Financial institutions are always looking for ways to combat newer fraud schemes—schemes that arise between fraud model developments and are not well-represented in the historical data. As I discussed in a recent post, one way to boost performance is to add analytics that are adaptive or self-learning.
Deploying adaptive models in a “cascade” architecture, where they act as a secondary analytic layer, strengthens conventional fraud detection. FICO client work shows that the cascade approach delivers significant lift when used with a strong neural network model built from an abundance of reliable, high-quality historical data. This holds true for large issuers with enough historical data to build a custom base model, as well as issuers who combine their data and produce a single consortium model.
A cascaded adaptive model is highly efficient because it works with the base neural network model without interfering with it. The adaptive layer is simply “bolted on,” and the base model remains untouched, continuing to produce highly predictive scores.
Keeping the layers separate also enables companies to turn up the “sensitivity dial” of their fraud detection to short-term changes in fraud behavior, while shielding the base model from the effects. Adjustments critical to detection today might not be relevant next week. For example, an adaptive model that captures just the past week of fraud/no-fraud case data will be very quick to adjust scoring to fraud burst behavior. An adaptive model with a longer-time horizon might provide broader protection across more fraud patterns, when an issuer is less concerned about burst activity.