Descriptions:
Developer Sharbel A. recounts how a Claude Opus-powered trading bot on Hyperliquid perpetuals grew $50 into $500 before slowly draining to zero across 814 trades—losing $193,666 in closed P&L and $11,520 in fees as market conditions shifted. Manual parameter tweaking made things worse, dropping the win rate from 19% to 12%. The video documents what went wrong and how he built a self-healing replacement.
Inspired by Andrej Karpathy’s recently released auto-research GitHub framework, the new system runs a continuous optimization loop: it prompts Claude or GPT to generate candidate trading strategies, backtests each against two full years of minute-by-minute Bitcoin, Ethereum, and Solana price data, automatically flags look-ahead bias by rejecting implausibly perfect results (e.g., an 11,000% P&L ratio that was simply reading future data), and locks in only strategies that outperform the current champion. As of filming, the optimizer had run 133 generations; the generation-61 strategy produced an 89% backtested profit from a $1,000 starting balance across all of 2025.
The OpenClaw implementation walkthrough covers the full pipeline architecture—strategy proposal, backtesting, bias detection, and generational comparison—making this a useful reference for developers building LLM-driven evaluation loops, automated hyperparameter search, or self-improving agent systems beyond the trading domain.
📺 Source: Sharbel A. · Published March 19, 2026
🏷️ Format: Hands On Build
