High-Dimensional Probabilistic Time Series Prediction Via WaveNet+t

Abstract

High-dimensional time series inference plays a crucial role in various fields (e.g., economic analysis, inventory analysis, electricity consumption, and stock market forecasting). However, classical time series models mostly deal with handling one dimension time series dataset with point estimate. In this work, we use a variant of WaveNet [9] in combination with a probability distribution (e.g., Student's t-distribution) for multivariate probabilistic time series prediction. WaveNet [9] and other related works [10], [11] used dilated causal convolutional neural networks (CNN) to extract the long/short term patterns from time series dataset. It also integrates residual network with the dilated causal CNN to solve the vanishing/exploding gradient problems and make models to be more expressive. Multi-step, probabilistic prediction for multivariate time series is generated by sampling from the conditional distribution (given input data) produced by the proposed WaveNet+t network. Our model demonstrates better or comparable performance on different real-world high dimensional time series dataset (e.g., Wikipedia with 9535 variables) when compared with the other state-of-the art multivariate probabilistic models.