01 / DATASET

A documented archive of BTC Up/Down market microstructure.

The dataset records CLOB quotes and order-book depth for Polymarket Bitcoin Up/Down markets with 5-minute and 15-minute windows. Understand the files before downloading them.

Collection pipeline

The collector receives CLOB WebSocket updates, maintains a local order book, and writes several layers for inspection and analysis.

CLOB WebSocket → local book → JSONL → Parquet → daily ZIP

The two market windows share the same data model. Use interval_min and the market timestamps to identify the window; do not rely on filenames alone.

What differs between 5m and 15m

Same data model

Both windows use ticks, depth, market_outcomes, and seen_markets joined by event_slug.

Different window length

A 5-minute market normally spans 300 seconds and a 15-minute market 900 seconds. Use interval_min, window_start_ts, and window_end_ts in the data.

File layers

LayerTypical pathPurpose
Live JSONLdata/live/*.jsonlAppend-oriented collection output for audit and incremental processing.
Hourly Parquetdata/hourly/*.parquetColumnar files for faster research queries.
Daily ZIPdata/daily/btc_*_YYYYMMDD.zipDaily distribution package containing the day's files.
Market mapseen_markets.jsonlMaps event_slug to market_id, condition_id, and UP/DOWN asset IDs.
Outcomesmarket_outcomes.parquetPost-event winner labels for evaluation and backtest scoring.

How files join

The primary research key is event_slug + ts. Join quotes and depth to market metadata and resolved outcomes with event_slug.

Asset IDs live in seen_markets rather than being repeated in every ticks and depth row. Preserve that join when building your own tables.

Open the schema reference →

What the archive does not claim

Quotes are not trades

This is an order-book record, not a complete trade-by-trade tape. It cannot prove that displayed liquidity was executed.

Depth is not a fill promise

A displayed size can change, disappear, or be ahead of your simulated order in the queue. Fill assumptions belong in the backtest model.