AI and machine learning methods in risk management

Early identification of credit and ESG risks

Since 2020, the global economy has been impacted by a series of exogenous shocks that have had a profound and lasting effect. These include the coronavirus crisis (Q1 2020 – Q1 2022), the war in Ukraine and the associated energy crisis (from February 2022), the sharp rise in inflation (peaking in 2022/2023), which the central banks then countered with a more restrictive monetary policy (ECB from mid-2022). Since March/April 2025, the US government under Trump has also kept the world on tenterhooks with its erratic and protectionist trade and customs policy. The events listed are not exhaustive. However, they show that banks and companies are facing major challenges: They find themselves in a highly volatile environment in which their business situation and, as a result, their credit rating can deteriorate quickly and unexpectedly.

This is precisely where early warning systems come into play, helping bank analysts and credit experts to identify companies and sectors that are particularly affected by deteriorating economic conditions at an early stage. For banks, early risk identification is a regulatory supplement to traditional ratings and enables the prompt identification of potential credit risks as part of ongoing risk monitoring and proactive risk management. In accordance with regulatory requirements, such as MaRisk, the effects of ESG risks must also be taken into account as part of early risk identification [see e.g. BaFin (2024), BTO 1.3.1, p. 38]. RSU’s Risk Guard early warning system uses a news-based analysis based on machine learning methods and AI for both use cases, among other things. Examples show that both financial risks and ESG controversies can be identified at an early stage in this way. In the ESG context, the systematic and automated evaluation of news also significantly broadens the analysts’ information base, making up in part for companies’ limited disclosure obligations.  

Overview of the early warning system

The Risk Guard early warning system (see Figure 1) automatically evaluates capital market data and quality economic news reports on a daily basis, thereby enabling a continuous review of a borrower’s credit situation. The models on which the early warning system is based forecast deteriorations in creditworthiness and defaults of both listed and unlisted companies with a lead time of up to one year. Anomalies in the early warning system can promptly trigger adjustments in rating or other systems (limit systems, etc.). The following section will briefly focus on RSU’s AI-supported news-based early warning system (NBF for short), which is used to identify credit and ESG risks.  

Overview of the early warning system

For model development in the context of AI, the first decision to make is which platform to use. For example, can a cloud-based large language model (LLM) such as Chat GPT, Gemini, etc. be used or should an open source LLM be used on an on-premise server? For development purposes, in addition to the important issue of data sovereignty, the costs incurred for particularly large amounts of data must also be taken into account, as cloud solutions are billed per token – both in development and in operation. With this in mind, we have set up a platform for our machine learning and AI models that basically enables both approaches for model development and operation. The news-based early warning system was developed on a dedicated server, as the large volumes of data would otherwise have resulted in very high costs. In addition, the use of an open source LLM in conjunction with machine learning methods led to very good results, as can be seen below. The ongoing operation of the news-based early warning system also takes place on a dedicated server.  

News reports as a data source

German- and English-language news feeds were commissioned from major providers for the news-based early warning system. One news provider covers the German press landscape almost completely. The second one supplies news for a large number of English-language newspapers. License agreements with the data providers ensure that copyrights, transparency of sources and thus a trustworthy database are maintained. This is an issue that is particularly important in the context of machine learning and AI and is a recurring theme.

Twenty years of news history with 24 million news items from more than 350 newspapers and magazines in German and English are currently available for the development of our models. More than 5000 news items are analyzed daily during operation.

Economic relevance of news items

AI and machine learning are used in many areas in the context of news-based early warning. The first step is to determine whether a news item is of economic relevance. Although it would be possible to filter according to certain newspaper sections, these original classifications are very imprecise and incomplete. There is a risk of missing relevant news that is reported in the local section of a newspaper, for example. To assess the economic relevance of a news item, a statistical classification procedure was therefore developed which ensures that all sections of a publication are covered.

Associating news reports with companies

News items with economic content identified as relevant must be assigned to the companies they concern. Deep learning methods are used for this. Company names are extracted from the respective articles using a specially trained neural network (Named Entity Recognition). If several company names appear in an article, they are sorted and filtered according to their salience. In this context, a higher salience means that one company name is considered more central to the article than another. This ranking is created using easily interpretable metrics such as the frequency or distribution of the company name in the text. In a final step, the company names extracted from the article are compared with a central company register in which company names are recorded in different variants, such as BMW, BMW AG or Bayerische Motorenwerke Aktiengesellschaft.     

Risk score, sentiment index and early warning signals  

To determine the risk score of a message, information from a large language model (LLM) and machine learning methods are combined with default information from the RSU data pool. The resulting risk scores for the individual reports are then aggregated into a sentiment index for the respective company, which has a discriminatory power, as measured by the “Area under the Curve” (AUC), of around 80 percent for the observed defaults in our data pool (see Figure 2). This is a level that only very good rating systems achieve.

Sentiment index for credit risks – case study

How a news-based early warning system works is described in more detail below for the example of Signa Holding GmbH:

If the sentiment index (green line) exceeds the signal zone, an early warning signal is generated and the analyst is informed by e-mail. In the application, a mouse click on the index displays the underlying news reports with their headlines and the words with the highest contribution to the index value. With a further mouse click, you can read the corresponding news item in full.

Signa Holding has been in the signal zones since the beginning of the coronavirus crisis. During the corona crisis, the business model, accounting techniques and the lack of up-to-date annual reports were discussed. Figure 3 shows further selected headlines.

Sentiment Index for ESG – Case study

The approach chosen for credit risks was also used in the ESG context. To determine the risk score, individual news items are assessed in terms of the criticality of the ESG-score relevant information they contain. The resulting risk scores for the individual news items are then aggregated into a sentiment index for the respective company.

Ferrero case study – see also Figure 4: On April 9, 2022, various newspapers reported that the food authorities in Belgium had ordered the closure of a chocolate factory due to salmonella contamination.

An overview of the news reports on that day can be seen in Figure 5 below.

The ESG Sentiment Index helps to identify ESG issues at companies, as in the case of Ferrero, and supports analysts in assessing ESG risks in the context of ESG ratings.

Conclusion and outlook

When selecting the infrastructure for development and operation for machine learning and AI approaches, in addition to the issue of data sovereignty, the costs incurred for particularly large amounts of data must be taken into account.

Examples show how both financial risks and ESG controversies can be identified from news reports for individual companies at an early stage using machine learning and AI methods. Machine learning and AI do not only play an important role in the context of early risk detection. There are many possible applications in credit risk management. Other areas of application include the automated evaluation of financial or sustainability reports and the integration with existing risk classification systems. With the methods described, previously manual activities can be standardised and, above all, scaled up. In the interest of modern risk management, it is to be hoped that the banking supervisory authorities will allow sufficient leeway for these innovative approaches to be used more widely.