📊 statistics

Separation-based causal discovery for extremes

The paper introduces Extreme Structural Causal Models (XSCMs), a new framework that uses transformed-linear algebra to model causal relationships among extreme values, enabling the use of separation-based discovery algorithms to identify underlying causal structures in heavy-tailed and tail-dependent data.

Original authors: Junshu Jiang, Jordan Richards, Raphaël Huser, David Bolin

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Junshu Jiang, Jordan Richards, Raphaël Huser, David Bolin

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 understand how a massive, chaotic storm system works. Most scientists use standard tools to study "normal" weather—the gentle breeze or the light rain. But when a super-hurricane hits, the rules of the game change. The wind doesn't just blow; it rips roofs off. The rain doesn't just fall; it floods entire cities.

In the world of data, we call these "extreme events." This paper, "Separation-based causal discovery for extremes," is essentially a new set of high-tech goggles designed specifically to see the "hidden wiring" of a storm, rather than just the calm weather.

Here is the breakdown of how it works using three simple analogies.

1. The Problem: The "Fair-Weather" Map

Imagine you have a map of a city showing how traffic flows. On a sunny Tuesday, traffic flows smoothly from the suburbs to the downtown area. But during a massive blizzard, the rules change: everyone stays home, or everyone rushes to the grocery store at once. If you use your "sunny day" map to predict the "blizzard" traffic, you will fail miserably.

Most current mathematical models (called SCMs) are "fair-weather" models. They are great at telling you how things relate when life is normal. But when things get extreme—like a stock market crash or a massive flood—the connections between variables change. The paper argues that we need a new kind of map, which they call an XSCM, built specifically for the "blizzard" moments.

2. The Solution: The "XSCM" (The Extreme Detective)

The authors created the XSCM. Think of this as a detective who doesn't just look at who is standing next to whom, but looks at who is pulling the strings during a crisis.

In a normal situation, two people might be standing near each other just by coincidence. But in an extreme situation (like a sudden panic in a crowded theater), if Person A moves, Person B moves instantly. The XSCM uses a special kind of "extreme math" (called transformed-linear algebra) to figure out if Person A is actually causing Person B to react, or if they are both just reacting to a third person.

3. The Method: "The Chain Reaction Test"

How do they actually find these connections? They use a method called "Separation-based discovery."

Imagine a long chain of dominoes. If you want to know if Domino 1 is connected to Domino 5, you don't have to look at them directly. You just look at the dominoes in the middle. If you "remove" (or account for) Dominoes 2, 3, and 4, and Domino 1 and 5 no longer have any influence on each other, then you know they weren't directly connected.

The researchers developed a way to do this "domino test" using Tail Dependence. Instead of looking at the average height of the dominoes, they only look at the moments when the dominoes are falling with massive, violent force. This allows them to build a "map of causes" that only matters when things go wrong.

Real-World Proof: Rivers and Money

To prove their "extreme goggles" actually work, they tested them on two things:

The Danube River: They looked at how water flows through a massive river system. Their model was able to correctly identify which parts of the river feed into others, even during extreme flooding, successfully mapping the "path of the water" through the landscape.
The Chinese Stock Market: They looked at massive, sudden spikes in trading activity. Their model was able to group different types of investments (like gold, oil, or bonds) and show how a "panic" in one category triggers a "panic" in another. It showed that during a crash, the "wiring" of the market changes, and money flows in specific, predictable directions.

Summary in one sentence:

While most math models study how things behave on a sunny day, this paper provides a new way to map out the "cause-and-effect" connections that only appear when a storm hits.

This paper, titled "Separation-based causal discovery for extremes," addresses the challenge of identifying causal relationships in complex systems during extreme events. Traditional Structural Causal Models (SCMs) often fail in these scenarios because the governing mechanisms of a system can change dramatically when variables enter their "tails" (extreme values).

Here is a detailed technical summary of the research:

1. The Problem: Causal Discovery in the Tails

Standard causal discovery methods (like those used in Gaussian settings) focus on average dependencies. However, in fields like finance, hydrology, and ecology, variables that appear uncorrelated during normal conditions often exhibit strong dependencies during extreme events (e.g., market crashes or massive floods).

Existing research in "extremal causality" has been limited by two main issues:

Metric Insufficiency: Previous methods relied on the "causal tail coefficient," which is inherently bivariate. This makes it difficult to establish the Causal Markov Property or Causal Faithfulness, which are necessary to connect graph separation (d-separation) to statistical independence.
Model Limitations: Most existing models struggle to handle the transition from undirected associations (extremal dependence) to directed causal influences (extremal causation) in a unified framework.

2. Methodology: The XSCM Framework

The authors propose a new class of models called Extremal Structural Causal Models (XSCMs).

A. Mathematical Foundation

The XSCM is built upon transformed-linear algebra for regularly varying random variables. Instead of standard addition and multiplication, the model uses specialized operators ( $\oplus$ and $\circ$ ) that preserve the "heavy-tailed" nature of the data.

Structural Equations: An XSCM defines each variable $X_i$ as a transformed-linear combination of its causal parents and an independent source of extreme noise $Z_i$ :
$X_i = (\alpha_i \circ Z_i) \oplus \sum_{j \in Pa(X_i)} (\beta_{j \to i} \circ X_j)$
Tail Dependence: The model assumes multivariate regular variation, specifically focusing on the case where the tail index $\alpha = 2$ .

B. The Partial Tail-Correlation Coefficient (PTCC)

To enable graph discovery, the authors utilize the PTCC ( $\gamma_{ij|S}$ ). This metric quantifies the extremal dependence between two variables after adjusting for the effects of a set of conditioning variables $S$ . This is the extremal analogue of partial correlation in Gaussian models.

C. Bridging Graph Theory and Statistics

The core technical innovation is the proof of the Tail Causal Markov and Tail Causal Faithfulness properties. The authors prove that:

$v_i$ and $v_j$ are d-separated in the DAG if and only if they are partially tail uncorrelated ( $\gamma_{ij|S} = 0$ ).

This equivalence allows the authors to adapt state-of-the-art constraint-based algorithms (like PCMCI+) directly to extreme data.

3. Key Contributions

New Model Class (XSCM): A mathematically rigorous SCM designed specifically for heavy-tailed, tail-dependent data.
Theoretical Guarantees: Proving that XSCMs satisfy the Markov and Faithfulness conditions relative to partial tail uncorrelation.
Unified Framework: The methodology works for both Directed Acyclic Graphs (DAGs) (for causal direction) and Undirected Graphs (for extremal Markov networks).
Algorithmic Adaptability: Enabling the use of existing separation-based discovery algorithms (like PCMCI+) by simply replacing the standard independence test with a PTCC-based test.

4. Results and Validation

The authors validated their method through three simulation scenarios and two real-world applications:

Simulations:
- Cross-sectional DAGs: The method accurately recovered graph skeletons and directions, with performance improving as sample size increased.
- Undirected Graphs: The method outperformed existing methods (e.g., Gong et al., 2024) by using smaller, more efficient separation sets rather than conditioning on all other variables.
- Time Series: The method proved robust in identifying contemporaneous and lagged causal effects, significantly outperforming previous time-series extremal models when contemporaneous effects were present.
Real-World Applications:
- Danube River Basin: The model successfully recovered the physical topology of the river (upstream to downstream) and identified "collider" structures and lagged effects between gauging stations.
- China’s Derivatives Market: The model identified causal clusters in high-frequency trading data. It revealed that extreme trading activity flows from high-turnover assets to lower-turnover ones and correctly clustered assets by their industrial categories.

5. Significance

This paper provides a vital bridge between Extreme Value Theory (EVT) and Causal Inference. By providing a framework that can distinguish between mere association and true causal direction during rare events, it offers powerful new tools for risk management in finance, disaster prediction in hydrology, and understanding systemic shocks in complex networks.