Evaluating Imputation under Missingness Mechanisms for Credit Risk Modeling

  • Marcelo José Wecchi UNIFESP

Resumo


A falta de dados é um desafio persistente na modelagem de risco de crédito, onde o mecanismo que gera a falta de dados molda tanto o desempenho preditivo quanto a justiça. Este estudo estabelece uma linha de base ao comparar cinco métodos de imputação – Média, Mediana, Moda, KNN e MICE – em três mecanismos (MCAR, MAR, MNAR) usando o conjunto de dados de crédito alemão. Os resultados mostram que o MICE tem melhor desempenho sob MCAR e MAR, enquanto a Moda se mostra mais robusta sob MNAR. Os resultados de justiça variam de acordo com o mecanismo, com o MAR geralmente produzindo disparidades mais baixas. No geral, os resultados demonstram que nenhum método de imputação único é universalmente ideal e que a seleção consciente dos mecanismos é essencial para equilibrar a precisão e a equidade nas aplicações de risco de crédito.

Referências

Caton, S., Malisetty, S., and Haas, C. (2022). Impact of imputation strategies on fairness in machine learning. Journal of Artificial Intelligence Research, 74:1011–1035.

Crook, J. and Banasik, J. (2004). Does reject inference really improve the performance of application scoring models? Journal of Banking and Finance, 28(4):857–874. Retail Credit Risk Management and Measurement.

Hardt, M., Price, E., and Srebro, N. (2016). Equality of opportunity in supervised learning. In Proceedings of the 30th International Conference on Neural Information Processing Systems, NIPS’16, page 3323–3331, Red Hook, NY, USA. Curran Associates Inc.

Hofmann, H. (1994). German credit data.

Khan, F. A., Herasymuk, D., Protsiv, N., and Stoyanovich, J. (2025). Still more shades of null: An evaluation suite for responsible missing value imputation.

Kwon, S., Jang, J., and Kim, C. O. (2025). Credit scoring using multi-task siamese neural network for improving prediction performance and stability. Expert Systems with Applications, 259:125327.

Leite, L. R. (2024). Comparative study of imputation methods for credit card delinquency prediction. Master’s thesis, USP - Universidade de São Paulo, São Carlos, São Paulo, Brazil.

Li, B., Gao, R., Chen, Z., and Wang, C. D. (2025). Inference of rejection in credit scoring using the nrmpi method: A non-random missing data approach. In Digital Finance, pages 75–95, Singapore. Springer Nature Singapore.

Nastoska, A., Jancheska, B., Rizinski, M., and Trajanov, D. (2025). Evaluating trustworthiness in ai: Risks, metrics, and applications across industries. Electronics, 14(13).

Nogueira, I. (2025). Balancing accuracy and explainability in credit scoring: a comparative study of machine learning models and imputation methods. Master’s thesis, Orebro University School of Business.

Putri, K. A., Hervani, N. M., Purwani, F., et al. (2025). Overcoming bias from missing values in microcredit data by comparing mice and imputation methods. International Journal of Multidisciplinary Reseach, 1(5):371–380.

Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3):581–592.

Sue, K.-L., Tsai, C.-F., and Tsau, H.-M. (2024). Missing value imputation and the effect of feature normalisation on financial distress prediction. Journal of Experimental & Theoretical Artificial Intelligence, 36(8):1467–1483.

van Buuren, S. and Groothuis-Oudshoorn, K. (2011). mice: Multivariate imputation by chained equations in r. Journal of Statistical Software, 45(3):1–67.

Zhang, M., Xiao, Z., Wang, N., Fu, X., Cheng, Y., Wang, W., and Zhao, J. (2026). Jump: Jointly utilizing missingness pattern in tabular data. In ICLR 2026 International Conference on Learning Representations, Rio de Janeiro.

Zhang, Y. and Long, Q. (2021). Fairness in missing data imputation.
Publicado
19/10/2026
WECCHI, Marcelo José. Evaluating Imputation under Missingness Mechanisms for Credit Risk Modeling. In: ENCONTRO NACIONAL DE INTELIGÊNCIA ARTIFICIAL E COMPUTACIONAL (ENIAC), 23. , 2026, Cuiabá/MT. Anais [...]. Porto Alegre: Sociedade Brasileira de Computação, 2026 . p. 635-645. ISSN 2763-9061. DOI: https://doi.org/10.5753/eniac.2026.30078.