Learning latent causal relationships in multiple time series

TL;DR

This paper introduces an unsupervised method to identify latent causal relationships in multiple time series by projecting data into components that maximize causality, demonstrated on synthetic, brain, and cryptocurrency data.

Contribution

It presents a novel technique to uncover latent causal structures in multivariate time series without supervision, using a linear mixture model and optimization of causality measures.

Findings

Successfully recovered causal relationships in synthetic data

Identified meaningful causal links in brain recordings

Detected causal patterns in cryptocurrency prices

Abstract

Identifying the causal structure of systems with multiple dynamic elements is critical to several scientific disciplines. The conventional approach is to conduct statistical tests of causality, for example with Granger Causality, between observed signals that are selected a priori. Here it is posited that, in many systems, the causal relations are embedded in a latent space that is expressed in the observed data as a linear mixture. A technique for blindly identifying the latent sources is presented: the observations are projected into pairs of components -- driving and driven -- to maximize the strength of causality between the pairs. This leads to an optimization problem with closed form expressions for the objective function and gradient that can be solved with off-the-shelf techniques. After demonstrating proof-of-concept on synthetic data with known latent structure, the technique is applied to recordings from the human brain and historical cryptocurrency prices. In both cases, the approach recovers multiple strong causal relationships that are not evident in the observed data. The proposed technique is unsupervised and can be readily applied to any multiple time series to shed light on the causal relationships underlying the data.