The use of predictive analytics for supply chain finance, part 1

LinkedinMail

What if you could imagine the ideal SCF world? There would be stable material flows across all channels and borders. We would have a fully digitalised workflow at financing level without any loss of time or other efficiency. Decisive criteria with regard to ESG issues would be fully taken into account. Last but not least: SCF players have optimum planning security thanks to the use of AI-based predictive analysis tools. This ensures maximum resilience, even in the event of major upheavals on the global markets.

Back to reality. Global material flows are unstable and will remain so in the future. We are (unfortunately) still relatively far away from comprehensive digitalisation of the entire SCF and consideration of all ESG factors. What is already available to us today, however, are technologies that help us to analyse existing data streams and use them to make predictions about potential financing risks and disruptions, as well as possible opportunities. We are talking here about machine learning (ML).

1. The basis: data, data, data

Supply chain relationships can be highly complex as they involve many different players on the supplier and customer side. According to an estimate by the Boston Consulting Group, more than 20 parties are usually involved in a typical trade finance transaction*. Accordingly, there is no shortage of data from all the related (sub-)networks: whether payment behaviour, demand trends, defaults, limit utilisation, etc. – the historical information is sometimes already available to SCF providers when it comes to long-term customer relationships. However, the availability of data does not automatically make it usable.

*Boston Consulting Group 2017

Excursus: of supercomputers and exaflops

Enormous computing power is required to process huge amounts of data – especially for training within AI models. It is therefore not surprising that the global demand for powerful computing machines has increased exponentially since the AI boom. Countries around the world are currently investing huge sums of money in the expansion of supercomputers – knowing full well that access to enormous computing power will be one of the strategic advantages in the future**. To give just one example: China is planning to increase its computing power to a total of 300 exaflops by 2025, with one exaflop comprising the power of two million laptops working in parallel. However, cloud computing and collaborations also enable smaller companies to train and utilise AI models with the necessary computing power (see also “Critical factors”).

**Guido Appenzeller, Matt Bornstein, and Martin Casado, Navigating the High Cost of AI Compute, Andreessen Horowitz, April 27, 2023, https://a16z.com/navigating-the-high-cost-of-ai-compute.

2. Agile systems for dynamic markets

The next logical step would be to generate knowledge from the existing databases that can help us make decisions. The aim here is to train a data model in such a way that it is able to output relevant labels such as “low risk”, “medium risk” or “high risk” (supervised learning) with the help of certain parameters (analogous to the evaluation of a customer by an employee). In a further development
stage, the system continues to learn independently and attempts to find and categorise relationships between the data (reinforcement learning). It can access not only “internal” available data from the individual management systems (such as payment behaviour and financial reports, etc.), but also “external” data such as general market trends, data from partners and information sources as well as information from the web (articles, social media). The ultimate aim is to map the dynamics of the markets in an agile, self-learning system.

Quelle: Relevant advanced technologies for trade and supply chain finance, Whitepaper by Commerzbank and Fraunhofer IML, 2022

3. I see what you don’t see!

In this case, mapping means, in particular, being able to recognise patterns that would not be visible to humans and/or previous methods. These patterns can then be interpreted and serve as the basis for further recommendations, warnings or decisions – for example, indications of possible risks that could arise from fraud. As already mentioned, the system’s ability to learn improves over time, leading to increasingly accurate results.

 

Click here to continue