Credit Scoring Final Project

1. Credit Scoring Model

In Microfinance Case
Data Science 4.0
Final Project
13.02.2026
Murataliev Ilias
Bishkek, Kyrgyzstan

2. Project Goal

• Build a credit scoring
model to predict
default
• Optimize cut-off level

3. Dataset Overview

• ~10,000 credit contracts
• Target: Default (1) / Non-default (0)
• Demographic, geographic and loan features
• Target balance 20/80

4. Feature Engineering

• Age instead of birth date
• Dates converted to months
• Seasonality: issue month and quarter
birth_date gender
age
issue_da
repaymen loan_amo interest loan_pur credit_p collater
region district
te
unt
_rate
pose
roduct al_type target
end_date t_type
issue_mon issue_quart
th
er
term_month
s

5. Models and Metrics

Models tested:
• Logistic Regression
• LightGBM
• CatBoost (with
optuna)
• XGBoost (with
optuna)
Metrics:
• AUC, Recall,
Precision
Data transform:
• One Hot Encoder
• Standard Scaler
Train/Test split:
• 0.7/0.3

6. Optimal cut off by KS

Logistic
Regression
AUC
XGBoost
CatBoost
LightGBM
Essemble
0.8143
0.8729
0.8736
0.8656
0.8747
Precision
(0/1)
0.94/0.48
0.95/0.55
0.95/0.56
0.93/0.57
0.94/0.57
Recall
(0/1)
0.77/0.81
0.83/0.83
0.84/0.81
0.82/0.81
0.85/0.80
F1 (0/1)
0.85/0.60
0.88/0.66
0.89/0.66
0.88/0.64
0.89/0.67
KS
0.58
0.65
0.65
0.62
0.65
Cut off
0.568
0.218
0.223
0.408
0.300

7. Optimal cut off by Profit

Assumptions:
• Loss on default = 100% of loan amount
• Income = All annuity payments – loan
amount
• Cut-off selected by maximizing profit
Logistic
Regression
XGBoost
CatBoost
LightGBM
Essemble
Max Profit
3.578
3.633
3.585
3.593
3.642
Cut-off
0.570
0.332
0.332
0.693
0.399
Approval
rate
0.657
0.756
0.754
0.785
0.758

8. Ensemble cut-off profit dependency

9. Ensemble profit approval rate dependency

10. Credit score visualization by proba

Probability of default
Grade
Score
Actions
Description
0.0 – 0.05
A
750 – 850
Auto-approval
Ideal clients. Minimal risk.
0.05 – 0.15
B
650 – 749
Approval
Good clients, minor issues.
0.15 – 0.35
C
550 – 649
Additional check
Grey zone. Additional
check.
0.35 – 0.60
D
450 – 549
More like a refusal
High risk. Can be approved
at a max allowed interest
rate.
> 0.60
E
300 – 449
Auto-refusal
Clients with a high
probability of default.

11. What can be improved

• More features in dataset: History of overdue,
Credit history, Employment info etc.
• More detail analyze from financial perspective
Cut-off criteria

12. Key Takeaways

• Gradient Boosting models showed best
performance
• Business-driven cut-off differs from KS cut-off
• Model is applicable for MFOs in Kyrgyzstan
• Solution can be scaled and deployed

13. Thank You