There are many ways to prevent chargebacks integrated into the best 3rd-party fraud solutions available to us today. These methods come to us in the form of a variety of features that can be used toward detection and prevention of the fraud that we seek to deter. But with so many options out there, it is hard to get a good understanding of all of the different defensive techniques and what they do.
If you are confused about how to effectively automate fraud prevention, you are in good company. Most fraud solutions have one or two key ways to prevent chargebacks that they promote and demonstrate. These ways have many buzzword terms to them. Machine learning, rule engines, behavioral analytics, graph network analysis, anonymous peer-to-peer networks and more… but what are they, what do they do and how can they help?
This article explains the most important ways to prevent chargebacks currently used by ecommerce fraud prevention specialists:
- Use Machine Learning
- Create and Update Chargeback Prevention Rule Engines
- Watch for Behavior that Indicates Chargebacks
- Graph Network Analysis
- Anonymous Peer-to-Peer Networks Dedicated to Chargeback Prevention
- Is There a One Stop Shop for Chargeback Prevention?
Use Machine Learning
One of the most most common ways to prevent chargebacks is machine learning. In fact, it is very difficult to find a chargeback prevention solution that does not have some form of machine learning within its capabilities.
I will try to explain machine learning in a very simple way. Machine learning is a complex set of computer algorithms that can take huge amounts of data and run them through many statistical analyses very quickly to determine an outcome based on patterns within the data. The models learn as more information is passed through them and additional patterns are recognised.
Machine learning is a very useful as a way to prevent chargebacks because it can process and run through large amounts of data so much faster than any human can, making calculation after calculation to put the pieces together in what is virtually no time at all. Everything is so fast that it basically spits out real-time results based on its observations.
Models can be completely self-guided or can be influenced to be more specific to desired schemes with the data it is fed. For example, you can have separate models trained to detect and action different fraud types as they use different indicating factors to determine their guilt. You could create a model that looks at payment fraud and also have a completely separate one to target account takeovers.
This feature can be set to react to different combinations of indicators but is typically used to produce a risk score. A client can then determine what scores trigger which actions and where along the customer journey they occur. Usually, the higher the score, the more risk there is.
Some vendors will even work things differently and offer a trust score, where they look to determine if an account is trustworthy as opposed to suspicious.
Machine learning has the ability to constantly evolve, very rapidly, just like fraudsters do. It can be useful in detecting and preventing whatever fraud type it is targeting if it is properly set up to do so.
The most common use for machine learning is payment fraud. The account, order, device, payment details and more are quickly processed through historical network and platform details to find patterns of risk and are then either actioned or scored. Score thresholds can be used to set accounts and orders to be passed freely, manually reviewed, forced through a verification process, passed to 3DS, etc…
Being real-time is a powerful element of machine learning because a lot of what we are able to do is reactive. Machine learning can find deviations, connections and patterns almost immediately.
Create and Update Chargeback Prevention Rule Engines
Rule engines allow the client to use a set of conditions that trigger the desired action. These are system-installed versions of old school processes for stopping chargebacks where analysts would look to see if accounts meet XYZ criteria to pass or fail transactions.
Chargeback prevention rules are often laid out like a Boolean formula with “If, And/Or, Then” statements. They are typically set to visibly outstanding values, velocities or specific identifying features with several aggregations throughout them. However, some rules can become quite complex, including the use of several in combination with one another to generate a single outcome.
Although stemming from an older technique, rules are tried and true, but they are static. Prevention rules need to be revisited and reanalysed to make sure that they are still effective and efficient. Fraudsters change their methods frequently and will strike just below set thresholds to circumvent the rules that have been put in place to stop them. As we have most clearly seen with this age of Covid, customers can also quickly change their patterns, so we have to monitor rules to make sure that we are stopping chargebacks while enabling exceptional customer experiences as well.
Rules are useful for every fraud type, especially when they are first created. They are reactive when put in place, but can become live if you can integrate them into your chargeback prevention systems. 3rd-party solutions will often include some form of rule engine for their clients. Some are more complex than others and some limit the data and conditions that are available for use. Ultimately though, in an ideal situation, if you can see it, you should be able to use it.
Rules are extremely important for newly exposed and experienced fraud and chargeback schemes. A rule may quickly be made and the results then fed through a machine learning model that will grow to include and build from the patterns revealed in the accounts triggering the rule. Ultimately, one would want both machine learning and a rule engine that can work together, with the rule results fed through the machine learning models to help train them.
Watch for Behavior that Indicates Chargebacks
Behavioral Analytics are a specific area of machine learning chargeback prevention. They look into how the system is interacted with to generate their outcome.
They will use such details as tapping and typing patterns, travel and input speed, gyroscopic details of the device, IP and device changes, language, location, copying and pasting and more. A plethora of details are used to determine if the accessing user is showing the classic signs of being a chargeback risk.
Behavioral analytics builds profiles to identify each customer and fraud scheme that can result in a chargeback. The profiles created can see how a site or app is navigated and information is entered into it to identify if an account has been accessed by the genuine account holder or not. They can look at all of the same factors and spot the patterns used by fraudsters on their own accounts, as well.
The strength of this feature is very apparent when it comes to bot detection and account takeovers. Velocity, device and navigation patterns are most visible within those kinds of attacks. However, the details available through behavioral analytics are powerful data that can be used well toward any fraud scheme they are set to defend against. There is a rise in hybrid fraud that is using both bots and human interactions to commit their crimes and chargebacks, and this form of machine learning is on the cutting edge of figuring that out.
Most providers will supply a rules engine with this kind of product, as well. As mentioned before, the combined ability to write rules and have machine learning that will complement each other increases the strength of chargeback and other preventative efforts.
Graph Network Analysis
This one is fun.
Have you ever watched those old detective shows where they would have a map with photos and a bunch of thumb tacks with threads of yarn connecting a bunch of them together in an intricately spun web? That is the exact thing that birthed this feature.
Graph network analysis takes identifying features of accounts, orders, payments, etc… and connects them to additional appearances upon the network. So, if you have one credit card that is used on twenty accounts, those connections are made and may be illustrated within their platform to allow for a visible path and pattern.
Clients can use what they can determine from across these connected features to help determine guilt or suspicion of a possible chargeback threat. Decisions may often be passed through entire networks as opposed to having to do so on an account-by-account basis. Rules may be created based on entire network insights. Connections could cluster fraud together depending on their shared data points.
Some providers connect on more elements than others and, although more is typically better, they must be able to provide a way to filter so that everything is not seen at once and can be decipherable. There is a lot of information that can be similar across a network, and not all of it necessary signals a chargeback threat. You need to be able to focus on those that matter to you or are most useful. Typically, these are different elements of account, payment and device details.
Outside of just being a fun way to detect and analyse fraud, graph network analysis is also good for machine learning fraud and chargeback prevention. As networks grow and connections are made to accounts that are deemed to be fraudulent, that data may be fed into the model to help train it to detect new patterns. Machine learning scores may be increased based on certain network elements as details connect to one another.
Although useful for stopping chargebacks, graph network analysis does not have its strength lean towards a specific fraud type. It can be found to be useful for any fraud scheme you may be looking into, as long as it is using the correct connecting features.
Anonymous Peer-to-Peer Networks Dedicated to Chargeback Prevention
One advantage that fraudsters have over us is their desire and ability to openly share amidst their nefarious circles. As fraud-fighters, we cannot be as open and free when communicating through our own communities. We keep many precise details tight to our chests for a variety of reasons.
Anonymous peer-to-peer networks allow companies to share information to help each other determine who is and is not a chargeback risk or known fraudulent actor without actually sharing with each other nor allowing anyone to know who shared what or what is happening on anyone else’s platform.
This is accomplished with the use of an intermediary. Data and details are shared to the provider who analyses them against their entire client base. The observations and machine learning results are communicated back to the initiating client with a recommendation or machine learning score based on the provider’s entire network. Some providers will be able to narrow its sights further, to only compare against companies that are within the same industry vertical. This is important, since chargeback risk for sure varies across industries.
The results may be interpreted and actioned as the company chooses to integrate it into their chargeback and risk strategy. Scores can be used to rate trust or risk, but no specific details from a company are ever shared with other clients. The results will not tell you any personal identifiable information and very limited details about the other company. You might get which industry they are in but that is about it.
From the details and reviews obtained from the other companies, a machine learning score is calculated for the client to action. Once you source a response, the details and results of your query can be included into the anonymous pool used to determine the chargeback risk or trust level of the next request using any of that same information..
Such sharing can be very useful to determine trust for an individual. If an account has existed on several platforms for 5 years with all of the same details and no fraud ever reported against it, one could probably conclude that they are genuine. Conversely, if a card is seen on 5 newly created accounts on 3 different platforms with 4 different names and addresses, then there is a good chance that there is some fraud associated with it, especially if there is a chargeback on any of those accounts.
Rule engines may be used to determine how to act upon chargeback or fraud risk recommendations and react to observations that could be fed through machine learning models to better score activities being shared throughout the anonymous community of companies.
Fraudsters do not tend to stick to one platform unless it is the only one they find success on. Often, they will attack several companies within the same pillar in the exact same manner. If that information can be shared across the affected companies, they may all be better able to defend against the chargeback and fraud onslaught. Knowing the same for good customers would help limit false positives. Participating companies get useful information that is beneficial for both of them simultaneously.
Is There a One Stop Shop for Chargeback Prevention?
When choosing a fraud solution, you want multiple layers of protection, especially those which will not interfere with a genuine customer’s journey. All of these features work behind the scenes to prevent chargebacks and fraud, acting only where it is necessary.
If everything is used appropriately, there will be no friction added to a customer outside of account creation verifications and the challenges sent upon risky logins. That is enough to reduce overall fraud risk and significantly reduce the level of chargebacks.
Unfortunately, for right now, there is no one-stop-shop solution available to merchants to prevent chargebacks or eliminate risk. Some come close, but none include all available features. That is why ultimately, it is best to create an omni-channel fraud prevention strategy specific to you.
This article was contributed by Shawn Colpitts, Senior Fraud Investigator at Just Eat Takeaway.com