Polymarket Intraday Data: 250ms Books, Minute Bars, and Hourly Prices Per Day

• At 250ms, a single market day holds 86,400 seconds x 4 frames = 345,600 book rows.
• At 1m, the same day is 1,440 price rows and 1,440 book rows — the dashboard cadence.
• At 5m/1h, the day is 288 / 24 rows — the study cadence.
• At 15m/4h, the day is 96 / 6 rows — the digest cadence for long-lived sports markets.

Crypto Up/Down lives at 250ms and 1m: each 5-minute contract is interned between a low and high bound, so the interesting record is the sub-minute frame.

Sports lives on the market clock: minute and hour resolution covers pre-match and in-play, and the 15m/4h digest cadence covers the series arc.

Every view shares the same timestamp grid and the same field names — switching granularity is a parameter, not a migration.

• A week of one Up/Down market at 250ms ~ 2.4M rows; at 1m ~ 10k rows — the storage split is 240-to-1 for the same truth.
• Full-depth snapshots at 250ms are richer than top-of-book: the archive's depth types capture book-wide frames, not just the top.
• Your pipeline should keep 250ms raw and derive minute bars; deriving 250ms back from 1m is impossible, deriving forward is trivial.

• Pull one market, one day, at 250ms: you should get 345,600 book rows; if you get 86,400, the resolution was misread as 1s.
• Derive the 1m bars from the same fetch and store them as a sibling file; keep the resolution in the file name so both readers stay unambiguous.
• Filter the 250ms rows into the final minute of each resolved contract and compare with the settlement flags; this is the fastest reproducible settlement drill.
• Run the same day across 250ms and 15m and render both mid charts; the fine one shows the reprice-seconds the coarse one hides.

Intraday data is where Polymarket behavior is actually visible: the same second that carries the cross-venue reference move (14:42:50 UTC in the recorded ETH study) is a single 250ms row, not a candle.

The granularity ladder is the honest framing: 250ms for behavior, 1m for trading, 5m/1h for structure, 15m/4h for digest — and every level is the same schema filters down to.

Never try to reconstruct fine from coarse; the archive sells the fine record precisely so you do not have to.

Your pipeline should keep 250ms raw in Parquet and derive all coarser views from it; the derive is one aggregate function and the raw never needs re-buying.

A one-week Up/Down study at 250ms is ~2.4M rows per market — DuckDB territory, comfortably laptop-queryable.

The intraday record is the archive's home: 345,600 book rows per market-day at 250ms, with every coarser cadence derived from the same frames so your 1m, 5m, 1h, 15m, and 4h views are never separate truths.

The honest habit — derive coarse, keep fine raw — costs one aggregate function and saves the impossibility of reconstructing seconds later. The 250ms raw is the one thing you cannot re-buy after the fact.

When you describe a market's day, choose the granularity that matches the claim: behavior anecdotes belong at 250ms, dashboard states at 1m, structure at the coarser steps. Mixed granularity is how wrong numbers creep into right pages.

A week of one market at 250ms is ~2.4M rows — eminently laptop-queryable in DuckDB, which makes the full intraday record a single-asset study you can genuinely own.

The decision rule for picking a cadence is one sentence: match the granularity to the fastest thing you claim to see. If your claim needs a second, you need 250ms; if it needs a day, 4h will serve.

Row math is the honest budgeting tool: a 250ms day is 345,600 rows and a week is ~2.4M — cheap in Parquet, uninteresting in Excel, and wasted on a chart. Keep fine rows in storage and coarse views on screen.

The archive's whole design is that every record is the same truth at every cadence; derive your views and you are always looking at one market, not five approximations of it.

For most teams this means a standing 250ms archive, a derived minute table, and a scheduled 15m/4h digest — the three layers that cover review, dashboard, and analysis without ever buying the same data twice.

• order book history data
• order book explained
• market data