# Source Two: PMData Archive

Prepared: 2026-07-03 UTC

## Purpose

PMData is the preferred source for the new execution-quality backtesting path because it provides daily ZIP archives of Polymarket market data with sub-second L2 event rows and trade-related events. It is the right primary input for the planned YES/NO ladder adapter and fill simulator.

Current local root:

```text
/home/Data_Collection_v2/data/pmdata
```

Current downloaded series:

```text
data/pmdata/btc-5m/YYYY-MM-DD/
```

Current downloaded ranges:

```text
2026-03-16 through 2026-03-30
2026-06-01 through 2026-06-30
```

## Data Types Downloaded

### `poly_l2`

This is the main backtesting source.

Observed properties:

- Daily ZIP archive by series/date.
- One parquet file per five-minute market when present.
- Event-level rows with book updates and trade-price events.
- Millisecond timestamps and local receive timestamps.
- Columns observed in the sample include market slug, event type, timestamps, book levels, best bid/ask, trade price/size/side, mirror fields, and winning outcome fields.

Use `poly_l2` for:

- Sub-second book replay.
- YES/NO ladder reconstruction.
- Maker queue and adverse-selection simulation.
- Strict taker VWAP walks against displayed depth.
- Markout and latency tests.
- Strategy backtests where fill quality matters.

### `polygon_trades`

This is an audit/reconciliation source, not the primary L2 replay source.

Use `polygon_trades` for:

- On-chain or wallet-level reconciliation.
- Activity sanity checks.
- Cross-checking market participation and realized trade activity.
- Investigating settlement and account behavior.

Do not use `polygon_trades` alone to simulate queue-aware maker fills. It does not replace the L2 book event stream.

## Current Downloads

The March test block downloaded:

- Series: `btc-5m`
- Dates: `2026-03-16` through `2026-03-30`
- Data types: `poly_l2`, `polygon_trades`
- Local disk usage: about `6.3G`
- `poly_l2` rows: `479,568,786`
- `polygon_trades` rows: `60,718,994`
- All 30 expected ZIPs exist.
- All ZIP integrity checks pass.

Detailed reports:

```text
reports/pmdata_btc_5m_20260316_20260330.md
reports/pmdata_btc_5m_20260316_20260330_audit.json
reports/pmdata_btc_5m_20260316_20260330_download_manifest.json
```

The June full-month block downloaded:

- Series: `btc-5m`
- Dates: `2026-06-01` through `2026-06-30`
- Data types: `poly_l2`, `polygon_trades`
- Local disk usage for `data/pmdata/btc-5m`: about `22G` including March and June
- ZIP bytes: `16,828,236,641`
- `poly_l2` rows: `1,534,252,982`
- `polygon_trades` rows: `70,121,307`
- All 60 expected ZIPs exist.
- All ZIP integrity checks pass.

Detailed reports:

```text
reports/pmdata_btc_5m_20260601_20260630.md
reports/pmdata_btc_5m_20260601_20260630_audit.json
reports/pmdata_btc_5m_20260601_20260615_download_manifest.json
reports/pmdata_btc_5m_20260616_20260630_download_manifest.json
```

## Known Data Quality Caveats

The March block has two short archives after retry:

- `2026-03-20 polygon_trades`: missing 3 five-minute files from `01:45-02:00 UTC`.
- `2026-03-23 poly_l2`: missing 98 five-minute files across `02:20-11:50 UTC` blocks.

The March 23 L2 gap matters for execution backtesting. Those windows must be excluded, patched from DOME, or explicitly labeled as fallback windows. Do not forward-fill through these gaps for fill-quality tests.

The June block has 12 short archives after retry:

- `poly_l2`: `2026-06-06`, `2026-06-17`, `2026-06-20`
- `polygon_trades`: `2026-06-03`, `2026-06-06`, `2026-06-10`, `2026-06-11`, `2026-06-12`, `2026-06-17`, `2026-06-19`, `2026-06-20`, `2026-06-22`

Use the generated missing-window catalog for exact five-minute gaps before running PMData-only execution tests.

The generated missing-window list is maintained in:

```text
catalog/pmdata_missing_windows.csv
```

## Strengths

PMData is strong for execution simulation:

- Sub-second event density.
- Daily archive structure.
- Parquet format suitable for fast selective reads.
- Includes book and last-trade-price events in one clock.
- Local receive timestamps allow latency assumptions to be tested.
- Better fit for queue-aware maker/taker simulation than 1 Hz reconstructed frames.

## Limitations

PMData still needs an adapter before it becomes the backtester input:

- The strategy wants explicit YES and NO ladders; PMData needs normalization into that internal representation.
- Full-depth behavior must be validated across complete days, not only one sample window.
- Missing archive windows must be part of the run mask.
- It does not directly replace DOME labels, Chainlink/Binance joins, and older reconstructed-frame tooling.
- Redistribution should remain private/internal unless the data provider terms explicitly permit public release.

## Strategy Role

Use PMData as the primary source for:

- `data/subsecond.py`
- fill simulation
- latency and markout analysis
- queue/hazard calibration
- production-grade backtest campaigns

Keep DOME as:

- baseline
- independent audit
- fallback for known gaps
- compatibility source for old loaders and labels

Final rule: PMData is the best execution source, but it is not a blind drop-in until the YES/NO ladder adapter and daily full-depth validation pass.