Explainable FinTech lending
Study on explainable AI for fintech lending, using Shapley values to improve credit scoring models for SMEs.
What it examines
This study addresses the gap in explainable AI for credit risk assessment in FinTech lending, proposing a model selection method using Shapley values to balance predictive accuracy and explainability, applied to a large dataset of European SMEs.
What it concludes
The research offers a method to create accurate and explainable credit scoring models, useful for banks, fintech companies, and regulators. Future research could explore alternative Shapley value approaches and the fairness of credit scores across different demographics.
Evidence objects
The research offers a method to create accurate and explainable credit scoring models, useful for banks, fintech companies, and regulators. Future research could explore alternative Shapley value approaches and the fairness of credit scores across different demographics.
key_findings bullet 1 · key_findings · validation V0
Raw abstract and provenance
Lending activities, especially for small and medium enterprises (SMEs), are increasingly based on financial technologies, facilitated by the availability of advanced machine learning (ML) methods that can accurately predict the financial performance of a company from the available data sources. However, despite their high predictive accuracy, ML models may not give users sufficient interpretation of the results. Therefore, it may not be adequate for informed decision-making, as stated, for example, in the recently proposed artificial intelligence (AI) regulations. To fill the gap, we employed Shapley values in the context of model selection. Thus, we propose a model selection method based on predictive accuracy that can be employed for all types of ML models, those with a probabilistic background, as in the current state-of-the-art. We applied our proposal to a credit-scoring database with more than 100,000 SMEs. The empirical findings indicate that the risk of investing in a specific SME can be predicted and interpreted well using a machine-learning model which is both predictively accurate and explainable.
Source row: 752 · abstract type: unknown