papylove/kalshi-perps-data

Kalshi Perps Market Data

Purpose: predicting whether a crypto perpetual future's ("perp") price on Kalshi will go up or down in the next minute. This dataset is the multi-timeframe price history + engineered technical features + news sentiment that feeds that prediction, collected automatically by an open-source trading bot: papykabukanyi/bettor.

This is the actual training data for the companion direction-classifier model, papylove/kalshi-perps-model, archived here so training never depends on only whatever recent window Kalshi's own API happens to still serve.

Two sources, honestly labeled (source column)

• Kalshi rows -- Kalshi's own live perp price, collected continuously since this bot started running. Kalshi's crypto perps only launched 2026-06-04, so this source can never cover more than a few weeks -- it is the ground truth for what the bot actually trades, but it is not deep history. These rows predate the source column and don't carry an explicit tag -- absence of a source value means Kalshi.
• source: coinbase -- the SAME underlying coins' spot price from Coinbase Exchange's public API (free, no key), used purely to give the model far more market experience than Kalshi's short lifetime allows. A perp's whole design tracks its underlying spot price closely, so this is a reasonable (not a perfect) proxy -- it is never used for live trading decisions, only for pretraining. Binance would offer even longer history but returns HTTP 451 (geo-blocked) from this bot's US-based servers, so Coinbase is the deepest free, unblocked source available. Coverage varies by coin -- confirmed live, all 16 coins are listed on Coinbase, but a coin's own history obviously can't exceed its actual existence (e.g. SUI, HYPE launched well under 4 years ago regardless of which exchange is asked).

Coverage

BTC, ETH, SOL, XRP, DOGE, LTC, BCH, LINK, SUI, NEAR, HYPE, kSHIB, ZEC have Kalshi-native history. (DOT, HBAR, and XLM are listed on Kalshi but have not yet produced any candlestick history as of this writing.) All 16 are covered by the Coinbase pretraining source.

Source

• Kalshi price data: Kalshi's own /margin/markets/{ticker}/candlesticks REST endpoint (1-minute and 60-minute OHLC) -- the actual tradable instrument's price, not an external index.
• Coinbase pretraining data: Coinbase Exchange's public /candles endpoint (free, no key, not geo-blocked) -- confirmed live to retain 2+ years of 1-minute candles and 5+ years of daily candles.
• News sentiment: free, no-key RSS feeds (Google News, CoinTelegraph, CryptoSlate, Decrypt) plus optional CryptoPanic/newsdata.io if a free key is configured, scored with simple keyword polarity. Only ever populated for the live Kalshi source -- there is no free historical news archive going back years, so the Coinbase pretraining rows honestly hold sentiment_score at 0.0 rather than fake a signal that doesn't exist.

Schema

Each row is one 1-minute observation for one instrument. Every timeframe from 1 minute up is derived directly from the 1-minute close series (pct_change/rolling windows) -- Kalshi's and Coinbase's candlestick APIs both cap out at 1-minute native resolution; no free source offers sub-minute (5s/15s/30s) history at all, so none is fabricated here.

Column	Description
ticker	Kalshi market ticker, e.g. KXBTCPERP
source	coinbase (pretraining only); absent/missing means Kalshi (live, traded)
ts	Unix timestamp (seconds, UTC) of this observation
close	Price at this minute
ret_1m, ret_3m, ret_5m, ret_10m, ret_15m, ret_30m	Trailing percent return over 1 to 30 minutes
trend_1h, trend_2h, trend_3h, trend_4h	Trailing percent return over 1 to 4 hours
dist_to_ma_15, dist_to_ma_30	Distance from the 15/30-minute moving average, as a fraction
volatility_5, volatility_15, volatility_30	Rolling std-dev of 1-minute returns over 5/15/30 minutes
trend_pct	Percent change over the trailing 6-hourly-candle window
sentiment_score	News sentiment for the underlying coin at collection time, in [-1, 1] (kalshi rows only; 0.0 for coinbase)
label_up	1 if price is higher 1 minute later, 0 if lower, null for rows too recent to know yet

All rolling/trailing features are computed leakage-free (only look backward from their own timestamp).

Update cadence

Kalshi rows: appended roughly every 15 minutes by the bot's scheduled collection job. Coinbase rows: a one-off (re-runnable) pretraining pull, not on a live schedule.

Disclaimer

Not affiliated with or endorsed by Kalshi or Coinbase. Provided for research/informational purposes as part of an open-source trading bot project. No guarantee of accuracy or completeness. Not financial advice.