In recent years, the assessment of debtor solvency has been characterised by external influences on companies, which have made the use of traditional methods for interpreting and evaluating existing data more difficult. Especially in such dynamic times, AI-based systems offer much greater flexibility in terms of adapting the interpretation of a wide variety of developments. At SCHUMANN, we focus on the analysis of internal payment experiences, which, in addition to other data sources, can provide important input for ratings and thus directly improve the information basis on which limit decisions are made.
In the following, we present our development approach and outline the requirements for the use of AI-based systems.
The first step for a company is to develop an appropriate data strategy for payment history analysis, which involves assessing the current state of document documentation. It must be clarified which data management dependencies exist and which interfaces already exist or need to be created. This includes documenting process dependencies and performance requirements in order to be able to respond flexibly to corresponding IT requirements at a later stage.
If these requirements are met, we offer the ability to support different business use cases based on three different models. At this point, it is important not to underestimate the amount of research and development that goes into data validation, quality assurance and other requirements that directly affect the performance of a model. It is important to define how irrelevant or inconsistent data will be handled and what individual adjustments will need to be made to the data model. This requires coordination at both business and technical levels to jointly eliminate unwanted side effects and achieve a robust prediction.
We have developed three different models that allow us to further optimise our customers' receivables management. Based on a debtor's historical payment behaviour, we have succeeded in developing a model that is able to predict very accurately the deviation of a payment from the due date. This enables not only risk assessment of individual documents, but also modelling of a company's liquidity. We have developed a wide range of processes to ensure the above aspects of data quality assurance and, in this respect, model forecasting, which also enables us to develop client-specific models.
We also analysed payment patterns to classify debtor behaviour over time and to identify changes. In this context, we developed a model that classifies debtors at time x (specified by the receipt of a payment) using a clustering approach. From the technical task, we were able to derive a functional definition of the clusters found, which allows us to distinguish between early, late, discount and on-time payers, as well as to identify unstructured payers. The advantage of this analysis is that it is possible to better assess the development of an individual debtor, as well as changes in the clusters of the entire debtor portfolio over time.
Figure 1 - Rule-based rating compared to AI-based rating
For a third model, we extended the overall picture of a debtor to determine an internal rating based on the probability of default. We were able to show that the AI system optimised traditional rule-based analysis for the development partner data pool. Fewer insolvent debtors were incorrectly assigned a good rating based on internal payment experience, but more insolvent debtors were correctly identified by the prediction (see Figure 1). Following the question of the feasibility of the forecast, we analysed different look-back and lock-ahead time windows and concluded that the internal payment experience offers its best performance for a forecast of 3 months.
We use a variety of approaches to ensure the traceability and transparency of both our modelling and the decision-making process for the models. We use best practice approaches to make the impact of individual variables on the output of the model measurable. This allows us to identify potential unintentional biases during model development and also ensures that the model output is causally traceable to the user.
Figure 2 – Integration of AI in limit decision processes
These models provide a bridge to the analysis of internal payment experience in the overall context of the variables used to support limit decisions. Any optimisation of the data quality of individual inputs of usual processes, such as internal payment experience in addition to credit bureau ratings or financial statements, always has the advantage of providing a more accurate and better overall picture (see Figure 2). Therefore, our research and development regarding payment experiences is the first step of refining the general process for assessing a debtor and making a limit decision based on a 3-month forecast in the future.
It is already clear that the most robust and accurate forecast possible in future will be based on many individual optimised data views of the debtor that will be automated through AI. The dynamic adaptability of AI systems and defined retraining processes also ensures that the data is up-to-date in the form of a constantly adapted assessment that can be consistently expanded in the future in line with external influences.
AI-based systems can help take your limit decision processes to a new level and make them future-proof. While there is of course a certain amount of effort involved in preparing data-driven processes, their use has clear advantages over traditional rule-based approaches in terms of objectivity and performance, and can support the decision maker with the best information currently available.
Tobias Nießner holds a PhD in Business Information Systems from the Georg-August University of Göttingen. He studied mathematics and business information systems with a focus on data science. During his PhD, he worked on risk use cases in accounts receivable management.
He joined SCHUMANN in 2023 as Product Owner to enable the product landscape with AI capabilities
Product Owner AI, SCHUMANN
