The evidence layer · documentation
Polymarket shows you the order book as it is now. It publishes no order-book history (a price-history endpoint exists; depth history does not) — when a market moves, Polymarket's own record of that moment is gone. We have been recording the full depth of every book change in the covered markets since January 2026, and the trade tape (fills of 5+ shares, platform-wide) since December 2025. This page documents exactly what we record — and what we don't.
This is the raw layer behind every TickTape verdict — for quants, researchers and agent builders who want to check our work.
The raw files are not for sale. This data is consumed as verdicts: four live per-call tools on Pricing — bought by agents via x402, no cards, no subscriptions. Humans can read. Agents can buy.
Known gaps: the trade tape lost 5 days in March 2026 (2026-03-06 to 2026-03-10, inclusive) to a disk incident; sports books are off-season since Jun 15 and resume in autumn. Machine-readable summary of this timeline: /coverage.json.
Gaps → verdict impact
A gap in the tape only matters if it touches what a verdict is computed from. Here is exactly how the gaps above propagate:
Wallet profiles are computed over a rolling 90-day window, recomputed nightly. The 5 days lost in March 2026 (2026-03-06 to 2026-03-10, inclusive) fell out of that window in early June — current profiles are computed entirely from gap-free tape. Historical studies that span March must account for it; the manifest marks it.
A wallet with fewer than 10 first entries in the current window returns an explicit NO_COVERAGE profile — never a verdict extrapolated from too little tape.
A per-profile coverage_pct (the share of a wallet's window actually covered by tape) is not currently computed. It is on the roadmap; until it ships, coverage is documented at the dataset level on this page and in /coverage.json, not per wallet.
One of ~600,000 book records per day
2026-07-10, 14:16:00 UTC — sixty seconds into a BTC 15-minute window. Left: the book as a trader saw it. Right: the raw record we store. The archive holds one of these for every book change of every covered window.
| Ask price | Size (shares) |
|---|---|
| spread 1c · mid 0.545 | |
| Bid price | Size (shares) |
{
"ts": 1783692960930,
"crypto": "btc",
"side": "up",
"window_slug": "btc-updown-15m-1783692900",
"time_remaining": 840,
"event_type": "book",
"bids": [
{"price": "0.54", "size": "0.65"},
{"price": "0.53", "size": "584.73"},
{"price": "0.52", "size": "614.54"},
... all levels, every update
],
"asks": [
{"price": "0.55", "size": "1278.27"},
{"price": "0.56", "size": "1981.64"},
...
]
}
NDJSON, hourly files, gzip. Loads with two lines of pandas.
Method — the daily rate is counted directly from the archive, not estimated. On 2026-07-14 (UTC, a full day): 299,465 crypto-window book events + 299,425 top-200 book events + 0 sports book events (off-season) = 598,890 order-book records. The trade tape added a further 4,193,397 fill records the same day. Sports books add to the total when the leagues resume. Counted 2026-07-15.
Proof of usefulness
We replayed every sub-$1.00 moment across 2,664 windows. Median lifetime: 1 millisecond. Zero lasted a second. One study, question closed — before risking a dollar.
Method — 2,664 BTC/ETH/SOL/XRP 15-minute up/down windows across seven full days sampled from Jan–Jul 2026, replayed from this book archive. Episode = best ask UP + best ask DOWN < $1.00 with explicit book sorting and at least 10 shares of executable depth on both legs. 43,297 episodes; median lifetime 0.001 s, longest 0.435 s; zero lasted ≥ 1 s. Computed 2026-07-12.
The same math on slower markets persists for a median of 65 seconds. Knowing which market types leave time to act is worth more than any signal.
Method — same sub-$1.00 complete-set condition, measured on our own live complete-set logs in slower markets (tennis, esports, MLB, multi-day events), not a replay. Episode count and exact observation period were not recorded in the original study — treat 65 s as an order of magnitude, not a precision estimate. Computed 2026-07-12.
Round-trip taker fees at mid-range prices exceed a 2-cent scalp target. Backtest against real books and 2026 fee curves before you build.
Method — Polymarket's 2026 taker-fee formula, fee per share = rate × p × (1−p), applied to both legs of a round trip (taker in, taker out); fees are per-share, so trade size does not change the per-share result. The exact entry/exit prices and category rate behind the 2.44c figure were not recorded in the original computation. The conclusion itself is arithmetic: at 50c prices the round trip costs rate/2 per share, so any taker rate above 4% exceeds a 2-cent target. Computed 2026-07-12.
The honest map
We sell judgment, not CSVs — so here is the raw-data market, competitors included. If a free source covers your need, use it.
| Source | What you get | Where | Our take |
|---|---|---|---|
| PMXT | Free tick-level order-book event streams for every live Polymarket asset — hourly Parquet (CC-BY 4.0), archive starts 2026-04-13 — plus a unified API across prediction-market venues, including execution | archive.pmxt.dev | If you only need raw books from April 2026 onward, start here. It's free and it's good. (Checked 2026-07-15.) |
| polymarketdata.co | 1-minute L2 snapshots, all markets, since Aug 2025 — $60–360/mo | polymarketdata.co | Broadest coverage at snapshot resolution. |
| MarketLens | Tick-level order-book history + backtesting SDK, free tier — $39–199/mo | marketlens.trade | Public coverage list (crypto, sports, weather, macro); collection start date not published. |
| Telonex | Tick-by-tick book depth on every change, trades, quotes, plus on-chain fills (wallet-level activity indexed from Polygon/BSC logs, on-chain history since Nov 2022) — Parquet, $99–199/mo | telonex.io | Serious off-chain + on-chain coverage. Off-chain collection start date not published on site. (Checked 2026-07-15.) |
| TickTape | Full-depth books joined with the identity-tagged trade tape — event-level sports books (Jan–Jun 2026), up/down ticks from Jan 2026 | not sold raw — consumed as verdicts | Others sell historical data. TickTape maintains the decision layer on top: wallet-attributed simulation with fees, slippage, out-of-sample gating and explicit verdicts — plus the pseudonym-tagged tape that links wallets to Polymarket usernames. |
PMXT and Telonex rows re-verified 2026-07-15 against their public sites; polymarketdata.co and MarketLens rows as checked 2026-07-13 (linked above). On-chain wallet addresses are public blockchain data and available from several vendors; what is unique to TickTape is the pseudonym-tagged tape and the verdict pipeline computed on top of it. Spot an error? Tell us at hello@ticktape.cc and we'll fix it — this table only works if it's fair.
From tape to verdict
The archive is not for sale — it is the computation substrate for the four paid tools. What you can buy is the judgment computed from it, per call, via x402.
The identity-tagged tape documented on this page is what the live preflight_copy verdict is computed from — $0.10 per call (or 5 credits) on the Agent API. It uses a fixed $5-entry / +2c-slippage simulation computed from TickTape's private trade tape; exact size-and-delay replay is not live. Wallets not yet covered return QUEUED, which is always free — typically ready in minutes, worst case about an hour. preflight_trade reads the live CLOB book at request time — the archive backs the copy and leaderboard products, research, and calibration of the simulation parameters. All four tools and their prices live on Pricing.
Test any tool free with ?sandbox=1 — a deterministic example response, unlimited, never charged. Machine docs: /llms.txt · /openapi.json · machine manifests: /agent.json · /coverage.json.
Published research built on findings from this page: encouraged, with attribution. For how this recording compares to the rest of the market — including the free option — see the comparison table above.
Four paid HTTP tools, USDC over x402 — no account, no key, no seat, no checkout. Point your agent at the docs and let it pay per call. Questions about the data or spot an error on this page: hello@ticktape.cc (support only — there is nothing to buy by e-mail).
Read the agent docs