How to improve credit modeling with AI
In this paper, Davis et al. developed a wide range of machine learning (ML) models to forecast consumer credit risk. These models include interpretable rule-based models (e.g., inductive logic programming and optimal trees) and black-box ML models (e.g., neural networks and random forests). They compare these models and find that neural networks outperform other linear and nonlinear models, reaching 74.75% accuracy. But the most important contribution of this paper is how the authors emphasize the “explainability” of these models by identifying the different stakeholders involved in credit risk management—loan companies, regulators, loan applicants, and data scientists—and provide explanations of the models according to their needs. In particular:
- For loan companies, they generate interpretable explanations for every prediction using two post hoc explainability tools, Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP), that are able to provide reasons for a loan denial.
- For regulators, they suggest a method of investigating the fairness of these models. Regulators may also be involved in investigating the model’s functionality in extreme scenarios, which is why the authors provide a methodology to perform stress tests to analyze the model’s behavior in extreme scenarios.
- For loan applicants, they generate a diverse set of suggestions that can help individuals deemed non-creditworthy become creditworthy.
In our results, we demonstrated that with the right tools, even black-box machine learning models are able to answer a series of important questions for credit risk modeling. These questions include: why does the model classify a data point in a certain way? What small changes in feature values could reverse the model’s classification of an individual? How does the model behave in extreme scenarios? What relationships did the model learn? Are the models biased? What is the minimal summary of the dataset?
Explainable Machine Learning Models of Consumer Credit Risk
Authors: Randall Davis, Andrew W. Lo, Sudhanshu Mishra, Arash Nourian, Manish Singh, Nicholas Wu, Ruixun Zhang
From: MIT
AI, fintech and the supply of consumer credit
In recent years, the volume of fintech loan originations has been characterized by explosive growth, especially among personal loans and mortgage loans. Dolson and Jagtiani, in this paper, explore the role of fintech lenders in expanding credit access to “underserved” consumers, namely minorities, consumers with low income, low credit scores, or thin credit histories, or consumers who have a history of being denied for credit. Towards this goal, they use fintech credit offers, rather than actual originations, which allows them to provide a detailed examination of the supply side of fintech loans. In addition they use a diverse and extensive set of measures of a consumer’s credit access which allows them to develop a more nuanced picture of the types of consumers receiving offers from fintech firms compared with traditional banks and traditional nonbank lenders. They note that the main difference between fintech, traditional banks and nonbanks is that the former use more data (including alternative data) and more complex artificial intelligence/machine learning modeling in their risk evaluation and pricing and in their fully digitized credit-decisioning process.
Overall, we find that the rapid entry and expansion of fintech firms into the personal lending and mortgage lending space have delivered potential benefits to subprime and other credit constrained consumers, allowing them more credit options through fintech (direct mail) credit offers.
Which Lenders Are More Likely to Reach Out to Underserved Consumers: Banks versus Fintechs versus Other Nonbanks?
Authors: Erik Dolson, Julapa Jagtiani
From: Federal Reserve Bank of Philadelphia
