A definition of social trading data, the seven rungs between a mention and a verified fill, and the four questions to ask any vendor selling it.
Centaur · Research
September 4, 2026 · 9 min read
Social trading data is structured, machine-readable information about market actions that identifiable traders disclosed in public. Each record carries an asset, a direction, the price at the moment the post went out, a timestamp, the trader's identity, and a link back to the original post.
It is not a sentiment score, a popularity ranking, a wallet label, or a copy-trading instruction. That definition is narrower than the way the phrase usually gets used, and the narrowness is the point.
Search for social trading and you will mostly find a retail product: a broker where you follow another investor and mirror their portfolio at the click of a button. That has been the dominant meaning of the phrase for over a decade, and it is a platform feature.
Social trading data is a data category. The buyer is a company rather than an individual investor, and the deliverable is a record you can query rather than a button you can press. Confusing the two sends you to the wrong vendor.
Most arguments in this category go wrong because they treat every public signal as the same kind of object. There are seven distinct rungs between an asset being mentioned and a trade being fully accounted for, and each supports a different claim.
Sentiment products live at rungs 1 and 2. Wallet analytics live at rung 6, for whatever share of activity happens on a chain. Social trading data lives at rungs 4 and 5, and Centaur is no exception.
That is a ceiling, and it is also a floor most vendors never reach. Rung 5 requires that somebody was already watching when the post went out, with the extraction and price capture running at that minute. It cannot be assembled afterwards from a timeline.
So the useful question to take into any vendor conversation is which rung the data actually occupies, and it decomposes into four:
A vendor at rung 4 selling rung 7 will struggle with all four.
| Data type | What it observes | What it can prove |
|---|---|---|
| Mention volume | Attention and visibility | What is being discussed, and by whom |
| Sentiment | Tone of the discussion | How a body of text is expressed |
| Mindshare | Share of total attention | Who is winning the conversation |
| Social trading data | A named participant's source-linked disclosure | What they said they were doing, when, and at what stated price |
| Wallet tracking | Blockchain activity attributed to an address | What that address did on-chain |
| Broker or exchange records | Executed orders and fills | What was actually traded in the account |
The structural difference sits in the middle column. Mention volume, sentiment and mindshare are aggregates computed over a body of text, so a source can only ever be attached to the corpus. A trade event is a single record, so the source attaches to the row, and it can be audited a line at a time.
Treating these as interchangeable is not just imprecise, it is measurably wrong. A 2024 study in the Journal of Financial Economics examined Twitter, StockTwits and Seeking Alpha together and found that attention and sentiment are different signals: attention is highly correlated across platforms while sentiment is not, and the two predict next-day returns in opposite directions.
Work on explicit recommendations has been more productive still. A 2025 paper in Electronic Markets used language models to extract stated buy and sell calls from X, Reddit, Stocktwits and Telegram, on the reasoning that a stated recommendation is a more precise object than a tone score. That is the same distinction as rung 3 against rung 2, reached independently.
Popularity, meanwhile, is a poor proxy for skill. A 2024 study in the Review of Accounting Studies examined roughly 36,000 tweets from 180 prominent crypto influencers across more than 1,600 assets, and found initial positive returns followed by significant negative longer-horizon returns, with the effect strongest among those proclaiming expertise and carrying the largest followings. Follower count measures distribution. It should never be used as a performance ranking.
In Centaur's own August data, the single highest seven-day return came from a post that was not a call. The account had written that they had bought a few days earlier and were already up around 50% from their entry.
Scored from the moment of that post, the position returned over 200%. Scored honestly, it does not belong in a performance ranking at all, because the trade was open before any reader could have acted on it.
About 3% of publicly posted opens in that window describe a position the trader already held. In aggregate it barely registers: filter them out and the cohort win rate moves from 54.6% to 54.8%.
But the flag concentrates in the tail, which is exactly where best-of content looks. A leaderboard built without it puts a non-call in first place.
That is the practical content of this category. It is not text that has been scraped and labelled. It is text classified for whether the action was stated or inferred, whether the position was being opened or merely described, and whether the record traces to a message the trader actually wrote. Skip that work and the output is confident and wrong.
One property of this data cannot be recovered after the fact. Posts get deleted, and the ones that disappear are not a random sample. Researchers studying trading discussion on X have documented the removal of posts carrying poor predictions, along with deleting and reposting to keep a timeline clean.
A dataset assembled from history is therefore a survivor's record, and it flatters everyone in it. One captured at the moment of posting, and scored against the price at that minute, is not. This is the argument behind the verification gap in public market commentary, and it is the one component of building this data that no budget can reconstruct retroactively.
It is also an open question in the research. No study we could find measures how much deletion inflates the apparent accuracy of public financial predictions, only that it happens. Answering it needs a record captured live and preserved, which is what Centaur has been accumulating.
Trading platforms use it as an activation layer, because a new user staring at an empty chart with no idea what to trade is a retention problem first (the case in detail).
Agent builders use it as context an agent cannot otherwise get: prices, news and on-chain flows are commodity inputs available to every agent in the market, while a per-trader record with the post attached lets an agent cite a source instead of asserting a conclusion.
Funds and research desks use it as a dataset to be tested rather than trusted, and value it for falsifiability rather than for any signal claim.
Publishing the limits is part of defining the category honestly, and it is worth noticing which vendors do not.
It does not prove execution. A public statement is a disclosure, not a fill, and any vendor presenting posts as proof that capital moved is describing rung 6 while holding rung 4.It does not carry position size. Direction, entry price and timing exist. Notional, weight and allocation do not, and cannot be inferred from a post.
It is not a complete record of anyone's trading. It covers what a trader chose to publish, on the venues being monitored. Private positions, hedges, other accounts and unannounced exits sit outside it.
And past returns do not select future performers. Centaur ran that test on its own data: rank traders on the first half of a ninety-day window, and the rank correlation with the second half is 0.074 across a hundred traders.
Selecting on win rate gives the same answer. The value of the record is that it lets a team run that test themselves, on their own cohort and horizon, rather than trusting a ranking they cannot audit.
If the job is measuring attention, buy attention data. If the job is tone, buy sentiment. If the job is a per-trader event you can attribute, audit and join to the rest of your stack, take the four questions above to whoever is selling it.
Centaur's answers are in the docs, with the endpoints walked through in What's Actually in the Centaur API.
