SD

Sanidhya Dash

M25AI1036 | M.Tech (AI)

Project Focus

Early Risk Detection & XAI

Faculty Guide

Dr. Binod Kumar

Key Technologies

Indian Institute of Technology, Jodhpur

Operational Pain Points

NPAs & Financial Losses: Delinquency leads to increased Non-Performing Assets and significant revenue erosion.
Rigid Rule-Based Systems: Static systems fail to adapt to evolving customer behavior and complex financial patterns.
Black-Box ML Models: Significant lack of explainability hinders trust and prevents effective manual overrides.
Regulatory Compliance: Compliance standards mandate transparent, fair, and fully interpretable AI decision systems.
Early Warning System: Absence of proactive identifiers to detect high-risk customers before actual default occurs.

15%

Avg. NPAs

$2.5T

Global Losses

30%

Default Risk

Technical Challenges

Imbalanced Data: Extreme scarcity of default cases compared to non-default transactions.

Data Leakage: Managing overlapping timeline features that bias model training results.

Trade-offs: Balancing high predictive accuracy with granular feature interpretability.

Temporal Analysis: Capturing shifting financial behaviors over extended sequences.

Proposed Solution Architecture

Phase 1

XGBoost + SHAP Explainer

Phase 2

LSTM Temporal Modeling

Ensemble

Hybrid Decision Framework

Objectives of the Project

Key goals and deliverables for the loan delinquency detection system

🤖

ML Models

Build predictive ML models for delinquency detection using advanced algorithms

🔍

Explainability

Integrate SHAP and LIME techniques for model interpretability

⚖️

Fairness

Evaluate fairness and bias in predictions across different groups

🔄

Hybrid Framework

Develop hybrid XGBoost + LSTM framework for enhanced accuracy

📊

Interactive Dashboard

Build interactive banking dashboard for risk monitoring

Key Features

✓

XGBoost & LSTM Integration

Combined for superior performance

✓

SHAP Explainability

Transparent AI predictions

✓

Real-time Monitoring

Live risk assessment

Dataset and Feature Engineering

Synthetic banking loan records with 55 original features and 11 engineered features

5,000 records

Dataset Details

5,000

Total Records

55 → 66

Features

725

Delinquent

4,275

Non-Delinquent

Delinquent: 725 (14.5%)

Non-Delinquent: 4,275 (85.5%)

Important Features Used

Annual Income

Primary income

Debt-to-Income

Ratio calculation

Loan Amount

Principal borrowed

Interest Rate

Annual percentage

Credit Utilization

Usage ratio

Delinquency Hist.

Past payments

Legal Issues

Bankruptcy/Tax

Installment Load

Payment ratio

Recent Inquiries

Credit checks

Feature Engineering

EMI to Income Ratio

Added

Credit Utilization Ratio

Added

Open Credit Ratio

Added

Delinquency Score

Added

Bankruptcy/Tax Issue

Added

Installment Burden

Added

Target Distribution

Delinquent14.5%

Non-Delinquent85.5%

Phase-1 Architecture and Model

XGBoost-based risk prediction model with SHAP explainability

Pipeline Status:

Active

Phase-1 Pipeline Flow

Data Ingestion

Cleaning

Leakage Removal

Feature Eng.

XGBoost

SHAP

Dashboard

Leakage Columns Removed

Balance

Paid Total

Principal

Interest

Late Fees

Why Leakage Removal Matters

Prevents unrealistic model performance (data leakage) and ensures fair, real-world prediction capability by removing post-repayment information that wouldn't be available at the time of inference.

Model Performance

Accuracy

93.9%

Recall

90.3%

Precision

45.2%

AUC

91.6%

Model Trained Successfully

Key Features

SHAP Explainability

Feature Importance Rankings

Real-time Monitoring Integration

Phase-1 Results - XGBoost Model Performance

High-performing risk prediction model with 93.9% accuracy and 91.6% AUC

Model Status:

Trained Successfully

Performance Metrics

93.9%

Accuracy

+2.1%

90.3%

Recall

+1.5%

45.2%

Precision

+3.2%

60.2%

F1-Score

+2.8%

91.6%

AUC

+1.9%

Model Interpretation

The XGBoost model demonstrates strong delinquency detection capability with high recall (90.3%) ensuring most risky customers are identified. The high AUC (91.6%) indicates excellent separation between risky and safe customers.

Strong delinquency detection: High recall means most risky customers are detected

High AUC score: Strong separation between risky and safe customers

Balanced performance: Good balance between precision and recall

Model Details

XGBoost Classifier

Gradient boosting with 100 estimators

Training Time2.3 minutes

Features Used66

Validation Split80/20

Class Distribution

Delinquent14.5%

Non-Delinquent85.5%

Explainability Techniques Used

SHAP Global

Model-wide feature importance analysis

SHAP Local

Individual prediction explanations

Feature Importance

Visual ranking of key features

Customer Risk

Personalized risk explanations

Value Addition

Transparency

95%

Trust

92%

Compliance

98%

Accuracy

94%

Top Features Driving Delinquency

Debt-to-Income Ratio - Primary risk indicator
Credit Utilization Ratio - Current usage
Delinquency Score - Historical behavior
Installment Burden - Monthly capacity
Historical Failed Payments - Failure count

Key Benefits

Transparent

AI predictions explained

Trust

Banking professionals

Compliance

Regulatory ready

Phase-2 Sequential Model

LSTM-based temporal modeling for capturing evolving customer behavior

6-month sequence

Why Phase-2 Was Needed

Behavior Changes

Customer risk patterns evolve dynamically

Evolving Risk

Static models miss temporal trends

Sequential Analysis

Repayment behavior tracking needed

Model Type

LSTM

Sequence

6 months

Features

12

Hidden Units

128

LSTM Architecture

Input

LSTM Layer 1

LSTM Layer 2

Output

Captures temporal dependencies in customer behavior over 6-month periods, analyzing sequential patterns in repayment history and credit utilization.

Sequence Features

Delinquency Score

Risk indicator

Credit Utilization

Usage ratio

Interest Rate

Annual rate

Recent Inquiry

Credit checks

Failed Payments

Historical data

Payment Amount

Monthly payment

Key Metrics

Sequence Length6 months

Input Features12

Phase-2 Results and Hybrid Framework

LSTM model performance and hybrid risk scoring system

Model Status:

Trained Successfully

LSTM Model Performance

47.3%

Accuracy

+5.2%

62.1%

Recall

+8.3%

16.0%

Precision

+2.1%

25.5%

F1-Score

+4.5%

53.5%

AUC

+6.2%

Hybrid Framework Formula

Final Hybrid Risk Score = 0.7 × XGBoost Risk + 0.3 × LSTM Risk

High Accuracy:
XGBoost provides 93.9%

Temporal Insights:
LSTM captures behavior

Balanced Scoring:
70/30 optimal weight

Hybrid Model Benefits

XGBoost Weight70%

LSTM Weight30%

Combined AUC91.6%

Combined Recall90.3%

Risk Distribution

High Risk15%

Medium Risk25%

Low Risk60%

Dashboard and Impact

Interactive banking dashboard for risk prediction and monitoring

System Status:

Active

Dashboard Features

Upload banking dataset

Predict delinquency risk

XGBoost, LSTM, Hybrid scores

SHAP explainability plots

Early warning alerts

Risk segmentation

Impact of the Project

Reduces NPAs

Improves loan recovery rates

Early Intervention

Prevents defaults proactively

Transparency

Enhances trust in AI decisions

Explainable Machine Learning Framework

Dataset and Feature Engineering

Phase-1 Architecture and Model

Phase-1 Results - XGBoost Model Performance

Phase-2 Sequential Model

Phase-2 Results and Hybrid Framework

Dashboard and Impact