FinTech Lending Default Risk: A Stacked Ensemble

Credit Risk Framework with SHAP-Based Interpretability

Asifa Iqbal1,* Shahid Mahmood2

1 School of International Languages, Zhengzhou University, Henan, China

2 School of Finance and Economics, Jiangsu University, Zhenjiang, People’s Republic of China

Received: January 03, 2026 Revised: February 25, 2026 Accepted: May 09, 2026 ⋆ Corresponding author

ABSTRACT

The quick expansion of digital lending platforms has, you know, put borrower default risk right at the center of

attention for FinTech investors, regulators, and researchers too. Even though machine learning is now quite often

used for credit risk prediction, many models still focus mostly on predictive accuracy rather than interpretability, and

also, single-algorithm setups don’t really take advantage of the extra value you’d get from using several different

classifiers together. In this work, we put forward the Stacked Ensemble Credit Risk Framework (SECRF). It’s

basically a multi layer architecture that merges base learners—logistic regression, random forest, and gradient

boosted trees—then routes them to a meta learner, plus an added layer for SHAP based post hoc interpretation and

calibration checkups. Across several evaluation perspectives, SECRF shows solid and fairly stable results, and in

cross validated experiments the area under the ROC curve stays higher than what single model alternatives typically

deliver. When it comes to what actually drives defaults, loan grade along with interest rate emerge as the main

predictors under all estimation setups, which lines up with how credit risk is theoretically priced. At the borrower

level, the FICO score and debt-to-income ratio also matter in a clear way, while macroeconomic proxies add extra

explanatory signals beyond what individual loan descriptors already capture. The interpretability layer is there on

purpose so the approach stays compatible with regulatory requirements, and so credit officers get insights they can

act on. Finally, the calibration analysis indicates that SECRF outputs are dependable, especially for risk based

pricing and provisioning decisions.

Keywords: FinTech Peer-to-peer lending Credit risk Default prediction Stacking ensemble SHAP Machine

learning

1. INTRODUCTION

Digital lending platforms basically show up and, like, completely

mess with the whole credit intermediation picture.

You have peer-to-peer , or P2P , lenders, buy now pay later

providers, and those marketplace lenders that originate credit

at scale well outside the usual banking sector, but somehow

without the institutional risk controls that banks worked out

over decades [1, 2]. When borrowers default here, it doesn’t

stay contained, the repercussions hit investors directly too,

and they also spill over into the broader financial stability

issues that [3] and [4] discuss. So predicting default becomes

not just, uh, a kind of private risk management need , but also

a macroprudential worry.

Machine learning has basically become the main methodological

route for credit risk classification, largely because

it can catch non linear feature interactions that models with

fixed, parametric forms just can’t really represent [5, 6]. Still,