On September 26, 2023 International Forum for China Financial Inclusion & Academic Summit of Digital Economy Open Research was held in Beijing. Zhang Xiaoyan, Professor and Associate Dean of Tsinghua University PBC School of Finance delivered a keynote speech entitled "Application and Challenge of Artificial Intelligence in the financial Field - Taking Large Language Model as an Example".
Professor Zhang summarized the application and challenges of LLM in the finance by cases. The LLM has shown surprising breakthroughs in natural language processing, whose outstanding ability in data analysis and decision support will promote the improvement of global productivity.
Zhang Xiaoyan giving a keynote speech
A Large Language Model is a large artificial intelligence model for natural language information processing. LLM is trained on the basis of large-scale text data to learn syntax, semantics and context information of the language, in order to understand and generate human language and perform multiple tasks, including Q&A, sentimental analysis, text generation, knowledge extraction, as well as abstract and summary. As a significant feature of LLM, the number of large parameters of Open AI GPT3.5 reaches 175 billion, and that of GPT4 is expected to reach 1 trillion. The closed-source version of InternLM from Shanghai AI Lab in China has 104 billion parameters, which is trained on a corpus containing 1.8 trillion tokens. The large number of parameters has brought breakthroughs in many of its capabilities. Zhong et al.(2023) showed that the ability of LLM has exceeded human levels in examinations, including SAT, GRE, China College Entrance Exam, bar Exam, and National Civil Service Exam.
Zhang Xiaoyan giving a keynote speech
Professor Zhang believes LLM technology is improving global economic production efficiency. A survey released by McKinsey in 2023 shows that LLM is being applied in many industries to improve industrial efficiency. Of all industries, generative AI is applied most in financial services and technology, while it is applied most in marketing, customer service management, research and development, and software development of all enterprise departments. LLM will contribute to economic growth. Its application in industries will lead to improvement in societal productivity, and will drive transformation in the job market, where some simple text processing jobs will be replaced and new jobs will open up.
China maintains a positive attitude and encourages artificial intelligence technologies such as LLM. The industry market is in orderly development and supervision is under continuous improvement.
Professor Zhang believes LLM’s potential impact in financial industry may be greater. In investment scenarios, LLM can be used for investment decisions, risk assessment, market analysis, document processing, automated customer service and public opinion analysis. Take macroeconomic analysis for example, the attitude of the US central bank towards monetary policy can be analyzed based on the LLM. In banking services, the application of generative AI run through all links of the industrial chain. Compared with other industries, the financial industry needs to process a variety of information, including news, analyst research reports, government policies, where LLM can extract and analyze massive texts more conveniently. The timeliness of financial information requires analysts to make quick decisions, and LLM can analyze the market and make suggestions in seconds. In addition, a large number of financial services require language communication. In the future, LLM may be able to answer customer inquiries, provide investment advice, and even partially replace human customer service and investment consultants.
Under the rapid development, the application of LLM in the financial field also faces some technical challenges. The lack of knowledge in the financial field (the LLM in the financial field is insufficient corpus). The cost of computing power required by LLM training is high. The financial information requires time efficiency while the training corpus of large language model may have a lag problem. The accuracy requirement of financial decision-making is high. There is also high dynamic problem in the financial field. For example, the professional vocabulary in the financial field is increasing dynamically, and the same term may have different meanings at different times.
Some solutions can address the current challenges. One is to re-train the LLM in the financial field based on the financial corpus; second, based on an open source large language model, the ability of the LLM is aligned with human needs through different fine-tuning techniques, and then a financial LLM in line with specific scenarios is built; thirdly, the word vector database and Langchain technology are combined to update the word vector database in time to make the LLM have the latest and accurate information.
Professor Zhang believed that although the LLM just emerged not long ago, it is placing a positive impact on the global and Chinese economy. Many business scenarios in the financial industry are very suitable for its application and implementation, despite many challenges. In order to standardize the healthy and orderly development of LLM in the finance, relevant regulatory authorities are needed to formulate rules and regulations in time to guide the healthy development of the industry.