Data Scientist question bank
100 questions. Filter by level and skill, track what you've solved, and see how many people get each one right.
- Typical income in a skewed dataset
- Meaning of a p-value
- Ice cream and drowning
- Counting orders per customer
- Customers with no orders
- Type I error
- Showing a distribution
- What standard deviation tells you
- Positive test for a rare disease
- Expected profit of a dice game
- Probability of A or B
- Values within two standard deviations
- Median of six numbers
- Z-score of an exam result
- Units of variance
- Exactly two heads in three flips
- Precision from a confusion matrix
- Metric when missing a case is costly
- High accuracy on rare events
- Why keep a test set
- High train score, low validation score
- Counting missing values in pandas
- pandas mean with a missing value
- value_counts with normalize
- Filtering groups in SQL
- COUNT of a column with NULLs
- Encoding a city column
- Scaling features for kNN
- Choosing the number of clusters
- Chart for a monthly trend
- A correlation of -0.9
- Reading a regression coefficient
- Survey sent only to app users
- What the central limit theorem says
- Sample size for a smaller effect
- groupby with transform
- Reading a 95% confidence interval
- Testing 20 metrics at once
- Latest order per customer
- Overall rate flips when split
- Predicting a heavily skewed target
- Studying only active users
- Users needed to detect a 1 point lift
- What statistical power means
- Comparing means with unequal variances
- Testing two categorical variables
- Bonferroni threshold for ten tests
- What an AUC of 0.8 means
- F1 from precision and recall
- Which prediction hurts log loss most
- RMSE with one large error
- Row count after a merge
- groupby with a missing key
- Ranking with ties in SQL
- NOT IN returns no rows
- Suspiciously strong churn feature
- Validating a forecasting model
- Making a trending series stationary
- Why a forest beats one tree
- Lowering the boosting learning rate
- Regularisation that zeroes coefficients
- PCA on unscaled features
- Using a model on rare positives
- Brier score for two predictions
- Halving a confidence interval
- Lift of the top decile
- A lift that fades after week one
- PM wants to ship an unclear win
- Checking results every day
- Coaching the worst stores
- Unexpected group sizes
- Using pre-experiment data
- Difference-in-differences assumption
- Coefficients that keep flipping
- Feature users retain better
- Click-through rate standard errors
- Propensity scores near 0 and 1
- Key assumption of an instrument
- A link that only appears in a subset
- Error spread grows with the prediction
- Direction of omitted variable bias
- AUC after squaring the scores
- Precision when prevalence drops
- Oversampling before cross-validation
- Target encoding without leakage
- NDCG at 3 for a ranking
- Treated drivers take control rides
- Benjamini-Hochberg discoveries
- Noisy feature, flatter slope
- Back-transforming log predictions
- Effect with an interaction term
- From log-odds to probability
- Shape after NumPy broadcasting
- Rolling mean at the edges
- LEFT JOIN that acts like INNER JOIN
- Average lifetime with active customers
- Great backtest, poor live forecast
- When the bootstrap breaks down
- Optimising time on site
- Big effect in one segment