Deep learning for credit risk

The challenge

A company leader in the European market for credit risk has been focusing on improving its statistical models since it was founded 40 years ago. They allow assigning a default probability to companies by analysing a tight selection of their financial an operations data. Its statistical models rely on logistic regressions applied to the output of data validation and data quality sub-systems.

In a world where big data offer ever more analysis and prediction potential, the challenge is to verify if, by using a larger amount of data, it is possible to improve the company’s credit risk prediction system.

Our approach

Axyon gathered all the financial data of target companies (financial module), normalized them and used them to develop a prediction model based on Deep Learning techniques. The client provided Axyon with a benchmark, representing the performance range (based on the AR metric, which evaluates the accuracy of the system by comparing true positive and false positive percentages) reached by their traditional statistical model on a test dataset. The development of Deep Learning models, evolved and trained using our technology, allowed for a sensible improvement of the benchmark’s performances in only two weeks of work.

Results

The system returns the default probability (in the following 6 months) for each sample.

The client’s system (benchmark) reached an average AR of 62,90% (minimum: 60%; maximum: 65%). The model developed by Axyon reached an average AR of 65,12% (minimum: 64,25%; maximum 65,99%).

The performance improvement can lead, for the final user, to a 1,6% reduction of bad debts.

Credit irsk

Conclusions

In this case, the model’s evolution and tuning period was deliberately reduced; additional improvements and on-going refinement will surely enhance the predictive capabilities over time.

Because of the very nature of Deep Learning (which performs optimally when provided with large amount of data), a further improvement potential is foreseen by feeding other modules’ data (other than the financial module) to the system. This improvement step would exceed the improvement reached in the single module.