How to use these practice questions
Data science assessments vary: some emphasise statistics, others programming, machine learning or applied case studies. These questions cover common foundations, but they do not represent a universal exam format. Check your course outline or assessment guidance to find out which topics and question styles you need.
Choose an answer for each question before reading the explanation beneath it. For every miss, name the underlying idea—not just the correct option—and add that idea to your revision list. A useful review question is: what information would make my chosen answer correct, and why is that not true here?
The set moves from data quality to modelling, statistical interpretation and querying. Each item is a single-best-answer question. The marked question screen shows how a practice item can include an explanation alongside the answer.
A single question on prediction-time data ends up looking like this:
A team is building a model to predict whether a customer will miss a loan payment in the next 30 days. One feature, last_payment_status, is populated only after the payment deadline has passed. What is the best response?
A predictor must be available at the time the model is used. A status recorded after the deadline leaks information about the outcome; strong correlation does not make it valid. Keeping it out of training but using it in testing would also make evaluation inconsistent.
The key check is the prediction-time boundary: could the system know this value when it needs to make the prediction? The strongest distractor is to keep a highly predictive feature, but predictive power from future information produces an unrealistic evaluation and a model that cannot work as intended.
Metrics and model evaluation
The next questions test whether you can select a metric that matches the cost of errors, and whether you can keep evaluation data separate from the fitting process.
A question on an imbalanced classification problem ends up looking like this:
A screening dataset contains 1,000 people, of whom 10 have the condition. A classifier labels every person negative and achieves 99% accuracy. Which metric directly measures the proportion of people with the condition that the classifier correctly identifies?
Recall is true positives divided by all actual positives. Here the classifier has zero recall despite 99% accuracy. Precision measures the proportion of predicted positives that are correct; it cannot describe how many actual cases were found.
This item distinguishes overall correctness from finding the positive class. Accuracy can conceal poor performance when one class is rare. Precision is a plausible distractor, but it answers a different question: when the model predicts positive, how often is it right?
A question on regression metrics ends up looking like this:
Two regression models have similar errors on most observations. One model makes a single very large error. Which metric is more strongly affected by that large error?
RMSE squares each residual before averaging, so a very large error contributes disproportionately; taking the square root does not remove that effect. MAE uses absolute residuals, so its penalty grows linearly. The strongest distractor, MAE, is often easier to interpret but is less sensitive to an outlier-sized error.
Match the metric to the decision. If large misses are especially costly, the squared-error penalty may be useful; if you want an error measure less dominated by extreme residuals, MAE may be easier to interpret.
A question on cross-validation ends up looking like this:
You use five-fold cross-validation to compare models. Each feature is standardised by subtracting its mean and dividing by its standard deviation. How should the scaling be performed to avoid information leaking from validation folds?
For each fold, estimate the scaling parameters from that fold's training data only, then apply those parameters to its validation data. Scaling the full dataset first lets validation-fold information influence preprocessing. The validation fold must not be used to fit its own scaler either.
Treat preprocessing as part of the model-fitting pipeline. The same rule applies to operations such as imputation and feature selection: fit them on training data, then apply them to held-out data.
Interpreting models and statistical results
These questions check the meaning of a coefficient, a p-value and a dimensionality-reduction choice. Focus on what each quantity says—and what it does not say.
A question on logistic-regression coefficients ends up looking like this:
In a logistic regression model, a predictor's coefficient is 0.7. Holding other predictors constant, what does a one-unit increase in this predictor imply for the odds of the outcome?
Logistic regression models log-odds, so a one-unit increase multiplies the odds by exp(0.7), approximately 2.01. The coefficient is not a direct change in probability, which depends on the starting probability and other predictors. A 70-percentage-point increase is therefore not justified.
Translate the coefficient from log-odds to an odds ratio by exponentiating it. Avoid describing that odds ratio as a fixed change in probability: the probability change depends on the starting point.
A question on hypothesis testing ends up looking like this:
A test of a null hypothesis returns p = 0.03. Which interpretation is correct, assuming the test's model and assumptions are appropriate?
The p-value is the probability, under the null hypothesis and the test assumptions, of results at least as extreme as those observed. It is not the probability that the null is true, nor the probability that the alternative is true. It also does not measure practical importance.
The strongest distractor is the claim that the p-value gives the probability the null is true. It does not: that would require a different analysis and assumptions about prior probabilities. Statistical evidence and practical importance should be assessed separately.
A question on principal component analysis ends up looking like this:
A dataset contains annual income measured in thousands of dollars and age measured in years. You plan to use principal component analysis based on variance. What is the most appropriate first step if both variables should contribute on a comparable scale?
PCA based on variance is affected by feature scale. Standardising puts variables on a comparable scale before calculating components. Multiplying income by 1,000 would make its numerical variance even larger, while removing a feature solely because of its units discards information.
Check whether scaling is appropriate for the analysis and the meaning of the variables. PCA on unscaled features can be dominated by the variable with the largest numerical variance, even when that variance mainly reflects its units.
A question on regularisation ends up looking like this:
Which regularisation method can set some fitted coefficients exactly to zero, making it useful for selecting a sparse set of features?
L1 regularisation adds a penalty based on the absolute values of coefficients and can drive some coefficients to exactly zero. L2 penalises squared coefficients and generally shrinks them towards zero without setting them exactly to zero. Scaling and one-hot encoding are preprocessing choices, not regularisation penalties.
Remember the distinction: L1 can produce sparse coefficients; L2 generally shrinks coefficients. Whether either improves a model should be judged with appropriate validation, not assumed from the method's name.
Experiments and working with data
The final two questions test cautious interpretation of an experiment and a basic grouped SQL query. In practice, confirm the study design and the exact schema before drawing conclusions or running code.
A question on confidence intervals ends up looking like this:
An A/B test estimates the difference in conversion rates (B minus A) as 1.1 percentage points, with a 95% confidence interval from -1.2 to 3.4 percentage points. Which conclusion is best supported by this interval?
The interval includes zero and values on both sides of zero, so it does not establish a direction at the corresponding two-sided 5% level. That is not proof of no effect. The observed estimate is 1.1 points, but it is not certain to equal the true effect; the usual confidence-interval interpretation is not a 95% probability statement about this fixed interval.
The strongest distractor is “no effect”: an interval spanning zero leaves uncertainty; it does not establish that the effect is exactly zero. Also distinguish the observed estimate from the range of values compatible with the data under the method.
A question on aggregation ends up looking like this:
A table named orders has one row per order, with columns customer_id and order_value. Which query returns the average order value for each customer?
Grouping by customer_id creates one group per customer, and AVG(order_value) calculates the mean within each group. Option B selects a customer alongside an aggregate without grouping by customer, so it does not produce the requested grouped result in standard SQL. Option D calculates a sum, not an average.
For grouped summaries, check that the grouping column matches the requested unit of analysis and that the aggregate matches the question. A sum and an average can both be valid summaries, but they answer different questions.
Turn misses into a revision plan
After completing the set, classify each miss: data leakage, metric selection, validation, model interpretation, inference or query logic. Revisit the relevant notes and then attempt a new question on that skill. Keep a short explanation of the mistake; memorising the option letter will not transfer to a differently worded problem.
A weak-area summary can help decide what to practise next. The percentages and attempt counts below represent a sample progress view, not exam statistics.
A review panel for these question topics ends up looking like this:
If the same idea is missed more than once, make the correction concrete: state the rule, then apply it to a new example. A card in a remediation tray might turn a repeated leakage mistake into a direct retrieval prompt.
A card prompted by a repeated mistake ends up looking like this:
You lost this mark twice: when fitting a scaler during cross-validation, which data supplies the mean and standard deviation?
How MySummaries helps
Build a revision board from your own data science notes, then use it to generate practice questions, flashcards and other revision formats. Review questions you miss and return to the relevant material on your board. Open MySummaries to organise your study materials.