For most of the past decade, a vendor specialising in natural-language processing (NLP) could sell a hedge fund something it could not easily make for itself. Sentiment scored from news, filings and earnings calls, provided as a clean feed, mapped to individual stocks and timestamped for point-in-time use. The work of pulling that signal out of the text sat with the vendor. Since 2023, that has changed. General-purpose large language models (LLMs) can now read text and return sentiment well enough that many funds assume they can build the pipeline themselves – and that assumption is changing what a data vendor can charge for.
BRAIN, a research firm founded by two computational physicists and active in alternative data since 2018, is close enough to that shift to describe it from the supply side. What its chief executive describes is not so much a threat to that business as a shift within it, away from data sold as a finished product, and towards the parts of the problem the models do not solve.What Changed When the Funds Got Their Own Models?
The shift starts with what buyers now think they can do without help. “Many clients initially assume that, with the advent of large language models, it is relatively straightforward to recreate an NLP pipeline internally,” Francesco Cricchio, chief executive and chief technology officer of BRAIN, tells Market & Alt Data Insight. On his account, the interest that once attached to sentiment feeds as standalone products has faded as that belief has spread.
His reply is not to defend the feeds but to point to where the value now sits. “What we demonstrate is that the real value lies in having datasets that are already production-ready, thoroughly validated, and easy to integrate into an existing research workflow.” The extraction, in other words, was never the hard part. The hard parts are the years of data infrastructure behind a feed, the relationships with news providers, and the assurance that a signal will behave the same way in production as it did in a backtest – none of which a general model supplies when a fund simply points it at a news source.
If the Extraction Is Easy, Where Did the Value Go?
Some of the value has moved to work that language models handle badly. BRAIN has increasingly taken on custom projects that apply machine learning to large numerical datasets, in trading and in credit-portfolio management, rather than selling standard text-based products off the shelf. Cricchio draws a firm line between the two. “While large language models have dramatically advanced the analysis of unstructured text, they are not generally the most appropriate technology for extracting predictive signals from high-dimensional numerical datasets,” he says. “Those applications still require specialised machine learning techniques, rigorous statistical validation, and domain expertise.”
That line matters commercially as well as technically. A fund may believe it can replicate a sentiment feed, but building a predictive model from raw numerical data unaided is a trickier proposition. The growing demand BRAIN sees is for exactly that job – complex numerical data turned into decision-support tools for investment and industrial use – and it is harder to commoditise because a general model cannot do it unaided.
Can Validation Be the Product?
The clearest sign of where the value has landed is that validation itself has become something firms will pay for. BRAIN built an internal cross-sectional validation platform to show off the investment uses of its own datasets, then extended it to test third-party data for others.
“The purpose of an external validation study is not to replace the client’s research process, but to demonstrate that a dataset can be integrated naturally into their existing research workflow and that it exhibits sufficiently promising historical relationships with forward stock returns to warrant further testing,” says Cricchio. Systematic funds already run their own platforms for finding alpha; they do not need an outsider to find their edge for them. What they will pay to cut is the cost of the first look – the research time and compute spent working out whether an unfamiliar dataset is worth integrating at all.
Behind the validation offer is a claim about control as well as rigour. BRAIN runs its language-model work on its own infrastructure rather than sending client material through outside providers, a choice it frames around confidentiality and governance. “In high-stakes domains such as finance, expert validation remains an essential component of any NLP workflow,” Cricchio says. For a data team weighing whether to run filings and internal documents through an outside model, where that data goes during analysis is a real concern, not a detail.
Where the argument is strongest, it rests on the buyer’s own economics. Testing a new dataset is expensive in the resources that funds guard most closely – researcher time and compute – and at first contact a prospective client often does not know a dataset’s structure, the features that can be drawn from it, or the right way to test it. A validation study lowers that barrier by showing, concretely, how the data becomes a signal.
The Narrower Business Underneath
The datasets that once defined an alternative-data vendor are becoming the entry point rather than the product, and the value has moved to the work around them: proving a signal is real, fitting it into a live workflow, and taking on the numerical problems a general model cannot. BRAIN describes its clients as predominantly medium-sized and large quantitative firms that want structured data fed straight into their research and production pipelines rather than shown through dashboards – buyers who were always going to do the modelling themselves, and who now need a vendor for the parts before and around that modelling rather than for the modelling itself.
That is a narrower, more research-like business than “alternative data vendor,” and it fits a wider move to treat sourcing as a proper institutional function rather than a procurement afterthought. Now that extraction is cheap, selling the feed alone is no longer enough. Maybe the vendors that thrive will be the ones that can show each buyer, dataset by dataset, that what they supply is worth the cost of taking it on.
Subscribe to our newsletter
