Machine Learning Is a Valuable Tool, Not a Silver Bullet
January 27, 2017
Online fraud attempts are up over the past year and only expected to rise further. This is a problem that’s not going away anytime soon, leaving many fraud managers and risk professionals searching for better solutions to avoid escalating costs.
The goal is simple: minimize fraud without impacting good customers, and do it as inexpensively as possible. If only the solution were just as simple.
Enter machine learning. This concept has become all the rage in fraud detection circles over the past few years. Talk to three different people and you’ll probably get three different definitions of machine learning, along with three different interpretations on what it can achieve. Herein lies much of the problem.
Some vendors have positioned machine learning as the answer to all your fraud problems. They would have you believe it negates or severely limits the need for any kind of rule-based strategy, manual review effort, or other fraud-detection techniques. While undeniably attractive, that line of thinking doesn't represent today's fraud environment.
My intent here is not to disparage machine learning. It is a powerful tool in the fight against fraud, and when used properly can both increase the accuracy of fraud detection while also lowering many of the operational costs associated with manual reviews. We’ve seen tremendous results with many of our clients when we introduced machine learning models, helping them to both catch more fraud and reduce customer insults, all while decreasing human review rates.
Much of this success, however, was attributed to a judicious approach that layered the machine learning model on top of many of the client’s existing fraud detection measures.
If you bought a fancy new alarm system for your home, would you stop locking the front door? Would you disable the outdoor motion lighting? Would you stop placing your valuables in a safe? The same mentality should be applied to utilizing machine learning as a part of your fraud strategy.
When considering adoption of a machine learning model, keep the following points in mind:
1. Any model is only as good as the data flowing into it. We’ve all heard the phrase, “Garbage in, garbage out.” That absolutely holds true here. If your data is not clean and accurately tagged, the model will suffer. Additionally, you need to ensure the right variables are being fed into the model in the first place to achieve the best results. The most effective models will utilize standard, traditional data as well as custom elements unique to the individual client, and then enrich those variables with community data from other merchants within a similar industry.
2. All models have limitations. Models predict the future based on the past. If the future looks very different than the past, the model may not perform accurately. Fraud attacks can come on rapidly, and patterns can change quickly. Legitimate business trends can also change, with seasonality, new product offerings, flash sales, etc. Models can be “overtrained” to look for certain common scenarios and can be slow to recognize or adapt to new behavior. As such, model performance must be regularly and carefully monitored. Using a machine learning model—even one that claims to be “unsupervised” or self-learning—doesn’t mean you can plug it in and walk away.
3. A multi-layered strategy is still the best approach. Layers are the best strategy when dressing for cold weather, and the same holds true for fraud prevention. Relying too heavily on any single methodology or technology may work in the short term but will likely fail over time. Combining a machine learning model with multiple layers of detection including custom fraud and policy rules is the safest, most efficient way to fight fraud. Finally, don’t make the mistake of thinking that machine learning will eliminate the need for manual review. While a good model may refine the subset of transactions requiring a review, it simply can’t detect everything that a human—particularly one with knowledge of a specific industry or business model—can discern. Human fraud analysts also provide valuable intelligence that can be fed back into machine learning models to improve future performance.
In summary, machine learning is a fantastic tool that can be a highly effective part of your fraud strategy, but don’t believe all the hype. Expecting it to replace all traditional means of fraud prevention and human intervention will likely leave you disappointed. Go ahead and install that new alarm system, but plan to keep locking the doors, using the safe, and turning on the lights.
About the Author
Clint Lowry is the Director of Global Product Management and Research for Accertify. He has over 15 years of experience working in payments, fraud, and technical product management. When he’s not traveling or building product roadmaps, he spends time with his family dodging scorpions and coyotes in Phoenix, Arizona.