This study addresses the challenge of high signal-to-noise ratios in financial sentiment analysis by introducing a hybrid, multi-stage AI framework. We combine the high-throughput capabilities of FinBERT with the deep contextual reasoning of Google Gemini to extract actionable intelligence from over 9,000,000 data points, including the U.S. Securities and Exchange Commission (SEC) filings and financial news. By applying our rigorous “Data Funnel” logic, we filter out noise from the massive dataset and surface a small set of high-conviction signals. These signals are executed on a historically dynamic universe of top S&P 500 constituents within a dollar-neutral long/short framework, integrated with macro-regime filters and technical trend confirmation. Our results over a 16-year testing period demonstrate a mean excess return of 51.02% per annum net of transaction costs, while achieving a Sharpe ratio of 1.06 and a Sortino ratio of 2.61. The significant divergence between Sharpe and Sortino ratios highlights the strategy’s positive skewness, effectively capturing upside volatility while limiting downside risk. Statistical robustness is confirmed by a Newey–West adjusted t-statistic of 4.01, indicating that the generated alpha is highly significant. This research provides a proof-of-concept for the use of Large Language Models (LLMs) as qualitative gatekeepers in quantitative finance, effectively bridging the gap between statistical NLP and human-like contextual understanding.
Stübinger et al. (2026) studied this question.