How a fintech client automated their processing while reducing credit risk
Our client Finiata is a fast growing financial startup which provides working capital to small and medium scale enterprises (SME) on the basis of their outstanding invoices/purchase orders (termed "factoring").
These invoices can range from a few thousand dollars to hundreds of thousands of dollars. Hence Finiata's ability to scale while reducing the volatility of their monthly profits depend heavily on their ability to put in place an automated process for evaluating the amount of capital that can be advanced to an SME based on the probability of non-payment of an invoice. Further, faster processing of incoming requests provides a key competitive advantage to the firm.
Finiata needed a way to quickly and automatically access the risk of each invoice that they got from their customers. Their requirement was for a machine learning framework to be architected and implemented such that
- Their internal teams could maintain, plug in and try alternative algorithms
- the algorithms could be extended and re-trained based on new data points
- new innovative indicators for default rates (developed by their financial experts on an ongoing basis) could be incorporated into the system
- Highly sophisticated frameworks such as this are challenging to build since leaking information from unseen data sets (which artificially boosts performance during testing, yielding good performance metrics that do not hold up in reality) creeps in subtly and is hard to detect.
Furthermore, a key objective for Greenlake was to improve the performance beyond a level that Finiata's data science team had been able to achieve in-house.
Greenlake has a strong expertise in being able to iterate rapidly in order to test multiple algorithms and approaches. In the span of a month, we designed a system that enabled comparison of multiple state of the art classification algorithms. We also went a step further and endowed the framework with the ability to search among a class of models to select the best model of that class.
This framework was used to search for and identify a classification system for default rate forecasting for Finiata. The performance of this system was tested on clients whose data had not been seen previously.
The solution achieved by Greenlake resulted in an increase of approximately 5% in the desired area under the curve (AUC) metric. The AUC is a highly demanding metric used in classification problems where one class (SMEs whose customers default on invoices) occurs far less often than another class (SMEs whose clients do not default). In such problems typical metrics such as accuracy are, unfortunately, ineffective in producing the desired results.
The resulting algorithm was put into production, allowing the customer to speed up their approval process significantly. The framework also achieved the desired objective of empowering their data science team to continue further R&D, algorithm comparisons and optimization. The system could be seamlessly retrained as more data became available and new features were formulated by client side financial subject matter experts