Bayesian approach to credibility theory and its use for calculating credibility premiums in simple cases

Welcome to Bayesian Credibility!

In this chapter, we are going to explore one of the most useful tools in an actuary's toolkit: Credibility Theory. Don’t let the name intimidate you! At its heart, credibility is just a way of deciding how much we should "trust" a specific piece of data versus our general knowledge.

Imagine you are insuring a fleet of delivery vans. You have general data for all delivery vans in the country, but you also have specific data for this one fleet. How do you set the price? Should you rely entirely on the fleet’s own history, or should you stick to the national average? Bayesian Credibility gives us a mathematical way to find the perfect balance between the two.

1. The Core Idea: The Credibility Formula

Before we dive into the math, let's look at the "Golden Formula" of this chapter. The Credibility Premium is always written as a weighted average:

\( P = Z\bar{x} + (1 - Z)\mu \)

Where:
• \( \bar{x} \) is the mean of the data we have observed (the Specific Experience).
• \( \mu \) is the mean of our prior beliefs (the Prior Knowledge or Manual Rate).
• \( Z \) is the Credibility Factor, a number between 0 and 1.

Quick Analogy: Imagine you are judging a cooking competition. One chef is a world-renowned professional (Prior Knowledge), but they just served you a slightly salty soup (Current Data). Because they are a pro, you give their past reputation a high weight and the one salty soup a low weight. That weight is your Z factor!

Key Takeaway:

If we trust our data a lot, \( Z \) will be close to 1. If the data is small or unreliable, \( Z \) will be close to 0, and we rely more on our prior beliefs.

2. The Bayesian Framework

In CS1, we use Bayesian Statistics to find this premium. Don't worry if Bayesian stats felt tricky in earlier chapters; the process here is very structured. We usually follow these steps:

Step 1: Identify the Prior Distribution of the parameter (usually called \( \theta \)). This represents our belief before seeing any data.
Step 2: Identify the Likelihood (the distribution of the data \( x \) given \( \theta \)).
Step 3: Calculate the Posterior Distribution using the data we've observed.
Step 4: Find the Posterior Mean. In Bayesian credibility, the posterior mean is our credibility premium.

Did you know? In many common cases, the posterior mean will automatically rearrange itself into the \( Z\bar{x} + (1-Z)\mu \) format. It’s like magic, but it’s actually just algebra!

3. Case Study 1: The Poisson-Gamma Model

This is the most common scenario in CS1 exams. It is usually used for the number of claims.

• The Situation: The number of claims follows a Poisson distribution with parameter \( \lambda \).
• The Prior: Our prior belief about \( \lambda \) follows a Gamma distribution with parameters \( \alpha \) and \( \beta \).
• The Data: we observe \( n \) years of data, and the total number of claims is \( \sum x_i \).

The Result:
The posterior distribution is also Gamma, with new parameters:
\( \alpha_{new} = \alpha + \sum x_i \)
\( \beta_{new} = \beta + n \)

The Credibility Premium:
The mean of a Gamma distribution is \( \alpha / \beta \). So, the posterior mean is:
\( \frac{\alpha + \sum x_i}{\beta + n} \)

Wait! How is that a credibility formula?
With a little algebra, we can rewrite it as:
\( P = (\frac{n}{\beta + n})\bar{x} + (\frac{\beta}{\beta + n})\frac{\alpha}{\beta} \)

Look closely:
• \( Z = \frac{n}{\beta + n} \)
• \( \bar{x} = \frac{\sum x_i}{n} \) (Sample mean)
• \( \mu = \frac{\alpha}{\beta} \) (Prior mean)

Common Mistake to Avoid:

Students often mix up \( n \) (the number of observations) and \( \sum x_i \) (the sum of the values). Remember: \( n \) affects the scale of our confidence (\( \beta \)), while the sum affects our estimate of the total (\( \alpha \)).

4. Case Study 2: The Normal-Normal Model

This model is often used for claim amounts (severity).

• The Situation: The data follows a Normal distribution \( N(\theta, \sigma^2) \).
• The Prior: Our belief about \( \theta \) follows a Normal distribution \( N(\mu, \tau^2) \).

The posterior distribution is also Normal. The posterior mean (our premium) is:
\( E[\theta | \text{data}] = \frac{\frac{n}{\sigma^2}\bar{x} + \frac{1}{\tau^2}\mu}{\frac{n}{\sigma^2} + \frac{1}{\tau^2}} \)

Simplifying to find Z:
If we divide the top and bottom by the denominator, we find:
\( Z = \frac{n}{n + \frac{\sigma^2}{\tau^2}} \)

Memory Trick:
Think of \( \sigma^2 \) as "noise" in the data and \( \tau^2 \) as the "spread" of our prior. If the "noise" (\( \sigma^2 \)) is very high, \( Z \) becomes smaller because we trust the messy data less!

5. Summary and Steps for Problem Solving

When you see a Bayesian credibility question in the CS1 exam, follow these steps:

1. Identify the Distributions: Is it Poisson-Gamma? Normal-Normal? Bernoulli-Beta?
2. Write down the Prior Mean: This is your \( \mu \).
3. Calculate the Posterior Mean: Use the formulas for the updated parameters (e.g., \( \alpha + \sum x \)).
4. Extract Z: Look at the coefficient in front of the sample mean \( \bar{x} \).
5. Sanity Check: Does \( Z \) increase as \( n \) increases? It should! More data should always mean more credibility.

Key Takeaways Checklist:

• Credibility Premium = Weighted average of data and prior.
• Conjugate Priors make the math easy because the posterior is the same family as the prior.
• As \( n \to \infty \), the credibility factor \( Z \to 1 \) (we eventually trust the data completely).
• As the prior variance increases, \( Z \to 1 \) (if our prior belief is vague, we rely more on the data).

Don't worry if the algebra feels heavy at first. Practice rearranging the posterior mean into the \( Z\bar{x} + (1-Z)\mu \) form a few times, and you'll start to see the pattern!