Latest Underwriting Model Marries Machine Learning and Human Artistry
Since 2009, the ZestFinance team has been working hard to give the underbanked access to lower cost credit to help them save billions of dollars in fees and costs. To achieve this, we’ve created an entirely different approach to underwriting by using machine learning and large-scale big data analysis. Over the past three years, we’ve more than doubled our team and hired some of the world’s best data scientists and mathematicians to sharpen our ability to analyze credit risk.
Today, we’re announcing Hilbert, our latest underwriting model. By combining advanced math and technology with human predictive modeling, Hilbert provides a more accurate depiction of a person’s ability to pay back loans. In fact, the new underwriting model offers a 54 percent lower default rate than the best-in-class industry score.
With Hilbert, approximately a quarter of the data that runs through ZestFinance’s underwriting models is based on new variables constructed by human predictive modelers. ZestFinance’s team of predictive modelers—with backgrounds in physics, computer science and mathematics—analyzes thousands of variables created by machine learning algorithms, modifies them based on patterns, trends, and unique insights, and feeds the modified variables into multiple big data models.
For example, in the case of bankruptcy (a strong signal used in underwriting decisions), a machine learning model can quickly calculate the number of years since a person filed bankruptcy, but wouldn’t know that people are only allowed to declare bankruptcy once every seven years. As such, is eight or nine better than seven? And, what does zero mean - that a person has never filed bankruptcy, or just not this year? Machines don’t know—people do.
When people think about big data, they typically focus only on number of transactions or outcomes. The real value in big data, however, is developing new signals—information about people—not the number of people you have information about. Arriving at those key insights requires human intervention and inferencing. Machines and data alone are not enough.
Since the company’s inception, we’ve increased net repayment by 90 percent over off-theshelf industry scores and more than doubled the number of underbanked Americans we can serve. Not bad for three years out of the gate. By integrating human modeling artistry through Hilbert, we’re another step closer to making more credit and loan options available to the people who need it most.
Founder & CEO, ZestFinance
November 19, 2012