Can bankruptcies be predicted algorithmically?

What if neural networks could be used to predict an organisation’s downfall? Philippe du Jardin, Professor at EDHEC Business School and coordinator of computing teaching, studies artificial intelligence models (“neurons”) that involve science with very real applications, namely in predicting risk. Meet a pioneer in the discipline who is now a member of the highly exclusive club of the world’s top 2% most cited researchers.

• This article was originally published in EDHEC Vox magazine #17, available in French and in English.

With the media lately saturated with news about artificial intelligence, it’s easy to forget one thing: the discipline didn’t emerge yesterday. The current surge comes in wake of more widespread use of large language models, but it builds on work that is often traced back to the 1950s and the figure of Alan Turing. Closer to home, Philippe du Jardin, Professor at EDHEC, is also an important contributor to the recent history of AI. The specialist in non-linear classification models — we’ll come back to this — took some time with us to do a bit of teaching on a complex subject: how AI can automate the prediction of corporate bankruptcies.

A little lesson in artificial intelligence

We know one thing when we take a seat across from Philippe du Jardin: his pet topic is predicting potential bankruptcies by using neural networks. Fortunately, the researcher is as good a teacher as he is passionate about his topic, and where to begin quickly becomes clear. Neural networks, he explains to us, “are units of calculation linked to one another by coefficients. Each neuron does a little bit of calculation based on what it receives beforehand and transmits — or doesn’t transmit — data to the next neuron.” Networks then offer a result based on this input data. 

Neural networks are special because they can be used to model nonlinear phenomena much better than traditional methods. In layman’s terms, they can model complex relationships between two objects. This modelling takes different shapes. With ChatGPT, you get a sequence of words; with Philippe du Jardin, you can predict whether an object belongs in a class: “Does this bone belong to a man or a woman? Is this customer pay consistently? Is this share likely to increase? Is this company going to go bankrupt?”

Neural networks must be trained to get such results. In the supervised learning phase, the network is forced to reach an expected result. If it’s wrong, the parameters are corrected, and so on. “We stop at a given moment. When we stop must be precisely calculated. If the Learning period is too short, the network will generalise poorly and will not be able to adapt. Conversely, if it has learned too much, it will be biased by the sample’s specificities, which Don't always correspond to those of the whole population,” Philippe du Jardin adds.

A fantastic tool for risk management

Trained in AI in the early 1980s, Philippe du Jardin first became interested in bankruptcies as a matter of convenience. “Every country publishes companies’ financial data. They’re accessible, inexpensive and standardised,” he explains. In doing so, he would make an entrance into the rapidly expanding field of credit ratings, with its many applications. “Every investor on the planet needs risk ratings to evaluate the accounts of the organisations they lend to. It’s a matter of interest to every creditor,” Philippe du Jardin says.

Bankruptcy models have existed since the 1960s. But as a computing enthusiast who also worked at several business schools, du Jardin stood out by incorporating real knowledge of economic phenomena into a field that was still purely algorithmic. “Statisticians and mathematicians dominated the discipline. Most of the time, they knew nothing about the problems being dealt with. Classifying flowers, companies or people amounted to much the same thing for them. Only the algorithm mattered,” Philippe du Jardin explains. Drawing on knowledge gained from the scientific literature, he then enriched the models. For example, some bankruptcies occur very gradually over ten years, whereas other companies sink within two years and others yo-yo constantly before closing down. 

“So, we know things about bankruptcy, and what interested me was using this information to improve the precision of models. I simulated bankruptcy trajectories that accounted for firms’ path towards their demise and used them as models. In short, I look for regular patterns in the data, based on knowledge we have a priori, and I build models based on those regularities,” du Jardin says.

An agentic future

Philippe du Jardin is banking on agentic AI for the future. He sees a near future in which autonomous agents will be able to collect data, reprocess it, choose parameters, and build models, but also analyse the results and make decisions. “For a long time, AI was for specialists because many of the tasks involved in creating a model were manual or involved a sort of statistical craft required to make choices. Recent progress now allows people to create models without being a specialist, because, in a way, specialists’ expertise has itself largely been automated. The next step is agents. Currently, a model creates a result, and this result is used by a person who analyses it, decides and then acts. This analysis-decision-action sequence is in the process of being automated, and agents are going to start doing all or part of the work currently done by people.” 

Replacing parts of human work raises ethical questions, but Philippe du Jardin prefers to refer them back to how we use algorithms, rather than to the Tools themselves. “The machine is neither bad nor good; ethics is involved when it comes to how we use algorithms. You can also do classification that improves people’s lives.” 

From algorithmic oracles to agentic tools, one thing is sure: the future is being written in new ways and in collaboration with machines that will transform our organisations.

Photo de and machines sur Unsplash