Machine learning’s role in the hunt for alpha

Our actively managed equity funds that employ machine learning (ML)—and their shareholders—are starting to benefit, according to Cesar Orosco, CFA, and Scott Rodemer, CFA, of Vanguard’s Quantitative Equity Group (QEG). The group develops quantitative models that attempt to replicate what a good fundamental investor would do, but systematically and at scale.

“There are a few ways machine learning has tended to boost alpha,” or marginal returns above those of a portfolio’s benchmark, said Rodemer, head of factor-based strategies. “For example, alpha exposures have become more dynamic; growth signals more price sensitive; and sources of excess returns more intertwined.”

As intended, neither portfolio turnover nor aggregate measures of portfolio risk, such as tracking error relative to a benchmark, have changed appreciably. 

Financial practitioners never have enough clearly actionable data, said Orosco, head of alpha equity investments. “Markets are incredibly noisy, so extracting the signal from the noise is quite the challenge.”

Through dogged research and testing, however, a QEG project team and members of Vanguard’s “AI Garage”—a community of artificial intelligence (AI)/machine learning practitioners from across the company—have found that using ML consistently added value in simulated portfolios. Complex relationships among economic and financial market conditions and the performance of individual companies and their stocks were clarified—if only a bit.  

To be sure, the introduction of machine learning to a portfolio management process is no guarantee of stronger results, and ML-based recommendations remain imperfect. “But we are a systematic manager,” said Orosco. “We diversify widely, and in our business, being right 55%–60% of the time over extended periods can add significant long-term value for investors.” 

QEG’s U.S. alpha team employs an ensemble model—really, a collection of dozens of models tracking tens of millions of underlying economic and financial market interactions. Overall, half the inputs come from the team’s traditional approach, refined over decades, and half from ML.

The ML-driven output doesn't rely on entirely new signals, however. Said Rodemer: “From the beginning we engrained fundamental reasoning into the model, an understanding of what adds value to a company and influences future returns—concepts like valuations, growth, and profitability—and embedded this intuition into a proprietary deep neural net architecture.” A key new element is ML’s continual interpretation of the macroeconomic environment, as context for company- and stock-specific changes.

While the PMs naturally remain responsible for all investment choices, they lean heavily on model signals to make virtually all buy, sell, and rebalancing decisions for three U.S. equity portfolios that are managed entirely by QEG—Vanguard Market Neutral Fund, Vanguard Strategic Equity Fund, and Vanguard Strategic Small-Cap Equity Fund. They also rely on the signals in co-managing their portions of Vanguard Explorer Fund and VVIF–Small Company Growth Portfolio. 

Before integrating machine learning into the management of Vanguard fund assets, Orosco and Rodemer and their colleagues set a high bar: They had to be able to understand why the algorithms would recommend any given investment. 

“We won’t trade on signals that we don’t understand,” said Rodemer. “Everything has to tie back to a fundamental rationale.”

That’s why the team built an “interpretability model” that gives PMs—as well as Vanguard’s independent portfolio review and risk-management teams—explanations of:

• The drivers of individual stock predictions.
• The aggregate-level interpretations of model behavior (such as the sensitivity of the signals to macroeconomic changes).
• The model health analytics (such as the volatility of scores and outliers in daily model changes), which help the team assess model inputs and outputs.

“We believe the foundations are solid, the experimentation was solid, and the improvements so far continue to be in line with expectations,” said Orosco. “So as long as that continues, we're going to be comfortable moving forward.” 

• Incorporating AI into our day-to-day (1:23 webcast excerpt, issued April 2024)
• A map to building factor-based portfolios (article, issued November 2023)
• How America innovates (15-page PDF, issued June 2022)