The Hong Kong Monetary Authority (HKMA) announced in November that it has developed a two-year roadmap to promote RegTech adoption in the Hong Kong banking sector. The HKMA cited “banks’ evolving business models, regulatory initiatives in response, and a challenging external environment” as factors driving the need for fresh technology solutions.

HKMA’s assessment holds equally true for capital markets. In securities regulation, changes include the growing number and diversity of listed companies, the increasing complexity of required disclosures (for example, on ESG topics), and a robust pipeline of first-time public market issuers fresh from a recent IPO.

All of these trends point towards a need for more sophisticated and responsive regulatory oversight tools. The adoption of digital standards and automation to improve regulatory capacity and effectiveness is already a minimum expectation. Experiments with ever more sophisticated regulatory AI, meanwhile, are rightfully gaining momentum.

Like it or not, data-heavy documents remain the lifeblood of regulatory practice, even as investors increasingly use real-time predictive analytics and market sentiment. Financial disclosures contain a treasure trove of crucial insights into the health and activities of reporting companies.

There have been efforts over the years to standardise financial reporting and make the volumes of data therein more accessible. Even the optimists, however, will admit that conveying a true and faithful picture of the quarterly ups and downs of the fortunes of a business takes more than a checklist.

The tone of the management’s narrative, an innocuous-looking acquisition, an easy-to-miss but crucial footnote, the biography of a new director – a few sentences might unlock essential insight into the affairs, stability and potential risks of a regulated company. Reading, contextualising and interpreting corporate filings is therefore a continuing focus of frontline RegTech efforts around the world.

AI can help regulators digest more information in less time, build data-rich knowledge maps of their markets, zero in on areas that require surveillance focus, and detect themes to guide their own strategy and resourcing. If well designed, an AI model’s accuracy improves with time, as it learns from its errors with some guidance from its teachers.

This is easier said than done. Even the most comprehensive regulatory filings are unstructured. Moreover, reading pages like a human is not enough. It is important to recognise the difference between disclosure and compliance. Determining whether what is written is accurate, or if required information has been excluded, is a formidable challenge that requires training AI models in the kind of contextualisation that can evade even the most experienced of human regulators.

Therefore, few RegTech AI models have so far succeeded in taking over human regulatory workloads in a consistent and rigorous manner.

At Hong Kong Exchanges and Clearing (HKEX) we set out to build an AI algorithm that could not only read unstructured data from the most complex of corporate filings, but also infer sufficient context to make an assessment about the substance of those documents.

The documents we chose were annual reports. We monitor annual reports of more than 2,500 companies listed on our markets to check that each company is disclosing the relevant information that the Listing Rules require of them. Historically, this has been done manually, using a thematic and sample-based approach.

We trained our AI to look for content which was rarely straightforward. For example, listed companies must disclose the attendance record of their directors in board meetings during the year. This information may be presented in a graphic, next to a board member’s photo, in prose, a table, or it may be missing from the report.

A change in board membership during the year could explain a poor attendance record, but so could a director’s absenteeism. We trained the model to read and analyse attendance data, as well as information related to more than 100 different Listing Rules.

The resulting platform has boosted the breadth, speed and accuracy of our ability to assess annual reports for compliance, leading to significant savings in manpower.

This AI model will help HKEX create even more valuable and comprehensive analyses, which can translate into effective guidance for how listed companies can improve their disclosures, and ultimately ensure higher quality information for investors. We plan to continue exploring further applications of our new system.

Our RegTech experiment taught us that by doing certain tasks previously entrusted to humans, AI can improve transparency, reporting quality and confidence in the securities markets, in line with our mission to ensure a resilient and robust market for all.