Building Autonomous AI Trading Pods on Hyperliquid

The Breakthrough

The author establishes a repeatable, agentic pipeline that uses Codex 5.5 to autonomously research, backtest, and deploy isolated trading strategies (pods) on Hyperliquid, filtering out overfitted models by enforcing a minimum Sharpe ratio of 1.2 through Monte Carlo simulations and walk-forward testing.

What Actually Worked

• Strategy Ideation: The author prompts Codex 5.5 to generate three distinct strategies based on asset-specific volatility and market structure, such as volatility-targeted breakouts or funding premium carries.
• Rigorous Validation: Before live deployment, the agent runs a robustness suite including walk-forward stress tests and Monte Carlo simulations to prevent overfitting. The author explicitly rejects strategies that fail to maintain a Sharpe ratio above 1.0 after accounting for trading frictions.
• Modular Deployment: The final strategy, a US late-session reversal, is deployed as an isolated pod using a 15-minute candle timeframe, a $50 entry, and no leverage, with an autonomous exit timer set for a two-hour hold period.
• UI Integration: The agent generates a simple, dark-mode HTML terminal interface to monitor live trade execution and account balance status, allowing the author to track multiple parallel pods simultaneously.

Before / After

• Initial Backtest: The 4-hour EMA breakout strategy showed a 50% net return with a high Sharpe ratio, which the author initially found promising.
• Robustness Check: After applying Monte Carlo and walk-forward stress tests, the Sharpe ratio dropped to 0.4, leading the author to reject the strategy as an overfitted artifact of a recent price window.
• Final Selected Strategy: The US late-session reversal pod achieved a Sharpe ratio of 1.12, which the author accepted for live testing despite noting that the walk-forward window still showed some loss probability.

Context

The author utilizes a pod-based theory where individual, low-expectation trading strategies run autonomously in parallel. By stacking multiple uncorrelated strategies, the goal is to achieve net profitability even if individual pods underperform. This approach aims to reduce the psychological burden of monitoring a single high-stakes trade by diversifying across several automated, set-and-forget logic loops.