The arithmetic
Sampling resolution sets a hard ceiling on what can be measured. To observe a phenomenon at all, you need samples separated by less than the phenomenon's own timescale — the same constraint that governs any sampled signal. In a market that exists for 300 seconds, a 60-second sample resolves nothing faster than two minutes, which is to say nothing at all.
| Observation window | 1-minute sampling | This archive (1.45s) |
|---|---|---|
| Full 5-minute round | 5 samples | 207 snapshots |
| Final 60 seconds before resolution | 1 sample | 41 snapshots |
| Final 10 seconds | 0 samples | ~7 snapshots |
The final row is the one that matters most. The last seconds before a binary market resolves are when spreads blow out, makers pull quotes and the book thins fastest. At 1-minute resolution that entire regime is invisible.
What 1-minute sampling rules out
Order flow imbalance
Imbalance predicts price movement over horizons of seconds. Sampled once a minute, the imbalance and the move it predicts land in the same bucket — the signal and its outcome become indistinguishable.
Spread dynamics
Measuring how the spread widens and recovers needs enough samples to see the recovery. Five points per round give a mean spread and nothing about its behaviour.
Book replenishment
How fast depth returns after a level is swept is a liquidity measure in its own right, and the whole event typically resolves inside a few seconds.
Pre-resolution liquidity withdrawal
Market makers pull quotes as a round approaches settlement. One sample in the final minute cannot tell you when the withdrawal started or how steep it was.
Market-making backtests
Fill simulation depends on queue position changing between samples. A minute of unobserved book evolution makes any fill assumption arbitrary.
Depth cannot be backfilled
Prices can be reconstructed after the fact from on-chain fills. Depth cannot. An order that rested at 0.53 for forty seconds and was cancelled without trading leaves no trace anywhere — not on chain, not in any trade feed. The only record is a snapshot taken while it was still resting.
This matters because Polymarket's own historical orderbook endpoint stopped producing new snapshots around 20 February 2026. The /book endpoint returns only the current book, and prices-history carries no depth at all. For any date after that cutoff, historical depth exists only where someone was recording live.
This archive has been recording BTC Up/Down books continuously since 24 April 2026, at 1.45-second resolution — 5.8M+ snapshots, ~50 price levels per side, both Up and Down books, 39 GB of JSONL.
Questions
What is tick-level orderbook data?
Orderbook snapshots taken fast enough that consecutive samples capture individual changes to the book rather than an aggregate of many. For Polymarket 5-minute markets that means sampling in seconds, not minutes. This archive samples every 1.45 seconds.
Why is 1-minute orderbook data not enough for 5-minute markets?
A round that lasts 5 minutes contains only 5 one-minute samples. Spread dynamics, order flow imbalance and the liquidity withdrawal that happens before resolution all occur on a scale of seconds, so a 1-minute sample averages them away entirely.
How many snapshots does each round contain?
About 207, taken roughly 1.45 seconds apart, each recording the full Up and Down book with a median of ~50 price levels per side.
Does Polymarket provide this data itself?
No. The CLOB /book endpoint returns only the current book, and prices-history returns prices without depth. The historical orderbook endpoint stopped producing new snapshots around 20 February 2026, so depth after that date exists only where someone recorded it live.
Can orderbook depth be reconstructed after the fact?
No. Trades can be replayed from on-chain fills, but resting orders that were cancelled without ever trading leave no record. Depth that was not captured at the time is permanently unavailable.