IMF Paper Examines Application of Machine Learning in Assessing Credit

The paper examines potential strengths and weaknesses of machine-learning-based credit assessment by presenting core ideas and the most common techniques in machine learning for the nontechnical audience; it also discusses the fundamental challenges in credit risk analysis. The paper presents the main elements of prudent lending in the context of the five Cs of credit and agency problems.  The five Cs of credit are capacity, capital structure, coverage, character, and conditions. The paper also discusses main machine learning tools and techniques and examines the strengths and weaknesses of machine learning-based lending in contrast with the traditional methods. 

The paper highlights that machine learning has certain weaknesses that should be taken into consideration when applying for credit risk assessment:

• Heavy reliance on learning from data, particularly in a context where the size of the sample is considerably larger than traditional ways of credit scoring, could result in noisy information playing a role in driving results of credit analysis and leading to financial exclusion of creditworthy applicants. The bias in the sample should be identified and avoided by analysts as much as possible to avoid digital financial exclusion.
• Machine learning may not capture structural changes in a timely manner, because arrival of informative data may be slow due to the lengthy process of default observation. This could negatively impact fintech lenders that rely on machine learning to assess borrowers without evaluating the relevance of data used for training the model for new applicants. 
• Borrowers may realize and counterfeit certain indicators that drive their credit score, thus decreasing the relevance of those features for new applicants.
• Machine learning is exposed to some of the key concerns in econometrics, most importantly the endogeneity and selection bias problem. The analyst should check the sample to ensure proper treatment of these issues and avoid superfluous results. Nonetheless, proper choice of risk drivers makes these issues less of a concern for credit risk assessment and default outcome prediction.

Overall, the paper concludes that fintech credit has the potential to enhance financial inclusion and outperform traditional credit scoring by leveraging nontraditional data sources to improve the assessment of the borrower’s track record, appraising collateral value, forecasting income prospects, and predicting changes in general conditions. Nevertheless, the emerging market economies may face some challenges in reaping the benefits while ensuring that the development of fintech credit does not expose systemic risks to the financial system. Given the central role of data in machine-learning-based credit analysis, it should be legally and technologically possible to gather digitalized data reliably from various sources and avoid noisy and biased data as much as possible. As a complement for high-quality data availability, cyber-security measures should be in place because of the sensitivity of credit information. Moreover, the machine learning analysts should consistently review machine learning models for credit rating to avoid potential weaknesses of naive application of machine learning. 

Related Link: Working Paper

Keywords: International, Banking, Financial Inclusion, Fintech, Machine Learning, Credit Risk, Big Data, Systemic Risk, IMF