Quant Convergence: Graham Rules Outperform AutoGluon on S&P 500

Quant Convergence, a paper by Augusto Eiji Yamazaki and Hugo Garrido-Lestache Belinchon submitted 23 June 2026, tested whether Benjamin Graham's classic value rules can restrain modern machine learning strategies on S&P 500 stocks. The authors trained models on 20 years of S&P 500 data and evaluated buy-and-hold performance over a four-year test period from March 2022 to March 2026.

What did the paper test and how?

The authors built three feature sets — pure Graham rules, modern market factors, and a combined mix — and ran highly complex algorithms (XGBoost and AutoGluon) plus Random Forest variants on 20 years of S&P 500 history, then applied a strict buy-and-hold strategy across March 2022 to March 2026 to compare realised returns and risk. The goal was to see if Graham's heuristics act as a mathematical "low-pass filter" to keep modern models from overfitting short-term noise.

How did the models perform in the March 2022–March 2026 test?

The headline numbers show the pure Graham Random Forest produced the highest total return at 232.13%, with a Calmar Ratio of 1.38, the AutoGluon model posted 222.68% total return but experienced a 39.78% maximum drop after buying volatile tech stocks just before the market crash, and the Combined Random Forest delivered 202.91% total return while recording the lowest maximum drop in the study at 34.53%. The paper therefore reports that the more complex AutoGluon capture of high returns came with a larger realised drawdown, whereas Graham-based signals yielded comparable or higher returns with lower measured risk in those experiments.

What exactly were the key metrics reported?

The study highlights three concrete outcomes from the four-year test window: AutoGluon returned 222.68% and suffered a 39.78% maximum drop; the pure Graham Random Forest returned 232.13% with a Calmar Ratio of 1.38; and the Combined Random Forest returned 202.91% with the lowest observed maximum drop of 34.53%. The paper frames these results as evidence that Graham's margin-of-safety rules can reduce risk-taking by modern algorithms.

Why it matters

The findings challenge the simple idea that more complex models automatically produce better real-world performance. In this experiment the highest-returning algorithm was not the most complex model in every respect, and the model that captured strong returns (AutoGluon) also concentrated into volatile positions and suffered a larger drawdown. For portfolio builders and quant researchers, that implies feature design — specifically value-based filters inspired by Graham — can materially change the risk-return profile, not just marginally improve backtest statistics.

What to watch

Look for replication of these results beyond the S&P 500 and for out-of-sample live trading evidence that the Graham-based filters reduce catastrophic drawdowns. Another concrete milestone will be comparisons that include the paper's XGBoost runs and fuller risk-return tables for alternative market regimes to see whether the reported patterns hold across different stress periods.

References: Quant Convergence: Bridging Classical Value Investing and Modern Factor Models for Systematic Equity Selection, Augusto Eiji Yamazaki and Hugo Garrido-Lestache Belinchon, submitted to arXiv 23 June 2026.