Contributed Session 13: Advanced Methods in Causal Analysis

This study investigates the causal relationships influencing financial returns of the European Union Emissions Trading System (EU ETS), a cornerstone policy instrument for reducing greenhouse gas emissions in Europe. Using data spanning January 2013 to April 2024, we explore the effectiveness of a recently introduced non-parametric tool, Differentiable Information Imbalance (DII), from the physics literature, to identify causal drivers of EU ETS financial returns. To benchmark the results, we compare DII with multivariate Granger causality, a traditional linear approach based on vector autoregressive (VAR) models. Both methods highlight significant causal variables, such as coal futures prices and the IBEX35 index. However, they also yield distinct causal sets, with the DII approach capturing non-linear relationships overlooked by the linear model. Through experiments on synthetic datasets, we demonstrate how the limitations of linear methods could account for these discrepancies. This work underscores the value of non-parametric techniques in capturing complex causal relationships in financial markets.

This paper studies a semiparametric quantile regression model which handles endogenous variables and random right censoring. The model assumes that the structural quantile of the outcome variable is linear with respect to the covariates, while the censoring is conditionally independent of the outcome, given the values of covariates. To address the endogeneity issue, we use instrumental variables. The instrument and the endogenous variable are supposed to be discrete, while the exogenous covariates can be either discrete or continuous. Our estimation process takes inspiration from the innovative presmoothing methodology outlined in [2]. We validate our approach through simulation studies.

Mediation analysis is widely used to assess if the effect of an exposure on a response of interest is conveyed by an intermediate variable called mediator. With the advent of big data, it is increasingly frequent to combine mediation analysis with machine learning, because of the high-dimensionality of the mediators and/or the covariates. In this work, we compare three machine learning approaches for es- timating natural mediational effects through a simulation study, showing that while Lasso-based methods perform well in linear settings, more flexible models such as Random Forests yield better results in non-linear or complex relationships where a high number of covariates are involved.