The ML.ARIMA_EVALUATE function

This document describes theML.ARIMA_EVALUATE function, which you can useto evaluate the model metrics ofARIMA_PLUS orARIMA_PLUS_XREG time seriesmodels.

Syntax

ML.ARIMA_EVALUATE(  MODEL `PROJECT_ID.DATASET.MODEL_NAME`,  [, STRUCT(SHOW_ALL_CANDIDATE_MODELS AS show_all_candidate_models)])

• PROJECT_ID: Your project ID.
• DATASET: The BigQuery dataset that containsthe model.
• MODEL_NAME: The name of the model.

• SHOW_ALL_CANDIDATE_MODELS: aBOOL value thatindicates whether to show evaluation metrics or an error message for eitherall candidate models or for only the best model with the lowestAkaike information criterion (AIC).

  When theSHOW_ALL_CANDIDATE_MODELS argument value isFALSE, theML.ARIMA_EVALUATE function returns evaluation metrics for only the bestmodel. When theSHOW_ALL_CANDIDATE_MODELS argument value isTRUE,metrics are returned for all candidate models, along with a possible fittingerror on a setting ofnon_seasonal_p,non_seasonal_d,non_seasonal_q,anddrift; this applies to both single time series training withauto.ARIMA and large-scale time series training cases.

  For large-scale time series forecasting training, regardless of the value oftheSHOW_ALL_CANDIDATE_MODELS argument, a single row is returned for atime series for which there is no a valid model. The error message explainsthe reason, and the values of all other columns areNULL.

  For single time seriesARIMA_PLUS orARIMA_PLUS_XREG models, the defaultvalue isTRUE. For large-scale time seriesARIMA_PLUS models, thedefault value isFALSE.

Output

ML.ARIMA_EVALUATE returns the following columns:

• time_series_id_col ortime_series_id_cols: the identifiers of a timeseries. Only present when forecasting multiple time series at once. The columnnames and types are inherited from theTIME_SERIES_ID_COL optionas specified in the model creation query.
• non_seasonal_p: anINT64 value that contains the order of theautoregressive part in a non-seasonal ARIMA model.
• non_seasonal_d: anINT64 value that contains the degree of firstdifferencing involved in a non-seasonal ARIMA model.
• non_seasonal_q: anINT64 value that contains the order of the movingaverage part in a non-seasonal ARIMA model.
• has_drift: aBOOL value that indicates whether the model has drift.
• log_likelihood: aFLOAT64 value that contains thelog-likelihood of the model.
• AIC: the AIC of the model.
• variance: aFLOAT64 value that contains the variance of the model.

• seasonal_periods: anARRAY<STRING> value that contains one or more ofthe following values:

  • DAILY
  • WEEKLY
  • MONTHLY
  • QUARTERLY
  • YEARLY
  • NO_SEASONALITY

• has_holiday_effect: aBOOL value that indicates whether the history datahas a holiday effect.

• has_spikes_and_dips: aBOOL value that indicates whether the history datahas spikes and dips.

• has_step_changes: aBOOL value that indicates whether the model hasstep changes.

• error_message: aSTRING value that contains the error message raised ifany time series fail in the model.

Thehas_holiday_effect,has_spikes_and_dips, andhas_step_changes columnsare only populated forARIMA_PLUS models that havedecompose_time_seriesenabled.

All of the columns are specific to the fittedARIMA models except for thefollowing columns:

• time_series_id_col
• time_series_id_cols
• seasonal_periods
• has_holiday_effect
• has_spikes_and_dips
• has_step_changes

When thenon_seasonal_d value is not1,has_drift is set toFALSE bydefault, becausehas_drift doesn't apply in those cases.

Example

The following example retrieves the evaluation metrics of the best model fromthe modelmydataset.mymodel in your default project:

SELECT*FROMML.ARIMA_EVALUATE(MODEL`mydataset.mymodel`,STRUCT(FALSEASshow_all_candidate_models))

What's next

• For information about model evaluation, seeBigQuery ML model evaluation overview.
• For more information about supported SQL statements and functions for timeseries forecasting models, seeEnd-to-end user journeys for time series forecasting models.