Open Source Python Library for Kalshi Prediction Market Data

Overview

Kalshi_Pull is an open source Python library for collecting market data from Kalshi, an exchange where contracts trade on real-world events. Each contract settles at one dollar if the event happens and zero if it does not, so a price reads directly as a market-implied probability.

The library collects three kinds of data: price history at daily, hourly, and minute frequency, every individual trade, and snapshots of the order book. It runs on a single ticker, a whole series, or the full committed catalog, and it can poll live markets on a schedule. The repository holds the library, the command line pullers, the live poller, the discovery scripts, and the ticker catalog.

Output is written to disk in one consistent layout, organized by data type, series, and date. Pulls resume from the last stored row and drop duplicates on write, so a store can be topped up on any schedule without gaps or repeats.

Coverage

The committed catalog covers 15 macro series, 524 events, and 4,065 unique tickers, measured from all_tickers.json in the repository. Discovery rebuilds the catalog for any series on the exchange.

Live polling runs over a hand-verified focus universe of 37 active tickers, pulling minute candles, trades, and order book snapshots on separate cadences.

How collection works

The exchange is the hard part. Kalshi splits its data across live and historical endpoints that name the same fields differently, serializes numbers as text, and drops settled markets off the live side. The library absorbs each of these, so settled and active markets read identically in the output.

What gets stored

Three schemas cover everything the library writes.

The data contract

• Timestamps are int64 UTC milliseconds in a column named ts_ms.
• Prices are float64 dollars between 0.0 and 1.0 and read directly as probabilities.
• Volumes and contract counts pass through exactly as the exchange reports them, and fractional counts are genuine fills.
• Where the API provides no value, the store holds NaN, never an invented zero.

Nine offline tests cover candle normalization, trade normalization, and the Parquet append round trip. They run against synthetic fixtures, with no credentials and no network.

Examples

The examples below show data the library collected, plotted and tabulated exactly as stored. Nothing is smoothed, filled in, or interpreted.

Both run on frozen copies of the output, collected 2026-08-09 through 2026-08-11 UTC and checked against MD5 manifests. Prices are stated as market-implied probabilities, which is what the exchange trades. The case workspace stays local.

Example one: US recession contracts

Four annual recession markets were pulled end to end: RECSSNBER-23, RECSSNBER-24, RECSSNBER-25, and KXRECSSNBER-26. The pull holds 1,678 daily candles across the four vintages, 8,890 hourly candles for the 2026 vintage, and 48,154 individual fills across the four tapes. Metadata recorded that the 2023, 2024, and 2025 vintages settled at no with value 0.0, and that the 2026 vintage remains active.

The RECSSNBER-23 peak close of 0.49 ties across 7 dates, 2022-12-31 through 2023-03-25, and the table reports the first.

The RECSSNBER-25 raw repricing entry is a documented zero-volume artifact. Excluding that date, its largest one-day repricing is -0.12 on 2025-04-10, volume 142,073.00.

Example two: the July CPI ladder

Kalshi lists one contract per threshold, so the July 2026 CPI event holds nine markets at strikes from 3.2 to 4.0 percent. All nine were pulled before the 2026-08-12 release, as a pre-event example. The pull holds 708 daily candles, 9,813 hourly candles, 53,250 minute candles, 12,460 fills, and 742 order book rows from three sweeps of the ladder.