Page 22 - Annual Report 2022 EN
P. 22

/ 22                                                                                                                                                                                                                                                                                                                      T ABLE OF C ONTENT S





                   The QFCRA’s Supervision and Authorisation division consists of the Authorisation, Financial Analysis and


                   Innovation (FAI), Bank and Insurance Supervision (BIS), Investment Manager, Advisor and Securities Supervision

                   (IMAS), and Anti-Money Laundering/Combating the Financing of Terrorism (AML/CFT) Supervision departments.






                   Authorisation





                   The Authorisation department is responsible                                               2023 statistics                                                                            SupTech and Innovation                                                                     The emergence of powerful LLMs also opened

                   for assessing and approving applications from                                                                                                                                                                                                                                   new avenues for the Supervisory Master

                   firms and individuals seeking authorisation from                                          Licenced Firms                                         2023                                The launch of OpenAI’s ChatGPT in November 2022                                            Data Management (MDM) system that was

                   the QFCRA to conduct regulated activities in or                                                                                                                                      and the chatbot’s upgrade to the GPT-4 model in March                                      being built over the course of the year.
                                                                                                             Regulated                                                65                                2023 meant a step change in the applicability of Artificial
                   from the QFC. The department also acts as the

                   interface between applicants and the Companies                                                                                                                                       Intelligence (AI) in processes across industries. Due to                                   Not only can the resulting supervisory Knowledge
                                                                                                             Non-regulated                                          1670                                the high semantic precision of the underpinning Large                                      Graph (KG) be used to ground a system with a natural
                   Registration Office of the QFCA regarding their
                   licensing or incorporation in the QFC.                                                                                                                                               Language Models (LLMs), many previously specialised                                        language user interface, the LLMs can be used to
                                                                                                             Total                                                  1734                                applications of Natural Language Processing (NLP) and                                      extract structured information from the text and

                   In 2023, the department met with prospective                                                                                                                                         Natural Language Understanding (NLU) could now be                                          populate the KG in a more automated manner than

                   applicants, reviewed their submitted regulatory                                                                                                                                      replaced by a single model. But while such models can                                      was previously possible. Work on both solutions
                   business plans, and involved the Supervision                                                                                                                                         retrieve and reproduce information from their training                                     continued at the end of the year, and are expected to

                   departments to ensure appropriate issues                                                                                                                                             data, solutions requiring high quality responses and                                       be ready to support day-to-day supervision starting

                   management after authorisation. During the year,                                                                                                                                     up-to-date information will still need to be grounded                                      in 2024. Outside of the two LLM-related projects,

                   1 fund manager, 1 investment advisory firm, and                                           Approved Individuals                                   2023                                to a datastore that is specific to the use-case.                                           FAI continued working with the supervisory teams

                   1 representative office were authorised.                                                                                                                                                                                                                                        to improve overall data processing and automating
                                                                                                             Total                                                   437                                Over the course of the year, QFCRA’s Financial Analysis                                    production of management information. The department

                   The number of individuals approved to                                                                                                                                                and Innovation Department (FAI) explored the potential                                     also increased its efforts in engaging supervisors in

                   perform controlled functions at authorised                                                                                                                                           of LLMs for supervision with two prototype solutions.                                      conversations about data to identify opportunities

                   firms decreased by 3 to 437 by year-end.
                                                                                                                                                                                                                                                                                                   and to upskill teams in data tools and theory.
                                                                                                                                                                                                        The first FAI application is a solution for AML
                                                                                                                                                                                                        compliance analysis. The tool was preloaded with

                                                                                                                                                                                                        AML regulations and laws on the one hand and

                                                                                                                                                                                                        individual supervised institutions’ relevant policies

                                                                                                                                                                                                        and procedures on the other. The LLM was prompted
                                                                                                                                                                                                        to answer a set of questions from a compliance

                                                                                                                                                                                                        perspective and return an analysis with references to

                                                                                                                                                                                                        pages and paragraphs from the source documents. The

                                                                                                                                                                                                        experience is promising – suggesting that, when used
                                                                                                                                                                                                        appropriately, such a tool can speed up such analysis

                                                                                                                                                                                                        significantly without sacrificing supervisory quality.
   17   18   19   20   21   22   23   24   25   26   27