🤖 AI

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

This research demonstrates that integrating Benjamin Graham's classic value investing principles with modern factor models effectively acts as a risk-control mechanism, enabling a hybrid Random Forest strategy to outperform complex AI models like AutoGluon by delivering superior returns with significantly lower drawdowns over a four-year period.

Original authors: Augusto Eiji Yamazaki, Hugo Garrido-Lestache Belinchon

Published 2026-06-24

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Augusto Eiji Yamazaki, Hugo Garrido-Lestache Belinchon

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the best apples in a massive, noisy orchard (the stock market). For decades, the smartest people have been using two very different tools to do this.

Tool 1: The Old-School Farmer (Benjamin Graham)
This is the approach from the 1940s. It's like a strict checklist. The farmer only picks apples that are:

Big and sturdy (not tiny saplings).
Have plenty of water in their roots (cash on hand).
Have a history of growing fruit every year (consistent profits).
Are sold at a fair, low price (not overpriced).

The farmer ignores the flashy, expensive "super-apples" that might grow fast but could rot in a storm. This is a low-pass filter: it blocks out the noisy, risky stuff and only lets the solid, safe stuff through.

Tool 2: The High-Tech Drone (Modern AI)
This is the new approach. It uses complex computers and machine learning (like XGBoost and AutoGluon) to scan the whole orchard. It looks at thousands of data points: how fast the wind is blowing, the color of the leaves, the history of the soil, and even the mood of the weather forecast. It tries to find hidden patterns that humans can't see.

The Experiment
The authors of this paper wanted to see which tool actually works better when the weather turns bad. They set up a race using 20 years of data from the S&P 500 (the biggest orchard in the US). They tested five different teams:

The Pure Farmer: Only using the old 1940s rules.
The Pure Drone: Only using modern, high-tech data.
The Hybrid: Trying to mix the old rules with the new tech.
The Super-Drone: A very complex AI that stacks many models together.
The Benchmark: Just buying everything in the orchard (the S&P 500 index).

They ran this race through a specific four-year period (2022–2026) that included a massive market crash (a "storm") followed by a recovery.

The Results: Who Won?

The High-Tech Drone (AutoGluon): This team made a lot of money (222% return), but it was a rollercoaster. Because it chased the "flashy" tech stocks, when the storm hit, it lost nearly 40% of its value. It was like a race car that goes fast but crashes hard.
The Pure Farmer (Graham Rules): This team made the most money overall (232% return). Even better, when the storm hit, they only lost 35%. They didn't crash as hard because they were holding sturdy, cash-rich companies.
The Hybrid Team: When they tried to mix the old rules with the new tech, the results got worse. The computer got confused, started chasing the risky stocks again, and made less money than the Pure Farmer.
The Super-Drone (XGBoost): This one failed completely in the test, making less money than just buying the whole market. It had "memorized" the past too well and couldn't handle the new storm.

The Big Lesson
The paper concludes that complexity isn't always better.

Think of the old 1940s rules not as "outdated," but as a safety harness. When you strap a high-powered AI engine into a safety harness (the Graham rules), it performs better than if you let the AI run wild.

The "margin of safety" that Benjamin Graham talked about isn't just a nice idea; it acts like a mathematical filter that stops the AI from getting too excited and buying risky, volatile stocks right before a crash.

In short:

Complex AI is like a race car: fast, but dangerous in a storm.
Old-school Value Investing is like a sturdy truck: slower to start, but it drives you safely through the storm and gets you to the finish line with more money in your pocket.
Mixing them is like putting a race car engine in a truck; it just makes the vehicle unstable.

The study proves that in the noisy, chaotic world of stocks, sticking to the simple, boring, fundamental rules of "buying good companies at a fair price" is actually the smartest way to use modern computers.

Technical Summary: Quant Convergence

Problem Statement

Modern quantitative finance increasingly relies on complex machine learning (ML) models, such as XGBoost and AutoML ensembles, to identify patterns in equity markets. However, the authors posit that as these models increase in complexity, they risk "overfitting" to short-term market noise and hype rather than identifying companies with genuine, long-term value. This leads to portfolios that may generate high absolute returns during bull markets but suffer catastrophic drawdowns during market shocks due to a lack of structural risk constraints. The core research question is whether Benjamin Graham's classical, rule-based value investing framework can serve as a mathematical "low-pass filter" to regularize modern AI models, preventing them from taking excessive risk while maintaining or improving predictive performance.

Methodology

The study employed a rigorous, 20-year backtesting protocol (March 2006 – March 2026) on the S&P 500 universe, utilizing a strict 80/20 temporal split to ensure no look-ahead bias. The training window covered 4,021 days, while the out-of-sample test window covered 1,006 days (March 2022 – March 2026), a period characterized by severe macroeconomic contraction followed by a concentrated rally.

Data and Feature Engineering

The researchers constructed three distinct feature sets to test against five different model architectures:

Strategy A (Pure Graham): 19 features based strictly on Graham's seven defensive criteria (e.g., Current Ratio ≥ 2, P/E ≤ 15, P/E × P/B ≤ 22.5), encoded as binary thresholds and supplemented with fundamental ratios.
Strategy B (Modern Quant): 15 features derived from vectorized operations, including risk-return metrics (CAPM Alpha, Beta, Sharpe), momentum signals, and a Benford's Law data-integrity signal.
Strategy C (Combined Hybrid): An aggregation of 36–40 features, including cross-domain interaction terms (e.g., Value × Momentum).

Model Architectures

Five models were evaluated using hyperparameter optimization via Optuna (Tree-structured Parzen Estimator) and TimeSeriesSplit cross-validation:

Random Forest (RF): Trained on Strategies A, B, and C.
XGBoost: Trained on the Combined Hybrid dataset (Strategy D).
AutoGluon: An automated ML framework configured to stack multiple models (Strategy E).

Execution Protocol

The study simulated a strict "buy-and-hold" strategy. At the start of the test period, each model selected the top 20 equities based on predicted outperformance probability. These were purchased with equal weights and held without rebalancing for the entire four-year duration.

Key Results

The out-of-sample results challenged the assumption that higher algorithmic complexity and unrestricted data yield superior risk-adjusted returns.

Absolute Returns: The Pure Graham Random Forest (Strategy A) achieved the highest total return of 232.13%, outperforming the S&P 500 benchmark (68.00%) and the unrestricted AutoGluon model (222.68%).
Risk and Drawdown: While AutoGluon generated high returns, it suffered a severe 39.78% maximum drawdown during the 2022 market shock, significantly worse than the benchmark's 24.00%. In contrast, the Pure Graham model experienced a drawdown of 35.01%, and the Combined Hybrid RF (Strategy C) achieved the lowest drawdown of 34.53%.
Risk-Adjusted Performance: The Pure Graham model achieved the highest Calmar Ratio (1.38), indicating superior return per unit of downside risk. The Combined Hybrid RF (Strategy C) followed with 1.22, while AutoGluon dropped to 1.16.
Complexity Penalty: When modern quantitative factors were mixed with Graham's rules (Strategy C), the total return dropped to 202.91% compared to the pure Graham model's 232.13%. Furthermore, the XGBoost model applied to the combined dataset (Strategy D) collapsed to a 56.79% return, underperforming the benchmark, indicating severe overfitting.
Statistical Significance: A Welch's t-test confirmed that the Pure Graham strategy's outperformance was statistically significant at the $\alpha = 0.10$ level ( $p = 0.098$ ) against the S&P 500. Conversely, the AutoGluon model failed to reject the null hypothesis ( $p = 0.106$ ), suggesting its returns were mathematically indistinguishable from random variance given its high volatility.

Significance and Claims

The paper claims that Benjamin Graham's "margin of safety" is not an outdated concept but a highly effective regularization technique for modern algorithmic trading.

Low-Pass Filtering: The study demonstrates that restricting ML models to classical balance sheet metrics acts as a mathematical low-pass filter, effectively removing financially fragile companies and preventing the model from overfitting to short-term momentum or volatility.
Dilution of Value: The research suggests that attempting to "modernize" Graham's rules by adding contemporary momentum factors actually dilutes their predictive power. The pure fundamental logic proved structurally superior to hybrid approaches in this specific market regime.
Structural Survival: The findings indicate that in the intersection of AI and finance, the most robust models are those anchored in fundamental economic reality. Complex, unrestricted ensembles may capture upside momentum but fail to provide structural survival during market shocks, whereas constrained fundamental models offer a superior risk-return profile.

The authors conclude that fundamental constraints are necessary to isolate true, statistically significant alpha, transforming Graham's 1940s defensive framework into a vital tool for managing risk in high-dimensional machine learning systems.