In a recent episode of Goldman Sachs “Exchanges”, Raffaele Savi, Global Head of BlackRock Systematic, discussed how artificial intelligence is redefining the architecture of systematic investing.
As data, compute, and model sophistication converge, Savi argued that AI is allowing quants to move beyond traditional factor-based models and build more adaptive, higher-resolution investment systems.
For much of the past two decades, systematic strategies leaned on established risk premia—value, momentum, quality—embedded in linear, stationary frameworks. These signals, while still relevant over long horizons, often break down under stress. “The first version of quant investing—what today you’d call factor investing—definitely has a place in long-term allocations,” Savi said. “But sometimes, it doesn’t have the dynamism to navigate what could be crowding periods like 2007 or a crisis like the GFC.”
That lack of dynamism has driven BlackRock’s systematic platform toward machine learning techniques capable of responding to faster, more complex market regimes. Rather than rely solely on predefined economic relationships, AI systems can extract signal from noisy, high-dimensional, and unstructured data—often in near real time. Savi emphasized that the recent evolution of large-scale models, particularly large language models (LLMs), has made this approach not only viable but scalable.
He outlined three properties that make modern AI, and foundation models in particular, materially different from earlier generations of quantitative tools: scale, accessibility, and robustness.
Scale refers to both the size of the models and the volume of data they can ingest. Citing Rich Sutton’s essay The Bitter Lesson, Savi noted that breakthroughs in AI have consistently come from increased computation applied to general-purpose learning algorithms—not from human-designed heuristics. “If you double the amount of data you’re training your models on—or multiply it by 10 or 100—you can beat any smart adjustment to your model,” he said. This has major implications for investing, where edge often derives from the ability to identify weak but persistent patterns in vast datasets.
Accessibility is the second differentiator. Unlike earlier generations of ML models, which required specialized engineering expertise to train and interpret, foundation models can now be used directly by domain experts. LLMs, in particular, allow portfolio managers, analysts, and researchers to interact with the models in natural language. “They speak our language,” Savi said. “It allows people to interact with models they couldn’t previously access or understand.” This opens the door for iterative workflows where human and machine intelligence can be combined in real time.
Robustness, or what Savi calls “safety,” is perhaps the most critical feature when applying AI in financial settings. While LLMs and deep learning systems are not inherently better at predicting tail events, they can be used to build more resilient investment architectures. For example, models can flag when decision boundaries are uncertain, adaptively reduce exposure when signals degrade, or simulate portfolio outcomes under alternate data scenarios. “AI might not get us any closer to the crystal ball,” Savi said, “but it can create layers of safety so that events—ultimately unpredictable—don’t derail the long-term outcomes clients are counting on.”
This shift is especially relevant in the context of private markets, where traditional quant inputs like prices, volumes, and accounting statements are sparse, delayed, or missing entirely. AI enables the use of alternative data—product reviews, search trends, job postings, satellite imagery, social media signals—to fill these gaps. “We’re using Instagram posts, sentiment analysis, product reviews… all of this data is available for every company,” Savi explained. “With the models today, you can start to build frameworks that bring quant investing into markets where it was never viable before.”
However, Savi cautioned that this doesn’t mean AI systems can be treated as black boxes. Even the most powerful foundation models require contextual framing, regulatory awareness, and portfolio integration to be usable in practice. “You’ve got to work every day to make your strategies better,” he said. “It’s about implementation, not just modeling.”
When asked about the skill sets needed to succeed in this new environment, Savi pointed not just to technical depth, but to the importance of intellectual flexibility. “Some of the best people I know are bizarrely open-minded on topics they know all about,” he said. In a domain increasingly driven by empirical testing and iterative learning, that kind of adaptability becomes a competitive advantage.
As the boundaries between discretionary insight and systematic execution continue to blur, Savi sees AI not as a replacement for human judgment, but as a mechanism for scaling and refining it. The key, he suggested, lies in designing systems that continuously learn—not just from data, but from interaction with the people who use them.
