Granger causality in the cryptocurrency market: a machine learning approach

This dissertation examines Granger causality and its role in identifying predictive relationships within the cryptocurrency market. First of all, the properties of Vector Autoregressions are outlined and their use in modeling the joint dynamics of multiple time series is discussed. Based on this foundation, the thesis analyzes Granger causality and several of its methodological extensions relevant for studying interactions among financial variables. Then, machine learning approaches that can serve as alternatives or complements to traditional causality testing are explored. In particular, shrinkage methods are examined, focusing on Lasso Regression and its variants, as well as an advanced alternative to these methods, the One Covariate at a Time Multiple Testing (OCMT) procedure. Moreover, the use of Neural Networks to capture potential nonlinear dependencies is investigated and related to causal detection. Finally, an empirical application using major cryptocurrencies evaluates how these methods detect directional predictability and interaction patterns.