What wallet tracking shows, and what it cannot

A dozen cards in this catalogue read Polymarket wallets, and most of them sell the reading rather than the record: smart money, informed flow, the account that was positioned before the news. The record underneath is real, public and free to check. Nearly everything printed on top of it is a model, and the distance between the two is the whole of this page.

Start with the fact that decides whether any of it applies to the venue you actually use. Wallet analytics exist because a fill on Polymarket is a transaction on a public chain. They cannot exist in that shape on a CFTC-regulated exchange, because what a regulated market publishes is market-level and what identifies an account goes to the regulator.

How it works

Why the flow is visible at all

Polymarket's exchange contract describes its own design plainly. The repository README calls it a hybrid-decentralised model in which "there is an operator that provides offchain matching services while settlement happens on-chain, non-custodially", swapping conditional-token ERC-1155 assets against an ERC-20 collateral asset.

Read that twice, because the half people assume is public is the half that is not. The order book is off-chain. Orders you placed, orders you cancelled, orders that never filled and the depth you were quoting leave no chain record at all. What settles on chain is the executed swap, plus the splits, merges and redemptions that move collateral in and out of positions. A wallet's "activity" is therefore a log of executions and never a log of intentions — which is why no dashboard here can tell you that an address tried to buy and missed.

What is indexable is what the contracts emit. Polymarket publishes the subgraph manifest for indexing "on-chain trade, volume, user, liquidity and market data", and every product in this category is reading the same events off the same chain. Differences between two dashboards are differences in aggregation, not in access — Dune dashboards and OrcaLayer see the same fills and disagree about what to call them.

And why it is not visible on a regulated exchange

A designated contract market publishes statistics, not accounts. 17 CFR 16.01 requires each reporting market to record and publish, for each business day, gross open contracts and total volume of trading by commodity and expiration, together with opening, closing and settlement prices. Nothing in it is per account.

The per-account layer exists, and it has a different destination. 17 CFR 17.00 provides that "each futures commission merchant, clearing member and foreign broker shall submit a report to the Commission for each business day" covering the special accounts it carries. To the Commission. Not to a dashboard, and not to you.

That is the whole reason Kalshi and Polymarket US have no wallet trackers. Their public surfaces are generous about markets and silent about accounts: the Kalshi API serves series, markets, order books, trades and candlesticks unauthenticated, while positions, fills and settlements sit behind your own credentials and are your own. Polymarket US serves markets, books and prices from a keyless public gateway, and gates everything about an account behind identity verification.

So when a product advertises wallet features on a regulated venue, it means one of three other things: aggregate market flow, your own authenticated account, or the Polymarket half of a cross-venue product. Which one it is is a question the product can answer in a sentence, and the cross-venue data comparison is the short version of how far the asymmetry goes.

What a "wallet" is, and how many of them one person is

Polymarket's archived v1 reference client spells the split out in its own configuration. The README — which now opens with "This repository has been archived and is no longer maintained" — documents funder as "This is your Polymarket Profile Address, where you send UDSC to" and signer as "This is your Private Key", alongside a numeric signature type where 0 is a browser wallet and 1 is Magic or email login. Two addresses, one trader, written into the client's first example.

The split survived the rewrite and got deeper. The current arrangement is a per-user deposit wallet — Polymarket signature type 3 — deployed gaslessly through a relayer that pays the gas, and a bare externally-owned account cannot trade on the v2 book at all; the Polymarket alpha bot card documents that from the other side, because it has to deploy one for you before it can trade.

The general mechanism is a standard. ERC-1271, Final since 2018, exists because "Externally Owned Accounts (EOA) can sign messages with their associated private keys, but currently contracts cannot", and it lets a contract validate signatures itself on the rule that "a signed action message is only valid when the signer is authorized to perform a given action on the behalf of a smart wallet". The acting address and the authorising key are different objects by design, the mapping between them is whatever the contract says it is, and it is nowhere published as a directory.

Four things follow, and each of them breaks a statistic:

• One person, several rows. Splitting size across addresses, or keeping strategies apart, is ordinary and costs nothing. Each address ranks separately with its own record.
• One row, several people. A fund, a shared account, or a copy-trading bot placing orders for subscribers all print as a single address. The Copy Trade tab on Polymarket Analytics is a Telegram bot that generates a wallet for you and charges 1% on every transaction, buys and sells alike — every subscriber's trade lands in one row.
• The signer is not evidence of the decider. A session key, a delegated signer or a relayer means the key that signed the order may belong to software, not to whoever chose the position.
• The population is not clean. OrcaLayer's April 2026 audit of 470 million trades reports that six in ten wallets with apparently flawless records were airdrop farmers. Its published rules classify a farmer numerically — an average buy price above 95 cents — and a bot at more than 500,000 trades averaging under 20 dollars, and it strips four Polymarket protocol router contracts platform-wide.

Why a leaderboard's profit is not a measurement of skill

The denominator is missing from the screen. OrcaLayer publishes three populations behind three pages: roughly 3.1 million wallets with at least one fill, 1.85 million clearing ten fills and 100 dollars of notional, and a narrower set with at least ten resolved markets that the smart-money rankings actually use. A top-50 table drawn from three million accounts is the right tail of a distribution that would have a right tail if every position were decided by a coin.

Sample size and concentration are separate problems, and one number hides both. A win rate counts resolved positions without weighting them, so an account can be right nine times in ten and be down; a profit figure can be one correct large position and forty small mistakes. Convexly composites three pillars into its Edge Score — a baseline-adjusted Brier score, the concentration of profit and loss, and the count of resolved positions — against a frozen 8,656-wallet reference cohort, precisely because the headline figure does not survive the other two.

The arithmetic is not shared, and nobody has to say which they used. Realised profit and an unrealised mark are different claims. Lot matching may be FIFO or something else. Multi-outcome markets need a correction that not every indexer applies. And splitting, merging or redeeming conditional tokens moves value in and out of a position with no trade to attribute it to, so a pipeline that only watches fills will misprice an address that uses them. Two dashboards reporting "P&L" for one address are answering two questions.

Completeness is a number almost nobody prints. Convexly marks every row as coming either from a full read rebuilt from the settlement record, block-anchored, or from a recent-trade window that omits positions already redeemed — and its board of 12 September 2026 had 31 of its 42 wallets read in full, with the ratio on the page rather than implied away. A census it generated on 22 August 2026 excludes a further 100 full reads that did not finish. That is what a measured figure looks like next to an unmeasured one.

The most useful evidence here is a vendor's own negative result. A July 2026 Convexly note points its skill-versus-luck filter at Convexly's own published top-50 leaderboard and reports that zero of the 35 readable wallets clear its own bar. A leaderboard and a skill claim are not the same object, and the clearest demonstration of that came from somebody who publishes both.

And the book you can see is not the book. Airavat scores wallets on one venue's public record; a trader hedging on an exchange the indexer does not cover, or running a second address, is scored on a fraction of their position and the score does not say so. Whether the price itself means what it looks like is a separate question, answered in where the liquidity comes from.

What "insider" means here, and what it does not

This is the claim that sells the category, and it is the one with the least behind it.

What a product computes is narrow and checkable. An address established a position of a stated size at a stated time, and the price later moved. Every element of that is on the chain and can be re-derived by anyone with a Polygon archive node.

What does not follow from it: who the address belongs to, whether it is one person or several, what that person knew, when they knew it, or whether any rule applied to them. None of those are in the data, and no amount of flow analysis puts them there.

Nothing in the category closes the gap. The three things that would make the label testable are the rule, the false-positive rate, and a retrospective over addresses flagged before the outcome was known. Polysights ships the claim as a screen called Insider Finder with none of the three published, and its own terms and privacy pages still read "Under Construction". Predicts.guru is more careful — its methodology notes warn that no single metric tells the full story and that some public profiles may be incomplete — and it still ranks by address, because address is all there is.

The law is narrower than the label, and it is worth reading rather than assuming. 17 CFR 180.1 prohibits manipulative devices, untrue or misleading statements of material fact, and acts operating as a fraud or deceit, in each case intentionally or recklessly. Its paragraph (b) reads, verbatim: "Nothing in this section shall be construed to require any person to disclose to another person nonpublic information that may be material to the market price, rate, or level of the commodity transaction, except as necessary to make any statement made to the other person in or in connection with the transaction not misleading in any material respect." Being better informed than the other side is not, by itself, the thing that rule reaches.

Where a genuine restriction exists, it is written down and it names categories of person. Series-level contract terms filed with the CFTC carry additional prohibitions on who may trade a particular series — employees of the source agency, holders of material non-public information — and they are readable per series, before you take a position, through the Kalshi API. A prohibition in a filing is a document you can check. A badge on a dashboard is a classifier you cannot.

The classifiers themselves show how far off the label can land. OrcaLayer removed four Polymarket protocol router contracts platform-wide after one of them surfaced through its own pipeline as an apparent insider. A router contract is not a person and knew nothing; it cleared the bar anyway, which is the most concrete published account in this catalogue of what a false positive looks like.

So the sentence the record supports is: this account established this position at this time, ahead of this move, on this venue. Anything beyond it is the reader's inference, and it is an inference about a human being — which is why this site writes the first sentence and not the second.

Latency, and the difference between a signal and a fact

Three clocks run through this category, and a screen that mixes them is not lying to you, it is showing you a pipeline.

Before confirmation. A transaction, in Ethereum's own developer documentation, "is then broadcasted to the network and added to a transaction pool consisting of all other pending network transactions", from which validators select what to include. Pending means broadcast, and broadcast means intent. Predexon sells exactly this as a channel — Polymarket fills seen in the Polygon mempool, which it describes as three to five seconds ahead of confirmation — and its own documentation says the quiet part: a mempool observation is not a trade, some never confirm, and anything built on the feed needs an explicit reconciliation path for the ones that do not.

At confirmation. The fill is in a block and the chain is the record. This is the only one of the three that is a fact.

At index. Every dashboard is behind the chain by its own amount, and the amounts are not comparable. OrcaLayer states 15 to 20 seconds as its own figure rather than as a commitment. Dune dashboards are batch-indexed and run minutes to hours behind. PolyWallet states its refresh interval three incompatible ways on a single page. And Brier.fyi still says on its about page that its downloader runs nightly, while its repository records the pipeline being paused on 23 January 2026 and nothing after the 24th — a stale chart is indistinguishable from a fresh one, and no product in this category displays its own staleness by default.

What it costs

Reading is mostly free. Reading reliably, and being allowed to do anything with what you read, is where the money is.

• Free, with no account: Predicts.guru runs the wallet checker, leaderboards, event analytics and the activity feed without registration; public Dune dashboards cost nothing to open; OrcaLayer gives five wallet lookups a day without an account and 20 with one.
• OrcaLayer: Pro at 9.99 USD a month lifts the lookup cap and adds alerts, a 25-wallet watchlist and CSV export; Premium at 19.99 USD a month is the developer tier, with REST at 600 requests a minute and a server-sent trade stream.
• Convexly meters the thing that actually costs, which is re-reading a wallet's full record on a schedule: Watchlist at 9 USD a month covers three wallets re-read every 30 days, Researcher at 49 USD covers ten on the same cadence, Trader at 499 USD covers fifty re-read every seven days. A single reconciled read of one named wallet is 19 USD once, delivered with the block it was reconciled at and the command that re-derives it.
• Predexon prices the raw material instead: Free, Dev at 49, Pro at 249 and Enterprise from 499 USD a month for request quota, with tick history billed separately per GiB of compressed Parquet downloaded — 40 USD for Polymarket, 80 USD for Kalshi.

The licence is the cost nobody budgets for. OrcaLayer's terms forbid reselling without written permission and forbid using the data to train, fine-tune or evaluate machine-learning models, commercially or otherwise. Convexly's self-serve licence is one named individual and forbids displaying, embedding or republishing output to third parties, your own clients included, without a separate agreement. Dune is internal use only below Enterprise, scheduled queries expire after three months on the Analyst plan, and the free plan cannot schedule at all. Polymarket Analytics licenses its free data for individual, non-commercial use and prohibits scraping. A research workflow that quietly becomes a product is a licence problem before it is a data problem.

And the cost of acting on a wallet is not a subscription. Polyrama's own terms say the copied wallet is an unrelated third party it does not vet, endorse or control, that past performance says nothing about the future, and that no cap or safeguard is promised to prevent loss. On top of that sit the fees on both legs of every mirrored position — what a trade actually costs — and the resolution risk of every market you inherit, which is who decides the outcome.

What you can do about it

Put four questions to any product in this category. Each has a short answer if the vendor has done the work, and an evasive one if they have not.

1. On which venue is this possible at all, and which half of your product am I looking at? If the answer covers a regulated exchange, ask whether it is aggregate market flow or my own authenticated account, because 17 CFR 17.00 sends the per-account data somewhere I am not.
2. Is the methodology published, and what counts as one wallet? Specifically: do you merge a deposit wallet with its signer, do you exclude router and protocol contracts, and how do you classify farming and bot activity? OrcaLayer answers this in numbers. Most do not answer it at all.
3. What is the false-positive rate on the label, and is there a retrospective? A flag applied to settled history is a description. A flag applied before the outcome, scored afterwards, is a claim. Ask which one I am buying.
4. Which clock is this screen on, and when did this number last move? Mempool, confirmed or indexed — and if the pipeline stopped, does anything on the page say so?

Then check the cheap things yourself, for nothing.

• Read one address on two products. Take a single Polygon address, pick the same window, and compare the profit figures on Predicts.guru and a public Dune dashboard. Where they differ, you have found the accounting assumption neither of them printed.
• Change the window and watch the table change. Twenty-four hours, thirty days and all-time produce three different top tens from the same events. If a ranking's story survives only one window, it was a property of the window.
• Look for a completeness ratio before you look at a rank. Convexly prints how many wallets on a board were read in full. Nobody else does, so on every other product the honest reading of a rank is "incomplete by an unstated amount".
• Date the product, not just the data. Open the changelog, the repository and the terms. A dashboard whose pipeline paused eight months ago still renders, still sorts and still looks current.
• On a regulated venue, read the filing instead of the flow. The contract terms for a series name who is prohibited from trading it, and they are published before the market opens. That is the only document in this whole subject that says anything checkable about who should not be in a market.

Then size the position on what the record actually supports. An address's history is evidence that an account has a record on one venue over one window. It is not evidence of a person, a skill or a source. The analytics and dashboards listing carries what each product reads and what it cannot see, which is the fastest way to find out whether the screen in front of you is reading a chain or reading a model.