ZestFinance 如何核保 (underwriting) 貸款

Leena Rao, ZestCash Debuts New Big Data Underwriting Models To Determine Consumer Credit Risk, TechCrunch, Apr 26, 2012  

Traditionally, lenders have used single regression models that analyze approximately fifteen points of data to make credit decisions. When ZestCash launched two years ago, it began applying big data analysis to underwriting for the first time. Now, instead of employing just one underwriting model, the ZestCash decisioning infrastructure can run dozens of models in parallel, returning loan decision results within seconds.

Basically, ZestCash used to run one big data underwriting model that looked at up to a thousand variables. This new infrastructure lets it run multiple big data models in parallel, so it can look at more data points to more accurately determine whether a person is a good candidate for credit. 

ZestCash now runs about 10 unique underwriting models simultaneously that consume thousands of raw data elements including third-party data and data collected from borrowers. The models then transform this data into tens of thousands of useful meta variables to assess key customer behaviors such as fraud, short-term and long-term credit risk, or the amount of money a borrower will likely repay. The models are then “ensembled” to arrive at a final underwriting decision that more accurately predicts credit risk. Some of this data includes cell phone contract credit history and rental information. 

Shawn Budde, ZestCash Co-founder and Chief Operating Officer adds that the new modeling has helped increase customers’ net payback by 20 percent.

John Lippert, Lender Charging 390% Uses Data to Screen Out Deadbeats, Bloomberg. October 4, 2014

For five years, Merrill has been harnessing oceans of online data to screen applicants for the small, short-term loans provided by his Los Angeles–based firm. Improvements in default rates have come in fractions of a percentage point. Now, on this 90-degree-Fahrenheit (32-degree-Celsius) July day, his researchers are claiming they can improve the accuracy of their default predictions for one category of borrower by 15 percentage points, Bloomberg Markets magazine will report in its November issue. ... 

Three of the most-digitized credit scorers for low-income borrowers are ZestFinance, San Francisco–based LendUp and Fort Worth, Texas–based Think Finance Inc. Advances in computer science allow these firms to collect thousands of facts on each loan applicant in a matter of minutes. That compares with the few dozen pieces of basic data -- mostly a borrower’s debt burden and payment history -- that San Jose, California–based Fair Isaac Corp. requires to compile the FICO score that is the basis of 90 percent of U.S. consumer loans. ...

The firm’s machines quickly organize individual facts about a loan applicant, including data that FICO doesn’t use such as annual income, into so-called metavariables. Some metavariables can be expressed only as mathematical equations. Others rank applicants in categories, including veracity, stability and prudence. 

An applicant whose stated income exceeds that of peers flunks the veracity test. A person who moves his residence too often is considered unstable. Someone who doesn’t read the terms and conditions attached to her loan is imprudent. 

One peculiar finding: People who fill out the ZestFinance loan application in capital letters are riskier borrowers than those who write in upper- and lowercase. Merrill says he doesn’t know why. ... 

After rejecting two-thirds of applicants, ZestFinance approves loans that average $600 for those that make the cut. The borrower pays $400 in interest for a six-month $600 loan. That computes to an annual percentage rate, or APR, of 390 percent -- at least four times more than the subprime credit cards offered by some banks. Borrowers have average annual incomes of $30,000 and many have poorly documented credit records or a history of defaults. ... 

Today, Merrill and his 60 ZestFinance employees use a smorgasbord of data sources to evaluate borrowers, starting with the three-page application itself. He tracks how much time applicants spend on the form and whether they read terms and conditions. More reflection, he says, indicates a greater commitment to repay. 

Merrill says he doesn’t scan social media postings. He does buy data from third-party researchers, including Atlanta-based L2C Inc., which tracks rent payments. One red flag: failure to pay mobile- or smartphone bills. Someone who is late paying his phone bill will be an unreliable debtor, he says. 

Once he’s organized his initial data sets into metavariables, he activates an ensemble of 10 different algorithms. 

An algorithm called Naive Bayes -- named for 18th-century English statistician Thomas Bayes -- checks whether individual traits, such as how long applicants have had their current bank account, help predict defaults. 

Another, called random forests, invented in 2001 by Leo Breiman at the University of California, Berkeley, and Adele Cutler at Utah State University, places borrowers in groups with no preset characteristics and looks for patterns to emerge. 

A third, called the hidden Markov model, named for 19th-century Russian math wizard Andrey Markov, analyzes whether observable events, such as lapsed mobile-phone payments, signal an unseen condition such as illness. 

The findings of the algorithms are merged into a score from zero to 100. Merrill won’t say how high an applicant must score to get approved. He says that in some cases where the algorithms predict a default, ZestFinance makes the loan anyway because the applicant’s income suggests he or she will be able to make up missed payments.