Welcome to Statistical Reserving!

In your earlier studies, you likely learned the Chain Ladder (Development) method. It’s a classic! But the basic Chain Ladder has one big limitation: it only gives you a single number (a "point estimate"). In the real world, regulators and company boards want to know, "How wrong could we be?"

In this chapter, we move from simple arithmetic to statistical models. We are going to look at Mack’s Model, the Poisson Model, and the Overdispersed Poisson (ODP) Model. These tools allow us to measure the uncertainty (variance) around our reserve estimates. Don’t worry if the formulas look intimidating at first—we’ll break them down piece by piece!

1. Mack’s Model: The "Stochastic" Chain Ladder

Think of Mack’s Model as giving the Chain Ladder method a statistical backbone. It doesn't change the reserve estimate itself, but it provides a way to calculate the Standard Error of those reserves.

The Three Big Assumptions

For Mack’s model to work, we assume three things about our cumulative losses \( C_{i,j} \) (where \( i \) is the accident year and \( j \) is the development year):

  1. Independence: Accident years are independent. What happens in 2022 doesn't affect what happens in 2023.
  2. Mean: The expected value of the next period depends only on the current period: \( E[C_{i, j+1} | C_{i,1}, ..., C_{i,j}] = C_{i,j} \cdot f_j \).
  3. Variance: The variance of the next period is proportional to the current period: \( Var(C_{i, j+1} | C_{i,1}, ..., C_{i,j}) = C_{i,j} \cdot \sigma_j^2 \).
Estimating the Parameters

To use Mack's model, we need to find two things for every development period \( j \):

1. Development Factors (\( f_j \)): These are the same ones you used in the basic Chain Ladder!
\( \hat{f}_j = \frac{\sum_{i=1}^{n-j} C_{i, j+1}}{\sum_{i=1}^{n-j} C_{i,j}} \)

2. Variance Parameters (\( \sigma_j^2 \)): This tells us how much the data "wiggles" around the development factor.
\( \hat{\sigma}_j^2 = \frac{1}{n-j-1} \sum_{i=1}^{n-j} C_{i,j} \left( \frac{C_{i, j+1}}{C_{i,j}} - \hat{f}_j \right)^2 \)

Did you know? If you only have one data point left for the final year (meaning \( n-j-1 = 0 \)), you can't use the formula above. Actuaries usually use a "rule of thumb" or an exponential decay pattern to estimate the final \( \sigma^2 \).

The Mean Square Error (MSE)

This is the "magic" of Mack's model. The Mean Square Error of Prediction (MSEP) for a specific accident year \( i \) is calculated using a recursive formula. While the full formula is long, remember that it consists of two parts:

MSEP = Process Variance + Estimation Variance

  • Process Variance: Randomness we can't avoid (the "luck of the draw").
  • Estimation Variance: Error because our calculated \( f_j \) might not be the true factor.

Quick Review: Mack’s Model gives the exact same reserve estimates as the Chain Ladder, but adds a way to measure uncertainty (Standard Error).

2. The Poisson Reserving Model

Now we shift gears into Generalized Linear Models (GLMs). Instead of looking at cumulative totals, we look at incremental losses (\( X_{i,j} \)).

In a basic Poisson model, we assume that the incremental loss \( X_{i,j} \) follows a Poisson distribution. However, there’s a catch: the Poisson distribution requires the Mean to equal the Variance (\( E[X] = Var(X) \)).

The Problem: Insurance claims are messy. Usually, the variance is much larger than the mean. This is called Overdispersion. Using a standard Poisson model would make us overconfident (it would underestimate the risk).

3. The Overdispersed Poisson (ODP) Model

To fix the Poisson model, we introduce a dispersion parameter (\( \phi \)). In the ODP model:

  • Mean: \( E[X_{i,j}] = \mu_{i,j} \)
  • Variance: \( Var(X_{i,j}) = \phi \cdot \mu_{i,j} \)

By including \( \phi \), we acknowledge that the data is "noisier" than a standard Poisson distribution.

Why Actuaries Love the ODP Model

Here is the most important "exam trick" to remember: The ODP model results in the exact same reserve estimates as the Chain Ladder method!

If an exam question asks for the reserve estimate using an ODP model, and you have a loss triangle, you can simply use the standard Chain Ladder approach unless told otherwise. The ODP model is just a formal statistical way to justify the Chain Ladder results.

Key Differences between Mack and ODP
  1. Mack is "non-parametric" (it doesn't assume a specific distribution shape).
  2. ODP is a GLM (it assumes a specific distributional form for the increments).
  3. Mack works on cumulative data; ODP works on incremental data.

Analogy Time! Imagine you are predicting how much it will snow tomorrow.
- Chain Ladder: "It will snow 5 inches."
- Mack/ODP: "It will likely snow 5 inches, but there is a 95% chance it will be between 2 and 8 inches."

4. Comparing the Models: MSEP Breakdown

In Exam ASTAM, you might be asked to calculate or explain the Mean Square Error of Prediction (MSEP). It’s vital to understand what goes into it.

\( MSEP[\hat{R}] \approx \sum Var(Process) + \sum Var(Estimation) \)

Common Mistake: Students often forget that when calculating the MSEP for the total reserve (all years combined), you can't just add the individual standard errors. You have to account for the fact that the same \( \hat{f}_j \) factors are used across all accident years, creating a correlation in our estimation error.

Quick Review Box:
1. Chain Ladder: Simple point estimate.
2. Mack: Adds variance to Chain Ladder using cumulative data.
3. ODP: A GLM that mirrors Chain Ladder results but allows for statistical testing.
4. MSEP: The sum of "Process" (randomness) and "Estimation" (parameter error).

5. Summary and Final Tips

Key Takeaways:
  • Mack’s Model relies on three assumptions (Independence, Mean proportionality, Variance proportionality).
  • \(\hat{f}_j\) in Mack’s is the same as the volume-weighted average development factor in Chain Ladder.
  • Overdispersion means Variance > Mean. The ODP model handles this using \( \phi \).
  • If a model is stochastic (like Mack or ODP), it means we can calculate a confidence interval or VaR (Value at Risk) for our reserves.
Exam Tips:
  • If you see a question about Mack's Variance (\( \sigma_j^2 \)), keep your table organized. Calculate all the \( f_j \) first, then all the \( \sigma_j^2 \).
  • Don't let the notation \( C_{i,j} \) vs \( X_{i,j} \) trip you up. C is for Cumulative (totals), X is for Incremental (the "slices").
  • Remember: ODP = GLM version of Chain Ladder. This connection is a favorite for conceptual questions!

Keep practicing those triangle calculations! The more you do, the more these formulas will feel like second nature. You've got this!