Why the Sub-model Architecture in Machine Learning is the next frontier in Fintech

The landscape of credit lending to small and medium-sized businesses (SMBs) has undergone a significant transformation with the advent of advanced technologies. Traditional credit assessment methods, often characterized by lengthy processing times and rigid criteria, have long been a bottleneck for SMBs seeking timely access to capital.

The ability to make accurate credit decisions efficiently represents a paradigm shift in financial services. This rapid decision-making capability not only enhances operational efficiency but also opens up new possibilities for SMBs to access credit at critical junctures of their business cycles. However, implementing such systems raises important questions about data usage, risk assessment accuracy, and the ethical implications of algorithmic decision-making in financial services.

I want to introduce a system capable of making credit lending decisions for SMBs with improved speed and accuracy. We will explore the architecture of this ML system, including its theoretical data sources, the pseudo algorithms employed, and the key parameters considered in the decision-making process. For reference, this architecture can be widely expanded to other problem sets and domains. The value proposition over traditional monoliths is evident in the model's ability to make quicker decisions and a more comprehensive evaluation of credit risk. 

For the sake of simplicity, let's look at an example where we're trying to evaluate the creditworthiness of a business. Let's keep the credit policy aside for a minute and evaluate the procedural/algorithmic process of evaluation. We're all familiar with how businesses are evaluated in a traditional monolithic pillar-based architecture. Let's look at the same problem with a different lens. 

In this new system, we apply a modular architecture, utilizing two primary submodels: the commercial submodel and the consumer submodel. This approach allows for a more nuanced and comprehensive evaluation of credit risk by separately assessing the business entity and the individual business owner.

The commercial submodel focuses on evaluating the creditworthiness of the business itself. It analyzes various aspects of the business's financial health and operational stability, including cash flow patterns, financial statements, business credit history and scores, industry performance, local economic factors, length of time in business, and external indicators of business health such as online reviews or foot traffic data.

Complementing the commercial assessment, the consumer submodel evaluates the personal creditworthiness of the business owner or primary stakeholder. This submodel considers factors such as personal credit scores, credit history, history of other business successes or failures, personal cash flow, and other assets. By incorporating these diverse personal financial indicators, the consumer submodel provides a comprehensive view of the individual's financial stability and business acumen, which can significantly impact the overall risk assessment of the loan.

The final credit decision is based on the integration of these two submodels. The system combines the outputs of the commercial and consumer submodels, weighing them according to their relative importance as determined by historical data analysis and ongoing performance evaluation. This modular approach offers several advantages, including flexibility to adjust weights based on specific application characteristics, precision in capturing nuances of both business and personal creditworthiness, adaptability to independently update each submodel, and transparency in explaining credit decisions for regulatory compliance and customer communication.

The submodel architecture allows for continuous refinement of each component independently, enabling the system to adapt more quickly to changing economic conditions or emerging risk factors in either the business or personal domains.

Final Scoring and Ongoing Monitoring

The integration of the commercial and consumer submodels culminates in a final score that serves as the cornerstone of the credit decision process. This final score is a sophisticated amalgamation of the outputs from both submodels, carefully weighted to reflect their relative importance in predicting loan performance. The system uses this comprehensive score to determine not only whether an applicant qualifies for credit but also to tailor the specific product offerings and payment terms.

Upon generating the final score, the system categorizes applicants into different risk tiers. These tiers correspond to various credit products, each designed to match the risk profile of the borrower. For instance, applicants with higher scores might qualify for larger loan amounts, lower interest rates, or more flexible repayment terms. Conversely, those with lower scores, while still qualifying for credit, might be offered products with more conservative terms to mitigate the perceived higher risk.

The process doesn't end with the initial credit decision. A crucial component of this technology-driven lending system is the ongoing monitoring of loan performance. As borrowers begin repaying their loans, the system meticulously tracks payment behaviors, focusing particularly on delinquencies. This real-time performance data is invaluable for continually refining the scoring models.

By analyzing the payment patterns of specific customer segments, the system can identify trends and correlations that might not have been apparent in the initial credit assessment. For example, it might discover that certain business types or geographical locations are showing higher than expected delinquency rates, or conversely, that some lower-scored segments are outperforming expectations.

This continuous feedback loop allows for fine-tuning of the scoring models. The weights assigned to various factors in both the commercial and consumer submodels can be adjusted based on actual loan performance data. If, for instance, the system observes that personal credit history is becoming a stronger predictor of loan performance for certain business types, it can increase the weight of this factor in the consumer submodel for those segments.

Moreover, this ongoing analysis helps in identifying new predictive factors that might not have been initially considered. For example, the system might discover that businesses with a strong social media presence tend to have better repayment rates, leading to the incorporation of social media metrics into the commercial submodel.

The adaptive nature of this system ensures that the credit assessment process remains dynamic and responsive to changing economic conditions and evolving business landscapes. It allows the lender to quickly adjust its risk appetite and lending strategies in response to observed performance trends. This not only helps in maintaining a healthy loan portfolio but also in identifying new market opportunities and underserved segments that might be less risky than initially perceived.

Furthermore, this continuous monitoring and adjustment process plays a crucial role in regulatory compliance and risk management. By consistently evaluating the performance of its credit models against actual outcomes, the lender can demonstrate to regulators the robustness and fairness of its lending practices. It also allows for quick identification and mitigation of any emerging risk factors, ensuring the long-term stability of the lending program.

Regulatory and Social Implications

The implementation of advanced algorithmic systems in credit decisioning brings forth significant regulatory considerations and social implications. Regulatory bodies are increasingly focusing on the use of artificial intelligence in financial services, with particular attention to fairness, transparency, and accountability.

A key concern is the potential for algorithmic bias. Models, if not carefully designed and monitored, can perpetuate or exacerbate existing biases in lending practices. To address this, our system incorporates regular bias testing, explainability mechanisms, and human oversight for borderline cases. These features help ensure compliance with regulatory expectations and promote fair lending practices.

Data privacy and security are paramount, given the vast amounts of personal and business data processed. Our system adheres to privacy-by-design principles, ensuring compliance with regulations such as GDPR and CCPA. Additionally, thorough documentation of model development, regular validation, and ongoing performance monitoring align with regulatory expectations for model risk management.

Conclusion

The modular architecture of the system, incorporating both commercial and consumer submodels, allows for a nuanced and comprehensive assessment of risk. By leveraging diverse data sources and advanced analytical techniques, the system can make informed lending decisions efficiently, a feat that was previously unattainable through traditional methods.

One of the key strengths of this system is its ability to continuously learn and adapt. Through ongoing monitoring of loan performance and regular backtesting, the model can refine its assessment criteria, ensuring that it remains relevant and accurate in the face of changing economic conditions and evolving business landscapes.

In conclusion, while technology-driven credit decisioning systems offer tremendous potential for improving efficiency and accessibility, their implementation must be approached with careful consideration of ethical implications and regulatory compliance. As these systems continue to evolve, ongoing research and dialogue will be crucial to ensure that they serve the needs of both lenders and borrowers while maintaining the integrity of the financial system and promoting inclusive economic growth.