Redefining a bank's risk assessment approach, leading to unprecedented levels of accuracy and efficiency.
SoFi (NASDAQ: SoFi) is a member-centric digital financial services platform, offering a comprehensive suite of products to its 5.2 million members. Through its three business segments – Lending, Financial Services, & its Technology Platform – SoFi aims to help people achieve financial independence by providing fast access to essential tools in one app.
Amidst a recovering economy, SoFi faced challenges in mitigating fraud and risk across its credit card and personal loan products, requiring data-driven solutions and machine learning model development. Through an existing relationship & a history of project success, SoFi turned to evolv for support & AI/ML Ops expertise.
Our team collaborated closely with the client to develop a comprehensive risk management data-science modeling solution. Our approach involved understanding the specific needs and requirements of SoFi’s business stakeholders. Through a meticulous evaluation process, we presented a range of options and recommendations for fraud and risk mitigation solutions.
To address fraud risk in the application funnels for personal loans and credit cards, we formulated effective strategies that would minimize potential threats. Additionally, we evaluated third-party vendor scores to identify the most suitable candidates for integration into SoFi’s fraud strategies.
The next step involved building machine learning models that would accurately identify and restrict risky credit card payments, as well as rate the riskiness of personal loan applications. Every model was rigorously assessed, including legacy models at SoFi, to ensure fair lending procedures and prevent any disparate impact on protected classes.
- Reduced held credit card payments from 20% to 2.5% by building a machine learning model that decreased the hold rate without increasing returned payments.
- Saved $1.2M in fraud losses by developing innovative fraud prevention strategies using ensemble learning algorithms, resulting in a 23% reduction in fraudulent credit card applications & effectively protecting investor interests.
- Promoted fairness and compliance by demonstrating model disparate impact on protected classes & implementing alternative model search algorithms, ensuring the fairest model for stakeholders while maintaining performance & business impact.
- Outperformed 3rd-party fraud models by creating in-house supervised learning models for personal loan application fraud detection, leading to a 4.8% increase in the model AUC metric.