Top 30 Forecasting Interview Questions and Answers
30 multiple-choice questions on Forecasting, drawn from 30 bites out of the 54 tagged Forecasting on Tezvyn. Answer them here or read straight down. Every question carries the correct option, why it is correct, and a link to the bite it came from.
30 questions. Pick an answer, or open “Show the answer” to read it.
Answers are graded in your browser. Nothing is saved, and no XP or streak is earned here. The app keeps score.
Question 1 of 30
In Scrum, how should the Scrum Team decide how much work to select for the Sprint Backlog during Sprint Planning?
Show the answer
Answer: a · By using their collective expertise and observed experience with the ordered Product Backlog items
The Scrum Guide defines Sprint Backlog selection as an empirical forecast based on the team's collective expertise and observed experience with the ordered Product Backlog. Citing velocity or capacity hours as required inputs is incorrect because the framework does not prescribe those tactics, and treating the forecast as a hard commitment contradicts Scrum's empirical foundation.
Read the full bite: How is work selected and forecasted for the Sprint Backlog?
Question 2 of 30
In a production time-series forecasting pipeline using rolling-origin validation, which approach correctly prevents future leakage?
Show the answer
Answer: a · Use backward-looking windows ending at t minus one and fit scalers exclusively on each training fold before transforming the matching validation fold
Backward-looking windows ending at t minus one ensure no future data enters features, and fitting preprocessing per training fold stops global statistics from leaking into validation. Option D is tempting because global scaling is standard in non-temporal ML, yet it embeds future information into every historical row before any split occurs.
Read the full bite: How do you prevent future leakage in time-series preprocessing?
Question 3 of 30
In which scenario is RMSE preferred over MAE for measuring forecast accuracy?
Show the answer
Answer: a · When large misses are disproportionately costly to the business
The card recommends RMSE when large forecast misses are disproportionately costly because squaring errors heavily penalizes large deviations. Option C is tempting because it describes an outlier scenario, but the card explicitly pairs one-time anomalies with MAE since it is robust and does not let a single error dominate the metric.
Read the full bite: How do you measure forecast accuracy and compare MAE to RMSE?
Question 4 of 30
When evaluating a sales forecast model where a large error is significantly more damaging to the business than a small one, which metric is most appropriate?
Show the answer
Answer: a · Root Mean Squared Error (RMSE), because it disproportionately penalizes large, costly errors.
RMSE is correct because by squaring errors before averaging, it gives much more weight to large mistakes, aligning with the business need to avoid them. While MAE is more interpretable, it treats the business cost of errors as linear, which is not the case here.
Read the full bite: How would you measure a sales forecast model's accuracy?
Question 5 of 30
When evaluating a sales forecast model, which situation most strongly suggests prioritizing Root Mean Squared Error (RMSE) over Mean Absolute Error (MAE)?
Show the answer
Answer: c · A few significantly large forecast errors lead to disproportionately severe business consequences, like production halts.
RMSE is preferred when large errors are disproportionately costly because its squaring step heavily penalizes them, as described in option C. Option B describes a linear cost scenario, where MAE would be more appropriate.
Read the full bite: How would you measure a sales forecast model's accuracy?
Question 6 of 30
A time series of company revenue shows a clear upward trend. To prepare this data for an ARIMA model, what is the most appropriate initial transformation to apply?
Show the answer
Answer: d · Perform first-order differencing on the series.
Differencing is the standard method to remove a trend and make a series stationary. While a log transform is a common distractor and can stabilize variance, it does not by itself remove the underlying trend.
Read the full bite: Explain time series stationarity and how to achieve it
Question 7 of 30
Which best describes a weakly stationary time series?
Show the answer
Answer: a · Constant mean and variance, with autocovariance depending only on lag
Weak stationarity requires constant mean and variance plus lag-only autocovariance. 'No trend' is necessary but not sufficient, and stationarity does not require independence.
Read the full bite: Stationarity in time series and why ARIMA needs it
Question 8 of 30
What is the primary reason for ensuring a time series is stationary before applying an ARIMA model?
Show the answer
Answer: b · Non-stationary data violates the fundamental assumptions of ARIMA, leading to inaccurate and unreliable forecasts.
The card emphasizes that ARIMA models assume stable statistical properties. Non-stationary data violates these assumptions, making the model's forecasts unreliable. While a constant mean is part of stationarity, it doesn't have to be zero, and stationarity ensures constant autocorrelation, not its complete removal.
Question 9 of 30
Which statement best explains why random train-test splits are unreliable for evaluating forecasting models?
Show the answer
Answer: b · They allow the model to train on data from time periods that occur after the test set observations.
Random splits ignore chronological order, allowing the model to learn from future observations and producing an unrealistic performance estimate. The most tempting distractor confuses the purpose of walk-forward validation: its primary goal is preventing temporal leakage, not simply increasing the number of test folds.
Read the full bite: Random split vs walk-forward validation in forecasting
Question 10 of 30
When forecasting daily data with both an upward trend and a stable weekly cycle, why would an analyst prefer Holt-Winters triple exponential smoothing over Holt's double exponential smoothing?
Show the answer
Answer: c · Double exponential smoothing estimates level and trend but lacks the seasonal equation needed to capture a repeating 7-day deviation.
Holt-Winters triple smoothing adds a seasonal equation with period m=7 to the level and trend updates, while Holt's double smoothing only handles level and trend. Option D is tempting but wrong because it repeats the common misconception that classical models cannot handle both trend and seasonality, when triple exponential smoothing is specifically designed as a classical baseline for this exact situation.
Read the full bite: Which classical baseline model handles weekly seasonality and upward trend?
Question 11 of 30
Sales data shows a steady upward trend and a recurring spike of about 15% each December. Which Holt-Winters configuration is most appropriate for this time series?
Show the answer
Answer: c · A model with a trend component and multiplicative seasonality.
The correct choice includes a trend component to capture steady growth and multiplicative seasonality because the spike is proportional (15%) to the sales level. An additive model would be for a fixed-amount spike (e.g., +1,000 units).
Read the full bite: Which model for forecasting with seasonality and trend?
Question 12 of 30
When forecasting inventory with an upward trend and weekly seasonal patterns where seasonal impact scales with overall demand, which Holt-Winters method is best?
Show the answer
Answer: a · Holt-Winters Multiplicative Seasonal Method
The card states that for demand forecasting, where seasonal fluctuations tend to scale with the series level, the multiplicative Holt-Winters method is often preferred. The additive method assumes a fixed unit seasonal lift, which is less realistic for growing demand.
Read the full bite: Forecasting inventory with trend and weekly seasonality?
Question 13 of 30
A model's residuals are found to be heavily skewed. Which method is most appropriate for generating a robust prediction interval for its forecasts?
Show the answer
Answer: d · Bootstrapping the residuals to simulate future outcomes, as this does not assume a normal error distribution.
The analytical method assumes normally distributed errors, which is violated when residuals are skewed. Bootstrapping is a non-parametric method that does not require this assumption, making it more robust.
Read the full bite: Describe two methods for generating prediction intervals
Question 14 of 30
If historical forecast residuals are heavily right-skewed and the target variable is bounded below by zero, which method for generating a 95% prediction interval is most appropriate and why?
Show the answer
Answer: c · Bootstrap residual resampling because it preserves empirical skewness and respects natural bounds through simulated sample paths
Bootstrap residual resampling builds empirical percentiles from simulated paths, naturally accommodating skewness and lower bounds without normality assumptions. Option B is tempting but wrong because confidence intervals quantify parameter uncertainty, not the range where a future observation will fall.
Question 15 of 30
When constructing prediction intervals, under which circumstance would the bootstrapping residuals method typically be preferred over the method assuming normally distributed errors?
Show the answer
Answer: c · When the forecast errors exhibit a clear non-normal distribution or heteroscedasticity.
The card explicitly states that bootstrapping residuals is "more robust if errors are not normally distributed" and should be chosen "When residual diagnostics show errors are not normally distributed or are not homoscedastic." Option B describes a confidence interval, which the card warns against confusing with a prediction interval.
Read the full bite: Describe two methods for generating prediction intervals
Question 16 of 30
Which approach best fits a beginner 30-day DAU forecast using timestamped event logs?
Show the answer
Answer: a · Start with a regression baseline using day-of-week and recent daily totals, evaluating with MAE
The card prescribes an interpretable regression baseline using temporal features and recent totals scored with MAE before considering complexity. Option B is wrong because the card explicitly red-flags both deep learning and real-time online inference as over-engineering for a simple 30-day batch forecast.
Read the full bite: What data and approach for a simple 30-day DAU forecast?
Question 17 of 30
What is the core flaw in forecasting annual users by compounding 5% weekly growth?
Show the answer
Answer: a · It assumes the weekly rate persists indefinitely while ignoring market saturation and seasonality
The core flaw is assuming a constant growth rate will continue unchecked despite market constraints, seasonality, and channel exhaustion. Distractor D is tempting because the 12x multiplier sounds alarming, but the arithmetic is correct; the problem is the unrealistic business assumption behind it.
Read the full bite: Why is extrapolating 5% weekly growth naive for annual forecasting?
Question 18 of 30
What is the fundamental nature of the "done" prediction provided by Actionable Agile Metrics?
Show the answer
Answer: b · A probabilistic forecast of completion, based on historical flow data and accounting for uncertainty.
Actionable Agile Metrics provide a probabilistic forecast, acknowledging uncertainty, rather than a single, fragile date, as stated in the card. Option A is a tempting distractor because it mentions using historical data for an average, but the card explicitly warns against "misleading averages" and advocates for understanding the true distribution of delivery times.
Read the full bite: Actionable Agile Metrics: Predicting 'Done'
Question 19 of 30
Which scenario most strongly suggests using a multiplicative time series decomposition model?
Show the answer
Answer: a · The magnitude of seasonal variations increases proportionally with the overall level of the series.
A multiplicative model is appropriate when the seasonal effect scales with the series' overall level, meaning its absolute magnitude grows as the trend grows. Option D describes the condition for an additive model, where seasonal effects have a constant absolute impact.
Read the full bite: Time Series Decomposition: Separating Signal from Noise
Question 20 of 30
When forecasting daily sales, how should you model a recurring campaign that gradually ramps up and fades down around its start date?
Show the answer
Answer: a · Use a windowed or decay-function regressor covering the ramp-up and fade periods
A windowed regressor captures the campaign's gradual effect shape, while a single-day binary flag misses the ramp-up and fade. Removing days destroys repeatable signal, and post-hoc adjustments break confidence intervals and scalability.
Read the full bite: How would you incorporate external events into a time-series forecast?
Question 21 of 30
Why is a prediction interval for next quarter's revenue typically wider than a confidence interval around the model's estimated growth parameter?
Show the answer
Answer: d · It captures both parameter uncertainty and the irreducible variability of an individual future outcome
A prediction interval must cover a single future realization, so it adds the outcome's own random variation on top of parameter uncertainty, making it wider. It is not merely a higher confidence level, and it explicitly includes residual error.
Read the full bite: Communicate forecast uncertainty with prediction intervals
Question 22 of 30
When reconciling divergent top-down and bottom-up forecasts, what is the most appropriate first step before applying any weighting or ensemble technique?
Show the answer
Answer: c · Map where the gap originates by examining the independent assumptions behind each model
Sophisticated reconciliation always begins by diagnosing the source of divergence, such as differing assumptions about market growth versus cohort retention, before any quantitative blending is applied. Jumping straight to inverse-variance weighting skips this structural diagnosis and can mask the true drivers of disagreement.
Question 23 of 30
How does linear regression determine the "best-fit" straight line through data points?
Show the answer
Answer: c · By minimizing the sum of the squared differences between observed and predicted values.
The card states that the algorithm "minimizes the total squared distance between the actual data points and the regression line," which is the definition of least squares estimation. While correlation is related to the strength of the fit, it is not the method used to determine the line itself.
Read the full bite: Linear Regression: Forecasting with a Straight Line
Question 24 of 30
For which scenario would applying the Bass Diffusion Model be LEAST appropriate?
Show the answer
Answer: b · Modeling the market penetration of a completely novel technology with no historical data.
The Bass Diffusion Model requires historical data from similar products to estimate its 'p' and 'q' coefficients. For a completely novel technology without market parallels, these essential parameters cannot be reliably estimated, making the model fundamentally unsuitable. While it is also not recommended for short-term forecasting (option D), that is primarily because simpler methods are more efficient, not due to a fundamental inability to apply the model's core mechanics.
Read the full bite: The Bass Diffusion Model: Innovators vs. Imitators
Question 25 of 30
Which statement accurately describes the core principle behind how Prophet forecasts time series data?
Show the answer
Answer: b · It models the time series as a sum of distinct components, including trend, seasonality, and holidays.
The card explicitly states, "Think of Prophet as a 'forecasting decomposer.' It assumes a time series signal is the sum of three main components: a general growth or decline trend, repeating seasonal patterns... and the effect of specific holidays." This directly describes option B. Option C is incorrect because Prophet's approach is to decompose the series into separate additive components, not to use a single, unified model for everything.
Read the full bite: Prophet: Time Series Forecasting for Seasonal Data
Question 26 of 30
Under which circumstance would time-series forecasting be LEAST reliable for predicting future outcomes?
Show the answer
Answer: a · The system being modeled is expected to undergo significant, unforeseen structural changes.
Time-series forecasting relies on the assumption that past patterns will continue into the future. If the underlying system experiences sudden, unpredictable changes, this assumption is violated, making the forecast unreliable. The other options describe conditions where time-series forecasting is typically effective.
Read the full bite: Time-Series Forecasting: Predicting the Future from the Past
Question 27 of 30
When would a Monte Carlo simulation be least appropriate for forecasting a system's outcome?
Show the answer
Answer: d · There is no reliable basis to define probability distributions for the key input variables.
The card explicitly states that Monte Carlo simulations are inappropriate when there is no reasonable basis for defining input probability distributions, as this leads to 'garbage in, garbage out.' Options A, B, and D describe scenarios or characteristics where Monte Carlo simulation is highly beneficial or inherent to its operation.
Question 28 of 30
To create a reliable forecast for completing 40 stories, which method best incorporates the team's historical delivery variability?
Show the answer
Answer: d · Run a Monte Carlo simulation using historical weekly throughput to generate a distribution of probable completion dates.
A Monte Carlo simulation is the correct method because it uses the full range of historical throughput data to model thousands of possible future outcomes, thus accounting for variability. Calculating a simple average provides a single-point estimate that ignores the reality of delivery variation.
Read the full bite: How would you create a probabilistic forecast for 40 stories?
Question 29 of 30
To provide a probabilistic forecast for a 40-story backlog of similarly-sized items, which approach is most effective?
Show the answer
Answer: a · Gather historical weekly throughput data, run a Monte Carlo simulation by randomly sampling from this data, and present completion dates with associated confidence levels (e.g., 50%, 85%).
The most effective approach for a probabilistic forecast involves using historical throughput data with a Monte Carlo simulation to generate a range of completion dates with associated confidence levels. Option C provides only a single, deterministic estimate, ignoring variability, while Option D uses story points which are less appropriate for similarly-sized items where throughput is a more direct measure.
Read the full bite: How would you create a probabilistic forecast for a backlog?
Question 30 of 30
A team needs to forecast when 40 similarly-sized stories will finish using ten weeks of historical throughput. Which approach best applies probabilistic forecasting?
Show the answer
Answer: c · Use the historical throughput values to run a Monte Carlo simulation, summing randomly drawn weekly throughputs until reaching 40 stories, and present the results as percentile delivery curves.
Monte Carlo resampling of historical throughput captures natural week-to-week variance and produces percentile curves that let stakeholders choose their risk appetite. Simply averaging throughput and adding a buffer still yields a single deterministic date that ignores the actual distribution of historical performance.
Read the full bite: How would you probabilistically forecast 40 stories using throughput data?
Could you explain these out loud?
That is what an interview actually tests. Tezvyn gives you questions like these with what the interviewer is really checking, the answer that lands, and the mistake that ends the conversation, in the four minutes before your next meeting.
The iPhone app is on the way
We are building it. Until it lands, nothing here is held back from you: every interview card, your saved cards, streaks and the job board all work in Safari, plus hundreds of free practice quizzes of thirty questions each. Sign in and it all carries over to the app the day it arrives.
Want it as an icon? Tap Share at the bottom of Safari, then Add to Home Screen. It opens full screen and the cards you have read stay available offline.