Using machine learning to build a robust and effective underwriting process in the face of the global financial crisis4 min read


Executive Summary

Over the past decade, artificial intelligence (AI) and machine learning (ML) methodologies have increasingly been adopted globally and are now used to complement decision making in a very wide set of use cases, from video suggestions to cancer detection. However, banks and non-bank lenders have been reluctant to jump on the AI train and are still relying on old scorecards methodologies to assess the creditworthiness of their customers. While these historical scorecards have proven their efficiency and simplicity of use in the past, they quickly fall short in times of high, global economic distress. Lenders then face the caveats of their risk assessment methodologies when they need it the most and struggle with time-consuming, manual processes needed to re-calibrate and re-evaluate their scorecards.

As of March 2020, JP Morgan Chase has temporarily stopped accepting loan applications from small businesses not guaranteed by the U.S. Small Business Administration (SBA). It is the first of America’s big banks to do so and will likely not be the last considering the skyrocketing demand from borrowers, who are trying to maintain staff salaries during the COVID-19 pandemic. Uncertain economic conditions during the coronavirus pandemic have transformed the economy to the point that traditional credit risk models no longer apply. This causes a problem for lenders, as they must re-evaluate their metrics for assessing potential borrowers. 

The promise offered by ML is the ability to quickly create new, adaptive models for a credit underwriting process; one that is adjusted for the current COVID-19 economy. Adaptive models are ideal for rapidly evolving crisis situations, as they can not only be specifically trained on historical data from previous crises, but can also ingest and be frequently re-trained on up-to-date data, making them more resilient to global economic shocks.

ML models created for this purpose will require transparency of process in their operations to see adoption from banks and regulators. In other words, the new metrics and processes behind ML-assisted credit assessment must be explainable to both lenders and borrowers, complete with a clear “how” and “why”. This transparency will also answer other challenges faced by lenders intending   reforming their traditionally constructed models , helping them to choose metrics that balance fairness and responsibility to business goals while avoiding algorithmic and systematic bias. This will foster trust between borrowers and lenders by removing the perception that credit underwriting processes are a mysterious black box that discriminates against certain borrowers for misattributed reasons. 

Slowing global economic activity and soaring unemployment rates are correlated with high non-performing loan (NPL) rates. Considering that these two factors are direct effects of governments imposing quarantine measures and setting a norm for the post-COVID economy, adaptive ML models must take both into account. The current coronavirus crash differs from previous recessions in that its economic impacts are a direct physical consequence of the virus rather than an imbalance in corporate balance sheets. Consequently, its effects will persist until entire industries are restructured to adjust for novel phenomena like the shut-in economy. While the dichotomy between “keeping the economy running” and “saving lives” is a false choice, businesses will still need  to experiment with new revenue models in a global economy that periodically alternate between total shutdown and cautious socialization. Until then, many small to medium enterprises (SMEs) will continue to apply for loans to stay afloat as waves of the pandemic continue to ripple across the globe.

Similarly, lenders will need to experiment and be flexible in building, testing and validating new credit risk models at an accelerated pace in order to iteratively adapt to evolving market conditions and shocks. Training new ML models on data collected from previous crises will help, as will acknowledging evolving chaos as a new normal that will persist at least until healthcare systems across the globe come to grips with the coronavirus. Now, more than ever, businesses and consumers alike are in dire need of loans and working capital, and yet lenders cannot rely on their existing credit underwriting models to accurately assess credit risk. ML-based credit risk models allow for the agility, flexibility, and accuracy needed for lenders to make better credit decisions during unforeseen economic uncertainty. Furthermore, an automated ML workflow incorporates an enormous quantity of data points, facilitating much more holistic and accurate predictions given the inclusion of a wider set of variables in the model, all along with the ability to automatically re-train and monitor the model with minimal human intervention

Flowcast monitors hundreds of macro-economic indicators and data points which are then fed into our ML models, making them reactive to global financial distress and robust enough to withstand the current global financial crisis. By using the bottom-up ML approach, Flowcast models are able to improve their assessments over time as they make connections between thousands of variables used to assess creditworthiness. On top of our models, our production-ready ML-workflows allows for faster development, as cycles of model training and validation using historical data are automated and, hence, can be done on a much more frequent basis. Through adaptive models, Flowcast is able to create a clearer, more nuanced picture of the evolving post-COVID-19 loan market.

About the author

You may also like