Statistical Tests for the Detection of the Arrow of Time in Vector Autoregressive Models

Pablo Morales-Mombiela, Daniel Hernández-Lobato, Alberto SuárezIJCAI 2013 · OpenReview

Abstract

The problem of detecting the direction of time in vector Autoregressive (VAR) processes using statistical techniques is considered. By analogy to causal AR(1) processes with non-Gaussian noise, we conjecture that the distribution of the time reversed residuals of a linear VAR model is closer to a Gaussian than the distribution of actual residuals in the forward direction. Experiments with simulated data illustrate the validity of the conjecture. Based on these results, we design a decision rule for detecting the direction of VAR processes. The correct direction in time (forward) is the one in which the residuals of the time series are less Gaussian. A series of experiments illustrate the superior results of the proposed rule when compared with other methods based on independence tests.
BibTeX
@inproceedings{moralesmombiela2013statistical,
  author = {Pablo Morales-Mombiela and Daniel Hernández-Lobato and Alberto Suárez},
  title = {{Statistical Tests for the Detection of the Arrow of Time in Vector Autoregressive Models}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2013},
  url = {https://openreview.net/forum?id=r1lBv4fO-S},
}

Loading related content...

Figure