PolyPolitics1000: Prediction Market Dataset

A curated dataset of 316 resolved political prediction markets from Polymarket, containing 63,331 5-minute OHLC candles with ground truth labels. Designed for machine learning research, this dataset includes filtered data to remove wash trading and comprehensive market metadata.

Dataset Structure

Files

• candles_5m_partial_cleaned.parquet (Recommended): Cleaned dataset with wash trading filtered. Contains 237 markets and 46,515 candles.
• candles_5m_partial.parquet: Original dataset (316 markets, 63,331 candles) including markets flagged for suspicious volume.
• markets_metadata_partial.csv: Metadata for all markets, including resolution status and themes.
• market_volume_stats.csv: Volume statistics and Benford's Law test results.
• suspicious_markets.csv: List of 79 markets filtered due to suspicious trading patterns.

Statistics

Metric	Cleaned Dataset	Original Dataset
Markets	237	316
Candles	46,515	63,331
Granularity	5-minute OHLC	5-minute OHLC
Label Coverage	100%	100%

Schema

Candles (candles_5m_partial_cleaned.parquet)

Field	Type	Description
condition_id	string	Market condition ID (primary key component)
market_id	string	Market identifier
token_id	string	Outcome token ID (2 per market)
bucket_start_ts	int	Unix timestamp (5-minute bucket start)
open, high, low, close	float	OHLC prices (0-1 range)
volume	float	Total volume traded
trade_count	int	Number of trades in bucket
label_outcome_index	int	Ground truth outcome (0 or 1)

Metadata (markets_metadata_partial.csv)

Field	Type	Description
condition_id	string	Market condition ID
market_id	string	Market identifier
token_id_0, token_id_1	string	Outcome token IDs
label_outcome_index	int	Ground truth outcome (0 or 1)
theme	string	Market category

Usage

Loading

import pandas as pd

# Load cleaned dataset (recommended)
df = pd.read_parquet("candles_5m_partial_cleaned.parquet")
metadata = pd.read_csv("markets_metadata_partial.csv")

Feature Engineering Example

# Sort and calculate features
df = df.sort_values(["condition_id", "token_id", "bucket_start_ts"])

# Moving averages and price changes
df["price_change"] = df.groupby(["condition_id", "token_id"])["close"].pct_change()
df["volume_ma_10"] = df.groupby(["condition_id", "token_id"])["volume"].transform(
    lambda x: x.rolling(10, min_periods=1).mean()
)

Data Quality

• Wash Trading: 79 markets (25% of original) were removed using Benford's Law analysis to filter artificial volume.
• Validation:
  • Prices strictly within range.[1]
  • OHLC constraints enforced (low <= open/close <= high).
  • No negative volumes or trade counts.

Model Performance (Baseline)

Random Forest results on the cleaned dataset:

• Accuracy: 60.42%
• ROC-AUC: 0.6328
• Brier Score: 0.2174

Citation

@dataset{polypolitics1000,
  title={PolyPolitics1000: Prediction Market Dataset},
  author={ValCtrl},
  year={2025},
  url={https://huggingface.co/datasets/valctrl/polypolitics1000}
}