FinTrustX: Explainable, Threshold-Aware Credit Risk Assessment on the Home Credit Default Risk Dataset
Year: 2026
Type: preprint
Abstract: Credit models usually get judged on one number, and that number hides most of what a lender cares about. This paper describes FinTrustX, an explainable credit-risk pipeline built on the Home Credit Default Risk data, and asks two questions: which of five model families discriminates best, and which decision threshold makes a model usable? We benchmark a decision tree, a random forest, XGBoost, CatBoost and a multilayer perceptron on 245 engineered features and 61,503 held-out applicants, 4,965 of whom defaulted (8.07%). XGBoost came out on top with ROC-AUC 0.779 and PR-AUC 0.275, about 3.4 times the no-skill baseline. The headline number is only half the story. At the conventional Rajveer Dept. of CSE (Cybersecurity) IILM University Greater Noida, India [email protected] FinTrustX is our attempt to build that whole path in one codebase, from raw applications to a ranked list of reasons. This paper reports what the pipeline does today, which is less than the platform name implies. Modelling, evaluation and the explanation step run. The serving layer, probability calibration and fairness audits are not finished. We keep our claims inside that boundary. The contributions are as follows. 0.50 cut-off XGBoost catches just 4.0% of defaults, because its scores are calibrated and few applicants reach 0.50 when only 8% default. Under a simple cost model its best cut-off falls from about 0.5 to 0.02 as a missed default becomes 1 to 50 times costlier than a wrong decline, tracking the 1/(1+r) rule. The other four models, whose raw scores overstate risk, miss that rule until they are recalibrated, and calibration does not lower their realised cost. At equal costs no model beats approving everyone by more than 0.8%. At the chosen cut-off, men are flagged 37.6% of the time against 22.9% for women. SHAP and LIME attributions feed a capped reason-code design for human review. Explanation quality, a gender- and age-free refit and a serving layer are not finished, and we say so.
Authors: Kashyap, anurag, Sharma, Hritik, Rajveer
Language: English
Subfield: Accounting
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225758718
Cites:
Cited by:
Related to: