Top 30 Amazon Data Analyst Interview Questions and Answers

I first investigate how many values are missing and whether the missingness affects important fields.

df.isnull().sum()

The treatment depends on the business context:

• Use median imputation for suitable numerical columns when outliers make the mean less representative.
• Consider removing records with missing critical identifiers when those records cannot be reliably linked.
• Use domain-appropriate methods for categorical variables.
• Avoid deleting records or filling missing values without understanding the impact.

For example, a missing customer ID may prevent reliable customer-level analysis, while a missing optional description may not affect the analysis.

Pandas provides pct_change() to calculate the percentage change between consecutive observations.

import pandas as pd

df["order_date"] = pd.to_datetime(df["order_date"])

monthly_revenue = (
    df.set_index("order_date")
      .resample("MS")["revenue"]
      .sum()
)

monthly_growth = monthly_revenue.pct_change() * 100

print(monthly_growth)

The code aggregates revenue by calendar month and calculates the percentage change from the previous month.

The first month's growth will be missing because there is no earlier month for comparison. If the dataset contains no records for some months, the analyst should ensure that the monthly series represents the intended reporting periods before interpreting the results.

The mean is the sum of all values divided by the number of observations. The median is the middle value after sorting the observations.

The mean can be strongly affected by extreme values, whereas the median is more resistant to outliers.

For example, if a small number of customers place exceptionally large orders, median order value may provide a more representative description of a typical order.

I would choose the measure based on the distribution, the business question, and whether the analysis concerns a typical observation or the overall average.

I would avoid loading the entire file into memory if it is too large for the available resources.

Possible approaches include:

• Reading the file line by line using a generator.
• Processing CSV files in chunks with Pandas.
• Using regular expressions to extract relevant log fields.
• Using distributed processing tools such as PySpark when the workload requires it.

For example:

import pandas as pd

for chunk in pd.read_csv("large_logs.csv", chunksize=100000):
    # Filter or aggregate the current chunk
    print(chunk.shape)

The right approach depends on file size, processing requirements, and available infrastructure.

A p-value is the probability of observing a test statistic at least as extreme as the one obtained, assuming the null hypothesis is true.

In an A/B test, the null hypothesis might state that there is no true difference in conversion rates between the control and treatment groups.

If the p-value is below a predefined significance threshold, such as 0.05, the result may be considered statistically significant under the test's assumptions.

However, a p-value is not the probability that the null hypothesis is true, nor does statistical significance automatically mean the result is practically important.

A p-value measures how compatible the observed data is with a specified null hypothesis under the test assumptions.

A confidence interval provides a range of plausible values for the parameter being estimated, along with a stated confidence level under the method's assumptions.

For example, an experiment may estimate a conversion-rate improvement and report a 95% confidence interval around that estimate.

The interval helps analysts evaluate both the precision of the estimate and whether the possible effect sizes are meaningful for the business.

I would begin by defining the primary metric, such as the percentage of eligible users who complete a purchase.

Then I would:

1. Verify the experiment's assignment process and sample counts.
2. Check for data-quality issues and sample ratio mismatch.
3. Estimate conversion rates for both groups.
4. Select an appropriate statistical test based on the metric, assignment unit, and assumptions.
5. Calculate the effect size and confidence interval.
6. Evaluate statistical significance alongside business impact and potential risks.

I would also check whether the experiment ran long enough to capture relevant behavioral patterns and whether repeated interim testing or multiple comparisons require special treatment.

Overfitting occurs when a model learns patterns specific to its training data, including noise, instead of generalizable relationships.

An overfitted model may perform well on training data but poorly on unseen observations.

Common ways to reduce overfitting include:

• Cross-validation.
• Holdout validation.
• Regularization.
• Simplifying the model.
• Using appropriate feature selection.

The goal is to build a model that generalizes to new data rather than merely memorizing historical observations.

Multicollinearity occurs when two or more independent variables in a regression model are strongly correlated.

It can make individual regression coefficients unstable and increase their estimated variance, making it harder to interpret each variable's independent contribution.

Possible responses include examining correlations, calculating the Variance Inflation Factor (VIF), removing redundant predictors when justified, or applying regularization methods such as Ridge regression.

The appropriate solution depends on whether the goal is interpretation, prediction, or both.

I would first examine the distribution and determine whether extreme orders are valid observations, data errors, or a meaningful part of customer behavior.

For heavily right-skewed order values, a small number of unusually large transactions can strongly influence the mean.

Possible approaches include:

• Comparing trimmed means.
• Considering a log transformation when appropriate.
• Using bootstrap methods to estimate uncertainty.
• Winsorizing at a prespecified percentile when justified.
• Conducting sensitivity analyses to determine whether a few extreme observations drive the conclusion.

I would clearly explain any change in the estimand. If the business decision specifically concerns raw mean revenue, I would not replace it with a different metric without explaining the trade-off.

I would first verify that the decline is real and that the tracking pipeline is working correctly.

Then I would break the problem into smaller stages:

1. Compare traffic and sign-ups with previous periods.
2. Segment results by region, device, acquisition channel, and visitor type.
3. Examine the funnel from landing-page visit to plan selection, payment, and confirmation.
4. Identify where the largest decline occurs.
5. Investigate possible causes such as payment failures, page performance, campaign changes, or user-experience issues.
6. Recommend an intervention and measure whether it improves the affected metric.

I would focus on identifying the controllable input metric responsible for the decline rather than simply recommending that the business increase sign-ups.

I would examine operational metrics across the fulfillment process, including:

• On-time delivery rate.
• Warehouse backlog.
• Picking and packing time.
• Shipment dispatch delays.
• Carrier handoff delays.
• Inventory availability and placement.

Next, I would segment the data by warehouse, product category, delivery route, and time period.

This would help determine whether the issue originates inside the fulfillment center, from inventory constraints, or during last-mile delivery. I would recommend immediate corrective action while also identifying longer-term process improvements.

I would evaluate whether the discount improves profitability rather than judging success only by sales volume.

My analysis would include:

1. Historical sales and demand patterns.
2. Product price elasticity.
3. Gross margin before and after the discount.
4. Inventory levels and holding costs.
5. Changes in units sold and average order value.
6. Cross-selling effects and potential changes in cart abandonment.

Where possible, I would use a controlled experiment or an appropriate comparison method to estimate incremental impact.

The final recommendation should consider incremental profit, inventory implications, and customer behavior rather than relying solely on revenue growth.

I would start by decomposing revenue into measurable business drivers.

For an e-commerce business, a simplified model is:

I would examine historical trends, seasonality, major promotional events, returns, and any known changes in customer behavior.

Depending on the data and forecasting requirements, I might use moving averages, seasonal decomposition, or a seasonal ARIMA model.

Finally, I would validate the forecast against a holdout period using appropriate error metrics such as MAE or RMSE, and explain the assumptions and uncertainty to stakeholders.

I would first establish a clear churn definition. For example, churn could mean subscription cancellation or no purchases for a specified period, depending on the business model.

Then I would:

• Compare churn across customer cohorts.
• Examine recency and frequency of purchases.
• Investigate payment failures and product-return behavior.
• Analyze engagement with relevant services.
• Identify patterns associated with customers who leave.
• Test targeted interventions where appropriate.

I would distinguish correlation from causation and evaluate whether proposed interventions reduce churn without creating undesirable effects elsewhere.

Output metrics measure the final business result, while input metrics measure factors that can influence that result.

For example:

Output metrics help determine whether the business is succeeding. Input metrics help teams identify actions that could improve performance.

In a business case, I would connect the outcome to specific operational drivers and explain how the team could measure the impact of changing them.

• Situation: Describe a project where an important business metric changed unexpectedly.
• Task: Explain what you were responsible for investigating.
• Action: Describe how you validated the data, examined relevant tables, identified the root cause, and communicated your findings.
• Result: State the actual business outcome, supported by measurable evidence where available.

Focus on the analytical steps you personally performed, not only the team's overall contribution.

Explain the deadline, the competing priorities, and the deliverable you were responsible for completing.

Describe how you assessed urgency, separated essential analysis from lower-priority work, communicated trade-offs, and checked the quality of the final output.

Conclude with the actual result: whether you met the deadline, what impact the work had, and what you learned.

Avoid claiming that you delivered successfully if the real outcome was different. A credible account of a difficult situation and the lessons learned is more useful than an exaggerated success story.

Describe a situation in which your analysis or professional judgment differed from a stakeholder's preferred approach.

Explain the evidence you used, how you presented your reasoning, and how you listened to the other person's concerns.

Then describe the decision that was made and how you supported the final direction.

This question can demonstrate Have Backbone; Disagree and Commit. The key is to show that you can challenge an assumption respectfully without turning the disagreement into a personal conflict.

Choose a genuine mistake with a clear consequence.

Explain what went wrong, your individual responsibility, and how you identified the underlying cause. Then describe the specific process change you introduced to reduce the chance of recurrence.

For example, an analyst might describe a reporting delay caused by unclear requirements and explain how a new intake and prioritization process improved subsequent deliveries.

Include evidence that the corrective action worked, if such evidence exists.

I would begin with the business question and the decision the stakeholder needs to make, rather than starting with statistical terminology.

I would explain the key finding in plain language, show the relevant evidence, and describe the limitations or uncertainty that could affect the conclusion.

For example, instead of simply reporting that a model has a particular R² value, I would explain what the model captures, what it does not explain, and how that affects its usefulness for the business decision.

I would then invite questions and confirm that the recommendation is understood.

Describe a real situation in which you noticed a problem that could affect a project's outcome but was not explicitly assigned to you.

Explain how you assessed the risk, coordinated with relevant colleagues, and took appropriate action without neglecting your core responsibilities.

Finish with the actual outcome and what the experience taught you about accountability.

This question tests Ownership and Deliver Results. Strong answers make your individual contribution clear and explain why you chose to act.