By Chris Logel, Regional Loss-Prevention Manager, Shoe Carnival
Like the assembly line and the machine age before us, automation is the future of online sales and merchants that have not already automated their fraud decisioning should at least be exploring how. Every order—good and bad—leaves behind an invaluable trove of data that can be analyzed and compared to other similar orders to help automate accept or reject decisions on future orders. And, any merchant can leverage these pieces of data to solve the puzzle of rooting out fraudulent orders without holding up legitimate ones.
The first step in this process is to sub-categorize orders into different profiles such as high-dollar, moderate, low-dollar, mobile, overnight, etc. Different rule sets and decisions can mean different things to different order profiles. Analyze orders from different profiles; take a sampling of good orders, suspect orders and chargeback orders. Once you get a sizable amount of orders you will see similarities in all different profile rules. These rules and similarities can be used as identifiers to make decisions on your orders.
Remember, all of these footprints are meant to be viewed as positive signs as well as negative. This information is invaluable to determine accepted orders as well as rejected orders.
Take a common sense approach to interpreting these rules and orders, always keeping in mind how consumers act in the real world. For example, would the average person spend more money in shipping than the item costs? Look at these samples as if you were the customer. Learn your order flow so that you can make necessary adjustments during known peak times so that more orders flow through with less friction. You might even want to establish special “holiday” rules you can turn on and off as needed so that your analysis is consistent with the changes made on your site during those times.
There are several true identifiers I have learned to pay very close attention to: order velocity, card-use velocity, large geolocation differences, time-zone inconsistencies and device ID inconsistencies. Fraudsters can be creatures of habit and in a rush at times, so look for patterns in velocity, addresses, names and emails (numbers mixed in with non-sensical letters or words, same obscure email domains). You must take the time and put forth the effort to build a scoring system you trust. Use all the tools and “footprints” available to create your accept or reject zones.
Putting Puzzle Pieces in Place
Of course, whether you automate fraud decisioning or not, manual reviews will remain part of the process. An intelligent system, however, can greatly reduce the number of orders that need to be reviewed manually—and the cost associated with them
Rule and scorecard maintenance is invaluable. They must be reviewed monthly so the most current evaluations and standards are being used. The time that orders spend unnecessarily in review could mean cart abandonment and a loss of the sale to a faster more automated competitor. Seventy-five to 90 percent of all manually reviewed orders eventually get accepted. So, all those good orders are being held up by a process that not only gives customers additional time to change their minds or go to a competitor that accepts their order faster and with less friction, but is unnecessary almost all the time.
According to industry data, 24-27 percent of all orders get manually reviewed. Even if you’re at a lower rate, however, there is still room to improve. Shoe Carnival was at a 12 percent review rate when I started developing my “puzzle theory.” I saw that each identifier that differs (or doesn’t) from how a legitimate consumer would shop online is a piece to the puzzle. These anomalies, or pieces, you see (or don’t see) come together to form a picture—accept or reject. After implementing that philosophy, we reduced our review rate to 4 percent in less than a year and only increased our chargeback percentage by .02.
It is important to look at all of the puzzle pieces, not just one. For example, legitimate last minute orders do occur. We have a tendency to assume that overnight orders are automatically suspect. I’ve found this to be a fallacy. A week before the big day or big event could call for an overnight order. Invest the time upfront to establish the rules based on your identifiers. And, review them consistently for any updates you need to make based on your order flow.
One of the biggest sources of “puzzle pieces” is analyzing your chargeback orders. I suggest analyzing this data quarterly so that you get an adequate sample to analyze and still have time before the next peak season to adjust your rules or scoring based on your results. Analyze the rules that hit most often in the chargeback orders and compare them to how many times they appear in good orders. Look for the lopsided results between the two. Analyze the rules you have that appear in chargeback orders. Do these rules appear in good orders also? If they do, is the rule really telling you anything? If they do not, then I would put more importance on this rule when it is seen in future orders as a red flag. Rules that overwhelmingly appear more in chargebacks than in accepts need to be analyzed and possibly adjusted.
This also works the other way as well. Rules that overwhelmingly appear in accepted good orders versus chargeback orders need to be reviewed and possibly adjusted to reduce the friction in the order decision process. Always remember, in analyzing this information, creating your “puzzle”, and developing your customized scoring system, use a scalpel and not a hatchet.
The Finished Puzzle
As with any complicated puzzle, it will take some time initially to analyze your rules and results. Historical footprints will provide information that will help you find the pieces you need to better automate your order decision process—the rules that historically align with good orders and those that have proven to be warning signs in rejected orders and chargebacks. The time that is put into this is a great investment that, if maintained regularly, will provide fewer rejected legitimate orders, faster acceptance of good orders and fewer orders sitting in a review queue waiting on the inevitable acceptance or the customer response that they found the same product on a competitor’s site faster.