The question
Can trades from Polymarket wallets that were skilled at the time be used as a trading signal that still makes money after everything a real copier pays: the delay, the spread, the slippage, thin order books and the fees?
Ground rules: wallets are chosen only from what was known at each date (no picking today's winners and replaying their past). Rules are fixed before the data that judges them. A rule that changes after the forward test starts becomes a new version; old versions keep their results. Nothing is tuned to make a backtest look good, and if the signal does not survive real execution the study says so — a negative result is a result. Every number is labelled with how it was priced: measured (our order walked through the real order book), observed (what other traders actually paid on the public record) or whale's price (the best case).
What we already know (priors, not conclusions)
- Copying costs a lot. Our real-price funds pay 44–53¢ more per $10 copy than the same copies at the whale's price (measured, 4–5 Oct). About 93% of that is the whale's price already being gone when the copy can look — the whale's own order took the cheap shares.
- The trades you can copy at the whale's price are the bad ones. In liquid books where the price did not move, the whales lost $1.14–$1.85 per $10 copy even at their own price (measured, one night).
- Fees are real and uneven. Polymarket charges the order that takes the price 0% (most sports) to 7% (crypto) × price × (1 − price), read from the chain.
- An earlier two-week study found no topic or time of day that held up, and only the under-20¢ rule survived — but it priced copies with average costs, so it is background, not evidence.
The data, checked live today
| Source | What it gives | How far back | Limits found | Used for |
|---|---|---|---|---|
| Polymarket activity (per wallet) | Every trade, redemption, merge and reward: price, shares, dollars, transaction id, time to the second | Back to the wallet's first trade (e.g. a followed whale from 17 Apr 2026) | 500 rows a call and at most 5,000 deep — so it is read in date windows | Wallet histories, metrics, who qualified when |
| Polymarket trades (per market) | Every trade print in a market: who, side, price, shares, time | The whole market | Newest first only: an old trade in a busy market is reached by paging through everything after it | What other traders actually paid after a whale (observed) |
| Polygon chain (exchange fill events) | Every fill with maker and taker, the fee charged and the exact order inside a block | Complete | Free public nodes return ~50 blocks (~100 s) a call; a free keyed node lifts that | Fees, maker vs taker, exact order, audits |
| Market data (Gamma) | Question, outcomes, scheduled end, result, close time, fee rule, fee type | Complete | No official category: we use the fee type (sports, politics, crypto, weather, finance, tech, culture, economics) plus our title rules | Results, categories, fees |
| Price history (CLOB) | The market price minute by minute | Verified for May 2026 markets | A quoted price, not something you could have traded at | How fast a whale's information reaches the price (+1 min to +24 h) |
| Order books (CLOB) | The buy and sell orders waiting | Now only — Polymarket keeps no history of order books | Past executable prices cannot be read | Live paper trading |
| Our recorder (VPS) | The real order book read 0, 1, 2, 5, 15 and 60 s after every signal of the 178 followed whales; 15 price levels a side | Since 4 Oct 2026 22:35 IST, ~750 signals a day | Followed whales only; up to 60 s; 15 levels may not cover $500–$1,000 | The only measured copy prices |
| Leaderboard and wallet P&L | Today's ranking and equity curves | — | Only today's survivors; curves do not start at zero and are sometimes flat copies | A cross-check, never the wallet pool |
Limits, plainly
- No historical order books exist. For any trade before 4 Oct 2026 we cannot measure what our order would have filled at. We can show the whale's price (best case), what other traders actually paid after the delay (observed — a real trade, but someone else's, only when one happened, at their size) and the minute-by-minute price. Decisions are taken only on measured fills; the rest is shown, labelled.
- Measured fills cover only the followed whales, who were picked with today's scores — a survivor list. The study adds wallets found without hindsight as a control.
- Delays past 60 s and orders past 15 book levels are not recorded yet (fix below).
- Some exits are not sales: merging both sides, conversions in linked markets, transfers. A redemption after the result is fine. Exits that cannot be seen are flagged, never invented.
- One person can run many wallets. "Independent" whales agreeing must pass a coordination check (same second, same funding source); it is a heuristic and is flagged as one.
- Times are to the second in the APIs; only the chain gives the exact order inside a block.
- Rate limits (about 10 calls a second to the order-book service, 5 to market data) make the first data pull take hours, not minutes. Raw data is saved once and never fetched twice.
How it will be built
1 · RawExactly what each source returned, never edited. Every file records its
address, parameters, time fetched and a checksum.
2 · ProcessedCleaned and joined tables: fills, markets, wallet positions, price
paths, trade prints, recorded order books.
3 · ResearchWho qualified when, the signals with their features, every simulated
copy, and each strategy's results — tagged with the data and code version.
The main tables
| Table | One row per | Key columns |
|---|---|---|
fills | wallet trade | wallet, tx, time, market, outcome token, side, price, shares, dollars, type (trade / redeem / merge / conversion) |
markets | market | condition, tokens, question, category, scheduled end, result, close time, fee rule over time, linked-market flag |
positions | wallet × outcome episode | entries, exits, exit kind (sold / redeemed / merged / unseen), resolution, P&L, held time |
wallet_asof | wallet × date | metrics using only earlier data (settled markets, ROI, win rate, drawdown, concentration, category mix, maker share…), qualified yes/no, quality score |
signals | qualified whale entry | price bucket, conviction (size ÷ wallet's median), category, wallet's edge in that category so far, other qualified whales on the same outcome within 1 min … 24 h |
copies | signal × delay × size × price source | our price (measured / observed / whale's), fill share, fees, exit, P&L, "could not be reproduced" reason |
strategies | strategy version | frozen rule text, start date, results by period |
The first experiment (MVP) — proposed
- Wallets (40): 20 from our followed list (marked as hindsight-picked) + 20 found blind — the biggest traders in a random sample of markets in June 2026, chosen before looking at what they did next — and a random-wallet control.
- Periods: develop on June–July 2026, choose rules on August, test the frozen rules once on September, and keep testing forward on the recorder from October.
- Who counts as a whale is decided each month from earlier months only.
- Copies: at the whale's price; after 30 s, 1 min, 5 min and 10 min from what traders actually paid and from the minute price; $10, $25 and $100 measured from the recorder wherever its books reach.
- Tests: the ten price buckets (is the under-20¢ effect real out of sample?), conviction, category, 2+ and 3+ whales agreeing, and how fast the price moves after a whale.
- Report: the ten questions of the brief, each answer marked measured / observed / whale's price, with bootstrap ranges and "remove the best 1, 3, 5, 10 trades / the best wallet" checks.
Not built until the experiment shows the pipeline works: the dashboard and the live paper-trading engine.
One change that makes the forward test measure what the brief asks
The recorder can also read the order book 5 and 10 minutes after each whale trade, keep 50 price levels instead of 15 (so $250–$1,000 orders can be walked), and price $10 to $1,000 orders from the same read. Measured fills at every delay and size would then build up every day. This needs one install on the VPS.
Decisions needed before the experiment runs
- Show pre-October delay results from observed trades (clearly labelled) while deciding only on measured fills?
- The wallet pool: 20 followed + 20 found blind + a random control?
- The periods: June–July / August / September 2026, then forward from October?
- The recorder upgrade (5 and 10 minutes, deeper books, $10–$1,000) on the VPS?
Paper research only. Nothing on this page places, or will place, an order.