Alternative Solutions - One could use Urchin Tracker and other Google web analytics to assess user behavior, but this won't provide real-time behavioral monitoring. Device Identification and Hot/Warm Lists can be used to identify past users that have displayed bad behavior in the past, but there is no true alternative to Behavioral Monitoring
Building this In-House - Most online businesses already monitor site behavior to some extent with Urchin trackers to see which pages users are visiting and where they are coming from. This could be built out into a low-cost, low-end form of Behavioral Monitoring, but the value of a software or service provider is that behavior is reported in (near) real-time so it can be identified and acted upon immediately. The provider may also provide an interface and ability to build rules and containers around users currently displaying bad behavior.
CUSTOM MODELING & ANALYTICS TECHNIQUE OVERVIEW
Custom Modeling and Analytics solution providers focus on providing tools or services related to the analysis of data to create custom models and blended solutions. The key word is modeling, and more specifically the building of custom models. These models could be multi-linear regression models, neural nets or Bayesian. There are several components to this type of solution. First, there are the modeling tools and the platform that provide a multitude of risk signals, complex algorithms and the overall system architecture for applying the analytic models, possibly including a user interface to create new and update existing models. The next component is the development and continued maintenance of the custom model and algorithms, which requires highly specialized resources to provide the data analysis, design, and the actual building and running of a model based solution. Providers may offer solutions that only provide the tools and platform (self-managed models), only the model design (consultants, contractors or data scientists to build the initial model), or they may offer the full suite of services. Key considerations when implementing or buying this functionality include:
What solution type is best for your organization: a self-maintained modeling platform and tool set or a full managed service where the vendor designs, trains and maintains custom models to sustain effective results?
If designing or maintaining models in-house, what are the true costs of this service (likely including full-time data scientists, analysts and/or statisticians in addition to costs to integrate and use the modeling platform)?
What signals and risk techniques can be applied and used by the analytic models? Can organizations provide custom data fields or signals? Can organizations integrate other third party services that provide a signal and utilize that signal within their models?
Does the provider offer the ability to create multiple custom models and/or segmented models? (i.e. multiple custom models depending on customer risk profile, product type, etc.)
Does the provider specialize in modeling for acquirers, card issuers or merchants, or do their solutions support all three?
Can the modeling algorithms capture and include behavioral signals, authorization signals or responses and post-transaction information?
Does the service provide a reporting dashboard to monitor and track model performance?
Does the service provide an interface for creating new models or updating existing ones? What level of technical skill is required to create or update models? (Can a fraud or risk analyst without a technical coding/programing background implement updates or new models?)
How long does it take to get the custom modeling solution up and running? This typically includes time for technical integration of the modeling platform, incorporating custom tools or signals, time to build the initial model(s), as well as time to test and train the model(s) with live data in a practice setting.
HOW DOES IT WORK?
The service provides the platform to execute one or multiple custom models, which may or may not include the design and maintenance required for these models to perform. The models themselves attempt to make predictive correlations between data elements and characteristics of a transaction or customer/cardholder profile. This can include many different factors, techniques and signals that the organization is able to provide, such as the distance between a shipping address and IP location, comparing the current transaction to customer/cardholder specific purchase patterns, and many other factors.
Each model is represented by an equation or algorithm that relies on many different variables, potentially hundreds. Each variable represents a risk signal, quantitative or qualitative piece of information, and how that variable affects the overall assessment of risk for the particular transaction is defined by each variable coefficient. The building of each model, including the weight of each variable coefficient, is crucial to the success of this service and should be conducted by a team of data scientists, statisticians and/or risk analysts. Models typically don't have rigid requirements or signals of risk, but rather more fluid thresholds and risk signals considered in the context of many other signals or model features.
Each variable may not provide a great signal of risk by itself, but when considered with hundreds of other variables may be meaningful. There can be a very high level of customization with each model in terms of what variables are included and the coefficients or weights applied to each variable. Typically historical data is used to assess the proper variables and predictors of risk as well as determining the appropriate coefficients. It is also standard to test and train models in a practice mode against live transactions to assess performance prior to being put into production.
A simplified example formula for a Multiple Linear Regression model is as follows where the dependent variable (Ƴ) represents the outcome or score and β equals the coefficient (weight) of each independent variable (χ): Ƴ = β0 + β1χ1 + β2χ2 … + βnχn
The two most common forms of custom modeling for fraud detection and prevention are Bayesian linear regression models and neural network models, which may also be referred to as artificial neural networks (ANNs).
HOW DO YOU USE THE RESULTS?
The results from the custom models could be a ranged score (such as from 1 to 100 or 1 to infinity) or it could be a simple pass or fail type of response. In both cases organizations will need to set up their systems to interpret these results and this will determine the outcome of a transaction. For example, an organization will determine what scores or thresholds result in declining a transaction, accepting a transaction or performing additional screening such as manual review.
Modeling provides good information, but organizations still have to build the logic into their system to handle the responses.
DID YOU KNOW
Custom Modeling and analytics as a fraud prevention technique are used by card issuers, merchants as well as acquirers and processors. It may also be referred to as modeling, analytic modeling, predictive models, neural nets or empirical models.
Issuers may use custom analytic models to assess the aggregate volume and patterns of card purchasing activity as well as cardholder purchase characteristics relative to historical and known patterns. When analytic models detect anomalies or unusual activity it may result in the issuer contacting the cardholder to confirm activity or not authorizing an attempted transaction due to suspected fraud.
Merchants may use custom models to provide a fully customized risk score or accept/deny decision. Analytic and modeling providers offer the platform for a merchant or other organization to include the specific data and variables they have available, while the risk weights and correlations of the data is specific to that merchant or organization.
Payment processors, acquirers, gateways and other payment service providers often use analytic models to monitor and detect risk within their merchant client portfolio. This is not limited to fraud risk as it typically also includes identifying changes in underwriting risk.
It is Important to Know That:
Models that are built and/or used solely in a one-merchant implementation don’t get the benefit of seeing consumer activity outside of their business. Some modeling solutions have the ability to incorporate shared, cross-merchant data.
It can be very difficult to set up and understand how to effectively design and implement custom models if building them in-house. Managing and maintaining custom models in-house require significant human capital and resources. Some custom modeling analytics providers may offer the platform but it is up to the organization to design or monitor and maintain their own models. Other providers will design and manage models as part of a full managed service offering.
Machine learning is often discussed with modeling and analytics and this term is generally used to refer to models that can be trained and updated automatically. However, the level of automated model training and adjustments can vary quite a bit based on the provider and will require the input of post-transaction final outcomes to recognize where better decisions could have been made. The level of learning supervision data scientist or statisticians are involved with will vary and organizations should be aware of any significant adjustments made to models. Machine learning does not imply that an organization can deploy a model and it will effectively train and adjust itself as needed perpetually. But machine learning can maintain strong model performance by detecting changes in risk patterns and updating models accordingly.
Join the Digital Trust & Safety Revolution with Sift
Sift is the leader in Digital Trust & Safety, empowering digital disruptors to Fortune 500 companies to unlock new revenue without risk. Sift dynamically prevents fraud and abuse through industry-leading technology and expertise, an unrivaled global data network of one trillion (1T) events per year, and a commitment to long-term customer partnerships. Global brands such as DoorDash, Poshmark, and Twitter rely on Sift to gain a competitive advantage in their markets. Visit us at sift.com, and follow us on LinkedIn.
KEY NOTES
Alternative Solutions - Model Based Fraud Scoring and Behavioral Monitoring typically rely on and incorporate forms of modeling. Model Based Fraud Scoring techniques are based on a proprietary model, but these are not client-specific, custom models. Behavioral Monitoring services may include modeling analytics to provide a behavioral score based on the many behavioral characteristics, but this is focused on user behavior not user provided data.
Building this In-House - The modeling platform and technology is something that can be built in-house, although this usually very expensive to develop and build from scratch. Even if an organization utilizes a modeling platform they can opt to build and manage the models in-house, if they have the skilled resources (such as statisticians, data scientists and risk analysts) to do so.
The Cost - Costs will vary based on the vendor and the level of service they offer. Typically access to the technology and platform for running the models is charged on a per transaction basis, but there can be additional costs for designing, maintaining and updating the models.
Sample Vendors - Features Analytics, Brighterion, AI Corp, ACI Worldwide, LexisNexis, ModelShop, FeatureSpace