Gianluca Bontempi (ULB) - "Machine learning for multivariate time series: from forecasting to causal inference"
Conventional approaches in times series literature are restricted to low-dimension series, linear methods and short horizons. Big data revolution is instead shifting the focus to problems (e.g. issued from the IoT technology) characterized by very large dimension, nonlinearity and long forecasting horizon. The presentation will discuss a number of settings where machine learning approaches may be used to deal with time series forecasting and causal understanding. The first part will focus on machine learning strategies for one-step-ahead and multi-step-ahead forecasting both in univariate and multivariate tasks. In particular we will discuss MIMO strategies for multi-step-ahead forecasting of univariate time series and DFML, a machine learning version of the Dynamic Factor Model (DFM), a successful forecasting methodology well-known in econometrics. The DFML strategy is based on a out-of-sample selection of the nonlinear forecaster, the number of latent components and the multi-step-ahead strategy. While accurate forecasting may be obtained by learning associative dependencies between different time instants and time series, an open challenge is how to discriminate between associative dependencies and effective causal relationships. This is particularly challenging in large-variate temporal settings (e.g. spatio-temporal time series) where the multivariate nature of interactions induces a significant correlation between most of the variables. The second part of the presentation will discuss how supervised classification techniques may be used to identify causal dependencies in time series once a proper set of context-dependent descriptors are introduced. The approach, called D2C (Dependency to Causality) performs three steps to predict the existence of a directed causal link between two variables in a multivariate setting: (i) it estimates the Markov Blankets of the two variables of interest and ranks its components in terms of their causal nature, (ii) it computes a number of asymmetric descriptors and (iii) it learns a classifier (e.g. a Random Forest) returning the probability of a causal link given the descriptors value.