Bilby, a Hong Kong-based startup, is using machine learning and large-language models to analyze government documents to help investors stay ahead of regulatory and policy shifts.
The business was founded by Ryan Manuel (pictured), a China analyst and researcher at institutions such as Hong Kong University, Australia National University and Oxford, as well as serving the Australian government.
Ryan has bootstrapped the business with ‘friends and family’ money, and says he hopes to avoid venture capital. “We need to sign some contracts,” he said, noting that J.P. Morgan is piloting the service.
He added the bank is using Bilby’s services across many teams across the organization, including capital markets, asset management, compliance, government relations, and trading.
Bilby has developed an API feed for quantitative traders and hedge funds that want customized updates and analysis fed directly into their systems. Chris Lee, who recently left Hong Kong Exchanges after working there for nine years as managing director for market development, has joined Bilby as its chief commercial officer, mainly to sell the API.
Tracking China
But customers can also use Bilby as a research tool.
The idea is not new: US-listed FiscalNote Holdings was founded by college students in 2013 (in a Motel 6) to do the same thing, initially focused on Washington, DC.
Manuel says Bilby’s competitive edge is its China expertise, although he is now also building models for India and other regional markets.
Although China’s government has been reducing the data it releases, Bilby’s AI doesn’t depend on that. Rather, Bilby’s team of data scientists have built a knowledge graph of China’s political leaders, bureaucrats, provincial and city leaders, and banking and corporate executives. It deploys machine learning to determine interrelationships, both professional and personal, all of it taken from public sources such as government training manuals.
The model provides weightings to relationships and to announcements or government instructions, to try to tell the difference between actual policy directions and what’s propaganda or spin.
“We aren’t providing quants with alpha,” Manuel said. “We’re telling them who’s issuing policy, and why.”
Putting knowledge to work
Analysts already do this in a limited way, relying on their own research and data. An AI like Bilby’s does this across 66,000 organizations, tracking more than 40,000 communications made online every week. These rules, circulars, interviews – any public information online – can be modeled, and then queried by an LLM.
The system can provide insight into policy direction, help identify people involved in personal networks, and provide visibility on who knows who, and who knows what.
An LLM might do some of this work on its own, but a trained AI is meant to identify which signals are important. An LLM relies on the corpus of information found online, where as a system like Bilby’s begins with its proprietary knowledge graph. That’s the key to unveiling personal networks and career histories, and their connections to how policies are made.
Manuel says this data will always be available: it’s how governments tell people what to do.
“Governments are the cleanest source of data for machine learning,” he said. “They’re constantly broadcasting instructions.”
The competition
The story of FiscalNote could presage the direction Bilby might want to take, assuming it gets off the ground. It has acquired media businesses focused on Washington news, such as CQ Roll Call, the daily following legislation on Capitol Hill, and Oxford Analytica.
It’s also developed tools or acquired media tracking ESG- and advocacy-related news, research and analyst briefings. In 2022 it acquired Aicel Technology, a South Korean data company specialized in emerging fintech markets.
FiscalNote is backed by many luminaries: Mark Cuban gave it seed money, and today its investors are institutions including Temasek, VCs, crypto businesses, and internet businesspeople like Jerry Yang and Steve Case. It’s got Army generals on its board. It has since expanded to cover international data sets.
All of which to say is that Bilby will need to ensure its Asia models can deliver traders a better edge, as the competition is sharp. This will be even truer when Bilby looks to enter Western markets.
But it can compete on price, or with more insightful analysis into the policy-making process. Its models are also designed from the outset to operate in very different political environments and languages. And its team, experienced in working in the region, may have a better understanding of local processes and cultural norms that are relevant to the model, versus competitors that grew up focused on the US.
More interesting, although not particularly relevant to the fintech’s success, is the impact these technologies will have on the role of human analysts.
“I’m a good China analyst; this is better than me,” Manuel said. “The human becomes the face to tell the story, to develop the investment thesis, and to be held accountable.”
