Case Study · Decision-Support Prototype

Interpretable Credit Risk Scoring

A proof-of-concept analysis of loan-default risk across 300K+ applications using LightGBM, SHAP explanations, threshold evaluation, and a stakeholder-facing Tableau dashboard.

LightGBM SHAP Python Tableau Optuna Home Credit Dataset Fintech · Risk

↗ Live Dashboard ⌥ GitHub Repo

300K+Loan Applicants

0.7602Validation ROC-AUC

5Core Visualizations

SHAPExplainability Layer

Live Dashboard

Interpretable Credit Risk Dashboard — Tableau Public

↗ Open full screen

Loading dashboard...

⚡ Fully interactive — use tabs, filters, and tooltips directly above · open in Tableau Public for best experience

0 — Overview

↗ Open in Tableau

Overview — Dashboard home with navigation to all 5 views

SHAP Global Importance — Top features ranked by mean absolute SHAP value across all applicants

Score by Actual Outcome — Predicted risk score distributions by true default vs. non-default

SHAP by Risk Group — Feature impact segmented across Low, Medium, and High risk bands

Feature Impact on Score — SHAP contribution vs. predicted probability for each applicant

Confusion Matrix & Metrics — Prediction outcomes and model performance breakdown for auditors

The Problem

The historical Home Credit Default Risk dataset is large, highly imbalanced, and suited to exploring how a classification model separates default and non-default outcomes.

The analytical challenge was broader than ranking performance: compare precision–recall trade-offs at different decision thresholds and translate model behavior into explanations and visual outputs that non-technical reviewers could inspect.

Approach

Data Cleaning & Feature Engineering

Handled missing values, encoded categorical variables, and addressed class imbalance in the raw Home Credit dataset. Structured outputs into analysis-ready formats.

PandasNumPyClass ImbalanceEncoding

LightGBM Modeling with Optuna Tuning

Trained and tuned a LightGBM classifier with Optuna, achieving a validation ROC-AUC of 0.7602. Later notebook experiments compare multiple decision thresholds to show the trade-off between default recall and precision.

LightGBMOptunaThreshold TuningAUC-ROC

SHAP Explainability Layer

Applied SHAP to generate both global feature importance and individual prediction explanations. Computed SHAP values by risk band so stakeholders can see how risk drivers differ across segments — not just the global picture.

SHAPGlobal ImportanceRisk SegmentationLocal Explanations

Tableau-Ready Exports & Dashboard

Structured model outputs into Tableau-ready CSV exports and built five core visualizations covering model results, feature importance, risk bands, and prediction outcomes.

Tableau PublicCSV ExportsDashboard Design

Key Findings

The LightGBM model achieved a validation ROC-AUC of 0.7602, providing a threshold-independent measure of ranking performance.

Threshold experiments demonstrate that improving default recall reduces precision; an operating threshold would require explicit business costs and risk-policy requirements.

EXT_SOURCE_2 and EXT_SOURCE_3 were among the strongest global predictors, while SHAP views show how feature contributions vary across modeled risk bands.

The Tableau dashboard reflects the model export used during dashboard development; later notebook threshold experiments can therefore produce different threshold-dependent counts.

Tableau-ready exports make model results and explanations reviewable without running the notebooks, supporting communication rather than automated lending decisions.

Decision-Support Value

The project demonstrates how model evaluation, threshold analysis, and SHAP explanations can be combined with a dashboard to support review and discussion.

It remains a portfolio proof of concept, not a deployed underwriting system or evidence of real-lender business impact.

Limitations

The dataset comes from a past Kaggle competition and may not represent a current lending population. Results depend on the selected threshold and assumed costs of false positives and false negatives.

ROC-AUC and SHAP explanations do not establish calibration, fairness, causality, or production readiness. Real lending use would require current representative data, fairness and calibration testing, governance, monitoring, human oversight, and regulatory validation.