Data Scientist Interview Guide

Sample Answer:

I discovered our most profitable segment had lowest retention. I created simple visualization showing revenue vs. churn, explained onboarding gaps, and proposed solutions. Framing as actionable strategy worked better than statistical analysis.

Sample Answer:

Marketing assumed younger users engaged more. I analyzed cohorts finding older users engaged 3x more. I presented data and proposed budget reprioritization, saving $200K annually.

Sample Answer:

I analyzed customer acquisition cost payback by channel. Organic search had 3-month payback vs. 8 months for paid. Budget shift to organic resulted in 25% CAC drop and 15% revenue growth.

Sample Answer:

User behavior data had 30% missing from tracking gaps. I investigated patterns, found older devices tracked poorly, and used predictive imputation. Sensitivity analysis confirmed results were robust.

Sample Answer:

Customer segmentation hit data quality blocker. I spent time cleaning rather than pushing through, finding segments only after fixing quality. Early communication prevented surprises.

Sample Answer:

Junior analyst built models without proper train-test splits. I paired with them teaching cross-validation and explaining data leakage. After one month, they independently built robust models.

Sample Answer:

I read research papers, implement methods in side projects, discuss with colleagues. I focus on practical papers—A/B testing, causal inference, time-series—directly applicable to my work.

Sample Answer:

Churn analysis revealed 30% payment failures for one processor. No one had noticed. Switching processors reduced failures by 95% and improved retention by 8%.

Sample Answer:

Define hypothesis, choose primary metric with success criterion, randomize 50/50 split, calculate sample size (power analysis), run 2+ weeks, analyze with confidence intervals. Check assumptions, avoid peeking.

Sample Answer:

Correlation does not equal causation. Establish causality with randomized experiments, instrumental variables, difference-in-differences, propensity score matching. No method is perfect; discuss limitations.

Sample Answer:

One-hot encode for low cardinality, target encode for high cardinality, group rare categories, learn embeddings. Be careful with target encoding in trees (overfitting risk).

Sample Answer:

Challenges: temporal dependency (use ARIMA, SARIMA), seasonality (decompose), non-stationarity (difference), external variables (include regressors). Evaluate via backtesting. Long-term forecasts have increasing uncertainty.

Sample Answer:

Techniques: train-test split (80/20), k-fold cross-validation, stratified for imbalanced, time-series cross-validation. Detect distribution shifts. Use same preprocessing to prevent leakage.

Sample Answer:

Analyze query plan for sequential scans, add indexes, use window functions, denormalize if needed, partition tables, cache results. Test on production volumes.

Sample Answer:

Tree models have built-in importance, permutation importance shows performance drop when feature shuffled, SHAP decomposes predictions. SHAP is most intuitive for stakeholders.