Credit Risk Modeling — Binary Default Prediction

---

Project Overview

This project builds a binary credit risk classifier on ~42,000 real-world borrower records to identify whether a loan applicant is High Risk (likely to default) or Low Risk (safe to approve).

The original problem was a 4-class classification (P1–P4). After deep data analysis, P3 was found to be a boundary class — 55.6% of P3 samples overlapped with P2 and 99.5% overlapped with P4 on key features. This made multi-class separation structurally impossible. The problem was redesigned as binary classification, which is also more aligned with real banking decisions.

---

Business Problem

Banks lose crores of rupees when high-risk borrowers default on loans. A model that can identify these borrowers before loan approval can prevent significant financial loss.

"Should we approve this loan application — or is this borrower likely to default?"

---

Dataset

Property	Detail
Source	Two case study Excel files (case_study1.xlsx, case_study2.xlsx)
Raw records	51,336
After cleaning	42,064
Features (final)	54
Target column	Approved_Flag → binary (0 = Low Risk, 1 = High Risk)
Class distribution	Low Risk 74%, High Risk 26%

---

Project Pipeline

Raw Data (2 files)
    → Data Cleaning (-99999 removal, column drops)
    → Merging on PROSPECTID
    → EDA (Chi-square, ANOVA, VIF analysis)
    → Feature Selection (54 features retained)
    → Feature Engineering (enq_pressure, deliq_income_ratio, PL_seeking, borrow_recency)
    → Label Encoding (categorical variables)
    → Binary Target Creation (P1+P2 = 0, P3+P4 = 1)
    → Train/Test Split (80/20)
    → Model Training (LR, RF, XGBoost, LightGBM)
    → Threshold Tuning
    → SHAP Explainability
    → Business Impact Analysis

---

Data Cleaning

• Identified -99999 as null placeholder across both datasets
• Dropped 8 columns with more than 10,000 missing values
• Removed remaining rows with -99999 values
• Final dataset: 42,064 records, 54 features
• Zero actual null values after cleaning

---

Why Binary (Not Multi-class)?

Multi-class (P1–P4) was attempted first. P3 recall was stuck at 40–47% despite:

• Class weights and SMOTE
• Threshold tuning
• Feature engineering (enquiry pressure, PL seeking score, etc.)

Root cause analysis revealed:

P3 samples overlapping with P2 range: 55.6%
P3 samples overlapping with P4 range: 99.5%

P3 is a "middle zone" class with no clear feature boundary. Binary conversion resolved this structurally.

---

Feature Engineering

Four new features were created based on domain logic and SHAP analysis:

Feature	Formula	Business Meaning
enq_pressure	enq_L3m / (time_since_recent_enq + 1)	Recent loan seeking intensity
deliq_income_ratio	max_recent_level_of_deliq / (NETMONTHLYINCOME + 1)	Delinquency relative to income
PL_seeking	pct_PL_enq_L6m_of_ever * PL_enq_L12m	Personal loan seeking behaviour
borrow_recency	1 / (time_since_recent_enq + 1)	How recently borrower applied

All four engineered features appeared in the top 15 SHAP features, confirming they added value to the model.

---

Models Trained

Model	Test Accuracy	High-Risk Recall	High-Risk Precision	ROC-AUC	Train-Test Gap
Logistic Regression	81.5%	76%	49%	0.8429	0%
Random Forest	85.8%	82%	63%	0.9069	1%
XGBoost	87.1%	80%	74%	0.9288	3%
LightGBM	87.6%	80%	75%	0.9318	3%

Final model: LightGBM — highest ROC-AUC (0.9318) and Average Precision (0.8706) among all models.

CatBoost's main advantage is native categorical handling. Since all features were pre-encoded, that advantage was unused. LightGBM outperformed on ROC-AUC and Average Precision in this setting.

---

Threshold Tuning (LightGBM)

Threshold	Recall	Precision	Accuracy
0.25	0.87	0.67	84%
0.30	0.83	0.72	86%
0.35	0.80	0.75	87%
0.40	0.78	0.77	87%
0.45	0.74	0.80	87%
0.50	0.72	0.82	88%

Chosen threshold: 0.35 — best balance for a lending context where missing a defaulter is costlier than a wrong rejection.

---

Final Model Performance

Metric	Value
Test Accuracy	87.6%
High-Risk Recall	80%
High-Risk Precision	75%
F1 Score	0.77
ROC-AUC	0.9318
Average Precision (AP)	0.8706
Train-Test Gap	~3%

---

ROC-AUC Comparison

Model	ROC-AUC	Avg Precision
Logistic Regression	0.8429	0.6839
Random Forest	0.9069	0.8280
XGBoost	0.9288	0.8648
LightGBM	0.9318	0.8706

---

SHAP Explainability

Top features driving High Risk prediction:

Rank	Feature	SHAP Value	Business Meaning
1	enq_L3m	1.20	Enquiries in last 3 months — strongest financial stress signal
2	Age_Oldest_TL	0.65	Older credit history = more trustworthy borrower
3	pct_PL_enq_L6m_of_ever	0.50	% of personal loan enquiries recently
4	deliq_income_ratio	0.46	Delinquency level relative to income
5	max_recent_level_of_deliq	0.42	Maximum recent delinquency severity
6	num_std_12mts	0.41	Standard accounts in last 12 months
7	enq_pressure	0.33	Engineered: recent loan seeking intensity

Key SHAP findings:

• High enq_L3m strongly pushes a borrower toward High Risk — borrower enquiring 3+ times in 3 months signals urgent financial need
• Low Age_Oldest_TL signals new-to-credit borrower — bank has limited credit history to assess risk
• High pct_PL_enq_L6m_of_ever means borrower is heavily seeking personal loans recently — stress signal
• Engineered features deliq_income_ratio and enq_pressure both appear in top 7 — feature engineering added real value

---

What the model is actually doing?

After seeing SHAP plots here is the conclusion.

enq_L3m is the major driver of the model with SHAP value of 1.20. Higher enquiry leads to risk according to Beeswarm plot — if a customer is enquiring for a loan a lot, it is risky to give him a loan.

Age_Oldest_TL is the second major driver. If the age of oldest account is greater, the customer is safer — because bank has some credit history information about them.

pct_PL_enq_L6m_of_ever is the third major driver with 0.50 SHAP value. Higher personal loan enquiry in recent months leads to higher default probability.

Single borrower explanation (index 1): Borrower has +1.97 SHAP for enq_L3m — enquired 3 times in 3 months. Age_Oldest_TL is 6 months — very new to credit. 50% of all their enquiries are for personal loans. Model predicted 3.369 vs dataset average of -1.705. Classified as High Risk.

---

Business Impact

Confusion Matrix (Test Set — 8,413 borrowers)

	Predicted Low Risk	Predicted High Risk
Actual Low Risk	5,404 (TN)	655 (FP)
Actual High Risk	468 (FN)	1,886 (TP)

Financial Impact

Metric	Value
Avg loan amount assumed	Rs. 5,00,000
Loss Given Default	40%
Cost per wrong rejection	Rs. 2,000
Loss avoided (test set)	Rs. 37.72 Cr
Loss still at risk (missed)	Rs. 9.36 Cr
Cost of wrong rejections	Rs. 13.1 Lakh
Net benefit (test set)	Rs. 37.59 Cr
Net benefit (42K portfolio)	Rs. 187.65 Cr

Risk Bucket Analysis

Risk Bucket	Borrowers	Actual Default Rate	Action
Very Low (0–25%)	5,409	5.8%	Auto Approve
Low-Medium (25–50%)	930	37.5%	Manual Review
Medium-High (50–75%)	627	59.8%	Conditional Reject
Very High (75–100%)	1,447	90.9%	Auto Reject

Model is able to classify 80 risky customers out of 100 actual risky customers. Borrowers with very high risk score (75–100%) have 90.9% actual default rate — almost every borrower in this bucket is genuinely risky. Borrowers in very low bucket (0–25%) have only 5.8% default rate — safe to approve directly.

---

Business Decision Rule

Probability Score    Action
─────────────────────────────────────
0% – 25%            Auto Approve
25% – 50%           Send for Manual Review
50% – 75%           Conditional Reject
75% – 100%          Auto Reject

---

Tech Stack

Language    : Python 3.10
Final Model : LightGBM (GPU accelerated)
All Models  : LightGBM, XGBoost, Random Forest, Logistic Regression
Libraries   : scikit-learn, pandas, numpy, matplotlib, seaborn, shap, lightgbm
GPU         : NVIDIA Tesla T4 (Google Colab)
Platform    : Google Colab

---

Repository Structure

credit-risk-classification/
│
├── Data_Cleaning_Model_Building.ipynb   ← Main notebook
├── case_study1.xlsx                     ← Raw data file 1
├── case_study2.xlsx                     ← Raw data file 2
├── lgbm_model.pkl                       ← Saved final model
├── shap_feature_importance.png          ← SHAP bar plot
├── shap_summary.png                     ← SHAP beeswarm plot
├── shap_waterfall.png                   ← Single borrower explanation
├── roc_curve.png                        ← ROC curve comparison
├── pr_curve.png                         ← Precision-Recall curve
└── README.md

---

How to Run

# 1. Clone the repository
git clone https://github.com/05Aniket598/credit-risk-classification.git

# 2. Install dependencies
pip install lightgbm xgboost catboost scikit-learn shap pandas numpy matplotlib seaborn openpyxl

# 3. Open notebook
jupyter notebook Data_Cleaning_Model_Building.ipynb

---

Author

Aniket Yadav B.Sc. Data Science & AI | Mumbai LinkedIn | GitHub