Hong Kong AI start-up tracks government policy patterns to predict major regulatory changes

Why is a government like a computer program? To a Hong Kong start-up, it’s because its activities can be mapped, converted into structured data, and studied using computers – to the extent that their next move can be predicted.

Bilby, which bills itself as an AI-powered decoder of government policy, uses machine learning to analyse policies of several major economies, predominantly China. By processing voluminous data, it can generate insights about what governments may intend to do – predicting regulatory changes ahead of the event, the company says.

“Governments are always telling everybody what to do, or what they are doing, because they have to communicate at some point with the public. And that is a very rich data source,” said Ryan Manuel, the company’s founder.

“The more that we can read those communications around the world, across different languages, across different cultures, the better off we will be.”

Bilby works by collecting millions of documents’ worth of publicly available data on government policy, ranging from official announcements at all levels of government to media coverage and information on over 100,000 listed and private companies.

It gathers this information in a centralised database, augments it with relationship maps and data on key individuals, labels it, and uses it to train its algorithms, including the ability to run sentiment analyses on documents and filter out propaganda-like language. It then turns the output into software programmes, such as a ChatGPT-like function that users can query.

Generative AI takes centre stage at twin Hong Kong tech fairs

While initially focused on China, Bilby has since rolled out tools for India, the Middle East, and Nigeria, with others in the pipeline. The company says it has predicted several notable regulatory changes ahead of their announcement, including China’s 2021 crackdown on the private tuition industry, and a similar crackdown last year on the pharmaceutical sector.

Manuel, a former data analyst and consultant turned policy adviser, including stints as a China analyst for the Australian government and as professor at the University of Hong Kong, launched Bilby in 2021. He named the company after an Australian marsupial, a desert-dwelling, rodent-like creature with outsize ears that give it unusually good hearing, akin to the company’s mission of ‘picking up’ regulatory signals that might otherwise be missed.

Bilby, one of the portfolio companies of the venture capital and accelerator firm Brinc, was a finalist in the Global Scaleup Competition at last year’s Hong Kong FinTech Week, and is one of three Asian start-ups in the current cohort of Endless Frontier Labs, an accelerator firm run by New York University’s business school. Its clients include investment banks and hedge funds.

While the commercial use of AI tools trained on massive amounts of data, such as OpenAI’s ChatGPT, has increased hugely, analysts have long flagged the quality of the data they are trained on as an area that warrants scrutiny.

As “a product of its methodology and circumstances”, data can be vulnerable to multiple challenges like including infrastructure limitations, structural biases, and ethical concerns, according to a group of experts writing in an APEC policy brief.

Any AI-based tools trained on data must “consider these limitations and, where necessary, correct for any methodological or structural bias,” the authors added.

While China has been putting out less statistical information, its government is not alone in this, said Manuel, describing a decrease in reliable data as a “worldwide problem”, with Saudi Arabia and Nigeria among other countries with a paucity of data. Statistical data is also different from the government activity – and statements about what it intends to do such as documents for consultation – which Bilby tracks, added Manuel.

Despite China’s reputation for opacity, “Xi Jinping is not giving less speeches than he used to … there are constant WeChat Posts, there’s official newspapers, there’s always information coming through,” said Manuel.

“There’s a decrease in reliable data, but governments still pump out a lot of narrative. The difference is AI allows a lot more interpretation of narrative at scale.”

He added that getting the timing of changes right was often harder than spotting the actual policy moves themselves.

“That is because it’s always based on a backwards model that looks at what was the time lag the last few times something similar happened, and it’s very rarely identical,” said Manuel.

Looking forward, China watchers can expect to see crackdowns on healthcare and the financial sectors and economic stimulus through local government bonds and credit expansion, said Manuel. Pressure from other countries regarding China’s overcapacity with electric vehicles, batteries, and photovoltaics will continue, with no domestic moves to reduce output.

“But the signal that everyone’s waiting for is a clear sign on real estate, which in some ways is the dog that has not barked yet,” he said.