The credit card industry is vast, with many companies servicing hundreds of millions of credit card holders. With so many customers, how do credit card companies manage everything? There are so many decisions to optimize, from determining whom to issue a credit card, how much credit to offer, what benefits to offer and when, how to protect from fraud, and much more.
Federal Reserve Bank of Atlanta data released in August 2019, for example, found 75.5% of consumers had at least one credit card, defined as a card that allows the cardholder to make a purchase by borrowing funds that will be paid back to the credit card company later.[1]
Machine Learning models are increasingly being used to both grow credit businesses and protect against increasingly sophisticated fraud attempts. This trend is bound to grow, as actors, both good and bad, continue to invest aggressively in this technology.
ElectrifAi machine learning models are pre-built from deep domain expertise across multiple verticals, particularly in credit card life cycles management and customer engagement. We have built industry leading technology for multiple top financial institutes since 2004, so we understand the complexities and risk for both issuer and borrower (consumers and businesses).
Let’s delve into a few of the top machine learning models that are vital to success across the entire credit life cycle. After reading through this list, please let us know how we can begin—or enhance—your machine learning journey!
Bust out fraud is when someone acts as a normal consumer would and then suddenly races up the charges on their card with no intention of paying. As the borrower uses the card in the mall, the gas station or online shopping, the Bust Out Fraud model identifies patterns of behavior and spikes in usage that precede a borrower walking away from the account without paying.
To prevent loss, some issuers may trigger a false alarm, preventing someone from using their credit until the issue can be resolved. Calling the customer and then preventing card usage is an unnecessary expense and puts a valid account at risk. False alarms also take time away from preventing real loss.
There are artificial intelligent approaches that do not use machine learning but apply a series of rules to detect fraud. Where machine learning helps is locating complex patterns that have appeared in real use and are too complex for rules-based systems.
Rules-based systems require constant tweaking and adjustment to try and tune it to the real world. Machine learning, on the other hand, is the process of using real life examples to do that tuning for you.
With much faster time-to-value and more accuracy, machine learning based bust out fraud detection minimizes false alarms, without the risk of increasing misses.
To find out more about how our machine learning models work, we created a video using our Bust Out Fraud model as an example. Check it out!
The traditional way issuers have looked at credit line management is to assign a score to certain behaviors of borrowers. For example, a consumer may have seven credit cards, and a rule-based system says the ideal number four. The consumer receives a lower score, independent of the actual underlying risk. Points are added or subtracted from the credit score based on certain behaviors issuers expect statistically align with people who have good credit.
Many factors are assigned points, and anything that raises suspicion may lose points. For example, if one suddenly takes out three credit cards, then despite a high current score, and ideal record of paying bills on time, the credit score may be reduced presuming risk. Machine learning has a much better ability to sort out cases where this has been true from those where it is not. A rules-based system may know that a sudden increase in available credit yields a negative outcome, perhaps 40% of the time. Machine learning looks at borrowers who resemble those who continue to pay from those who have not, and can recover many of the remaining 60% of borrowers who are not a risk.
A rules-based system that adds or subtracts points is going to miss out on important differences between these groups. If someone is a career professional and has consistently had jobs but takes out a few credit cards, they shouldn’t be penalized as much as someone without much employment or credit history. One can always add more rules, but creating those rules gets very complicated. Machine learning effectively derives those rules from real data, including complex interactions between thousands of factors, far beyond the capability of any analyst to derive from the data through inspection.
With machine learning, you can segment the population far more effectively. The Credit Line Management model looks for people who have similar behaviors and history, and determines credit worthiness based on those similar people. Those who are like others with good credit scores will receive a higher credit limit. Because people are multi-faceted and many factors would have to be considered to find a true similarity, machine learning is the only practical way to achieve advanced results.
Cross-Sell and Upsell applies to many products that a credit agency or bank provides for customers. Customer segmentation effectively determines the type of offer that is most attractive to a customer, resulting in a much higher likelihood to purchase. Maybe it’s a different type of credit card program, such as a reward program for travel as opposed to points for shopping online. Understanding the propensity of someone to buy each product is critical to making the product a success.
If a person has a lower risk disposition, trying to sell them a higher risk product will not work. But selling a useful product in their comfort zone is much more likely to be successful. You must understand not only what you think is in their best financial interest but what they see as their own best financial interest.
That’s where machine learning shines. What a human may see as a likely scenario due to their own background and what they would think someone is interested in, a machine will only look at the facts—actual data with driven decisions. Machine learning can process hundreds of thousands of factors that go into finding similarities between customers to understand what really makes them tick and what product would truly get them to buy.
When an account goes into collection, there are a lot of rules around what a lender is allowed to do, such as how often they’re allowed to reach out and contact the delinquent customer. Lenders should be allowed to collect the money that hasn’t been paid, but sometimes there is a fine line between collecting and harassing.
One of the things that can make a company much more efficient in the collections process is to understand who they should be contacting on that account, what method they should be using, and when they should contact. If you try to reach a customer to let them know they are behind on a bill but it’s 10am on a Tuesday and you know they are at work, there is little chance you’ll get them to pay. But if you were to contact that person at 7pm, you’re much more likely to get that person to go online and pay.
Getting in contact with the right person at the right time to pay their bills relies on you understanding that person. The best time of day and the best method to do so can be accomplished by using the Party Contact Likelihood Prediction model. The model takes in different factors and tries to understand the customer’s behaviors and predict when they are likely to be available, what method they’re likely to respond to, and even who on the account to contact.
In many cases, you can have sleeping (inactive) customers. They were using their credit cards but for whatever reason they’ve stopped using it. Maybe they got another credit card and they’re using that one. But it could even be as simple as they put your card at the back of their wallet and now, they see this other card and they use that one more frequently. Or maybe the other card offers them better rewards.
The question is, how do you wake that sleeping customer and win them back? This requires understanding the customer. What’s driving them from using your card? What would drive them toward using your card again? You must understand how to reengage the customer and get what has been a profitable customer back.
Machine learning lets you easily look at the past offers you’ve given, look at the people who responded well to those offers, and do a detailed granular customer segmentation. Looking at the customers who are like that sleeping customer, see what worked and didn’t work for the similar customers. And then you can target that customer with inexpensive advertisements to bring them back into the habit of using your card again.
At ElectrifAi, we have a lot of in-depth knowledge in the credit card industry. We have built many models, looked at tons and tons of data, and worked with many top credit card companies and financial institutions. Because we have done this multiple times and have really dug into the problems this industry faces from many angles, we understand how credit card holders think and their disposition. It’s that type of industry and real-world experience and that set us above other machine learning companies.
The models we have listed in this blog are just a slight dip in the vast library of models we have built over many years. And this list is ever-growing. As we now begin to see the light at the end of the 2020 tunnel and we begin to envision what 2021 may bring, we believe credit management will be very different.
2020 has been very disruptive to many people’s lives financially for various reasons. Because of that, a lot of the old rule-based systems and ways of doing things just won’t apply. Having machine learning capabilities can quickly pick up on all the changes in the market and really understand the new normal. This is especially critical for recovery in 2021.
Let us know how we can help you recover from 2020 and look forward to all the incredible things 2021 can bring! Contact us today for a custom demo.
[1] https://www.creditcards.com/cr...
