Deep Extreme Mixture Model for Time Series Forecasting

Abstract

Time Series Forecasting (TSF) has been a topic of extensive research, which has many real world applications such as weather prediction, stock market value prediction, traffic control etc. Many machine learning models have been developed to address TSF, yet, predicting extreme values remains a challenge to be effectively addressed. Extreme events occur rarely, but tend to cause a huge impact, which makes extreme event prediction important. Assuming light tailed distributions, such as Gaussian distribution, on time series data does not do justice to the modeling of extreme points. To tackle this issue, we develop a novel approach towards improving attention to extreme event prediction. Within our work, we model time series data distribution, as a mixture of Gaussian distribution and Generalized Pareto distribution (GPD). In particular, we develop a novel Deep eXtreme Mixture Model (DXtreMM) for univariate time series forecasting, which addresses extreme events in time series. The model consists of two modules: 1) Variational Disentangled Auto-encoder (VD-AE) based classifier and 2) Multi Layer Perceptron (MLP) based forecaster units combined with Generalized Pareto Distribution (GPD) estimators for lower and upper extreme values separately. VD-AE Classifier model predicts the possibility of occurrence of an extreme event given a time segment, and forecaster module predicts the exact value. Through extensive set of experiments on real-world datasets we have shown that our model performs well for extreme events and is comparable with the existing baseline methods for normal time step forecasting.