Coyevans/polymarket-kalshi-scoresync-orderbook-sample

Polymarket x Kalshi Orderbook Archive - Free Score-Synced Sample

A free excerpt of the ZenHodl Polymarket & Kalshi Historical Orderbook Archive, published so you can VERIFY the two non-reconstructable features before you buy -- the things competitors (Telonex, PolymarketData) do not ship:

1. Per-row score-sync - every Kalshi row carries the live game state next to the quote: home_score, away_score, score_diff, period, time_remaining, game_state.
2. Cross-venue - the same games are priced on both Polymarket and Kalshi on one timeline (a separate cross-venue matched-book sample is published on HuggingFace/Zenodo).

These are score-synced top-of-book quotes (bid / ask / mid / spread) joined to live game state -- not a multi-level depth ladder. The non-reconstructable part is the join: you cannot retroactively attach live game state to historical prices, you had to capture both at once.

The 30-second verify

Open kalshi_scoresync_PHI_at_TOR_HIGHLIGHTS.csv. It's one market — the Kalshi "Philadelphia wins" moneyline for a real NBA game (PHI @ TOR, a 116–115 overtime thriller). Find a row where score_changed = True and watch mid (PHI's win probability) reprice with the scoreboard: it tracks the lead all game and swings the full 0.36–0.65 through the overtime finish, on a tight ~2¢ book. That join — the live score / period / clock sitting on every quote row — is what can't be rebuilt after the fact, and what the full archive gives you across 110M Kalshi + 25.6M Polymarket rows.

Files

• kalshi_scoresync_PHI_at_TOR_2026-01-11.parquet - one clean market (PHI moneyline, yes = Philadelphia wins), 1,844 captured snapshots, all 26 columns.
• kalshi_scoresync_PHI_at_TOR_HIGHLIGHTS.csv - same market, annotated + downsampled to read by eye.
• DATA_CARD.json - provenance, coverage, data-quality notes, license.

Data quality (the honest version)

Real live capture, warts disclosed: ~33s median snapshot cadence; top-of-book quotes only — yes_bid_size/yes_ask_size are empty (no per-level depth or sizes on this archive); and ~0.2% of in-game rows show a transient score-feed hiccup (a momentary 0). Drop those with a one-line filter on non-decreasing score:

df = df.sort_values('timestamp')
df = df[df['home_score'].eq(df['home_score'].cummax()) & df['away_score'].eq(df['away_score'].cummax())]

This is one market of the full multi-market archive (each game also has the other side's moneyline, spreads and totals).

Full archive

~110M Kalshi + ~25.6M Polymarket score-synced snapshots, Jan 2026 window (known gaps disclosed in-product). Buy: https://zenhodl.net/products Methodology + honest analysis: https://zenhodl.net/blog/are-prediction-markets-calibrated

Not a trading signal. These markets are efficient (we prove it ourselves) -- this is research / backtest / model-training data. License: https://zenhodl.net/data-license