Best Polymarket Historical Data APIs & Datasets in 2026

Match the one bullet that describes your input, then read the matrix below for source-linked facts — this is not a ranking.

• Free official prices — Polymarket /prices-history and /batch-prices-history
• Current L2 books — Polymarket GET /book and CLOB WebSocket feeds
• Historical L2 over REST — PolyOrderbooks, PolymarketData, DepthFeed, Predexon
• Tick Parquet pipelines — Telonex, Predexon ticks, polyReplay archive
• Multi-venue research — Telonex, DepthFeed, Predexon
• On-chain analytics — Dune SQL for volumes, flows, and trader PnL (no off-chain L2)
• BTC/ETH Up/Down backtesting — PolyBackTest for full-depth on that segment
• Fine Up/Down backtests on BTC/ETH/SOL — PolyTest for sub-second-style capture on three coins
• Crypto API-first backtests — PolyOrderbooks at 250ms capture

Public pricing at time of verification (September 24, 2026). Table order is not a ranked recommendation — use the Best for column and summaries below to match your workflow.Scroll →

Summaries below are grouped for readability — not a #1–#7 ranking. Each provider fits different workflows; see the matrix for side-by-side facts.

Every provider on this page advertises coverage, resolution and price. None of them, including us, publishes an accuracy figure. So check it yourself. Ask for a sample and run one test: a real order book can never have a bid at or above the best ask.

import pandas as pd

df = pd.read_parquet("sample.parquet")          # any provider's L2 sample

def crossed(row):
    if not row["bids"] or not row["asks"]:
        return False
    return max(p for p, _ in row["bids"]) >= min(p for p, _ in row["asks"])

bad = df.apply(crossed, axis=1)
print(f"{bad.mean() * 100:.2f}% crossed")       # a real book is never crossed

This matters most for providers that ship deltas rather than snapshots. Rebuilding a book by seeding from one snapshot and applying every subsequent change sounds exact, and is not. We measured it across 381,093 checkpoints on a public Polymarket archive: the replay disagreed with the archive's own snapshots 67.8% of the time, and 6.4% of the reconstructed books were crossed. Extra levels outnumbered missing ones 4.2 to 1, which is what a lossy event stream looks like rather than a faulty replay.

Snapshot-based providers are not exposed to that particular failure, but they share its cause: a websocket carrying 24,000 events a second drops messages, and a collector that never re-reads the book from the venue will drift. The only defence is reconciling against the live API on a timer, and the only way to tell from outside is to test the output.

Applied to our own data that test returns 0.0% on the final minute before settlement, measured on 671 snapshots. We publish the number because we would rather you checked than took our word for it. Run the same test on any sample you are offered, including ours.