Holistic credit risk assessment and the goal of safe unbiased modelling

In this Asia Risk webinar, experts in artificial intelligence (AI) and machine learning examined the growing applications being seen in the field and their merit in credit risk modelling

During a panel hosted by S&P Global Market Intelligence and Asia Risk at the Risk Japan conference held in June, Hiroyuki Yoshizawa, executive director and head of product, data, valuation and analytics at S&P Global Market Intelligence, and Yutaka Sakurai, head, RP tech at the AI Finance Application Research Institute, examined the current state and the future of holistic credit risk assessment.

Yoshizawa began the session by asking about the growing role of AI and quantum computing analysis in this field.

The role of AI and machine learning in credit risk modelling

Sakurai explained how AI and machine learning are compatible with credit modelling. In an example comparing the fields of trading and asset management, he noted that, with trading, you cannot always get ahead by following the same pattern. It is important to switch approaches and strategies, and adapt your methods depending on the other side’s strategy and behaviour. While machine learning can help you discover statistical information about the specific point, ever-changing information means it can be difficult for humans to adapt in real time. Asset management is much more stable and therefore better suited to machine learning. Sakurai added that many people have tried to apply machine learning to trading but inevitably end up losing money and giving up.

Sakurai also pointed to the importance of alternative data in credit risk management, as this has been a growing trend in recent years. Machine learning applies to this alternative data not typically included in financial statements and can deliver a higher degree of accuracy in credit risk management. He added that there has been significant progress on this front across North America and Europe.

On the retail credit front, some transactions in Japan, such as deposits, have been studied with reasonably consistent results. However, alternative data is still currently limited in Japan, and Sakurai says little research is being carried out to push this industry forward.

Yoshizawa reiterated Sakurai’s point that, compared with market risk projections, credit risk modelling is more compatible with machine learning. Following this, he foresees that AI-powered credit risk analysis for individual companies will soon become a growing trend, which is an exciting prospect.

Tying the discussion to his work, Yoshizawa explained how S&P Global Market Intelligence offers data to global financial institutions such as hedge funds and asset management firms, as well as noting that clients often carry out credit risk modelling and investment decisions to move forward with. S&P is currently focused on how it can deliver data in a way that provides the most added value to its users. It is for this reason that data accuracy, data transparency, and the disclosure of raw data and data cataloguing are incredibly important.

Yoshizawa added that focusing strictly on credit risk, the most-discussed challenges concern the ability to present historical and current data in a uniform format that works for most companies as well as the ability to maintain input data. When looking at just a single company, it’s helpful to map the historical corporate data, such as the evolution of different categories of the business, and various capital ties such as split-ups, mergers and acquisitions. This mapping makes it much easier to keep up with any developments.

Finally, Yoshizawa explained the link between historical financial data and real-time market data – in particular, the bridge between the quarterly data and the real-time data often used in models. Both for debt and equity, the data can indicate the volatility of a company or the probability of bankruptcy and the financial viability of a company. He then raised the question of how machine learning can use all available input data to help with company decision-making.

Sakurai and Yoshizawa agreed that market data can be easy to access. That said, Sakurai stressed the need to keep in mind the many types of market data and methods of combination. Japan is relatively late in the game in terms of data collection, but it needs to begin now. Once the data has been collected correctly, there are infinite options for combining and sorting it, so Sakurai explains that improved pairing and use of data will inevitably lead to better results. Both Sakurai and Yoshizawa make the point that there is significant room for improvement as well as more dialogue in this data space now to offer new suggestions for data processing.

Modelling and ethics

Yoshizawa questioned whether modelling can help implement a more ethical decision-making process and if we can discuss the ways to adjust it to account for a range of variables.

Sakurai noted that people typically ask the opposite: whether modelling will harm ethics. He stresses that ethics can vary depending on country, noting that the US and Europe have set the tone on ethics for the rest of the world, specifically with regard to equality and anti-discrimination. Applying AI and machine learning to data can often lead to discrimination and, as a result, is already being addressed in certain countries, including the US.

Sakurai presented the example of retail credit, where machine learning can produce a higher credit risk for ethnic minorities, despite the same set of data entered. He noted that, when this occurs, it is possible to obtain different outputs based on gender, age or ethnicity. If we can define this issue as discrimination, then it becomes clear that action must be taken to remedy the situation. That said, this issue is complex.

Yoshizawa concurred that there is currently much debate surrounding the ethics of data on minorities and gender, and it can be interesting to examine simulations to understand the potential consequences on the market and on regulations. Yoshizawa believes the market does a good job reflecting the data obtained through scenario simulations, and that, as data selection and modelling improve, many existing ethical problems can be mitigated.

On the broader topic of environmental, social and governance, Sakurai added that prejudicial results can appear from the data because the chosen fields or standards have a tendency to provide more favourable or negative findings depending on the industry to which they are applied. He believes people who make decisions based on these criteria must understand that these discriminatory results can occur and therefore cannot be fully trusted. Models can be improved but will never be perfect, so Sakurai questioned whether it is necessary to consider additional standards for trading and investment decisions.

Yoshizawa further added that, in adapting the models, a very careful approach must be taken. With regard to the capital markets, consensus has been that market participants needed more and more data to generate profits. However, we are now at a point where there is an abundance of data, highlighting the emphasis on quality over quantity and on polishing data intelligence.

Further clarifying what he means by taking ‘a careful approach’, Yoshizawa noted this approach should include various types of data, a clear relationship between the data and the output, and the coding and technology used by data scientists to make sense of it all.

Sakurai added to the notion of a careful approach, suggesting a quick turnaround cannot be the only approach, because you run the risk of creating a situation with more clutter and less accurate data. Bringing the discussion back to the use of alternative data, Sakurai noted that alternative data has a much shorter history compared with traditional data, so it’s difficult to conduct research in a uniform way.

Yoshizawa ended this section by noting that clutter and model adjusting are going to remain a major part of this discussion for some time yet. While alternative data is not heavily used in the financial markets at this time, the industry should find more applications for market data as there is still a lot of data out there to harness.

Forgotten items in credit risk assessment

The webinar closed with a discussion on forgotten items in credit risk assessment. Sakurai noted that, especially in Japan, it is how higher interest rates affecting credit risk. While in the US and Europe, more people have been seriously considering this pressing issue, Japan hasn’t experienced high interest rates for more than 30 years, and so the correlation between interest rates and credit risk isn’t a focus. That said, Sakurai explained that higher interest rates influence credit risk in myriad ways: banks become less inclined to lend and funds do not circulate as well, borrowers must deal with higher interest payments, and the spreads in the commodities markets can quickly increase. He added that now is the time to begin taking serious consideration of this issue in Japan.

Yoshizawa concluded that data can be much more useful when the following elements are well understood: data consistency, a general idea behind input and output and the technological contribution of data scientists.

Finally, Sakurai noted that now we understand the critical importance of data, it is time to become more efficient in data collection and get more creative. Most important is the understanding that this is a long-term game, and the only way to make progress in this field is to continue with trials and to accumulate more experience with data.

Sponsored content