ML_Techniques_I Week 3 Session 1

An exogenous variable in forecasting is one that is

• Determined inside the system and influences itself

• Always a lag of the target variable

• Independent of other variables in the system but affects the output

• Ignored during model fitting

A 1-D Convolutional Neural Network (CNN) is especially useful for forecasting because it can:

• Guarantee stationarity

• Extract local temporal patterns with shared filters

• Learn global hierarchical relationships

• Model infinite long-range dependencie

The term forecast horizon refers to:

• The interval between observations

• The length of the training window

• The number of features in the model

• How far into the future predictions are required

Creating lag features and rolling means for a univariate series is an example of

• Feature engineering steps

• Dimensionality reduction

• Hyper-parameter optimisation

• Differencing

Multivariate time-series forecasting involves

• Modelling only seasonal components

• Predicting many future horizons of one series

• Using multiple correlated variables to predict one or more targets

• Clustering similar sequences without a target

Which property must a series satisfy for most classical statistical models (like ARIMA to be directly applicable without transformation?

• Normality

• Weak stationarity

• Uniform spacing

• High frequency

Seasonal differencing is most often used to

• Stabilise variance

• Remove linear trend

• Eliminate repeating seasonal patterns

• Create exogenous regressors

Resampling an irregularly spaced series to a uniform time grid is essential because it

• Eliminates all missing values automatically

• Increases model complexity

• Guarantees the data are stationary

• Ensures each observation represents the same time interval, enabling most forecasting algorithms