu/user/nicolasc/publications/Kawai-CHI24.pdf
Khan, Dimitrov and Dietze (2025). Characterization of Tweet Deletion Patterns in the Context of Covid-19 Discourse and Polarization. ACM. https://doi.org/10.1145/3720553.3746663
