Comparing AI-Driven Trading Strategies on Kalshi

Strategy Comparison and Performance

The author tested two AI models, Fable 5.1 and Astra, by tasking them to develop a trading strategy for the 15-minute BTC up/down market on Kalshi. Fable 5.1 produced a more sophisticated strategy that incorporated volatility measurements, distance from target, and a fallback entry at the 7-minute mark, while capping bets at 14% of equity. Astra generated a simpler, less effective strategy that checked BTC price at the 3-minute mark and allocated a flat 20% of available cash without dynamic sizing or risk management.

Results and Live Deployment

During an 18-hour paper trading simulation with a $1,000 bankroll, Fable 5.1 achieved a nearly 20% return, while Astra resulted in a 60% loss. Astra’s higher win rate was negated by poor position sizing and excessive trading fees totaling $237, compared to $72 for Fable. The author subsequently deployed the Fable strategy on a Hostinger KVM2 VPS, using an API-driven setup via Codeex to automate the bot 24/7. In a live test with a $100 bankroll, the bot achieved a 77% return over 72 hours, though the author noted that much of this gain was correlated with broader market beta rather than pure alpha.

Implementation Mechanics

To automate the trading bot, the author utilized a Hostinger VPS configured with 8 GB of RAM and two CPU cores. The deployment process involved:

• Adding a Hostinger API key to a local environment file within the repository.
• Using Codeex to authenticate with the Hostinger API to provision the server.
• Automating the upload of the Python bot and strategy files via SSH keys generated during the deployment process.