Status: UNPUBLISHED draft. Do not enroll learners until the P5 capstone milestone placeholder (
data-scientisttrack,is_required=False) is flipped to required. Ref: DS track seedbackend/scripts/seed_career_track_data_scientist.py.
1. Objective
An end-to-end, interview-defendable capstone in ONE chosen domain (fintech risk, e-commerce growth, or SaaS retention): frame the business problem, analyze messy data, ship at least one predictive or causal artifact, and present the recommendation as if to a hiring manager.
2. Deliverables
- Scoping memo (≤1 page): decision to inform, success metric, constraints.
- EDA notebook: data quality audit + ≥ 3 decision-relevant findings.
- Model or causal artifact: trained/evaluated model OR experiment analysis with power/sample-size reasoning.
- Stakeholder memo (≤2 pages): recommendation, risks, what you would do next.
- 10-minute recorded walkthrough defending all of the above.
3. Constraints
- One domain only; depth beats breadth.
- All claims traceable to a table/figure/model output in the submission.
- At least one "what would change my recommendation" sensitivity check.
4. 100-Point Rubric
| # | Criterion | Points |
|---|---|---|
| C1 | Scoping: decision, metric, and constraints crisply defined | 10 |
| C2 | Data quality audit (missingness, joins, outliers, definitions) | 15 |
| C3 | EDA: ≥ 3 findings that actually move the recommendation | 15 |
| C4 | Model/causal artifact: correct methodology, honest evaluation | 20 |
| C5 | Sensitivity check: one credible challenge to the recommendation | 10 |
| C6 | Stakeholder memo: clear recommendation + risks + next steps | 15 |
| C7 | Walkthrough: defends choices, answers "why not X" credibly | 10 |
| C8 | Reproducibility (pinned env, re-runnable notebook/scripts) | 5 |
| Total | 100 |
Pass: ≥ 80 overall AND C4 ≥ 14/20 (methodology is a hard gate).