According to the Federal Reserve , nearly 40% of adults in the U.S. sought credit in 2020, only slightly fewer than those who applied in the previous pre-pandemic year; among those who applied more than 1 in 10 were denied credit or were approved for less than they had sought. The reasons behind these denials are many, however, the same report found that those who were denied or who were approved for limited credit often fell in lower-income and education brackets as well as along racial lines.
While many lenders have used artificial intelligence for years to detect fraud and provide real-time insight into commercial activities, banks and non-traditional lenders are putting AI to work to identify new ways to predict which applicants are creditworthy.
Traditionally, lenders consult a prospective borrower’s credit score, which weighs criteria including the borrower’s payment history, how much credit they currently have (also known as their “credit utilization”), the average age of existing credit accounts, the types of current credit owed, and the number of new credit inquiries. In other words, it takes credit to get credit.
Artificial intelligence and deep learning algorithms enable lenders to leverage troves of data not captured in a traditional credit score. For example, a recent college graduate might not yet have much of a track record in repaying their student loans (a credit type that is reflected in the traditional credit score), but might have a years-long history of on-time mobile phone payments and a steady, well-paying job. Other non-traditional criteria include how long an applicant has lived in the same place, their use of payment apps and on-time utility payments, among others.
In addition to using AI to expand credit eligibility to groups previously assessed as too risky (without demonstrably widening the lender’s risk profile), financial institutions also have deployed deep learning models to “backtest” credit decisions. This process looks at credit denials to assess whether someone was declined erroneously, or whether there is bias in the system that needs to be corrected. Many organizations conduct this type of refresh review after a certain percentage of declines are registered.
Much of this non-traditional data lies in unstructured or inconsistent formats – customer databases, rental agreements and the like. In fact, as much as 90% of data is unstructured, and few organizations were capable of deriving value from it. Until recently, an organization needed to hire a staff of data scientists knowledgeable in AI to develop the deep learning models required to parse this data. With the availability of H2O Hydrogen Torch, however, nearly anyone can create deep learning models that can process images, video and text processing without coding. H2O Hydrogen Torch enables companies to unlock new opportunities to transform their business and their industry using public or private data sources.
To parse text-based or natural language data, H2O Hydrogen Torch can be trained to classify and regress text and tokens, perform span prediction, sequential analysis and metric learning. Language use cases also include sentiment prediction and sequential analysis to summarize large text blocks. M odels can be packaged for deployment in other computing environments or directly into H2O MLOps for production usage.
Because AI can analyze data more quickly than humans and can detect patterns among criteria that might not be obvious to a bank underwriter, one European company that provides credit options for small and medium-sized businesses looking to finance capital purchases has leveraged H2O.ai’s Driverless AI solution to provide 20% more applicants with credit and to offer them the financial products that best meet their needs. The solution helped the company improve its risk predictions, increased fair lending and bolstered the overall customer experience.
To learn more about how financial services companies are leveraging H2O.ai to modernize their businesses, please visit our website .