Maastricht University, Netherlands
Title: Causal Inference from Multivariate Time Series: Principles and Problems
Abstract: In time series analysis, inference about cause-effect relationships among multiple time series is commonly based on the concept of Granger causality, which exploits temporal structure to achieve causal ordering of dependent variables. One major and well known problem in the application of Granger causality for the identification of causal relationships is the possible presence of latent variables that affect the measured components and thus lead to so-called spurious causalities. This raises the question about whether Granger causality is an appropriate tool for causal learning; indeed, there are many researchers that deny any such claim.
To answer the question in more depth, we present a graph-theoretic approach for describing and analysing Granger-causal relationships in multivariate time series that are possibly affected by latent variables. It is based on mixed graphs in which directed edges represent direct influences among the variables while dashed edges—directed or undirected—indicate associations that are induced by latent variables. We show how such representations can be used for inductive causal learning from time series and discuss the underlying assumptions and their implications for causal learning. Finally we will discuss non-Markovian constraints imposed by latent variable structures and how these can be exploited for causal inference.
Background
Research Interests and Projects
- Causal inference for multivariate time series with latent variables
In applications involving multiple time series, the concept of Granger causality is frequently used to describe the causal relationships among the variables. Although the interest in Granger causality has been increased, for instance, in neuroscience, its suitability as a measure for causality is much disputed. The objective of my research is to clarify the meaning and limitations of Granger causality and to develop algorithms for causal inference from time series data that take these limitations into account. - Dependence analysis in multiple heterogeneous time series
The objective of this project, carried out in cooperation with Ratheon BBN Technologies (USA), is to predict/detect rare events of significance on a population level from multiple heterogeneous openly available data sources (including e.g. social network data). Within this task, our project focuses on the development and investigation of methods for aggregating and extracting relevant information from multiple heterogeneous time series where particular attention will be given to approaches based on Granger causality, dynamic factor modelling, and model averaging. This project is part of the open source indicator (OSI) program financed by the US government agency IARPA. PhD student: Carlos A. Moreno - Electricity spot price modelling
Electricity spot prices feature a number of stylized facts such as mean reversion, strong seasonality, and frequent occurrence of short periods of extreme prices. These price spikes occur more often in markets with compulsary market participations such as the Australian NEM market. The objective of this project is to develop new approaches for modelling and forecasting electricity spot prices and spikes. PhD student: Dennis Tuerk - Semi-parametric dynamics factor models for non-stationary processes
Current approaches for fitting dynamic factor models to nonstationary time series are based on dynamic principal components analysis in the frequency domain. These approaches are fully nonparametric and depend strongly on the chosen bandwidths for smoothing over frequency and time. In this project, a semi-parametric approach in which only parts of the model are allowed to be time-varying is investigated. PhD student: Anne van Delft