Dataset guide

What is in the archive?

How the BTC Up/Down order-book dataset is collected, organized, joined, and split between 5-minute and 15-minute markets.

Collection pipeline

CLOB WebSocket → local order book → event-driven JSONL → hourly Parquet → daily ZIP

Each market window is represented by event_slug. Both intervals share the same model; duration changes from 300 to 900 seconds.

Data layers

Raw JSONL

Closest to collection time; useful for audits.

Hourly Parquet

Columnar files for filtering and joins.

Daily ZIP

Convenient distribution unit for one day.

Joining records

Use event_slug + ts as the primary time-context pair. Do not join rows only by wall-clock time.

Timing: ticks are event-driven best-quote changes, not fixed-frequency samples.

5m versus 15m

Dimension5-minute15-minute
Window300 seconds900 seconds
FocusShort-horizon quote responseLonger intrawindow liquidity
SchemaSame fields and order-book representation

Download

Daily ZIPs, hourly Parquet, and discovery records are posted in the Telegram archive ↗.