Explore the insights of Granger causal networks and how they unravel direct and indirect feedback in time series analysis.
The exploration of Granger causal networks opens new avenues in the analysis of structural vector autoregressive models (SVARs). This article aims to elucidate the intricacies of direct and indirect effects that exist among temporal variables through a non-parametric variable selection process, thereby enriching our understanding of feedback mechanisms.
Over the past decade, vector autoregressive (VAR) models have emerged as staples in econometrics, widely adopted by analysts, researchers, and policymakers alike. These models facilitate a range of econometric analyses, including impulse response studies, forecasts of endogenous variables, and assessing cross-correlation among temporal variables.
Despite their popularity, VAR models do not inherently provide mechanisms for delineating the effects of endogenous variables into direct and indirect components. For instance, understanding whether a particular variable affects another directly or only through an intermediary poses a computational challenge. This complexity arises because traditional methods primarily measure aggregate impacts through orthogonal impulse-response functions without the ability to independently trace effects through each equation in the VAR setup, creating a significant analytical hurdle.
In addressing these challenges, the concept of causality network graphs has surfaced as a powerful visualization tool in panel data research. These graphs allow for a clearer representation of how variables interconnect within a dataset, facilitating easier inference measurements. Through this article, we will construct a Granger causal network graph (G(e,d)) using live data, illustrating how these networks can clarify the causal relationships between different endogenous variables.
For our example, we will focus on the log returns of the QQQ ETF, which tracks the NASDAQ-100 index, along with other significant indicators such as Relative Strength Index (RSI), Bollinger Percent (B(pctB)), trading volume, price ranges, and the SPY ETF. The purpose is to derive insights into how these variables correlate with the price movements of QQQ, while also ensuring the stationarity of our time series data.
Initially, one straightforward approach to building G(e,d) could involve looking for highly correlated variables (correlation coefficient > |0.6|). This leads to a visually appealing network graph where all variables appear interconnected. However, it is crucial to remember that correlation does not imply causation, as revealed in our initial tests.
Advancing from basic correlation, we can utilize pairwise Granger causality to construct a directed network graph. This technique highlights the directional influence of one variable over another based on their temporal lags. Here, each edge in G(e,d) reflects whether the lags of the independent variable significantly contribute to the variation in the dependent variable, beyond its own lags.
By employing a restricted and unrestricted model comparison with a 1% p-value threshold, our bivariate Granger causal network emerges. Notably, despite a seemingly weak correlation between RSI and QQQ log returns, the Granger causality indicates that RSI significantly impacts QQQ with a 99% confidence level.
More interestingly, QQQ log returns are influenced by indirect relationships, such as volume affecting range, which in turn impacts QQQ log returns. This highlights the importance of recognizing intermediate feedback paths that exist within the networks.
However, measuring Granger causality in isolation often fails to account for the complex interdependencies present in the dataset. While network graphs visually represent multiple feedback paths, they do not provide comprehensive insights into how aggregate causality splits between direct and intermediate effects.
To overcome such limitations, some might propose incorporating all relevant variables into a single VAR model. Though this approach can effectively capture aggregate Granger causality, it still struggles to delineate direct influences from intermediate feedback in isolation.
The challenge remains to balance this complexity while adhering to accurate interpretations of structural causality. Structural causality must consider all feedback loops and produce insights that accurately reflect the data's underlying dynamics, distinguishing between direct and indirect effects.
The insights gleaned from causality network graphs represent an underutilized resource in econometric analysis. As we investigate the topology of these graphs, new avenues for understanding datasets emerge. Specifically, employing a conditional Granger causality graph methodology can facilitate a more granular analysis of feedback mechanisms.
Transitioning the focus from frequency domain analysis to adjacency matrices could enable a deeper comprehension of variable relationships. The advancement of this methodology possesses the potential to refine SVAR modeling, where variable selection substantially relies on temporal correlations.
Ultimately, this exploration of Granger causal networks underscores the significant value they bring to modeling complex economic systems. As researchers and analysts, recognizing and leveraging these insights can lead to more robust econometric applications.
For those intrigued by the concepts discussed here, the author encourages ongoing discourse in this field. Insights drawn from Granger causal networks can sparking new thoughts and methodologies in time series analysis.
About the author
Vedant Bedi is an Analyst at Mastercard, engaged with the NAM portfolio development team. He holds a Bachelor’s degree in Mathematics and Economics from NYU (magna cum laude, Class of 2019) and is passionate about data science, econometrics, and their applications in finance. As a member of the Phi Beta Kappa (NYC chapter), the oldest academic honors society in the U.S., his commitment to academic rigor remains steadfast.
Share your thoughts or insights on the exciting field of time series analysis using network graphs and Granger causality.