Actuarial Risk Modeling

Actuarial and Financial Risk Modeling

Overview

Apply statistical modeling to uncertain outcomes where distributional assumptions, exposure, dependence, tail behavior, calibration, and decision consequences matter. The skill is methodology-first: it teaches model selection and evidence, not a particular library or rating formula.

When to Use

Load this skill when the task involves:

• insurance pricing, claims, reserving, solvency, risk classification, or experience rating;
• claim frequency, severity, pure premium, medical expenditure, loss, or event-time outcomes;
• linear, generalized linear, two-part, count, survival, panel, longitudinal, or tail models;
• financial returns, volatility, portfolio loss, risk measures, or scenario output;
• calibration, forecast evaluation, backtesting, model comparison, or assumption diagnosis;
• explaining model results, uncertainty, limitations, or use controls to decision-makers.

When Not to Use

• Use the Decision Entry Points table above for adjacent work. In brief: data-scientist owns general statistical, causal, experimental, and machine-learning methodology; financial-modeling owns deterministic operating, SaaS, fundraising, and cash-flow models.
• Use a named tool skill for operating a forecasting, database, or modeling platform.
• Do not present output as licensed actuarial, investment, legal, accounting, or regulatory advice. Escalate consequential decisions to qualified practitioners and applicable standards.

Decision Entry Points

Core Workflow

1. Frame the decision. State the decision, audience, horizon, unit of observation, estimand or forecast target, action threshold, and cost of false positives and negatives. Separate descriptive, predictive, and causal questions.

2. Write the data contract. Define grain, exposure or offset, outcome support, observation and development windows, censoring/truncation, policy or account boundaries, leakage risks, missingness states, and provenance.

3. Profile before modeling. Inspect distributions, zeros, negatives, skew, tail concentration, dependence, repeated entities, time ordering, category sparsity, exposure balance, and data-quality exceptions. Use scripts/risk_preflight.py for a read-only first pass.

4. Choose the simplest defensible model family. Match the outcome and data-generating structure before comparing algorithms. Load references/model-families.md for the decision table.

   Model-selection quick pick: counts with exposure → count GLM with an offset; zero plus positive loss → two-part/frequency-severity; event time with censoring → survival; ordered observations → dynamic/time-series model; tail decision → tail-aware or quantile model plus stress sensitivity. Load references/model-families.md before choosing a specific distribution or link.

5. Fit without contaminating evaluation. Treat transformations, imputation, feature selection, calibration, resampling, and hyperparameter choices as part of the fitted procedure. Fit them only on the permitted training partition.

6. Diagnose and challenge. Check residual structure, link and variance assumptions, overdispersion, zero inflation, leverage, collinearity, separation, calibration, dependence, censoring, tail fit, and sensitivity to plausible alternatives. A convergence flag is not validation.

7. Validate for use. Use grouped, blocked, or rolling splits when the deployment boundary demands them. Report point accuracy, probabilistic scores, calibration, ranking, tail or aggregate-loss behavior, stability across segments, and uncertainty. Use scripts/temporal_split_audit.py to audit time-ordered partitions.

   Validation-design quick pick: exchangeable observations → random holdout; repeated entities or clusters → grouped split; ordered deployment → blocked or rolling-origin split; overlapping development or labels → gap/embargo; extensive tuning or candidate comparison → nested validation. Load references/validation-and-calibration.md before fixing the final design.

8. Compare and govern. Prefer a transparent model unless a more complex one earns its complexity on the decision-relevant metric and remains stable, interpretable enough, and monitorable. Record assumptions, overrides, limitations, approvals, and rollback or review triggers.

9. Communicate the decision. Use templates/model-report.md and state what was observed, inferred, assumed, estimated, not identified, and not tested. Include units, intervals, scenario definitions, diagnostics, and a plain-language recommendation.

Required Distinctions

• Frequency is not severity. A count model and a positive-loss model have different supports, exposures, diagnostics, and aggregation rules.
• Prediction is not causation. A useful rating variable is not automatically a fair causal explanation or a permitted classification factor.
• Calibration is not discrimination. A model can rank well while producing systematically wrong probabilities.
• Backtesting is not proof. Historical success can reflect regime, selection, leakage, or unavailable information.
• Uncertainty is layered. Separate sampling error, parameter uncertainty, process variance, model-form uncertainty, scenario uncertainty, and data-quality uncertainty.
• A reserve or risk estimate is a decision input. It is not an objective fact independent of horizon, assumptions, and intended use.

Minimum Analysis Contract

Before presenting a recommendation, state: the decision and estimand; row grain, horizon, exposure, and information cutoff; candidate model family and why its support/dependence assumptions fit; validation boundary and metrics; uncertainty and sensitivity; limitations and permitted use. If one of these is unknown, label it as an unresolved input rather than silently choosing a convention.

Failure-Mode Quick Map