Kalshi AI Trading Bot vs PredictPM

What Kalshi AI Trading Bot does

Open-source Python toolkit for automated trading on Kalshi, built around the question most bot repos skip: do you actually beat the book? An LLM reads each candidate market with a news summary, scores a probability through any model on OpenRouter, and the trade is sized at quarter-Kelly under a risk governor with a daily-loss limit and a drawdown kill switch. Every decision goes into a local journal; the edge harness then pulls Kalshi's own settlements, scores your settled trades by Brier score, log-loss and realised win rate against the price you paid, and refuses to call it an edge on fewer than ten trades that resolved after you bought. A policy step turns that record into a pre-trade gate that blocks categories you have lost money in. Ships a paper mode, a Streamlit dashboard, a Claude skill and an MCP server; MIT-licensed at github.com/ryanfrigo/kalshi-ai-trading-bot.

• `cli edge`: Brier score, log-loss and edge against the price you paid, scored on Kalshi's settlements and limited to trades that resolved after entry
• `cli improve`: re-derives a pre-trade gate from your settled record — blocks losing categories, warns on a losing side, shrinks overconfident estimates
• Risk governor on every live order: daily-loss limit, drawdown kill switch, manual halt, per-position cap, quarter-Kelly sizing
• Paper mode, and dry-run by default — `trade` and `close` only send orders with `--live`
• Signed Kalshi REST and WebSocket client, SQLite journal, Streamlit dashboard

• Three example strategies ship, and the README says plainly that none is the answer. AI Directional asks one model per market for a probability and trades the difference; the README corrects its own earlier “5-model ensemble” claim, because the ensemble code is quarantined and not wired in. Safe Compounder needs no model: it rests NO orders a cent under the ask where YES last traded at 20¢ or less and the edge clears 5¢ — many small wins, and a rare loss that erases a run of them. Beast Mode drops the guardrails, and the README says running it lost real money.
• The author publishes the live account it runs on, losses included. On Sep 25, 2026 equity was $901.50, 69.7% below a $2,975.25 peak, and 170 settled markets had netted +$128.56 at a 66% win rate. That account is also traded by hand, and the verdict on the toolkit's own journal reads “insufficient data”: one settled journaled trade, none yet scorable. Treat it as a measuring instrument without a proven edge, which is how it describes itself.
• The code is real: 141 Python files with a test suite and CI, 124 commits since July 2025 from the author and two outside contributors, the last on Sep 26, 2026. Requests to Kalshi are signed with RSA-PSS from a key file on your machine, and outbound calls go to Kalshi, the model provider you configure and public RSS feeds. sentry-sdk is listed as a dependency but never initialised.