Can machine learning predict the probabilities of default?

Can machine learning predict the probabilities of default?

According to Moore’s law, computing power doubles up each two years. This performance increase in computing power makes machine learning increasingly efficient each year, and widely applicable. But does this also apply to credit risk issues?

The idea of neural networks, for example, was introduced in the 1940s but abandoned in the 1970s due to the lack of computing power. Today, the neural network is among the most used machine learning algorithms. With today’s strong, and ever growing, computer power we can benefit from the higher prediction accuracy from the more complex machine learning models in credit risk as opposed to the current standard – the logistic regression. Even the required explainability for credit risk models can be obtained due to recent developments in machine learning. In this light Zanders, together with students from the Erasmus University, have conducted a case study in which a shadow rating model is formed for predicting the probabilities of default by leveraging on machine learning techniques.

Baseline model

In the case study the shadow rating model uses the ratings of Moody’s, Fitch and S&P. By considering the financials over the years from a vast array of banks a corresponding rating for the bank in each year is formed. The baseline model used is the logistic regression, because of inherent interpretability and simplicity . During the case study several machine learning techniques were tested to improve the accuracy of predicting the probability of default while maintaining high interpretability.

Better prediction accuracy

We found that almost all tested machine learning models gave significantly better results in terms of prediction accuracy. The best performing machine learning model outperformed the logistic regression by 7% when looking at predicting exactly the right rating as compared to the logistic regression. This increases further to 11% when we accept that the rating might be one classification off. Overall Random Forest or stochastic gradient decent models in combination with explanation methods such as SHAP significantly increase performance while maintaining the interpretability.

In conclusion

It is possible to leverage on machine learning techniques to improve model accuracy significantly. When these machine learning techniques are combined with explanation methods, the influence of each variable on decision of the final rating become interpretable. Hence machine learning could be an improvement in comparison to the logistic regression in predicting probability of default. With more computing power and data available this field will likely only grow in the future. The time is there to double down on machine learning with credit risk.



Dumitrescu, Elena-Ivona and Hué, Sullivan and Hurlin, Christophe and tokpavi, sessi, Machine Learning or Econometrics for Credit Scoring: Let’s Get the Best of Both Worlds (January 15, 2021).
Lundberg, S. M. & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in neural information processing systems, 4765–4774.