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

Modern quantitative finance increasingly relies on complex machine learning models like XGBoost and AutoGluon to identify stock market patterns. However, these models often suffer from overfitting, where they memorize short-term market noise rather than identifying companies with genuine, long-term value. This research investigates whether applying Benjamin Graham’s classic 1940s value investing rules—such as strict balance sheet requirements and conservative valuation multiples—can act as a mathematical "low-pass filter" to constrain modern AI and improve portfolio performance.

The authors constructed three distinct feature sets: pure Graham rules, modern quantitative market factors, and a hybrid combination of both. These were tested against various machine learning architectures, including Random Forests and AutoGluon, using 20 years of S&P 500 data. To ensure real-world applicability and avoid look-ahead bias, the researchers employed a strict 80/20 temporal split, using the final four years (March 2022 to March 2026) as an out-of-sample test window. The models were evaluated based on absolute returns, maximum drawdown, and the Calmar Ratio, which measures risk-adjusted performance.

The results demonstrate that algorithmic complexity does not guarantee superior performance. While the unrestricted AutoGluon model achieved high absolute returns (222.68%), it experienced a severe 39.78% drawdown during market volatility. In contrast, the "Pure Graham" Random Forest model achieved the highest overall return (232.13%) with significantly lower risk, yielding a superior Calmar Ratio of 1.38. The study concludes that Graham’s "margin of safety" remains a powerful regularization tool, effectively preventing AI from taking on excessive risk and outperforming both the benchmark S&P 500 and more complex, unconstrained models.