This is the complete source code for the Kalshi trading bot I run on real money — documented trade by trade on the Rebel Edge Substack.

What the bot does

It trades BTC, ETH, XRP, and NEAR UP/DOWN contracts on Kalshi every 15 minutes using two independent AI models. Both models must agree at 55%+ confidence before a dollar moves. 91.9% of signals get rejected. The trades it does take have closed at a 55.3% win rate across 132 live positions — with XRP hitting 79%.

How it works

• Chronos-T5-Small (Amazon, 46M parameters) generates 100 probabilistic price forecasts from raw OHLCV history

• LightGBM classifier reads 26 engineered features across 8 categories — momentum, trend, volume, volatility, structure, oscillators, time encoding, and live Kalshi market data

• Dual-confirmation filter requires both models to agree on direction and confidence before any trade is placed

• Kelly Criterion (0.25 fraction, $10 cap) sizes every position

• Self-retraining loop refits the model every 6 hours and every 20 closed trades

• SQLite logging captures every signal, skip reason, and trade outcome for full auditability

What you get

• Start with the README file

• Full Python source code for the live trading bot

• Paper trading mode enabled by default — PAPER_MODE = True in one file

• Flip to PAPER_MODE = False when your data says you're ready

• SQLite schema with 5 tables including skipped signals

• Hetzner VPS deployment stack: cron, systemd, rotating logs

Who this is for

Builders who want a working algorithmic trading system they can understand, modify, and improve — not a black box. You should be comfortable with Claude Code and need a subscription. An account to Kalshi is preferable. Setting up a cloud server (e.g. Hetzner) is recommendable but it can run on your machine also (machine must be always on then). Basic familiarity with trading concepts helps but is not required.

Live results

Date July 13. $440 portfolio · +22.5% from $360 starting balance · 134 trades · 56.0% win rate. Updated live on the Rebel Edge dashboard. Paper bot running in parallel: +183% simulated across 431 trades.

Not financial advice. Past performance does not guarantee future results.