WorksheetsBA 9
Total questions: 64
Worksheet time: 32mins
Name
Class
Date
1.
Which item is listed as a key concept for Advanced Time Series Analysis and Forecasting?
a)
Principal component analysis (PCA)
b)
Handling holidays and exogenous variables
c)
k-means clustering
d)
Decision trees
e)
Support vector machines
2.
According to the slides, which business goal is an example of Strategic Planning in temporal analytics?
a)
Estimating employee satisfaction from a one-time survey
b)
Long-term trend analysis for future demand (e.g., tourism demand for 2030)
c)
Randomly splitting customers into test groups
d)
Optimizing a static warehouse layout once per decade
e)
Choosing a logo color for a brand refresh
3.
In the "data cube" view of business time series, what best describes a multivariate time series?
a)
A single vector such as total GDP over time
b)
A vector of multiple variables over time, such as sales, price, and temperature
c)
A series with no missing values
d)
A series observed only once per year
e)
A series that must be stationary by definition
4.
In the generative view of a time series, which component represents stochastic noise or residual variation?
a)
Trend-cycle
b)
Seasonality
c)
Remainder
d)
Hierarchy constraint
e)
Exogenous regressor
5.
What is the key assumption behind a multiplicative decomposition model?
a)
Seasonal fluctuations are constant in absolute size regardless of the level
b)
Seasonal effects are unrelated to the trend
c)
Seasonal fluctuations scale proportionally with the level of the series
d)
The series must have no trend
e)
Noise must be zero-mean and independent across time
6.
Why is STL (Seasonal-Trend decomposition using Loess) used instead of classical decomposition in some cases?
a)
It forces seasonality to be constant forever
b)
It allows the seasonal component to change over time
c)
It requires the series to be strictly stationary
d)
It removes the need for any preprocessing
e)
It can only be applied to annual data
7.
How do the weights behave in exponential smoothing compared with a simple moving average?
a)
All past observations receive equal weight forever
b)
Weights increase for older observations
c)
Weights decay exponentially as observations get older
d)
Only the first observation receives weight
e)
Weights depend only on calendar month, not recency
8.
Simple Exponential Smoothing (ETS(A,N,N)) is intended for which type of data?
a)
Clear trend and clear seasonality
b)
Trend but no seasonality
c)
No clear trend and no seasonality
d)
Multiple seasonalities (hourly and weekly)
e)
Only irregular tick data
9.
Holt's linear trend method (ETS(A,A,N)) is designed for data with which pattern?
a)
Trend but no seasonality
b)
Seasonality but no trend
c)
No trend and no seasonality
d)
Only random shocks with no persistence
e)
Only fixed holiday spikes with no other structure
10.
In the damped trend method, what is the role of the damping parameter (between 0 and 1)?
a)
It amplifies the trend more and more over time
b)
It forces the forecast to be constant from the first step
c)
It gradually reduces the influence of the trend as the forecast horizon increases
d)
It removes the seasonal component
e)
It converts an additive model into a multiplicative model
11.
Holt-Winters seasonal method (Additive) ETS(A,A,A) uses how many state components?
a)
One: level only
b)
Two: level and trend
c)
Three: level, trend, and seasonal
d)
Four: level, trend, seasonal, and holiday
e)
Five: level, trend, seasonal, residual, and exogenous
12.
Holt-Winters seasonal method (Multiplicative) ETS(M,A,M) is used when seasonality is best described as:
a)
A constant amount added each season
b)
Proportional changes that scale with the level of the series
c)
Completely absent
d)
Randomly changing without any pattern
e)
Only caused by measurement error
13.
In the ETS framework, which set of components is part of the model taxonomy?
a)
Error, Trend, Seasonal
b)
Nodes, Edges, Graph
c)
Bias, Variance, Overfitting
d)
Precision, Recall, F1
e)
Shards, Replicas, Partitions
14.
In ARIMA(p,d,q), what does the "I" (d) component represent?
a)
Regression on past values
b)
Differencing to achieve stationarity
c)
Regression on past errors
d)
Multiplicative seasonality
e)
Automatic variable selection
15.
A time series is stationary if which property is true according to the definition on the slides?
a)
Its mean and variance change over time in a predictable way
b)
Its statistical properties such as mean and variance do not change over time
c)
It must contain seasonality
d)
It must be multivariate
e)
It must be measured at irregular intervals
16.
In the Augmented Dickey-Fuller (ADF) test, the null hypothesis corresponds to:
a)
The series is stationary
b)
The series has a unit root (is non-stationary)
c)
The series has no autocorrelation
d)
The series has no seasonality
e)
The series has zero variance
17.
What is the main purpose of first differencing in ARIMA modeling?
a)
Remove linear trend
b)
Remove weekly seasonality
c)
Create holiday dummy variables
d)
Increase the number of observations
e)
Guarantee that residuals are white noise without diagnostics
18.
An AR(p) model expresses the current value primarily as a function of:
a)
Only the current error term
b)
Past forecast errors (shocks)
c)
Past values of the series
d)
Only seasonal indices
e)
Only exogenous regressors
19.
An MA(q) model expresses the current value primarily as a function of:
a)
Past values of the series
b)
Past forecast errors (shocks)
c)
Only the trend component
d)
Only the seasonal component
e)
Only the mean level with no randomness
20.
Which diagnostic tool measures correlation between the series at time t and time t-k including direct and indirect effects?
a)
PACF (Partial Autocorrelation Function)
b)
ACF (Autocorrelation Function)
c)
ADF test
d)
MASE
e)
MAPE
21.
In the Box-Jenkins ARIMA procedure, what is the main goal of the diagnose step?
a)
Choose the holiday calendar
b)
Ensure residuals look like white noise
c)
Force the series to have a trend
d)
Maximize MAPE
e)
Remove all outliers without checking
22.
Why can a standard univariate SARIMA model fail for moving holidays like Easter or Ramadan?
a)
It assumes seasonality is fixed to the calendar month and cannot represent moving holiday timing
b)
It requires exogenous variables in all cases
c)
It cannot handle any form of trend
d)
It only works on quarterly data
e)
It always overfits by design
23.
Dynamic regression (ARIMAX / Reg-ARIMA) combines linear regression with what additional structure?
a)
A decision tree on the residuals
b)
An ARIMA model for the regression residuals
c)
A k-means cluster model for the residuals
d)
A neural network for the regressors
e)
A moving average on the calendar month only
24.
When modeling holidays with dummy variables, what does a basic holiday dummy column represent?
a)
The exact sales increase caused by the holiday
b)
Whether a given date is the holiday (1) or not (0)
c)
The day of week index only
d)
A rolling mean of past sales
e)
A stationary transformation of the series
25.
Why is Facebook Prophet presented as an alternative approach for business forecasting?
a)
It requires no holiday information
b)
It is designed to handle holidays and regressors easily
c)
It only works for stationary data
d)
It cannot handle outliers
e)
It forbids any seasonality terms
26.
Why is standard k-fold cross-validation with random shuffling considered illegal for time series?
a)
It increases the training size too much
b)
It destroys temporal dependence and can leak future information into training
c)
It makes models too simple
d)
It always reduces variance
e)
It guarantees unbiased forecasts
27.
In time series cross-validation, what characterizes an expanding window strategy?
a)
The training window size stays constant
b)
The training window grows to include all available history
c)
The test set is randomly sampled
d)
The model is trained once and never retrained
e)
The window ignores the most recent data
28.
In time series cross-validation, what characterizes a rolling window strategy?
a)
The training set grows without bound
b)
The training set keeps a constant size by dropping the oldest observations
c)
The test set is shuffled
d)
The horizon is always one year
e)
The model never adapts to changes
29.
What is a key flaw of MAPE mentioned in the slides?
a)
It is always defined even when actual values are zero
b)
It penalizes negative errors more than positive ones
c)
It is undefined when the actual value is zero
d)
It cannot be computed on daily data
e)
It requires stationarity
30.
How is MASE interpreted according to the slides?
a)
MASE < 1 means the model is worse than the seasonal naive benchmark
b)
MASE < 1 means the model is better than the seasonal naive benchmark
c)
MASE = 0 means the model has zero variance
d)
MASE > 1 means the model is always accurate
e)
MASE is only meaningful for stationary series
31.
In the Australian domestic tourism case study, which modeling approach is said to fit multiplicative seasonality natively?
a)
ETS(M,A,M)
b)
ARIMA without transformations
c)
A simple moving average
d)
MAPE-based selection
e)
A stationary white-noise model
32.
In the Rossmann store sales case study, which factor is listed as an exogenous variable affecting sales?
a)
Promo (marketing)
b)
Color of the store logo
c)
Store aisle width
d)
Cash register model number
e)
Customer shoe size
33.
According to the summary table, what is the main focus of ARIMA compared with ETS?
a)
Trend and seasonality components
b)
Autocorrelations between lags
c)
Holiday list management
d)
Hierarchical aggregation constraints
e)
Only regression coefficients
34.
In the forecasting workflow, which step comes immediately after benchmarking with MASE of a seasonal naive model?
a)
Validate with rolling cross-validation
b)
Visualize the series
c)
Fit ETS and ARIMA models (auto-selection via AICc)
d)
Add holiday dummy variables
e)
Select the model with lowest test MASE
35.
A retailer has hourly demand data showing patterns by time of day and day of week. Based on the slides, what modeling implication follows from this frequency?
a)
Hourly data cannot contain any seasonality
b)
High-frequency data can introduce multiple seasonalities that must be considered
c)
Only annual seasonality is possible at high frequency
d)
Hierarchical aggregation constraints disappear at high frequency
e)
Stationarity is guaranteed when sampling is hourly
36.
A series has seasonal peaks that become larger as the overall level rises (for example, summer peaks grow as total demand grows). Which decomposition or seasonal form is most consistent with this behavior?
a)
Additive seasonality (constant absolute seasonal swings)
b)
Multiplicative seasonality (seasonal swings scale with the level)
c)
No seasonality (N)
d)
A purely moving-average model with equal weights
e)
A model that assumes seasonality is constant forever
37.
If a business believes the seasonal shopping pattern itself is changing over the years (for example, December effects evolving), which decomposition approach from the slides is designed for this situation?
a)
Classical decomposition with fixed seasonality
b)
STL decomposition with seasonality allowed to change over time
c)
First differencing (d=1) only
d)
AR(p) without any seasonal terms
e)
Randomly shuffled k-fold cross-validation
38.
Why can exponential smoothing be more responsive to recent changes than a simple moving average of the last k observations?
a)
It assigns equal weight to all k observations
b)
It assigns higher weight to older observations
c)
It uses weights that decay as observations get older, emphasizing recent data
d)
It removes all autocorrelation by construction
e)
It requires differencing before use
39.
You observe a series with a clear upward trend but no repeating seasonal pattern. Which ETS method is the best match described in the slides?
a)
SES: ETS(A,N,N)
b)
Holt: ETS(A,A,N)
c)
Holt-Winters additive: ETS(A,A,A)
d)
Holt-Winters multiplicative: ETS(M,A,M)
e)
A stationary white-noise model
40.
For long-term sales budgeting, the slides note that trends rarely continue linearly forever due to competition and saturation. Which method explicitly addresses this with a damping parameter?
a)
SES: ETS(A,N,N)
b)
Holt-Winters additive: ETS(A,A,A)
c)
Damped trend: ETS(A,Ad,N)
d)
ARIMA with no differencing
e)
MAPE-based baseline forecasting
41.
Automatic ETS model selection via AICc is described as balancing goodness of fit against model complexity. What does this imply about selecting a model with a slightly better fit but many extra parameters?
a)
It is always preferred because fit is the only criterion
b)
It can be rejected if the added complexity is not justified by fit improvement
c)
It is preferred only when the series is stationary
d)
It is preferred only when holidays are present
e)
It is invalid unless cross-validation is randomized
42.
The slides contrast ETS and ARIMA philosophies. Which statement matches that contrast?
a)
ETS focuses on autocorrelations between lags, while ARIMA focuses on trend and seasonality components
b)
ETS focuses on trend/seasonality components, while ARIMA focuses on autocorrelations between lags
c)
Both focus only on holiday dummy variables
d)
Both require stationarity of the original series
e)
Neither can be combined with regressors
43.
A series shows a linear trend. In ARIMA terms, which preprocessing step is directly motivated by the slides to address this before modeling autocorrelations?
a)
Apply first differencing (set d=1) to remove the trend
b)
Randomly shuffle observations to remove dependence
c)
Use MAPE as the loss function to remove trend
d)
Drop the most recent observations to avoid peeking
e)
Replace the series with a constant mean
44.
An ADF test on a series returns a p-value of 0.08. Using the hypotheses described in the slides, what conclusion is most appropriate?
a)
Reject the unit-root null; the series is stationary
b)
Fail to reject the unit-root null; the series is likely non-stationary
c)
Conclude the series has no seasonality
d)
Conclude the model residuals are white noise
e)
Conclude ETS must be used instead of ARIMA
45.
You inspect ACF and PACF to identify an AR or MA order. According to the rules of thumb shown, which pattern points to an AR(p) model?
a)
ACF cuts off after lag q; PACF decays
b)
ACF decays; PACF cuts off after lag p
c)
Both ACF and PACF cut off immediately at lag 0
d)
Both ACF and PACF decay with no structure
e)
Only ADF p-values determine p and q
46.
In the Box-Jenkins procedure, the diagnose step requires residuals to be white noise. If residuals still show serial dependence, what does the procedure imply you should do?
a)
Accept the model because the fit is already good
b)
Return to identification/estimation to adjust the ARIMA specification
c)
Switch to standard shuffled k-fold validation
d)
Stop and report MAPE only
e)
Remove all exogenous variables by default
47.
Easter moves between March and April across years. Based on the slides, which modeling choice best addresses this limitation of fixed-calendar SARIMA seasonality?
a)
Use SARIMA only and assume Easter always occurs in March
b)
Add holiday dummy variables via dynamic regression (ARIMAX / Reg-ARIMA)
c)
Apply only first differencing and ignore holidays
d)
Use a simple moving average with equal weights
e)
Randomly shuffle data before fitting ARIMA
48.
The slides mention holiday "lead-up" and "hangover" effects (for example, sales spike before Christmas). What is the recommended implementation strategy?
a)
Use one dummy for the holiday date only
b)
Create multiple dummy variables representing lead and lag windows around the holiday
c)
Replace holidays with a single annual seasonal component
d)
Assume holiday effects are absorbed by the AR terms automatically
e)
Remove all holiday periods from the dataset
49.
Why is random shuffling in standard k-fold cross-validation considered invalid for time series forecasting, according to the slides?
a)
It reduces the size of the training set
b)
It destroys temporal dependence and can leak future information into training
c)
It makes ETS models impossible to fit
d)
It forces multiplicative seasonality
e)
It guarantees overly high error estimates
50.
If you suspect structural breaks and want the model to "forget" old, potentially irrelevant history, which time series cross-validation strategy is emphasized as helpful on the slides?
a)
Expanding window, because it always uses all history
b)
Rolling window, because it keeps a constant training size and drops older data
c)
Random shuffle k-fold, because it mixes all regimes
d)
Single holdout split without retraining
e)
MASE scaling, because it removes breaks
51.
A retail dataset includes many zero-sales days because stores are closed. Given the metric discussion on the slides, which evaluation approach is most defensible?
a)
Prefer MAPE because zeros do not affect it
b)
Prefer MASE because it avoids the zero-division issue that breaks MAPE
c)
Prefer only RMSE because it is always scale-free
d)
Avoid any benchmark and report only training error
e)
Use shuffled k-fold validation with MAPE
52.
Your model achieves a test MASE of 1.3. According to the slides' interpretation, what should you conclude?
a)
The model is better than the seasonal naive benchmark
b)
The model is worse than the seasonal naive benchmark
c)
The model is perfect on all horizons
d)
The model is guaranteed to generalize
e)
The series must be stationary
53.
A company has a steadily growing sales series where seasonal swings stay roughly the same in absolute units (for example, about +/- 100 units each year) even as the level rises. Which decomposition assumption best matches this pattern?
a)
Multiplicative decomposition, because seasonality scales with the level
b)
Additive decomposition, because seasonal magnitude is roughly constant
c)
No decomposition, because any seasonality implies multiplicative effects
d)
Seasonality must be modeled only with AR terms
e)
The series is stationary by definition
54.
In the damped trend method, how would forecasts typically differ when the damping parameter is very close to 1 versus clearly below 1?
a)
Close to 1: trend influence persists longer; below 1: the trend contribution flattens more quickly with horizon
b)
Close to 1: forecasts immediately become constant; below 1: forecasts explode
c)
Close to 1: removes all trend; below 1: adds a new seasonal pattern
d)
Close to 1: makes the series stationary; below 1: makes it non-stationary
e)
Close to 1: forces multiplicative error; below 1: forces additive error
55.
A time series shows increasing variability as the level increases (variance grows over time). Which Box-Jenkins step from the slides is specifically intended to address this before differencing and identification?
a)
Randomly shuffle the series to average out variance changes
b)
Apply a Box-Cox transform such as a log transform to stabilize variance
c)
Skip transformation and increase AR order p
d)
Compute MAPE and use it as a variance stabilizer
e)
Use rolling cross-validation to remove heteroscedasticity
56.
You run the ADF test and obtain a p-value slightly above 0.05. In the ARIMA selection procedure shown, what is the most consistent next action?
a)
Assume stationarity and skip differencing
b)
Apply differencing and re-test until the stationarity test passes
c)
Increase the seasonal period to force p-value below 0.05
d)
Switch to shuffled k-fold validation to validate stationarity
e)
Replace the series with a moving average of the last k points
57.
Suppose the ACF cuts off after lag 2 while the PACF decays gradually. Using the diagnostic rules on the slides, which model family is most suggested and with what order?
a)
AR(2)
b)
MA(2)
c)
ARMA(2,2) is required in all cases
d)
A unit-root model with d=2 only
e)
SES with no trend and no seasonality
58.
A store-sales dataset has weekly seasonality (stores closed on Sundays), annual holiday effects, promotions, and many zero-sales days. Based on the Rossmann case study slide, which modeling design is most aligned with the recommended approach?
a)
Univariate ARIMA with no regressors, because seasonality is fixed
b)
Dynamic regression (ARIMAX) with dummies for promo/day-of-week/holidays and AR errors for remaining dependence
c)
Simple moving average, because zeros make models unstable
d)
Classical decomposition only, because it separates components
e)
Prophet without any holiday list, because holidays are implicit
59.
You need a validation strategy that both avoids look-ahead bias and adapts to structural breaks by down-weighting or discarding older history. Which combination best matches the slides?
a)
Standard shuffled k-fold with 80/20 split
b)
Rolling window time series cross-validation
c)
A single random holdout split repeated many times
d)
Expanding window with no retraining
e)
Stationarity testing with ADF only
60.
Which statement best captures the trade-off between expanding-window and rolling-window time series cross-validation as presented on the slides?
a)
Expanding window is cheaper because the training set stays constant in size
b)
Rolling window always uses all available history, while expanding window discards it
c)
Expanding window uses all history but can be computationally expensive as the training set grows
d)
Rolling window is illegal because it drops observations
e)
Both strategies require random shuffling to be valid
61.
MAPE is described as asymmetric, penalizing positive errors more than negative ones. In practice, what is a key risk of using MAPE to compare two forecasting models?
a)
It can favor a model that tends to under-forecast because over-forecasts are penalized more
b)
It can only be computed for quarterly data
c)
It automatically corrects for look-ahead bias
d)
It is unaffected by zero actual values
e)
It forces the use of multiplicative seasonality
62.
In the Australian tourism case study, ARIMA is said to require a log transform to handle multiplicative seasonality, while ETS(M,A,M) fits it natively. What is the most defensible interpretation of this comparison?
a)
ARIMA cannot model seasonality at all
b)
When seasonal swings scale with the level, a transformation may be needed for ARIMA, whereas an ETS model with multiplicative seasonality can represent the pattern directly
c)
ETS cannot represent trend, so it needs ARIMA to add trend
d)
Log transforms are only used to remove seasonality, not to address scaling
e)
The choice is irrelevant because AICc ignores seasonality
63.
Prophet is presented as easier for holiday integration and more robust to outliers, but less statistically rigorous than ARIMA. Which scenario best justifies choosing Prophet, based on that trade-off?
a)
A setting where strict unit-root testing and Box-Jenkins diagnostics are the primary requirement
b)
A setting where quick business forecasting with many known holidays and outliers is needed and ease-of-use is prioritized
c)
A setting with no holidays and no outliers where only autocorrelations matter
d)
A setting where random shuffling is acceptable for validation
e)
A setting where the series must be univariate by law
64.
The workflow slide recommends selecting the model with the lowest test MASE after time series cross-validation. What failure mode is this step designed to reduce compared with choosing the model with the lowest training error?
a)
Look-ahead bias and overfitting to the training history
b)
The need to identify seasonality
c)
The requirement that the series be multivariate
d)
The need for any preprocessing
e)
The need to compute a benchmark
100 %
