Designing topological data to forecast bankruptcy using convolutional neural networks

Over the past two decades, convolutional neural networks (CNN) have been used to successfully perform classification tasks, such as image recognition, that rely on translation-invariant, locally and spatially correlated data. However, with certain classification problems, such as corporate bankruptcy forecasting, data are made up of unordered columns of numbers, and the distance between two columns is meaningless. Indeed, such data most often represent a firm’s financial situation measured at a given point in time, and the variables that can embody this situation are not structured in a way that would make them suitable to be used with a CNN. Since CNN represent a class of classification techniques that is very efficient and as there are virtually no methods that make it possible to adapt accounting data to CNN properties, in this article we propose a technique that aims to transform data into a topological space—a space endowed with a spatial structure—in order to study the ability of CNN to forecast corporate bankruptcy and financial failure. The results of our study show that CNN-based models designed with this technique lead to models that are more accurate than traditional ones.