Descriptions:
Michael Automates puts OpenAI’s GPT-5.2 thinking model through four trading-specific challenges: converting a TradingView indicator to a full strategy, improving an existing strategy by adding independent short logic, generating a novel strategy from scratch, and reverse-engineering a closed-source indicator. The video references GPT-5.2’s published benchmarks — 55.6% on software engineering tasks and 70.9% on knowledge work challenges — before testing whether those numbers translate to real trading code generation.
The methodology is highly specific and reproducible: chart data is exported from TradingView as CSV files, Pine Script indicator source code is copied directly from the editor, and trade history exports are attached so the model has full context for backtesting and strategy comparison. For the short-logic improvement task, GPT-5.2 is instructed to develop entirely independent shorting logic for the existing Gshian Channel strategy (which only traded long), with an explicit goal of improving P&L while reducing maximum drawdown.
All four tasks are run directly against results previously obtained from Gemini 3 on identical inputs, making the comparison empirical rather than benchmark-only. The creator identifies at least one specific task type — a particular code challenge category — where GPT-5.2 measurably outperforms all other tested models, providing viewers with actionable guidance on when to use this model versus Claude, Gemini, or other alternatives.
📺 Source: Michael Automates · Published December 18, 2025
🏷️ Format: Benchmark Test
