1-second aggregates: columns and preview

Conventions

Time: ts is timestamp[ns, UTC] and marks the start of the bar: bars are left-closed, [ts, ts + 1 s).
The clock: Bars are stamped on the collector's receive clock, not on the venue's. It is the only clock on every message of every venue and it is the same physical clock for all of them, so joining two venues on ts is exact; and it is when the information was actually available to a participant. event_lag_ms returns what this costs: the median of receive time minus venue time over the bar. Subtract it for venue time. A large value means the messages arrived late in a burst.
Sizes: Every column ending in _size, _size_NN, _depth_* or quote_volume is in the quote currency (USDT, USD, USDC), never in the base asset and never in contracts. Base size is recoverable: divide by the price in the same row. volume and taker_buy_volume are in the base asset, as the names usually mean.
Complete calendar: Every second of the day is present: a second in which nothing happened is a row with n_messages = 0. The exception is Polymarket, which has rows only while a market is quoted.
Carried forward, and not: A silent bar keeps the last quote and the last book — the quote is still the standing quote — and says how stale it is in quote_age_ms, book_age_ms and n_messages; filter on those. book_valid is not that filter: it says the book was provably in sync when last updated, not that it was updated recently. Trade columns are never carried: no trade is trades = 0 and a null price.
Nulls: A missing value is a parquet null. There are no NaNs standing in for nulls and no zeros standing in for “unknown”.
Prices are float64: float32 would round at the scale the data measures.
Provenance: The footer carries schema_version, source_file, source_sha256, producer, bucket_rule and clock. The generation time is deliberately absent, so rebuilding from the same input gives the same bytes.

Polymarket

polymarket.<coin>.<market>_<day>.v4.1s.parquet · table pm_1s

One row per second per outcome token, both the Up and the Down token, while the market is quoted — so the table is far shorter than a day (about 270 thousand rows for a 5-minute coin). Prices are probabilities between 0 and 1. Keyed and sorted by (ts, condition_id, asset_id).

Preview

Loading the preview…

Columns (82)

Market

Column	Type	Meaning
ts	timestamp[ns, UTC]	Start of the bar, on the collector's receive clock.
condition_id	string	Market id.
asset_id	string	The outcome token: a decimal uint256, it is a string and has to stay one.
outcome	string	What this token pays on: `Up` or `Down`.
coin	string	The coin, e.g. `btc`.
interval_min	int16	Market length in minutes, `5` or `15`.
market_start_ts	timestamp[ns, UTC]	Start of the market's own window.
market_end_ts	timestamp[ns, UTC]	End of the market's own window.
winning_outcome	string	The settled outcome, the same on every row of a market. A label from the future — never a feature.
tick_size	double	Current price step; it changes during a market's life (`0.01` to `0.001`).
quote_currency	string	Always `USDC`.

Quote

Column	Type	Meaning
bid_price	double	Best bid at the end of the bar: a probability, 0 to 1.
bid_size	double	Notional at the best bid, USDC.
ask_price	double	Best ask.
ask_size	double	Notional at the best ask.
mid_open	double	Mid at the start of the second.
mid_high	double	Highest mid.
mid_low	double	Lowest mid.
mid_close	double	Mid at the end of the second.
mid_twap	double	Time-weighted mid.
spread_twap	double	Time-weighted spread in probability points — basis points mean nothing near a price of zero.
bbo_count	int32	Best-quote updates in the bar.
quote_age_ms	int32	Age of the quote at the end of the bar.

Executions

Column	Type	Meaning
open	double	First execution price in the bar; null when nothing traded.
high	double	Highest execution price.
low	double	Lowest execution price.
close	double	Last execution price.
vwap	double	Volume-weighted average execution price.
volume	double	Shares traded.
quote_volume	double	Notional traded, USDC.
taker_buy_volume	double	Shares bought by takers.
taker_buy_quote_volume	double	The same, USDC.
trades	int32	Number of executions.

Order book, at the end of the bar

Column	Type	Meaning
bid_price_NN / bid_size_NN	double	The ten best bid levels, best first (`00` … `09`): price and notional.
ask_price_NN / ask_size_NN	double	The ten best ask levels.
bid_size_total	double	The whole bid ladder, summed. The book of a market is small and known entirely.
ask_size_total	double	The whole ask ladder, summed.
bid_levels	int32	Number of price levels on the bid side.
ask_levels	int32	Number of price levels on the ask side.

Quality

Column	Type	Meaning
n_messages	int32	Messages for this token in the bar.
book_updates	int32	Changes to resting orders — not trades: it counts orders placed and pulled without trading.
event_lag_ms	int32	Median of receive time minus the venue's time.
book_age_ms	int32	Time since the last book update at the end of the bar.
book_valid	bool	The replayed book was in sync.

Left out on purpose

Market announcements and the catalogue — new_market covers other series' markets and the catalogue is one shared dump; both are in the raw files, which are sold separately.

Notes

The book is trimmed to the venue's own stated top: a fill does not always delete a level, so a book built from level updates alone would keep bids above the real best bid until the next full snapshot. Without the trim the replayed top disagreed with the venue on 0.18% of updates; with it, on 2 in 3.7 million.

Binance

binance.<coin>_<day>.v4.1s.parquet · table cex_1s

One row per second per instrument — spot and perpetual as separate rows, a complete calendar of 86,400 seconds each, so 172,800 rows a day. Keyed and sorted by (ts, symbol, instrument_class).

Preview

Loading the preview…

Columns (83)

Identity

Column	Type	Meaning
ts	timestamp[ns, UTC]	Start of the bar, on the collector's receive clock.
exchange	string	`binance`, `bybit`, `okx` or `coinbase`.
symbol	string	The venue's own name: `BTCUSDT`, `BTC-USDT-SWAP`, `BTC-USD`.
instrument_class	string	`spot` or `perp`. Binance and Bybit call both their spot pair and their perpetual `BTCUSDT`, so `symbol` alone is not a key.
base	string	The base asset, e.g. `BTC`.
quote_currency	string	`USDT` or `USD`. Treated as the same 1.0; no conversion is applied.
contract_size	double	Base asset per contract: `1.0`, except on OKX swaps (see the OKX notes).
size_unit	string	Always `quote_notional`: every `_size` column is in the quote currency.

Quote

Column	Type	Meaning
bid_price	double	Best bid at the end of the bar.
bid_size	double	Notional standing at the best bid, quote currency.
ask_price	double	Best ask at the end of the bar.
ask_size	double	Notional standing at the best ask.
mid_open	double	Mid price at the start of the second.
mid_high	double	Highest mid within the second.
mid_low	double	Lowest mid within the second.
mid_close	double	Mid price at the end of the second.
mid_twap	double	Time-weighted mid over the second: keeps the shape of the second that the last value loses.
spread_twap_bps	double	Time-weighted spread, basis points.
bbo_count	int32	Best-quote updates in the bar.
quote_age_ms	int32	Age of the quote at the end of the bar: tells a quiet market from a dead feed.

Trades

Column	Type	Meaning
open	double	First trade price in the bar. Null when nothing traded — the previous close is not carried forward.
high	double	Highest trade price.
low	double	Lowest trade price.
close	double	Last trade price.
vwap	double	Volume-weighted average trade price.
volume	double	Traded volume in the base asset.
quote_volume	double	Traded volume in the quote currency. Not derivable from `volume` exactly, so both are kept.
taker_buy_volume	double	The part of `volume` where the taker bought, base asset.
taker_buy_quote_volume	double	The same, quote currency.
trades	int32	Number of trades. `0` when none.
taker_side_inferred	bool	`true` on Coinbase only: the taker side is derived there, not given by the protocol.

Order book, at the end of the bar

The book is replayed from the raw messages. Depth is computed on the whole replayed book, not on the ten levels.

Column	Type	Meaning
bid_price_NN / bid_size_NN	double	The ten best bid levels, best first (`00` … `09`): price, and notional at the level in the quote currency.
ask_price_NN / ask_size_NN	double	The ten best ask levels, the same.
bid_depth_1bps	double	Notional resting within 1 basis point of the mid, bid side.
bid_depth_2bps	double	Within 2 basis points.
bid_depth_5bps	double	Within 5 basis points.
ask_depth_1bps	double	Within 1 basis point of the mid, ask side.
ask_depth_2bps	double	Within 2 basis points.
ask_depth_5bps	double	Within 5 basis points.
book_span_bps	double	How far the maintained book reaches, the narrower side. Read every depth column against it: a venue that keeps few levels cannot have depth beyond its reach, and that is a limit of the subscription, not a thin market.

Quality

Column	Type	Meaning
n_messages	int32	Messages of any kind for this instrument in the bar. `0` means silence.
event_lag_ms	int32	Median of receive time minus the venue's time over the bar's messages; null if none carried a venue clock.
sequence_gaps	int32	Sequence discontinuities detected in the bar.
book_valid	bool	The replayed book was provably in sync when last updated. Read it with `book_age_ms`.
book_age_ms	int32	Time since the last depth update, at the end of the bar.

Notes

Binance publishes no book snapshot on its websocket. Where the capture has the REST snapshot the book is exact; where it does not, the book columns are null with book_valid = false rather than filled with a guess.

Bybit

bybit.<coin>_<day>.v4.1s.parquet · table cex_1s

The same table as Binance: spot and perpetual as separate rows, 172,800 rows a day.

Preview

Loading the preview…

Columns (83)

The same table as Binance — 83 columns, listed there. What differs is in the notes below.

OKX

okx.<coin>_<day>.v4.1s.parquet · table cex_1s

The same table: spot and swap as separate rows, 172,800 rows a day.

Preview

Loading the preview…

Columns (83)

The same table as Binance — 83 columns, listed there. What differs is in the notes below.

Notes

OKX swaps quote in contracts and contract_size records the conversion: 0.01 BTC, 0.1 ETH, 1 SOL, 100 XRP, 1000 DOGE per contract. Every size here is already converted to notional.
OKX keeps 400 levels, which on the BTC swap span roughly 8 basis points: *_depth_5bps is complete, anything wider would be censored. book_span_bps states the reach of every bar.

Coinbase

coinbase.<coin>_<day>.v4.1s.parquet · table cex_1s

The same table, spot only: 86,400 rows a day. There is no BNB file.

Preview

Loading the preview…

Columns (83)

The same table as Binance — 83 columns, listed there. What differs is in the notes below.

Notes

Coinbase does not publish the taker side. Its trade side behaves like the maker's, so taker_buy_volume is derived and every row carries taker_side_inferred = true.
Coinbase sends a book snapshot on connect and on reconnect, not on a schedule. Before the first one in a capture the book columns are null with book_valid = false.

Chainlink

chainlink.<coin>_<day>.v4.1s.parquet · table oracle_1s

The Chainlink feed that settles the markets and its two smoothed views: one row per second, a complete calendar of 86,400 rows. There is no order book and no tape. Keyed and sorted by (ts, symbol).

Preview

Loading the preview…

Columns (10)

Oracle

Column	Type	Meaning
ts	timestamp[ns, UTC]	The oracle round, not the arrival time. A round is published about 1.7 s after the second it is stamped with.
symbol	string	`BTC/USD`.
price	double	The unsmoothed oracle price.
twap_30s	double	30-second rolling mean, published on the same feed. The 5-minute markets settled on it before 2026-08-14.
twap_60s	double	60-second rolling mean: what the markets settle on now.
price_e18	decimal(38, 18)	`price` exactly as published on chain, unrounded.
twap_30s_e18	decimal(38, 18)	`twap_30s`, unrounded.
twap_60s_e18	decimal(38, 18)	`twap_60s`, unrounded. Use the `_e18` columns wherever the value settles money.
event_lag_ms	int32	Median of receive time minus the round's time: about 1,700 ms.
n_topics	int8	How many of the three series arrived for that round; `3` is complete.

Left out on purpose

The exchange-price series (crypto_prices) — The exchanges' own spot prices are in their own files.

Notes

ts is the round, not the arrival. Joining this table to an exchange table on ts and acting on the result uses information about 1.7 seconds before it existed: shift by event_lag_ms for a point-in-time join.

Limits worth knowing

Feed outages are real and marked, not hidden — Sockets drop. When one does, n_messages falls to zero, sequence_gaps counts the discontinuity where the venue numbers its updates, and book_valid goes false until the book can be proven again.
USD and USDC are treated as 1.0 USDT — No conversion between quote currencies is applied; quote_currency says which one a row is in.