📊 statistics

Regression modeling of multivariate precipitation extremes under regular variation

To address the EVA2025 data challenge, this paper proposes a computationally efficient two-stage regression strategy based on regular variation to estimate the occurrence and intensity of multivariate precipitation extremes using CESM2 Large Ensemble data.

Original authors: Rishikesh Yadav, Arnab Hazra

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rishikesh Yadav, Arnab Hazra

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

The "Extreme Weather Detective" Approach: Predicting the Unpredictable

Imagine you are a weather detective trying to predict how often a "once-in-a-century" flood might happen. The problem is, you can’t wait 100 years to see if you’re right! You only have a few decades of data, and the most extreme events—the massive, city-flooding storms—are so rare that they barely show up in your records at all.

This paper describes a clever mathematical strategy used by a team (the "DesiBoys") to solve this exact problem. They participated in a global data challenge (EVA2025) to predict extreme rainfall using complex climate models.

Here is how they did it, broken down into simple ideas:

1. The Problem: The "Ghost" in the Data

In statistics, extreme events are like ghosts. You know they exist, and you can see their faint footprints (small rainstorms), but the actual "ghost" (the massive flood) is rarely seen.

Traditional math models are like a flashlight that only works for things happening right in front of you. If you try to use them to look into the dark distance where the "ghosts" live, the light fades out, and the math breaks.

2. The Solution: The "Connect-the-Dots" Strategy

Instead of trying to guess what the ghost looks like in the dark, the researchers used a method called Regular Variation.

Think of it like this: Imagine you are watching a child grow. You can easily measure their height at age 5, 6, and 7. You notice a pattern: every year, they grow about 2 inches. Even though you haven't seen them at age 30, you can use that "growth pattern" to make a very educated guess about how tall they will be.

The researchers did the same thing with rain:

Step A (The Footprints): They looked at "sub-asymptotic" levels—basically, moderate rainstorms that happen often enough to measure accurately.
Step B (The Pattern): They used a mathematical rule (Regular Variation) that says, "If the small storms follow this specific pattern, the giant storms MUST follow a related pattern."
Step C (The Extrapolation): They drew a line (a regression model) from the moderate storms out into the "darkness" to predict the intensity of the massive ones.

3. The "Safety Net": The Block Bootstrap

When you make a guess about the future, you need to know how much you can trust it. If you guess a person will be 6 feet tall, you might be 90% sure, or only 50% sure.

To figure out their "certainty," they used a technique called Block Bootstrapping.
Imagine you have a deck of cards representing different years of weather. Instead of just shuffling them randomly, you grab "blocks" of cards (like a whole month or a whole year) and reshuffle those blocks. This keeps the "story" of the weather intact (because weather in January is related to weather in February) while allowing them to run thousands of "what-if" scenarios. This tells them: "We think the flood will be this big, and we are 95% sure it will fall within this range."

4. Why does this matter?

The team didn't just win a trophy (they took 2nd place!); they proved that you don't need massive, super-expensive supercomputers to make great predictions.

By using this "two-stage" approach—measuring the small stuff to predict the big stuff—they created a way to help cities, farmers, and disaster planners prepare for the "ghosts" of climate change before they actually arrive.

In short: They used the predictable patterns of "normal" heavy rain to build a mathematical bridge into the world of "extreme" disasters, allowing them to see into the future of our changing climate.

Technical Summary: Regression Modeling of Multivariate Precipitation Extremes under Regular Variation

1. Problem Statement

The paper addresses the challenge of estimating the occurrence and intensity of extreme precipitation events using large-scale ensemble climate data. Specifically, it focuses on the EVA2025 Data Challenge, which utilized the CESM2 Large Ensemble Community Project (LENS2).

The core difficulty lies in the statistical nature of extremes: they are heavy-tailed, spatially and temporally dependent, and sparse. Traditional Gaussian-based regression models fail to capture the tail behavior of such data. The objective was to predict three specific "target quantities" (expected counts of extreme events) at very high thresholds that exceed the range of the observed data, while providing 95% confidence intervals.

2. Methodology

The authors propose a two-stage regression strategy grounded in the mathematical framework of multivariate regular variation.

Stage 1: Empirical Estimation at Sub-asymptotic Levels:
Instead of attempting to fit complex multivariate extreme value distributions (like max-stable processes) directly to the highest extremes, the authors first calculate empirical counts of the target events at "moderate" or sub-asymptotic thresholds ( $u_l$ ). This ensures a sufficient sample size for stable empirical estimation.
Stage 2: Regular Variation-Based Regression:
The authors leverage the theoretical property that for a multivariate regularly varying vector, the joint exceedance probability follows a power law: $P(Y > u) \approx C u^{-\alpha}$ .
They transform this into a linear regression framework:
$\log(\text{Expected Counts}) = C' - \alpha \log(u)$
By regressing the empirical counts (from Stage 1) against the log-transformed thresholds, they estimate the scaling constant ( $C$ ) and the shape parameter ( $\alpha$ ). These parameters are then used to extrapolate and predict the expected counts at the much higher thresholds required by the competition.
Uncertainty Quantification:
To provide confidence intervals, the authors employ a nonparametric block bootstrap procedure. They treat "years" as blocks to preserve any potential temporal dependence within a year while respecting the observed independence between different years and climate model runs.

3. Key Contributions

Bridging Theory and Practice: The paper provides a method that connects the theoretical precision of regular variation (asymptotic scaling laws) with the scalability and interpretability of regression modeling.
Computational Efficiency: Unlike likelihood-based multivariate extreme value models or max-stable processes, which are computationally intensive for high-dimensional datasets, this two-stage approach is "computationally straightforward" and highly scalable.
Robustness to Climate Ensemble Data: The framework successfully handles the specific characteristics of the LENS2 dataset, such as the lack of long-term trends in precipitation intensity and the independence across different model runs.

4. Results

The methodology was validated through the EVA2025 competition, where the team ("DesiBoys") achieved joint second place in the final rankings.

Point Estimation: The model performed exceptionally well in predicting the frequency of extreme events, notably securing 1st place in Task 3 (predicting consecutive-day extremes).
Regression Fit: The authors demonstrated high goodness-of-fit ( $R^2$ values up to 0.98) when regressing log-counts against log-thresholds, confirming the validity of the regular variation assumption.
Interval Estimation: While the point estimates were highly accurate, the confidence intervals (via block bootstrap) were less competitive (ranking 5th–7th). The authors attribute this to the inherent difficulty of using non-parametric methods to capture the uncertainty of highly skewed extreme distributions.

5. Significance

This work is significant for climate risk assessment and hydrometeorology. It demonstrates that complex, high-dimensional multivariate extreme data can be modeled effectively without the prohibitive computational cost of traditional extreme value models. By providing a reproducible framework that balances theoretical rigor with practical feasibility, the paper offers a tool for researchers to extrapolate the likelihood and severity of rare, high-impact weather events (like floods or landslides) under varying climatic conditions.