Time series¶

Batch predictions for time series models work without any additional configuration. However, in most cases you need to either modify the default configuration or prepare the prediction dataset.

Time series batch prediction settings¶

The default configuration can be overridden using the timeseriesSettings job configuration property:

Here is a complete example job:

{
    "deploymentId": "5f22ba7ade0f435ba7217bcf",
    "intakeSettings": {"type": "localFile"},
    "outputSettings": {"type": "localFile"},
    "timeseriesSettings": {
        "type": "historical",
        "predictionsStartDate": "2020-01-01",
        "predictionsEndDate": "2020-03-31"
    }
}

An example using Python SDK:

import datarobot as dr

dr.Client(
    endpoint="https://app.datarobot.com/api/v2",
    token="...",
)

deployment_id = "..."

input_file = "to_predict.csv"
output_file = "predicted.csv"

job = dr.BatchPredictionJob.score_to_file(
    deployment_id,
    input_file,
    output_file,
    timeseries_settings={
        "type": "historical",
        "predictions_start_date": "2020-01-01",
        "predictions_end_date": "2020-03-31",
    },
)

print("started scoring...", job)
job.wait_for_completion()

Prediction type¶

When using forecast mode, DataRobot makes predictions using forecastPoint or rows in the dataset without a target. In historical mode, DataRobot enables bulk predictions, which calculates predictions for all possible forecast points and forecast distances within predictionsStartDate and predictionsEndDate range.

Requirements for the scoring dataset¶

To ensure the Batch Prediction API can process your time series dataset, you must configure the following:

• Sort prediction rows by their timestamps, with the earliest row first.
  • If using multiseries, the prediction rows must be sorted by series ID then timestamp.
• There is no limit on the number of series DataRobot supports. The only limit is the job timeout as mentioned in Limits.

Single series forecast dataset example¶

The following is an example forecast dataset for a single series:

Multiseries forecast dataset example¶

If scoring multiple series, the data must be ordered by series and timestamp: