Descriptions:
This video documents a hands-on experiment in which the creator uses Claude Code to autonomously build, test, and refine algorithmic trading strategies overnight — entirely from a single prompt with no further human intervention. The AI agent iterates through multiple indicator designs (Keltner EMA, MFI Regime, Momentum Squeeze) and self-evaluates each one against a custom backtesting engine built on VectorBT, a Python-based vectorized backtesting library.
The testing methodology is notably rigorous for a YouTube tutorial: strategies are evaluated on both in-sample and out-of-sample candlestick data, with final candidates required to survive Monte Carlo bootstrapping (randomized candlestick sequences). Key metrics tracked include Sharpe ratio, Sortino ratio, maximum drawdown, and win rate. The final overnight output achieved a 0.66 Sharpe ratio, 19% max drawdown, and 45% win rate across 409 trades — with near-identical in-sample and out-of-sample results, suggesting limited overfitting.
Viewers interested in agentic AI workflows for quantitative finance will find practical takeaways around how to structure constraints for an AI coding agent, how to design a backtesting harness that an LLM can interact with programmatically, and what statistical thresholds are worth enforcing before running a strategy live. The creator also references Signal Swap for connecting Python-generated signals to live exchanges via TradingView.
📺 Source: Trade Tactics · Published February 25, 2026
🏷️ Format: Workflow Case Study
