4

I created a model on big query, Is it possible to include additional columns as external regressors ?

For example I'd like to include Date, Users, page per session, bounce rate etc. for forecasting users.

create or replace model bqml_tutorial.create_model
options
(model_type= 'ARIMA_PLUS',
time_series_timestamp_col='Date',
time_series_data_col='Users',
auto_arima=True,
data_frequency = 'AUTO_FREQUENCY',
decompose_time_series= True)
as
select Date, cv as Users from `bqml_tutorial.cvrate` ORDER BY Date
Cylldby
  • 1,783
  • 1
  • 4
  • 17
Vishvas Chauhan
  • 240
  • 2
  • 10

1 Answers1

1

Looking at the documentation this is currently not available. The ARIMA_PLUS model that you can train in BigQuery already does a lot of things (seasonality study, outlier removal, missing data interpolation, etc). But in terms of external regressors, you cannot add specific columns for training your model.

The only additional data you can put into your model is the holiday information (with HOLIDAY_REGION option). Which is already awesome !

Note that you can train models for multiple time series at the same time by specifying the columns you want to forecast with the TIME_SERIES_ID_COL parameter. But that will get you forecasts for all these columns from independant models (therefore the effect from one column on the other will not be modelled).

Cylldby
  • 1,783
  • 1
  • 4
  • 17
  • That is helpful. Is there any model suggestion for other columns dependencies training and predicting 1 target col? – Vishvas Chauhan May 24 '21 at 16:43
  • For time series, the only dedicated kind of model available in BQ is the ARIMA model, which can serve as a baseline model in your case. As for investigating the effect of other columns, maybe you can have a look at XGBOOST and AUTOML models ? But be careful of your methodology ! – Cylldby May 25 '21 at 14:36