Time Series Analysis — Questions and Answers

Question 1: What is 'forecast combination,' and why is it used with exponential smoothing methods?

Using exponential smoothing to combine input variables
Selecting the single best model and discarding others
Applying different smoothing levels to different series
Averaging forecasts from multiple models to reduce prediction error variance (Correct answer)

Correct answer: Averaging forecasts from multiple models to reduce prediction error variance

Combining forecasts from several models often outperforms any single model by reducing the risk of betting on one specification.

Question 2: What is the purpose of differencing in an ARIMA model?

To remove seasonality only
To achieve stationarity (Correct answer)
To increase autocorrelation
To normalize residuals

Correct answer: To achieve stationarity

Differencing removes trends and unit roots, transforming a non-stationary series into a stationary one.

Question 3: What does GARCH add beyond the basic ARCH model?

A deterministic trend in volatility
Additional AR terms in the mean equation
Lagged conditional variance terms in addition to lagged squared errors (Correct answer)
Seasonal heteroscedasticity

Correct answer: Lagged conditional variance terms in addition to lagged squared errors

GARCH includes past conditional variances as predictors, allowing the model to capture longer-memory volatility persistence with fewer parameters than high-order ARCH.

Question 4: In time series forecasting, what is the 'horizon' h?

The seasonal period of the series
The number of model parameters
The number of steps ahead being forecast (Correct answer)
The length of the training data

Correct answer: The number of steps ahead being forecast

The forecast horizon h specifies how many time periods into the future the forecast extends (e.g., h=12 for 12-step-ahead monthly forecasts).

Question 5: What does the PACF plot help identify in an ARIMA model?

The seasonal period
The differencing order 'd'
The MA order 'q'
The AR order 'p' (Correct answer)

Correct answer: The AR order 'p'

The PACF cuts off sharply after lag p for a pure AR process, making it useful for selecting the AR order.

Question 6: What is N-BEATS, and in what context is it used?

A deep learning architecture for univariate time series forecasting based on backward and forward residual links (Correct answer)
A batch exponential smoothing algorithm
A neural network for natural language processing applied to time series text
A non-parametric bootstrapping method for time series

Correct answer: A deep learning architecture for univariate time series forecasting based on backward and forward residual links

N-BEATS (Neural Basis Expansion Analysis for Time Series) is a deep learning model that uses stacked fully connected blocks with interpretable basis functions for forecasting.

Question 7: What is ARCH (Autoregressive Conditional Heteroscedasticity) in time series modeling?

A model for series with conditional mean changes
A model where the variance of the error term depends on past squared errors (Correct answer)
An autoregressive model with changing coefficients
A heteroscedastic extension of the MA model

Correct answer: A model where the variance of the error term depends on past squared errors

ARCH models the time-varying variance by making it a function of past squared residuals, capturing volatility clustering in financial data.

Question 8: Which of the following best describes a 'trend-stationary' process?

A process with no trend at all
A process that is stationary after removing a deterministic trend (Correct answer)
A process with a stochastic trend only
A process that is stationary after differencing

Correct answer: A process that is stationary after removing a deterministic trend

A trend-stationary process has a deterministic trend; once that trend is subtracted, the residuals are stationary.

Question 9: What condition must be satisfied for an AR(p) model to be stationary?

Variance must be zero
The mean must equal zero
All roots of the characteristic polynomial must lie outside the unit circle (Correct answer)
All roots must lie inside the unit circle

Correct answer: All roots of the characteristic polynomial must lie outside the unit circle

Stationarity of an AR model requires that characteristic polynomial roots have modulus greater than 1 (outside the unit circle).

Question 10: An ARIMA(0,1,0) model is equivalent to which simple forecasting method?

Exponential smoothing
Simple moving average
Random walk (Correct answer)
Linear regression

Correct answer: Random walk

ARIMA(0,1,0) applies one difference with no AR or MA terms, which is the definition of a random walk model.

Question 11: Over-differencing a time series typically introduces which problem?

Heteroscedasticity
Trend reappearance
Perfect stationarity
Negative autocorrelation at lag 1 (Correct answer)

Correct answer: Negative autocorrelation at lag 1

Over-differencing adds unnecessary variance and often creates a strong negative autocorrelation at the first lag.

Question 12: What does the 'I' in ARIMA stand for?

Integrated (Correct answer)
Indexed
Iterative
Inverse

Correct answer: Integrated

The 'I' in ARIMA stands for Integrated, referring to the differencing applied to make the series stationary.

Question 13: What is the primary risk of minimizing MAPE for model selection?

MAPE ignores large errors
MAPE penalizes over-forecasts more than under-forecasts, creating asymmetric bias toward lower predictions (Correct answer)
MAPE is undefined for large series
MAPE always favors ARIMA over ETS

Correct answer: MAPE penalizes over-forecasts more than under-forecasts, creating asymmetric bias toward lower predictions

MAPE's asymmetry arises because a 50% over-forecast results in 50% error, but a 50% under-forecast gives 100% error, pushing models to systematically under-predict.

Question 14: The Yule-Walker equations are used to:

Relate the theoretical PACF of an AR(p) process to its parameters (Correct answer)
Estimate MA parameters from the sample ACF
Compute the spectral density from the PACF
Test for the presence of unit roots using autocorrelations

Correct answer: Relate the theoretical PACF of an AR(p) process to its parameters

The Yule-Walker equations express the autocovariances (and thus ACF) of an AR(p) process as a linear function of the AR parameters, enabling parameter estimation.

Question 15: In an ARIMA(p,d,q) model, what does the parameter 'p' represent?

Order of the autoregressive component (Correct answer)
Order of the moving average component
Number of differences
Seasonal period

Correct answer: Order of the autoregressive component

The parameter 'p' specifies the order of the autoregressive (AR) component, indicating how many lagged values are used.

Question 16: Which Python class in statsmodels implements the full Holt-Winters exponential smoothing model?

ets()
ExponentialSmoothing (Correct answer)
SimpleExpSmoothing
HoltWinters

Correct answer: ExponentialSmoothing

statsmodels.tsa.holtwinters.ExponentialSmoothing provides the full ETS model including additive and multiplicative trend and seasonality options.

Question 17: What is 'dynamic time warping' (DTW) used for in time series analysis?

Computing cross-correlations at multiple lags
Differencing two series to find cointegration
Measuring similarity between two time series that may be shifted or distorted in time (Correct answer)
Aligning seasonal indices across years

Correct answer: Measuring similarity between two time series that may be shifted or distorted in time

DTW finds the optimal alignment between two sequences by allowing flexible time shifts, making it a robust distance measure for time series classification and clustering.

Question 18: What does 'invertibility' mean in the context of an MA model?

The model parameters are negative
The forecast errors are zero
The MA model can be expressed as an infinite AR model (Correct answer)
The series can be reversed in time

Correct answer: The MA model can be expressed as an infinite AR model

An invertible MA model has roots outside the unit circle and can be rewritten as an infinite-order AR process.

Question 19: Which plot is most commonly used to determine the MA order 'q' in an ARIMA model?

Residual plot
Spectral density plot
Autocorrelation Function (ACF) (Correct answer)
Partial Autocorrelation Function (PACF)

Correct answer: Autocorrelation Function (ACF)

The ACF plot shows a sharp cutoff after lag q for a pure MA process, helping identify the MA order.

Question 20: What is the main advantage of LSTM networks over ARIMA for time series forecasting?

LSTMs are always faster to train
LSTMs produce interpretable coefficients
LSTMs can capture non-linear dependencies and complex patterns across long sequences (Correct answer)
LSTMs always require less data

Correct answer: LSTMs can capture non-linear dependencies and complex patterns across long sequences

LSTM (Long Short-Term Memory) networks learn non-linear temporal dependencies from data without requiring explicit model specification, potentially outperforming linear models on complex series.

Question 21: In a business, prosperity, recession, and depression are examples of this.

Irregular Component
Cyclical Component (Correct answer)
Secular Trend
Seasonal Component

Correct answer: Cyclical Component

Prosperity, recession, and depression are all stages of the business cycle, which represents the cyclical component of a time series. These are medium-term fluctuations in economic activity that typically last longer than a year but are not as long-term as a secular trend. They reflect the ebb and flow of the economy, driven by various economic factors.

Question 22: What is the Holt-Winters method designed to handle?

Non-seasonal stationary series only
Series with only cyclical components
Series with no trend
Time series with both trend and seasonality (Correct answer)

Correct answer: Time series with both trend and seasonality

Holt-Winters extends Holt's method with a third equation for the seasonal component, handling series that have both trend and repeating seasonal patterns.

Question 23: What is Granger causality in time series analysis?

A test whether past values of series X improve predictions of series Y beyond Y's own past (Correct answer)
A measure of cross-correlation between two series
A test for unit roots in bivariate systems
A cointegration test

Correct answer: A test whether past values of series X improve predictions of series Y beyond Y's own past

Granger causality tests whether lagged values of X add statistically significant predictive power for Y after controlling for Y's own lags.

Question 24: What causes a forecast to be 'biased'?

High variance in forecast errors
Non-normality of residuals
Systematic over- or under-prediction where the mean error is consistently non-zero (Correct answer)
Large absolute errors

Correct answer: Systematic over- or under-prediction where the mean error is consistently non-zero

Bias occurs when a model consistently forecasts too high or too low, resulting in a non-zero mean error (ME) across forecasts.

Question 25: What distinguishes the ETS(A,N,N) model?

It is equivalent to ARIMA(1,1,1)
It is simple exponential smoothing with additive errors, no trend, and no seasonality (Correct answer)
It has multiplicative seasonality
It is Holt-Winters with additive errors

Correct answer: It is simple exponential smoothing with additive errors, no trend, and no seasonality

ETS(A,N,N) denotes additive Error, None Trend, None Seasonality — the state-space representation of simple exponential smoothing.

Question 26: What does a significant Ljung-Box test on ARIMA residuals indicate?

The model fits perfectly
Stationarity is confirmed
Seasonality is absent
Residual autocorrelation remains — model is inadequate (Correct answer)

Correct answer: Residual autocorrelation remains — model is inadequate

A significant Ljung-Box p-value means residuals are not white noise, suggesting the ARIMA model has not captured all structure.

Question 27: Which metric penalizes model complexity more heavily than AIC?

MAE
R-squared
RMSE
BIC (Bayesian Information Criterion) (Correct answer)

Correct answer: BIC (Bayesian Information Criterion)

BIC applies a stronger penalty for additional parameters than AIC, especially with larger sample sizes.

Question 28: What is 'mstl()' used for in modern time series analysis?

Multiple Seasonal-Trend decomposition using Loess for multiple seasonal periods (Correct answer)
Moving Standard Time Lag
Mean Seasonal Trend Level
Multi-Step Trend Linearization

Correct answer: Multiple Seasonal-Trend decomposition using Loess for multiple seasonal periods

MSTL extends STL to handle series with multiple seasonalities (e.g., hourly data with daily and weekly patterns simultaneously).

Question 29: When should multiplicative decomposition be preferred over additive?

When the series is stationary
When residuals are normally distributed
When seasonal variation grows proportionally with the trend level (Correct answer)
When the series has no trend

Correct answer: When seasonal variation grows proportionally with the trend level

Multiplicative decomposition suits series where seasonal swings increase as the series level increases, reflecting proportional variation.

Question 30: What is a unit root in a time series?

A characteristic root equal to 1, indicating non-stationarity (Correct answer)
A seasonal component with period 1
A constant mean value
A root that stabilizes variance

Correct answer: A characteristic root equal to 1, indicating non-stationarity

A unit root means the characteristic equation has a solution of exactly 1, which causes the series to be non-stationary with a stochastic trend.

Question 31: How are the smoothing parameters (α, β, γ) typically estimated in exponential smoothing models?

By setting them manually to fixed values like 0.5
By solving a system of linear equations
By using cross-validation on test data only
By minimizing the sum of squared one-step-ahead forecast errors (Correct answer)

Correct answer: By minimizing the sum of squared one-step-ahead forecast errors

Parameters are optimized numerically to minimize in-sample forecast error, most commonly the sum of squared errors (SSE).

Question 32: What is 'forecast reconciliation' in hierarchical time series?

Adjusting forecasts at different levels so they are mutually consistent (Correct answer)
Combining multiple model forecasts into one
Removing bias from individual series forecasts
Choosing the best model for each series independently

Correct answer: Adjusting forecasts at different levels so they are mutually consistent

Hierarchical reconciliation ensures that forecasts for sub-series sum to forecasts for aggregate series, maintaining logical consistency across levels.

Question 33: What is the Prophet model, and by whom was it developed?

A decomposable additive time series model developed by Facebook (Meta) (Correct answer)
A Bayesian ARIMA variant by Amazon
A seasonal VAR model by the US Census Bureau
A neural network model developed by Google

Correct answer: A decomposable additive time series model developed by Facebook (Meta)

Prophet (2017) decomposes series into trend, seasonality, and holiday effects using an additive model, optimized for business forecasting with automatic changepoint detection.

Question 34: Ratio-to-trend analysis is a typical approach for determining

Represent graphical curve
Deseasonalize data (Correct answer)
Remove multicollinearity
Take moving average

Correct answer: Deseasonalize data

Ratio-to-trend analysis is a method used to isolate and measure the seasonal component in a time series. It involves calculating the ratio of actual observed values to the trend values, then averaging these ratios for corresponding periods to derive seasonal indices. These indices are then used to remove the seasonal effect from the original data, a process known as deseasonalization.

Question 35: If the ACF of a time series shows a sinusoidal pattern with significant spikes at regular intervals, what does this most likely indicate?

The series contains a seasonal component (Correct answer)
The series is purely autoregressive
The series exhibits heteroskedasticity
The series is integrated of order 1

Correct answer: The series contains a seasonal component

Regular sinusoidal spikes in the ACF at fixed intervals (e.g., lags 12, 24, 36 for monthly data) indicate the presence of seasonality.

Question 36: What are 'impulse response functions' (IRFs) in VAR analysis?

Functions measuring parameter sensitivity
Functions showing how each variable responds over time to a shock in one variable (Correct answer)
Functions showing seasonal responses to external inputs
Forecasting functions for long horizons

Correct answer: Functions showing how each variable responds over time to a shock in one variable

IRFs trace the dynamic effect of a one-unit shock to one variable on itself and all other variables in the system over subsequent time periods.

Question 37: We can't find trend values of some things using the moving average method.

Middle Period
Starting and End Periods (Correct answer)
End Periods
Starting Periods

Correct answer: Starting and End Periods

The moving average method calculates the average of data points over a specified period. Consequently, it cannot compute trend values for the very beginning or very end of the time series because there aren't enough preceding or succeeding data points to form a complete average window. This results in a loss of data at both ends of the series.

Question 38: What is the key difference between in-sample (training) error and out-of-sample (test) error?

They measure the same thing
In-sample error measures fit on data used for estimation; out-of-sample measures true predictive accuracy on unseen data (Correct answer)
In-sample error is always larger
Out-of-sample error is computed before training

Correct answer: In-sample error measures fit on data used for estimation; out-of-sample measures true predictive accuracy on unseen data

In-sample error can be overly optimistic due to overfitting; only out-of-sample evaluation on a held-out test set truly measures forecast performance.

Question 39: When both the ACF and PACF of a stationary series show gradual decay, which model class is most likely appropriate?

Random walk
Pure AR(p)
ARMA(p,q) with both p>0 and q>0 (Correct answer)
Pure MA(q)

Correct answer: ARMA(p,q) with both p>0 and q>0

When both the ACF and PACF tail off gradually without cutting off, an ARMA model with both AR and MA components is typically indicated.

Question 40: What does 'forecast encompassing' mean?

When forecasts are combined with equal weights
When a model uses all available predictors
When one forecast contains all the information of another forecast, making combination unnecessary (Correct answer)
When multiple forecasts produce identical values

Correct answer: When one forecast contains all the information of another forecast, making combination unnecessary

If forecast A encompasses forecast B, then forecast B adds no useful information beyond A, indicating A is the superior standalone model.

Question 41: Which of the following is a key difference between the ACF and the PACF?

The ACF only applies to stationary series, while the PACF does not
The PACF is always non-negative, while the ACF can be negative
The PACF removes indirect effects of intermediate lags, while the ACF does not (Correct answer)
The ACF is bounded between 0 and 1, while the PACF is not

Correct answer: The PACF removes indirect effects of intermediate lags, while the ACF does not

The key difference is that the PACF isolates the direct relationship at lag k by controlling for all shorter lags, whereas the ACF includes all indirect effects.

Question 42: Which statistical test is most widely used to test for a unit root (non-stationarity)?

Ljung-Box test
Breusch-Pagan test
Shapiro-Wilk test
Augmented Dickey-Fuller (ADF) test (Correct answer)

Correct answer: Augmented Dickey-Fuller (ADF) test

The ADF test checks for the presence of a unit root; a low p-value (< 0.05) allows rejection of the null hypothesis of non-stationarity.

Question 43: What does a 'subseries plot' reveal in seasonal decomposition analysis?

The trend component only
The mean and variation within each season, showing seasonal patterns and changes over time (Correct answer)
The autocorrelation at each seasonal lag
Residual distribution by season

Correct answer: The mean and variation within each season, showing seasonal patterns and changes over time

A subseries plot groups data by season (e.g., all Januaries, all Februaries) and plots them over time, revealing both the seasonal pattern and whether it has changed.

Question 44: For a SARIMA model ARIMA(1,1,1)(1,1,1)[12], what does the [12] represent?

12 AR lags
12 MA lags
Seasonal period of 12 months (Correct answer)
12 differencing steps

Correct answer: Seasonal period of 12 months

The [12] denotes the seasonal period, meaning the seasonal pattern repeats every 12 time steps (monthly data with annual seasonality).

Question 45: A researcher finds that the PACF of a residual series from a fitted model has a significant spike at lag 4. What does this suggest?

An additional AR(4) term or seasonal term may be needed (Correct answer)
The model should be re-estimated with fewer parameters
The residuals are white noise and the model is adequate
The series needs to be log-transformed

Correct answer: An additional AR(4) term or seasonal term may be needed

A significant spike in the residual PACF at lag 4 indicates unexplained autocorrelation at that lag, suggesting an AR(4) or seasonal AR component may improve the model.

Question 46: Which metric is commonly minimized to select between competing exponential smoothing models?

MAPE only
AICc (corrected Akaike Information Criterion) (Correct answer)
Adjusted R-squared
MAE only

Correct answer: AICc (corrected Akaike Information Criterion)

AICc corrects AIC for small sample sizes, making it the preferred criterion for ETS model selection in automated forecasting.

Question 47: Which information criterion is commonly used for ARIMA model selection?

AIC (Akaike Information Criterion) (Correct answer)
Durbin-Watson statistic
R-squared
Mean Absolute Error

Correct answer: AIC (Akaike Information Criterion)

AIC balances model fit and complexity, penalizing models with more parameters to avoid overfitting.

Question 48: What is the Box-Jenkins methodology primarily used for?

Systematic identification, estimation, and diagnostic checking of ARIMA models (Correct answer)
Calculating moving averages
Computing seasonal indices
Only visualizing time series data

Correct answer: Systematic identification, estimation, and diagnostic checking of ARIMA models

Box-Jenkins is a three-step iterative process: identify model orders, estimate parameters, then check residual diagnostics.

Question 49: When the movements in the time series are _____, the straight line is fitted to the time series:

Nonlinear
Upward
Irregular
Linear (Correct answer)

Correct answer: Linear

A straight line, or linear trend, is fitted to a time series when the data points exhibit a consistent, additive increase or decrease over time. This method assumes a constant rate of change and is appropriate for modeling linear relationships. If the movements were nonlinear, a curved line or more complex model would be required to accurately represent the data.

Question 50: This stage is characterized by an increase in the production of goods that leads to prosperity:

Boom
Recovery (Correct answer)
Recession
Depression

Correct answer: Recovery

In the context of business cycles, the Recovery stage follows a depression or recession. It is characterized by an increase in economic activity, rising production of goods and services, and growing employment. This upward trend signifies a return to growth and eventually leads to a period of prosperity.

Question 51: Which function in Python's statsmodels library is commonly used to fit ARIMA models?

time_series_arima()
ARIMA() (Correct answer)
fit_arima()
arima_fit()

Correct answer: ARIMA()

The statsmodels ARIMA() class provides a full interface for specifying and fitting ARIMA models to time series data.

Question 52: The Ljung-Box test statistic is used to:

Estimate the order of differencing needed
Measure the strength of seasonal autocorrelation
Test whether a group of autocorrelations are jointly zero (Correct answer)
Test whether a single autocorrelation is significant

Correct answer: Test whether a group of autocorrelations are jointly zero

The Ljung-Box Q statistic tests the null hypothesis that the first m autocorrelations are jointly equal to zero, useful for model diagnostics.

Question 53: What is a periodogram in spectral analysis?

A histogram of time series values
A plot of forecast errors over time
A plot of the squared amplitude of each frequency component against frequency (Correct answer)
A plot of autocorrelation vs lag

Correct answer: A plot of the squared amplitude of each frequency component against frequency

The periodogram estimates the spectral density by computing and squaring the Fourier transform coefficients at each frequency.

Question 54: What is the key difference between AR and MA components in ARIMA?

AR uses future values; MA uses current values
AR reduces variance; MA reduces bias
AR is for stationary series; MA is for non-stationary
AR uses past observations; MA uses past forecast errors (Correct answer)

Correct answer: AR uses past observations; MA uses past forecast errors

The AR component models current values as a linear function of past observations, while MA models them as a function of past errors.

Question 55: How many components are there in time series data?

6
7
2
4 (Correct answer)

Correct answer: 4

Time series data is typically decomposed into four main components: Trend, Seasonal, Cyclical, and Irregular (or Random) components. The Trend represents long-term progression, Seasonal reflects short-term periodic fluctuations, Cyclical describes medium-term oscillations around the trend, and Irregular accounts for unpredictable variations. Understanding these components helps in analyzing and forecasting time series data.

Question 56: What is the 'Nyquist frequency' in spectral analysis of a discrete time series?

The fundamental frequency of the seasonal cycle
Half the sampling frequency — the highest frequency detectable in discrete data (Correct answer)
The frequency at which the periodogram is maximized
The frequency corresponding to a unit root

Correct answer: Half the sampling frequency — the highest frequency detectable in discrete data

The Nyquist frequency is fs/2 (where fs is the sampling rate), representing the maximum frequency that can be unambiguously detected without aliasing.

Question 57: The term "observation set" refers to a collection of observations taken at regular intervals of time.

Data
Geometric Series
Array data
Time series data (Correct answer)

Correct answer: Time series data

Time series data is defined as a sequence of observations recorded at successive, equally spaced points in time. This regular interval is a defining characteristic, distinguishing it from other types of data. Examples include daily stock prices, monthly sales figures, or annual population counts, all of which are collected sequentially over time.

Question 58: Why is splitting time series data for train/test using random sampling incorrect?

It increases variance in test set performance
It reduces training data size unnecessarily
It prevents parameter estimation
It causes data leakage — future observations in training allow the model to 'see ahead' (Correct answer)

Correct answer: It causes data leakage — future observations in training allow the model to 'see ahead'

Time series data has temporal dependencies; random splitting breaks this order, allowing future information to leak into training, producing unrealistically optimistic results.

Question 59: In ARIMA modeling, what does 'd=2' imply?

There are 2 MA terms
The seasonal period is 2
The series was differenced twice to achieve stationarity (Correct answer)
There are 2 AR terms

Correct answer: The series was differenced twice to achieve stationarity

d=2 means second-order differencing was applied, removing both linear and quadratic trends from the series.

Question 60: What does Holt's linear trend method add compared to simple exponential smoothing?

Automatic outlier detection
A third equation for seasonality
A multiplicative error term
A second smoothing equation to track the trend component (Correct answer)

Correct answer: A second smoothing equation to track the trend component

Holt's method introduces a separate trend equation with its own smoothing parameter β, allowing the model to follow linear trends.

Time Series Analysis

A comprehensive assessment covering statistical methods for analyzing time-ordered data, including ARIMA modeling, forecasting techniques, and frequency-domain spectral analysis. Commonly tested in graduate statistics, data science, and econometrics programs.

Questions

90 min

Time Limit

70%

To Pass

Checking for saved progress...

Exam Rules

You can skip questions and return to them later
Flag questions for review before submitting
No feedback shown until you submit the entire exam
Unanswered questions count as wrong — answer everything
10 pretest questions are mixed in and don't affect your score
Timer auto-submits when time runs out
Your progress is auto-saved every 30 seconds