My analytics team has been working on some exciting new technologies that improve fraud detection while reducing declines of legitimate transactions. One such research innovation boosts performance of traditional fraud models, which use a single customer’s past purchases to determine whether a current transaction is consistent with that customer’s historical behavior patterns. The new analytic technology goes further by examining patterns of transactional behavior across millions of customers to continuously update “behavior archetypes” for each customer in real time. As such, the analytics provide a more complete and textured view of customer behavior.
Our new analytic technology leverages a large-scale consortium-wide data set of historical behavior and, with some modern Bayesian methods, computes data-driven archetypes based on patterns across many customers. Think of these archetypes as the “atoms” in the periodic table of customer behaviors. Real customers are “compounds,” composed of a flexible mixture of various archetypes. At a given point in time, for instance, a customer might be 50.5% home improver, 11.8% tech lover, and so on. This approach contrasts with typical customer segmentation, where each person is forced into one or another hard category, often by external demographic variables or arbitrary rules.
The archetypes are core probabilistic behavior patterns that predictively distinguish any given customer’s typical behavior from the crowd. They are not based a few selected “representative” customers chosen from the data set, nor do they depend on an expert to manually insert domain knowledge, in contrast to typical customer segmentation strategies.
Because these analytics work in the stream, the archetype allocations are updated in real time with each new transaction. This allows the fraud model to determine whether any new, previously unseen transactional behavior should be viewed as fraudulent or simply new behavior. The two figures below graphically illustrate the concept.
How can these new analytics determine that previously unseen transactional behavior is "normal" even if the customer has never performed that specific transaction before? The analytics compare it with equivalent transactions from other customers that are similar to the one being scored.
Let’s say there’s a cardholder with frequent domestic hotel and airline purchases that falls heavily into a “traveler” archetype. One day, there’s a high-amount foreign taxicab transaction, the first of its type for the cardholder. Taken on its own, such a transaction may be considered quite risky. But because the “traveler” archetype weights this as high-probability transaction, the model lowers the score and avoids a false positive.
When we tested this new technology on the US Credit model of FICO® Falcon® Fraud Manager, we saw substantial improvement across all performance metrics. In the graphic above, I’ve highlighted the relative improvement in fraud detection and relative reduction in false positive ratio at a couple account false positive ratios. While the technology is not yet on our Falcon roadmap, it has exciting potential as we find new ways to continuously drive improvements in our fraud solutions.