HKEX targets company disclosure gaps with artificial intelligence

As part of its three-year strategic plan, one of the Hong Kong Exchange and Clearing’s (HKEX’s) goals was to use technology to modernize its core functions. In the past two years, some examples of this work include its next-generation post-trade platform, stock connect programs, and the buildout of its data marketplace platform.

While work on some of those projects is ongoing, one of the solutions that HKEX has already put in place is an artificial system called Jura, which is intended to improve corporate governance and disclosures.

There are more than 2,500 listed companies on HKEX, and each issuer is required to publish an annual report, presenting their financial results, business performance, and management commentary to the general public. As the primary regulator of stock market participants, HKEX monitors issuers’ annual reports and other commentary documents and checks if they are compliant with the exchange’s listing rules.

Previously, HKEX conducted this exercise on a thematic basis, focusing on major topics, such as environmental, social, and governance (ESG) compliance. It would then issue guidance and interpretation to issuers on what companies were reporting well and what they weren’t, so the next time they reported, issuers could correct those shortcomings.

In mid-2020, HKEX introduced Jura within its listing division to determine if issuers are compliant, and automatically review all annual reports.

Dealing with Unstructured Data

HKEX and its listing division built Jura in collaboration with Beijing Paoding Technology (PAI Tech), an artificial intelligence (AI) company specializing in financial semantic understanding, including document intelligence and regulatory technology. Together they used a combination of natural language processing (NLP) and deep learning techniques to create models that can read, understand and interpret all elements of an annual report, often in varying, unstructured formats.

“HKEX refined some existing natural language algorithms specifically for this regtech use case. You’ll find a lot of people in the market are using NLP to look at company disclosures—that’s nothing new. But usually, people look for specific things like the dividend payout, what was the date of the annual general meeting, or the bond prospectus, where the information is kind of structured,” Lukas Petrikas, head of HKEX’s Innovation and Data Lab, tells WatersTechnology.

Annual reports, however, are very much unstructured documents, as companies can produce them in whatever formats they like. First-time listers, especially, can take creative approaches to these documents.

“You’ll see things like mascots with speech bubbles, or charts that are very colorful. It’s not like there’s a standard for how you say certain things. You can use thousands of choices of words to say the same thing. And sometimes it’s a graph, sometimes it’s text, sometimes it’s a picture, and sometimes it’s a chart,” Petrikas says.

The system downloads annual reports and supplementary announcements, locates disclosures within the content corresponding to each listing rule, and deduces whether issuers are compliant. It then retains the assessment and compliance analysis.

Petrikas says its developers trained Jura for eight to nine months by tagging new reports and exposing the model to different listing rules. HKEX used more than 3,800 English language annual reports and more than 400 items of corporate communication to train the system to assess 140 different types of corporate disclosure and compliance. Jura can now find inconsistencies in annual reports and issuers’ communications in company announcements.

Petrikas says it’s the “silences” in the annual reports that are most difficult for AI to pick up. “Whether they’re false negatives or true negatives, you don’t know unless you check some other document. For example, if a company says, ‘During this year, the company did not issue any warrants for shares’—how do we know if that’s true? What if somebody remembered that actually, back in May, the company announced it had issued some warrants to its senior employees, and they forgot to mention it in the annual report?”

Mistakes do happen, especially with new issuers and companies that are disclosing for the first time. The annual report is a big document, often 300 to 400 pages long, and its writers often omit important details. “Our job is to find that,” Petrikas says.

In that kind of situation, Jura would send an alert that the company had in fact issued a warrant in May, and return the announcement so that a human reviewer could look into it.  

Benjamin Quinlan, CEO and managing partner of Quinlan & Associates, a Hong Kong-based strategic consulting firm, says while he doesn’t know the specific mechanics of Jura’s algorithm, it will deliver huge efficiencies if it can identify reporting discrepancies with greater speed and accuracy.

“Accelerating corporate governance checks is a major plus for companies, especially when the whole industry is very manual, racking up huge bills with lawyers, accountants, and others. Moreover, many current processes are prone to human error,” he says.

However, the strength of the algo driving the technology is key, he says. If it’s not well-trained, it can create a lot of unnecessary “red flags” that end up producing more work for the user in the end.

“It will take time for these solutions to show their worth. The use of NLP, optical character recognition and other forms of AI helping to cut through laborious human tasks are becoming common practice across many industries. I see this as a positive development overall, as it allows a company’s employees to focus their efforts on higher, value-added activities while streamlining compliance costs,” Quinlan says.

Beyond Disclosures

HKEX’s Petrikas says Jura can also be deployed elsewhere within the exchange. Extending beyond the current use case will involve more training.

“[Jura] is something that we can apply across other use cases because the examples I used are as hard as it gets,” he says. The nirvana for NLP models is having a high recall and high precision rate, he adds; it’s usually a trade-off between the two. 

“You might achieve high precision and low recall, and a human has to check. Or you have high recall and low precision, in which case the system probably picks up a lot, but it’s hard to say when it didn’t pick it up. As a regulator, it’s the things that you might have missed that’s always worrying. So to get to a place where we have both high recall and high precision—within the 90% range—with the system has been a huge achievement,” he says.

According to a HKEX case study detailing the use of the Jura AI system to assess the annual reports of listed companies, the overall accuracy rates for location of annual report disclosures and issuer compliance recommendations for the training set reached 90% and 92%, respectively.

It then tested the model against 50 previously unseen reports, and the accuracy rates were 84% and 85%, respectively. HKEX aims to improve performance through a regular review of data generated from user verification.

“The amount of learning we had in terms of applying AI, I don’t think we could have picked a better use case [for regtech], because it was so challenging to get it right. Now that we’ve got it right, other use cases are easy. So now if we want to look at the simpler stuff like companies announcing convertible bond issuance—looking at the conversion price, the maturity, the coupon—that’s very easy because we’ve now done something that’s much harder to capture,” Petrikas says.

The system would need a little bit more training on, for example, convertible bonds. “And then, the infrastructure is there, the dashboards, the analysis, it’s all there. So it’s really just training the AI model incrementally,” he says.

Future areas of research include extending the platform to results announcements and other types of regular corporate communications. That said, HKEX is also committed to refining and enhancing Jura in the current use case.