BigQuery ML forecast using ARIMA_PLUS (mostly) ignores the holiday effect

I'm trying to forecast daily sales for a business using the past 5 years of daily sales history. This business has very clear holiday sales patterns, including zero sales on Easter/Thanksgiving/Christmas and unusually high sales on Valentine's Day, Mother's Day, etc.

Only two dates in the resulting model that have a non-zero holiday effect: December 23 and President's Day. I cannot figure out why.

I'm using the ARIMA_PLUS model type in BigQuery ML, with the holiday_region set to 'US'. The data is simply date and sales.

CREATE OR REPLACE MODEL `model_name`
OPTIONS(MODEL_TYPE='ARIMA_PLUS',
    time_series_timestamp_col='date',
    time_series_data_col='sales',
    data_frequency='DAILY',
    holiday_region='US') AS
( SELECT date, sales FROM `table_name`);

An example of an obvious forecasting error is that April 9, 2023 (Easter) is forecasted as a normal sales day (instead of zero), and April 17, 2023 is forecasted to be zero. The business was closed on April 17, 2022 due to Easter, so the model is clearly ignoring the Easter holiday and providing an inaccurate forecast as a result.

I used the EXPLAIN_FORECAST function to see if any days had a holiday effect, which is where I found it was just December 23 and President's Day. There were adjustments for those two days across each of the five years of history. The holiday_effect field was zero for the rest of the dates.

SELECT * FROM ML.EXPLAIN_FORECAST(MODEL `model_name`, STRUCT(14 AS horizon, 0.8 AS confidence_level));

Any idea why the model ignores all holidays except these two? Is this a defect in BigQuery ML, or is there something wrong with my setup?

The dip on 2023-04-17 might be a result of YEARLY seasonality effect.

In your model creation setting, SEASONALITY is default to AUTO. Given that your data frequency is DAILY and time series length is longer than 2 years, it will automatically infer the SEASONALITY as WEEKLY and YEARLY. It is indeed the YEARLY seasonality that captures the dip from 2022 this time.

To workaround that, you can set SEASONALITY = ['WEEKLY'] in OPTIONS clause explicitly to avoid modeling with YEARLY seasonality.