Take-Home Tactics — The Hidden Rubric

# Take-Home Tactics — The Hidden Rubric Most DA roles include a take-home. You're given a dataset and 2–5 days. The rubric is rarely shared. It's almost always: 1. Did you answer the question that was asked? (40%) 2. Is your code clean and reproducible? (20%) 3. Did you communicate the result clearly? (25%) 4. Did you go beyond the basic ask? (15%) If you optimize for those four, you score much higher than candidates who spent the same time on a fancier model. ## (1) Answer the question that was asked Read the prompt twice. Most candidates answer a *related* question — usually a more interesting one — instead of the actual one asked. Don't be that candidate. If the question is "what factors are correlated with churn?", don't build a predictive model. They asked for correlation, not prediction. Re-read the prompt at the end of your analysis to make sure your answer matches. If your write-up's headline doesn't directly answer the question, rewrite the headline. ## (2) Clean, reproducible code - Single notebook or script that runs end-to-end without manual steps. - Clear cell/section structure: load data → clean → explore → analyze → conclude. - Variable names that mean something (`active_users`, not `df_temp_3`). - A README with 2 lines: "how to run this" and "what's in this". Reviewers grade reproducibility heavily — if your code doesn't run on their machine, they DQ you regardless of the analysis quality. Test it in a clean Python environment before submitting. ## (3) Clear communication A 1-page summary at the top of your submission. The summary should answer: - What was the question? - What's the answer? (Lead with this.) - How confident are you? Why? - What would you do next if you had more time? This 4-bullet summary is what the hiring manager reads first. Often it's all they read. If your headline answer is buried on page 7, you've already lost. ## (4) Going beyond — the right way The temptation: more charts, fancier models. The right move: domain insight. - If the data shows churn correlates with low usage, don't stop there. Suggest *which usage thresholds* matter, ranked by lift on churn. - If signups are dropping, don't just say "investigate". Pre-run the most likely hypotheses and rule some out. - Spot something the prompt didn't ask about but is interesting. Mention it in 2 sentences under "Bonus observations". Going beyond looks like "I noticed X while doing the main analysis — worth investigating later." Not "here's a 5-model ensemble I built unprompted." ## Time allocation (for a 4-day take-home) - **Day 1:** Read the prompt twice. Load data. EDA. Form 2–3 hypotheses. - **Day 2:** Main analysis. Get to a *draft* answer. - **Day 3:** Refine. Add the bonus observations. Clean code. - **Day 4:** Write the summary. Test reproducibility. Re-read the prompt. Submit. Candidates who burn day 1 on EDA and day 4 on the model usually fail to write a tight summary. Inverted: start with the answer, polish backwards. ## The submission checklist Before you hit send: - [ ] Headline answer matches the prompt's question word-for-word. - [ ] 1-page summary at the top. - [ ] Code runs end-to-end in a fresh environment. - [ ] No commented-out cells, no exploratory junk in the final notebook. - [ ] One "bonus" observation worth 2 sentences. - [ ] README with 2 lines. ## Takeaway A take-home is a structured test, not an open creative writing prompt. Hit the rubric. Save the creativity for the bonus observations.