Giorgio Baldassarri Hoger von Hogersthal
Global Head,
Analytical Innovation & Development Group,
S&P Global Market Intelligence
Published: October 10, 2023
Artificial intelligence offers many benefits to companies in terms of automation, risk management and efficiency gains, but it needs to be employed carefully to mitigate potential challenges, especially when dealing with small datasets.
The use of artificial intelligence, and machine learning techniques in particular, is very promising in the field of credit scoring and modern portfolio surveillance, where it can help with detailed analysis and interpretation of large datasets.
We believe the use of artificial intelligence, in connection with firms' alternative datasets (i.e., digital fingerprints) can help refine the credit risk assessment and generate more accurate and timely signals for credit risk management and investment purposes.
Artificial intelligence (AI) has risen to the top of corporates' agenda as a versatile tool to automate repetitive tasks, speed up and scale the analysis of big, often unstructured and noisy datasets, and distill relevant information into actionable insights, be it for corporate strategy, operational, investment or risk management purposes.
But what does AI really think of itself? Figure 1 shows the pros and cons that OpenAI's ChatGPT summarized for us when prompted with the following question: "Should AI be used by companies?" None of the answers came as a surprise. However, one critical aspect — "Works poorly with little data" — was omitted by ChatGPT, and we added this item to the graphic for completeness. AI does not work well in the absence of large and relevant datasets. It seems an obvious pre-requisite for harnessing AI's power, but it is often neglected by human practitioners and, thus, by AI itself. In fact, this omission points to the intrinsic limitation of current AI-based technologies: none of them can come up with an original contribution, i.e., something beyond the data they are fed. The role of model creators and users of AI techniques thus remains central, and these individuals must ensure that AI produces actionable outputs that do not defy intuition and general understanding.
The challenge of data, in terms of both quantity and relevance, and the related quality of AI-based systems surfaces in all its glory when applying machine learning techniques in predictive credit risk analytics, for example, for the generation of early warning signals of potential creditworthiness deterioration of a credit risk portfolio. To mitigate these issues, early warning frameworks used by credit risk managers at financial and non-financial corporations tend to include multiple indicators that can be unreliable. These factors — and their accompanying issues — include:
Market signals based on equities, bonds and CDS market prices. They cover only publicly traded companies and are, at times, affected by noisy market-gyrations, driven by supply and demand conditions rather than credit-risk fundamentals.
Company fundamentals that are updated infrequently, and at best quarterly for publicly listed companies, do not lend themselves to generating timely signals. To make things worse, many private companies do not publish audited financial reports.
Payment behavioral models that monitor abrupt changes in the payment of utility bills and loans, or usage of credit lines. These often provide delayed indicators of a potential default event when it is already close to crystallizing.
Sentiment from news, which can be limited for smaller companies that rarely feature in mainstream media, and problematic for larger companies, which could be targeted by misinformation.
Changes in members of the board and senior executives. For small and midsize enterprises (SMEs) which are often run by family members, friends and relatives, the frequent change of directors may simply reflect family events (e.g., marriage, retirement, exitus) rather than genuine credit deterioration.
Fines reports are infrequent and not necessarily related to genuine credit distress, although they may trigger a liquidity crunch that can precipitate financial matters.
To go beyond classical credit risk automation, risk managers need access to alternative signals that complement their analysis by solving for the limited coverage and timeliness of more traditional approaches when applied to SMEs.
One significant opportunity comes from analyzing digital footprints (e.g., website activity, traffic) that companies "leave" in the internet daily. In fact, the share of SMEs that have established a website has been steadily increasing in the past 10 years, accelerating during the COVID-19 pandemic to exceed 75% in the European Union (and in the United States), and is projected to reach 95% in the next few years. A similar trend, albeit less pronounced, is expected in emerging market economies across the globe as they embrace new technology.
Recent studies on retail purchasers' creditworthiness point to the strong predictive power of the digital footprints they leave online, e.g., customer's email address, device used to browse the internet, hour of day of purchase.1
In a similar vein, at S&P Global Market Intelligence, we are leveraging AI to collect, cleanse, screen, test, and ultimately integrate firms' digital footprints in our credit risk models, and to create a modern portfolio surveillance framework in which a series of automatic and timely signals will help users to assess, monitor and effectively manage credit risk.
As shown in Figure 3:
Step 1: We start from a baseline credit-risk assessment that incorporates company fundamentals, industry- and country-risk scores, and market indicators (such as equity price and volatility, bond spreads, etc.).
Step 2: We then apply a series of overlays to factor in qualitative considerations (including sustainability aspects), payment behavior, sentiment, digital footprints and cyber risk. This refined view provides the most complete, automated, and up-to-date view of a firm's creditworthiness.
Step 3: Macro-economic scenarios project a company's credit risk profile under multiple (benign, negative or stressed) conditions.
Step 4: A set of pre-defined and customizable thresholds trigger multiple early warning signals of an imminent or potential credit risk deterioration, all the way down to default, thus enabling users to filter risky companies or reduce their credit limits.
Step 5: Users can also identify potentially promising business opportunities.
In our case, AI plays a critical role not only in helping to seamlessly monitor alternative datasets on millions of firms (e.g., website activity and trend, hosting environment), but also in generating the company's credit score, and in adding natural-language-based context about the output score and underlying risk drivers by extracting and distilling relevant content from research documents produced by S&P Global Ratings' economists and analysts.
A further distinctive benefit of AI is its ability to dynamically learn from human interaction. In a future step on our journey to leveraging AI, the credit risk signals will be automatically emailed to our users, who will be able to provide feedback, validating or discarding them. In this process, AI will learn and adjust its criteria to increase a signals' accuracy by removing unwanted alarms before they reach the user.
We believe the use of AI will keep growing and become more and more pervasive, as it opens the way to solve new problems in more efficient ways by automating repetitive, quantitative tasks, thereby freeing up humans to focus on the visionary, conceptual, creative, and even emotional elements of work.
What is the relationship between machine learning and AI? While the terms machine learning and AI are often used interchangeably, they are not exactly the same thing. AI represents the ability of a computer to replicate cognitive processes typical of a human being. Machine learning is a practical application of AI, in which mathematical models and processes are applied to data and enable a computer to learn and improve its outputs based on its own experience.2
Why is AI important in credit risk management? AI leverages machine learning credit scoring and scaling to thousands of counterparties, saving time, resources and costs to the adopting firm. AI can help identify relevant research that elucidates the economic, country and industry background in which a given counterparty operates, and the risk drivers of its potential creditworthiness deterioration.
How is AI used in credit risk management? One useful application is in the generation of early warning signals for credit risk portfolio surveillance. Generated signals can help risk analysts focus on companies at risk, digging further before confirming and taking action on a specific company. The use of AI can refine the signals and ensure they are relevant and improve accuracy over time, based on relevant feedback.
1"On the rise of FinTechs – credit scoring using digital footprints", T. Berg, V. Burg, A. Gombovic, and M. Puri, The Review of Financial Studies, Volume 33, Issue 7, July 2020, Pages 2845–2897.
2"Artificial Intelligence (AI) vs. Machine Learning (ML)," Microsoft™ Azure (2023).
