Time Series quiz

Question 1

What is a time series and what makes it different from tabular data?

Accepted Answer

Temporally ordered observations with dependency between consecutive values

Answer

Independent identically distributed samples with timestamps

Answer

Only applicable to stock prices

Answer

Any sequence of numbers

Question 2

What is stationarity in time series?

Accepted Answer

Constant mean, variance, and autocovariance over time

Answer

Always increasing series

Answer

No seasonal patterns

Answer

Only applies to financial data

Question 3

What is autocorrelation and partial autocorrelation?

Accepted Answer

ACF: total correlation at lag k; PACF: partial correlation removing intermediate lags

Answer

They are the same metric

Answer

ACF measures seasonality only

Answer

PACF is always higher than ACF

Question 4

What is SARIMA and when is it used?

Accepted Answer

Extends ARIMA with seasonal AR/MA/differencing terms for periodic patterns

Answer

Replaces ARIMA for all tasks

Answer

Only handles annual seasonality

Answer

Doesn't use differencing

Question 5

What is time series cross-validation?

Accepted Answer

Walk-forward: train on past, test on future, expand; respects temporal order

Answer

Standard k-fold with time-based stratification

Answer

Shuffle data then split

Answer

Use last 10% as validation

Question 6

What is a rolling window feature in ML for time series?

Accepted Answer

Statistics over fixed past window (mean/std/lag) to create tabular features from time series

Answer

Average over the entire series

Answer

Future values averaged

Answer

Exponential smoothing of features

Question 7

What is Prophet and what makes it suitable for business forecasting?

Accepted Answer

Additive model with piecewise trend + Fourier seasonality + holiday effects; automatic changepoints

Answer

Neural network for time series

Answer

Uses ARIMA internally

Answer

Only for daily data

Question 8

What is temporal convolutional network (TCN) for time series?

Accepted Answer

Causal dilated 1D convolutions with exponentially growing dilation for large receptive field

Answer

Standard 2D convolution on time-frequency spectrogram

Answer

Requires sequential computation like LSTM

Answer

Only for classification tasks

Question 9

What is a vector autoregression (VAR) model?

Accepted Answer

Each variable regressed on its own and all other variables' past lags simultaneously

Answer

One ARIMA per variable independently

Answer

Only for two time series

Answer

Uses only own past, not cross-series lags

Question 10

What is cointegration in time series?

Accepted Answer

Linear combination of non-stationary series is stationary; they share long-run equilibrium

Answer

Both series are stationary

Answer

Correlation coefficient is constant over time

Answer

Only applies to stationary series

Question 11

What is anomaly detection in time series?

Accepted Answer

Points deviating from expected patterns; detected via z-score, forecast error, or reconstruction error

Answer

Only defined by z-score thresholds

Answer

Requires labeled anomaly examples always

Answer

Only works on stationary series

Question 12

What is the N-BEATS architecture for time series?

Accepted Answer

FC blocks with backcast+forecast outputs; trend/seasonality stacks with polynomial/Fourier bases

Answer

LSTM with attention

Answer

Uses ARIMA internally

Answer

Only for multivariate forecasting

Question 13

What is the Transformer for time series (Informer, Autoformer)?

Accepted Answer

Sparse attention or auto-correlation based Transformers for long time series forecasting

Answer

Standard Transformer unchanged for time series

Answer

Only for classification tasks

Answer

Replace time series with images first

Question 14

What is the naïve forecast and seasonal naïve forecast?

Accepted Answer

Naïve: last value; seasonal naïve: same period last year; baseline for comparison

Answer

Always the best model

Answer

Same as exponential smoothing

Answer

Uses moving average

Question 15

What is changepoint detection in time series?

Accepted Answer

Identifies times where statistical properties (mean, variance, trend) change

Answer

Detects seasonality only

Answer

Same as anomaly detection

Answer

Only for stationary series

Question 16

What is conformal prediction for time series?

Accepted Answer

Distribution-free intervals with guaranteed coverage using calibration residual quantiles

Answer

Requires Gaussian residuals

Answer

Only for point forecasting

Answer

Based on bootstrap confidence intervals

Question 17

What is the matrix profile for time series?

Accepted Answer

Distance of each subsequence to its nearest neighbor; enables motif/discord/anomaly detection

Answer

Moving average of the series

Answer

Covariance matrix of lags

Answer

Distance between full series pairs

Question 18

What is point forecast vs probabilistic forecast?

Accepted Answer

Point: single value; probabilistic: intervals/quantiles/distributions quantifying uncertainty

Answer

Probabilistic means using ML models only

Answer

Prediction intervals only for Gaussian errors

Answer

MAPE evaluates probabilistic forecasts

Question 19

What is the difference between AR and MA processes?

Accepted Answer

AR: weighted past values, PACF cuts off at p; MA: weighted past errors, ACF cuts off at q

Answer

AR uses past errors; MA uses past values

Answer

Both have the same ACF pattern

Answer

ARMA = AR + AR

Question 20

What is the global vs local forecasting approach?

Accepted Answer

Local: one model per series; global: one model across all series for cross-series sharing

Answer

Global models are always worse

Answer

Local models require more data

Answer

They are identical for long series

Question 21

What is feature importance for time series ML models?

Accepted Answer

Gain/permutation importance or SHAP values show which temporal features (lags, rolling stats) drive predictions

Answer

Only log-likelihood based

Answer

Cannot be computed for time series

Answer

Only applies to linear models

Question 22

What is the autocorrelation function (ACF) and how is it plotted?

Accepted Answer

Correlation vs. lag plot; spikes outside ±1.96/√T are significant; cuts off at q for MA, tails for AR

Answer

Total variance vs lag

Answer

Only used for stationary series of known distribution

Answer

ACF = PACF always

Question 23

What is the Box-Jenkins methodology?

Accepted Answer

Identify (ACF/PACF) → estimate → diagnose residuals; repeat until white noise residuals

Answer

Only identification step needed

Answer

Estimate first then identify

Answer

Uses cross-validation not AIC

Question 24

What is Dynamic Time Warping (DTW)?

Accepted Answer

Finds optimal elastic alignment between series via dynamic programming; handles time shift/stretch

Answer

Simple point-to-point distance

Answer

Only for same-length series

Answer

Requires normalization always

Question 25

What is a seasonal decomposition and how is STL better?

Accepted Answer

STL uses LOESS for robust, flexible decomposition; allows time-varying seasonality; outlier-resistant

Answer

Classical decomposition is always better

Answer

STL requires stationarity

Answer

Only for multiplicative models

Question 26

What is the ljung-box test and what does it check?

Accepted Answer

Tests if residual autocorrelations are collectively zero (white noise); p>0.05 = good fit

Answer

Tests for normality of residuals

Answer

Detects unit roots

Answer

Same as Durbin-Watson test

Test yourself