Credibility theory Study Guide

📖 Core Concepts Credibility Theory – Actuarial method that blends different sources of data to produce a risk‑premium (expected claim amount). Credibility Factor (Z) – A weight \(0\le Z\le 1\) that measures how much trust to place in a particular data source; higher when the source is less uncertain. Best Linear Approximation – Choose \(Z\) so the weighted estimate minimizes mean‑squared error to the true Bayesian predictive mean. Bayesian Credibility – Uses class probabilities (priors) and the conditional probability of the data to compute posterior class probabilities; the final estimate is a posterior‑weighted average of class statistics. Bühlmann Credibility – Decomposes total variance into: Variance of hypothetical means (between‑class variation). Expected process variance (within‑class variation). The relative sizes of these variances determine \(Z\). Actuarial Credibility (Frequentist view) – Combines a small‑sample estimate \(X\) with a larger, less specific estimate \(M\) to improve accuracy; \(Z\) balances sampling error of \(X\) against modeling error of \(M\). Premium Setting – Depends on portfolio homogeneity: Homogeneous → single premium = overall mean. Heterogeneous → group‑specific premiums; avoid adverse selection (low‑risk customers leaving). --- 📌 Must Remember Credibility‑weighted premium: \(\displaystyle \text{Premium}=Z\,\mu+(1-Z)\,\theta\). Interpretation of \(Z\): \(Z=1\) → trust only the specific experience (\(\theta\)). \(Z=0\) → trust only the overall experience (\(\mu\)). Bühlmann \(Z\) formula (conceptual): \(Z=\dfrac{\text{Var}(\text{hypothetical means})}{\text{Var}(\text{hypothetical means})+\text{E}[\text{process variance}]}\). Posterior class probability: \(\displaystyle P(\text{class }i \mid \text{data})=\frac{Pi\;Li}{\sumj Pj Lj}\) where \(Pi\) = prior class probability, \(Li\) = conditional likelihood. Adverse selection occurs when a uniform premium over‑charges low‑risk and under‑charges high‑risk groups. --- 🔄 Key Processes Compute Credibility‑Weighted Estimate Obtain overall estimate \(\mu\) (e.g., portfolio mean). Obtain specific estimate \(\theta\) (e.g., employer’s experience). Determine \(Z\) (via variance balance or Bayesian posterior). Plug into \(Z\mu+(1-Z)\theta\). Bayesian Credibility Updating Assign prior class probabilities \(Pi\). Calculate conditional data probabilities \(Li=P(\text{data}\mid\text{class }i)\). Compute overall data probability \(P(\text{data})=\sumi Pi Li\). Find posterior probabilities \(Pi^{\text{post}} = \frac{Pi Li}{P(\text{data})}\). Weight each class statistic by \(Pi^{\text{post}}\) and sum. Bühlmann Credibility Weight Determination Estimate between‑class variance \(\sigma^2{\text{b}}\) (variance of hypothetical means). Estimate within‑class variance \(\sigma^2{\text{w}}\) (expected process variance). Compute \(Z = \dfrac{\sigma^2{\text{b}}}{\sigma^2{\text{b}}+\sigma^2{\text{w}}/n}\) where \(n\) = number of observations for the specific group. --- 🔍 Key Comparisons Bayesian vs. Frequentist Credibility Bayesian: incorporates prior class probabilities; posterior weighting is explicit. Frequentist (Actuarial): blends a small‑sample estimate with a larger estimate; \(Z\) chosen to balance sampling vs. modeling error. Low‑Variance vs. High‑Variance Class (Bühlmann) Low within‑class variance: small contribution to process variance → higher \(Z\). High between‑class variance: large contribution to overall variance → also raises \(Z\) because the class is distinct from the overall mean. Homogeneous vs. Heterogeneous Portfolio Homogeneous: single premium works; no need for credibility weighting. Heterogeneous: requires group‑specific premiums and credibility adjustments to avoid adverse selection. --- ⚠️ Common Misunderstandings “Higher variance always means lower credibility.” Wrong: In Bühlmann, high between‑class variance actually increases credibility for that class because the class deviates from the overall mean. “Credibility factor of 0.5 means equal trust in both sources.” Only true if the two sources have comparable variances; otherwise 0.5 may be the result of variance balance, not equal informational value. “Bayesian credibility ignores data.” Incorrect; the data enter through the conditional likelihoods and reshape the posterior probabilities. --- 🧠 Mental Models / Intuition “Pull‑the‑rope” model – Imagine the overall mean and the specific mean tied together by a rope. The credibility factor tells you how far you pull the specific mean toward the overall mean: more uncertainty (weak rope) → pull further toward the overall mean. “Variance seesaw” – Think of the two variance components (between vs. within) as weights on a seesaw. The side with the larger weight dominates the tilt, giving that source a larger credibility share. --- 🚩 Exceptions & Edge Cases Very Small Group (n ≈ 1) – Process variance dominates; \(Z\) may be near 0, forcing reliance on the overall estimate. Extremely Homogeneous Classes – When within‑class variance ≈ 0, the class’s own data are perfectly reliable → \(Z\) approaches 1, even with few observations. Prior probabilities of zero – If a class’s prior is zero, its posterior remains zero regardless of data; the class is effectively excluded from the estimate. --- 📍 When to Use Which Use Bühlmann credibility when you have multiple risk classes with observable within‑class claim data and you can estimate both between‑ and within‑class variances. Use Bayesian credibility when prior class information (e.g., expert classification) is available and you want to update it with observed data. Use simple credibility weighting (Zμ+(1‑Z)θ) for single‑group premium calculations where only an overall benchmark and a specific experience estimate exist. --- 👀 Patterns to Recognize “Z close to 1” → Look for: low within‑class variance, high between‑class variance, or many observations for the group. “Uniform premium in a heterogeneous portfolio” → Flag potential adverse selection question. “Posterior probability proportional to prior × likelihood” – Classic Bayes’ theorem pattern; if a term appears, expect a posterior weighting step. --- 🗂️ Exam Traps Distractor: “Higher overall variance always reduces credibility.” → Wrong; only within‑class (process) variance reduces credibility; between‑class variance can increase it. Distractor: “Credibility factor is chosen arbitrarily.” → Incorrect; \(Z\) is derived from variance calculations or Bayes’ theorem, not guessed. Distractor: “If a class has zero prior probability, its data can still dominate the estimate.” → False; zero prior → zero posterior regardless of data. Distractor: “Adverse selection is avoided by charging a higher uniform premium.” → Misleading; a uniform premium exacerbates adverse selection; the remedy is risk‑based (credibility‑adjusted) pricing.