📊 statistics

Granger Causality in Extremes

This paper introduces a rigorous, model-free mathematical framework for detecting Granger causality specifically in extreme events by leveraging the causal tail coefficient, which outperforms existing methods in handling non-linear, high-dimensional time series and hidden confounders while successfully identifying causal links in financial and weather data.

Original authors: Juraj Bodik, Olivier C. Pasche

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Juraj Bodik, Olivier C. Pasche

Original paper licensed under CC BY 4.0 (http://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

Imagine you are trying to figure out what causes what in a chaotic world. Usually, statisticians look at the "average" behavior of things. They ask, "On a typical Tuesday, does rain make the river rise?" This is like studying the weather on a calm, sunny day. It tells you a lot about normal life, but it misses the drama.

This paper, "Granger Causality in Extremes," argues that when things go crazy—when the stock market crashes, a hurricane hits, or a river floods—we need a different set of tools. The authors propose a new way to find the "root cause" of these extreme events, rather than just looking at what happens on average.

Here is the breakdown of their ideas using simple analogies:

1. The Problem: The "Average" Blind Spot

Imagine a forest. Most of the time, the trees sway gently in the breeze. If you study the forest by looking at the average wind speed, you might conclude that the wind doesn't really do much. But if a massive storm hits, the trees might snap, and the whole ecosystem changes.

Old Method: Looks at the gentle breeze (the "body" of the data). It might miss the fact that a specific type of wind only breaks trees when it gets to hurricane strength.
The Paper's Idea: We need to focus specifically on the "stormy days" (the extremes) to see what actually causes the damage.

2. The Solution: The "Tail" Detective

The authors introduce a concept called Granger Causality in Extremes. Think of it as a detective who only investigates crimes that happen during a full moon (the extreme events).

They use a tool called the Causal Tail Coefficient.

The Metaphor: Imagine you are watching two people, Alice and Bob.
- Normal Causality: You ask, "Does Alice usually make Bob happy?"
- Extreme Causality: You ask, "When Alice is extremely angry, does that guarantee Bob becomes extremely angry?"
- If the answer is "Yes, every single time," then Alice is the cause of Bob's extreme anger. The paper provides a mathematical way to measure this "guarantee."

3. The Superpower: Ignoring the "Noise"

One of the biggest headaches in finding causes is confounders. These are hidden third parties that mess up your logic.

The Analogy: Imagine you see that ice cream sales and shark attacks both go up in July. A bad detective might say, "Ice cream causes shark attacks!" But the real cause is the hidden confounder: Hot weather. Hot weather causes people to buy ice cream and go swimming (where sharks are).

Usually, to fix this, you need to measure everything (temperature, humidity, etc.). But in the real world, you often can't measure everything.

The Paper's Discovery: The authors found a magic trick. When you are looking at extreme events (like a massive heatwave), the "hidden confounder" often disappears from the equation.
Why? If Alice and Bob are both reacting to a hidden "Hot Weather" variable, but Alice's reaction is so extreme that it overwhelms the weather, you can tell Alice is the direct cause of Bob's extreme state without needing to measure the weather. The paper proves that for extreme events, you can often ignore the hidden noise and still find the true cause.

4. The Test: Does it Work?

The authors didn't just write theory; they built a computer program to test it.

The Simulation: They created fake worlds with complex rules, including some with "hidden spies" (confounders) and some with "heavy tails" (extreme outliers).
The Result: Their new method was faster and more accurate than the current "state-of-the-art" methods. It correctly identified the causes in extreme situations where other methods got confused or failed.

5. Real-World Examples

The paper tested their method on two real-life scenarios:

Swiss Rivers and Rain: They looked at precipitation (rain) and river discharge (water flow).
- The Finding: They could tell exactly which rivers would flood if it rained heavily at a specific mountain source. They found that rain in one spot caused extreme flooding downstream, but didn't necessarily cause extreme flooding in a different, distant valley. This helps predict floods more accurately.
Cryptocurrency: They looked at 14 different cryptocurrencies.
- The Finding: They identified Bitcoin and Litecoin as the "leaders." When Bitcoin had an extreme drop or spike, it almost always caused extreme movements in other coins (like Ethereum or Cardano) shortly after. This helps traders understand who is driving the market during a crash or a boom.

Summary

In short, this paper says: "Don't just study the calm days to understand the storm."

They created a new mathematical lens that zooms in specifically on the most extreme, volatile moments. This lens is better at ignoring hidden distractions (confounders) and is faster at finding the true "trigger" of a crisis, whether it's a financial crash or a natural disaster. They have even made their code open-source so others can use this "extreme detective" tool.

Technical Summary: Granger Causality in Extremes

1. Problem Statement

Causal discovery in time series is critical for understanding dynamic systems, particularly during periods of high volatility and extreme events. However, state-of-the-art methods predominantly focus on causality within the "body" of the distribution (e.g., causality in the mean or variance). These classical approaches often fail to capture mechanisms that manifest exclusively during extreme events, such as the impact of extreme stock returns on other assets or extreme precipitation on river floods. Furthermore, standard causal inference is highly sensitive to hidden confounders, and existing methods for extreme event analysis often rely on strong assumptions regarding regular variation, tail equivalence, or specific parametric models that limit their applicability to both heavy- and light-tailed variables.

The paper addresses the need for a framework that formally defines and infers Granger-type causality specifically within the context of extreme events, capable of handling non-linear, high-dimensional series, and robust to hidden confounding under specific tail conditions.

2. Methodology

2.1 Theoretical Framework: Causality in Extremes

The authors introduce two novel definitions of causality in extremes for a bivariate process $(X, Y)$ , conditioned on a set of variables $Z$ (which may include past values of $X$ and $Y$ and potential confounders). These definitions rely on the causal tail coefficient, a generalization of the coefficient introduced by Gnecco et al. (2020).

Let $F$ be a distribution function with $F(x) < 1$ for all $x$ . The causal tail coefficient adjusted for $Z$ is defined as:
$\Gamma_{X \to Y|C} := \lim_{v \to \infty} \mathbb{E}[F(Y_{t+1}) \mid X_t > v, C^{-X}_t]$
where $C^{-X}_t$ represents the information set excluding the past of $X$ .

The paper defines two forms of extreme causality:

Tail Causality ( $X \xrightarrow{tail} Y$ ): The upper tail of $X$ causes the tail of $Y$ if the conditional expectation above differs from the baseline expectation $\mathbb{E}[F(Y_{t+1}) \mid C^{-X}_t]$ .
Extreme Causality ( $X \xrightarrow{ext} Y$ ): An extreme in $X$ causes an extreme in $Y$ if the coefficient equals 1 almost surely. This implies that an extreme event in $X$ makes an extreme event in $Y$ certain (probability approaches 1).

The framework extends to non-unit lags (max-lag $p$ ), instantaneous effects, both tails (using $|X|$ and $|Y|$ ), and bounded supports.

2.2 Equivalence and Robustness

The authors establish formal equivalences between these new definitions and classical concepts:

Under Assumption A1 (structural dependence does not vanish in the tails), $X \xrightarrow{ext} Y \iff X \xrightarrow{tail} Y \iff X \xrightarrow{G} Y$ (Granger causality).
Robustness to Hidden Confounders: A key theoretical contribution is the demonstration that, under regular variation assumptions (specifically the Grey assumption) and stationarity, the adjustment for confounders $Z$ can be ignored when testing for extreme causality. Specifically, $\Gamma_{X \to Y|C} = 1 \iff \Gamma_{X \to Y|\emptyset} = 1$ . This allows the detection of extreme causal links without explicitly modeling high-dimensional confounders, provided the extreme behavior of $X$ is sufficiently driven by its own innovations.

2.3 Estimation and Inference

The paper proposes a non-parametric, model-free inference procedure:

Estimator: A sample-based estimator $\hat{\Gamma}_{X \to Y|Z}$ is constructed by averaging $F(y_{t+1})$ over a subset of time steps $S$ where $X_t$ is extreme (exceeding a threshold $\tau_n$ ) and confounders $Z_t$ (and potentially $Y_t$ ) are non-extreme.
Set Selection: Two sets are proposed for $S$ $S$ :
- $S_1$ : Conditions on $X_t$ being extreme and $Z_t$ (and $Y_t$ ) being below a fixed threshold.
- $S_2$ : Conditions on $X_t$ being extreme and $(Y_t, Z_t)$ falling within a small ball around a specific point $(y_0, z_0)$ .
Consistency: Theorem 2 and Theorem 3 prove that these estimators are consistent as sample size $n \to \infty$ .
Algorithm: A classification algorithm (Algorithm 1) compares the estimated coefficient against a baseline. If $\hat{\Gamma} \approx 1$ , it concludes extreme causality; otherwise, it concludes no extreme causality.
Multivariate Extension: Algorithm 2 extends this to estimating a full causal graph for $m$ variables. It leverages the robustness property to first perform pairwise screening (ignoring confounders) and then refine the graph by conditioning on parents, achieving a time complexity of $O(m^2 n \log n)$ .
Testing: A moving block bootstrap procedure is used to construct confidence intervals for tail causality tests, preserving temporal dependencies.

3. Key Results

3.1 Theoretical Contributions

Equivalence: The paper proves that under mild structural assumptions (A1/A2), causality in extremes is equivalent to classical Granger and structural causality.
Confounding Robustness: Theorem 1 establishes that for regularly varying processes satisfying the Grey assumption, the presence of hidden confounders does not alter the detection of extreme causality ( $\Gamma=1$ ) if the direct pathway dominates the confounding pathway in the tails.
No-Free-Lunch Theorem: The authors demonstrate that without restricting the model class (e.g., to structural causal models), no valid-level test exists for Granger non-causality or non-causality in extremes, highlighting the necessity of the structural assumptions used.

3.2 Empirical Performance

Simulations: Comparative studies against state-of-the-art methods (PCMCI with RobustParCorr and GPDC) across VAR and GARCH models (both Gaussian and heavy-tailed) show that the proposed method:
- Achieves significantly lower average model error in almost all settings, particularly in high-dimensional and heavy-tailed scenarios.
- Is computationally more efficient (e.g., ~6 seconds vs. >1 hour for PCMCI-GPDC on a 20-variable dataset).
- Demonstrates robustness to hidden confounders where other methods fail.
Real-World Applications:
- Hydro-meteorology: Applied to Swiss river discharge and precipitation data. The method successfully identified that extreme precipitation at a source station causes extreme discharge in downstream stations (extreme causality), while showing only tail causality for stations where the causal link is probabilistic or blocked by geography.
- Cryptocurrency: Analyzed 14 high-frequency cryptocurrency returns. The framework identified Bitcoin and Litecoin as primary drivers of extreme market movements, uncovering coherent causal structures that align with market knowledge (e.g., Bitcoin as a leading indicator).

4. Significance and Claims

The paper claims to provide the first formal framework for Granger causality in extremes that bridges the gap between predictive causality and extreme event analysis. Its significance lies in:

Handling Extremes: Offering a tool specifically designed for high-impact, low-probability events where classical mean-based methods are insufficient.
Robustness: Providing a theoretical guarantee that extreme causal links can be detected even in the presence of hidden confounders, a major challenge in observational causal inference.
Practical Utility: Delivering a fast, non-parametric, and open-source implementation that outperforms existing state-of-the-art methods in accuracy and speed, applicable to fields ranging from finance to climate science.

The authors note that while the framework is powerful, it relies on specific tail assumptions (regular variation, Grey assumption) and structural properties. They acknowledge open questions regarding the extension of these robustness results to Kesten-Goldie assumptions and the development of faster, theoretically justified alternatives to the bootstrap testing procedure.