---
name: Data Deliverable Checklist
description: Hit every literal requirement a data-science deliverable/grade checks for.
---
# Data Deliverable Checklist

Data-science deliverables (feature docs, analyses, model designs) are graded on the feature set, the leakage analysis, and the encoding approach. Run this checklist before submitting.

## Universal checks
- **Named addressee** if the brief implies one.
- **Word count** inside any band.
- **Labeled sections** using the brief's exact names.

## Feature engineering specific
- **State the prediction target** explicitly (e.g. "churn within 30 days").
- **Propose at least the number of features the brief requires** (e.g. ≥6) — count them.
- **For EACH feature**: name, source columns, derivation logic, and WHY it's predictive. All four. A feature without the "why" loses the point.
- **Include a time-based feature** (days since last login, tenure) — graders check for this explicitly.
- **Address data leakage** explicitly: state that no feature uses information unavailable at prediction time. Call out any feature that could leak (e.g. "order_count_30d uses the full window which is known only at period end — we use order_count as of prediction timestamp").
- **Address categorical encoding** for every categorical column (country, plan_tier): one-hot for low cardinality, target/frequency encoding for high cardinality.
- **Address missingness**: how each feature handles nulls.
- **State the evaluation approach** (time-based split, metric, baseline).

## Final pass
Re-read the brief; tick each required feature attribute (name+source+logic+why), the leakage statement, and the encoding approach. Omitting the "why predictive" or the leakage analysis are the most common point-losers.