Introduction
Since the advent of finance, professionals have always faced the problem of how to objectively and effectively assess customer credit ratings. Taking the history of credit scoring development in Mainland China as an example: In the 1980s and 1990s, Shenzhen introduced paper loan certificates to cope with credit risk management and loan financing issues, which later evolved into credit reports of the People’s Bank of China; From 2015, the Credit Reporting Center also started providing “digital interpretation of personal credit reports” services. However, due to the lack of sustainable improvement mechanisms and other factors, it was difficult to promote [1]. After nearly 30 years of development, the rating results of these traditional credit scoring systems (hereinafter referred to as the traditional systems) are still not stable enough, and their credibility is occasionally less than optimal, which can even cause disputes, such as errors in personal credit evaluations that can occur when applying for credit cards or bank loans.
Struggles of the Traditional System
Scholars and analysts have pointed out that the traditional system, which relies on linear analysis of various indicators and subjective judgment results, has its shortcomings. For example, as shown in below example [2], the composite score of each company is the product of the score of each indicator and the reference weight, and then summed up.
| Indicator | Reference Weight | Enterprise A | Enterprise B | Enterprise C |
| Debt Asset Ratio | 50% | 9 | 8 | 8 |
| Quick Ratio | 15% | 10 | 9 | 10 |
| Operating Cash Flow | 15% | 6 | 8 | 8 |
| Executive Changes | 20% | 10 | 7 | 8 |
| Score | —– | 8.9 | 7.25 | 7.5 |
From the table, we can see that Enterprise A has the highest score, while Enterprise B has the lowest. However, if we choose a decision tree, that is, among the four indicator scores mentioned above, if one item is below 7, then the enterprise is close to default or has a low credit rating. We would then consider Enterprise C to have the highest credit rating, and Enterprise A to have the lowest, because its operating cash flow indicator score is only 6.
From this example, we find that the modeling of the traditional system is too simple and idealized, and it is far from the complex actual situation; or it relies on subjective judgment, and lacks credibility and persuasiveness. In fact, the traditional system is relatively simple and strongly dependent on basic assumptions. Once the basic assumptions are biased, the reference value of the results will be greatly reduced. Even more fatal is that rating agencies sometimes struggle to maintain the necessary objectivity and neutrality. The above example also shows that the rating results of the traditional system lack foresight or guidance.
Merging AI and Credit Scoring Systems
As early as 2018, the Security Research Institute of CAICT1 believed that artificial intelligence (AI) technology can improve the efficiency and accuracy of credit assessment and risk control [3]. When facing massive and diverse data, the traditional system’s analysis is too simple, and the amount of information it can process is limited. However, current AI technology can integrate large amounts of unconventional data such as text, pictures, videos, and quickly analyze emerging data such as consumption data and internet transaction behavior data. The credit scoring system based on artificial intelligence (hereinafter referred to as the new system) will also eliminate the influence of subjective factors to a certain extent. Scholars have also proven from an academic perspective that the new system is superior to the traditional system [4].
Many domestic and foreign enterprises, rating agencies and government institutions have already begun to promote and adopt the new system:
- Deloitte China has adopted Natural Language Processing (NLP) technology and developed a bond risk platform called “Deloitte iBonds” with real-time warning and monitoring functions. In the first 10 months of 2018, this platform achieved a 100% accuracy rate for warning alerts in the Chinese bond market [5].
- Zest AI, a company in US, has developed a new system. The approval rate for customer applications requiring credit rating has increased by 20% to 30%. With the application approval rate unchanged, the customers’ bad debts and default incidents have decreased by 30% to 40% [6]. Compared to the traditional system, the new system can find customers with better credit and help service providers better avoid defaults and other situations.
- Neurensic has utilized the new system to identify behaviors that pose risks to trading companies, while also automatically detecting and tracking high-risk activities [7].
- In November 2019, HKMA2 released two guiding documents, “High-level Principles on Artificial Intelligence” [8] and “Consumer Protection in respect of Use of Big Data Analytics and Artificial Intelligence by Authorized Institutions” [9]. These have encouraged Hong Kong enterprises to apply the new system to a certain extent.
- The ASTRI 3has also combined AI with fintech to develop new models for micro, small and medium enterprises, helping them to apply for loans [10].
Implementation of AI and Assistance of Large Language Models
While the new system has many advantages, the training of artificial intelligence requires a large amount of data and computational resources, which can be a barrier. Therefore, we can use pre-trained artificial intelligence models, such as ChatGPT, to help us develop new systems. Mainstream large language models already support document uploads and the use of the Code Interpreter plugin, which can be combined with specific requirements for analysis, generating naturally fluent results. As shown in Figure 1, the use of large language models to help build new systems mainly involves two steps: pre-processing and pre-analysis of data, and the selection and construction of models.
The following case study [11] attempts to use ChatGPT to develop a new system, analyzing the provided data and building a new system based on this. As shown in Figure 24, ChatGPT can also propose some more macro and practical analysis results and views based on the data, providing references for technical personnel.
However, new systems require more data to be more accurate and customized, some of the data may be sensitive, requiring us to carefully consider how to obtain these data in a compliant and legal manner. Furthermore, user data is first uploaded to the developer’s server, processed by the large language model, and then the results are returned. This involves a third party in the data access and analysis process. In today’s world where data privacy and security are increasingly valued, we need to consider the associated risks.
Future Prospects
We can see that enterprises need large language models with strong privacy and high performance to develop new systems. On August 23, 2023, OpenAI launched ChatGPT Enterprise [12], which provides unlimited, more powerful GPT-4 access, and data encryption transmission services, considering both the performance of the large language model and providing better choices for building new systems. Scholars have also published articles in Nature, stating that large language models, including GPT, can help optimize existing credit scoring systems [13]. We can foresee that soon; more enterprises and rating agencies will adopt the new system.
Reference
[1] “A Decade Retrospective on the ‘Private Investigation Industry Management Act'”, WANG Lu, 2023.09.01
[3] “Artificial Intelligence Security White Paper”, Security Research Institute of CAICT, 2018.09
[4] JIANG Minghui (2017). “Optimizing Personal Credit Assessment to Promote Innovation in Social Governance”. Social Governance Review. (9), 66-71
[5] “Artificial Intelligence for Credit Risk Management”, Deloitte, 2020
[6] Zest AI
[7] Neurensic
[8] High-level Principles on Artificial Intelligence, Hong Kong Monetary Authority, 2019.11.1
[11] “Using ChatGPT to Automate Your Data Analysis”, Yu Shu Zhi Lan, 2023.05.14
[12] Introducing ChatGPT Enterprise, OpenAI, 2023.08.28
[13] What large language models like GPT can do for finance, Rita Cucchiara, 2023.07.05
The work described in this article was supported by InnoHK initiative, The Government of the HKSAR, and Laboratory for AI-Powered Financial Technologies (AIFT).
(AIFT strives but cannot guarantee the accuracy and reliability of the content, and will not be responsible for any loss or damage caused by any inaccuracy or omission.)
- CAICT: China Academy of Information and Communications Technology ↩︎
- HKMA: Hong Kong Monetary Authority ↩︎
- ASTRI: Hong Kong Applied Science and Technology Research Institute ↩︎
- Original picture available at: https://sspai.com/post/79800 ↩︎
