Open Banking: Multi-Layered Self-Calibrating Models

Difficulties in accessing high-quality historical data or anticipated systematic shifts in data are barriers to the development of supervised machine learning (ML) models. This challenge provides a rationale to create models that are able to learn patterns and variances online as the data is streamed. These models are known as self-calibrating models, and they can be used for outlier detection, which is often a strong proxy for financial crime events. A good example of effective use of self-calibrating models is for open banking, where labeled historical data is not yet available and the evolution of open banking in terms of adoption is still unclear. Without adequate data, supervised ML models such as neural networks can’t be built, therefore self-calibrating models provide a valuable alternative. Following the first white paper, Open Banking: Streaming Analytics for Behavioral Profiling, this white paper focuses on technology that uses the data refined through our behavioral analytics methods to produce a score indicative of fraud or non-fraud. This technology is called multi-layered self- calibrating (MLSC) models and was developed and patented by FICO to determine how unusual the transaction is based on quantiles of key feature detectors often derived based on expert knowledge to determine risky behaviors in production.