Case Study · Flagship Project

Interpretable Credit Risk Scoring

A machine learning pipeline that predicts loan default probability across 300K+ applicants — then makes every prediction explainable to risk managers, auditors, and product teams using SHAP and Tableau.

LightGBM SHAP Python Tableau Optuna Home Credit Dataset Fintech · Risk

↗ Live Dashboard ⌥ GitHub Repo

300K+Loan Applicants

0.25Optimized Threshold

6Dashboard Views

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

Many loan applicants — particularly those with limited or no formal credit history — are denied financing not because they're high risk, but because traditional scoring models can't assess them. The Home Credit dataset represents exactly this population: 300K+ applicants where standard bureau data is thin or absent.

The challenge isn't just prediction accuracy. In regulated financial environments, a model that says "denied" without explanation is unusable. Risk managers need to know why. Auditors need documentation. Product teams need actionable thresholds. A black-box ML model, no matter how accurate, fails all three.

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 a LightGBM classifier with automated hyperparameter optimization via Optuna. Tuned the decision threshold to 0.25 to maximize recall — prioritizing detection of actual defaulters over false 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 all model outputs into clean CSV exports purpose-built for Tableau. Built a 6-view interactive dashboard that lets risk teams explore predictions, SHAP values, and model performance without touching any code.

Tableau PublicCSV ExportsDashboard Design

Key Findings

A threshold of 0.25 rather than the default 0.50 significantly improved recall — catching more actual defaulters at the cost of some precision. In credit risk, missing a true default is far more costly than a false positive.

External credit scores (EXT_SOURCE_2, EXT_SOURCE_3) and credit annuity ratio emerged as the strongest global predictors — consistent with domain intuition, which validates the model's behavior to risk managers.

SHAP values revealed that the same features drive risk differently across Low, Medium, and High risk bands — a finding that wouldn't surface from global importance alone, with direct implications for segmented lending policy.

The confusion matrix shows 843 true positives vs. 4,122 false negatives at threshold 0.25 — a deliberate trade-off favoring default detection in a class-imbalanced dataset (91% non-default).

Exporting structured CSVs from the modeling pipeline into Tableau made the model's outputs accessible to compliance, product, and executive audiences — no code required.

Business Implication

This project demonstrates that interpretability isn't a trade-off against accuracy — it's a requirement for deployment. A model that risk teams don't trust won't be used, regardless of its AUC score.

The SHAP + Tableau layer transforms a machine learning output into a decision support tool that risk managers can act on, auditors can review, and executives can understand. In a regulated lending environment, that's the difference between a proof of concept and a production-ready system.