📊 statistics

A Robust Bootstrap Test for Bidirectional Granger Causality via Unexplained Mutual Information

This paper proposes a robust bootstrap test for bidirectional Granger causality in Gaussian VAR models that utilizes unexplained mutual information, an intersection-union procedure, and Huber M-estimation to address gaps in finite-sample validity, causal directionality distinction, and outlier sensitivity.

Original authors: zouhaier dhifaoui

Published 2026-09-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: zouhaier dhifaoui

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of economics and finance, researchers are constantly trying to map the invisible threads that connect different moving parts of a system. When two things happen over time, like the price of a stock and the fear level of the market, a fundamental question arises: does one actually drive the other, or are they just moving together by chance? For decades, statisticians have used a specific method called Granger causality to answer this. The logic is simple but powerful: if knowing the past history of one series helps you predict the future of another better than you could without it, then the first series is said to "cause" the second in a predictive sense. This is not about deep philosophical cause and effect, but rather about the flow of information. In many real-world scenarios, this relationship is not a one-way street; the two series often push and pull on each other simultaneously, creating a complex loop of influence. However, proving that this two-way street exists, especially when the data is messy or contains extreme, unusual events, has remained a difficult challenge for standard statistical tools.

A new study by Zouhaier Dhifaoui addresses this challenge by refining how scientists test for these two-way relationships. The researcher focused on a common problem where standard tests fail: they often mistake a one-way influence for a two-way loop, or they break down completely when the data contains outliers—those rare, extreme spikes that happen frequently in financial markets. To solve this, the author developed a new testing procedure that is both more accurate at distinguishing the direction of influence and more resistant to being thrown off by extreme data points. The method relies on a concept called unexplained mutual information, which essentially measures how much uncertainty remains in one series after accounting for the other. If the uncertainty drops significantly when you add the history of the second series, it suggests a causal link. The study shows that this measure is mathematically equivalent to older, well-known statistical tools, but the author realized that simply looking at the numbers was not enough. Because the numbers always go up slightly just by chance, even when there is no real connection, a simple threshold does not work. Instead, the researcher built a simulation-based approach that repeatedly reshuffles the data to create a realistic picture of what random noise looks like, allowing for a much fairer comparison.

The core of the work involves a two-step process to ensure the findings are genuine. First, the test checks if there is any connection at all between the two series. If a connection is found, the second step is crucial: it must determine if the influence flows both ways or just one. The study demonstrates that many traditional methods fail here, often declaring a two-way relationship when only one direction is actually active. By using a specific combination of tests, the new method successfully isolates true bidirectional causality. Furthermore, the researcher introduced a "robust" version of the test designed to handle contaminated data. In the real world, data sets often contain errors or extreme outliers that can distort standard calculations. The robust version uses a different mathematical approach that limits the influence of these extreme points, ensuring the results remain reliable even when the data is imperfect.

To prove the method works, the author ran thousands of computer simulations. These experiments showed that the new test correctly identified when there was no connection, avoiding false alarms that plagued older methods. More importantly, when the simulations included extreme outliers, the standard test failed to detect real connections, while the robust version continued to perform well. The study also confirmed that the method works for systems with different levels of complexity, not just simple ones. Finally, the researcher applied the new tool to real-world financial data, looking at the daily relationship between the S&P 500 stock index and the VIX, a measure of market volatility. The analysis covered daily data from January to April 2026. The results revealed a clear, one-way influence: changes in stock returns helped predict changes in market volatility, but the reverse was not true. This finding supports the idea that when stock prices fall, it mechanically increases financial leverage and risk, driving up volatility, rather than volatility changes driving the stock prices down.

This research provides a clearer, more reliable way to understand how different economic forces interact. By fixing the flaws in how we detect two-way relationships and by making the tests resilient to messy data, the study offers a stronger foundation for analyzing complex systems. The work does not claim to solve all problems in economics, but it provides a sharper lens for seeing the true direction of influence in the data we already have. For researchers and analysts, this means they can now distinguish between a simple one-way push and a complex two-way dance with greater confidence, even when the data is noisy. The study concludes that while the new robust method is slightly less sensitive when data is perfectly clean, it becomes the superior choice whenever the real-world messiness of outliers is present, making it a vital tool for modern financial analysis.

Technical Summary: A Robust Bootstrap Test for Bidirectional Granger Causality via Unexplained Mutual Information

Problem Statement
The paper addresses three specific limitations in detecting bidirectional Granger causality within Gaussian Vector Autoregression (VAR) models:

Lack of Valid Finite-Sample Tests: While the natural information-theoretic version of the causality statistic (based on mutual information) exists, there is no valid finite-sample test for it.
Ambiguity in Joint Restrictions: Standard joint restriction tests (e.g., testing if all cross-lag coefficients are zero) cannot distinguish between unidirectional causality and genuine bidirectional causality. Rejecting the joint null only implies causality exists in at least one direction.
Sensitivity to Outliers: The underlying covariance estimators used in classical Gaussian VAR models are highly sensitive to extreme observations, a common issue in fields like finance.

Methodology
The proposed framework introduces a statistic called Unexplained Mutual Information (UMI) and couples it with a null-imposing moving-block bootstrap and an intersection-union procedure.

Unexplained Mutual Information (UMI):
- The author defines UMI as the mutual information between a VAR model's residuals and the observed variables.
- Mathematically, this is expressed as a log-determinant ratio of residual covariance matrices.
- Key Identity: The paper establishes that this UMI statistic coincides exactly with Geweke's (1982) measure of linear feedback. When scaled by the sample size, it is identical to the classical Gaussian Likelihood-Ratio (LR) statistic. Consequently, under the null hypothesis, it follows a $\chi^2$ distribution asymptotically.
- The Sampling Issue: The sample ratio of covariance determinants is bounded below by 1 regardless of whether the null hypothesis holds. Therefore, simply observing a value $>1$ is not a valid test; the statistic must be compared against its actual sampling distribution.
Null-Imposing Moving-Block Bootstrap:
- To handle finite-sample dependence and avoid relying solely on asymptotic $\chi^2$ approximations, the paper proposes a residual-based moving-block bootstrap.
- Mechanism: The bootstrap imposes the joint null hypothesis ( $H_0$ : no cross-series feedback) by construction. It estimates the restricted model (own-lags only), resamples the restricted residuals using overlapping blocks, and regenerates synthetic data recursively.
- This ensures the bootstrap distribution reflects the null hypothesis exactly, allowing for the calculation of valid $p$ -values for the joint test.
Intersection-Union Test (IUT) for Bidirectionality:
- To specifically test for bidirectional causality (rather than just any causality), the paper employs an intersection-union procedure.
- The null hypothesis of "no bidirectional causality" is rejected only if both directional null hypotheses ( $H_{2 \to 1}$ and $H_{1 \to 2}$ ) are individually rejected.
- This approach isolates genuine bidirectional causality, preventing the misinterpretation of unidirectional effects as bidirectional.
Robust Variant:
- To address outlier sensitivity, a robust version of the test is developed.
- Estimation: Ordinary Least Squares (OLS) is replaced by Huber's M-estimator (iteratively reweighted least squares).
- Covariance: The Gaussian sample covariance is replaced by a robust location-and-covariance estimator based on a multivariate $t$ -distribution (Lange et al., 1989).
- The bootstrap algorithm remains structurally identical, utilizing these robust fits at every step.

Key Results (Monte Carlo Evidence)

Size Control: The asymptotic LR test is shown to be moderately oversized in small samples (e.g., $N=50, 100$ ), whereas the bootstrap test controls the size closely at or below the nominal level.
Bidirectional Distinction: The joint test frequently rejects unidirectional alternatives (high false positive rate for bidirectionality), whereas the Intersection-Union (IU) test maintains correct size for unidirectional cases, successfully distinguishing them from bidirectional ones.
Higher-Order VARs: The procedure performs consistently across VAR(1), VAR(2), and VAR(3) designs, maintaining correct size and power.
Robustness: Under 20% outlier contamination, the standard Gaussian test suffers a collapse in power. At $N=100$ , the standard test's power drops from 83.3% to 10.0%, while the robust test falls from 46.7% to 16.7%, ending up more powerful than the standard test once outliers are present. At $N=50$ , the robust test shows a smaller relative decline (from 16.0% to 4.0%) compared to the standard test (from 34.0% to 10.0%), demonstrating greater stability in the presence of outliers, although its absolute power (4.0%) remains lower than the standard test (10.0%) in this specific small-sample contaminated scenario.

Application
The method is applied to the joint dynamics of S&P 500 returns and the VIX (market-implied volatility) using daily data from January to April 2026.

Findings: The robust bidirectional test rejects the null of no causality. However, the directional decomposition reveals significant causality running from returns to volatility (supporting the leverage effect hypothesis) but not vice versa.
Implication: In this specific sample, the volatility-feedback hypothesis is not supported, while the leverage effect is dominant.

Significance and Claims
The paper claims to close the gap between information-theoretic measures of causality and operational finite-sample testing. Its primary contributions are:

Theoretical Unification: Demonstrating that the Unexplained Mutual Information is a rescaling of the classical Gaussian LR statistic, providing a familiar $\chi^2$ benchmark while highlighting the need for resampling due to the statistic's bounded nature.
Methodological Rigor: Providing a valid finite-sample test for the joint hypothesis via a null-imposing bootstrap, which is superior to asymptotic approximations in small samples.
Logical Precision: Offering a practical Intersection-Union framework to rigorously test for bidirectional causality, a distinction often blurred in standard joint tests.
Practical Robustness: Delivering a robust variant that maintains statistical power in the presence of outliers, making the method applicable to financial time series where extreme observations are common.

The author notes that while the robust test is less efficient on clean data, it is the preferred choice when contamination is a realistic concern. Future work is suggested to extend the framework to non-Gaussian innovations and higher-dimensional systems.

Technical Summary: A Robust Bootstrap Test for Bidirectional Granger Causality via Unexplained Mutual Information

More like this