Machine Learning: Perfect Fit for Predicting Credit Risk
IN TODAY’S AS-A-SERVICE CLOUD MODELS, NO EXPERTISE REQUIRED TO DEPLOY OR USE
To test the efficacy of machine learning as a vital tool in assessing credit risk, a data architect fed historical census data into a machinelearning model. He was trying to gauge the correlation between credit worthiness and income levels. The model also included data on zip codes, public v. private sector employment, management v. rank and file, and so on.
But the machine learning model determined that the most influential factor was simpler: Marital status.
The results were as unpredictable as they were surprising. The presumption was that income and some of those other factors would weigh heavily in determining credit risk. But the machine learning model determined that the most influential factor was simpler: Marital status. As this research revealed, machine learning can unveil patterns and relationships in data that traditional means of statistical analysis often overlook, and without any human intervention or bias.
Similarly McKinsey scanned 10,000 resumes received in an earlier round of hiring at the firm, and then fed that and other data into a machine-learning program to determine which of the recruits were likely to accept the company’s employment offers – based solely on data. The predictions of the machine-learning program ‘strongly correlated’ with the actual results of the recruits who accepted offers from McKinsey.
Machine learning is hot
“The use of machine learning techniques can help banks improve the predictability of credit early-warning systems by up to 25%.”
Given anecdotes like these and many others, it is no surprise that in its 2017 Top Ten Technology Trends report, Gartner listed AI & Advanced Machine Learning as the number one trend. Gartner cited financial services as a vertical that can leverage machine learning to model real-time data to improve credit risk decisions and prevent fraud. McKinsey went even a step further in its praise of machine learning for determining credit risk, saying, “The use of machine learning techniques can help banks improve the predictability of credit earlywarning systems by up to 25%.”
If you are new to machine learning or feel daunted by its seeming complexity, think of it this way. Computers, which are inherently dumb, get their smarts from programmers that feed them instructions, which the computer follows explicitly. With machine learning, a computer is equipped with highly complex algorithms - which are similar to programs - along with mountains of data. Only the computer then acts and learns to ‘think’ on its own as it discerns patterns and anomalies in that data. In fact, the more data it is fed over time, the more the machine ‘learns’. It is the increasing need for automation in credit decisioning helping to drive the creation of sophisticated machine language algorithms.
These machines can even have fun.
Last year a machine learning program called AlphaGo from Google DeepMind beat the world’s top player of Go, a highly complex and ancient Chinese board game. AlphaGo was not programmed to play Go, but rather learned to play on its own as the games progressed.
Several vertical markets, including financial services, are already heavily using machine learning techniques. In financial services, machine learning has taken firm root in fraud detection, showing capabilities to not only detect fraud that has happened but, increasingly, to actually predict when fraud is likely to occur so actions can be taken to prevent it. Machine learning does so by recognizing patterns of transactions or behaviors that indicate fraud – patterns that might well have revealed themselves to human detection methods. Machine learning is particularly adept and finding these previously hidden patterns.
Producing a fount of business benefits
One study last year pegged the annual savings reaped by credit card issuers deploying machine learning fraud detection systems at $12 billion annually.
Machine learning can foster better credit customer relations. All too often credit cards are ‘flagged’ for what static models perceive as abnormal behavior, such as a purchase in a foreign country. Machine learning instead can perceive this as ‘normal’ behavior if the same card was used to book an airline flight, a hotel room in the same foreign land, and perhaps pay for a meal at an airport.
Also, marketers have given machine learning a warm embrace. They use it to crunch through enormous volumes of customer data, augmented by third party data on consumer patterns and behaviors, to develop far more targeted and therefore cost-effective online marketing campaigns as well as in-store promotions. Rent The Runway, a start up seeking to disrupt the venerable business of selling high-end designer dresses, leverages an as-a-service machine learning solution to accurately target customers seeking to rent rather than buy this costly apparel.
The bottom line is that machine learning programs are capable of ingesting and analyzing tidal volumes of data, and any combination of variables you want to stick into the program (income, zip code, banking history, etc.). Given that data volumes by most estimates are doubling every two years in most businesses, machine learning has arrived at a perfect time given its ability to handle these massive data volumes. Moreover, machine learning does this work tirelessly on commodity –meaning inexpensive – hardware.
There
are great benefits to the as-a-service model.
All you need to get started is your data
More importantly, the ability for businesses such as lenders to access and leverage the benefits of machine learning requires nothing more than what they are already awash in – namely data. A great number of third party providers has emerged offering machine learning as-a-service, often over trusted and highly secure cloud providers such Amazon Web Services.
There are great benefits to the as-a-service model. The key elements of, say, a credit risk determination model built for one lender can be easily ported to another lender, and then be modified and tailored to fit individual lender needs and requirements. Further, a lender need not invest in additional hardware, software, or hard-to-find, sophisticated IT staff to create the models and reap the benefits of machine learning. All you really need is data. And finally business professionals, such as the credit risk manager at a typical lender, can easily use the machine learning models offered by the better third party providers, lessening the dependence on IT.
For a number of reasons, machine learning is poised to become a significant technology in credit risk determination. For one thing, the lending market itself is undergoing significant change and disruption, with well-funded, non–traditional lenders – so-called fintechs - muscling into the market backed by the latest digital solutions.
No longer is it acceptable to take days or even weeks to approve or deny a loan application. Increasingly it must be done nearly in real-time, which implies high level of automation at the lender.
“Machine learning is perfect for building models to predict risk, identify correlations, and categorize individuals and activities.”
– NIK ROUDA, SENIOR ANALYST AND MACHINE LEARNING SPECIALIST AT ESG
“Machine learning is perfect for building models to predict risk, identify correlations, and categorize individuals and activities,” says Nik Rouda, senior analyst and machine learning specialist at ESG. “Its models can be trained on the vast amounts of transactional data and customer profiles.”
Also lenders are being nudged and cajoled by government forces to lend responsibly to smaller borrowers, including small businesses and individuals. The data needed to properly conduct credit worthiness for these constituents is far different than loaning to bigger borrowers. More automation is needed to lessen the dependence on expensive manual processes, which are not cost-effective when dealing with
More automation is needed to lessen the dependence on expensive manual processes.
small loans. Simply relying heavily on statistical tools of traditional credit reporting agencies to serve the so-called ‘under-banked’ will not work going forward. By contrast, machine learning can actually categorize the growing list of non-traditional borrowers, such as the rising tide of millenials, that are often not catered to and who are highly accustomed to helping themselves with various self service digital tools.
Machine learning not only automates and therefore expedites the credit risk determination process, but also improves the accuracy of the process by factoring in a far wider range of relevant data than static statistical models could ever process. But perhaps most importantly, machine learning can leverage new data as it arrives to change the parameters of credit worthiness. For example, gentrification of downtrodden neighborhoods is happening very rapidly in some major US cities. Thus the weighting of zip codes and street addresses should change accordingly. In a static statistical model of credit worthiness, they don’t. With machine learning, the algorithm recognizes the broader demographic changes underway, factoring that into the overall credit calculus.
Machine learning not only automates and therefore expedites the credit risk determination process, but also improves the accuracy of the process by factoring in a far wider range of relevant data than static statistical models could ever process.
“Most services have set up data migration or loading functions to work with common sources, so it shouldn’t be too much work to import data to cloud services.”
– NIK ROUDA, SENIOR ANALYST AND MACHINE LEARNING SPECIALIST AT ESG
A rich value proposition made easy and simple
As in any business, time is money. With machine learning, non-credit worthy applicants can be more quickly eliminated from consideration with no human intervention. Weeding out less credit-worthy loan applicants earlier means no costly and time-consuming queries to third party credit rating firms for these applicants. The technology also gives lenders greater assurance of the credit worthiness of those to whom credit is extended, owing to the far more extensive array of datasets analyzed to reach a credit approval status.
The cloud-based machine learning as-a-service model for modernizing and streamlining credit risk determination is easy. Literally nothing more is needed other than a continuous stream of data, fed into the system by way of very user-friendly interfaces that any business analyst can use. Plus the security of customer data is virtually guaranteed because all personal information sent to the machine learning model is either encrypted or left out. The machine doesn’t care who the data is from, rather only what information is actually in the data.
According to ESG’s Rouda, “Most services have set up data migration or loading functions to work with common sources, so it shouldn’t be too much work to import data to cloud services. That said, good data governance is important in any setting, as bad or badly managed data will train the model to be inaccurate.”
This is simplicity that rivals the instant cake mixes marketing under the Betty Crocker brand in the 1950s by General Mills. All you had to do was add water and an egg and reap the output from the oven an hour later. So it is with machine learning for credit risk determination when served up in an as-a-service model. Just add data, as much as you have, and the rest is pure business benefit.