A trading bot that admits it loses money

What it does

This repo is a Python scaffold for automating trades on Kalshi prediction markets. It wraps the exchange’s REST and WebSocket APIs in a signed client, ingests market data into SQLite, and exposes a CLI and Streamlit dashboard to run and monitor strategies. Three example bots ship with the kit: an LLM-driven directional trader, a math-only “safe compounder” that hunts for NO-side edge, and an aggressive “Beast Mode” the authors explicitly warn against using with real money.

The interesting bit

The README is refreshingly honest: it admits the example strategies lose money, clarifies that the “AI” bot calls a single LLM (not an ensemble), and notes that past performance dissolves quickly in prediction markets. That candor is rarer than a profitable edge.

Key highlights

• Pluggable OpenRouter client with model swapping and a daily cost circuit breaker.
• Built-in paper trading and SQLite telemetry for every signal, decision, and P&L tick.
• Risk helpers include fractional Kelly sizing, stop-loss/take-profit exits, and drawdown breakers.
• “Safe Compounder” strategy places resting maker orders on NO-side markets without touching an LLM.
• Unwired scaffolding for multi-agent debate and sentiment analysis sits in src/agents/ and src/data/ for forking.

Caveats

• The AI directional strategy is not a multi-model ensemble; parallel agent voting in src/agents/ is still unwired.
• Several directories contain “scaffolding” rather than fully integrated features.
• Beast Mode exists mainly as a cautionary tale and is explicitly not recommended for live trading.

Verdict

Grab this if you want a well-documented starting point for Kalshi automation and appreciate a maintainer who treats “past results may vary” as a feature, not a bug. Skip it if you need a turnkey money printer—the README will tell you explicitly that you won’t find one here.