Features Research at Two Sigma
“What is a feature” is the type of philosophical question we pontificate over lunch, up there with “what is a factor”. To me, features are simply facts about the world worth noting. There are the obvious ones: a stock’s price movement over the past week, a company’s level of news coverage relative to its history, or even the specific day of the week a press release was published. But then there are also the more subtle ones: how many times did the CEO say “um” in that conference call, how many cars are parked in a store’s parking lot, or just how big were the discounts this year at a retail store? What we do at Two Sigma is not just collect data for data’s sake. Rather we layer human knowledge and intuition on top of the process, thinking creatively about what any stream of data can reveal and then shaping that data into predictive features by asking differentiated questions. And ultimately, it’s the creativity of those questions that drive our success.
In other words, as raw data has become more widely available and commoditized, the real edge has migrated from simply what data you have, toward what you actually do with that data.
On the Complex Features Engine team at Two Sigma, that work happens through close collaboration between researchers, dedicated feature engineers and a broader set of engineers that build the tools to make it happen. We intentionally bring in smart, curious minds from multidisciplinary, diverse backgrounds. The result of this diversity is a long list of individually small signals, each of which may seem small, but like dots of a Seurat painting, start to paint a clearer picture when you put them all together.
Creative Science: Rigor Without Rigidity
I’ve always believed that feature research is as much a creative exercise as it is a scientific one. Our process is often grounded in hypothesis formation: you start with a prior about why something might matter—and then you test it. If the data doesn’t confirm your prior, you don’t try to rationalize it away. You simply move on as this scientific method dictates.
We do this work repeatedly, each time with a differentiated idea – maybe a new data set, a new algorithm or a new take on how we put it all together. These differences we inject in the process from our varied background help drive orthogonality, which is the lifeblood of the alpha modeling process. Rather than spending a disproportionate amount of time trying to optimize a single model to perfection, we champion a diversity of approach where we’re constantly sourcing new ideas and hypotheses to test.
With AI, Anything Can Be Language
My background is in natural language processing (NLP), which means I’ve been thinking about text as data for over 15 years. Large Language Models (LLMs) completely change the equation because they make it possible to generate textual data, not just collect it. Each output an LLM produces can itself become a new data set to analyze.
For a firm like ours that has spent well over a decade studying how text can help predict markets, the arrival of models that can produce rich, queryable text about virtually any company or event represents a dramatic expansion of the fuel that drives many of our models, creating an unprecedented expansion of the research landscape.
Equally important is what LLMs do to the economics of exploration. Ideas that previously would have required months of upfront technical work—such as analyzing CEO microexpressions during earnings calls—can now be explored far more quickly and cheaply. This means our researchers can finally work through a backlog of great ideas they had previously shelved as too costly to pursue, again leading to a sort of renaissance of productivity.
Ultimately, for AI in an industry like ours where diversity of ideas is the point, it is important to think of these tools as an amplification of what we’re good at as humans, which is originality and orthogonality. There are certainly risks to over-automation; if designed poorly, AI tools could produce increasingly homogeneous results across researchers who might otherwise bring distinct perspectives to the same data. Our goal at Two Sigma is to build tools that make each researcher more themselves—not tools that converge everyone toward the same answer.
Listen to the Full Conversation
Listen to the full Flirting with Models episode to hear more on the mechanics of features research, our approach to AI infrastructure at Two Sigma, and what makes a great quant in the age of LLMs.