AWS Partner Network (APN) Blog
Fast, Accurate, Alternate Credit Decisioning Using ElectrifAi’s Machine Learning Solution on AWS
By Vaibhav Sabharwal, Sr. Solutions Architect – AWS
By Samir Agarwal, EVP, Global Products and Channels – ElectrifAi
 |
| ElectrifAi |
 |
Amazon Web Services (AWS) offers rich and powerful artificial intelligence (AI) and machine learning (ML) services that help customers innovate faster with a comprehensive set of tools. These technologies are easy to use, come with pay-as-you-use convenience, and do not require prior ML expertise. As a result, customers are building new experiences while getting more value from their data.
In the financial services sector, lending departments traditionally undergo a tedious review process while evaluating small and medium-sized enterprise (SME) loan applications. This level of scrutiny is time-consuming, as the financial services provider manually reviews the SME’s profile, documents provided, and their risk level based on historical data.
Such a system is not designed to scale as the volume of applications grows, and keeping the process free from human biases and maintaining consistency in the decision process can pose a challenge.
Infusing machine learning into core business processes such as credit scoring creates a competitive edge for banks and financial services institutions. It does not require a data science team, expertise, or platform rollout.
In this post, we’ll discuss an ML-based credit-decisioning model built by ElectrifAi in collaboration with AWS. The model rapidly determines the creditworthiness of a SME, and data-driven, actionable insights reduce the overall processing cost and are consistent and free from any potential human biases.
The model also flags suspicious data, thereby strengthening the lender’s confidence. It also scales to meet high demands, and the higher acceptance rates can result in increased revenue. In addition, there’s a reduction in bad debt because the model can gauge the likelihood of default.
ElectrifAi is an AWS Machine Learning Competency Partner that is working to change the way businesses work through ML, driving cost reduction as well as process and performance improvement.
Common Data Model with Customized Data Layers
At the core of ElectrifAi’s solution is a common data model that includes a list of data fields with minimum requirements, such as customer application data, credit product performance data, and company information data. It can be mapped to other banking and financial services (BFS) data sources.
Once the data mapping is done, the pre-trained model can be applied and measured on first-party client data. The customized data layers provide the ability to ingest additional data sources from clients, and models can be retrained and fine-tuned on the enhanced data source to boost performance.
The intelligence extracted from these data points creates a holistic view of the SME. Publicly available information, social media sentiment, and other relevant data sources can be incorporated. The model continuously learns and fine-tunes itself, keeping financial institutions resilient in challenging times.
The credit-scoring algorithm can quickly identify low-risk SMEs and recommend them for approval while also flagging SMEs who appear not to be creditworthy. As a result, business experts don’t need to spend too much time on straightforward cases, allowing them to focus on applicants that require their expertise and qualitative assessment.
Figure 1 – Architecture diagram.
Real-Time Credit Decisioning
Below is the process flow to make real-time credit decisioning:
- Customer fills out the credit card application or loan application on the bank’s website.
- The request is received by Amazon API Gateway and sent to AWS Lambda for processing.
- The credit-decisioning model is hosted on Amazon SageMaker as an endpoint for real-time inference. Lambda makes a REST API call to the SageMaker endpoint.
- The ElectrifAi/SageMaker model is the brain of the architecture and is trained to calculate credit risk score for different customer profiles.
- Based on input parameters, the model calculates the risk profile score and sends the score back to Lambda.
- The output from the model is stored in an Amazon Simple Storage Service (Amazon S3) bucket.
- The Lambda function sends back the approved/denied response based on the credit risk profile score.
- A decision notification (email or text message) is sent to the customer using Amazon Simple Email Service (Amazon SES).
- Custom dashboard is built using Amazon QuickSight on the application status.
Batch Credit Decisioning
Below is the process flow to make credit decisioning for multiple customers:
- The back-office team uploads files containing potential customer profiles for the credit card or loan applications.
- The file is transferred to Amazon S3 over Secure Shell File Transfer Protocol (SFTP) using AWS Transfer Family.
- AWS Glue transforms the input file in the required format and stores the formatted output in an S3 bucket for processing.
- Lambda creates an Amazon SageMaker batch transform job to process the records.
- The job reads data from the S3 bucket, uses the ElectrifAi/SageMaker model to calculate the credit risk score, and stores the response in an output S3 bucket.
- A notification is sent to customers pre-qualified for the credit card/loan instrument using Amazon SES.
- The back-office team can review the application decision status on a QuickSight dashboard.
Figure 2 – Sample QuickSight dashboard.
Performance Monitoring, Security, and Compliance
When the ML model is deployed in production, performance monitor workflows run on a regular basis. The output includes model performance reports and model performance alerts.
Performance reports contain accuracy and stability metrics to evaluate whether the model performs well. Model retrain workflows can be triggered based on the client’s requirement to meet the suggested threshold, if needed.
The model adopts and extends security guidance within Amazon SageMaker’s shared responsibility model by protecting the infrastructure from known security vulnerabilities. In addition, the model and its artifacts are always encrypted at rest and in transit.
Finally, the extract, transform, load (ETL) process provides a record of actions taken by the model, user, role, or any services to enable auditing for customer privacy compliance.
Conclusion
The alternate credit decisioning solution detailed in this post is about using the power of machine learning to bring financial fairness by making credit available to applicants without sizeable credit history but with low-risk profiles to be eligible for credit and loan products.
The ElectrifAi/SageMaker solution enhances the customer experience with quick approvals without incurring additional manual processing costs. When customers trust their financial institutions, they remain engaged and have a reason to keep coming back for more.
We have shown how ElectrifAi’s high-value, domain-specific, pre-built ML models available on AWS accelerate business outcomes in weeks. The same architecture can be extended for other financial services use cases such as performing real-time fraud detection, increasing credit line, and offering customized products based on spending patterns, and used in other industries like telecommunications, healthcare, and media and entertainment.
To learn more about model hosting on Amazon SageMaker, refer to the AWS documentation on deploying models for inference.
.
.
ElectrifAi – AWS Partner Spotlight
ElectrifAi is an AWS Machine Learning Competency Partner that is working to change the way businesses work through ML, driving cost reduction as well as process and performance improvement.