Credit Risk Prediction with LightGBM, Hyperopt and SHAP

Project Overview

Business Context

Credit risk refers to the potential loss a lender may face when a borrower fails to repay a loan or meet contractual obligations. The goal of a credit risk assessment is to determine if potential borrowers are creditworthy, have the means to repay their debts, and to minimize credit risk. This project involves building a classification model for default prediction using LightGBM, optimizing hyperparameters with Hyperopt, and using SHAP for model explainability.

---

Aim

The aim of this project is to predict loan defaulters and reduce the risk of financial loss by analyzing credit history, employment, and demographic data.

---

Data Description

The dataset includes information on 143,727 borrowers, including attributes such as employment type, work experience, income, dependents, total loans, and total payments.

---

Tech Stack

• Language: Python
• Libraries: pandas, numpy, matplotlib, seaborn, scikit-learn, lightgbm, hyperopt, shap

---

Approach

1. Data Reading
2. Data Processing
   • Drop Columns
   • Split Data
   • Define Label
   • Roll Rate Analysis
   • Window Roll Analysis
3. Feature Engineering
   • Label
   • % Amount Paid as interest in past Loan Repayment
   • % of Loans defaulted in the last 2 years
4. Exploratory Data Analysis (EDA)
   • Univariate Analysis
     • Numerical Summary: Min, Max, Mean, Median, etc.
     • Categorical Summary: Top, Unique, Count, etc.
   • Bivariate Analysis
     • Correlation Plot
     • Box Plots
5. Target Encoding
6. Feature Selection
   • Random Forest
   • Decision Tree
7. ML Model Development
   • LightGBM
   • Hyperparameter Tuning using Hyperopt
8. Model Evaluation
   • ROC AUC
   • PRAUC
   • Score Distribution
9. Feature Importance
   • Split and Gain
   • SHAP
10. Class Rate Curve and Right Threshold

---

Modular Code Overview

1. input: Contains the raw data for analysis, in this case, credit_risk_data.csv.
2. documents: Contains supporting learning material.
3. lib: A reference folder containing the original iPython notebook used in lectures.
4. ml_pipeline: Contains functions in various Python files for processing the data and training the model.
5. output: The folder where the trained model is saved.
6. engine.py: A script that calls the functions in the ml_pipeline to run the entire process and save the model.
7. requirements.txt: Lists required libraries and versions.
8. readme.md: Instructions for running the code.

---