CASE STUDY

Mitigating Default Loan Risks With Microsoft Azure-based Loan Risk Analytics Solution For A Finance Organization

Customer Profile
Location:
Central US
Industry:
Finance
Vertical:
Risk Management
Duration:
3 Months
Technologies Used
Azure Data Factory
Azure Data Lake Storage
Azure Databricks
Cosmos DB
Azure Kubernetes Service
Power BI
Key Outcomes
76%
Accuracy in predicting default loans
79%
Recall rate
4x
Improved loan risk analysis

Problem Statement

Financial organizations are at risk of default or bad loans. This type of credit default risk arises due to a borrower’s failure to repay a loan. Lending companies need to assess the risk of bad debts in order to avoid holding risky assets.
Credit risk analytics is not new to the fintech sector, but the manual approach has been replaced by an advanced machine learning approach. Our finance sector client relied upon manual credit risk analytics and faced massive financial losses. The major issues were dependency on human experts and the risk of getting out of sync with market changes.
The company was facing 38% cost overheads with constant rework, inefficient data management, and cumbersome reporting. Manual loan risk assessment also required more time, affecting the efficiency of loan processing cycles. We noticed the following technical and operational issues with the fintech client.
Operational
  • Dependency on human experts
  • Risk of getting out of sync
  • Time consuming manual analytics 
Technical
  • Inability to produce data-driven insights
  • No sync with dynamic trends, micro markets and latest data

Solution

Effective Credit Risk Analytics With Microsoft Azure ML, Data Bricks and XGBoost
Credit risk assessment has transformed surprisingly with data-driven technologies. ML brings higher accuracy, faster results, and data-driven insights to analyze default credit risks and helps to prevent financial setbacks to the lending organization.
After analyzing our client’s preferences and technology stack, we suggested a Microsoft Azure Databricks solution using XGBoost and logistic regression. We built and deployed two models and noticed that the XGBoost-based model delivered superior results in terms of accuracy.
The dataset for experiments comprises different variables such as age, income, employment length, loan purpose, loan amount, loan grade (categorical variable from A to G), interest rate, loan-to-income ratio, previous default, and loan status (target variable).
Loan grade is assigned on the basis of factors such as credit history, quality of the collateral and likelihood of repayment of principal and interest. E.g. Grade A denotes loans with lower expected returns, lower expected losses, and lower interest payments whereas grade G denotes loans with higher expected returns, higher potential losses, and high interest rates. 
With data exploration and data preprocessing operations on the dataset, we came across the following key outcomes—
  • Dataset primarily consisted of non-defaulters, which meant high class-imbalance
  • Loan grades A and B are most common, whereas F and G grades are seldom
  • Home renters defaulted more as compared to homeowners
Initially, we created a logistics regression model with borrower income data and a binary target variable - possibility of defaulting - yes / no. The model worked well with linear data, but with non-linear data patterns, the logistic regression model had limitations. Hence we decided to go with the XGBoost model.
XGBoost is a decision-tree-driven ensemble ML algorithm based on a gradient boosting framework. We chose the XGBoost algorithm due to its compatibility with Azure clusters and our client’s technology preference - Python and Microsoft Azure. 
With the above results and other experiments out of the scope of the project, our experts finalized the XGBoost algorithm with its better F1 score and precision accuracy. The XGBoost-based model had the highest AUC (~0.65), and it performed the best in classifying loan applications in both default and no-default classes, the objectives our client was keen on.
Attri’s loan risk assessment solution helped our fintech client to determine the verdict of loan applications. The solution classified loan applicants at risk as possible defaulters and denied such risky loan applications with further insights, which can be viewed in Power BI.
The model ensures a great degree of explicability for suggested at-risk applicants. The credit risk assessment solution is scalable to incorporate dynamic creditworthiness factors - microeconomic conditions, market fluctuations, and credit score changes. 
Attri delivered a battle-tested ML solution that helped the client to optimize their processes and drive higher ROI . We also strengthened the client’s in-house team with UX training, documentation, and complete knowledge transfer.

Results

  • Enhanced credit risk assessment
  • Faster loan risk assessment processing
  • Minimum manual dependencies and intervention to analyze bad debts. 
  • Reduced risk of holding risky assets

Summary

A financial organization implemented Attri’s machine learning solution for loan risk assessment. The XGBoost-based ML solution helped our client in finding loan applicants at risk and denying such applications. The solution reduces manual efforts and time to evaluate bad loans. The solution also incorporates market changes affecting the creditworthiness of an applicant.

Mitigate Loan Default Risks By Saying No To Defaulters

Schedule A Meeting
arrow icon