Purpose
Convert raw experience data into a defensible actual-to-expected analysis — computing A/E ratios by cohort, testing credibility, isolating statistically significant deviations from expected, and recommending specific assumption updates with supporting rationale. Solves the problem of translating scattered claims and exposure data into an assumption-setting decision an actuary can sign off on.
When to Use
Invoke this skill when the user asks to:
- Run an annual mortality or morbidity assumption review
- Benchmark a book of business against SOA VBT, 2015 VBT, or an industry morbidity table
- Compute actual-to-expected (A/E) ratios and test whether deviations are credible
- Support a product repricing with an updated mortality or morbidity basis
- Investigate an emerging trend (deteriorating or improving experience) by cohort
- Validate whether a reserve or pricing table still fits recent experience
Inputs Required
Before starting, collect:
- Experience data: Actual deaths/claims and exposures — by policy year, attained age band, gender, and smoker status if available (count-based and/or amount-based)
- Standard table basis: The expected table (e.g., 2015 VBT ANB/ALB, RP-2014, SOA industry LTC/DI/CI tables) and any applied selection/ultimate split
- Line of business: Term life, whole life, DI, LTC, critical illness, group, etc.
- Exposure basis: Count-based, face-amount, or annualized-benefit exposure
- Study period and prior A/E: The observation window and last review's A/E if available
- Mortality improvement assumption: Any projection scale currently applied (e.g., MP-2021, Scale G2)
- Credibility standard: Target credibility method and threshold (default: Limited Fluctuation, 90% within ±5%, ~1,082 expected claims for full credibility)
If credibility standard or improvement scale is missing, apply the defaults noted above and state the assumption explicitly in the output.
Instructions
Step 1: Structure the Exposure and Expected Basis
Reconcile the data before analysis:
- Confirm exposure units match the expected table (count vs. amount) — an amount-weighted A/E against a count-based expected is invalid
- Apply the correct select/ultimate split by duration; do not compare ultimate-basis experience against a select-period expected
- Apply any in-force improvement/projection scale to bring the base table to the study midpoint
- Bucket into cohorts: policy duration (1, 2–5, 6–10, 11–15, 16+), attained age band, gender, and risk class
- Flag any cohort with material data gaps or exposure below a usable floor
Step 2: Compute A/E Ratios by Cohort
For each cohort and in aggregate:
- A/E = Actual claims (or amount) / Expected claims (or amount)
- Compute both count-based and amount-based A/E where data allows — divergence between the two signals anti-selection or large-claim concentration
- Report Actual, Expected, A/E, and expected claim count per cell (the credibility driver)
- Aggregate up the hierarchy (cell → age band → duration → total) so partial credibility can borrow from adjacent cells
Step 3: Test Statistical Credibility
Assess whether each A/E is reliable enough to act on:
- Full credibility (Limited Fluctuation): expected claims ≥ ~1,082 (count) for 90%/±5%; scale threshold for the chosen p/k
- Partial credibility factor Z = √(n / n_full), capped at 1.0
- Credibility-weighted A/E = Z × observed A/E + (1 − Z) × prior/complement — state the complement source (prior study, aggregate, or table = 1.00)
- Flag cells where Z < 0.30 as "directional only — do not set assumptions on this cell alone"
- Note where amount-based A/E is credible but count-based is not, or vice versa
Step 4: Flag Significant Deviations
Isolate what actually matters:
- Materiality screen: credible A/E deviating more than ±5% from 1.00 (or from prior assumption)
- Statistical screen: apply an approximate confidence interval on A/E (Poisson-based on expected claims) — flag cells where 1.00 falls outside the interval
- Classify each flag: deterioration (A/E > 1 and rising), improvement (A/E < 1 and falling), anti-selection pattern (early-duration A/E elevated), wear-off/reversion (select-period pattern normalizing), or large-claim noise (amount-A/E spike, count-A/E stable)
- Distinguish signal from noise: a single high-severity claim in a low-exposure cell is not a trend
Step 5: Trend Analysis Across the Study
- Compare current-period A/E to prior study A/E by cohort to identify direction and slope
- Separate genuine trend from mix shift — a rising aggregate A/E driven by cohort mix (aging book) is not the same as underlying deterioration
- For mortality, assess whether the applied improvement scale is over- or under-shooting recent experience
Step 6: Recommend Assumption Updates
For each credible, material, flagged cohort:
- Current basis → recommended basis (as an A/E load, table swap, or improvement-scale change)
- Magnitude and direction of the change
- Rationale tied to credibility (Z), materiality, and trend direction
- Where to hold (insufficient credibility → retain current assumption, monitor)
- Downstream impact note: pricing, reserves, or embedded-value implication at a directional level Be decisive on credible cells; be explicitly conservative where credibility is thin.
Output Format
- Structure: A/E results in a table by cohort, followed by prioritized deviation flags and a ranked list of assumption recommendations
- Length: 900–1,300 words plus the A/E table(s)
- Format: Tables for A/E and credibility; structured bullets for flags; numbered recommendations
- Sections: Data & basis reconciliation → A/E table by cohort → Credibility assessment → Significant deviation flags → Trend analysis → Recommended assumption updates → Monitoring plan
- Tone: Consulting actuary — precise, evidence-first, credibility-aware, willing to recommend "hold and monitor" when the data doesn't support a change
Quality Checklist
Before delivering output, verify:
- [ ] A/E is computed on a consistent basis (count vs. amount) matching the expected table
- [ ] Every A/E cell reports its expected claim count and credibility factor Z
- [ ] Select/ultimate and improvement-scale adjustments are applied and stated
- [ ] Deviation flags pass both a materiality (±5%) and a statistical (interval) screen
- [ ] Each assumption recommendation cites credibility, materiality, and trend direction
- [ ] Low-credibility cells (Z < 0.30) are marked "hold/monitor," not used to reset assumptions
Notes
- The most common error is acting on aggregate A/E movement that is actually a cohort mix shift — always decompose before recommending a change
- Amount-based A/E is more volatile than count-based; a spike there with stable count-A/E usually means large-claim concentration, not a mortality trend
- Where credibility is partial, always name the complement of credibility explicitly — an unstated complement makes the credibility-weighted A/E indefensible in review
- For LTC, DI, and CI, incidence and termination should be evaluated separately; a blended morbidity A/E can mask offsetting errors
How to use this skill
This file is a Claude Code skill in Anthropic's standard format — markdown with YAML frontmatter at the top.
Installing in Claude Code:
~/.claude/skills/<folder-name>/SKILL.md— auto-loads in every session anywhere on your machine (recommended for skills you'll use across multiple projects)<your-project>/.claude/skills/<folder-name>/SKILL.md— auto-loads only when working in that project (recommended for project-specific skills)
Restart Claude Code or open a new session to register.
Using in Claude.ai, ChatGPT, Gemini, or other AI tools:
- Drag this file into a new chat, or copy and paste its contents
- Send your task or request as your next message — the AI will apply the skill where it fits
For Claude Code skill details: https://docs.anthropic.com/en/docs/claude-code/skills