Continuous-time modeling and bootstrap for Schnieper's reserving

This paper proposes a continuous-time stochastic framework for Schnieper's reserving model that utilizes a Poisson measure for claim arrivals and Brownian motion for cost fluctuations, enabling a robust bootstrap method to estimate the full predictive distribution of reserves while naturally ensuring non-negativity and accounting for asymmetry.

The Gist

Imagine you are running a massive insurance company. Every year, you have to answer a very tricky question: "How much money do we need to set aside today to pay for accidents that have happened but haven't been fully counted yet?"

In the insurance world, this is called reserving. It's like trying to guess the final bill for a dinner party while the guests are still eating and the kitchen is still cooking.

This paper, written by Nicolas Baradel, proposes a new, smarter way to calculate that bill. It takes an old, classic recipe (Schnieper's model) and upgrades it with a "continuous-time" engine that handles uncertainty much better.

Here is the breakdown using simple analogies:

1. The Two Types of "Surprise" Bills

The author starts by splitting the unknown future costs into two distinct buckets, like sorting laundry into "whites" and "colors":

• Bucket A: The "New" Claims (True IBNR). These are accidents that have happened but no one has reported to the insurance company yet. It's like realizing you have a flat tire, but you haven't called the tow truck yet.
• Bucket B: The "Growing" Claims (IBNER). These are accidents that have been reported, but the estimated cost keeps changing. Maybe a car crash looked like a $5,000 repair, but after the mechanic looked closer, it's actually $15,000. Or maybe they found a cheaper part, and it's only $3,000.

2. The Old Way vs. The New Way

The Old Way (Discrete Steps):
Traditional models look at data like a flipbook. They check the numbers once a year (or once a quarter). They take a snapshot, make a guess, and move to the next snapshot.

• The Problem: Life doesn't happen in snapshots. It happens in a continuous flow. Also, old methods sometimes get silly results, like predicting a negative cost (which is impossible) or assuming the future looks exactly like a bell curve (a normal distribution), which often underestimates rare, huge disasters.

The New Way (Continuous-Time Modeling):
The author suggests treating the flow of money like a river, not a series of buckets.

• The River of New Claims: New accidents arrive randomly, like raindrops hitting a pond. The author uses a mathematical tool called a Poisson process to model these random "splashes."
• The River of Changing Costs: Once a claim is in the system, its cost fluctuates up and down like a boat on waves. The author uses Brownian motion (the same math used to model how dust particles jitter in the air) to model these cost changes.

3. The "Bootstrap" Magic Trick

The paper introduces a method called Bootstrap. Imagine you are trying to predict the weather for next year, but you only have data from the last 10 years.

• The Trick: Instead of just making one guess, you create 1,000 "parallel universes." In each universe, you slightly tweak the weather patterns based on what you know, then run the simulation forward.
• The Result: You end up with 1,000 different possible futures. You can then see the whole picture: "In 95% of these universes, the cost is low. But in 5% of them, it's huge." This gives you a much better safety net than just guessing a single average number.

4. Why This New Model is Better

The author's new "Continuous-Time Bootstrap" has three superpowers that the old methods lack:

1. No Negative Numbers: You can't have a negative cost for a car repair. Old models sometimes accidentally calculated negative reserves. This new model is built so that the numbers physically cannot go below zero.
2. It Respects the "Tail": Insurance is all about the rare, catastrophic events (the "fat tails" of the distribution). This model naturally accounts for the fact that sometimes, things go very wrong, without needing to force-fit the data into a perfect bell curve.
3. It Handles "Zero" Claims: If a claim is fully paid and settled, the cost stops. The new model handles this "stopping" behavior naturally, whereas older models sometimes struggle to know when to stop the simulation.

5. The Real-World Test

The author tested this on real data from a motor insurance company (cars).

• They compared their new "River" model against the old "Flipbook" models.
• The Result: The new model gave a more realistic view of the risk. It showed that while the average cost might be similar to old methods, the risk of a huge surprise was captured much more accurately.

The Bottom Line

Think of this paper as upgrading the navigation system of an insurance company.

• Old GPS: "You will arrive in 2 hours." (It gives one number and ignores traffic jams).
• New GPS: "There is a 90% chance you arrive in 2 hours, but there is a 10% chance of a massive traffic jam that could take 5 hours. Here is the map of all possible routes."

By using continuous math and simulating thousands of "what-if" scenarios, this method helps insurance companies set aside the right amount of money—not too little (which would bankrupt them in a crisis) and not too much (which would waste money that could be invested elsewhere).

Technical Summary

Here is a detailed technical summary of the paper "Continuous-time modeling and bootstrap for Schnieper's reserving" by Nicolas Baradel.

1. Problem Statement

The paper addresses the challenge of modeling and estimating Incurred But Not Reported (IBNR) reserves in insurance, specifically within the context of excess-of-loss reinsurance.

• The Decomposition: It focuses on Schnieper's model, which decomposes IBNR reserves into two distinct components:
  1. True IBNR: Claims that have occurred but have not yet been reported.
  2. IBNER (Incurred But Not Enough Reported): Changes in the estimated ultimate cost of claims that have already been reported.
• Limitations of Existing Methods: While Schnieper's original discrete-time model provides unbiased estimators, subsequent stochastic extensions (e.g., non-parametric bootstraps) often struggle with:
  • Generating negative reserve values during simulation.
  • Failing to capture the inherent asymmetry of claim distributions.
  • Ignoring the continuous nature of claim development and cost fluctuations.
  • Requiring ad-hoc adjustments to enforce non-negativity.

2. Methodology

The author proposes a continuous-time stochastic framework that unifies Schnieper's discrete assumptions with modern stochastic differential equations (SDEs).

A. The Continuous-Time Model

The aggregate claims process $(C_t^i)$ for accident year $i$ is modeled as the unique strong solution to an SDE driven by two independent mechanisms:

1. Poisson Measure (for True IBNR): A random Poisson measure $\Pi^i$ drives the arrival of new claims. This captures the discrete, jump-like nature of new claim reporting.
2. Brownian Motion (for IBNER): A Brownian motion $W^i$ drives the fluctuations in the cost of already reported claims. This captures the continuous re-estimation of ultimate costs.

The governing SDE is:
$$C_t^i = C_s^i + \int_s^t \int_{\mathbb{R}_+^*} z \Pi^i(du, dz) - \int_s^t \delta_u C_u^i du + \int_s^t \tau_u \sqrt{C_u^i} dW_u^i$$

• Drift ($\delta_u$): Represents the decay or development of reported claims.
• Diffusion ($\tau_u$): Represents the volatility of cost re-estimation.
• Jump Term: Represents the arrival of new claims with size $Z$.

B. Consistency with Schnieper's Framework

The paper rigorously proves that this continuous-time model satisfies the original assumptions of Schnieper's discrete model (Assumptions H1, H2, H3) regarding conditional means and variances.

• The discrete parameters ($\Lambda_j, \Delta_j, \Sigma_j, T_j$) are derived as integrals of the continuous coefficients ($\lambda, \delta, \tau$) over development periods.
• A key theoretical contribution is the branching property: The process can be decomposed into independent branches starting from each jump time, allowing for exact simulation.

C. Bootstrap Methodology

To estimate the predictive distribution of reserves, the author develops a two-step parametric bootstrap procedure:

1. Parameter Bootstrapping:
   • Simulate the increments (new claims $N$ and cost changes $D$) conditional on observed data using the estimated continuous-time parameters.
   • Re-estimate the Schnieper parameters ($\hat{\Lambda}, \hat{\Delta}, \hat{\Sigma}, \hat{T}$) from these simulated increments to capture estimation error.
2. Process Bootstrapping:
   • Using the bootstrapped parameters, simulate the future development of the lower triangle (unobserved claims) iteratively.
   • This captures process error (stochastic variability).

• Distribution of Jump Sizes ($Z$): Since aggregated data does not reveal individual claim sizes, the paper proposes estimating the ratio $E[Z^2]/E[Z]$ via weighted linear regression. A Gamma distribution is then fitted to $Z$ to complete the model specification.

3. Key Contributions

1. Theoretical Unification: Successfully bridges Schnieper's discrete decomposition with continuous-time stochastic calculus, providing a coherent parametric structure for both new claims and cost re-estimation.
2. Intrinsic Non-Negativity: Unlike discrete bootstraps that can generate negative cumulative claims (requiring truncation), the proposed SDE solution is guaranteed to remain non-negative by construction.
3. Asymmetry and Bounds: The model naturally accounts for the asymmetry of claim distributions and respects intrinsic bounds (e.g., cost reductions cannot exceed the current reserve) without artificial adjustments.
4. Exact Simulation: Leveraging the branching property and explicit distributional representations (Proposition 3.12), the method allows for exact simulation of the reserve distribution, avoiding discretization errors common in Euler-Maruyama schemes.

4. Results (Numerical Case Study)

The method was tested on a real-world motor third-party liability excess-of-loss dataset from Schnieper (1991).

• Comparison: The continuous-time bootstrap was compared against:
  • Schnieper's model with a Log-normal approximation.
  • Schnieper's model with a standard non-parametric bootstrap.
  • A Time-series bootstrap approach.
• Performance Metrics:
  • MSEP (Mean Squared Error of Prediction): The continuous-time bootstrap yielded an MSEP of ~43.1% of the point estimate, comparable to the Log-normal method (42.9%) but slightly higher than the standard Schnieper bootstrap (38.2%).
  • Tail Risk (99.5% Quantile): The continuous-time method produced a 99.5% quantile excess of 136.7%. This is significantly lower than the Log-normal approximation (165%) but higher than the standard bootstrap (103%).
• Sensitivity Analysis: The results showed sensitivity to the assumed mean of the jump size $E[Z]$. Lower values of $E[Z]$ (implying a heavier-tailed distribution for claim sizes) increased the MSEP and tail quantiles.

5. Significance

• Robustness: The framework eliminates the risk of generating negative reserves, a common failure mode in discrete stochastic reserving methods.
• Flexibility: It allows for the incorporation of individual claim data (if available) to refine the jump size distribution, while remaining valid for aggregated triangle data.
• Practical Application: By providing a mathematically rigorous way to simulate the full predictive distribution, it offers insurers a more reliable tool for capital allocation and risk management, particularly for the tail risks associated with IBNR and IBNER.
• Extension of Literature: It extends the continuous-time philosophy previously applied to Mack's model (by [2]) to the more complex, two-component Schnieper model, filling a gap in the actuarial literature.

In conclusion, Baradel presents a sophisticated stochastic model that respects the structural assumptions of Schnieper's decomposition while leveraging the mathematical advantages of continuous-time processes to ensure non-negativity and capture realistic reserve dynamics.