Daniel Lasaga

Dan is a Chief Data Scientist for Regulatory & Legal Support products in Deloitte Risk & Financial Advisory. His skills and experience sit on the vertices of technology, communications, and management. He has former experience in a wide variety of industries including manufacturing, utilities, banking, and ecommerce. His work has taken him in a wide arch stretching from hands on algorithmic programming to high-level interactions with decision makers. He has had the opportunity to publish eminence work on machine learning as well as develop and teach trainings for executives and fellow practitioners. Through Deloitte Advisory, he has had the opportunity to work on exciting projects combining his talents and expertise in structural analysis, data modelling, software engineering and management to tackle fraud detection, manufacturing and supply chain risk and other forms of risk for clients.

ABSTRACT

Applications of MDNs for Anomaly and Risk Detection in Financial Data

The mercurial nature of fraudsters and the complex systems within which they work render traditional classification models insufficient to capture bad behavior with, often leaving data scientists with unsupervised modeling techniques. We will introduce Mixture Density Networks (MDNs) as a form of one-class classification modeling focusing on the abstraction of normal and periphery activity that is highly tuned to both expectation and variability of where fraud or risk may be. MDNs predict probability distributions of potential output values based on input features. The distribution of outputs can be compared to actual outputs to determine outliers informed by the complexities of the input data via a neural network. MDNs have shown promising results in detecting anomalies and rare-events in complex systems in recent studies. They can be particularly useful predicting a continuous variable in the theoretical parameter space when experimental data has large uncertainties. We will describe the approach of using MDNs and the comparison to traditional anomaly detection techniques to detect business risk or unexpected behavior.