Unwrapping black box models A case study in credit risk

The past two decades have witnessed the rapid development of machine learning techniques, which have proven to be powerful tools for the construction of predictive models, such as those used in credit risk management. A considerable volume of published work has looked at the utility of machine learning for this purpose, the increased predictive capacities delivered and how new types of data can be exploited. However, these benefits come at the cost of increased complexity, which may render the models uninterpretable. To overcome this issue a new field has emerged under the name of explainable artificial intelligence, with numerous tools being proposed to gain an insight into the inner workings of these models. This type of understanding is fundamental in credit risk in order to ensure compliance with the existing regulatory requirements and to comprehend the factors driving the predictions and their macro-economic implications. This paper studies the effectiveness of some of the most widely-used interpretability techniques on a neural network trained on real data. These techniques are found to be useful for understanding the model, even though some limitations have been encountered.