GPT-4 Turbo beats S&P 500 in AI trading study with highest returns

A new research paper from the Georgia Institute of Technology provides one of the most rigorous comparisons of large language models (LLMs) in financial trading to date. Author Geofrey Ntale evaluated five prominent models—GPT-4 Turbo (OpenAI), Claude 3 Opus (Anthropic), Gemini 1.5 Pro (Google), Llama 3 70B (Meta), and the domain-specialized FinGPT—across four structured tasks: candlestick pattern recognition from OHLCV data, directional signal generation (BUY/SELL/HOLD), simulated backtesting with execution pipelines, and financial report comprehension.

The experimental framework used quantitative metrics including Sharpe ratio, maximum drawdown, Sortino ratio, information coefficient, F1-score, and BLEU score. GPT-4 Turbo delivered the highest annualized return and Sharpe ratio among general-purpose LLMs, while FinGPT showed competitive risk-adjusted performance thanks to its domain-specific fine-tuning. Both models outperformed a passive S&P 500 benchmark in simulated backtesting. The study also identified persistent failure modes: numerical hallucination (especially in candle pattern recognition), context-window limitations when processing multi-day sequences, and inconsistent performance during sideways market regimes.

The key takeaway for professionals: LLMs hold genuine promise for AI-assisted trading, but robust deployment requires careful task decomposition, rigorous backtesting, and domain-aware fine-tuning. The paper is available on arXiv (2607.15414) as a Master's research thesis from Georgia Tech.