od2961/polymarket-full-market-dataset

Polymarket Full Market Dataset

A complete, point-in-time archive of Polymarket: every event and market served by Polymarket's public Gamma API — from the platform's launch in 2020 through the snapshot date — with final resolution outcomes, full metadata (questions, rules, tags, timing, order-book flags), and dense daily OHLC price candles for every outcome token, assembled from the CLOB price-history API.

One snapshot = one self-contained JSON file. One row = one market, with its event context, resolution, and complete daily price series embedded.

At a glance (snapshot 2026_07_12_235011)

Market rows	1,788,703
Distinct events	679,884
Markets with candle series	1,787,257 (99.92%)
Embedded daily candle rows	38,263,931
Markets with a derived resolution	1,715,580 (95.9%)
Time span	2020-10-02 → 2026-07-12 (snapshot date)
File size	44.9 GB (uncompressed JSON)
Schema	polymarket_market_export_v2

Every snapshot ships with a .summary.json (same header metadata + one example row) and machine-readable coverage reports (polymarket_daily_candles_status.json, etc.).

Reading the four panels: (1) market creation exploded in 2026 — driven by high-frequency recurring series (hourly/5-minute crypto up-or-down, sports); the peak month alone added 336K markets. (2) Those series dominate the category mix ("Up or Down" + Sports + Esports + Tennis ≈ 75% of all markets). (3) Volume is bimodal: ~484K markets never traded ($0), while the traded bulk clusters at $100–$10K lifetime volume and 65 markets exceed $100M. (4) Coverage by creation cohort — this panel counts real trade prices (non-null closes), a stricter bar than having a candle series (99.9% of markets have one): pre-2023 cohorts have few real prices because CLOB price history doesn't exist for that era (structural, not a collection gap), and the 60–85% band in 2025–26 cohorts reflects ultra-short markets (5-minute/hourly series) whose lifetime is below the daily price-history fidelity. The resolution-rate drop in the newest cohort is simply markets that are still open.

How this snapshot was verified

Coverage claims are audited, not assumed. For this snapshot:

• Events — exact: the archive's event ids were diffed against a full enumeration of Gamma's closed-events and open-events keyset feeds (670k+ closed, ~9.7k open). Missing: zero. (14 events exist here that Polymarket has since deleted upstream — retained deliberately; this is an archive.)
• Markets — sampled exact: 2,498 randomly sampled events (over-weighted toward recently created ones) were re-fetched live from Gamma and every listed child market checked against the archive: 0 missing of 6,698, 0 closed-flag mismatches, 0 missing outcome prices.
• Candles — internal audit: 0 duplicate (token, interval, day) keys across partitions; 99.92% of markets carry a complete dense daily series and 100% of 3.55M tokens have daily rows; 99.99+% of closed markets carry outcome prices. (This required repairing a systematic upstream quirk: retroactively-listed markets carry start_time > end_time in Gamma metadata and were invisible to the standard backfill — their histories were recovered via window-independent fetches.)
• File-level: the shipped JSON is re-parsed end-to-end after export; header row_count/markets_with_candles must match streamed reality exactly; ids must be unique; row ordering must match the declared row_ordering. A 348-row sample was additionally compared against the live API (zero regressions).

File format

Each snapshot is a single JSON object, laid out so that you never need to load 44 GB at once:

{
  "dataset_name": "polymarket_full_market_dataset",
  "row_schema_version": "polymarket_market_export_v2",
  "generated_at": "...",
  "row_count": 1788703,
  "markets_with_candles": 1757791,
  "row_ordering": {...},
  "row_shape": {...},          ← full machine-readable schema
  "rows": [
    {...one COMPACT market object per line...},
    {...}
  ]
}

The header is pretty-printed; each element of rows is exactly one line. The fast path is line-oriented — no streaming-JSON library required:

import json

def iter_markets(path):
    with open(path, "rb") as f:
        in_rows = False
        for line in f:
            s = line.strip()
            if not in_rows:
                if s == b'"rows": [':
                    in_rows = True
                continue
            if s in (b"]", b"],", b"}"):
                continue
            yield json.loads(s.rstrip(b","))

for row in iter_markets("polymarket_full_market_dataset_2026_07_12_172952.json"):
    ...

Rows are ordered by model-facing times.open_time ascending (nulls last), then market_created_time, then market_id — a coarse chronological pass over markets.

Row schema

Each row contains these blocks (full field-level types in the header's row_shape):

Block	Contents
ids	market_id (Gamma integer id), condition_id (0x… CLOB hash), event_id, clob_token_ids (one ERC-1155 token id per outcome)
question / question_sources	canonical question text and every source it was derived from (market title, event title, memory case)
rules	primary rules text, market/event descriptions, resolution source URL, resolver address
resolution	derived outcome: winner_index / winner_label / winner_yes, resolution_method, resolution_time, UMA status fields, strict_settlement_yes
outcomes	outcome labels, count, is_binary, latest prices per outcome
times	every timestamp: market/event created/start/end/updated, close_time, and the model-facing open_time used for ordering
market	title, slug, descriptions, active/closed flags, volumes, raw_meta extras (best bid/ask, closed time, automation flags)
event	parent event: title, slug, category, tags (full Gamma tag objects), series id, raw_meta (ticker, images)
trade_metrics	volume aggregates (total/24h/1wk/1mo/1yr, AMM vs CLOB), spread, best bid/ask, last trade price
order_book	order-book capability flags (accepting orders, funded, ready, restricted, RFQ)
market_closing_rules	close/end timing details, UMA end date, liveness, activation flags
market_memory	enrichment from the research pipeline's memory store — question/rules snapshot, category, entity tokens, baseline probability, price-trace stats. Only present for a ~47k-market research subset (see caveats)
candles	the full dense daily price series (see below)
candle_summary	count, empty-row count, intervals present, latest candle, volume stats
availability	has_candles, has_resolution, has_market_memory_case — flags for quick filtering

Resolutions

Polymarket has no separate "resolutions" API; outcomes are encoded in final prices. The resolution block is derived at export time:

• price_extreme (94% of resolved rows): some outcome's final price pinned to ≥ 0.999 (winner) — the normal on-chain settlement signature.
• deterministic_winner (~2%): market closed/resolved without pinned prices; the max-price outcome is taken, with YES/NO tie-breaking. Treat as slightly weaker evidence (includes 50-50 voids).
• strict_settlement_yes is a conservatively-derived boolean settlement label, present only for the memory-case subset.

Markets that are still open (or closed without usable prices) have availability.has_resolution = false.

Candles — read this before using the price series

Dense daily candles (period_interval_minutes = 1440) target one row per CLOB token per UTC day across the market's listing window, built from the CLOB prices-history endpoint:

• Multi-outcome markets have one series per outcome token, interleaved in candles (group by token_id; each token's series is time-ordered).
• Days with trades carry real OHLC (close = probability in [0,1] for that outcome).
• Days without trades are synthetic: either carry-forward of the last close (raw_meta.carry_forward = true) or fully empty placeholders (raw_meta.empty = true, empty_reason = "no_prior_trade"). This is by design — the series is dense so that day-indexed joins work — but it means you must check raw_meta before treating a candle as a market observation.
• A small set of markets additionally carries finer close-window candles (period_interval_minutes ∈ {15, 60}) near market close.
• volume/liquidity are usually null at daily fidelity (the endpoint returns prices only).

Strict full-window coverage (every token, every expected day) is tracked in polymarket_daily_candles_status.json; markets can be metadata-complete but candle-sparse — filter with availability.has_candles and candle_summary.

Examples

A binary recurring crypto market (the highest-frequency market family on Polymarket — hourly/5-minute up-or-down series):

{
  "ids": {"market_id": 2603340, "event_id": 610611,
          "condition_id": "0xc19b389c49ce4bc0d5e266273890ede2e55c159d6a20b5abdc984d7fb4579600",
          "clob_token_ids": ["34004…7473", "18197…0081"]},
  "question": "Bitcoin Up or Down - June 21, 12AM ET",
  "event": {"title": "Bitcoin Up or Down - June 21, 12AM ET", "category": "Crypto",
            "slug": "bitcoin-up-or-down-june-21-2026-12am-et"},
  "outcomes": {"is_binary": true, "labels": ["Up", "Down"], "latest_prices": [0.0, 1.0]},
  "resolution": {"winner_index": 1, "winner_label": "Down", "winner_yes": false,
                 "resolution_method": "price_extreme",
                 "resolution_source": "https://www.binance.com/en/trade/BTC_USDT",
                 "resolution_time": "2026-06-21T05:00:00+00:00",
                 "uma_resolution_status": "resolved"},
  "times": {"open_time": "2026-06-19T04:01:59+00:00", "close_time": "2026-06-21T05:00:00+00:00"},
  "market": {"closed": true, "volume": 32553.71},
  "candle_summary": {"count": 6, "empty_candle_count": 2, "interval_minutes_present": [1440]},
  "candles": [
    {"token_id": "18197…0081", "outcome": "Down",
     "end_period_ts": "2026-06-19T23:59:59.999999+00:00",
     "open": null, "close": null,
     "raw_meta": {"empty": true, "empty_reason": "no_prior_trade",
                  "source": "polymarket_prices_history_dense_daily"}},
    {"token_id": "18197…0081", "outcome": "Down",
     "end_period_ts": "2026-06-20T23:59:59.999999+00:00",
     "open": 0.5, "high": 0.5, "low": 0.5, "close": 0.5, "raw_meta": {"…": "…"}},
    "… one row per token per day …"
  ],
  "availability": {"has_candles": true, "has_resolution": true, "has_market_memory_case": false}
}

Note the two tokens (Up/Down): each gets its own daily series; the pre-listing day is an empty placeholder; the resolution (Down, price_extreme) matches latest_prices [0.0, 1.0].

A categorical sports market — same shape, non-YES/NO labels:

{
  "ids": {"market_id": 2603415, "event_id": 610636},
  "question": "Dallas Wings vs. Connecticut Sun",
  "event": {"category": "Sports", "slug": "wnba-dal-conn-2026-07-02"},
  "outcomes": {"labels": ["Dallas Wings", "Connecticut Sun"], "latest_prices": [1.0, 0.0]},
  "resolution": {"winner_index": 0, "winner_label": "Dallas Wings", "winner_yes": true,
                 "resolution_method": "price_extreme",
                 "resolution_source": "https://www.wnba.com/scores"},
  "market": {"closed": true, "volume": 170961.86},
  "candle_summary": {"count": 30, "empty_candle_count": 2}
}

Two weeks of daily prices per team-token; the winning token's final close pins to ~1.0.

Usage recipes

All resolved markets in a window, as lightweight JSONL:

out = open("resolved_2026H1.jsonl", "w")
for row in iter_markets(SNAPSHOT):
    res, times = row["resolution"], row["times"]
    if res.get("winner_index") is None:
        continue
    t = res.get("resolution_time") or times.get("close_time") or ""
    if "2026-01-01" <= t < "2026-07-01":
        out.write(json.dumps({
            "market_id": row["ids"]["market_id"],
            "question": row["question"],
            "labels": row["outcomes"]["labels"],
            "winner": res["winner_label"],
            "close_time": times["close_time"],
            "volume": row["market"]["volume"],
        }) + "\n")

A per-day probability panel for one market:

from collections import defaultdict
series = defaultdict(dict)
for c in row["candles"]:
    if c["period_interval_minutes"] != 1440 or c["close"] is None:
        continue
    day = c["end_period_ts"][:10]
    series[c["outcome"]][day] = float(c["close"])

⚠️ Leakage warning: this is one row per market including the final resolution. For training forecasters, never show a model fields computed after your decision time (resolution, outcomes.latest_prices, final candles, market.closed, volumes-at-close…). For leakage-aware supervised splits, prefer a temporal panel built from the candles, or the upstream export_polymarket_temporal_panel.py.

Caveats and known limitations

• 1,377 rows have condition_id: null — early pre-CLOB / draft markets; they also lack tokens and candles. Metadata-only.
• market_memory covers only ~47k markets (a research-pipeline subset built in March 2026). availability.has_market_memory_case flags it; all other enrichment is present for every row.
• Synthetic candle rows (empty / carry-forward) are part of the dense series by design — check raw_meta.
• No tick data. Candles are the only price history; Polymarket does not serve historical ticks, and a server outage May 6 → Jul 11 2026 makes sub-daily data for that window unrecoverable.
• A dense series is not the same as real trade data. 99.9% of markets have a complete daily series, but for pre-2023 listings (no CLOB history exists for that era) and many ultra-short 2025–26 markets (lifetime below daily fidelity) the series is entirely placeholder rows. Filter on close IS NOT NULL / candle_summary.empty_candle_count when you need actual prices.
• Snapshots accumulate in this repo; use latest pointers/README to find the newest. Older snapshots are kept for reproducibility.
• Deleted-upstream events are retained (14 as of this snapshot).

Provenance

Snapshots are exported from a PostgreSQL archive continuously ingested from Polymarket's public APIs (gamma-api.polymarket.com for events/markets/metadata, clob.polymarket.com/prices-history for candles). All writes are idempotent upserts keyed on Gamma ids; resolution fields refresh whenever a closed market is re-observed. The archive → export path is a single REPEATABLE READ transaction, so header counts and rows are mutually consistent.

Snapshot history

Snapshot	Rows	With candles	Candle rows	Size
2026_07_12_235011	1,788,703	1,787,257	38,263,931	44.9 GB
2026_07_12_172952	1,788,703	1,757,791	37,699,627	44.6 GB
2026_05_01_221213	1,004,540	982,741	24,239,343	25.2 GB
2026_04_30_215103	1,004,540	470,775	11,279,708	17 GB
earlier April snapshots	1,004,540	partial	—	4.9–15 GB

Data © Polymarket, collected from public APIs; provided for research use.