Juan4561/pmq-tier1-l2-nba-moneyline

PMQ Tier1 NBA L2 Moneyline Capture

Public, periodically refreshed copy of the PMQ Tier1 Polymarket NBA moneyline market WebSocket capture. The data is designed for reproducible inspection of real order-book tape and for future development that needs to understand the exact raw-to-normalized layout without reverse engineering the pipeline.

The canonical source remains the internal PC data lake under C:\Users\sshuser\pmq\data. This Hugging Face repository is a public copy of selected non-secret Bronze/Silver evidence. It must not be treated as the mutable source of truth.

Repository Layout

.
|-- README.md
|-- dataset_manifest.json
|-- capture_runs.json
|-- capture_runs/
|   `-- <capture_run_id>/
|       |-- manifest.json
|       |-- status.json
|       |-- metrics.json
|       |-- capture_plan.json
|       |-- data_verification.json
|       `-- summary.md
`-- data/
    |-- bronze_market_ws/
    |   `-- capture_run_id=<capture_run_id>/bronze_market_ws.parquet
    |-- silver_orderbook_events/
    |   `-- capture_run_id=<capture_run_id>/silver_orderbook_events.parquet
    `-- silver_l2_orderbook_levels/
        `-- capture_run_id=<capture_run_id>/silver_l2_orderbook_levels.parquet

Every exported Parquet file is filtered by capture_run_id. Do not interpret a shared date partition as evidence for one run. Always use the explicit capture_run_id=* path or row column when auditing coverage.

Capture Unit

capture_run_id is the atomic collection window. The internal collector runs in rotating windows so the static WebSocket seed can be refreshed. A single capture_run_id should be read as one continuous capture attempt with one capture plan, one run manifest, and one set of normalized rows.

The most important metadata files are:

• capture_runs/<capture_run_id>/capture_plan.json: selected NBA moneyline markets, outcomes, token IDs, scheduling, and market-status context used to seed the stream.
• capture_runs/<capture_run_id>/data_verification.json: run-specific Bronze and Silver quality checks.
• dataset_manifest.json: publication-level index with exported row counts, time ranges, file counts, and the capture runs included in the current public copy.

Core Identifiers

• market_id: internal market identifier. In PMQ this is the Polymarket conditionId.
• condition_id: same identity as market_id where present; kept for compatibility with Polymarket naming.
• token_id: Polymarket CLOB token/asset ID for one outcome side.
• outcome: human-readable outcome/team label when known.
• event_id: Polymarket/Gamma event identifier when available.
• capture_run_id: run/window that produced the row. This is required for clean joins and run-specific audits.

Typical join keys:

• Bronze to Silver audit: capture_run_id + raw_content_hash/content_hash
• Token-level top-of-book to L2: capture_run_id + market_id + token_id
• Same raw book frame in L2: capture_run_id + token_id + raw_content_hash
• Same normalized snapshot time: market_id + token_id + snapshot_ts

Timestamp Semantics

• ingested_at: collector write/ingestion time.
• event_timestamp: raw event timestamp from the WebSocket payload when present.
• source_event_ts: normalized source timestamp from the exchange/book payload.
• snapshot_ts: normalized book observation time used by L2 rows.
• scheduled_start_ts: scheduled game start/kickoff for the market.
• seconds_to_start: scheduled_start_ts - snapshot_ts in seconds. Positive values are pre-game. Negative values are after scheduled start and should be treated as in-play/out-of-domain for pre-game-only models.
• partition_date: physical storage partition date, not a modeling feature by itself.

Bronze timestamp columns may be strings because Bronze preserves raw ingestion shape. Silver timestamp columns are normalized timestamp types where possible.

Tables

bronze_market_ws

Path:

data/bronze_market_ws/capture_run_id=<capture_run_id>/bronze_market_ws.parquet

Granularity: one row per raw market WebSocket message/frame captured by the collector.

Purpose: immutable source-of-truth audit trail. Use this table when validating message shape, replaying normalization decisions, or checking whether the Silver rows came from actual WebSocket frames.

Important columns:

• capture_run_id: source run/window.
• ingested_at: time the collector wrote the raw message.
• message_type: raw WebSocket event type, for example book or price_change.
• event_timestamp: timestamp extracted from the raw payload when available.
• content_hash: hash of the raw payload content.
• payload_json: raw payload as JSON text. Numeric blanks or source quirks are intentionally preserved here.

silver_orderbook_events

Path:

data/silver_orderbook_events/capture_run_id=<capture_run_id>/silver_orderbook_events.parquet

Granularity: lightweight token-level top-of-book observation rows derived from Market WebSocket book-like messages.

Purpose: fast quote/tape inspection where full book depth is not needed. This is the table to use for quick best-bid/best-ask coverage checks, spread checks, and simple executable quote availability.

Common columns:

• market_id, token_id, capture_run_id
• event_type, event_timestamp, ingested_at
• best_bid, best_ask
• best_bid_size, best_ask_size
• last_trade_price
• spread
• payload_json

Rows are append-only observations. They are not globally deduplicated latest state. For live-style research, select rows by bounded time windows and use as-of logic.

silver_l2_orderbook_levels

Path:

data/silver_l2_orderbook_levels/capture_run_id=<capture_run_id>/silver_l2_orderbook_levels.parquet

Granularity: one row per visible L2 order-book level, per side, per token, per snapshot/update. A single raw book frame normally expands into many silver_l2_orderbook_levels rows.

Purpose: dense microstructure tape: depth, cumulative liquidity, price ladder, spread reconstruction, and quote-quality analysis.

Important columns:

• market_id, condition_id, event_id, league_key
• token_id, outcome, capture_run_id
• snapshot_ts, source_event_ts, ingested_at
• side: bid or ask.
• price, size, notional
• level_index: zero-based rank on that side. 0 is best bid or best ask.
• cumulative_size, cumulative_notional: cumulative depth across that side from best to worse levels.
• best_bid, best_ask, last_trade_price, tick_size
• book_hash, raw_content_hash
• is_snapshot, is_delta_applied
• capture_action: e.g. snapshot-level or replayable update action.
• market_status, scheduled_start_ts, seconds_to_start

Interpretation details:

• book snapshots are expanded by side and sorted from best to worse price.
• price_change rows can represent replayable deltas. size = 0 can mean a level deletion and should not be dropped blindly.
• is_delta_applied = false means the row is an observed update/snapshot row, not a full reconstructed book after applying all prior deltas.
• This table intentionally has no future labels, winner fields, model probabilities, or PnL columns.

Suggested Reading Pattern

Download only the files you need. Example:

from huggingface_hub import snapshot_download
import duckdb

root = snapshot_download(
    repo_id="Juan4561/pmq-tier1-l2-nba-moneyline",
    repo_type="dataset",
    allow_patterns=[
        "dataset_manifest.json",
        "capture_runs.json",
        "data/silver_l2_orderbook_levels/**/*.parquet",
    ],
)

con = duckdb.connect()
l2 = con.execute(
    '''
    select
      capture_run_id,
      market_id,
      token_id,
      outcome,
      snapshot_ts,
      seconds_to_start,
      side,
      level_index,
      price,
      size,
      cumulative_size,
      best_bid,
      best_ask
    from read_parquet(?)
    where side in ('bid', 'ask')
      and level_index <= 4
      and seconds_to_start > 0
    order by snapshot_ts
    limit 100
    ''',
    [root + "/data/silver_l2_orderbook_levels/**/*.parquet"],
).fetchdf()

Data Quality Expectations

The publisher is intentionally conservative:

• Rows are exported only from local Tier1 Bronze/Silver Parquet and filtered by capture_run_id.
• Run metadata is copied with the Parquet rows so the selection and verification context travels with the data.
• Raw logs, environment variables, wallet credentials, and API secrets are not published.
• The public copy is append/refresh style. It may replace exported Parquet files for the same capture_run_id as a live capture window grows, but it does not rewrite the internal canonical raw data.

Research Use Notes

• For PMQ internal modeling, this Tier1 tape is final-test / validation evidence, not external training data. Do not tune local thresholds on this public copy and then claim clean Tier1 performance.
• For pre-game models, filter seconds_to_start > 0 and keep time-based splits.
• Do not randomly split repeated rows from the same market_id or event.
• If you build features from L2 rows, preserve market_id + token_id + snapshot_ts or decision_ts granularity. Collapsing only by market_id mixes timestamps and invalidates live-style simulations.
• Treat after-start rows (seconds_to_start < 0) as in-play data. They require a separately validated in-play model contract.

Target Repo

Target repo: Juan4561/pmq-tier1-l2-nba-moneyline.