By name: modeling-strategy-review description: Choose, review, or debug data science modeling strategy. Use when Codex needs to decide between inference, prediction, causal, clustering, topic modeling, graph, time-series, or baseline approaches; review model assumptions; prevent leakage and overfitting; design train/validation/test splits; interpret coefficients or metrics; or calibrate claims against evidence.
Modeling Strategy Review
Match the model to the question type and the data-generating constraints.
Domain Context Requirement
Use the Domain Context Contract as controlling context for model choice. Adapt target definition, feature provenance, validation split, metric, interpretability level, deployment assumptions, and failure analysis to the stakeholder decision and domain constraints. If the contract is missing for a business/domain task, do not finalize modeling recommendations.
Strategy
- Restate the domain contract fields that constrain modeling: unit, target/KPI, decision, success metric, operational constraints, and prohibited claims.
- Start from the framed question, not from a preferred algorithm.
- Choose the simplest baseline that can answer the question and the stakeholder decision.
- Define the target, features, unit of prediction or inference, and time boundary.
- Decide whether the goal is explanation, estimation, prediction, discovery, ranking, clustering, or compression.
- Pick validation that matches the deployment, hidden-test, operational, or inference scenario.
- Check assumptions and failure modes before interpreting results.
Required Checks
- Baseline comparison exists.
- Train/validation/test split avoids leakage and temporal contamination.
- Feature availability matches prediction time.
- Metrics match the cost of false positives, false negatives, calibration, ranking, or estimation error.
- Uncertainty is reported for inferential estimates.
- Causal claims require a causal design, not only a statistically significant model.
- Error analysis covers slices, outliers, missingness, and relevant groups.
Output Template
Domain context used:
Question type:
Modeling goal:
Baseline:
Candidate methods:
Split/validation:
Metrics:
Assumptions:
Leakage checks:
Diagnostics:
Interpretation limits:
Recommended next action: