naratron33/fable5-kalshi-l2

Kalshi L2 order book archive

Recorded L2 book snapshots, trades, settlements and market metadata from Kalshi, captured continuously by the Fable5 recorder. There is no public historical L2 order book data for Kalshi — it exists only from the moment someone records it, which is what this is.

Tables

Hive-partitioned Parquet, series=…/date=…. Read with DuckDB, pandas or polars.

table	what it holds
book	best bid/ask, spread, and every level of both sides, per 60s interval
trades	the public tape, including taker_side — the maker/taker attribution
settlements	determined/settled outcomes with result and settlement_value
market_meta	tick grid (price_ranges) over time, and close times — change-data-capture, see below
coverage	which hours are present, missing, or damaged

Read coverage first

Recording is continuous but not gap-free — a host reboot or a sleep leaves a hole, and a backtest will run straight across one without complaining.

SELECT hour_utc, present, degraded, note
FROM 'coverage/**/*.parquet'
WHERE NOT present OR degraded
ORDER BY hour_utc;

Depth is complete, not a top-of-book sample

yes_levels and no_levels hold the whole book — every resting price level on each side, best first, not a top-N slice. Measured across 1,255 live books: median 14 levels per market across both sides, p90 28, max 67.

This matters most for fill models. The size resting behind your quote is what decides whether a resting order would really have been filled, and a truncated ladder looks identical to a thin market. Schema 3 and earlier published only the top five levels a side, which discarded about half of every level recorded — check schema_version before mixing days.

SELECT ts_ms, len(yes_levels) AS yes_depth, len(no_levels) AS no_depth
FROM read_parquet('book/**/*.parquet', hive_partitioning = true)
WHERE ticker = 'KXRAIN-26SEP01-DC' ORDER BY ts_ms LIMIT 5;

market_meta is change-data-capture

A row exists only where something changed, so a market whose grid has not moved has no row in that day's partition. Reading one date= partition therefore under-reports, and reading all of them over-reports. What you want is the latest observation at or before your day:

SELECT ticker, price_level_structure, price_ranges_json, close_time
FROM (
  SELECT *, ROW_NUMBER() OVER (PARTITION BY ticker ORDER BY observed_ts_ms DESC) rn
  FROM read_parquet('market_meta/date=*/*.parquet', hive_partitioning = true)
  WHERE date <= '2026-08-27'
) WHERE rn = 1;

fable5.bridge.parquet.market_meta_asof(root, day) does exactly this, and is the supported path — schema 2 published the whole table into every partition, so a plain WHERE date = … returned metadata from days in the future.

Units

column suffix	meaning
_scaled on a price	10,000 = $1.00 (four decimals; Kalshi runs 12 tick grids)
_scaled on a size	100 = one contract (counts are fractional to 0.01)
_ms	epoch milliseconds, UTC
yes_bid / yes_ask / mid	dollars, as floats, for convenience only

Use the scaled integers for arithmetic. The float columns exist for reading.

Kalshi publishes only bids, on both sides, so yes_ask is derived as 1.00 − best NO bid. Re-deriving that yourself is the most common source of sign-flip bugs in Kalshi code; the published columns already have it right.

Getting it

Partitioning is by series, which makes one series-day a single ~17 KB read — but it also means the archive is thousands of small files. A plain snapshot_download of the whole thing can hit HTTP 429 on a free account (measured 2026-08-31). Two ways round it, both tested:

Pull only what you need — the right default for backtesting one series:

from huggingface_hub import snapshot_download

root = snapshot_download(
    "fable5-kalshi-l2", repo_type="dataset",
    allow_patterns=["README.md", "coverage/**", "market_meta/**",
                    "book/series=KXRAIN/**", "trades/series=KXRAIN/**"],
)

Or take everything, with fewer parallel requests:

root = snapshot_download("fable5-kalshi-l2", repo_type="dataset", max_workers=4)

Cached and resumable either way, so a second call fetches only what changed. Pin revision="..." to make a result reproducible.