Credit Risk Scoring & Dashboard

Author: Or Ben‑Haim
Stack: Python · PyTorch · Scikit‑learn · Pandas · Tableau

Overview

Predicting loan risk grades from applicant features and exploring results in a Tableau dashboard.
Data is sourced from Kaggle and processed into a dataset that feeds the dashboard. The model is a fully‑connected neural network (PyTorch) for multi‑class classification, evaluated both technically and with business‑oriented views.

Links

• Kaggle dataset: https://www.kaggle.com/datasets/laotse/credit-risk-dataset
• Public Tableau dashboard: https://public.tableau.com/views/LoanGradePredictionAnalysis/LoanGradesPopulationInsightsModelvsActual

Repository Structure

• data/raw/ – Original Kaggle CSV
• data/processed/ – Processed dataset used by the dashboard (loan_predictions_final_clean.xlsx)
• dashboard/ – Tableau workbook (Loan Grade Prediction Analysis.twb) + screenshot
• notebooks/ – _End_to_End_Credit_Risk_Scoring_From_Data_Cleaning_to_Dashboard (cleaning → modeling → export)
• reports/End-to-End Credit Risk Scoring From Data Cleaning to Business Dashboard with a Neural – Final PDF report

Model

• Type: Fully‑connected neural network (feed‑forward MLP) in PyTorch
• Task: Multi‑class loan grade classification
• Label handling: Collapsed rare grades to 4 classes (A–D) for stability
• Highlights: ~88% accuracy after collapsing; most confusion between adjacent grades (A↔B, C↔D)

How to Reproduce

1. Open notebooks/End to End Credit Risk Scoring.ipynb and run the steps to produce data/processed/loan_predictions_final_clean.xlsx.
2. Open dashboard/Loan Grade Prediction Analysis.twb in Tableau Desktop (or use the public link above) to explore Model vs Actual views.

Results (short)

• Balanced performance across classes after label merge (~88% overall accuracy).
• Dashboard surfaces error patterns and the impact on loan amounts and interest rates.

Data Source

Kaggle — Credit Risk Dataset by laotse: https://www.kaggle.com/datasets/laotse/credit-risk-dataset

License

MIT © 2025