Summary
The Algovibes channel set out to definitively disprove AI-powered crypto trading — and ended up with results more interesting than expected. Using Anthropic’s Claude API, the creator built a rigorous backtesting framework that fed the model two years of Bitcoin 5-minute candlestick charts. For each decision point, Claude receives only the previous 8 hours of price action (96 candles) rendered as a candlestick image, then returns a single JSON object: a direction (long, short, or flat), a confidence score, and a one-sentence reason. Every decision is evaluated against the actual 24-hour outcome, with 0.08% roundtrip trading costs applied.
The video includes a detailed code walkthrough covering the full pipeline — data loading from Parquet, chart rendering with mplfinance, base64 encoding for the Anthropic API, caching to CSV to avoid redundant API spend, and parallel execution. The creator candidly addresses the limitations of passing visual charts versus structured OHLCV data in JSON or text form, flagging it as less precise but intentionally matching the “internet-style” use case that circulates widely on social media.
Results are compared against random decisions and a simple buy-and-hold strategy. Whether Claude edges out randomness is the central tension the video builds toward, making it a useful reference for anyone curious about the genuine signal — or lack thereof — in using multimodal LLMs for quantitative financial decision-making.
📺 Source: Algovibes · Published June 13, 2026
🏷️ Format: Benchmark Test
