4.6.3.1 Forecasting Task Types

This topic describes the types of forecasting models supported by AutoML.

Before You Begin

• Review the list of supported Forecasting Models.

You can create the following types of forecasting models.

Univariate Models

In a univariate model, you define one numeric column as an endogenous variable, specified as a JSON_ARRAY. This is the target column that AutoML forecasts. For example, you forecast the rainfall for the next month by using the past daily rainfall as an endogenous variable.

Multivariate Models

In a multivariate model, you define multiple columns as endogenous variables, specified as a JSON_ARRAY. You must define one of these columns as the target column (the column with ground truth values). For example, you forecast the rainfall for the next month by using the past rainfall, temperature highs and lows, atmospheric pressure, and humidity. The target column is rainfall.

Univariate and Multivariate Models with Exogenous Variables

You have the option to define exogenous variables for univariate and multivariate models. These columns have independent, non-forecast, predictive variables. For example, you forecast future sales and use weather conditions like rainfall and high and low daily temperature values as exogenous variables.

Selecting Forecasting Models

To specify which models that are considered for training, use the model_list option and enter the appropriate model names. If only one model is set for model_list, then only that model is considered. Review the list of supported Forecasting Models and which type of model they support, univariate endogenous models, univariate endogenous models with exogenous variables, and multivariate endogenous models with exogenous variables. .

If the model_list option is not set, then ML_TRAIN considers all supported models during the algorithm selection stage. If options includes exogenous_variables, all supported models are still considered, including models that do not support exogenous_variables.

For example, if options includes univariate endogenous_variables with exogenous_variables, then ML_TRAIN considers NaiveForecaster, ThetaForecaster, ExpSmoothForecaster, ETSForecaster, STLwESForecaster, STLwARIMAForecaster, SARIMAXForecaster, and OrbitForecaster. ML_TRAIN ignores exogenous_variables if the model does not support them.

Similarly, if options includes multivariate endogenous_variables with exogenous_variables, then ML_TRAIN considers VARMAXForecaster and DynFactorForecaster.

If options also includes include_column_list, this forces ML_TRAIN to only consider those models that support exogenous_variables.

What's Next

• Learn more about Prediction Intervals.

• Learn how to Train a Forecasting Model.