Chapter 16 Quiz: Time Series Forecasting

Chapter 16 Quiz: Time Series Forecasting

Multiple Choice

Question 1. Which of the following best distinguishes seasonality from cyclicality in a time series?

(a) Seasonality has larger amplitude than cyclicality.
(b) Seasonality has a fixed, known period; cyclicality has a variable, unknown period.
(c) Seasonality is caused by external factors; cyclicality is caused by internal factors.
(d) Seasonality can be modeled; cyclicality cannot.

Question 2. A retail company's December sales are consistently 25% above the annual average, regardless of whether the annual average is $1 million or $5 million. Which decomposition model is most appropriate?

(a) Additive, because the percentage is constant.
(b) Multiplicative, because the seasonal effect is proportional to the level of the series.
(c) Neither — seasonal effects that are proportional cannot be modeled with standard decomposition.
(d) Either additive or multiplicative would work equally well in this case.

Question 3. What does the "I" in ARIMA stand for, and what does it address?

(a) "Iterative" — the model iterates multiple times to find optimal parameters.
(b) "Integrated" — it refers to differencing the data to achieve stationarity.
(c) "Interpolated" — it fills in missing values in the time series.
(d) "Independent" — it ensures that residuals are independent of each other.

Question 4. In an ARIMA(2, 1, 0) model, which of the following is true?

(a) The model uses two past error terms and one differencing step.
(b) The model uses two past values, one differencing step, and no past error terms.
(c) The model uses no past values, one differencing step, and two seasonal periods.
(d) The model uses two differencing steps, one past value, and no past errors.

Question 5. In Holt-Winters exponential smoothing, what does the alpha parameter control?

(a) The flexibility of the trend component.
(b) The speed at which the seasonal pattern adapts to new data.
(c) The weight given to recent observations when estimating the level of the series.
(d) The number of seasonal periods to model.

Question 6. Which of the following was NOT identified in the chapter as a reason for Prophet's popularity in business?

(a) It handles missing data and outliers gracefully.
(b) It consistently achieves the highest accuracy of any forecasting method.
(c) It produces interpretable component plots.
(d) It generates prediction intervals by default.

Question 7. A Prophet model's changepoint_prior_scale parameter is set to a very high value (e.g., 0.5). What effect does this have?

(a) The trend becomes very smooth and resistant to change.
(b) The trend becomes highly flexible and can change direction frequently.
(c) The seasonal patterns become more pronounced.
(d) The prediction intervals become narrower.

Question 8. You are adding an external regressor to a Prophet model. Which of the following is a valid external regressor for a 30-day forecast?

(a) Actual daily temperature (observed after the fact).
(b) The company's planned promotional calendar for the next 30 days.
(c) GDP growth for the quarter (not yet reported).
(d) Competitor sales data (not available until the quarter ends).

Question 9. In the M4 forecasting competition, which finding was most notable?

(a) Deep learning methods consistently outperformed statistical methods.
(b) Statistical methods outperformed ML methods on most individual series, but ensembles performed best overall.
(c) ARIMA was the single best model across all series types.
(d) Prophet achieved the highest accuracy of any individual model.

Question 10. Tom built an LSTM that fit training data nearly perfectly but performed worse than Prophet on test data. This is an example of:

(a) Underfitting — the model was too simple for the data.
(b) Overfitting — the model memorized training data patterns, including noise.
(c) Data leakage — future information was accidentally included in training.
(d) Structural break — the data-generating process changed between training and testing.

Question 11. What is the primary difference between walk-forward validation and random k-fold cross-validation?

(a) Walk-forward validation uses more data for training.
(b) Walk-forward validation always trains on past data and tests on future data, respecting temporal order.
(c) Random k-fold is more appropriate for time series data.
(d) Walk-forward validation does not produce prediction intervals.

Question 12. MAPE is considered an unreliable metric when:

(a) The forecast horizon is longer than one year.
(b) Actual values are frequently near zero, causing percentage errors to explode.
(c) The time series has strong seasonality.
(d) Multiple models are being compared.

Question 13. A company reports that their forecast accuracy improved from 85% to 89%. Which of the following is the MOST important question to ask before accepting this claim?

(a) What forecasting algorithm was used?
(b) At what level of aggregation and forecast horizon was accuracy measured?
(c) How many data scientists worked on the model?
(d) Was the model deployed in production?

Question 14. In hierarchical forecasting, why is it generally better to forecast at an aggregated level and disaggregate rather than forecasting each individual series?

(a) Aggregated models always produce more accurate forecasts.
(b) Individual series often have sparse, intermittent demand with low signal-to-noise ratios, making direct modeling unstable.
(c) Individual forecasting requires more expensive hardware.
(d) Aggregated forecasting does not require historical data.

Question 15. A prediction interval of 80% means:

(a) The model is 80% accurate.
(b) There is an 80% probability that the actual value will fall within the interval.
(c) The model explains 80% of the variance in the data.
(d) 80% of the model's parameters are statistically significant.

Question 16. Athena's supply chain VP initially asked Ravi for "just one number" instead of a prediction interval. According to the chapter, why is this problematic?

(a) A single number is harder to compute than an interval.
(b) A single number provides false precision and prevents planners from managing inventory risk across scenarios.
(c) A single number is always less accurate than an interval.
(d) Regulatory requirements mandate the use of prediction intervals.

Question 17. Which of the following best describes "forecast accuracy theater"?

(a) Using dramatic visualizations to present forecast results.
(b) Reporting accuracy metrics that look favorable but do not reflect the operational accuracy that matters.
(c) Running forecasting models in a test environment but never deploying them.
(d) Using multiple models and selecting whichever produced the best accuracy on the test set.

Question 18. A company's forecasting model produces a quarterly revenue forecast of $47,832,419. The model's MAPE is 12%. What is the best way to present this forecast?

(a) "$47,832,419 in expected revenue."
(b) "$48M in expected revenue, with a range of approximately $42M to $54M."
(c) "Revenue between $47M and $48M."
(d) "$47,832,419 ± 12%."

True / False

Question 19. A stationary time series has a constant mean, constant variance, and no seasonal pattern.

(a) True
(b) False

Question 20. The Moving Average (MA) component in ARIMA refers to using a simple moving average of past values as the forecast.

(a) True
(b) False

Short Answer

Question 21. Professor Okonkwo states: "The point of forecasting isn't to be right. It's to be usefully wrong." In three to four sentences, explain what this means and why it matters for business decision-making.

Question 22. Ravi Mehta tells his team: "The model is maybe 20 percent of the work. The data pipeline is 80 percent." Explain what he means by this and provide two specific examples of data pipeline challenges in the context of demand forecasting.

Question 23. Describe a scenario in which a simple moving average forecast would outperform Prophet. What characteristics of the data or the business context would favor the simpler approach?

Question 24. A forecasting team reports an overall MAPE of 10% for their demand model. However, the supply chain team complains that stockouts have actually increased since the model was deployed. Propose three possible explanations for this contradiction.

Question 25. Explain the difference between a forecast and a target. Why is confusing the two dangerous for supply chain planning?

Answer Key

(b) — Seasonality has a fixed, known period; cyclicality has a variable, unknown period.
(b) — Multiplicative decomposition is appropriate because the seasonal effect is proportional to the level (25% of the mean, which grows as the mean grows).
(b) — "Integrated" refers to differencing the data to achieve stationarity. The "d" parameter in ARIMA(p,d,q) specifies the number of differencing steps.
(b) — ARIMA(2,1,0): p=2 (two past values in the autoregressive component), d=1 (one round of differencing), q=0 (no moving average component using past errors).
(c) — Alpha controls the weight given to recent observations when estimating the level (current position) of the series. Higher alpha means the model reacts more quickly to recent changes.
(b) — Prophet does not consistently achieve the highest accuracy. Its popularity stems from practical usability features (handling missing data, multiple seasonalities, holiday effects, interpretable components, and default prediction intervals).
(b) — A high changepoint_prior_scale allows the trend to change direction more frequently and abruptly. This increases flexibility but can lead to overfitting.
(b) — The promotional calendar is known in advance and can be provided for the future period. Actual temperature, unreported GDP, and unavailable competitor data cannot serve as external regressors because their future values are not known at forecast time.
(b) — The M4 competition found that statistical methods outperformed ML/DL methods on most individual series, but the top overall performers were ensembles combining statistical and ML approaches.
(b) — Overfitting. The LSTM memorized the training data, including noise and one-time patterns, resulting in poor generalization to new data.
(b) — Walk-forward validation respects the temporal ordering of data by always training on past observations and testing on future observations, preventing information leakage from the future.
(b) — MAPE becomes unreliable when actual values are near zero because the percentage calculation produces extremely large (and misleading) error values. WMAPE is preferred in such cases.
(b) — The aggregation level and forecast horizon are critical context. Accuracy at the total-company level is always higher than at the SKU-store level due to error cancellation. Without this context, the improvement claim is uninterpretable.
(b) — Individual SKU-store series are often sparse and intermittent, making direct modeling statistically unstable. Aggregated series have more data points and clearer signal, enabling more reliable pattern detection.
(b) — An 80% prediction interval indicates that in repeated sampling, the actual value would fall within the interval approximately 80% of the time.
(b) — A single point forecast creates the illusion of certainty, preventing planners from building safety stock buffers and contingency plans for the range of plausible outcomes.
(b) — Forecast accuracy theater occurs when metrics are reported in ways that flatter the forecasting team (e.g., using favorable aggregation levels or horizons) without reflecting the accuracy that matters for operational decisions.
(b) — Given 12% MAPE, the forecast should be rounded to a level of precision consistent with its accuracy ($48M) and presented with a range ($42M-$54M) that reflects the uncertainty.
(a) True — A stationary series has constant statistical properties (mean, variance, autocorrelation structure) over time.
(b) False — The MA component in ARIMA models the relationship between the current value and past forecast errors (residuals), not a simple moving average of past values.
Sample answer: Forecasting will never predict the future exactly — every forecast is "wrong" in the sense that it will not match the actual outcome precisely. A forecast is "usefully wrong" when it quantifies the magnitude and direction of its uncertainty through prediction intervals, enabling decision-makers to plan for a range of outcomes rather than a single scenario. In business, the value of a forecast lies not in its point accuracy but in its ability to reduce the cost of uncertainty — by informing safety stock levels, contingency plans, and resource allocation. A well-calibrated 80% prediction interval that correctly contains the actual value roughly 80% of the time is far more useful than a precise-sounding point forecast that provides no information about risk.
Sample answer: Ravi means that the forecasting model itself (choosing between Prophet, ARIMA, etc., tuning parameters, evaluating accuracy) is a relatively small portion of the total work required to deploy a production forecasting system. The majority of effort goes into the data pipeline — the infrastructure that collects, cleans, validates, and transforms raw data into model-ready format. Two specific examples: (1) Athena's legacy POS system produces data with a 48-hour lag and frequent missing values, requiring automated cleaning and imputation logic; (2) incorporating external regressor data (promotional calendars, holiday schedules, weather feeds) requires building integrations with multiple source systems, each with its own format and update cadence.
Sample answer: A simple moving average could outperform Prophet for a time series with no trend, no seasonality, and no external influences — for example, daily interbank exchange rate movements over a very short horizon, or the price of a commodity in a highly efficient market. In such cases, there is no temporal structure for Prophet to exploit, and the overhead of fitting trend, seasonality, and holiday components introduces noise. Additionally, for very short series (fewer than two seasonal cycles), Prophet does not have enough data to estimate seasonal parameters reliably, and a moving average serves as a more robust default.
Sample answer: Three possible explanations: (1) The 10% MAPE is measured at an aggregated level (e.g., category or regional), but stockouts occur at the SKU-store level where errors may be much larger. (2) The MAPE is symmetric but stockouts are asymmetric — the model's over-predictions and under-predictions cancel out in the MAPE calculation, but only under-predictions cause stockouts. (3) The safety stock levels were not updated to reflect the new model's prediction intervals — if planners are still using old safety stock rules, improved forecast accuracy does not automatically translate to fewer stockouts.
Sample answer: A forecast is a statistical estimate of what is likely to happen based on data, models, and assumptions. A target is a business objective — what the organization wants to happen based on strategic goals. Confusing the two is dangerous because it introduces bias: if the sales team's compensation depends on hitting a revenue target, there is pressure (conscious or unconscious) to adjust the "forecast" upward to match the target. This inflated forecast then drives supply chain planning, leading to overproduction, excess inventory, and wasted resources. Keeping forecasts and targets separate allows each to serve its proper function: forecasts inform planning with honest estimates, and targets motivate performance with aspirational goals.