High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula\n Processes

David Salinas, Michael Bohlke‐Schneider, Laurent Callot, Roberto Medico, Jan Gasthaus

arXiv (Cornell University) · 2019 · 67 citations · 0 references

DOIFull text

Open access

Concepts

Abstract

Predicting the dependencies between observations from multiple time series is\ncritical for applications such as anomaly detection, financial risk management,\ncausal analysis, or demand forecasting. However, the computational and\nnumerical difficulties of estimating time-varying and high-dimensional\ncovariance matrices often limits existing methods to handling at most a few\nhundred dimensions or requires making strong assumptions on the dependence\nbetween series. We propose to combine an RNN-based time series model with a\nGaussian copula process output model with a low-rank covariance structure to\nreduce the computational complexity and handle non-Gaussian marginal\ndistributions. This permits to drastically reduce the number of parameters and\nconsequently allows the modeling of time-varying correlations of thousands of\ntime series. We show on several real-world datasets that our method provides\nsignificant accuracy improvements over state-of-the-art baselines and perform\nan ablation study analyzing the contributions of the different components of\nour model.\n