Shuvo, Md Sabbir Hossen and Alam, Mohammad Kowshik and Hasan, Md Majedul (2024) Decision Shock in Financial AI Systems: A Perturbation-Based Analysis of Prediction Instability and Outcome Sensitivity. American Journal of Technology Advancement, 1 (8). ISSN 2997-9382

Text AJTA_Decision+Shock+in+Financial.pdf Download (822kB)

Abstract

Financial institutions have begun making use of Artificial Intelligence (AI) systems for things such as automated loan approval, credit risk evaluation and borrower assessment. While machine learning models deliver high predictive power, many financial AI systems are still susceptible to prediction instability, which occurs when the prediction probabilities and final lending decisions of financial AI change significantly depending on the small variation of borrower financial information. This volatile performance raises questions about fairness, transparency, robustness and financial stability of automated decision-making processes. This paper introduces an analytical approach to the study of decision shock and outcome sensitivity within the context of financial AI models through a perturbation method. This study examines what happens to the prediction results and decision boundaries when the financial variables of the borrowers are perturbed by some control. The sensitivity of the models is tested for near boundary conditions by making small changes in borrower characteristics, such as an increase or decrease in income, debt ratio changes, loan amount changes, and interest rate changes within borrower profiles. The Lending Club loan dataset is used to train machine learning models such as Random Forest, Logistic Regression, and XGBoost to emulate real-world lending practices. Probability shift analysis, instability scoring and decision boundary flip detection are the measures of prediction instability. The framework also identifies the financial variables that cause the largest decision shocks, and assesses the financial consistency of AI-generated approvals, even when making small variations to the features. Experimental results will show that even with small financial changes, highly accurate financial AI systems can still yield unpredictable and inconsistent results. This study underscores the need for robust assessment of financial AI models, going beyond accuracy measures and paving the way for trustworthy and explainable financial AI systems. In the increasingly dynamic and technologically advanced banking and fintech landscape, the suggested perturbation-based approach can help financial institutions build more trustworthy models, minimize overly automated decision-making, and contribute to fairer lending processes.

Item Type:	Article
Uncontrolled Keywords:	Financial Artificial Intelligence, Prediction Instability, Decision Shock Analysis, Perturbation-Based Learning, Loan Approval Systems and Explainable Financial AI
Subjects:	H Social Sciences
Depositing User:	admin eprints
Date Deposited:	01 Jul 2026 00:49
Last Modified:	01 Jul 2026 00:49
URI:	http://eprints.umsida.ac.id/id/eprint/16697

Actions (login required)

View Item