Prophet is a forecasting model maintained by Meta. See theProphet paperfor algorithm details and thedocumentation for more information about the library.

LikeBigQuery ML ARIMA_PLUS,Prophet attempts to decompose each time series into trends, seasons, and holidays,producing a forecast using the aggregation of these models' inferences. Animportant difference, however, is that BQML ARIMA+ uses ARIMA to model the trendcomponent, while Prophet attempts to fit a curve using a piecewise logistic orlinear model.

Google Cloud offers a pipeline for training a Prophet model and a pipeline forgetting batch inferences from a Prophet model. Both pipelines are instances ofVertex AI Pipelines fromGoogle Cloud Pipeline Components (GCPC).

Integration of Prophet with Vertex AI means that you can do the following:

• Use Vertex AIdata splitting andwindowing strategies.
• Read data from either BigQuery tables or CSVs stored inCloud Storage. Vertex AI expects each row to have the sameformat asVertex AI Forecasting.

Although Prophet is a multivariate model, Vertex AI supportsonly a univariate version of it.

To learn about the service accounts this workflow uses, seeService accounts for Tabular Workflows.

Workflow APIs

This workflow uses the following APIs:

• Vertex AI
• Dataflow
• BigQuery
• Cloud Storage

Train a model with Prophet

Prophet is designed for a single time series. Vertex AI aggregatesdata by time series ID and trains a Prophet model for each time series. Themodel training pipeline performs hyperparameter tuning usinggrid searchand Prophet's built-in backtesting logic.

To support multiple time series, the pipeline uses aVertex AICustom Training JobandDataflowto train multiple Prophet models in parallel. Overall, the number of modelstrained is the product of the number of time series and the number of hyperparameter tuning trials.

The following sample code demonstrates how to run a Prophet model training pipeline:

job = aiplatform.PipelineJob(    ...    template_path=train_job_spec_path,    parameter_values=train_parameter_values,    ...)job.run(service_account=SERVICE_ACCOUNT)

The optionalservice_account parameter injob.run() lets you set theVertex AI Pipelines service account to an account of your choice.

The pipeline and the parameter values are defined by the following function.

(    train_job_spec_path,    train_parameter_values,) = utils.get_prophet_train_pipeline_and_parameters(    ...)

The following is a subset ofget_prophet_train_pipeline_and_parameters parameters:

Dataflow parameters

The following is a subset ofget_prophet_train_pipeline_and_parameters parameters for Dataflow customization:

Because Prophet training jobs run on Dataflow, an initialstartup time of 5 - 7 minutes occurs. To reduce additional runtime, you can scaleup or scale out. For example, to scale up, change themachine type fromn1-standard-1 toe2-highcpu-8. To scale out,increase the number of workers from1 to200.

Data split parameters

The training pipeline offers the following options for splitting your data:

You define the data split parameters inget_prophet_train_pipeline_and_parameters as follows:

Window parameters

Vertex AI generates forecast windows from the input data using a rolling windowstrategy. If you leave the window parameters unset, Vertex AI usesthe Count strategy with a default maximum value of100,000,000. The trainingpipeline offers the following rolling window strategies:

You define the window parameters inget_prophet_train_pipeline_and_parameters as follows:

Make inferences with Prophet

The Vertex AImodel training pipeline for Prophet createsone Prophet model for each time series in the data. The inference pipelineaggregates input data by time series ID and calculates the inferences separatelyfor each time series. The pipeline then disaggregates the inference results tomatch the format ofVertex AI Forecasting.

The following sample code demonstrates how to run a Prophet inference pipeline:

job = aiplatform.PipelineJob(    ...    template_path=prediction_job_spec_path,    parameter_values=prediction_parameter_values,    ...)job.run(...)

The pipeline and the parameter values are defined by the following function.

(    prediction_job_spec_path,    prediction_parameter_values,) = utils.get_prophet_prediction_pipeline_and_parameters(    ...)

The following is a subset ofget_prophet_prediction_pipeline_and_parameters parameters: