Bayesian (greatest accuracy) credibility

Welcome to Bayesian Credibility: The "Gold Standard" of Prediction!

Hello there! Welcome to one of the most intellectually rewarding chapters in Exam ASTAM. If you have ever felt like you were guessing how much to charge an insurance client, Bayesian Credibility (also known as Greatest Accuracy Credibility) is here to save the day.

In the previous sections, you might have looked at Limited Fluctuation Credibility, which uses somewhat arbitrary "standards." Bayesian Credibility is different. It uses rigorous probability theory to combine what we think (prior knowledge) with what we see (actual data) to predict the future. Don't worry if it seems heavy on the math at first—we will break it down into simple steps!

1. The Core Philosophy: "Prior" vs. "Posterior"

Imagine you are meeting a new friend. Before they speak, you have a "prior" opinion of them based on their reputation (this is the Prior Distribution). After you hang out with them for a few hours (this is the Data), you update your opinion. This updated opinion is your Posterior Distribution.

In actuarial terms:
1. Prior Distribution \(\pi(\theta)\): Our belief about the risk parameter \(\theta\) before seeing any claim data.
2. Model Distribution (Likelihood) \(f(x|\theta)\): The probability of seeing specific claim data \(x\), given a specific risk level \(\theta\).
3. Posterior Distribution \(\pi(\theta|\mathbf{x})\): Our updated belief about the risk level after seeing the claim data \(\mathbf{x}\).
4. Predictive Distribution: The distribution of the next claim, which we use to set the premium.

Key Formula: The Posterior

The math behind updating our belief follows Bayes' Theorem:
\( \pi(\theta|\mathbf{x}) = \frac{f(\mathbf{x}|\theta)\pi(\theta)}{\int f(\mathbf{x}|\theta)\pi(\theta) d\theta} \)

Pro-Tip: In exam questions, you usually don't need to solve that scary integral in the denominator! You just need to recognize the shape (the kernel) of the distribution in the numerator.

Summary: Bayesian credibility is the process of updating our initial assumptions using real-world observations to produce a "best guess" for future claims.

2. The Step-by-Step Bayesian Recipe

If you get a question asking for the "Bayesian Premium" or "Predictive Mean," follow these steps:

Step 1: Identify the Likelihood. Write down the probability of the observed data. If you have \(n\) years of data, this is usually \(f(x_1|\theta) \cdot f(x_2|\theta) \cdots f(x_n|\theta)\).

Step 2: Identify the Prior. This is \(\pi(\theta)\), usually given in the problem (e.g., "Theta follows a Gamma distribution").

Step 3: Find the Posterior. Multiply the Likelihood by the Prior. Ignore any constants that don't involve \(\theta\). Look at the resulting formula and see if it looks like a distribution you know (like Gamma or Normal).

Step 4: Calculate the Predictive Mean. This is the expected value of the next claim, given the updated (posterior) parameters.
\( E[X_{n+1}|\mathbf{x}] = E[E[X_{n+1}|\theta] | \mathbf{x}] \)

Did you know? The term "Greatest Accuracy" is used because this method minimizes the Mean Squared Error (MSE). It is mathematically the most accurate linear or non-linear predictor possible!

3. Conjugate Priors: The "Shortcuts"

In the heat of the exam, you won't have time to do long calculus. This is where Conjugate Priors are your best friends. If the prior and the likelihood belong to certain "matching" families, the posterior will be in the same family as the prior!

Common Pairs to Memorize:
1. Poisson (Likelihood) + Gamma (Prior) \(\rightarrow\) Gamma (Posterior)
2. Binomial (Likelihood) + Beta (Prior) \(\rightarrow\) Beta (Posterior)
3. Exponential (Likelihood) + Inverse Gamma (Prior) \(\rightarrow\) Inverse Gamma (Posterior)
4. Normal (Likelihood) + Normal (Prior) \(\rightarrow\) Normal (Posterior)

Example: Poisson-Gamma Shortcut

If claims \(X_i\) are Poisson with mean \(\lambda\), and \(\lambda\) follows a Gamma distribution with parameters \((\alpha, \theta)\):
The posterior \(\lambda|\mathbf{x}\) is also Gamma with:
New \(\alpha'\) = \(\alpha + \sum X_i\)
New \(\theta'\) = \(\frac{\theta}{n\theta + 1}\) (if using the \( \theta \) scale) or New \(\beta'\) = \(\beta + n\) (if using the \( \beta \) rate scale).

Common Mistake: Watch out for which version of the Gamma distribution the exam uses (\(\theta\) vs \(\beta\)). Always check the formula sheet provided in the exam!

4. Bayesian vs. Bühlmann Credibility

You might wonder: "When do I use the big Bayesian formula and when do I use the \(Z \bar{X} + (1-Z)\mu\) formula?"

Here is the magic: For many common distributions (the Linear Exponential Family with conjugate priors), the Bayesian result and the Bühlmann result are exactly the same!

Quick Review Box:
- Bayesian: Uses the full distribution (Exact).
- Bühlmann: Uses only the first two moments (Mean and Variance).
- The Link: If the problem asks for the "Bayesian Premium" for a Poisson-Gamma or Normal-Normal case, you can often use the simpler Bühlmann \(Z\) formula and get the right answer faster!

5. Summary and Key Takeaways

Key Points to Remember:
- Bayesian credibility uses Prior knowledge + New Data = Posterior knowledge.
- The Posterior Distribution is proportional to Likelihood \(\times\) Prior.
- Use Conjugate Priors to skip the integration steps.
- The Predictive Mean is your "Greatest Accuracy" estimate for the next period's claims.

Don't worry if this seems tricky at first! The math can be intimidating, but the logic is very consistent. Practice identifying the "Prior" and "Likelihood" in word problems, and the rest will fall into place. You've got this!