Blog post
Building Blocks to Improve Fraud Detection and Enhance Modeling
SentiLink
Published
September 18, 2023
High-quality data is foundational for precise models targeting fraud nuances and resulting economic outcomes. Intelligence around identity, fraud and risk brings resilience and lift to models that allow data teams to root out a broad array of evolving issues. Data teams ingesting attributes into their models benefit from building blocks that are generated and optimized to target business-specific problem areas. SentiLink Facets, which are available for access via our single end-point API, are purpose-built, flexible attributes designed to help tackle our partners’ hardest fraud problems.
Before we dive into this solution, let’s take a broader look at the importance of high-performing attributes.
What are Attributes?
Attributes are feature-extracted data points used to interpret model inputs and improve the predictive performance of a model. These data points have many uses and are especially powerful when applied to fraud and risk modeling, as well as managing the decision flow for consumers applying for financial products.
Why are Attributes Valuable?
Effective attributes support targeting fraud vectors and behaviors specific to the financial products and demographics of a given financial institution beyond general fraud trends. SentiLink partners have particular needs, and we help them reach their targets based on their unique situation.
Use Cases
Based on our initial work with FI partners assessing attributes for their models, we identified the following five primary use cases for teams interested in incorporating these building blocks for model lift.
Targeting Specific Partner Variables
Attributes support FIs seeking to predict specific behaviors not directly targeted by existing 3rd party fraud solutions, such as fraudulent first payment default or illegitimate bank disputes.
Identify Good Populations for Reduced Friction
Attributes also enable FIs to surface non-fraudulent, high-performing populations – allowing for policies to increase approval rates or expedite the approval process for these applicants.
Deepen Understanding of Risky but not Clearly Fraudulent Applicants
FIs can dive deeper into the riskiness of applications that do not necessarily meet the criteria for an auto-decline but are also demonstrably outside the auto-approve standards. For SentiLink’s Synthetic Abuse and ID Theft Scores, which range from 1 to 999, a score in the 300 to 600 range may require additional clarity before proper treatment is applied. Attributes can provide additional context to more definitively route applications for further review or verification. For example, if a borderline application returns an attribute indicating the phone has never been associated with the applicant, the FI can route the application for phone verification.
Enhance Recall for ID Theft
ID theft presents itself differently in each FI’s application pool and process (e.g., some verify email or phone, and some do not). Attributes can boost fraud capture from an ID theft scoring solution for the FI’s particular situation. ID theft solutions on the market tend to be either very targeted without capturing all nuances of ID theft, or too broad and catch a lot of ID theft but with high false-positive rates. Attributes offer flexibility to increase recall and flag additional cases or narrow the capture and increase precision.
Manage Model Governance Issues
For large and heavily regulated FIs with regulatory concerns, attributes provide ingestible data points that are explainable and directly identifiable. In these cases, we encourage partners to employ attributes as an initial solution while working towards integrating SentiLink Scores into their governance process to achieve the full power of Scores and attributes together.
What are SentiLink Facets?
SentiLink Facets are detailed and informative data points derived from identity matching and linking as well as account opening activity within the applications we see in our network – in other words, our complete library of the attributes described above. When ingested into a model, these data points have proven to be predictive of various types of fraud and risky behaviors.
Identity intelligence comprises the foundation of meaningful attributes. Facets extract this intelligence from Manifest, our comprehensive catalog of identities. The additional processing work of logically merging and matching the raw identity inputs ensures we are making sense of even the most convoluted identities.
Network activity intelligence, derived from application submissions across a broad consortium of financial institutions and fintechs, fills in the next layer necessary for producing valuable attributes. SentiLink’s network, containing over 1B applications, establishes a detailed composite of applicant activity patterns. We aggregate and rationalize applications into clusters on multiple dimensions to ensure our attributes trace risky behavior patterns to their proper source.
Additional metadata also provides Facets with more context on phones, emails, and IPs. No matter the application details, we optimize our attributes for predictiveness with the same feature intelligence ingested into our best-in-class scoring models.
How are Facets Used?
SentiLink's full catalog of Facets includes 300+ attributes, which we offer in two formats, depending on the specific situation of the partner in question:
all_attributes
, which contains all of our Facet attributes. This option is suitable for partners who have already built and are maintaining their own fraud prevention models and/or fraud rules.essentials
, a smaller bundle containing a subset of Facets that will be provided along with recommended rules for their implementation.
SentiLink may also consult with partners to develop custom bundles for partners who have specific requests and/or partners who collect limited PII.
Why use Facets?
Our list of attributes is growing as we learn more about the problems our partners face. With Facets, not only are you receiving data points that do not require feature extraction, you receive attributes that are:
- Predictive: proven to add significant lift to models targeting partner-specific labels that generic models do not capture.
- Smart: SentiLink’s approach to interpreting raw data to generate highly insightful attributes is unique and optimized for performance.
- Proprietary: our identity catalog and network of leading financial institutions and fintechs enhance models with intelligence distinctive to SentiLink.
- Built for model use: all attributes are numeric or boolean to be easily implemented in new or existing models.
Partnering with SentiLink
We take a consultative approach to understanding our FI partners' needs to close gaps and accomplish goals specific to each partner. Those partnering with SentiLink and using our attributes can expect the following:
- Attribute Selection: our Analytics team can provide analysis on the most predictive attributes for specific needs.
- Funnel Placement: guidance on using Facets in different parts of the same funnel to achieve the best results.
- Model Implementation: our Data Science team can provide high-level guidance on how to implement our attributes in models.
- Fraud Expertise: our Fraud Intelligence Team can show how certain fraud trends manifest themselves in a population.
- Trend Analysis: as part of a retrostudy and as part of an ongoing partner relationship, SentiLink can perform analytics using a partner’s internal labels to return insightful observations on a given population or specific moments in time (e.g., after making changes to their product or process, during a fraud attack, etc.).
In this partnership, we will also work with you to source additional attributes our current list may not cover.
Conclusion
For financial institutions, the better the data, the more effectively their models can identify specific fraud patterns. SentiLink supports FIs with attributes built on identity, fraud and risk insights to keep models ready to handle evolving fraud. If you’re interested in learning more about Facets, reach out to us. We can conduct a retrostudy with your team to determine which attributes best fit your applicant population and goals.
Note: The product and use cases described above are not intended to capture credit risk, as SentiLink supplies GLBA data, not FCRA data.
Related Content
Blog article
November 20, 2024
When Does "Cloning" Work for Reject Inference?
Read articleBlog article
November 19, 2024
Consumer Portfolio Services Partners with SentiLink to Enhance Fraud Prevention
Read articleBlog article
November 14, 2024