This case study details the successful implementation of an automated credit decisioning system for one of our clients, a small loan company. The project’s goal was to modernize the client’s traditional credit evaluation processes using data-science, enhancing the speed, accuracy, and fairness of their credit decision-making.

Client background

Prior to our engagement, their process was largely manual, involving significant human intervention that led to slower decision times and potential for inconsistency.

Project objectives

• Increase Efficiency: Reduce the time required to reach credit decisions.

• Improve Accuracy: Enhance the predictive accuracy of credit default risks.

• Ensure Fairness: Minimize biases in credit evaluations to promote fairness across all applicant demographics.

• Ensure explainability : Fully able to explain the decision makin process

Solution and Design

The solution involved several key phases:

1. Data Integration: We integrated data from multiple sources, including financial histories, transaction records, and demographic information, ensuring comprehensive data collection.

2. Model Development: Utilizing Python’s data science libraries (Pandas, Scikit-learn, xgboost), we developed several predictive models:

   • Logistic Regression for baseline predictions.
   • Gradient Boosted Trees and Random Forests for more nuanced decision paths.

3. Model Deployment: The models were deployed within the client’s AWS cloud infrastructure, using FastApi providing real-time credit scoring capabilities.

4. Continuous Monitoring and Updates: We established a framework for ongoing monitoring and regular updates to the models to adapt to new data and emerging trends.

Implementation

Implementation details included:

• Data Preparation: Data cleansing and preparation were performed using Python, ensuring high-quality inputs for model training.

• Model Training and Validation: Models were trained on historical data, with rigorous validation using techniques like cross-validation to ensure they generalised well on unseen data and stratritified sample to ensure well balanced sample

• Integration into Existing Systems: The models were integrated into the client’s existing loan application platforms through APIs, allowing for seamless, real-time decision-making.

• Training and Support: Key client personnel were trained on the system’s functionality, including how to interpret model outputs and perform routine system checks.

Results

The deployment of the automated credit decisioning system yielded significant improvements:

• Decision Speed: Credit decisions were much faster than the previous manual system enabling the team to drastically increase the volume of applicant.

• Accuracy Improvement: The accuracy of predicting credit default improved by approximately 25%, reducing financial risks.

• Reduction in Operational Costs: By automating the decision process, the client saved significantly on operational costs associated with manual reviews.

Challenges and Solutions

• Data Privacy: Rigorous data encryption and anonymization techniques were employed to protect sensitive information.

• Bias Mitigation: Regular audits were conducted to check for biases in the models, and adjustments were made to ensure fairness across different demographic groups.

Conclusion

The implementation of automated credit decisioning systems exemplifies how advanced data science techniques can transform traditional business operations. This project not only enhanced the efficiency and accuracy of credit scoring but also reinforced the agency’s commitment to fairness.

As data science continues to evolve, its integration into credit scoring will likely deepen, leading to more innovative and refined approaches to risk management in finance. This synergy not only benefits lenders through reduced default risks but also facilitates broader access to credit for consumers by allowing for more nuanced assessments of creditworthiness.

Some key players in the market are Provenir, Zest AI, Scienaptic, Alloy.