🔢 mathematics

Gerber–Shiu Analysis of a Dependent Portfolio Risk Model with a Constant Dividend Barrier and Extended-Exponential Premium Income

This paper develops a dependent portfolio risk model with a constant dividend barrier and extended-exponential premium income to derive Gerber-Shiu functions for mechanism-specific ruin probabilities, demonstrating through theoretical analysis and empirical application to French motor insurance that incorporating inter-claim duration dependence improves claim severity modeling and enhances ruin risk assessment.

Original authors: Halim Zeghdoudi

Published 2026-09-29

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Halim Zeghdoudi

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

Insurance is a system of shared risk, where many people pay small amounts into a common pool to cover the rare, large losses of a few. For the companies managing these pools, the fundamental question is one of survival: can the money coming in from premiums ever be enough to cover the money going out for claims, especially when those claims arrive at unpredictable times and in unpredictable sizes? If the outflow of cash ever outpaces the inflow for too long, the company faces ruin, a state where it can no longer pay its debts. To understand this delicate balance, researchers study risk models, which are mathematical descriptions of how money flows in and out of an insurance portfolio over time. These models help insurers set prices, decide how much capital to keep in reserve, and determine when it is safe to pay dividends to shareholders. However, real-world insurance is messy. Premiums do not always arrive on a strict schedule, and the size of a claim often depends on how long it has been since the last one occurred. A long quiet period might suggest a buildup of risk, while a sudden cluster of claims could indicate a changing environment. Understanding these hidden connections is crucial for predicting when a company might fail.

In a recent study, a researcher named Halim Zeghdoudi tackled these complexities by building a new kind of risk model that treats an entire insurance portfolio as a single, dynamic system. Instead of assuming that every claim is independent or that premiums arrive at a steady, predictable rate, this model allows for randomness in both the timing and the size of incoming money. It also introduces a specific way to link the size of a claim to the time that has passed since the previous one. The core idea is that the history of the portfolio matters; the length of the quiet period before a claim arrives changes the likelihood of that claim being small or large. To make sense of this, the study uses a tool called the Gerber–Shiu function, which acts like a detailed scorecard. It doesn't just tell you if a company will go bankrupt; it calculates the probability of that happening, how deep the hole will be, and, crucially, which type of event caused the collapse. This allows the researcher to separate the risk of a few massive, rare disasters from the risk of many smaller, frequent losses.

The study focuses on a specific scenario where an insurance company pays out dividends to its owners whenever its cash reserves grow above a certain fixed limit, known as a barrier. This strategy creates a tension: paying out dividends keeps shareholders happy in the short term but leaves the company with less money to absorb future shocks. The researcher wanted to see how this strategy interacts with the randomness of premiums and the dependence between claim timing and size. To do this, the model assumes that premium payments follow a flexible pattern that can vary more than a simple average would allow, and that claims fall into one of two distinct categories. One category represents smaller, more frequent claims, while the other represents larger, more severe events. The model determines which category a new claim belongs to based on how long it has been since the last claim. If a long time has passed, the model suggests the next claim is more likely to be one of the severe type.

To test how this system behaves, the researcher first ran a series of computer simulations using a set of standard numbers to represent a typical insurance portfolio. These experiments showed that the amount of money a company starts with is the most powerful factor in its survival. When the initial reserve is low, the chance of ruin is high, but as the reserve grows, the risk drops sharply. The simulations also confirmed that receiving premiums more frequently helps the company stay safe, while having claims arrive more often increases the danger. Interestingly, the study found that the shape of the premium distribution matters. Even when the average amount of money coming in stays the same, changing the pattern of how those payments arrive can slightly alter the risk of ruin. The analysis also explored the effect of the dividend barrier. A lower barrier, which triggers payouts sooner, leads to a higher risk of ruin because the company keeps less cash in reserve. A higher barrier acts as a safety buffer, reducing the risk of failure but delaying the return of money to shareholders.

The second part of the study moved from theory to reality by applying the model to a real dataset from a French private motor insurance portfolio. The researcher used records of premiums and claims from over 50,000 policies covering the years 2003 and 2004. After cleaning the data, the study focused on 606 specific instances where a claim occurred, looking at the time that had passed since the previous claim and the size of the new claim. The results were striking. The data clearly showed that claims were not all the same; they fell into two very different groups. One group consisted of a small number of claims that were extremely large, averaging about 12,573 euros, while the other group contained the vast majority of claims, which were much smaller, averaging about 1,539 euros. The model successfully captured the rule connecting these groups to time: claims that followed a longer waiting period were significantly more likely to belong to the severe group.

When the researcher compared this new model to older, simpler ways of looking at the same data, the results were nuanced. The new model, which accounts for the time between claims, performed better than a basic model that assumed all claims were the same. It also performed slightly better than a model that assumed there were two types of claims but ignored the time between them. The improvement over the simpler two-type model was modest, suggesting that while the time between claims does provide useful information, the main driver of the difference in claim sizes is simply that there are two distinct types of accidents happening. The study concluded that the proposed model offers a practical way to understand these risks. It confirms that the time between claims holds clues about the severity of the next one, and it provides a clear method for calculating how different strategies, like paying dividends early or late, affect the long-term safety of the insurance company. The findings suggest that insurers can use this kind of detailed analysis to make more informed decisions about how much capital to keep and when to share profits, balancing the need for safety with the desire for growth.

Technical Summary: Gerber–Shiu Analysis of a Dependent Portfolio Risk Model

Problem Statement
This paper investigates an aggregate insurance portfolio risk model that departs from the classical Cramér–Lundberg framework by incorporating three complex features simultaneously: stochastic premium income, dependence between claim severity and the preceding inter-claim duration, and a constant dividend barrier strategy. The primary objective is to analyze the probability of ruin and the associated discounted penalty functions (Gerber–Shiu functions) within this setting. Specifically, the study aims to distinguish the contributions of different claim-severity regimes to the overall insolvency risk, recognizing that claim sizes may not be independent of the time elapsed since the last claim.

Methodology
The model is constructed as follows:

Stochastic Premiums: Premium receipts arrive according to a Poisson process with intensity $\lambda$ . The premium amounts follow an extended-exponential distribution (introduced by Gómez et al., 2014), which is a mixture of an exponential and an Erlang-2 distribution. This phase-type structure is chosen to allow for greater variability in premium sizes than a standard exponential model while maintaining analytical tractability.
Dependent Claim Mechanism: Claims arrive via an independent Poisson process with intensity $\lambda_1$ . The severity of the $i$ -th claim depends on the inter-claim duration $V_i$ . A latent, claim-specific benchmark $M_i \sim \text{Exp}(\lambda_2)$ determines the regime: if $V_i > M_i$ , the claim belongs to regime 1; otherwise, it belongs to regime 2. This results in conditional probabilities where longer claim-free periods increase the likelihood of the first regime.
Dividend Barrier: A constant barrier $b$ is imposed. Any surplus exceeding $b$ is immediately paid out as dividends.
Gerber–Shiu Analysis: The author defines mechanism-specific Gerber–Shiu functions, $\Phi_i(u)$ , which represent the expected discounted penalty associated with ruin caused specifically by a claim from regime $i$ . The total penalty function is the sum $\Phi(u) = \Phi_1(u) + \Phi_2(u)$ .
Analytical Derivation:
- Integral equations are derived for $\Phi_1$ and $\Phi_2$ using first-event decomposition.
- The exponential–Erlang mixture representation of the premium distribution allows the premium-barrier operator to be decomposed into auxiliary functions satisfying first-order differential relations.
- Under the assumption that claim severities in both regimes are exponentially distributed and the penalty function is unity, the system is reduced to a finite-dimensional linear differential system.
- This leads to a matrix-exponential representation for the mechanism-specific discounted ruin probabilities, $\Psi_i(u)$ .

Key Contributions

Model Formulation: The paper formulates a unified risk model combining stochastic premiums, duration-dependent claim severities, and dividend barriers, addressing a gap where these elements are often studied in isolation.
Mechanism-Specific Decomposition: It introduces a decomposition of the Gerber–Shiu function that explicitly tracks which severity regime causes ruin, allowing for a granular analysis of insolvency risk sources.
Analytical Tractability: By leveraging the phase-type (exponential–Erlang) structure of the premium distribution, the author derives a compact matrix-exponential solution for the ruin probabilities under exponential claim severities, avoiding the need for high-order scalar differential equations.
Empirical Validation: The model is applied to a real-world French private-motor insurance portfolio (Damage guarantee), fitting the extended-exponential distribution to premiums and the duration-dependent two-regime model to claim severities.

Results

Numerical Sensitivity Analysis:
- Initial Reserve: Ruin probabilities decrease non-linearly as the initial surplus increases, with a sharper decline at lower capital levels.
- Premium Parameters: Increasing the premium arrival intensity ( $\lambda$ ) reduces ruin risk. The shape parameter $\alpha$ of the extended-exponential distribution also reduces ruin risk; however, a mean-preserving analysis reveals that this effect is driven by both the increase in the mean premium amount and the change in distributional shape.
- Claim Parameters: Increasing the claim arrival intensity ( $\lambda_1$ ) increases ruin risk. This parameter also alters the unconditional probabilities of the two severity regimes, adding complexity to the sensitivity analysis.
- Dividend Barrier: A lower barrier leads to earlier dividend payments but significantly increases ruin probability due to reduced capital buffers.
Empirical Application:
- The duration-dependent model significantly outperforms a single exponential claim model in terms of AIC and BIC.
- Compared to a constant two-exponential mixture (which allows for heterogeneity but not duration dependence), the duration-dependent model yields slightly lower AIC and BIC values ( $\Delta \text{AIC} \approx 1.60$ ). The author characterizes this improvement as moderate, suggesting that while inter-claim duration provides additional information, much of the severity heterogeneity is captured by the two-regime structure itself.
- The fitted model identifies two distinct regimes: Mechanism 1 (rare, high severity, associated with longer waiting times) and Mechanism 2 (frequent, low severity, associated with shorter waiting times).

Significance and Claims
The paper claims that its primary significance lies in providing a tractable analytical framework for analyzing ruin in portfolios where claim severity depends on the time since the last claim and premiums are stochastic. The decomposition into mechanism-specific Gerber–Shiu functions offers a practical tool for insurers to understand the specific drivers of insolvency (e.g., distinguishing between frequent small claims and rare large claims triggered by long gaps).

The author is modest regarding the empirical findings. They do not claim the duration-dependent model is a universal description of motor insurance data. Instead, they conclude that the model is empirically supported as a useful extension that captures additional information regarding claim severity, though the improvement over a standard heterogeneous mixture is not overwhelming. The study emphasizes that the model is best interpreted at the portfolio level, where premium receipts and claims arise from different policies and do not need to follow a strict one-to-one temporal order.

Limitations
The paper acknowledges several limitations:

The assumption of Poisson arrival processes for both premiums and claims, which may not capture seasonality or clustering.
The reliance on exponential claim severities for the explicit matrix-exponential solution (though integral equations exist for general distributions).
The use of annual premium data rather than transaction-level data, which approximates the premium arrival intensity.
The lack of historical dividend data in the empirical dataset, meaning the dividend barrier is treated as a scenario parameter rather than an estimated policy.

More like this