This dataset contains historical prediction market microstructure data. Access is gated to track usage and ensure compliance with data terms. Provide your contact information to request access.

Sayf Episodes: Prediction Market Microstructure Data for Agent Training

5,583 settled prediction market episodes with entry-time orderbook snapshots, executed trades with CLV (Closing Line Value) labels, and settlement outcomes across Kalshi and Polymarket.

Why this dataset exists

Teams building trading agents for prediction markets need structured historical data. The two options today are PredictionMarketBench (4 episodes) and raw API scraping. Neither provides historical orderbook depth at trade time, which is permanently lost if nobody captured it in real time.

Sayf captured it. Every episode includes the bid/ask/spread/depth at entry, the closing line at settlement, and the CLV grade (how much the entry price beat or missed the close). 468 episodes include actual executed trades with per-fill CLV.

What's in each episode

episode/
  meta.json           -- event metadata, category, venue, outcome, settlement
  orderbook.parquet    -- entry-time orderbook snapshot (bid, ask, spread, depth, liquidity bucket)
  trades.parquet       -- executed trades with CLV labels (may be empty)

meta.json fields

Field	Type	Description
schema_version	int	Always 1
event_id	string	Kalshi ticker or Polymarket condition ID
category	string	Market category (moneyline, totals, spread, soccer, tennis, etc.)
venue	string	kalshi or polymarket
timestamp	string	ISO 8601 observation timestamp
outcome	string	yes or no
settlement_price	float	Final closing mid price
closing_locked	bool	Whether the closing line was frozen before event start
clv_method	string	How the closing line was captured
event_start	string	When the underlying event began
entry_depth_dollars	float	Total dollar depth at entry
liquidity_bucket	string	thick (>$5K), medium ($500-$5K), thin (<$500), or unknown
trade_count	int	Number of executed trades on this ticker
has_trades	bool	Whether any trades were executed

orderbook.parquet columns

Column	Type	Description
timestamp	string	Observation time
bid	float	Best bid at entry
ask	float	Best ask at entry
mid	float	Midpoint at entry
spread	float	Bid-ask spread
depth_dollars	float	Total dollar depth
liquidity_bucket	string	thick / medium / thin / unknown
yes_depth_5lvl	float	Dollar depth on yes side (5 levels)
no_depth_5lvl	float	Dollar depth on no side (5 levels)
book_imbalance	float	Order book imbalance ratio
closing_bid	float	Bid at close
closing_ask	float	Ask at close
closing_mid	float	Mid at close (settlement reference)
closing_spread	float	Spread at close

trades.parquet columns

Column	Type	Description
timestamp	string	Trade execution time
side	string	Buy or sell side
price	float	Execution price
size	int	Contract count
clv_cents	float	CLV in cents (entry price minus closing line)
closing_mid_price	float	Closing mid used for CLV
outcome	string	Event outcome
category	string	Market category
entry_bid	float	Bid at trade time
entry_ask	float	Ask at trade time
entry_spread	float	Spread at trade time
entry_depth_dollars	float	Dollar depth at trade time
liquidity_bucket	string	Liquidity classification
clv_exec	float	Execution-adjusted CLV

Dataset statistics

Metric	Value
Total episodes	5,583
Episodes with trades	468
Venues	Kalshi (4,903), Polymarket (680)
Categories	15 (moneyline, totals, spread, tennis, soccer, props, etc.)
Outcomes	yes: 2,444 / no: 3,139
Liquidity	thick: 1,883 / medium: 1,936 / thin: 1,086 / unknown: 680
Total size	~90 MB
Format	Parquet + JSON
Collection period	Summer 2026

Quick start

import json
import pandas as pd
from pathlib import Path

# Load one episode
episode_dir = Path("episodes/KXMLBGAME-26AUG091610LADAZ-LAD")

meta = json.loads((episode_dir / "meta.json").read_text())
orderbook = pd.read_parquet(episode_dir / "orderbook.parquet")
trades = pd.read_parquet(episode_dir / "trades.parquet")

print(f"Category: {meta['category']}, Outcome: {meta['outcome']}")
print(f"Entry: bid={orderbook['bid'].iloc[0]:.3f}, ask={orderbook['ask'].iloc[0]:.3f}")
print(f"Trades: {len(trades)}")

Honest disclosure

Sayf's own trading track record on this data is negative: -4.3 cents mean CLV across taken trades. We are not selling signals or alpha. We are selling the environment. The CLV labels describe what happened structurally (did the entry beat the close?), not what you should trade.

The irreplaceable value is the historical orderbook depth. Kalshi does not publish historical L2 data. If nobody captured the bid/ask/depth at trade time, it is permanently gone. Sayf captured it continuously during the collection period.

Computed labels (CLV, liquidity buckets) are mechanically reproducible from public tape. The raw orderbook snapshots are not.

Citation

If you use this dataset in research, please cite:

@dataset{sayf_episodes_2026,
  title={Sayf Episodes: Prediction Market Microstructure Data for Agent Training},
  author={Khan, Sadaf},
  year={2026},
  url={https://huggingface.co/datasets/sadaftkhan/sayf-episodes}
}

License

This dataset is provided under a custom license. See LICENSE for terms. The data may not be redistributed without permission.

Contact

sadaf@alsayf.ai | alsayf.ai

Downloads last month: 6

Total file size:

94.1 MB