📊 statistics

Parameter Estimation and Seasonal Modification of the Fractional Poisson Process with Application to Vorticity Extremes over the North Atlantic

This paper introduces a new quantile-based parameter estimation method and a distance-weighted seasonal modification for the fractional Poisson process, demonstrating their effectiveness through simulations and an application to modeling relative vorticity extremes in the North Atlantic-European region.

Original authors: Merle Mendel, Roland Fried

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Merle Mendel, Roland Fried

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 the next big storm will hit. In the old way of thinking (the "Standard Poisson Process"), storms were like buses arriving at a stop on a strict schedule. If a bus came, the next one was equally likely to arrive in 1 minute or 1 hour, with no memory of the past. It was random, but very "even."

However, in the real world—especially with weather—storms often come in clusters. One big storm hits, and then another follows quickly, followed by a third. It's like a bus that, once it arrives, suddenly brings a whole fleet of buses right behind it.

This paper introduces a better way to model these "storm clusters" and offers a new tool to measure them. Here is the breakdown in simple terms:

1. The New Tool: The "Fractional Poisson Process"

The authors use a mathematical model called the Fractional Poisson Process (FPP).

The Old Way: Assumes storms arrive with "exponential" timing (like a fair coin flip every second).
The New Way: Uses something called the Mittag-Leffler distribution. Think of this as a "heavy-tailed" distribution. In plain English, it acknowledges that while long gaps between storms happen, "clumps" of storms are much more common than the old model predicts. It's like realizing that after a bus arrives, there's a high chance three more will show up immediately.

2. The Problem: How to Measure the Clumps?

To use this new model, you need to know two numbers (parameters) that describe the weather:

The "Tail" (β): How "clumpy" is the weather? Is it one storm every few days, or a week of storms followed by a month of calm?
The "Scale" (σ): How long, on average, are the gaps between these clumps?

The paper argues that the old ways of calculating these numbers are either too slow (like trying to solve a Rubik's cube by hand) or not accurate enough.

3. The Solution: A New "Ruler" (Quantile-Based Estimation)

The authors propose a new method to find these two numbers.

The Analogy: Imagine you have a bag of marbles of different sizes (your storm data). You want to guess the average size and the spread.
- Old Method 1 (Log-Moments): You weigh the marbles. It's fast, but if you have one giant boulder in the bag, it throws off your average.
- Old Method 2 (Maximum Likelihood): You try to fit a perfect mold to every single marble. It's incredibly accurate but takes forever to do.
- The New Method (Quantile-Based): Instead of weighing everything or fitting a perfect mold, you pick a few specific marbles (quantiles) from the bag—say, the 10th smallest, the 50th, and the 90th. You measure the distance between where these marbles should be in a perfect bag and where they actually are. You then adjust your numbers until the distance is as small as possible.

Why is this cool? It's a "Goldilocks" solution. It's almost as accurate as the slow, perfect mold method, but it runs as fast as the simple weighing method. The authors tested this with thousands of computer simulations and found it works very well, especially for "heavy-tailed" data (weather that clumps).

4. Adding the "Season" Factor

Weather isn't the same all year round. Winter storms are different from summer thunderstorms.

The Old Approach: Some researchers only looked at winter data because that's when the "clumping" was obvious. They treated the whole year as if it were just winter.
The New Approach: The authors made the model seasonal. They let the "clumpiness" and "gap size" change every single day of the year.
How they did it: They used a "sliding window" with a special weighting system (like a spotlight). When calculating the weather pattern for July 15th, they looked at data from June and August, but gave July 15th the brightest spotlight. As they moved further away in time, the spotlight got dimmer. This allowed them to see how the weather behavior shifts smoothly from winter to summer and back.

5. The Real-World Test: Storms over the North Atlantic

The authors applied their new model to real data: Relative Vorticity Extremes (a fancy way of saying "strong spinning winds" that indicate cyclones) over the North Atlantic and Europe.

What they found:

Winter, Spring, and Fall: The storms definitely "clump." The new model showed that when one storm hits, another is very likely to follow soon. The "ordinary" model (the old bus schedule) failed to predict this.
Summer: The weather behaves more like the old, "even" model. The storms are more random and less clustered.
The Result: By using their new seasonal model, they could predict the probability of a second storm hitting within 72 hours much better than the old models. In the spring, the old model said, "Don't worry, the next storm is far away." The new model said, "Actually, there's a good chance another one is coming tomorrow."

Summary

This paper is about building a better weather prediction tool.

They created a new, fast, and accurate math trick to measure how much weather events "clump" together.
They updated the model to change with the seasons, realizing that winter storms clump together while summer storms are more random.
They proved that for the North Atlantic, ignoring the seasons and the clumping leads to bad predictions, but their new method captures the reality of stormy weather much better.

It's essentially upgrading the weather forecast from a "random guess" to a "smart pattern recognizer" that knows when to expect a stormy week versus a calm one.

Technical Summary: Parameter Estimation and Seasonal Modification of the Fractional Poisson Process

Problem Statement
The standard Poisson process assumes exponentially distributed return times between events, implying independent occurrences. However, many real-world phenomena, particularly in meteorology (e.g., extratropical cyclones, solar flares), exhibit temporal clustering and heavy-tailed return times that the standard model cannot capture. The Fractional Poisson Process (FPP) addresses this by replacing exponential return times with Mittag-Leffler distributed ones, governed by a tail parameter $\beta \in (0, 1]$ and a scale parameter $\sigma > 0$ . While the FPP offers a more flexible framework, existing parameter estimation methods for the Mittag-Leffler distribution face trade-offs between computational efficiency, accuracy (Mean Squared Error), and robustness. Furthermore, real-world applications often involve strong seasonal patterns (e.g., winter storm clustering) that static FPP models with constant parameters fail to address adequately.

Methodology
The paper proposes two primary methodological advancements:

New Quantile-Based (QB) Estimation:
The authors introduce a new estimator for Mittag-Leffler parameters that minimizes the sum of squared distances between the empirical and theoretical Cumulative Distribution Functions (CDF) at selected quantiles. This approach combines the Cramér-von Mises (CM) distance minimization with the Quantile Least Squares (QLS) concept.
- The objective function is defined as: $\sum_{i=1}^r (\alpha_i - F_{\beta,\sigma}(\hat{q}_{\alpha_i}))^2$ .
- Through simulation, the authors determine that using $r=5$ quantiles (specifically $\alpha = \{0.1, 0.3, 0.5, 0.8, 0.925\}$ ) offers an optimal balance between estimation efficiency and computational cost.
- Theoretical properties, including consistency and asymptotic normality, are established under specific conditions regarding the chosen quantiles.
Seasonal Modification via Distance-Based Weighting:
To handle seasonality, the authors extend the homogeneous FPP to allow parameters $\beta$ and $\sigma$ to vary periodically over the calendar year.
- Parameters for a specific day $t_0$ are estimated using a weighted version of the chosen estimator (Log-Moment, Maximum Likelihood, QLS, or the new QB).
- Weights are assigned using an Epanechnikov kernel based on the temporal distance between the observation day and $t_0$ , accounting for the cyclical nature of the year.
- This non-parametric approach imposes minimal assumptions on the form of seasonality.

Simulation Study and Comparative Results
An extensive simulation study ($324$ settings across varying $\beta$ , $\sigma$ , and sample sizes $n$ ) evaluates the new QB estimator against Log-Moment (LM), Maximum Likelihood (ML), CM, and QLS estimators.

Accuracy (MSE): The ML estimator generally yields the lowest MSE. However, the QB estimator demonstrates high relative efficiency compared to ML (median relative efficiency of $0.76$ for $\beta$ and $0.86$ for $\sigma$ ). Both QB and CM significantly outperform the widely used LM estimator and the QLS estimator (which performs poorly for tail parameter estimation in heavy-tailed settings).
Computational Efficiency: The ML and CM estimators are computationally intensive, with times increasing substantially with sample size. In contrast, the QB estimator's computation time remains relatively stable and is significantly faster than ML (up to $2,500$ times faster) and CM (up to $8,000$ times faster).
Robustness: Sensitivity curve analysis indicates that QB, CM, and QLS estimators are more robust to contamination than the LM estimator. The QB estimator maintains robustness while offering a favorable trade-off between speed and accuracy.

Application: Vorticity Extremes in the North Atlantic
The proposed seasonal FPP is applied to relative vorticity extremes (850 hPa level) in the North Atlantic-European region (1940–2023) using ERA5 reanalysis data. Extremes are identified via the peaks-over-threshold method.

Seasonal Patterns: The analysis reveals distinct seasonal variations. Clustering behavior (indicated by $\beta < 1$ ) is most pronounced in winter, spring, and fall. In contrast, summer exhibits behavior closer to an ordinary Poisson process ( $\beta \approx 1$ ), particularly over the western North Atlantic.
Spatial Variability: Stronger clustering is observed near the British Isles and the northwest French coast compared to the open ocean.
Model Performance: When comparing the seasonal FPP to a seasonal ordinary Poisson process, the FPP better captures the empirical 0.75 quantile of return times and the probability of additional extremes occurring within 72 hours. The ordinary Poisson process consistently underestimates the probability of clustering, especially during spring.

Significance and Claims
The paper claims that the proposed seasonal FPP framework provides a necessary extension for modeling extreme weather events that exhibit both heavy-tailed return times and seasonal variability. The primary contribution is the introduction of the Quantile-Based (QB) estimator, which the authors argue is a robust, computationally efficient alternative to Maximum Likelihood estimation, making it suitable for large-scale applications involving many datasets (such as gridded meteorological data).

The study demonstrates that ignoring seasonality or assuming a constant heavy-tailed distribution can lead to inaccurate risk assessments. By allowing parameters to vary daily, the model successfully captures the transition from clustered behavior in winter/spring to more random behavior in summer. The authors conclude that the FPP is superior to the ordinary Poisson process for modeling the temporal clustering of vorticity extremes, particularly when seasonal dynamics are accounted for.