Scientists have proposed a new method for predicting company bankruptcy based on machine learning.

The authors of the new work from the Higher School of Economics taught the algorithm to divide companies into two types that are stable and continue to exist and those that go bankrupt within a certain period of time.

To train AI, the authors of the paper used historical data on successful and bankrupt firms. The main indicators that the algorithm paid attention to: business performance, as well as the conditions and patterns under which the company began to develop or, conversely, ceased to exist.

The authors note that the task of predicting the likelihood of a company closing or not is highly dependent on the underlying data. And there is an imbalance in them, since according to statistics, bankruptcy is rare, so there are many more surviving enterprises in training kits.

The share of bankrupt companies is about 5-10%, so AI does not always understand what signs and conditions led to bankruptcy. Therefore, the researchers took a different approach: they created a method that is less sensitive to imbalances in the original information.

It trains many separate classification algorithms, from which the most effective ones are then selected and combined to achieve the highest prediction accuracy.

Since the model is based solely on the financial performance of companies, its results are also true in the extreme conditions of the global COVID-19 pandemic. In the future, interest in using machine learning methods will only grow, and we believe that at some point they will completely replace traditional methods of predicting company bankruptcy.

Yuri Zelenkov, Professor, Department of Business Informatics, HSE