Quant Convergence: Bridging Classical Value Investing and Modern Factor Models for Systematic Equity Selection

A new study finds that Benjamin Graham’s classic value-investing rules can act as a guardrail for modern machine-learning models, preventing them from chasing risky market noise and reducing catastrophic drawdowns during turbulent periods. The paper, posted to arXiv on June 23, 2026, tested whether Graham’s principles could serve as a mathematical “low-pass filter” for complex algorithms [1]. Researchers built three feature sets—pure Graham rules, modern market factors, and a combination of both—and evaluated them against XGBoost and AutoGluon models using 20 years of S&P 500 data [1]. A strict buy-and-hold strategy was applied over a four-year test period from March 2022 to March 2026 [1]. The pure Graham Random Forest delivered the highest overall return of 232.13% with a Calmar Ratio of 1.38, indicating substantially lower risk [1]. By contrast, the AutoGluon model posted a 222.68% return but suffered a 39.78% drop after purchasing volatile technology stocks just before a market crash [1]. The Combined Random Forest, which blended momentum signals with Graham’s rules, returned 202.91% while recording the lowest maximum drawdown of any model tested, at 34.53% [1]. The results suggest that more complex algorithms do not always win; Graham’s “margin of safety” remains an effective mechanism for curbing AI-driven risk-taking [1]. The study’s approach contrasts with pure technical analysis, which attempts to forecast price direction from past market data and whose efficacy is disputed by the efficient-market hypothesis [3]. By embedding fundamental-valuation constraints directly into the feature-engineering stage, the researchers forced models to prioritize companies with durable value rather than short-term price momentum [1][3]. The paper was submitted by Hugo Garrido-Lestache Belinchon and is available as a preprint, meaning it has not yet undergone formal peer review [1]. Preprint servers such as arXiv assign a Digital Object Identifier to each submission, making the work citable while it awaits journal evaluation [4].