Financial Signal Processing and Machine Learning

Chapter 4Temporal Causal Modeling

Prabhanjan Kambadur¹, Aurélie C. Lozano² and Ronny Luss²

¹Bloomberg LP, USA

²IBM T.J. Watson Research Center, USA

4.1 Introduction

Discovering causal relationships in multivariate time series data has many important applications in finance. Consider portfolio management, where one of the key tasks is to quantify the risk associated with different portfolios of assets. Traditionally, correlations amongst assets have been used to manage risk in portfolios. Knowledge of causal structures amongst assets can help improve portfolio management as knowing causality – rather than just correlation – can allow portfolio managers to mitigate risks directly. For example, suppose that an index fund “A” is found to be one of the causal drivers of another index fund “B.” Then, the variance of B can be reduced by offsetting the variation due to the causal effects of A. In contrast simply knowing that “A” is correlated with “B” provides no guidance on how to act on index “B,” as this does not mean that the two indexes are connected by a cause-and-effect relationship; hedging solely based on correlation does not protect against the possibility that correlation is driven by an unknown effect. Moreover, causal structures may be more stable across market regimes as they have more chance to capture effective economic relationships.

In order to mitigate risks effectively, we need several enhancements to mere causality detection. First, we need to be able to reason ...