CMStatistics 2021: Start Registration
View Submission - CFE
A1675
Title: Statistical causality: What more can we learn about our data Authors:  AB Zaremba - University College London, Quantitative Risk Solutions Lab (United Kingdom) [presenting]
Gareth Peters - Heriot-Watt University (United Kingdom)
Abstract: A novel testing framework for statistical causality has been developed in general classes of multivariate nonlinear time series models. Our framework allows us to study causality in the trend, volatility, or both, and accommodates a range of structural features that are important when modelling financial time series, including long memory. However, we want to emphasise the distinction between choosing an exact model and finding a causal relation. With our framework, we test the dependence with regards to the structures that are specified, and we show a range of examples of good performance even for misspecified models. We then focus on the example of commodity futures data and show the usefulness of testing for causality under different model specifications as a way to explore the data and the potentially complex dependence relationships. We point out what can be learned from comparing statistical causality tests based on GPs to analogous tests based on linear regression, explain what makes the latter overconfident, and show a breakdown of steps that can adjust for the overconfidence.