Credit Risk Analytics

Problems such as credit scoring, fraud detection, churn prediction, credit limit definition, and financial behavior forecasting (among others) are constant challenges for banks and financial institutions, which permanently research for more accurate results and ways to decrease business-related risk when providing services. Most of these problems can be tackled by using machine learning to classify users who are likely to, for example, not pay their bills on time or commit fraud. In this chapter, the quantum machine learning side of these scenarios will be explored, using a permanent benchmark with classical counterparts for most of the cases.

In the current economic situation, where the stability of the markets is unpredictable and the way people work is always changing (thanks to the rise of the “gig economy”), it is harder to increase a credit product portfolio and cover a larger number of customer cohorts without increasing the risk for businesses...

The relevance of credit risk analysis

With the objective of providing a broader context and understanding of the relevance of addressing classification problems in the finance sector, for this part of the book, it is important to define some core concepts, even from a high-level perspective. The term “credit risk” in the context of this chapter is the chance that a lender will lose money if a borrower doesn’t pay back a loan by a certain date. As the credit card business has grown quickly, as illustrated in Figure 6.1, and the financial players have grown over the years, the challenge of expanding the scope of targeted people requires more sophisticated underwriting systems. This puts a big portion of financial institutions at risk if the means to assess this risk are not accurate enough.

Given the situation previously described, it is often necessary to look at the credit risk of customers who have little or no credit history to expand the current client segments...

Data exploration and preparation to execute both ML and QML models

As mentioned before, in this chapter, we will walk you through the implementation of hybrid quantum-classical algorithms and how they behave in a real-world scenario in finance, but before you start playing with them in a professional setup, you should think – or at least review – some the following concepts.

Data enrichment refers to the process of enriching or supplementing an existing dataset with extra information. Data enrichment in the context of credit scoring systems is the use of additional data sources to supplement extra variables and features that could come from a credit bureau or a non-traditional source (e.g., mobile data mining) in order to increase the accuracy of credit risk assessments.

By incorporating additional data sources like public records (digital footprints), social media behavior, financial history, open finance, and other alternative data sources, data enrichment can...

Implementation of classical and quantum machine learning algorithms for a credit scoring scenario

Applying machine learning and quantum machine learning for credit scoring challenges requires the development of a prediction model that can properly determine an individual’s or company’s creditworthiness. Typically, this procedure, as shown in the steps described previously, includes data collection, data enrichment, data preparation, feature engineering, feature selection, model selection, model training, model evaluation, and subsequently, deployment. In this section, we will cover most of the previous concepts and procedures, assuming that the data is already encoded to numerical variables and the feature has been selected.

Data preparation

First, the data needs to be loaded. This data will come in one of the more well-known formats in the industry, which is CSV. The information that will load into the notebook, as previously detailed, is in a classical format...

Quantum Support Vector Machines

The Support Vector Classifier (SVC) and Quantum Support Vector Classifier (QSVC) are the first models that will be used to look at our synthetic dataset, and we will see how a quantum algorithm versus a classical algorithm can work to find potential defaulters. One of the most widely used techniques is known as Support Vector Machines (SVM) (Hearst et al., 1998), which make use of hyperplanes in order to find separable spaces within our data regime. These hyperplanes are responsible for separating our N-dimensional information into different spaces, trying to maximize the margin between samples from the regions split by the hyperplane itself. By softening this margin constraint and allowing some samples to be misclassified, we allow the model to generalize from the dataset itself. This softened version is what we will call an SVC.

Thanks to the abstraction level that Python libraries such as scikit-learn provide, its usage is as simple as calling...

Conclusion

As mentioned earlier, even if accuracy is a common measure from the classification report that most people will look at, the way to treat this kind of imbalanced data scenario is to compare the models using a balanced accuracy score, or AUC score, which in this case are the same, since it is a binary classification challenge.

Figure 6.8 – A comparison of classification results between classical and hybrid quantum-classical methods

At first glance, the results do not appear to be conclusive about the benefits of using hybrid quantum-classical for classification problems that the finance sector may face. However, the purpose of the exercise in this chapter is to get people to think about their own business challenges and do more research, since we can see that quantum machine learning could be at least equal or slightly better than classical ML methods (e.g., QSVC versus SVC). When any incremental benefit is achieved in terms of...

Further reading

• Agarap, A. F. (2018). Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375.
• Assefa, S. A., Dervovic, D., Mahfouz, M., Tillman, R. E., Reddy, P., & Veloso, M. (2020, October). Generating synthetic data in finance: opportunities, challenges and pitfalls. In Proceedings of the First ACM International Conference on AI in Finance (pp. 1–8).
• Bishop, C. M. (1994). Neural networks and their applications. Review of scientific instruments, 65(6), 1803–1832.
• Bro, R., & Smilde, A. K. (2014). Principal component analysis. Analytical methods, 6(9), 2812–2831.
• Crook, J. N., Edelman, D. B., & Thomas, L. C. (2007). Recent developments in consumer credit risk assessment. European Journal of Operational Research, 183(3), 1447–1465.
• Crouhy, M., Galai, D., & Mark, R. (2000). A comparative analysis of current credit risk models. Journal of Banking & Finance, 24(1-2), 59–117.
...