The application of artificial intelligence (AI) in credit scoring is fundamentally transforming the lending industry. Lenders are leveraging advanced AI algorithms and substantial amounts of data to predict credit risk more accurately, expand credit access for small businesses and underbanked/unbanked individuals, and increase operational efficiency. This article discusses the role of AI in credit scoring models and recent developments in this field.
Image credit: Golden Dayz/Shutterstock
Importance of AI in Credit Scoring
Credit scores are crucial for everyone as they control the cost of loans, housing decisions, and employment. The conventional credit scoring models used by lenders primarily depend on a limited number of data points, including credit history length, outstanding debt, payment history, and current income, while determining the creditworthiness of borrowers.
Although these models are effective for several borrowers, they are not suitable for individuals with non-traditional sources of income/thin credit lines. AI credit scoring models can quickly identify data patterns and make more accurate predictions of the creditworthiness of all types of borrowers by analyzing an extensive range of data points, including non-traditional data sources.
The decisions of AI credit scoring models primarily depend on user behavior analytics, work experience, transaction analysis, credit history, total income, and different additional real-time factors. Thus, using AI credit scoring models can accelerate the application process of borrowers and assist lenders in making better lending decisions as they provide more individualized and sensitive credit score assessments.
Specifically, these models allow more individuals with adequate income potential to access financial services. The inclusion of non-traditional data sources, such as utility bill payments and social media activities, assists in developing a comprehensive credit risk profile of a borrower, leading to better and more accurate loan decisions.
Additionally, AI credit scoring models can automate several aspects of the lending process, which improves operational efficiency by reducing the need for manual underwriting, leading to an improved borrower experience and faster loan approvals/rejections.
Concerns with AI in Credit Scoring
Data security and privacy are the major concerns of using AI in credit scoring, as AI credit scoring models primarily rely on vast amounts of data. Lenders must collect and use data in compliance with applicable regulations and laws and implement effective cybersecurity measures to secure sensitive borrower data.
AI credit scoring models can perpetuate bias/discrimination if the data used to train the models is biased, which is another significant challenge. Lenders must carefully consider potential biases in their AI model training data and implement steps to mitigate potential discrimination in lending decisions.
AI algorithms are often difficult to interpret and complex, which increases the difficulties for lenders to explain lending decisions to borrowers. A clear explanation of AI credit scoring models and lending decisions to borrowers is crucial to ensure trust and transparency in AI models.
Limited control of financial consumers over the AI model outcomes owing to the existing scope of financial consumer protection law and data protection law, financial exclusion due to lack of data from traditionally excluded groups, insufficient oversight over the use of alternative data for AI credit scoring, and regulatory arbitrage in lending markets, are the other major disadvantages of AI in credit scoring that create challenges for financial regulators.
Financial regulators can adopt a set of solutions and tools to address these problems to effectively protect the end users of AI credit scoring models/consumers and promote access to finance. A testing supervisory process can be implemented for AI algorithm-based credit scoring models to promote fair lending.
The regulators can provide the right consumers to know the AI model outcomes, including the opinion data and inferences, to ensure digital self-determination. This solution can allow consumers impacted by algorithmic credit scoring to challenge and verify the decisions made by AI credit scoring models.
Using AI credit scoring, financial institutions, and other lenders can establish a level playing field. Additionally, the sandbox can be used as a test environment for lenders to generate data on traditionally excluded groups, such as minorities and low-income groups, in a controlled environment.
Moreover, open finance schemes can promote data portability and sharing initiatives for credit scoring in a financial regulatory authority-controlled environment.
A New Approach Towards AI-based Credit Scoring
Random forest (RF) and support vector machine (SVM) techniques have received significant attention from financial institutions for credit scoring owing to the flexibility of these techniques to identify different data patterns. Both techniques are black-box models that are sensitive to hyperparameter settings.
In a paper published in the journal Entropy, researchers proposed the use of harmony search (HS) to form a hybrid HS-RF model for hyperparameter tuning to provide reliable feature importance to ensure good model performance and a hybrid HS-SVM for simultaneous hyperparameter tuning and feature selection to select proper hyperparameters and explain attributes based on reduced features.
Feature importance derived by RF can be used for model explanation, while feature selection performed on SVM enable explanation with reduced features, which allows flexibility in modeling while ensuring high accuracy in classification similar to traditional statistical modeling.
SVM and RF were also hybridized with a modified HS (MHS) to form MHS-SVM and MHS-RF to improve the computational efficiency while maintaining a performance similar to HS and grid search (GS) with a shorter computational time.
MHS consisted of four major modifications compared to standard HS, including Elitism selection in place of random selection during memory consideration, dynamic exploitation and exploration operators in place of original static operators, extra termination criteria to achieve faster convergence and a self-adjusted bandwidth operator.
All four proposed hybrid models were compared with standard statistical models across three datasets commonly used in credit scoring studies. MHS and parallel computing effectively reduced the computational time of proposed hybrid models. Computational results displayed that MHS-RF was the most effective among all models based on computational time, model explainability, and model performance.
Recent Developments
In a recently published work in Progress in Artificial Intelligence, researchers performed a comparative result analysis and assessed the impact of feature selection approaches on different classification approaches with credit-scoring datasets.
Infinite latent feature selection (ILFS), feature selection via Eigenvector centrality (ECFS), Relief, feature selection via concave minimization (FSV), least square (LS), unsupervised discriminative feature selection (UDFS), feature selection by local learning-based clustering (LLCFS) and correlation coefficients-based feature ranking (CFS)-based feature selection with linear discriminant analysis (LDA), naive Bayes (NB), time delay neural network (TDNN), k-nearest neighbors (KNN), decision tree (DT), extreme learning machine (ELM), faster-1 ELM (ELM-1), faster-2 ELM (ELM-2), RF, partial decision tree (PART), multi-layer perceptron (MLP), radial basis function neural network (RBFN), SVM with radial kernels (SVM-R), SVM-P, linear regression analysis (LRA) and sequential minimal optimization (SMO) classification approaches were utilized in this study.
Seven real-world credit scoring datasets, including German categorical and numerical, Taiwan, Japanese, Bankruptcy, Bank-marketing, and Australian datasets, were utilized for validation. The results were compared based on classification accuracy, specificity, sensitivity, and G-measure.
The study results demonstrated that TDNN and RF were the best and second-best classification approaches with the most credit-scoring datasets, respectively, while LLCFS-/CFS-based feature selection approaches were more efficient and improved the performances of both classifiers. UDFS feature selection approach demonstrated the best performance with most classification approaches. The feature also improved the classification performances compared to performances with other features. These results were compared with those of different state-of-the-art credit scoring prediction model approaches, including rule-based classification, ensemble, hybrid, and classification models.
TDNN displayed better performance with all credit scoring datasets than hybrid and model-based classification credit scoring approaches. Overall, neighborhood rough set (NRS)-based feature selection and layered ensemble classification was the best performer on credit scoring datasets, while TDNN displayed the second-best performance.
To summarize, AI credit scoring models are revolutionizing the lending industry by increasing the accuracy of lending decisions and accelerating the decision-making process. However, AI governance guidelines, fair lending regulation, AI regulation, unbiased data, and better algorithms are required to make AI more effective for determining the creditworthiness of borrowers.
References and Further Reading
Goh, R. Y., Lee, L. S., Seow, H., Gopal, K. (2020). Hybrid Harmony Search–Artificial Intelligence Models in Credit Scoring. Entropy, 22(9), 989. https://doi.org/10.3390/e22090989
Tripathi, D., Edla, D.R., Bablani, A., Shukla, A. k., Reddy, B.R. (2021). Experimental analysis of machine learning methods for credit score classification. Progress in Artificial Intelligence, 10, 217–243 (2021). https://doi.org/10.1007/s13748-021-00238-2
Moss, J. (2022). A New Approach to Regulating Credit-Scoring AI. [Online] Available at https://www.theregreview.org/2022/06/07/moss-new-approach-to-regulating-credit-scoring-ai/ (Accessed on 30 July 2023).
The Role of Artificial Intelligence and Machine Learning in Credit Scoring [Online] Available at https://chronicle.creditinfo.com/2023/04/26/the-role-of-artificial-intelligence-and-machine-learning-in-credit-scoring/ (Accessed on 30 July 2023)