Dunning Solutions: Machine Learning Decline Salvage vs. Rules-Based Decline Salvage
How Does Machine Learning Compare with Rules-Based Expert Systems when Applied to Recovering Declined Credit Card Transactions?
You’re probably asking this question because you’ve discovered several companies have a decline salvage service, and you’ve read descriptions of AI, rules, machine learning, and you don’t know how to evaluate the difference between these approaches. This article will clarify what those terms mean and address whether there is a superior approach.
In keeping with our philosophy of not being coy, we’ll jump directly to our position. We, at FlexPay, believe there is an unequivocal winner: automated decline salvage driven by Artificial Intelligence, which is informed by Machine Learning.
What is Automated Decline Salvage?
Let’s be clear about what we’re discussing. Automated decline salvage is the process of retrying declined credit card transactions, changing something about the way the transaction is processed, with the intention of presenting the transaction in a way that the issuer’s algorithm expects, resulting in an approval.
The challenge is in deciding what to change. The most obvious option is to try on another day at another time, but there are dozens of other techniques a system can use.
What is AI and What Does it Mean for Decline Salvage?
This is where Artificial Intelligence comes in. AI, in general, is a software application that makes decisions over a range of options. This doesn’t mean that the software is self-aware, only that the application is attempting to make presumably optimal decisions. For decline salvage, AI will decide on the timing mentioned above as well as on how and whether to apply the myriad other adjustments available for getting a transaction approved.
Most companies develop their AI by building up a set of rules to govern what to change under what conditions. Humans, based on their knowledge of how the world works and specifically how banks consider transactions, write these rules in the form of “if this happens, then do that.” The humans that built these rules into our competitors’ declines salvage systems are experts, and rules-based AI’s of this type are called Expert Systems.
What are Machine Learning Models?
Machine Learning Models (MLM) take a different approach. In this case, humans build software models that consume historical behaviour. With decline salvage MLM, this means feeding the application transaction histories at a minimum. At FlexPay, we have access to hundreds of millions in historical transactions that have run through our engine as well as billions of credit card transactions provided to us by technology partners. In addition, we tap an array of other subscription and public domain sources to append data to our existing transaction data. It is truly big data!
With both approaches, smart, well-informed humans work with software to create systems that respond to stimulus – details of a transaction, with a response – retrying the transaction in a new way. One of the challenges is that there are a massive number of stimuli for every declined transaction.
Decline Salvage is Complex
Consider for a moment what we know about a declined transaction. We know a bunch of demographic information about the cardholder. For example, because the transaction comes with address information, we can append census data. Census data can tell us information about typical ages, incomes, employment status, etc. We know the value of the transaction. We know how many months the subscription has been in place. We know when the transaction was initially attempted, what merchant bank, what issuing bank, and how the merchant is classified by the merchant bank (it’s by the MCC). By virtue of access to so many credit card transactions, we know how the issuing bank has historically dealt with transactions of this type. We know the location of the merchant, that of the cardholder’s address, and from that, we know something about the general propensity of the bank to approve transactions from that combination.
Here’s a little tip: a transaction is more likely to be approved if both the merchant and the consumer are in the same state.
The Challenges with Rules-Based Expert Systems
Expert Systems make a valiant attempt to weed out the most important of those factors and trigger the appropriate response. But, even by reducing the number of data points considered, any system attempting to respond appropriately to all that information would need thousands of rules. As the availability of relevant data increases, the rules-based AI must become as complicated as the system it is managing.
This phenomenon is precisely why issuing banks do such a poor job at identifying the fraudulent transaction and have a false positive ratio of 13 to 1 on declining fraudulent transactions. Ironically, one of the main reasons why we need decline salvage is the same reason that the most popular decline salvage approach, Expert Systems, is fatally flawed.
There is a place for Expert Systems. When the number of inputs and outcomes are small, knowable, and change infrequently, then this approach can be efficient.
How do Machine Learning Models Compare?
The advantages are well known to anyone who tackles this kind of problem. Machine Learning can sift through billions of records to find hidden, often counter-intuitive, correlations. Unlike rules-based systems, ML can be continuously updated as new information becomes available. In a world where $400 billion in transactions are declined as false positives, any insight leading to the smallest improvement in performance can have a massive impact. Under conditions of increasingly complicated and interrelated information, ML models improve while Expert Systems drown in complexity.
Why doesn’t Everyone Build MLM Driven AI?
We’re confident that all enterprises engaging in decline salvage would use ML if they could. But, to do so effectively, you would need access to an enormous wealth of historical transaction information. This can be difficult to acquire, and when available, say from the enterprise’s own operations, the data is only from the perspective of the merchant. To truly succeed, you would want data from the issuers. With that, your model will have a more profound understanding of the operation of the underlying rules-based fraud detection systems.
You need the mathematical, business, and computer science knowledge in-house that can mine the data, seed, build, and provide continuous training to the models. This requires people with specialized skills and Ph.D. level credentials. To find and retain these individuals, you need an exciting project and a culture that is conducive to their professional growth and the satisfaction of their academic interests.
Your business must be willing to commit. Anything you read about machine learning will tell you that it takes time. For a company to grow and generate revenue, it takes a substantial leap of faith to make the years-long investment into development and research teams required to produce effective models.
Your choice of AI represents the kind of tool you use for solving the salvage problem. Choosing rules-based AI to solve the decline salvage problem is like using a screwdriver to hammer in a nail. If the nail is small and there aren’t many of them, you can get away with it. Otherwise, you should use the machine learning hammer.
Contact FlexPay to see how we can help you today!