Use custom holidays in a time-series forecasting model

This tutorial shows you how to do the following tasks:

Create an ARIMA_PLUS time-series forecasting model that uses only built-in holidays.
Create an ARIMA_PLUS time-series forecasting model that uses custom holidays in addition to built-in holidays.
Visualize the forecasted results from these models.
Inspect a model to see which holidays it models.
Evaluate the effects of the custom holidays on the forecasted results.
Compare the performance of the model that uses only built-in holidays to the performance of the model that uses custom holidays in addition to built-in holidays.

This tutorial uses the bigquery-public-data.wikipedia.pageviews_* public tables.

Required permissions

To create the dataset, you need the bigquery.datasets.create IAM permission.
To create the connection resource, you need the following permissions:
- bigquery.connections.create
- bigquery.connections.get
To create the model, you need the following permissions:
- bigquery.jobs.create
- bigquery.models.create
- bigquery.models.getData
- bigquery.models.updateData
- bigquery.connections.delegate
To run inference, you need the following permissions:
- bigquery.models.getData
- bigquery.jobs.create

For more information about IAM roles and permissions in BigQuery, see Introduction to IAM.

Costs

In this document, you use the following billable components of Google Cloud:

BigQuery: You incur costs for the data you process in BigQuery.

To generate a cost estimate based on your projected usage, use the pricing calculator. New Google Cloud users might be eligible for a free trial.

For more information, see BigQuery pricing.

Before you begin

Sign in to your Google Cloud account. If you're new to Google Cloud, create an account to evaluate how our products perform in real-world scenarios. New customers also get $300 in free credits to run, test, and deploy workloads.

In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

Go to project selector

Make sure that billing is enabled for your Google Cloud project.

Enable the BigQuery API.

Enable the API

In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

Go to project selector

Make sure that billing is enabled for your Google Cloud project.

Enable the BigQuery API.

Enable the API

Create a dataset

Create a BigQuery dataset to store your ML model:

In the Google Cloud console, go to the BigQuery page.

Go to the BigQuery page
In the Explorer pane, click your project name.
Click View actions > Create dataset.
On the Create dataset page, do the following:
- For Dataset ID, enter bqml_tutorial.
- For Location type, select Multi-region, and then select US (multiple regions in United States).
  
  The public datasets are stored in the US multi-region. For simplicity, store your dataset in the same location.
- Leave the remaining default settings as they are, and click Create dataset.

Prepare the time-series data

Aggregate the Wikipedia page view data for the Google I/O page into a single table, grouped by day:

Go to the BigQuery page.

Go to BigQuery

In the SQL editor pane, run the following SQL statement:

CREATE OR REPLACE TABLE `bqml_tutorial.googleio_page_views`
AS
SELECT
  DATETIME_TRUNC(datehour, DAY) AS date,
  SUM(views) AS views
FROM
  `bigquery-public-data.wikipedia.pageviews_*`
WHERE
  datehour >= '2017-01-01'
  AND datehour < '2023-01-01'
  AND title = 'Google_I/O'
GROUP BY
  DATETIME_TRUNC(datehour, DAY)

Create a time-series forecasting model that uses built-in holidays

Create a model that forecasts daily page views for the Wikipedia "Google I/O" page, based on page view data before 2022 and taking built-in holidays into account:

Go to the BigQuery page.

Go to BigQuery

In the SQL editor pane, run the following SQL statement:

CREATE OR REPLACE MODEL `bqml_tutorial.forecast_googleio`
  OPTIONS (
    model_type = 'ARIMA_PLUS',
    holiday_region = 'US',
    time_series_timestamp_col = 'date',
    time_series_data_col = 'views',
    data_frequency = 'DAILY',
    horizon = 365)
AS
SELECT
  *
FROM
  `bqml_tutorial.googleio_page_views`
WHERE
  date < '2022-01-01';

Visualize the forecasted results

After you create the model using built-in holidays, join the original data from the bqml_tutorial.googleio_page_views table with the forecasted value from the ML.EXPLAIN_FORECAST function, and then visualize it by using Looker Studio:

Go to the BigQuery page.

Go to BigQuery

In the SQL editor pane, run the following SQL statement:

SELECT
  original.date,
  original.views AS original_views,
  explain_forecast.time_series_adjusted_data
    AS adjusted_views_without_custom_holiday,
FROM
  `bqml_tutorial.googleio_page_views` original
INNER JOIN
  (
    SELECT
      *
    FROM
      ML.EXPLAIN_FORECAST(
        MODEL `bqml_tutorial.forecast_googleio`,
        STRUCT(365 AS horizon))
  ) explain_forecast
  ON
    TIMESTAMP(original.date)
    = explain_forecast.time_series_timestamp
ORDER BY
  original.date;

In the Query results pane, click Explore data, and then click Explore with Looker Studio. Looker Studio opens in a new tab.
In the Looker Studio tab, click Add a chart, and then click the time series chart:

Place the chart on the report.
On the Setup tab of the Chart pane, click Add metric and select adjusted_views_without_custom_holiday:

The chart looks similar to the following:

You can see that the forecasting model captures the general trend pretty well. However, it isn't capturing the increased traffic related to previous Google I/O events, and it isn't able to generate an accurate forecast for
1. The next sections show you how to deal with some of these limitations.

Create a time-series forecasting model that uses built-in holidays and custom holidays

As you can see in Google I/O history, the Google I/O event occurred on different dates between 2017 and 2022. To take this variation into account, create a model that forecasts page views for the Wikipedia "Google_I/O" page through 2022, based on page view data from before 2022, and using custom holidays to represent the Google I/O event each year. In this model, you also adjust the holiday effect window to cover three days around the event date, to better capture some potential page traffic before and after the event.

Go to the BigQuery page.

Go to BigQuery

In the SQL editor pane, run the following SQL statement:

CREATE OR REPLACE MODEL `bqml_tutorial.forecast_googleio_with_custom_holiday`
  OPTIONS (
    model_type = 'ARIMA_PLUS',
    holiday_region = 'US',
    time_series_timestamp_col = 'date',
    time_series_data_col = 'views',
    data_frequency = 'DAILY',
    horizon = 365)
AS (
  training_data AS (
      SELECT
        *
      FROM
        `bqml_tutorial.googleio_page_views`
      WHERE
        date < '2022-01-01'
    ),
  custom_holiday AS (
      SELECT
        'US' AS region,
        'GoogleIO' AS holiday_name,
        primary_date,
        1 AS preholiday_days,
        2 AS postholiday_days
      FROM
        UNNEST(
          [
            DATE('2017-05-17'),
            DATE('2018-05-08'),
            DATE('2019-05-07'),
            -- cancelled in 2020 due to pandemic
            DATE('2021-05-18'),
            DATE('2022-05-11')])
          AS primary_date
    )
);

Visualize the forecasted results

After you create the model using custom holidays, join the original data from the bqml_tutorial.googleio_page_views table with the forecasted value from the ML.EXPLAIN_FORECAST function, and then visualize it by using Looker Studio:

Go to the BigQuery page.

Go to BigQuery

In the SQL editor pane, run the following SQL statement:

SELECT
  original.date,
  original.views AS original_views,
  explain_forecast.time_series_adjusted_data
    AS adjusted_views_with_custom_holiday,
FROM
  `bqml_tutorial.googleio_page_views` original
INNER JOIN
  (
    SELECT
      *
    FROM
      ML.EXPLAIN_FORECAST(
        MODEL
          `bqml_tutorial.forecast_googleio_with_custom_holiday`,
        STRUCT(365 AS horizon))
  ) explain_forecast
  ON
    TIMESTAMP(original.date)
    = explain_forecast.time_series_timestamp
ORDER BY
  original.date;

In the Query results pane, click Explore data, and then click Explore with Looker Studio. Looker Studio opens in a new tab.
In the Looker Studio tab, click Add a chart, click the time series chart, and place the chart on the report.
On the Setup tab of the Chart pane, click Add metric and select adjusted_views_with_custom_holiday.

The chart looks similar to the following:

As you can see, the custom holidays boosted the performance of the forecasting model. It now effectively captures the increase of page views caused by Google I/O.

Inspect holiday information

Inspect the list of holidays that were taken into account during modeling by using the ML.HOLIDAY_INFO function:

Go to the BigQuery page.

Go to BigQuery
In the SQL editor pane, run the following SQL statement:
```
SELECT *
FROM
  ML.HOLIDAY_INFO(
    MODEL `bqml_tutorial.forecast_googleio_with_custom_holiday`);
```
The results show both Google I/O and the built-in holidays in the list of holidays:

Evaluate the effects of the custom holidays

Evaluate the effects of the custom holidays on the forecasted results by using the ML.EXPLAIN_FORECAST function:

Go to the BigQuery page.

Go to BigQuery

In the SQL editor pane, run the following SQL statement:

SELECT
  time_series_timestamp,
  holiday_effect_GoogleIO,
  holiday_effect_US_Juneteenth,
  holiday_effect_Christmas,
  holiday_effect_NewYear
FROM
  ML.EXPLAIN_FORECAST(
    model
      `bqml_tutorial.forecast_googleio_with_custom_holiday`,
    STRUCT(365 AS horizon))
WHERE holiday_effect != 0;

The results show that Google I/O contributes a large amount of holiday effect to the forecasted results:

Results from the ML.EXPLAIN_FORECAST function.

Compare model performance

Use the ML.EVALUATE function to compare the performance of the first model created without custom holidays and the second model created with custom holidays. To see how the second model performs when it comes to forecasting a future custom holiday, set the time range to the week of Google I/O in 2022:

Go to the BigQuery page.

Go to BigQuery

In the SQL editor pane, run the following SQL statement:

SELECT
  "original" AS model_type,
  *
FROM
  ml.evaluate(
    MODEL `bqml_tutorial.forecast_googleio`,
    (
      SELECT
        *
      FROM
        `bqml_tutorial.googleio_page_views`
      WHERE
        date >= '2022-05-08'
        AND date < '2022-05-12'
    ),
    STRUCT(
      365 AS horizon,
      TRUE AS perform_aggregation))
UNION ALL
SELECT
  "with_custom_holiday" AS model_type,
  *
FROM
  ml.evaluate(
    MODEL
      `bqml_tutorial.forecast_googleio_with_custom_holiday`,
    (
      SELECT
        *
      FROM
        `bqml_tutorial.googleio_page_views`
      WHERE
        date >= '2022-05-08'
        AND date < '2022-05-12'
    ),
    STRUCT(
      365 AS horizon,
      TRUE AS perform_aggregation));

The results show that the second model offers a significant performance improvement:

Results from the ML.EXPLAIN_FORECAST function.

Clean up

In the Google Cloud console, go to the Manage resources page.
Go to Manage resources
In the project list, select the project that you want to delete, and then click Delete.
In the dialog, type the project ID, and then click Shut down to delete the project.