agentic-polymarket

38,915 settled Polymarket binary event markets with full hourly price curves, question text, resolution terms, and ground truth outcomes. Prepared for research on "getting LLM agents to trade on prediction markets."

Companion code (backtest env + agent trading interface): see RSI-economy/shadow-market.

What this dataset solves

Historical price series cannot be used directly as backtest targets —— a recording does not react to agent behavior: any quantity fills at any price, buys and sells at the same price, own trades do not move price, and positions carry no capital cost. The result is that certain strategies perform well in

backtest but are not executable in reality (typical pattern: only entering near implied probability 0 or 1 to capture terminal residual value).

Files

File	Contents
data/markets.jsonl	One market per line: path (hourly VWAP anchor curve) / outcome / T_start / T_end / category / slug / trade count and volume
data/questions.jsonl	condition_id -> question / resolution_criteria / tick, 100% coverage
data/index.json	Construction pipeline and selection funnel
splits.json	Frozen train/val/test market id table + sha256
depth_table.json	Depth table D(m,k) and half-spread δ(m), bucketed by mid price, median
market_params.json	Shadow market mechanism parameters (impact strength / decay / spread etc.)
Polymarket-v{1,2}/daily_aligned/*.parquet	Raw trade-by-trade (1364 daily files, 19.31 GB)

Polymarket-v{1,2}/ is the raw material for building the derived files above: a 24-column wide table, one row per trade, with condition_id / price / usdc_amount / taker_direction / resolution status / lifecycle timestamps. Not needed if you only want the market store; needed if you want to change the construction pipeline (different aggregation granularity, different filters, or recompute depth yourself).

# Download derived data only (157 MB)
from huggingface_hub import snapshot_download
snapshot_download("tennant/agentic-polymarket", repo_type="dataset",
                  allow_patterns=["data/*", "*.json"])

# Download one day of raw trades
from huggingface_hub import hf_hub_download
hf_hub_download("tennant/agentic-polymarket", repo_type="dataset",
                filename="Polymarket-v1/daily_aligned/2026_03_23.parquet")

# Download only 2026 raw trades
snapshot_download("tennant/agentic-polymarket", repo_type="dataset",
                  allow_patterns=["Polymarket-v*/daily_aligned/2026_*.parquet"])

Scale

• Raw archive 19.31 GB / 1364 daily files (2022-11-21 .. 2026-08-22)
• Scanned 476,956,714 rows of trade-by-trade data (YES leg; all legs ~950 million)
• Retained 38,915 markets: settled + has ground truth + YES/NO binary + trades >= 80
• Excluded Up or Down short-period price candles (not predictable by research, would inflate agent scores without teaching anything)

Frozen splits (temporal, not random)

split	n	T_end range	base rate
train	9,020	2023-01-21 .. 2026-04-01	40.4%
val	1,933	2026-04-01 .. 2026-06-12	42.2%
test	1,934	2026-06-12 .. 2026-08-22	36.2%

sha256 = 9930e963b1ec75613a62dac9393169929adc9609f6306ae796217feb73c14b36

Why temporal split: (1) Polymarket often splits an event into multiple markets differing only by date; random splits would let test answers be inferred from train. (2) Pre-trained models have a knowledge cutoff; using early markets for tuning and late markets for evaluation is the only way to distinguish "actually can predict" from "remembered the outcome."

Data reliability

The anchor curve is the foundation of the entire environment. The verification method checks calibration of the terminal anchor price: bucketed actual resolution rates must be monotonic and close to the bucket center.

Terminal 10% anchor      Actual resolution rate
0.0-0.1  n=22,433    0.000
0.3-0.4  n=   681    0.217
0.5-0.6  n=   470    0.623
0.9-1.0  n= 9,914    0.999

Brier = 0.0196   (constant-guess 0.5 = 0.2500)

Monotonic and well-calibrated. This check also catches reversed outcome labels, confused YES/NO legs, and misaligned timestamps.

Known limitations (must read)

Depth table is not a historical snapshot. Polymarket does not provide historical order books (/book only returns the current state), and no public dataset contains book archives. Therefore D(m,k) and δ(m) are calibrated by live sampling then frozen (1,233 two-sided books). Depth is aggregate liquidity; calibrating today does not leak any historical market outcome —— anachronism is only in microstructure fidelity, not in information. However, any conclusion relying on precise slippage magnitudes must explicitly state this limitation.

tick is not globally 0.001. Measured adjacent-level spacing: only 21% fall on consecutive 0.001, 59% are 10 ticks apart (0.01 grid). So tick is determined per-market; questions.jsonl carries a tick field.

High crypto exposure. ~13k of 39k markets are Crypto/Bitcoin/Ethereum/Price Action. To study transferable retrieval/reasoning, filter them out, then require >=40 price points and

=$10k volume to get ~6,400 "research-tractable" markets.

Source and license

Derived from TimeSeventeen/Polymarket-v1 and -v2 (CC BY 4.0) daily_aligned partitions, question and resolution criteria from Polymarket Gamma API. This dataset is likewise released under CC BY 4.0.

Note: the two repos have a 26-day overlap from 2026-04-03..04-28, but v2 files are nearly empty in the overlap window (2026_04_15: v1 = 4,387,421 rows, v2 = 2 rows). During construction, the larger byte count wins per day; 26/26 days were awarded to v1.

Usage

import json

markets = [json.loads(l) for l in open("data/markets.jsonl")]
questions = {json.loads(l)["id"]: json.loads(l)
              for l in open("data/questions.jsonl")}
splits = json.load(open("splits.json"))

val_ids = set(splits["ids"]["val"])
val = [m for m in markets if m["id"] in val_ids]

m = val[0]
print(questions[m["id"]]["question"])
print(f"{len(m['path'])} price points, outcome={m['outcome']}")