Purpose
Translate a described prediction problem and dataset profile into a ranked shortlist of appropriate modeling approaches — with the assumptions, interpretability trade-offs, validation strategy, and regulatory fit spelled out — so the analyst starts building the right model instead of defaulting to whatever is familiar or reaching for a black box that won't pass review.
When to Use
Invoke this skill when the user asks to:
- Decide which model to use for a specific prediction, classification, or risk problem
- Compare modeling approaches (e.g., GLM vs. gradient boosting, Cox vs. accelerated failure time)
- Choose a model that satisfies interpretability or regulatory constraints
- Recommend a survival, frequency-severity, or time-to-event modeling approach
- Justify a modeling method choice to a review board, actuary, or regulator
- Diagnose why a chosen model class isn't fitting and what to switch to
Inputs Required
Before starting, collect:
- Prediction target: What is being predicted, and its type — continuous, binary, count, time-to-event, multiclass, ordinal, or probability
- Business/risk objective: The decision the model informs (e.g., pricing, reserving, fraud flagging, churn, claim propensity)
- Dataset profile: Row count, feature count, feature types (numeric/categorical), sparsity, and whether observations are independent or grouped/panel
- Data quality notes: Missingness pattern, class imbalance, censoring/truncation, outliers, known measurement issues
- Interpretability requirement: Fully transparent (coefficient-level explanation required) / partially interpretable (feature-level explanation acceptable) / black box acceptable
- Regulatory or governance context: Any constraints — insurance rate filing, fair lending, model risk management, audit requirements, or none
- Deployment constraints: Latency, retraining frequency, monobank/production environment limits, or none
If the target type or interpretability requirement is missing, ask one clarifying question — these two inputs drive the entire recommendation.
Instructions
Step 1: Classify the Prediction Problem
Pin down the problem type before naming any model. Map to one of these families:
- Continuous outcome → regression family (linear, GLM Gamma/Tweedie, quantile, GAM, tree ensembles)
- Binary/probability outcome → classification family (logistic GLM, penalized logistic, gradient boosting, GAM)
- Count outcome → count family (Poisson, Negative Binomial, zero-inflated, Tweedie for frequency-severity)
- Time-to-event with censoring → survival family (Kaplan-Meier, Cox PH, AFT, discrete-time hazard, survival GBM)
- Multiclass/ordinal → multinomial or ordinal logistic, softmax GBM
- Grouped/hierarchical data → mixed-effects / GLMM, or credibility-weighted approaches
Explicitly name which family applies and why, based on the target type stated. If the data is panel/grouped or heavily censored, flag that it overrides the naive choice (e.g., "counts with excess zeros → zero-inflated, not plain Poisson").
Step 2: Apply the Interpretability Filter
Cut the candidate space by the stated interpretability requirement:
- Fully transparent required (rate filings, fair lending, regulated pricing): Restrict to GLMs, GAMs with monotonic constraints, penalized regression, decision trees with bounded depth. Exclude unconstrained ensembles.
- Partially interpretable acceptable: Allow gradient boosting / random forests with SHAP or partial dependence, monotonic-constrained GBMs, and Explainable Boosting Machines (EBM/GA2M).
- Black box acceptable: Open the full space — XGBoost/LightGBM/CatBoost, neural nets — but still rank by data-fit, not just accuracy ceiling.
State which candidates were eliminated by this filter and why. In regulated insurance contexts, note that a monotonic-constrained GBM or GLM+GAM is often the practical ceiling.
Step 3: Match to Dataset Characteristics
Refine the shortlist against the data profile:
- Small n / high dimensionality: Favor penalized regression (Lasso/Ridge/Elastic Net) over deep ensembles; ensembles overfit small samples
- Large n, many interactions: Favor gradient boosting — captures nonlinearity the analyst won't specify manually
- Heavy class imbalance: Note resampling, class weights, or focal loss; recommend PR-AUC over ROC-AUC as the evaluation metric
- High-cardinality categoricals: Favor CatBoost / target encoding / GLM with credibility grouping over one-hot into a linear model
- Missingness: If informative, recommend explicit missing-indicator or native-NA handling (LightGBM/XGBoost); if MAR, note imputation strategy
- Censoring/truncation (insurance claims, policy duration): Enforce survival methods — a standard regressor on observed durations is biased
Step 4: Produce the Ranked Shortlist (3-5 Candidates)
For each candidate, provide a structured block:
- Model: Named method (e.g., "Tweedie GLM," "Cox Proportional Hazards," "Monotonic-constrained LightGBM")
- Fit rationale: 2 sentences on why it suits this problem and data
- Key assumptions: The assumptions that must hold (e.g., proportional hazards, log-link variance structure, independence) and how to check each
- Interpretability level: What the analyst can explain to a stakeholder with this model
- Watch-outs: The specific failure mode for this pairing (e.g., "PH assumption violated for time-varying covariates")
Rank by best overall fit given the constraints — not by raw predictive power. Put the recommended primary first.
Step 5: Recommend Validation Strategy
Specify how to test the chosen approach honestly:
- Split scheme: Random k-fold vs. time-based holdout (mandatory for temporal drift) vs. grouped CV (mandatory for panel data to prevent leakage)
- Evaluation metric matched to target: RMSE/MAE (continuous), deviance/log-loss (GLM), PR-AUC (imbalanced binary), C-index / integrated Brier score (survival), Gini/lift (insurance ranking)
- Assumption diagnostics: The specific checks for the recommended model (residual plots, PH tests, calibration curves, dispersion test)
- Overfitting guard: Early stopping, regularization path, or nested CV for hyperparameter tuning
Step 6: Name the Fallback
State the second-choice approach and the trigger to switch to it — e.g., "If PH tests fail, move from Cox to an AFT or discrete-time hazard model," or "If the GLM underfits nonlinear interactions, escalate to a monotonic-constrained GBM with SHAP for the filing narrative."
Output Format
- Structure: Problem classification → interpretability filter → data-fit refinement → ranked shortlist (structured blocks) → validation strategy → fallback
- Length: 900–1,300 words
- Format: Structured document — prose for reasoning, numbered blocks for the shortlist, bulleted lists for validation and diagnostics
- Sections: Problem Classification, Interpretability Filter, Data-Fit Refinement, Ranked Model Shortlist, Validation Strategy, Fallback Plan
- Tone: Senior quantitative analyst — precise, assumption-aware, decisive without overclaiming; names methods exactly and refuses to hand-wave over trade-offs
Quality Checklist
Before delivering output, verify:
- [ ] The prediction target type is explicitly classified into a model family before any method is named
- [ ] 3–5 candidate models are provided, each with fit rationale, assumptions, interpretability level, and a specific watch-out
- [ ] Candidates eliminated by the interpretability filter are named, with the reason stated
- [ ] The evaluation metric recommended matches the target type (no ROC-AUC for continuous, no RMSE for survival)
- [ ] Censoring, imbalance, grouping, or small-n conditions in the input actually change the recommendation — not ignored
- [ ] A named fallback and a concrete switch trigger are provided
Notes
- In regulated insurance and lending contexts, a monotonic-constrained GBM with SHAP or a well-specified GLM/GAM is usually the practical interpretability ceiling — recommend accuracy gains that survive the filing, not the leaderboard.
- Censoring is the most common silently-broken assumption: any time-to-event target with incomplete follow-up must go to a survival method, even if the analyst framed it as regression.
- When frequency and severity both matter (claims modeling), consider whether a single Tweedie GLM or a two-part frequency-severity split serves the business decision better — this depends on whether the two components need separate governance.
- Prefer the simplest model that meets the accuracy and interpretability bar; escalate model complexity only when a diagnostic shows the simpler class is underfitting.
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