📊 statistics

A new mixture model for spatiotemporal exceedances with flexible tail dependence

This paper proposes a novel mixture model and simulation-based inference framework using random forests to effectively estimate spatiotemporal streamflow exceedances by flexibly capturing both asymptotic dependence and independence in extreme event tails.

Original authors: Ryan Li, Emily C. Hector, Brian J. Reich, Reetam Majumder

Published 2026-02-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ryan Li, Emily C. Hector, Brian J. Reich, Reetam Majumder

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 predict when a river will overflow its banks and cause a flood. You have data from many different towns along the river, recorded every single day for 60 years. The problem is that floods are rare, weird, and complicated. Sometimes, if one town floods, a town 50 miles away floods too. Other times, they flood completely independently. Sometimes, if it floods today, it floods tomorrow; other times, the river calms down immediately.

Traditional math models for these events are like trying to fit a square peg into a round hole. They usually assume that if a flood happens, the "worst-case scenario" is always linked to other worst-case scenarios (like a domino effect that never stops). But in the real world, extreme events often lose their connection as they get more extreme or as you look at different places and times.

This paper introduces a new, flexible "mixture" model that acts like a smart chef's kitchen rather than a rigid factory line. Here is how it works, broken down into simple concepts:

1. The Four "Flavors" of Floods (The Mixture)

Instead of forcing the river to behave in just one way, the authors created a model that mixes four different "flavors" of behavior, like a smoothie made of different fruits:

Flavor A (The Global Storm): If this flavor is dominant, a flood in one town means a flood in every town, on every day. Everything is linked.
Flavor B (The Local Storm): If this is dominant, a flood in one town means a flood in nearby towns, but only today. It doesn't last long or spread far.
Flavor C (The Long-Lasting River): If this is dominant, a flood today means a flood tomorrow, but it doesn't necessarily affect towns far away.
Flavor D (The Independent Splash): If this is dominant, a flood in one town is a total fluke. It has nothing to do with other towns or other days.

The model's job is to figure out how much of each flavor is in the mix. Is the river mostly "Global Storm" or mostly "Independent Splash"? The answer changes depending on the data.

2. The "Censoring" Trick (Ignoring the Boring Stuff)

Floods are rare. 99% of the time, the river is just flowing normally. If you try to teach a computer to predict floods using all the data (including the boring, calm days), the computer gets confused.

The authors use a trick called censoring. Imagine you are a detective looking for a specific criminal. You don't interview every person in the city; you only interview people who match a specific description (e.g., "wearing a red hat").

In this model, they ignore all the days where the water level is low.
They only look at the "extreme" days (the red hats).
This makes the math much cleaner and focuses the model strictly on the dangerous events.

3. The "Black Box" Problem (Why Computers Get Stuck)

Usually, to find the right "flavor mix," statisticians use a complex formula (likelihood) to calculate the odds. But for this specific type of river model, that formula is so complicated that even supercomputers can't solve it without taking years. It's like trying to solve a Rubik's cube that has 1,000 sides.

4. The Solution: The "Forest of Guessers" (Random Forests)

Since they can't solve the math equation directly, they use a clever workaround called Simulation-Based Inference.

Step 1: The Practice Run. They tell the computer to "pretend" to be the river. They generate thousands of fake river datasets using different random "flavor mixes."
Step 2: The Summary. For each fake dataset, they calculate a few simple summary numbers (like "how often did Town A and Town B flood together?").
Step 3: The Teacher. They use a machine learning tool called a Random Forest. Think of this as a classroom of 100 different experts. Each expert looks at the summary numbers and guesses, "Based on this pattern, the river must be 40% Flavor A and 60% Flavor D."
Step 4: The Real Deal. Once the classroom of experts is trained on the fake data, they take the real river data, calculate the summaries, and ask the experts: "What's the mix for this real river?"

The experts (the Random Forest) are much faster and cheaper to run than the old math formulas, and they are surprisingly accurate.

5. What Did They Find?

They tested this on real river data from the US (1965–2024).

The Result: The model found that for daily river flows in this region, the "Independent Splash" flavor is the strongest.
What that means: If a river floods in one town today, it doesn't guarantee a flood in a neighboring town tomorrow. The extreme events are mostly local and short-lived.
Why it matters: This helps engineers and city planners know exactly where to build levees and when to issue warnings. If the events were always linked (Global Storm), you'd need to prepare every town at once. Since they are mostly independent, you can be more targeted and efficient.

Summary

The authors built a flexible, multi-flavored model to understand river floods. Because the math was too hard to solve directly, they taught a computer "forest" of experts to recognize patterns in fake flood data, so it could accurately guess the behavior of real floods. This helps us predict disasters better and save money and lives.

1. Problem Statement

The paper addresses the challenge of modeling extreme streamflow events (flooding) at high temporal resolutions (e.g., daily maxima). Existing methods face several limitations:

Temporal Independence Assumption: Traditional Max-Stable Processes (MSPs), the standard for spatial extremes, assume temporal independence. This fails for daily data where a single flood event spans multiple days.
Rigid Tail Dependence: Most MSPs assume asymptotic dependence (the probability of joint extremes remains positive as thresholds increase). However, many environmental processes exhibit asymptotic independence (dependence vanishes as events become more extreme) or a mix of both. Selecting a model with the wrong tail structure leads to poor extrapolation.
Computational Intractability: Exact likelihood inference for spatial extremes models (especially with censoring for peaks-over-threshold) is computationally impossible for moderate-to-large numbers of sites ( $n$ ).
Limitations of Existing Mixtures: Previous mixture models (e.g., Huser & Wadsworth, 2019) handle spatial dependence flexibility but lack temporal dependence modeling or require manual selection of dependence structures.

2. Methodology

A. The Spatiotemporal Mixture Model

The authors propose a latent process model, $X(s, t)$ , defined as a convex combination of four distinct processes, each representing a different asymptotic dependence regime:
$X(s, t) = \lambda_1 R_{ST}(s, t) + \lambda_2 R_S(s, t) + \lambda_3 R_T(s, t) + \lambda_4 W(s, t)$
Where $\sum \lambda_k = 1$ and $\lambda_k \geq 0$ . The components are:

$R_{ST}$ (Space-Time Dependent): A Brown-Resnick process with spatiotemporal correlation, representing asymptotic dependence in both space and time.
$R_S$ (Space Dependent): A Brown-Resnick process dependent in space but independent in time.
$R_T$ (Time Dependent): A Brown-Resnick process dependent in time but independent in space.
$W$ (Independent): An inverted Brown-Resnick process, representing asymptotic independence in both space and time.

The observed streamflow $Y(s, t)$ is linked to the latent process via a transformation $Y = G(X)$ , where the marginal distribution of $Y$ is modeled using a Generalized Pareto Distribution (GPD) for values exceeding a threshold (Peaks-Over-Threshold approach).

B. Censoring Mechanism

To focus on extremes, the model employs a censoring mechanism. Observations below a threshold $Q_\tau$ are censored. This allows the model to estimate marginal and joint probabilities of extreme events using only the tail data, preventing non-extreme values from biasing the dependence structure estimation.

C. Simulation-Based Inference (SBI) with Random Forests

Since the likelihood function is intractable for the mixture model (especially with censoring), the authors use Simulation-Based Inference (SBI):

Data Generation: Simulate datasets from the model across a dense grid of plausible parameter values (specifically the mixture weights $\lambda$ and range parameters $\rho_S, \rho_T$ ).
Summary Statistics: Reduce the simulated data to summary statistics, primarily empirical $\chi$ -coefficients (measuring tail dependence) binned by spatial and temporal distances, and conditional correlations.
Machine Learning Surrogate: Train Random Forests (instead of the more common deep learning approaches in SBI) to map the summary statistics to the model parameters.
- Rationale: Random forests are computationally cheaper, require fewer tuning parameters, and offer high interpretability with minimal loss of accuracy compared to deep learning for this specific task.
Estimation: Apply the trained Random Forests to the summary statistics of the real observed data to estimate the parameters.

3. Key Contributions

Flexible Spatiotemporal Dependence: The model is the first to simultaneously allow for flexible asymptotic dependence/independence in both space and time within a single mixture framework. It can capture four distinct regimes: (Space-Time Dependent), (Space Only), (Time Only), and (Independent).
Novel Inference Framework: The paper introduces a Random Forest-based SBI approach for spatial extremes, moving away from composite likelihood approximations (which suffer from bias and inefficiency) and deep learning surrogates.
Theoretical Characterization: The authors provide a theorem characterizing the asymptotic dependence of the mixture model, showing that the dominant mixture weight ( $\lambda_j$ ) dictates the asymptotic regime (e.g., if $\lambda_1$ is dominant, the process is asymptotically dependent in space and time).
Practical Application: Demonstrated on real-world USGS streamflow data, proving the framework's feasibility for high-resolution environmental data.

4. Results

Simulation Study

Parameter Recovery: The model successfully recovered the true mixture weights ( $\lambda$ ) and range parameters ( $\rho_S, \rho_T$ ) across four different simulation settings (dominant space-time, space, time, and independence).
Regime Identification: The method consistently identified the correct asymptotic dependence regime based on the dominant $\lambda$ parameter.
Performance: While bias was low for $\lambda$ , range parameters showed some bias under strong dependence, but coverage probabilities for confidence intervals were generally near nominal levels (95%).

Real Data Analysis (USGS Streamflow)

Data: Daily streamflow from 30 stations in the Western/Midwestern US (1965–2024), restricted to May–July.
Findings:
- The estimated mixture weight for the independent component ( $\lambda_4$ ) was the largest (0.534), followed by the time-dependent component ( $\lambda_3$ ).
- Conclusion: Daily streamflow in this region is asymptotically independent in both space and time. This implies that as flood events become more extreme, the probability of simultaneous extreme flooding at distant locations or different days approaches zero.
- Range Parameters: Spatial dependence is limited to ~1/4 of the domain, and temporal dependence is limited to ~15 days.
- Model Fit: The empirical $\chi$ -statistics from the data closely matched the model's predictions, validating the fit.

5. Significance and Future Work

Engineering and Policy: By accurately modeling the tail dependence structure, this model improves the reliability of flood risk assessments at unobserved sites and times. This is critical for designing civil engineering solutions (e.g., levees, dams) and developing targeted flood response policies.
Methodological Shift: The use of Random Forests for SBI offers a robust, computationally efficient alternative to deep learning for parameter estimation in complex spatial models.
Future Directions:
- Improving simultaneous estimation of marginal and dependence parameters.
- Adapting the model to allow for prediction at unobserved sites (spatial interpolation), which is currently a limitation.
- Extending theoretical derivations to general cases where the mixture components are fully spatiotemporal processes.

In summary, this paper presents a robust, flexible framework for modeling extreme streamflow that overcomes the rigidity of traditional Max-Stable Processes and the computational hurdles of exact inference, providing a practical tool for understanding and mitigating flood risks.