Anomaly Detection, Density Estimation

For my MSc thesis (see downloads section), “Anomaly Detection using Variational Autoencoders (VAE)”, I developed two, generative model based, methods for anomaly detection. I applied them to several models and achieve new state of the art results on the ECG200 dataset from the UEA and UCR Time Series Classification Repository. Both methods, one using standard and another conditional VAEs, were later, independently, published in the literature by different research teams.

For my work in financial forecasting I have further developed these methods and combined them with flow based density estimators to try to identify abnormal observations, obtain uncertainty estimates for predictions and infer possible inflection or regime change points in price trajectories.

Variational Autoencoders, conditional generative models, probability density estimation, inverse autoregressive flows, anomaly detection, out-of-distribution estimation, likelihood ratios