make_card.py

"""Write README.md (the dataset card) from the manifests and the Parquet files."""

import json
from collections import Counter
from pathlib import Path

import pyarrow.parquet as pq

REPO = "lowskillcoding/pdata-prediction-markets"
KINDS = ("resolved", "markets-open", "events-open")
here = Path(__file__).parent
man = {k: json.loads((here / "manifests" / f"{k}.json").read_text()) for k in KINDS}
tables = {k: pq.read_table(here / "data" / f"{k}.parquet") for k in KINDS}
as_of = man["resolved"]["as_of"][:10]
snapshot = as_of


def fields(kind: str) -> str:
    rows = "\n".join(f"| `{k}` | {v} |" for k, v in man[kind]["fields"].items())
    return "| Field | Meaning |\n|---|---|\n" + rows


def by_venue(kind: str) -> str:
    c = Counter(tables[kind].column("source").to_pylist())
    return ", ".join(f"{s} {n:,}" for s, n in c.most_common())


n = {k: tables[k].num_rows for k in KINDS}
legs = sum(len(x) for x in tables["resolved"].column("markets").to_pylist())

card = f"""---
license: cc-by-4.0
pretty_name: "pdata.world prediction markets (Polymarket, Kalshi and 6 more venues)"
language:
- en
tags:
- prediction-markets
- polymarket
- kalshi
- manifold
- forecasting
- finance
- tabular
size_categories:
- 100K<n<1M
configs:
- config_name: resolved
  data_files: data/resolved.parquet
  default: true
- config_name: markets-open
  data_files: data/markets-open.parquet
- config_name: events-open
  data_files: data/events-open.parquet
---

# pdata.world prediction markets

Prices, volumes and results for prediction markets on eight venues — Polymarket, Kalshi,
Manifold, Myriad, Limitless, Predict, Opinion and Gemini — in one schema. The data is
collected by [pdata.world](https://pdata.world), which polls each venue every 10 to 60
minutes, and is published under CC BY 4.0.

This snapshot was taken on {as_of}. It has three parts:

| Config | Rows | What one row is |
|---|---|---|
| `resolved` (default) | {n['resolved']:,} events, with their top {legs:,} markets (at most 10 each) | An event pdata stopped tracking after it closed, with each top market's reported result |
| `markets-open` | {n['markets-open']:,} | A market open for trading: price, 24h move and volume |
| `events-open` | {n['events-open']:,} | An open event, which groups the markets of one question |

```python
from datasets import load_dataset

resolved = load_dataset("{REPO}", "resolved", split="train")
markets = load_dataset("{REPO}", "markets-open", split="train")

# or with pandas
import pandas as pd
df = pd.read_parquet("hf://datasets/{REPO}/data/markets-open.parquet")
```

## Rows per venue

- **resolved:** {by_venue('resolved')}
- **markets-open:** {by_venue('markets-open')}
- **events-open:** {by_venue('events-open')}

## Read this before using the numbers

- `probability` and prices are on a 0–1 scale.
- **Volume is not comparable across venues.** `volume` and `volume_24hr` are in each venue's
  own unit: shares for Polymarket and contracts for Kalshi (each pays $1 at resolution),
  play-money mana for Manifold, and each other venue's figure as it reports it. pdata's
  report [How eight venues count volume](https://pdata.world/research/prediction-market-volume)
  explains each one.
- **In `resolved`, a market's `probability` is the last price pdata saw**, which is usually
  after the market settled. It records the result; it is not a forecast. For prices before
  close, use the API's history endpoint while a market is live (history is deleted 10 days
  after close) — see [How accurate were prediction-market prices a day before the
  result?](https://pdata.world/research/prediction-market-accuracy) for one way to do it.
- **`result` is null where the venue never reports one.** Of the {legs:,} archived markets,
  {sum(1 for x in tables['resolved'].column('markets').to_pylist() for m in x if m['result'] is not None):,}
  carry a result.
- **A gap in `resolved`:** events that closed between 2026-06-02 and 2026-08-11 are missing,
  because they were pruned before a backup was taken. Events closed in the 10 days before
  the snapshot are in neither file yet.
- pdata groups Kalshi's sibling events for one game under a single group event
  (`is_group`), whose volume is the sum of its children's.

## Fields

### resolved

{man['resolved']['description']}

{fields('resolved')}

### markets-open

{man['markets-open']['description']}

{fields('markets-open')}

### events-open

{man['events-open']['description']}

{fields('events-open')}

Timestamps are UTC. The manifest each file was built from is in [`manifests/`](manifests/).

## Updates and versions

A new snapshot is added monthly; each is tagged with its date (this one is `{snapshot}`),
so `load_dataset("{REPO}", "resolved", revision="{snapshot}")` always returns the same
rows. For daily data, the same three files are rebuilt every night at
<https://api.pdata.world/bulk/> as gzipped JSON lines.

## Other ways to get the data

- REST API, no key: <https://api.pdata.world/docs>
- MCP server for AI assistants: `https://api.pdata.world/mcp`
- Read-only Postgres with price history:
  `postgresql://pdata_readonly:pdata_public_2026@pg.pdata.world:6433/pdata_new`
  (the password is published on purpose)
- Python client and examples: <https://github.com/noskillcoding/pdata-client>
- Research built on this data: <https://pdata.world/research>

## Licence and citation

CC BY 4.0. Credit "Data: pdata.world" with a link to <https://pdata.world>. This snapshot is
archived on Zenodo with DOI [10.5281/zenodo.23211611](https://doi.org/10.5281/zenodo.23211611). The underlying
market data remains subject to each venue's own terms. pdata.world is independent and not
affiliated with any venue.

```bibtex
@misc{{pdata_world_prediction_markets_{snapshot.replace('-', '_')},
  title        = {{pdata.world prediction markets: prices, volumes and results from eight venues}},
  author       = {{{{pdata.world}}}},
  year         = {{{snapshot[:4]}}},
  note         = {{Snapshot {snapshot}}},
  doi          = {{10.5281/zenodo.23211611}},
  howpublished = {{\\url{{https://huggingface.co/datasets/{REPO}}}}}
}}
```
"""
(here / "README.md").write_text(card)
print("README.md", len(card), "chars")