It's dangerous to go alone. Take this.

One hour of this archive is 1.26 GiB and about 119.3 million rows. Excel will not open it. Google Sheets will not open it. That is not a reason to leave. It is a reason to bring something that can.

You do not have to download anything. Parquet files are columnar, which means a query can reach into one over the network and pull back only the columns it needs. Every example below runs against the file where it sits, on an ordinary laptop, in seconds. The rest of the hour never moves.

And you do not have to know SQL. Each step gives you a sentence to paste into ChatGPT, Claude, Gemini or whatever you use, and then the query it should hand back. Read the two side by side for a few minutes and you will start recognising the shape. That is most of the skill.

Getting set up. Four steps, about five minutes.

Nothing here needs an account, and nothing you type can break your computer.

winget install DuckDB.cli

brew install duckdb

duckdb

SET TimeZone = 'UTC';

That is the whole setup. From here on, every box below works the same way: hit Copy, paste it into that same window, press Enter. Each query ends with a semicolon, which is how DuckDB knows you have finished typing, and each box is one paste.

Ctrl+V may do nothing. The old Windows console does not use it: right-click pastes instead. Windows Terminal takes Ctrl+Shift+V. On a Mac it is Command+V as usual. This catches almost everybody once.

Five things to try

All five use the same single hour, 2026-08-28T16, which the audit records as complete: sixty minutes out of sixty, with capture reaching both edges.

SELECT event_type, count(*) AS rows
FROM 'https://archive.pendulumflow.com/v3/2026-08-28/16/2026-08-28T16.parquet'
GROUP BY 1 ORDER BY rows DESC;

SELECT question, slug
FROM 'https://archive.pendulumflow.com/v3/2026-08-28/16/2026-08-28T16.parquet'
WHERE event_type = 'new_market' AND question IS NOT NULL
GROUP BY ALL
LIMIT 20;

SELECT strftime(timestamp_received, '%H:%M:%S') AS at,
       round(price, 3) AS price,
       round(size, 2)  AS size,
       side,
       '0x' || lower(hex(market)) AS condition_id
FROM 'https://archive.pendulumflow.com/v3/2026-08-28/16/2026-08-28T16.parquet'
WHERE event_type = 'last_trade_price'
ORDER BY size DESC
LIMIT 10;

SELECT date_trunc('minute', timestamp_received) AS minute,
       round(avg(best_bid), 3) AS bid,
       round(avg(best_ask), 3) AS ask
FROM 'https://archive.pendulumflow.com/v3/2026-08-28/16/2026-08-28T16.parquet'
WHERE event_type = 'best_bid_ask'
  AND market = (SELECT market FROM 'https://archive.pendulumflow.com/v3/2026-08-28/16/2026-08-28T16.parquet'
                WHERE slug = 'ethereum-above-3000-on-september-4-2026' LIMIT 1)
GROUP BY 1 ORDER BY 1;

CREATE TABLE names AS
SELECT DISTINCT '0x' || lower(hex(market)) AS condition_id, question, slug
FROM read_parquet(list_transform(range(0, 24), lambda h:
       printf('https://archive.pendulumflow.com/v3/2026-08-28/%02d/2026-08-28T%02d.parquet', h, h)))
WHERE event_type = 'new_market' AND question IS NOT NULL;
SELECT count(*) AS named FROM names;

SELECT n.question, round(t.size, 2) AS size, round(t.price, 3) AS price
FROM 'https://archive.pendulumflow.com/v3/2026-08-28/16/2026-08-28T16.parquet' t
JOIN names n ON n.condition_id = '0x' || lower(hex(t.market))
WHERE t.event_type = 'last_trade_price'
ORDER BY t.size DESC
LIMIT 10;

When you get stuck

You will, and it is almost always one of four things. Hand your assistant this page along with the error and it will usually spot it at once.

• "syntax error at or near" with nothing after the "at or near". An empty command reached the prompt, usually a stray blank line in a paste. The statement above it almost certainly ran; try the next box and see.
• Ctrl+V does nothing. Right-click pastes in the old Windows console; Windows Terminal uses Ctrl+Shift+V.
• duckdb is not recognised, right after you installed it. On Windows the installer changes your PATH, and a terminal only reads that when it starts. Close the window and open a new one.
• Timestamps are out by a few hours. You skipped SET TimeZone = 'UTC'.
• A market id prints as a list of numbers. It is raw bytes. Wrap it in lower(hex(...)).
• Everything is null. Columns belong to event types. bids and asks exist only on book rows; price and size only on trades and price changes. Filter on event_type first.
• It is slow. Ask for fewer columns. A query naming three columns downloads three columns; SELECT * downloads all twenty-five.

What you are actually holding

Not a summary, and not a daily close. This is the order book itself, every quote and every trade, recorded as it happened by several machines at once and merged. You do not have to take our word for any of it: the audit publishes what we measured, including where we fall short, and the format notes describe every column in every era. The older V1 and V2 eras are third-party mirrors and are thinner, and the difference is measured rather than asserted.

It is free and it is CC BY 4.0. If you build something with it, we would love to see it.